I S R U P T IN

DE THERAP RUG DIGSC

EUTIC D
N

OV
U
IMM

ERY

OX I I E S
J A X NGL E

OT T U D
CI T Y
A SI
CY T

KIN
ON N

S
O

ST E R
UD AT I
Y TO ELEASE EVALU IMMU
TEST EFFICACY AND

USE A SINGLE HUMAN PBMC DONOR USE MULTIPLE HUMAN PBMC DONORS
to perform dose ranging studies and to recapitulate donor-to-donor variability
to allow lead selection. you’d expect to find in the clinic.

CYTOKINE RELEASE BUILD ON THE OPTIMAL IMMUNODEFICIENT

PLATFORM

EVALUATION STUDIES to investigate your specific targets of interest.

Utilize this platform to evaluate

immunomodulation of human immune cells
YOUR HUMANIZED PLATFORM
against human specific targets. This assay is ideal for evaluating relative cytokine levels from
provides a fast, accurate, and reliable method to immunomodulating therapies and for assessing
anti-inflammatory therapies.
screen novel immunotherapies for efficacy and
cytokine release syndrome.

SHAPE YOUR DISCOVERY AT

JAX.ORG/CRES
 Reprint compendium
CANCERCELBO2020

Cancer Cell
Cancer Cell Best of 2020

Best of 2020

Most widely read papers of 2020

Cancer immunotherapy can save lives
Partner with our experts to accelerate your breakthroughs
Explore Cancer Applications
We are committed to providing researchers with the right
tools and applications to solve the complex puzzle that
is cancer. In addition to our immunoassays, recombinant
proteins, and flow cytometry reagents, we provide pathogen-
tested GoInVivo™️ antibodies that block immune checkpoints
and TotalSeq™️ antibodies for single-cell proteogenomics
analysis. By collaborating together, we can further our
understanding of cancer and how to effectively eliminate it.
Learn More: biolegend.com/en-us/cancer
World-Class Quality | Superior Customer Support | Outstanding Value
Our helpful website provides:
• On-demand webinars, including a recent talk by Dr. Vijay
Kuchroo investigating co-inhibitory receptors in anti-
tumor immunity.
• Scientist-supported reagents for applications, such as flow
cytometry, microscopy, and multiomics.
• Informative and educational content covering the basic
tenets of cancer biology.
BioLegend products are manufactured in an ISO 13485:2016-
certified facility to ensure the highest quality standards.
biolegend.com
Be Brilliant. Single Molecule
Localization / Voltage Sensitive
Dyes / Lightsheet Microscopy /
Spinning Disk Confocal / Live Cell
Fluorescence / Total Internal
Reflection Fluorescence / Calcium
Imaging / Super Resolution /
Expansion Microscopy /
Fluorescence Resonance Energy
Transfer / Structured Illumination /
Multi-dimensional Imaging /
Optogenetics / Point Accumulation
for Imaging in Nanoscale
Topography. Do Dim Things.

H A M A M AT S U C A M E R A S . C O M
Foreword
Cancer Cell provides a high-profile forum for major advances across cancer research and oncology. Cancer
Cell reaches a broad audience of cancer biologists and clinicians to facilitate the translation of biological
discoveries to clinical cancer care. Articles published in Cancer Cell cover a wide range of cancer research,
including clinical trials, translational research, and basic cancer biology studies. We introduced the Best of
Cancer Cell series in 2012 to highlight some of the exciting articles published in the previous year, and we are
pleased to now present the tenth installment, Best of Cancer Cell 2020, which will be presented at the virtual
2021 AACR annual meeting.

For this edition, we selected two reviews and eight full-length articles that strongly captured our readership’s
attention and are among our most-accessed content in 2020. In addition, these papers represent some of the
most exciting areas in cancer research: they allow us to better understand the complexity and heterogeneity
of cancer as well as its progression and response to therapies, and they have strong implications for patient
diagnosis, stratification, and treatment. These articles also showcase a range of cancer models, as well as
cutting-edge technologies and computational algorithms used in cancer research. Finally, we included one of
our series of Voices, in which several experts share techniques used to visualize cancer in the lab and in the
clinic.

We hope that you enjoy reading this collection. We recognize that no single measurement can be accurately
indicative of “the best” research papers over a given period of time. This is especially true for recently published
research, because the community hasn’t had enough time to fully appreciate the relative importance of the
findings. However, we are confident that you will find this collection informative and exciting.

We are grateful for the generosity of our sponsors, who helped to make this reprint collection possible. You can
access the entire Best of Cell Press collection online at www.cell.com/bestof. Visit www.cell.com/cancer-cell
to learn about the latest findings that Cancer Cell has had the privilege to publish and www.cell.com to find
other high-quality cancer-relevant papers published in the full portfolio of Cell Press journals. Please feel free
to contact us at cancer@cell.com to tell us about your latest work or to provide feedback. We look forward to
working with you in 2021 and beyond!

The Cancer Cell editorial team

For information for the Best of Cell Press Series, please contact:
Jonathan Christison
Program Director, Best of Cell Press
e: jchristison@cell.com
p: 617-397-2893
t: @CellPressBiz

Cancer Cell ll
OriGene TrueMAB 8.25 X 11_Nature_PRINT.indd 1 12/19/18 2:47 PM
Accelerate Your Antibody Development
and Still Have Peace of Mind
Gain More Insights, Make Better Decisions.
Sartorius accelerates development of breakthrough therapeutics with
groundbreaking, comprehensive solutions for lab filtration, cell/protein
analysis, and more. Our products and technologies provide rapid,
multifactorial results to enable deeper insights and ensure the choice,
quality, and integrity of your biopharmaceutical product.

Advance time to results with:

ƒ Process streamlining
ƒ Multiplexing
ƒ Information-rich data
ƒ Biologically relevant insights

www.sartorius.com/mab-development
Cancer Cell
Best of 2020

Voices
Visualizing cancer Matthew Krummel, Kenneth Hu, Elisabeth G.E. de Vries,
Jacky G. Goetz, Diana Passaro, Sriram Subramaniam,
Tom Misteli, Melissa Skala, Jose Javier Bravo-Cordero,
and Julie L. Sutcliffe

Reviews
Intratumor heterogeneity: The Rosetta Stone of therapy Andriy Marusyk, Michalina Janiszewska, and Kornelia Polyak
resistance

Engineering CAR-T cells for next-generation cancer Mihe Hong, Justin D. Clubb, and Yvonne Y. Chen
therapy

Articles
Integrated molecular and clinical analysis of 1,000 Scott Ryall, Michal Zapotocky, Kohei Fukuoka, Liana Nobre,
pediatric low-grade gliomas Ana Guerreiro Stucklin, Julie Bennett, Robert Siddaway,
Christopher Li, Sanja Pajovic, Anthony Arnoldo,
Paul E. Kowalski, Monique Johnson, Javal Sheth,
Alvaro Lassaletta, Ruth G. Tatevossian, Wilda Orisme,
Ibrahim Qaddoumi, Lea F. Surrey, Marilyn M. Li,
Angela J. Waanders, Stephen Gilheeney, Marc Rosenblum,
Tejus Bale, Derek S. Tsang, Normand Laperriere,
Abhaya Kulkarni, George M. Ibrahim, James Drake,
Peter Dirks, Michael D. Taylor, James T. Rutka,
Suzanne Laughlin, Manohar Shroff, Mary Shago,
Lili-Naz Hazrati, Colleen D’Arcy, Vijay Ramaswamy,
Ute Bartels, Annie Huang, Eric Bouffet,
Matthias A. Karajannis, Mariarita Santi, David W. Ellison,
Uri Tabori, and Cynthia Hawkins

MYC drives temporal evolution of small cell lung cancer
subtypes by reprogramming neuroendocrine fate
Abbie S. Ireland, Alexi M. Micinski, David W. Kastner,
Bingqian Guo, Sarah J. Wait, Kyle B. Spainhower,
Christopher C. Conley, Opal S. Chen, Matthew R. Guthrie,
Danny Soltero, Yi Qiao, Xiaomeng Huang,
Szabolcs Tarapcsák, Siddhartha Devarakonda,
Milind D. Chalishazar, Jason Gertz, Justin C. Moser,
Gabor Marth, Sonam Puri, Benjamin L. Witt,
Benjamin T. Spike, and Trudy G. Oliver

ll
 Sharpen your senses.
FINE SURGICAL INSTRUMENTS FOR RESEARCH™

FINESCIENCE.COM • +1 800-521-2109 / +1 650-349-1636 • INFO@FINESCIENCE.COM

A probabilistic classification tool for genetic subtypes
of diffuse large B cell lymphoma with therapeutic
implications
Dendritic cell paucity leads to dysfunctional immune
surveillance in pancreatic cancer
Treatment-induced tumor dormancy through
YAP-mediated transcriptional reprogramming
of the apoptotic pathway
Predicting drug response and synergy using a deep
learning model of human cancer cells
Conserved interferon-γ signaling drives clinical response
to immune checkpoint blockade therapy in melanoma
George W. Wright, Da Wei Huang, James D. Phelan,
Zana A. Coulibaly, Sandrine Roulland, Ryan M. Young,
James Q. Wang, Roland Schmitz, Ryan D. Morin,
Jeffrey Tang, Aixiang Jiang, Aleksander Bagaev,
Olga Plotnikova, Nikita Kotlov, Calvin A. Johnson,
Wyndham H. Wilson, David W. Scott, and Louis M. Staudt
Samarth Hegde, Varintra E. Krisnawan, Brett H. Herzog,
Chong Zuo, Marcus A. Breden, Brett L. Knolhoff,
Graham D. Hogg, Jack P. Tang, John M. Baer, Cedric Mpoy,
Kyung Bae Lee, Katherine A. Alexander, Buck E. Rogers,
Kenneth M. Murphy, William G. Hawkins, Ryan C. Fields,
Carl J. DeSelm, Julie K. Schwarz, and David G. DeNardo
Kari J. Kurppa, Yao Liu, Ciric To, Tinghu Zhang,
Mengyang Fan, Amir Vajdi, Erik H. Knelson, Yingtian Xie,
Klothilda Lim, Paloma Cejas, Andrew Portell,
Patrick H. Lizotte, Scott B. Ficarro, Shuai Li, Ting Chen,
Heidi M. Haikala, Haiyun Wang, Magda Bahcall,
Yang Gao, Sophia Shalhout, Steffen Boettcher,
Bo Hee Shin, Tran Thai, Margaret K. Wilkens,
Michelle L. Tillgren, Mierzhati Mushajiang, Man Xu,
Jihyun Choi, Arrien A. Bertram, Benjamin L. Ebert,
Rameen Beroukhim, Pratiti Bandopadhayay, Mark M. Awad,
Prafulla C. Gokhale, Paul T. Kirschmeier, Jarrod A. Marto,
Fernando D. Camargo, Rizwan Haq, Cloud P. Paweletz,
Kwok-Kin Wong, David A. Barbie, Henry W. Long,
Nathanael S. Gray, and Pasi A. Jänne
Brent M. Kuenzi, Jisoo Park, Samson H. Fong,
Kyle S. Sanchez, John Lee, Jason F. Kreisberg, Jianzhu Ma,
and Trey Ideker
Catherine S. Grasso, Jennifer Tsoi, Mykola Onyshchenko,
Gabriel Abril-Rodriguez, Petra Ross-Macdonald,
Megan Wind-Rotolo, Ameya Champhekar, Egmidio Medina,
Davis Y. Torrejon, Daniel Sanghoon Shin, Phuong Tran,
Yeon Joo Kim, Cristina Puig-Saus, Katie Campbell,
Agustin Vega-Crespo, Michael Quist, Christophe Martignier,
Jason J. Luke, Jedd D. Wolchok, Douglas B. Johnson,
Bartosz Chmielowski, F. Stephen Hodi, Shailender Bhatia,
William Sharfman, Walter J. Urba, Craig L. Slingluff,
Jr., Adi Diab, John B.A.G. Haanen, Salvador Martin Algarra,
Drew M. Pardoll, Valsamo Anagnostou, Suzanne L. Topalian,
Victor E. Velculescu, Daniel E. Speiser, Anusha Kalbasi,
and Antoni Ribas
 See sequencing in a new light
Explore the NextSeq™ 2000

ZZZLOOXPLQDFRP1H[W6HT
For Research Use Only. Not for use in diagnostic procedures.

© 2020 Illumina, Inc. All rights reserved.
The PD-1/PD-L1-checkpoint restrains T cell immunity
in tumor-draining lymph nodes
Floris Dammeijer, Mandy van Gulijk, Evalyn E. Mulder,
Melanie Lukkes, Larissa Klaase, Thierry van den Bosch,
Menno van Nimwegen, Sai Ping Lau, Kitty Latupeirissa,
Sjoerd Schetters, Yvette van Kooyk, Louis Boon,
Antien Moyaart, Yvonne M. Mueller, Peter D. Katsikis,
Alexander M. Eggermont, Heleen Vroman,
Ralph Stadhouders, Rudi W. Hendriks, Jan von der Thüsen,
Dirk J. Grünhagen, Cornelis Verhoef, Thorbald van Hall,
and Joachim G. Aerts

On the cover: Featured on the cover is a cross-section of visually striking images published on
the cover of Cancer Cell in 2020. Images are courtesy of (from left to right): Kristen E. Labbe,
Catríona M. Dowling, Hua Zhang, and Mahnoor Arshad (volume 37, issue 1); Bianca Dunn (volume
37, issue 6); Samantha Walker (volume 38, issue 2); Dylan Mortimer (volume 38, issue 1); and
Himanshu Brahmbhatt, Jennifer MacDiarmid, and Martin Hale (volume 37, issue 3).
 APRIL 10-15
MAY 17-21

We look forward to
seeing you in 2021!

Register now for the world’s most comprehensive

annual meeting dedicated to the research,
prevention, detection, and treatment of cancer.

★ Become a Member!
Join the AACR and receive a discount on registration.

VISIT AACR.ORG/AACR2021 FOR MORE Continuing Medical Education Activity -

AMA PRA Category 1 Credits™ available
INFORMATION AND TO REGISTER!

AACR.ORG ★ #AACR21

AACR.ORG/AACR2021
2001015Yy

2001015Yy_MCS_21AMVirt_Ad_Cell_8.375x10.875_3.indd 2 12/7/20 9:42 AM

1:41 PM JUL 05, 2023

THE MOMENT YOUR

RESEARCH MOVES FROM
BENCH TO BEDSIDE_

WITH BD AS YOUR PARTNER IN IMMUNO-ONCOLOGY FROM BENCH TO BEDSIDE. BD is your partner in immuno-oncology
from discovery research to clinical applications, with a range of high-performance solutions designed to give you high-quality,
reliable research data for even the most complex experiments. More solutions. More answers. More data you can trust. Discover
more with BD at your side. Discover the new BD.

Learn more at bd.com/Immuno-oncology

For Research Use Only. Not for use in diagnostic or therapeutic procedures.
BD, San Jose, CA, 95131, U.S.
BD and the BD Logo are trademarks of Becton, Dickinson and
Company or its affiliates. © 2019 BD. All rights reserved. (0619)
Recommend Trends in Cancer
to your librarian today!
With your recommendation, your library can review the opportunity
for ongoing access to seminal articles in Trends in Cancer.

What does Trends in Cancer have to offer?

• Concise and engaging expert commentary articles that address key frontline research topics
and cutting-edge advances in the rapidly changing field of cancer discovery and medicine.
• Unique platform for multidisciplinary information, discussion and education that is valuable
for scientists, clinicians, and policy makers, as well as patients and advocates.
• Latest opportunities of basic, translational and clinical findings, industry R&D, technology
and innovation, ethics, or cancer policy and funding equally presented and debated in an
authoritative but reader-friendly format.

Voices
Visualizing Cancer
Imaging has had a profound impact on our ability to understand and treat cancer. We invited some experts to
discuss imaging approaches that can be used in various aspects of cancer research, from investigating the
complexity and diversity of cancer cells and their environments to guiding clinical decision-making.
Matthew Krummel and Kenneth Hu
University of California, San Francisco
Elisabeth G.E. de Vries
University of Groningen
ll
Linking Single-Cell RNA-Seq and Live Imaging
The tumor microenvironment represents significant heterogeneity in space, with the
localization of a given cell being just as important as its transcriptional state in deter-
mining tumor progression and response to therapy. We recently introduced ZipSeq,
a method for on-demand spatial barcoding of live cells for mapping scRNA-seq data.
Following a series of spatially modulated illuminations which uncage DNA for hybridiza-
tion, combinations of surface-bound barcodes define multiple regions of interest for
transcriptomic analysis.
Doing so allows us to overlay scRNA-seq information on top of imaging data,
revealing genes whose expression varies in space within a given cell population. For
example, we recently described the gradients in cell composition and gene expression
in a lymph node, while application to a tumor model revealed a progression of myeloid
and lymphoid cell differentiation correlated with tumor infiltration depth.
Extending this approach to patient biopsies will shed light on mechanisms that result
in non-responsiveness to immunotherapy. Designating tumors as immunologically hot
or cold may not capture the whole story; tumors can consist of varied regions of
immune activity. ZipSeq will allow us to dissect the composition of cell compartments
within hot and cold regions, revealing cell-cell interactions that drive formation of this
heterogeneity. Moving forward, increased spatial resolution of this approach combined
with other spatial transcriptomic techniques will augment tissue imaging, answering
a question any microscopist has encountered: ‘‘The biology in that area looks really
interesting; I wonder what’s going on there?’’
Molecular Imaging and Cancer Drug Development
Numerous cancer medicines are designed to bind a specific target. In general, we only
have information about their blood kinetics but not about the tissue distribution of the
drugs or the whole-body distribution of their targets. To comprehensively visualize their
distribution, drugs can be radiolabeled and imaged non-invasively with single-photon
or positron emission tomography (PET), which can be highly informative. Using this
approach with trastuzumab, the monoclonal antibody targeting HER2, we detected
more tumor lesions than with conventional imaging. Low tracer tumor uptake meant
that the patient had a shorter treatment benefit from the trastuzumab drug conjugate
T-DM1. Moreover, administering an HSP90 inhibitor, known to lower HER2 tumor
expression, reduced radiolabeled trastuzumab tumor uptake.
Increasingly, antibodies are engineered to have new features—for example, bispe-
cific antibodies. These antibodies bind two epitopes, not necessarily with the same
affinity, and most are designed to bind tumor and immune cells. PET imaging with
the labeled antibody can provide insights into the consequences for drug distribution
of targeting distinct epitopes with one compound.
In the field of immunotherapy, molecular imaging can contribute as well. A small
study showed that radiolabeled PD-L1 antibody tumor uptake before treatment with
PD-1 antibody was related to response and patient survival. Another example is using
a tracer against non-tumor cells critically involved in generating the antitumor immune
response, such as CD8+ T cells, which may support proper timing and dosing of new
immunotherapeutic approaches.
Cancer Cell 38, December 14, 2020 ª 2020 Elsevier Inc. 753
 ll
Voices

Resolving Metastasis with Intravital CLEM

Metastases are resistant to multiple therapies and are responsible for the large
majority of cancer-related deaths. Yet, the cellular and molecular mechanisms driving
the formation of these secondary tumors remain only partially solved. The current
challenge is to improve our understanding of metastasis and identify machineries
that are ideal targets for anti-metastatic strategies. It is therefore of utmost impor-
tance to dissect, at the highest resolution possible, tumor cell behavior in vivo. In
collaboration with experts in metastasis and high-resolution imaging, we recently de-
signed and applied correlative tools to link the dynamic and functional recordings of
tumorigenic events in vivo to their most-detailed ultrastructure. This technique, called
intravital correlative light and electron microscopy (intravital CLEM), combines the
power of intravital imaging with electron microscopy, and it can be applied to several
Jacky G. Goetz
Fédération de Médecine Translationnelle de
animal models, such as mice or zebrafish. This technique can be used to study tumor
Strasbourg growth and invasion, priming of metastatic niches via extracellular vesicles, and
mechanisms of arrest and metastatic extravasation. It can capture subcellular
features (cellular protrusions, trafficking machineries, and organelles) or nanoscale
objects (such as extracellular vesicles) that are ‘‘invisible’’ to classical intravital
imaging approaches. It is versatile, complementary to other high-end approaches,
and likely to unravel key metastatic programs that could lead to therapeutic targeting
in the near future.

It Takes a Village to Raise a Cancer

Cancer is by definition a genetic disease and has classically been studied from this
perspective. In fact, how cancer cells misbehave is largely driven by their genetic alter-
ations. Nevertheless, the surrounding environment is key for keeping cancer cells alive.
A wide range of stromal cells, vascular structures, extracellular matrix, nerve fibers, and
immune components populate most tissue environments and regulate their homeo-
static balance.
The advent of intravital microscopy has revolutionized the way we study the tumor
microenvironment. The bone marrow was one of the first tissues to be imaged at
high resolution over time, thanks to the easy access to the calvarium through minimal
surgery. Our group has used this approach to observe fluorescent hematopoietic cells
floating through dark bone cavities. Then, lighting up one niche cell type after another
Diana Passaro
revealed a dynamic and complex multicellular unit supporting normal and malignant
Université de Paris, Institut Cochin
hematopoiesis.
The primary application of intravital microscopy is the study of dynamic behavior of
cancer cells within their native environment. In parallel, researchers have implemented
protocols, probes, and pipelines to study complex features such as metabolite flow,
intercellular exchange, and vascular functionality.
This multiparametric map of a tissue can be astonishing and raise a wide range of
questions. Reducing the dimensionality of this intricate net is the next challenge of
the field, with image processing algorithms and pattern recognition approaches
opening novel avenues in our understanding of cancer biology.

Focused Ion Beam Scanning Electron Microscopy

Understanding the hierarchical organization of molecules and organelles within the
interior of large eukaryotic cells is a challenge of fundamental interest in cell biology
and cancer research. About 15 years ago, we began developing a strategy for 3D
imaging of cells and tissue by combining iterative removal of material from the surface
of a bulk specimen using focused ion-beam milling with imaging of the newly exposed
surface using scanning electron microscopy. We originally coined the phrase ‘‘ion abra-
sion scanning electron microscopy’’ to describe this method, but we and others
switched later to labeling this more simply as focused ion beam scanning electron
microscopy (FIB-SEM). The level of detail in the 3D images obtained with FIB-SEM is
about an order of magnitude higher than what can be achieved with confocal micros-
copy. 10 nm-sized gold particles and quantum dot particles with 7 nm-sized cores
Sriram Subramaniam
The University of British Columbia
can also be detected in single cross-sectional images, allowing imaging in conjunction

754 Cancer Cell 38, December 14, 2020

 ll
Voices

with tagging of specific proteins. Revelations about the nature of organization of

membranes have emerged in almost every instance where FIB-SEM has been applied,
including insights into the organization of mitochondria in muscle tissue, organization of
membranes at virological synapses, and the nature of contact zones between mito-
chondria and internal membranes such as the endoplasmic reticulum in melanoma
cells. FIB-SEM methods can be used to provide a deeper structural understanding of
cancer biology, enabling detailed insights into differences in subcellular architecture
between normal and cancer cells.

High-Throughput Imaging: Seeing More Is Seeing More

The discoveries made using cellular imaging are countless. Yet, traditional light
microscopy methods are limited in that they are typically low-throughput and often
require prior knowledge of the interrogated pathways. High-throughput imaging
(HTI) is revolutionizing cellular imaging. HTI methods use high-capacity, high-preci-
sion, automated microscopes that allow rapid imaging of large numbers of samples,
acquisition of complex datasets, and computational image analysis methods to quan-
titatively capture morphological phenotypes, enabling powerful new experimental
strategies for cancer studies. As HTI is inherently a single-cell method and thousands
of cells can be observed in a sample, rare events, such as the presence of sparse
stem cells, the stochastic behavior of cells in a population, or features of tumor
heterogeneity, can be detected. Conversely, large numbers of cellular targets can
Tom Misteli
be interrogated simultaneously in a sample by using sets of hundreds or thousands
National Cancer Institute, USA
of antibodies or DNA probes. HTI also makes possible large-scale small molecule,
RNAi, or CRISPRi screens, which can use any phenotypic difference between normal
and cancer cells as an assay. The full potential of HTI has not been reached;
combining high-throughput with super-resolution imaging, developing new types of
imaging probes, adapting HTI to tissue imaging, and using artificial intelligence will
make previously unseen patterns of cellular and tissue organization visible and
measurable. HTI is the next wave of microscopy and has transformed cellular imaging
from a descriptive candidate approach into a powerful unbiased discovery tool in
cancer research.

Imaging Cellular Metabolic Diversity in Cancer

The autofluorescent redox cofactors reduced nicotinamide adenine dinucleotide
(phosphate) (NAD[P]H) and oxidized flavin adenine dinucleotide (FAD) are label-
free indicators of metabolic activity in cells. Optical metabolic imaging (OMI)
uses two-photon fluorescence lifetime microscopy of NAD(P)H and FAD to quan-
tify cell redox state and enzyme-binding activity within intact 3D samples. This
approach is advantageous because fluorophores that are already present in the
cells can be used to monitor metabolism with single-cell resolution. OMI has
been developed with image analysis tools and population density models to quan-
tify cellular heterogeneity and correlate it with tumor growth and treatment. These
tools predict treatment response at early time points in vivo and in patient-derived
tumor organoids by monitoring dynamic changes in cell subpopulations that are
Melissa Skala
responsive and resistant to treatment. Recently, these technologies have quanti-
Morgridge Institute for Research
fied metabolic diversity within distinct cell populations including tumor cells, fibro-
blasts, and immune cells. OMI can provide new insights into metabolic differences
between responsive and resistant cells within the same tumor, changes in the
relative abundance of responsive and resistant cells during treatment, and spatial
distributions of responsive and resistant cells within the tumor. OMI methods are
important for testing and refining drugs to target treatment-resistant minority cell
subpopulations that can ultimately result in tumor recurrence and to promote an
anti-tumor microenvironment.

Cancer Cell 38, December 14, 2020 755

 ll
Voices

Intravital Microscopy—A Window into Metastasis

Conventional microscopy only provides static images at the time the sample was
collected, not capturing their dynamic behavior, which is required to fully understand
complex processes like metastasis. The last 20 years have been a revolution in our
understanding of tumors, and the use of intravital microscopy (IVM) to image living
animals has contributed to understanding the dynamic behavior of cancer cells in
real time. IVM of small animals (from zebrafish embryos to mice) has been used to study
different aspects of tumor metastasis, including invasion, intravasation, and extravasa-
tion, and to define tumor phenotypes present in vivo (i.e., motile versus non-motile cells)
and how they relate to one another. The development of new fluorescent proteins as
well as biosensors has allowed investigators to interrogate interactions between
different cell types, expanding the possibility of studying the contribution of multiple
Jose Javier Bravo-Cordero
Icahn School of Medicine at Mount Sinai
cellular compartments to metastasis. The combination of mouse models with implanted
anatomical imaging windows has improved the temporal and spatial resolution of IVM,
facilitating the study of the same tumor region for several days and tracking tumor cells
for long periods of time, a technique that has now been extended to metastatic organs.
The next challenge will be to further develop preclinical imaging approaches used in
models and translate them into a clinical setting to be able to acquire real-time informa-
tion of tumor features in patients. This approach could potentially guide future clinical
decisions to better treat primary and metastatic cancer based on direct observation
of tumors in patients.

Molecularly Targeted Cancer Imaging via avb6

With the recognition that personalized medicine can dramatically impact patient
outcomes, there is an obvious need for novel positron emission tomography (PET)
molecular imaging agents that measure clinically relevant targets that are surrogates
or predictors of disease. A rapidly growing body of literature identifies one such target
as the integrin avb6, an epithelial-specific cell surface receptor that is generally unde-
tectable in healthy adult epithelium but significantly upregulated in a wide range of
epithelial-derived cancers. avb6 is also recognized as a prognostic indicator for several
challenging malignancies and correlates with metastatic phenotype. Therefore, avb6 is
an attractive molecular target for both detection and treatment of cancer.
We and others have developed radiolabeled peptides to non-invasively image integ-
rin avb6 expression in vivo using PET. Our group has developed a 18F-avb6-binding-
Julie L. Sutcliffe
University of California, Davis
peptide (18F-avb6-BP) with high affinity and selectivity for the integrin avb6 and favorable
pharmacokinetics in tumor-bearing mice and non-human primates, and we have trans-
lated this imaging agent into patients with breast, colon, lung, and pancreatic cancer.
18
F-avb6-BP was well-tolerated, with PET/CT images showing low background uptake
in common sites of metastatic disease and demonstrating significant uptake in primary
lesions and metastases, including sub-centimeter lesions. Beyond the immediate clin-
ical impact of 18F-avb6-BP PET/CT on pretreatment molecular imaging, this ligand may
also serve as a therapeutic delivery platform for some of the most lethal cancers facing
patients today.

756 Cancer Cell 38, December 14, 2020

Cancer Cell

Review

Intratumor Heterogeneity: The Rosetta

Stone of Therapy Resistance
Andriy Marusyk,1 Michalina Janiszewska,2 and Kornelia Polyak3,4,*
1Department of Cancer Physiology, H Lee Moffitt Cancer Center and Research Institute, Tampa, FL 33612, USA
2Department of Molecular Medicine, The Scripps Research Institute, Jupiter, FL 33458, USA
3Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA 02215, USA
4Department of Medicine, Harvard Medical School, Boston, MA 02115, USA

*Correspondence: kornelia_polyak@dfci.harvard.edu
https://doi.org/10.1016/j.ccell.2020.03.007

Advances in our understanding of molecular mechanisms of tumorigenesis have translated into knowledge-
based therapies directed against specific oncogenic signaling targets. These therapies often induce dra-
matic responses in susceptible tumors. Unfortunately, most advanced cancers, including those with robust
initial responses, eventually acquire resistance to targeted therapies and relapse. Even though immune-
based therapies are more likely to achieve complete cures, acquired resistance remains an obstacle to their
success as well. Acquired resistance is the direct consequence of pre-existing intratumor heterogeneity and
ongoing diversification during therapy, which enables some tumor cells to survive treatment and facilitates
the development of new therapy-resistant phenotypes. In this review, we discuss the sources of intratumor
heterogeneity and approaches to capture and account for it during clinical decision making. Finally, we
outline potential strategies to improve therapeutic outcomes by directly targeting intratumor heterogeneity.

Sources of Intratumor Heterogeneity
Cellular phenotypic heterogeneity within tumors is a complex,
multifactorial phenomenon, which integrates genetic, epige-
netic, and environmental inputs (Figure 1).
Genetic Heterogeneity
Genetic heterogeneity is the most studied and best understood
aspect of intratumor heterogeneity (ITH), although our under-
standing is still far from being sufficiently complete (McGranahan
and Swanton, 2017). Despite mounting challenges, gene and
mutation-centric focus remains at the core of molecular
oncology, and the majority of cancer researchers would prob-
ably still agree with the famous statement by Bert Vogelstein:
‘‘The revolution in cancer research can be summed up in a single
sentence: cancer is, in essence, a genetic disease’’ (Vogelstein
and Kinzler, 2004). Discovering tumor-associated genetic muta-
tions and interrogating their functional and clinical impact has
been the major focus of cancer research over the last few
decades, and the recent revolution in DNA sequencing technol-
ogies enabled an outpour of studies documenting startling intra-
tumor genetic heterogeneity.
Even though eukaryotic cells replicate their DNA with
astounding fidelity, the mechanism is not entirely error free.
Every time a cell divides, a few mutational errors in the form of
nucleotide substitutions and small deletions are introduced
even in the absence of internal and external mutagens (Zhang
and Vijg, 2018). Owing to the constant turnover of tumor cells
and the large size of tumor cell populations, some of these
stochastic mutational hits inevitably affect genes with known
cancer relevance, leading to the activation of oncogenes and
inactivation of tumor suppressors. Whereas the sufficiency of
these baseline replication error rates to account for carcinogen-
esis remains a subject of debate (Beckman and Loeb, 2006;
Tomasetti and Vogelstein, 2015; Tomlinson et al., 1996), many
tumors display increased mutation rates, due to either increased
external or internal mutagen exposure. These include UV-related
mutagenesis in skin cancers; tobacco-related mutagenesis in
oral, lung, and bladder cancers; reactive oxygen species-
induced mutagenesis in a wide range of malignancies, and de-
fects in DNA repair mechanisms, which have been detected in
certain tumors (e.g., deficiencies in DNA mismatch repair and
APOBEC pathways) (Alexandrov et al., 2013; Fox et al., 2013;
Kandoth et al., 2013).
Small-scale genetic mutations remain best understood and
easiest to study, but they represent only a tip of the iceberg of
genetic heterogeneity. The vast majority of spontaneous human
cancers display aneuploidy, which is linked with chromosomal
instability (CIN)––an increased rate of genomic mutational errors
involving loss, gains, and translocations of large fragments of
genomic DNA (Bakhoum and Cantley, 2018). Aneuploid cells
commonly emerge following whole-genome doubling (WGD)
due to mitotic failure leading to tetraploidization (Dewhurst
et al., 2014). WGD is an early event commonly observed across
most human cancer types and it is associated with poor prog-
nosis (Bielski et al., 2018). Compared with point mutations, these
large-scale genomic events are much more likely to impact
cellular phenotypes. Although reduced cell fitness is the most
common outcome of large-scale genomic changes, CIN enables
much faster rates of genomic diversification, thus speeding up
evolution during both tumor development and cancer therapies
(Tang and Amon, 2013). At the same time aneuploid cells also in-
fluence their microenvironment and may trigger a specific anti-
tumor immune response, which in turn may facilitate immune
evasion through upregulation of immune suppressive mecha-
nisms (Senovilla et al., 2012).
The functional relevance of genomic heterogeneity in tumor
progression and therapy resistance remains poorly understood.
The prevalent approach is to reduce the impact of large-scale
gains to changes in copy numbers (and related expression

Cancer Cell 37, April 13, 2020 ª 2020 Elsevier Inc. 471

changes) of individual ‘‘driver’’ genes, with known cellular func-
tions. Although in cases of complete loss-of-functional alleles
or focal massive amplification this approach is justified, in
many cases the impact of copy number alteration on the expres-
sion of an individual gene is relatively subtle. Most of these large
genomic changes involve many genes, leading to significant im-
pacts on gene regulatory networks, which cannot be reduced to
individual ‘‘drivers’’ (Tang and Amon, 2013). Furthermore, in
contrast to point mutations, which are almost always irreversible,
large-scale DNA copy number and structural variations (with the
exception of homozygous deletions) are typically much less sta-
ble, due to much higher rates of genomic mutations in aneuploid
cancer cells (10 2 per chromosome per generation versus 10 7
per gene per cell division) (Lengauer et al., 1997), which might
significantly complicate subclonal reconstruction analyses.
Notably, karyotypic analyses revealed significant cell-to-cell
variability of chromosomal numbers in cancer cell lines that
maintained ostensibly stable average karyotypes (Li et al.,
2009). The instability of copy number variations is most pertinent
for extrachromosomal DNA (ecDNA), which has been shown to
mediate inheritance of many driver genes across multiple can-
cers (deCarvalho et al., 2018; Turner et al., 2017; Verhaak
et al., 2019). Paralleling plasmids in bacteria, this extrachromo-
somal inheritance enables massive focal amplification of an
oncogene, while avoiding fitness penalties, associated with
genomic changes in larger pieces of DNA. Whereas the exact
mechanisms and inheritance patterns remain to be elucidated,
even entirely stochastic distribution of ecDNA can dramatically
accelerate tumor evolution and ITH (Verhaak et al., 2019).
Although genetic heterogeneity within populations of tumor cells
is shaped by the several mechanisms of genetic diversification
described above, they are generally thought to be lacking a power-
ful diversification mechanism observed in natural populations
throughout all taxa of life––(para)sexual recombination. However,
the assumption of strict asexuality of cancers might be incorrect.
472 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Figure 1. Sources of Intratumor
heterogeneity
Intratumor heterogeneity represents integration of
inputs from genetic, phenotypic, and microenvi-
ronmental heterogeneity, in turn increasing the
odds of both pre-existence of tolerant and resis-
tant subpopulations, and the ability to evolve new
adaptations.
Several publications, as well as our unpub-
lished work have documented evidence of
spontaneous cell fusions between geneti-
cally distinct cancer cells, or between can-
cer and non-cancer cells within the tumor
microenvironment (Gast et al., 2018; Ja-
cobsen et al., 2006; Lu and Kang, 2009).
Genomes of polyploid cells are often un-
stable (Duelli et al., 2007; Storchova and
Pellman, 2004), which could lead to addi-
tional diversification associated with ploidy
reduction, similar to a fusion-mediated
ploidy conveyor in mature hepatocytes
(Duncan et al., 2010) or asexual recombi-
nation observed in Candida albicans (Zhang et al., 2015). As the
mechanistic underpinning of spontaneous cell fusions remains
poorly explored, it is unclear whether it is related to the cell-in-
cell phenomenon, cell cannibalism and entosis that have been
observed in numerous cancer types upon various stress condi-
tions (Fais and Overholtzer, 2018). Whereas the relevance of this
experimentally observed phenomenon is difficult to access in pri-
mary tumors (apart from cases of tumors in recipients of tissue
transplants, such as described in case reports of a bone marrow
transplant recipients) (LaBerge et al., 2017; Lazova et al., 2013),
mathematical modeling suggests that fusion-mediated recombi-
nation can further enhance diversity in populations of tumor cells
(Miroshnychenko et al., 2020).
One reason why progress has been rather slow with under-
standing the links between genetic heterogeneity and tumor
evolution is the paucity of suitable experimental models that
reproduce the ITH of human tumors. Conventional genetically
engineered mouse models (GEMMs) are dominated by those
driven by powerful combination mutations within a single cell,
providing convenient and reproducible experimental systems
to study molecular mechanisms. However, these models rarely
display the degree of subclonal and cellular genetic heterogene-
ity seen in spontaneous cancers. However, next-generation
GEMMs are designed to more accurately mimic human cancers,
including ITH (Kersten et al., 2017). Among others, the KPC
mouse model of pancreatic adenocarcinoma (PAD) driven by
mutant Kras and Trp53 transgenes displays extensive subclonal
heterogeneity largely driven by large-scale copy number alter-
ations targeting genes with known functional relevance in human
PAD (Niknafs et al., 2019). Another promising GEMM-based
approach to understand the impact of ITH is to combine low-
penetrance oncogenic drivers with a source of genetic diversifi-
cation, such as Sleeping Beauty transposons (Mann et al., 2016).
The use of such models will enable more detailed studies
assessing the functional relevance of subclonal interactions in
Cancer Cell
Review
tumor evolution, which then enable the design of improved
therapeutic strategies for heterogeneous tumors.
Epigenetic Heterogeneity
Studies on tumor heterogeneity have primarily focused on
cancer cell genomes even though selection forces act on
phenotypes rather than genotypes. Despite the critical impor-
tance of cancer-associated mutations on clinically relevant
phenotypic features, such as responses to growth signaling,
proliferation, and death, epigenetic (we use the term in a broader
sense, covering all non-genetic phenotypic determinants) mech-
anisms have a greater impact on tumor cell phenotypes (Flava-
han et al., 2017). For example, the gene expression and epige-
netic profiles of genetically abnormal CD44+CD24– progenitor-
like cancer cells isolated from primary breast tumors shows
closer resemblance to the profiles of normal CD44+CD24– mam-
mary epithelial cells than to that of more differentiated
CD44 CD24+ cancer cells from the same tumor; the same is
true for CD44 CD24+ cells (Shipitsin et al., 2007). These data
suggest that differentiation state-related epigenetic programs
have a dominant impact in shaping phenotypes compared with
cancer-related genetic aberrations. Similar observations have
been made in brain tumors, which represent one of the best ex-
amples of cancer stem cell-driven malignancies (Liau et al.,
2017; Neftel et al., 2019; Patel et al., 2014). However, more differ-
entiated cancer cells can give rise to stem cell-like cancer cells
through epithelial-to-mesenchymal transition (EMT) triggered
by certain microenvironmental signals, such as hypoxia (Mohlin
et al., 2017) or interaction with stromal cells (Karnoub et al.,
2007). Thus, tumor cells display high epigenetic and phenotypic
plasticity. This phenotypic plasticity is likely to be key for the phe-
nomenon of drug-tolerant persistence, as well as the ability of
cells to acquire resistance through stable non-genetic changes
in gene expression, as we will elaborate below.
Although epigenetic heterogeneity could be key in acquisition
of traits of profound clinical importance, such as therapeutic
resistance (Shaffer et al., 2017) or metastatic dissemination
(Roe et al., 2017), accounting for non-genetic heterogeneity is
far more challenging since, in contrast to genetic heterogeneity,
the phenotypes of tumor cells can be highly plastic. Epigeneti-
cally defined phenotypic traits range from essentially hard-wired
ones, which can be conceptualized as epi-mutations, such as
silencing of key tumor suppressor genes, mediated by DNA hy-
permethylation on one end of the spectrum, to noise-driven cell-
to-cell differences that dissipate within a few cell divisions on the
other end of the spectrum (Marusyk et al., 2012). Whereas these
two extremes are conceptually the easiest to deal with, the
majority of non-genetic phenotypic heterogeneity, reflecting
integration of microenvironmental inputs and stochastic events
and differentiation status, falls in the gray zone in terms of
heritability.
Studies assessing epigenetic heterogeneity within tumors
have been largely focusing on DNA methylation, since this is
technically less challenging to measure even at single-cell level
than chromatin modification (Mazor et al., 2016). Because of
the reversible nature of epigenetic modifications, it was not clear
if they can be used to define subclones, track tumor evolution,
and assess intratumor subclonal and cellular heterogeneity.
However, studies in prostate cancer (Brocks et al., 2014), glioma
(Mazor et al., 2015), and chronic lymphocytic leukemia (CLL)
(Gaiti et al., 2019) have demonstrated that inferring tumor evolu-
tion based on genetic and DNA methylation patterns largely
overlaps. Furthermore, quantitative measures of ITH in Barrett’s
esophagus based on DNA methylation and genetic mutations
were both predictive of the progression to esophageal carci-
noma, implying that diversity is an inherent tumor characteristic
and heritable epigenetic changes have similar impact on tumor
evolution as genetic ones (Merlo et al., 2010).
ITH for DNA methylation has consistently been observed in
regulatory regions that affect the transcription of genes relevant
to the disease process. For example, in prostate cancer high ITH
is observed for enhancers bound by the androgen receptor
(Brocks et al., 2014), a ligand-dependent transcription factor
driving the expression of many genes essential for the survival
and proliferation of prostate tumor cells. Similarly, in CLL binding
sites for transcription factors with known functional relevance,
such as NFKB1 and MYBL1 displayed lower epimutation poten-
tially due to the protection of these sites from epimutations by
transcription factor binding (Gaiti et al., 2019). Advances in tech-
nologies have enabled the single-cell multiomic profiling of tu-
mors that allows the direct analyses between genetic and epige-
netic alterations and gene expression profiles. A recent study
performing single-cell RNA sequencing and single-cell assay
for transposase-accessible chromatin sequencing (ATAC-seq)
profiling of mixed-phenotype acute leukemias have demon-
strated common malignant signatures despite extensive hetero-
geneity across patients and within individual cases, and identi-
fied the RUNX1 transcription factor as a key regulator of these
programs (Granja et al., 2019).
Because tumor cells display clear parallels with differentiation
hierarchy-related phenotypic heterogeneity in normal tissues,
the cancer stem cell framework has been, and still remains, a
popular paradigm to conceptualize the ITH. However, pheno-
typic ITH cannot be reduced to differentiation hierarchies, as tu-
mor cells display far greater cell-to-cell variability compared with
their differentiation state normal counterparts (Jenkinson et al.,
2017; Landau et al., 2014). This increased phenotypic variability
likely reflects the impact of intratumor genetic heterogeneity, as
well as greater variability in contextual signals provided by aber-
rant and less-structured (compared with normal tissues) tumor
microenvironments. In addition to increased plasticity and the
diversifying inputs of genetic and microenvironmental heteroge-
neity, this increased cell-to-cell variability might also be a
consequence of global epigenetic changes, generally observed
in cancer cell epigenomes (such as global hypomethylation of
CpG islands) (Feinberg, 2007), and can be summarized as higher
entropy of the cancer epigenome (Jenkinson et al., 2017).
Microenvironmental Heterogeneity
Although heterogeneity of genotypes within tumors translates
into diversity of phenotypes, phenotypes are not simple func-
tions of genotypes. Indeed, all normal cells share nearly identical
wild-type genomes, which encode highly diverse phenotypic
manifestations. Phenotypic diversity in normal cells reflects
developmental processes triggered by responses to microenvi-
ronmental cues. Whereas tumor cells often display aberrant re-
sponses to microenvironmental signals (e.g., suppression of
death from anoikis), they still retain a large repertoire of normal
responses. An extreme example is the ability of some tumor cells
to become a functional part of normal tissues, when embedded
Cancer Cell 37, April 13, 2020 473

in normal microenvironments. For example, human breast can-
cer cell lines co-injected with normal mammary epithelial cells
into the mammary fat pads will become part of a functional mam-
mary epithelium (Bussard and Smith, 2012).
On the other hand, tumor formation entails not only genetic
and epigenetic transformation of normal cells, but also highly
aberrant microenvironments. Normal tissues are organized in
functional and structural units with near-equal access to
vasculature, which provides oxygen and nutrients while
removing waste products. For example, breast epithelium is
organized as two layers––basal and luminal layers––with cells
facing either basal membrane, or lumen (Figure 2). As blood
and lymphatic vasculature resides in sparsely populated con-
nective tissue, both cell layers obtain nearly equal access to
fully diffusible factors (such as oxygen), while cellular access
to contact signaling from the extracellular matrix (ECM) com-
ponents of basal membrane, as well as to stroma-produced
cytokines, is primarily confined to the basal layer. Perturba-
tions in normal tissue architecture due to aging and chronic
inflammation contribute to increased cancer risk and tumor
progression (Fane and Weeraratna, 2019), and tissue organi-
zation is completely lost in invasive and metastatic tumors.
These changes result in cellular and paracrine interactions
not observed in healthy tissues, which contribute to selection
and further diversify cancer cell phenotypes. Although some
tumor cells are still in contact with stroma-derived ECM, its
composition is altered compared with normal tissues (Naba
et al., 2012), and many tumor cells are separated from the
stroma by more than one cell layer. Moreover, blood and
lymphatic vasculature in tumors are disorganized with signifi-
cant functional, spatial, and temporal heterogeneity (Carmeliet
and Jain, 2000; Stacker et al., 2014). These perturbations in
both stromal and epithelial organization lead to significant
474 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Figure 2. Microenvironmental
Heterogeneity
In normal tissues (such as normal breast depicted
here), most epithelial cells experience similar
concentrations of nutrients, oxygen, and growth
factors. However, structural disorganization of
epithelia, stroma, and vasculature leads to signifi-
cant inequality in concentrations of these factors
(one factor is illustrated). Furthermore, differences
in diffusion and consumption rates of different
factors combinatorially increase variability in
microenvironmental cues directly influencing
phenotypic heterogeneity and creating distinct
selection forces.
spatial and temporal variability in nutri-
ents, oxygenation, growth factors, and
pH (Korenchan and Flavell, 2019;
Yuan, 2016), in turn providing diverse
and abnormal contextual signals, which
are absent in healthy normal tissues. As
sensing environmental cues integrates a
multitude of inputs, existence of
different spatial and temporal gradients
of microenvironmental components
might lead to dramatic, combinatorial
diversification of contextual signals,
thus inducing phenotypic ITH, reflective
of cellular responses to these contextual signals, rather than
specific well-defined phenotypes (Figure 2).
In addition to diversification of contextual cues, structural and
microenvironmental disorganization in carcinogenesis also cre-
ates relatively consistent new microenvironments, leading to
predictable association of tumor cell phenotypes with these
microenvironmental features. For example, local acidification
at the tumor edge drives mesenchymal transition, mediating tis-
sue invasion (Estrella et al., 2013). There is a striking similarity be-
tween the arrangement of cancer cells around blood vessels and
vegetation around waterways in desert regions (Alfarouk et al.,
2013). In addition to directly influencing tumor cell phenotypes,
predictably distinct microenvironmental habitats could provide
distinct selective pressures (Gatenby and Gillies, 2008). Several
studies have linked genetic heterogeneity with distinct habitat lo-
cations (Gillies et al., 2018; Hoefflin et al., 2016; Lloyd et al., 2016)
and have shown that spatial distribution of genetically distinct tu-
mor cell populations correlates with poor clinical outcome (Jan-
iszewska et al., 2015).
Microenvironmental heterogeneity also involves heterogeneity
in immune cell infiltration, which is of obvious importance for im-
munotherapies. Leukocytes are frequently one of the most abun-
dant cell types within tumors and their highly mobile nature can
lead to rapidly changing spatial heterogeneity that can create
immunologically active or silent niches. Because T cells can
directly eliminate certain cancer cells or even populations of can-
cer cells with specific markers, it is not surprising that the fre-
quency and location of T cells have directly been linked to sub-
clonal heterogeneity in cancer. Furthermore, since T cells are
activated by specific tumor neoantigens, many of which are
generated by tumor-specific mutations, the location of T cells
with specific T cell receptors (TCRs) also varies within tumors
and correlates with the number and types of mutations. For
Cancer Cell
Review
example, in non-small cell lung cancer (NSCLC) assessing muta-
tions and TCR diversity in different regions of the same tumor
demonstrated that the abundance of expanded intratumor ubiq-
uitous TCRs is associated with the number of nonsynonymous
mutations (Joshi et al., 2019). At the same time the number of
expanded regional TCRs correlated with the number of regional
nonsynonymous mutations.
In serous ovarian cancer metastases, higher epithelial CD8+
T cell infiltration was associated with lower tumor ITH presum-
ably due to immune-mediated elimination of certain subclones
present as minor subpopulations (Zhang et al., 2018). Subclonal
neoepitopes have been shown to be more immunogenic than
clonal ones (Jimenez-Sanchez et al., 2017), potentially explain-
ing these results. In some cases, immune evasion was due to
HLA (human leukocyte antigen) loss of heterozygosity in the tu-
mor cells coupled with upregulation of immune checkpoint inhib-
itors. Interestingly, while tumor regions with high CD8+ and CD4+
T cell densities had high TCR diversity, this was not associated
with tumor cell ITH implying that T cell and tumor cell cellular
heterogeneity are related but distinct features.
Thus, phenotypes of tumor cells are shaped by an integration of
genetic, epigenetic, and microenvironmental inputs (Figure 1).
Whereas we have introduced these inputs separately, they are
highly interrelated. For example, microenvironmental inflammation
has been linked with increased CIN and increased epigenetic plas-
ticity (Colotta et al., 2009; Grivennikov et al., 2010; Korkaya et al.,
2011). On the other hand, CIN has been linked to the induction of
inflammatory signaling through the cGAS-STING anti-viral
pathway (Bakhoum and Cantley, 2018), thus modulating microen-
vironments, including immune responses. The resulting pheno-
typic ITH translates to heterogeneity in the sensitivity of tumor cells
to anti-cancer drugs and cytotoxic immunity. In addition, some of
the heterogeneous features of the tumor microenvironment can
directly modulate therapeutic responses as described below.
Quantitative Assessment of ITH
ITH in Tissue Samples
Even though all types of ITH have been linked to poor patient out-
comes and therapeutic resistance in many different tumor types,
assessing heterogeneity in human tissue samples is still a major
challenge, which partially explains why ITH is still not commonly
used to guide clinical decisions. One of the best examples for
this is ITH for ERBB2 (encoding HER2) amplification in HER2+
breast cancer driven by oncogenic HER2 signaling. HER2-tar-
geted therapies have been very successful for HER2+ tumors,
but they are only effective when the tumor cells uniformly ex-
press and are dependent on HER2. However, in many cases
HER2 expression and ERBB2 amplification is not homogeneous
within tumors and this has been associated with shorter disease-
free survival (Rye et al., 2018). Assessing HER2 status by immu-
nohistochemistry and fluorescence in situ hybridization (FISH)
has long been used in the clinic for the identification of patients
who will most likely benefit from HER2-targeted therapies. How-
ever, due to the increasing recognition of ITH, HER2 heterogene-
ity, defined as HER2 positivity by FISH in >5% and <50% of
tumor cells, has also been incorporated into the report. Consid-
ering that HER2 status is determined by automated counting of
FISH signal in 50 single cells, this assay could relatively easily
be expanded into a more quantitative assessment of cellular
heterogeneity for HER2, and several clinical trials are ongoing
(NCT02326974) to determine if this could potentially improve
the clinical management of patients. Preliminary analysis of the
data suggests that patients with higher pre-treatment cellular
heterogeneity for ERBB2 amplification are less likely to have a
complete pathologic response to neoadjuvant pertuzumab and
T-DM1 treatment (Filho et al., 2019).
Quantitative assessment of cell-to-cell variation in the expres-
sion of a therapeutic target at the protein level is much harder to
evaluate, as it is usually measured using immunostaining and
manually scored by pathologists. Scoring discordance between
individuals and variation between different batches of staining
contribute to limited reporting of the observed heterogeneity.
However, technological advances in multiplex immunostaining
techniques, digital pathology, and machine learning are paving
the way to more accurate reporting of ITH.
Decreased expression of the target is only one of the many
mechanisms by which tumors evade therapy. Many known resis-
tance mechanisms are related to genetic alterations that are iden-
tified based on unbiased genome-wide analyses. Several compu-
tational approaches have been developed to use next-generation
sequencing read frequency to assess ITH and infer clonal evolu-
tion of a tumor. Multiregional sequencing of renal carcinoma, glio-
blastoma, and lung cancer have all showed significant divergence
between distinct areas of the same tumor (Gerlinger et al., 2012;
Thrane et al., 2018; Yates et al., 2015; Zhang et al., 2014).
Although evolutionary trajectories can be drawn between distinct
clones at different sites, clinical benefit of these analyses remains
to be determined. The relatively high cost of exome-sequencing
makes it challenging to perform these analyses for all patients. A
large clinical trial (NCT01888601) was initiated in 2014 for NSCLC
patients, aiming at multiregional and longitudinal sampling of
primary and recurrent tumors to determine the impact of ITH on
clinical outcomes. Initial results from this study show that ITH for
copy number alterations and genomic instability are associated
with increased risk of recurrence or death (Jamal-Hanjani et al.,
2017). Moreover, immune cell heterogeneity within tumors points
to a correlation between T cell clonality and tumor antigen diver-
sity in different regions of the same tumor (Reuben et al., 2017).
Advances leading to decrease of sequencing costs and stream-
lined bioinformatic analysis will be required for the more wide-
spread use of genome-wide approaches for routine ITH assess-
ment for patient prognostication.
Regional sampling of heterogeneous tumors remains one of
the largest challenges in clinical diagnostics. In certain tumor
types, such as lung or breast cancer, biopsies from several
distinct regions may be taken. However, in glioblastoma and
pancreatic cancer, additional sample collection is associated
with higher risk for the patient. Multiregional sequencing has
invariably showed regional heterogeneity in different tumor
types, raising many questions. How representative is a single
biopsy? How many biopsies would we need to collect to cover
the extent of heterogeneity of a large highly aggressive tumor?
An important tool in addressing these seemingly impossible is-
sues is ‘‘liquid biopsy’’: sampling blood and other body fluids.
ITH in Liquid Biopsies
Blood drawn from cancer patients contains circulating tumor
cells (CTCs) and circulating tumor DNA (ctDNA) (Rossi and Igna-
tiadis, 2019). Liquid biopsies can easily be repeated weekly or
Cancer Cell 37, April 13, 2020 475

biweekly, allowing for unprecedented longitudinal monitoring of
cancer. Analysis of genomic alterations present in CTCs or
ctDNA has been successfully used to follow therapy response
over time and identify resistance in breast (Aceto et al., 2014;
Dawson et al., 2013), colon (Misale et al., 2012; Siravegna
et al., 2015), and lung (Anagnostou et al., 2019) cancer.
Moreover, blood samples can also be used to assess the hetero-
geneity of the T cell antigen receptor repertoire, which could help
the design of more effective adoptive T cell immunotherapies
and monitoring their efficacy. The liquid biopsy paradigm was
also applied to cerebrospinal fluid (CSF) of brain tumor patients.
Although not as easy as blood collection, ctDNA in CSF is asso-
ciated with glioma progression and shorter survival. A recent
study has shown that CSF ctDNA recapitulated the genotype
of tumor biopsy and could thus be used for genotype-directed
therapy monitoring (Miller et al., 2019). Yet, shedding of CTCs
and ctDNA into the bloodstream is neither fully understood,
nor uniform in all patients. Additional studies are required to
dissect the biology of CTC and tumor DNA shedding and to
determine how well ctDNA represent tumor profiles in order to
validate liquid biopsy as a safe and robust way to monitor
changes in ITH during treatment.
Single-Cell Profiling
In recent years the development of single-cell sequencing tech-
nologies has enabled studies of ITH at increased resolution.
Thousands of cells can now be barcoded and profiled, revealing
phenotypically or genetically distinct subpopulations present in a
tumor. Phenotypic heterogeneity can also be assessed by high
multiplexing capacity of cytometry-time-of-flight (Spitzer and
Nolan, 2016), multiplexed ion beam imaging (MIBI) (Angelo
et al., 2014), and cyclic immunofluorescence (cycIF) (Lin et al.,
2015) techniques that allow for the quantification of multiple pro-
tein expression in individual cells. The advantage of MIBI and cy-
cIF is that they are performed in intact tissue samples maintain-
ing tumor topology and cellular context. Although all these
techniques are very useful in identifying distinct subpopulations
of cells, especially in the tumor microenvironment, it is much
476 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Figure 3. Evolution of Therapeutic
Resistance
Heterogeneous primary tumors likely contain
subpopulations of cancer cells with pre-existing
partial or full resistance to therapy due to genetic or
epigenetic mechanisms. Genetically, resistant
cancer cells can also outgrow from epigenetically
plastic persisters in part due to therapy-induced
alterations. Interaction with the microenvironment
also contributes to phenotypic heterogeneity and
associated therapeutic resistance within tumors.
In most advanced-stage tumors, both genetic and
epigenetic resistance is likely to be present, lead-
ing to multiple different resistance mechanisms.
harder to clearly deconvolute complex
phenotypes of tumor cells. In the case of
brain tumors, single-cell profiling revealed
four major states associated with distinct
developmental pathways hijacked by
cancer cells (Neftel et al., 2019; Patel
et al., 2014). Yet, in many other tumor
types these distinctions are not as clear.
However, phenotypic diversity may provide more insight into
the genetic and epigenetic changes associated with treatment,
as therapy has a more immediate effect on cellular phenotypes.
Because drug resistance can occur via different mechanisms
due to genetic, epigenetic, and phenotypic heterogeneity
(Figure 3), single-cell multi-omics and parallel deconvolution of
all the traits of individual cancer cells would be needed to
integrate all the variables. First approaches of integration of
genetic and epigenetic measures at single-cell resolution show
that transcriptional and epigenetic plasticity of certain subpopu-
lations may be associated with their lineage history and underly-
ing driver mutation (Gaiti et al., 2019). Multi-omics approaches
will certainly lead to better understanding of the complex biology
of the tumor ecosystem. However, more targeted approaches to
identify clinically relevant variables will likely be needed as well.
Heterogeneity in Therapy Resistance
The majority of advanced, metastatic cancers remain incurable,
even in those cases when available therapies are capable of
eliminating the vast majority of tumor cells. Pre-existing ITH in-
creases the odds of at least some tumor cells to survive ther-
apy-induced elimination, while ongoing diversification of tumor
cell phenotypes during treatment enables tumor cells to adapt
to therapy-imposed selective pressures, leading to de novo
resistance and eventual relapse. Here, we discuss the impact
of ITH on resistance to the two most promising and effective
(when properly matched) types of therapies: targeted therapies
and immunotherapies, both of which are commonly combined
with chemotherapy. ITH has been shown to be associated with
poor outcome and decreased response to cancer treatment by
multiple groups in multiple human cancer types implying a
universal role in therapeutic resistance (Almendro et al., 2014;
Morris et al., 2016).
Heterogeneity in Therapy Resistance-Targeted Therapy
Small-molecule inhibitors, directed against abnormal signaling
of mutated kinases, are commonly used as frontline therapies
for cancers with druggable driver lesions, such as lung cancers
Cancer Cell
Review
with mutations in epidermal growth factor receptor (EGFR) and
anaplastic lymphoma kinase (ALK), or melanomas with mutant
BRAF (Zhang et al., 2009). Despite the often-spectacular initial
tumor responses, and continuous development of more efficient
inhibitors, resistance emerges with near inevitability in
advanced, metastatic cancers. In many cases of clinical relapse,
resistance is associated with specific genetic mutations, such as
genomic amplification of the therapy target, point mutations that
allosterically reduce the ability of the drug to block enzymatic
activity, or amplification/mutation of other genes, enabling tumor
cells to sustain oncogenic signaling (Lovly and Shaw, 2014).
Most of these mutational changes are considered to be sufficient
to confer full resistance, underlying clinical relapse. In this case,
probability of pre-existing or de novo resistance could be
inferred using mathematical models, involving target population
size (numbers of tumor cells capable of self-renewal), mutational
probability, and proliferation/death rates (Bozic et al., 2013; Foo
and Michor, 2014; Wodarz and Komarova, 2005). Assessing
tumors at different stages of progression and building
mathematical models based on these data could also be used
to predict probable tumor evolutionary paths toward resistance
(Angelova et al., 2018; Khan et al., 2018). The reality, however,
might be more complicated, as development of resistance might
be multifactorial (Hong et al., 2019; Shaffer et al., 2017; Vander
Velde et al., 2019), in which case pre-existing diversity and
ongoing diversification, which fuel natural selection, rather than
presence of specific mutations, might be essential for the evolu-
tion of resistance (Merlo et al., 2006). The relationship between
genetic diversity and evolution is more complicated for copy
number variations due to CIN: while providing a powerful source
of diversification, the majority of large-scale chromosomal
changes are disadvantageous (Tang and Amon, 2013). Both
experimental studies and analysis of clinical data support the
existence of a ‘‘sweet spot’’ of CIN-related diversity, where
intermediate levels, which, balance evolvability with the ability
to maintain fitness of tumor cell populations, are most dangerous
(Andor et al., 2016; Godek et al., 2016). This constraint most
likely does not apply to copy number changes mediated by
ecDNA, which avoids fitness cost of large-scale chromosomal
changes, while providing a highly dynamic focused diversifica-
tion mechanism, enabling populations of tumor cells to quickly
find optimum under therapy-imposed selection pressures
(Nathanson et al., 2014).
Despite the clear evidence of genetic mechanisms of resis-
tance, in many cases no known genetic drivers could be identi-
fied. Whereas it is formally possible that resistance in these
cases could be attributed to uncharacterized mutations, com-
monality of long remission times, and a growing body of exper-
imental evidence suggests that resistance can arise through
epigenetic mechanisms, involving semi-stable changes in gene
expression mediated by chromatin remodeling (Hinohara et al.,
2018; Knoechel et al., 2014; Liau et al., 2017; Risom et al.,
2018; Sharma et al., 2010; Shu et al., 2016). A seminal paper
by Sharma et al. (2010) implied stochastic, reversible epigenetic
transition into a distinct drug-tolerant phenotypic state, enabling
tumor cells to survive therapy, but incapable of supporting
robust net positive growth rates to drive tumor relapse. A number
of subsequent studies demonstrated the relevance of this phe-
nomenon toward multiple therapeutic modalities across many
cancer types, including breast and brain tumors (Hinohara
et al., 2018; Hong et al., 2019; Knoechel et al., 2014; Liau
et al., 2017; Risom et al., 2018; Sharma et al., 2010; Shu et al.,
2016). At least in some cases, phenotypic transition toward
drug tolerance might be induced by the treatment (Goldman
et al., 2015; Pisco and Huang, 2015), although the ability to un-
dergo phenotypic transition is likely to be limited to only certain
subpopulations of tumor cells (Hong et al., 2019; Shaffer
et al., 2017).
Drug tolerance is often thought as a distinct, well-defined
phenotypic state that might be defined by similar molecular un-
derpinnings across multiple cancer types, such as dependency
on certain metabolic pathways (Hangauer et al., 2017). However,
our lineage tracing-based studies revealed that tolerance toward
ALK inhibitors develops from heterogeneous pre-existent sub-
populations which differ in their ability to survive or grow under
different inhibitors (Vander Velde et al., 2019). Likewise, func-
tional phenotypic heterogeneity was recently implied in
chemo-resistance in pancreatic cancers (Seth et al., 2019), sug-
gesting that tolerance might reflect ITH in general, rather than a
well-defined phenotypic state. At the same time, certain
epigenetic states may confer ‘‘universal’’ drug resistance. For
example, high expression and activity of the KDM5 family of his-
tone demethylases has been linked to multiple different types of
therapeutic resistance in multiple different cancers. In lung can-
cer, high KDM5A expression is associated with EGFR inhibitor
resistance (Sharma et al., 2010); in melanoma, quiescent cells
required for tumor propagation that also play a role in resistance
to chemotherapeutic agents and BRAF inhibitors have high
KDM5B levels (Roesch et al., 2013). A recent study in breast can-
cer investigating mechanisms by which high KDM5B activity
leads to endocrine resistance determined that higher KDM5B
levels are associated with higher cell-to-cell transcriptomic het-
erogeneity (Hinohara et al., 2018). Thus, mutant and abnormally
expressed epigenetic regulators may influence tumor progres-
sion and therapeutic responses via their effect on cellular epige-
netic and phenotypic heterogeneity and, therefore, may repre-
sent good combination agents even if their single-agent
efficacy is modest.
Whereas drug tolerance is likely responsible for the ability of
tumor cells to survive therapy, leading to minimal residual dis-
ease, clinical relapse requires acquisition of stronger resistance
phenotypes, capable of maintaining net positive tumor growth in
the presence of treatment. Bona fide resistance can develop
from drug-tolerant cells through acquisition of new genetic mu-
tations, thus requiring new genetic diversification (Figure 3).
Notably, acquisition of EGFR T790M gatekeeper mutations by
tolerant cells leads to more robust tumor cell phenotypes. These
cells are less sensitive to T790M targeting third-generation
EGFR inhibitors compared with T790M mutations on the back-
ground of therapy naive cells (Hata et al., 2016). By maintaining
a reservoir of viable tumor cells, tolerance enables acquisition
of multiple distinct resistance mechanisms, thus supporting
more heterogeneous relapse, which reduces chances of sec-
ond-line therapy success. Similarly, in colorectal carcinomas
treatment with EGFR and BRAF inhibitors downregulates DNA
repair pathways and upregulates error-prone polymerases in
drug-tolerant persister cells, therefore increasing mutation rates
and the likelihood of resistance (Russo et al., 2019).
Cancer Cell 37, April 13, 2020 477

On the other hand, full resistance can also develop from
weakly resistant subpopulations through non-genetic mecha-
nisms. A case in point is a recent study on acquisition of resis-
tance to BRAF inhibitors, where transient expression heteroge-
neity in resistance-associated genes, such as AXL and EGFR,
enabled survival of subpopulations of tumor cells, but develop-
ment of fully resistant phenotypes involved acquisition of multi-
ple, partially coordinated gene expression changes (Shaffer
et al., 2017). A similar phenomenon has been described in breast
and lung cancers (Hong et al., 2019; Vander Velde et al., 2019).
These studies shed a new light on the association between
‘‘stemness’’ and resistance. Although EMT and stemness are
often considered to be proximal causes of therapy resistance
(Holohan et al., 2013; Singh and Settleman, 2010), phenotypic
plasticity per se might be the true culprit, by enabling cancer
cells to access resistance phenotypes by rewiring of gene regu-
latory networks (Pisco and Huang, 2015).
Whereas the majority of studies focus on cell-autonomous
resistance mechanisms, a growing body of evidence suggests
that cell-to-cell and microenvironmental interactions could
dramatically impact drug sensitivity. For example, growth factors
that can be produced by both tumor cells and normal cells within
tumor microenvironment have been shown to partially or
completely de-sensitize tumor cells to many types of tyrosine
kinase inhibitors across numerous cancer types (Wilson et al.,
2012). Similarly, drug sensitivity could be dramatically reduced
by signaling, mediated by interaction of tumor cells with the
ECM (Hirata et al., 2015). Whereas in typical in vitro studies,
every tumor cell can be equally impacted, microenvironmental
heterogeneity in vivo likely leads to heterogeneity in environmen-
tally mediated therapy resistance, where only some of the tumor
cells can evade the impact of therapy (Marusyk et al., 2016).
Although the impact of microenvironmentally mediated resis-
tance on relapse remains unclear (as tumor cells remain sensitive
outside of the borders of protective niches), the phenomenon
likely contributes to residual disease, providing reservoir of tu-
mor cells capable of developing cell-autonomous resistance
mechanisms (Hirata and Sahai, 2017; Meads et al., 2009). The
link between microenvironmental heterogeneity is not limited to
paracrine signals, cell-to-cell, and cell-ECM contact. Given the
well-documented impact of hypoxia, acidity, and nutrient status
on therapy sensitivity (Gillies et al., 2012), metabolic microenvi-
ronmental heterogeneity likely provides an important modulator
of tumor responses to therapies, a notion supported by in silico
modeling (Robertson-Tessi et al., 2015).
A standard approach for a biomedical research paper entails
reducing a phenomenon, such as therapy resistance, to a single
mechanism. However, resistance might entail a combined
impact of multiple mechanisms operating both within single cells
(Hong et al., 2019; Shaffer et al., 2017; Vander Velde et al., 2019)
and across different functionally distinct subpopulations (Cha-
bon et al., 2016; Piotrowska et al., 2015). Thus, the development
of clinical resistance is likely to involve not only different sources
of pre-existing ITH, but also different inputs for additional diver-
sification. For example, resistance to EGFR inhibitors in a cell line
model of EGFR mutant lung cancer can develop from both pre-
existing resistance-conferring mutations, but also from drug-
tolerant cells, acquiring diverse resistance mechanisms (Hata
et al., 2016; Ramirez et al., 2016). Although the role of microen-
478 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
vironmental diversification remains poorly explored, its inputs
likely shape both genetic and epigenetic diversification.
Sufficiently complete understanding of resistance cannot be
accomplished exclusively with mainstream reductionistic meth-
odological approaches, but instead requires integrative studies
and development of new tools.
Heterogeneity in Therapy Resistance––Immunotherapy
Immune surveillance is one of the key microenvironmental
constrains that all tumors must overcome in order to progress
and grow. Immunoediting refers to the process by which the im-
mune system impacts tumor evolution and it is classified into
three phases: elimination, equilibrium, and escape (Dunn et al.,
2004). Studies in experimental models and patients with
compromised immune system have demonstrated that most
neoplastic cells are eliminated by the immune system before
they become clinically relevant. However, tumor cells can have
variable expression of neoantigens and may display differences
in signaling pathways related to immunity, thus ITH for these
traits will limit the efficacy of antitumor immune responses. As
a consequence, all clinically relevant tumors have already
escaped immune surveillance by various mechanisms.
Immunotherapy is thought to be relatively unaffected by ITH,
since tumor neoantigens are usually not functional cancer
drivers, thus they are not expected to be impacted by cancer
therapy. However, recent data demonstrate that ITH influences
the success of immunotherapy as well. Acquired resistance to
immunotherapy can develop due to ITH for neoantigens, antigen
presentation, and interferon signaling (Havel et al., 2019; Kalbasi
and Ribas, 2019). Tumor mutational burden (TMB), defined as
the number of somatic mutations within tumors, is one of the
strongest predictors of response to immune checkpoint inhibi-
tors, with high TMB in general associated with better response
(Ribas and Wolchok, 2018). However, higher TMB can also in-
crease ITH and recent data in both experimental models and
patients show that tumors with high subclonal ITH are less likely
to respond to immune checkpoint inhibitors than more homoge-
neous tumors (Anagnostou et al., 2017; Gejman et al., 2018;
McGranahan et al., 2016; Milo et al., 2018; Wolf et al., 2019).
Resistance to immunotherapy due to ITH could be due to multi-
ple mechanisms. Higher ITH is associated with higher risk of hav-
ing a resistant clone. At the same time, subclonal mutations may
not trigger such an effective immune response as clonal ones
(Anagnostou et al., 2017; Gejman et al., 2018; McGranahan
et al., 2016; Milo et al., 2018; Wolf et al., 2019), although there
are also data suggesting that subclonal neoantigens are more
immunogenic (Jimenez-Sanchez et al., 2017). Thus, ITH has
prognostic and predictive value for immunotherapy as well.
Microenvironmental and spatial heterogeneity are of key
importance for immune therapies as well. For example, attrac-
tion of immune cells is mediated by paracrine action of chemo-
kines and cytokines, key suppressors of immune responses
CD73 and IDO1 act in a paracrine manner, and hypoxia (Noman
et al., 2015) and acidification (Pilon-Thomas et al., 2016) can
potently suppress cytotoxic immunity. The location of immune
cells within tumors is also highly heterogeneous, which is likely
influenced by stromal cells and is associated with probability
of response to immune checkpoint inhibitors (Hirata and Sahai,
2017). Based on the frequency and location of immune cells,
tumors can be classified as immune cold (immune desert),
Cancer Cell
Review
peritumoral, stromal restricted, and inflamed (Gruosso et al.,
2019; Keren et al., 2018; Rosenthal et al., 2019). Although this
classification was applied to tumors in different patients reflect-
ing intertumor heterogeneity, even within the same tumor there is
significant spatial heterogeneity for tumor-infiltrating lymphocyte
frequency, with some regions being fully infiltrated and others
devoid of leukocytes. In metastatic urothelial cancer a fibroblast
transforming growth factor b signature correlated with immune
exclusion and resistance to immune checkpoint inhibitors,
both of which were improved by combined treatment with
TGFBR inhibitors and anti-PD-L1 antibody (Mariathasan et al.,
2018). Thus, decreasing tumor cell and microenvironmental
ITH would likely improve the efficacy of immunotherapy as well.
Accounting for Heterogeneity in Therapeutic Decision
Making
Given the profound importance of ITH shaping therapy
responses, it could be beneficial to explore its utility as a predic-
tive marker. Perhaps more importantly, considering strategies to
reduce ITH might lead to improvement of long-term therapeutic
responses, prolonging remission and increasing the odds of
driving tumors toward extinction. Given the current focus on spe-
cific drivers of resistance, the main effort is directed toward the
detection of resistance-conferring mutations in tissue and liquid
biopsies, with the idea that these data should shed light on the
emergence of resistance, potentially informing therapeutic inter-
ventions to intercept the relapse. However, given the well-estab-
lished link between ITH and evolvability of tumors, complemen-
tary consideration of ITH could provide an independent
predictive marker, which could also be factored in when formu-
lating the initial treatment strategy, as well as informing decisions
on changing therapies during remission.
For example, in HER2+ breast cancer, tumors with higher ITH
for HER2 are less likely to respond to treatment (Rye et al., 2018).
Figure 4. Optimal Therapies for
Heterogeneous Tumors
The effective treatment of heterogeneous tumors
requires optimized therapy to minimize the evolu-
tion of therapeutic resistance. There are three
general approaches by which this can be ach-
ieved: (1) combination therapies targeting different
tumor subpopulations, different dependencies, or
both cell-autonomous and non-cell-autonomous
functions; (2) therapies combining tumor-targeting
agents with compounds that reduce ITH, such as
HDAC or BET inhibitors; and (3) adaptive therapy
when ITH is maintained and the competition of
drug-resistant and drug-sensitive cells is limited by
alternating therapy and no treatment.
Therapeutic resistance is likely caused by
cancer cells that lack HER2 amplification
and overexpression making them resis-
tant to HER2-targeted therapies. Thus,
the combination of HER2-targeting and
non-targeted agents (e.g., chemotherapy
and phosphatidylinositol 3-kinase/AKT
inhibitors) is predicted to be the most
beneficial in tumors with high ITH. Indeed,
a recent clinical trial (NCT03248492)
testing trastuzumab deruxtecan (DS-8201), an anti-HER2 anti-
body linked to topoisomerase inhibitor, in metastatic HER2+
breast cancer patients who failed previous treatment has given
positive results (Modi et al., 2019). Although DS-8201 is also a
HER2-targeting antibody, its anti-tumor effects do not require
HER2 dependency, since it simply serves as a chemotherapy
delivery agent. Thus, it is also efficacious against tumor cells
that lack HER2 amplicon but still have HER2 expression, which
makes it more efficacious in heterogeneous HER2+ tumors.
This example clearly shows benefits of understanding the cancer
cell dependencies and their combinations in improving efficacy
of treatments for heterogeneous tumors. Mathematical modeling
can also help with predicting optimal drug combinations and
treatment schedules that are most likely to work in heteroge-
neous tumors (Rockne and Scott , 2019; Chakrabarti and Mi-
chor, 2017; Gallaher et al., 2018; Leder et al., 2014). Indeed,
several clinical trials have been designed based on these models
and following evolutionary principles and some have shown
promising results (Zhang et al., 2017).
In addition to guiding therapeutic decision making, many as-
pects of ITH could be considered therapeutic targets (Figure 4).
Whereas little can be done to prevent stochastic mutations
occurring during replication, maximizing tumor debulking
through surgery or irradiation should reduce ITH, stemming
from all of the inputs. Moreover, at least some types of ITH could
be targeted more directly. Given the links between inflammation,
CIN, and phenotypic plasticity (Colotta et al., 2009; Grivennikov
et al., 2010; Korkaya et al., 2011), use of anti-inflammatory
agents as therapy explicitly directed against ITH might warrant
exploration. Similarly, phenotypic plasticity underlying acquisi-
tion of tolerance and resistance can be disrupted by epigenetic
inhibitors, such as those targeting HDACs (Sharma et al.,
2010), BET bromodomain proteins (Knoechel et al., 2014; Risom
et al., 2018), or histone demethylases (Hinohara and Polyak,
Cancer Cell 37, April 13, 2020 479

2019; Hinohara et al., 2018; Liau et al., 2017). Although current
strategies primarily focus on short-term cytotoxic/cytostatic ef-
fects of the drugs, consideration of longer-term effects, medi-
ated by reduction of ITH might be warranted, even when the
agents are not effective in the short term. Reducing microenvi-
ronmental heterogeneity might also be beneficial, both in the
short and long term. Although anti-angiogenic agents did not
live up to the initial promise of eliminating tumors by blocking
their access to oxygen/nutrients, the reverse idea––normalizing
blood vasculature using lower doses of anti-angiogenic
drugs––might still prove to be clinically useful (Jain, 2014). In
addition to improving drug delivery, which is currently the main
rationale behind the strategy, normalization of blood vasculature
could also reduce microenvironmental heterogeneity, while also
inhibiting selection for more invasive and aggressive subpopula-
tions (Robertson-Tessi et al., 2015).
Concluding Remarks
Despite the recent explosion of interest toward cancer evolution
and ITH, our knowledge in this area remains rudimentary. Mov-
ing from observational studies toward deeper understanding of
ITH and using this knowledge to improve clinical outcomes
would require not only the development of novel methodologies
to quantify different types of ITH, but also new frameworks to
conceptualize and integrate this knowledge. Achieving these
goals would necessitate moving from reductionistic approaches
of molecular oncology toward multidisciplinary studies, involving
meaningful integration of clinical observations, bioinformatical
analyses, and systems biology approaches with experimental
and mathematical modeling.
ACKNOWLEDGMENTS
We thank Dr. Daria Miroshnychenko for her help with designing Figure 2. This
work was supported by the National Cancer Institute R35CA197623 (to K.P.),
R00CA201606 (to M.J.), Susan G. Komen Breast Cancer Foundation
CCR17481976 (to A.M.), and the Breast Cancer Research Foundation (to K.P.).
DECLARATION OF INTERESTS
K.P. is a member of the scientific advisory boards of Farcast Biosciences and
Acrivon Therapeutics.
REFERENCES
Aceto, N., Bardia, A., Miyamoto, D.T., Donaldson, M.C., Wittner, B.S.,
Spencer, J.A., Yu, M., Pely, A., Engstrom, A., Zhu, H., et al. (2014). Circulating
tumor cell clusters are oligoclonal precursors of breast cancer metastasis. Cell
158, 1110–1122.
Alexandrov, L.B., Nik-Zainal, S., Wedge, D.C., Aparicio, S.A., Behjati, S., Bian-
kin, A.V., Bignell, G.R., Bolli, N., Borg, A., Borresen-Dale, A.L., et al. (2013).
Signatures of mutational processes in human cancer. Nature 500, 415–421.
Alfarouk, K.O., Ibrahim, M.E., Gatenby, R.A., and Brown, J.S. (2013). Riparian
ecosystems in human cancers. Evol. Appl. 6, 46–53.
Almendro, V., Cheng, Y.K., Randles, A., Itzkovitz, S., Marusyk, A., Ametller, E.,
Gonzalez-Farre, X., Munoz, M., Russnes, H.G., Helland, A., et al. (2014). Infer-
ence of tumor evolution during chemotherapy by computational modeling and
in situ analysis of genetic and phenotypic cellular diversity. Cell Rep. 6,
514–527.
Anagnostou, V., Forde, P.M., White, J.R., Niknafs, N., Hruban, C., Naidoo, J.,
Marrone, K., Sivakumar, I.K.A., Bruhm, D.C., Rosner, S., et al. (2019). Dy-
namics of tumor and immune responses during immune checkpoint blockade
in non-small cell lung cancer. Cancer Res. 79, 1214–1225.
480 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Anagnostou, V., Smith, K.N., Forde, P.M., Niknafs, N., Bhattacharya, R., White,
J., Zhang, T., Adleff, V., Phallen, J., Wali, N., et al. (2017). Evolution of neoan-
tigen landscape during immune checkpoint blockade in non-small cell lung
cancer. Cancer Discov. 7, 264–276.
Andor, N., Graham, T.A., Jansen, M., Xia, L.C., Aktipis, C.A., Petritsch, C., Ji,
H.P., and Maley, C.C. (2016). Pan-cancer analysis of the extent and conse-
quences of intratumor heterogeneity. Nat. Med. 22, 105–113.
Angelo, M., Bendall, S.C., Finck, R., Hale, M.B., Hitzman, C., Borowsky, A.D.,
Levenson, R.M., Lowe, J.B., Liu, S.D., Zhao, S., et al. (2014). Multiplexed ion
beam imaging of human breast tumors. Nat. Med. 20, 436–442.
Angelova, M., Mlecnik, B., Vasaturo, A., Bindea, G., Fredriksen, T., Lafontaine,
L., Buttard, B., Morgand, E., Bruni, D., Jouret-Mourin, A., et al. (2018). Evolu-
tion of metastases in space and time under immune selection. Cell 175, 751–
765.e716.
Bakhoum, S.F., and Cantley, L.C. (2018). The multifaceted role of chromo-
somal instability in cancer and its microenvironment. Cell 174, 1347–1360.
Beckman, R.A., and Loeb, L.A. (2006). Efficiency of carcinogenesis with and
without a mutator mutation. Proc. Natl. Acad. Sci. U S A 103, 14140–14145.
Bielski, C.M., Zehir, A., Penson, A.V., Donoghue, M.T.A., Chatila, W., Armenia,
J., Chang, M.T., Schram, A.M., Jonsson, P., Bandlamudi, C., et al. (2018).
Genome doubling shapes the evolution and prognosis of advanced cancers.
Nat. Genet. 50, 1189–1195.
Bozic, I., Reiter, J.G., Allen, B., Antal, T., Chatterjee, K., Shah, P., Moon, Y.S.,
Yaqubie, A., Kelly, N., Le, D.T., et al. (2013). Evolutionary dynamics of cancer in
response to targeted combination therapy. eLife 2, e00747.
Brocks, D., Assenov, Y., Minner, S., Bogatyrova, O., Simon, R., Koop, C.,
Oakes, C., Zucknick, M., Lipka, D.B., Weischenfeldt, J., et al. (2014). Intratu-
mor DNA methylation heterogeneity reflects clonal evolution in aggressive
prostate cancer. Cell Rep. 8, 798–806.
Bussard, K.M., and Smith, G.H. (2012). Human breast cancer cells are redir-
ected to mammary epithelial cells upon interaction with the regenerating mam-
mary gland microenvironment in-vivo. PLoS One 7, e49221.
Carmeliet, P., and Jain, R.K. (2000). Angiogenesis in cancer and other dis-
eases. Nature 407, 249–257.
Chabon, J.J., Simmons, A.D., Lovejoy, A.F., Esfahani, M.S., Newman, A.M.,
Haringsma, H.J., Kurtz, D.M., Stehr, H., Scherer, F., Karlovich, C.A., et al.
(2016). Circulating tumour DNA profiling reveals heterogeneity of EGFR inhib-
itor resistance mechanisms in lung cancer patients. Nat. Commun. 7, 11815.
Chakrabarti, S., and Michor, F. (2017). Pharmacokinetics and drug interactions
determine optimum combination strategies in computational models of cancer
evolution. Cancer Res. 77, 3908–3921.
Colotta, F., Allavena, P., Sica, A., Garlanda, C., and Mantovani, A. (2009). Can-
cer-related inflammation, the seventh hallmark of cancer: links to genetic
instability. Carcinogenesis 30, 1073–1081.
Dawson, S.J., Tsui, D.W., Murtaza, M., Biggs, H., Rueda, O.M., Chin, S.F.,
Dunning, M.J., Gale, D., Forshew, T., Mahler-Araujo, B., et al. (2013). Analysis
of circulating tumor DNA to monitor metastatic breast cancer. N. Engl. J. Med.
368, 1199–1209.
deCarvalho, A.C., Kim, H., Poisson, L.M., Winn, M.E., Mueller, C., Cherba, D.,
Koeman, J., Seth, S., Protopopov, A., Felicella, M., et al. (2018). Discordant in-
heritance of chromosomal and extrachromosomal DNA elements contributes
to dynamic disease evolution in glioblastoma. Nat. Genet. 50, 708–717.
Dewhurst, S.M., McGranahan, N., Burrell, R.A., Rowan, A.J., Grönroos, E., En-
desfelder, D., Joshi, T., Mouradov, D., Gibbs, P., Ward, R.L., et al. (2014).
Tolerance of whole- genome doubling propagates chromosomal instability
and accelerates cancer genome evolution. Cancer Discov. 4, 175–185.
Duelli, D.M., Padilla-Nash, H.M., Berman, D., Murphy, K.M., Ried, T., and Laz-
ebnik, Y. (2007). A virus causes cancer by inducing massive chromosomal
instability through cell fusion. Curr. Biol. 17, 431–437.
Duncan, A.W., Taylor, M.H., Hickey, R.D., Hanlon Newell, A.E., Lenzi, M.L., Ol-
son, S.B., Finegold, M.J., and Grompe, M. (2010). The ploidy conveyor of
mature hepatocytes as a source of genetic variation. Nature 467, 707–710.
Dunn, G.P., Old, L.J., and Schreiber, R.D. (2004). The three Es of cancer immu-
noediting. Annu. Rev. Immunol. 22, 329–360.
Cancer Cell
Review
Estrella, V., Chen, T., Lloyd, M., Wojtkowiak, J., Cornnell, H.H., Ibrahim-Ha-
shim, A., Bailey, K., Balagurunathan, Y., Rothberg, J.M., Sloane, B.F., et al.
(2013). Acidity generated by the tumor microenvironment drives local invasion.
Cancer Res. 73, 1524–1535.
Fais, S., and Overholtzer, M. (2018). Cell-in-cell phenomena in cancer. Nat.
Rev. Cancer 18, 758–766.
Fane, M., and Weeraratna, A.T. (2019). How the ageing microenvironment in-
fluences tumour progression. Nat. Rev. Cancer 20, 89–106.
Feinberg, A.P. (2007). Phenotypic plasticity and the epigenetics of human dis-
ease. Nature 447, 433–440.
Filho, O.M., Viale, G., Trippa, L., Li, T., Yardley, D.A., Mayer, I.A., Abramson,
V.G., Arteaga, C.L., Spring, L., Waks, A.G., et al. (2019). HER2 heterogeneity
as a predictor of response to neoadjuvant T-DM1 plus pertuzumab: results
from a prospective clinical trial. J. Clin. Oncol. 37, 502.
Flavahan, W.A., Gaskell, E., and Bernstein, B.E. (2017). Epigenetic plasticity
and the hallmarks of cancer. Science 357, https://doi.org/10.1126/science.
aal2380.
Foo, J., and Michor, F. (2014). Evolution of acquired resistance to anti-cancer
therapy. J. Theor. Biol. 355, 10–20.
Fox, E.J., Prindle, M.J., and Loeb, L.A. (2013). Do mutator mutations fuel
tumorigenesis? Cancer Metastasis Rev. 32, 353–361.
Gaiti, F., Chaligne, R., Gu, H., Brand, R.M., Kothen-Hill, S., Schulman, R.C.,
Grigorev, K., Risso, D., Kim, K.T., Pastore, A., et al. (2019). Epigenetic evolution
and lineage histories of chronic lymphocytic leukaemia. Nature 569, 576–580.
Gallaher, J.A., Enriquez-Navas, P.M., Luddy, K.A., Gatenby, R.A., and Ander-
son, A.R.A. (2018). Spatial heterogeneity and evolutionary dynamics modulate
time to recurrence in continuous and adaptive cancer therapies. Cancer Res.
78, 2127–2139.
Gast, C.E., Silk, A.D., Zarour, L., Riegler, L., Burkhart, J.G., Gustafson, K.T.,
Parappilly, M.S., Roh-Johnson, M., Goodman, J.R., Olson, B., et al. (2018).
Cell fusion potentiates tumor heterogeneity and reveals circulating hybrid cells
that correlate with stage and survival. Sci. Adv. 4, eaat7828.
Gatenby, R.A., and Gillies, R.J. (2008). A microenvironmental model of carci-
nogenesis. Nat. Rev. Cancer 8, 56–61.
Gejman, R.S., Chang, A.Y., Jones, H.F., DiKun, K., Hakimi, A.A., Schietinger,
A., and Scheinberg, D.A. (2018). Rejection of immunogenic tumor clones is
limited by clonal fraction. eLife 7, https://doi.org/10.7554/eLife.41090.
Gerlinger, M., Rowan, A.J., Horswell, S., Math, M., Larkin, J., Endesfelder, D.,
Gronroos, E., Martinez, P., Matthews, N., Stewart, A., et al. (2012). Intratumor
heterogeneity and branched evolution revealed by multiregion sequencing.
N. Engl. J. Med. 366, 883–892.
Gillies, R.J., Brown, J.S., Anderson, A.R.A., and Gatenby, R.A. (2018). Eco-
evolutionary causes and consequences of temporal changes in intratumoural
blood flow. Nat. Rev. Cancer 18, 576–585.
Gillies, R.J., Verduzco, D., and Gatenby, R.A. (2012). Evolutionary dynamics of
carcinogenesis and why targeted therapy does not work. Nat. Rev. Cancer 12,
487–493.
Godek, K.M., Venere, M., Wu, Q., Mills, K.D., Hickey, W.F., Rich, J.N., and
Compton, D.A. (2016). Chromosomal instability affects the tumorigenicity of
glioblastoma tumor-initiating cells. Cancer Discov. 6, 532–545.
Goldman, A., Majumder, B., Dhawan, A., Ravi, S., Goldman, D., Kohandel, M.,
Majumder, P.K., and Sengupta, S. (2015). Temporally sequenced anticancer
drugs overcome adaptive resistance by targeting a vulnerable chemo-
therapy-induced phenotypic transition. Nat. Commun. 6, 6139.
Granja, J.M., Klemm, S., McGinnis, L.M., Kathiria, A.S., Mezger, A., Corces,
M.R., Parks, B., Gars, E., Liedtke, M., Zheng, G.X.Y., et al. (2019). Single-
cell multiomic analysis identifies regulatory programs in mixed-phenotype
acute leukemia. Nat. Biotechnol. 37, 1458–1465.
Grivennikov, S.I., Greten, F.R., and Karin, M. (2010). Immunity, inflammation,
and cancer. Cell 140, 883–899.
Gruosso, T., Gigoux, M., Manem, V.S.K., Bertos, N., Zuo, D., Perlitch, I., Saleh,
S.M.I., Zhao, H., Souleimanova, M., Johnson, R.M., et al. (2019). Spatially
distinct tumor immune microenvironments stratify triple-negative breast can-
cers. J. Clin. Invest. 129, 1785–1800.
Hangauer, M.J., Viswanathan, V.S., Ryan, M.J., Bole, D., Eaton, J.K., Matov,
A., Galeas, J., Dhruv, H.D., Berens, M.E., Schreiber, S.L., et al. (2017). Drug-
tolerant persister cancer cells are vulnerable to GPX4 inhibition. Nature 551,
247–250.
Hata, A.N., Niederst, M.J., Archibald, H.L., Gomez-Caraballo, M., Siddiqui,
F.M., Mulvey, H.E., Maruvka, Y.E., Ji, F., Bhang, H.E., Krishnamurthy Radhak-
rishna, V., et al. (2016). Tumor cells can follow distinct evolutionary paths to
become resistant to epidermal growth factor receptor inhibition. Nat. Med.
22, 262–269.
Havel, J.J., Chowell, D., and Chan, T.A. (2019). The evolving landscape of bio-
markers for checkpoint inhibitor immunotherapy. Nat. Rev. Cancer 19,
133–150.
Hinohara, K., and Polyak, K. (2019). Intratumoral heterogeneity: more than just
mutations. Trends Cell Biol. 29, 569–579.
Hinohara, K., Wu, H.J., Vigneau, S., McDonald, T.O., Igarashi, K.J., Yama-
moto, K.N., Madsen, T., Fassl, A., Egri, S.B., Papanastasiou, M., et al.
(2018). KDM5 histone demethylase activity links cellular transcriptomic hetero-
geneity to therapeutic resistance. Cancer Cell 34, 939–953.e9.
Hirata, E., Girotti, M.R., Viros, A., Hooper, S., Spencer-Dene, B., Matsuda, M.,
Larkin, J., Marais, R., and Sahai, E. (2015). Intravital imaging reveals how BRAF
inhibition generates drug-tolerant microenvironments with high integrin beta1/
FAK signaling. Cancer Cell 27, 574–588.
Hirata, E., and Sahai, E. (2017). Tumor microenvironment and differential re-
sponses to therapy. Cold Spring Harb Perspect. Med. 7, https://doi.org/10.
1101/cshperspect.a026781.
Hoefflin, R., Lahrmann, B., Warsow, G., Hubschmann, D., Spath, C., Walter,
B., Chen, X., Hofer, L., Macher-Goeppinger, S., Tolstov, Y., et al. (2016).
Spatial niche formation but not malignant progression is a driving force for in-
tratumoural heterogeneity. Nat. Commun. 7, ncomms11845.
Holohan, C., Van Schaeybroeck, S., Longley, D.B., and Johnston, P.G. (2013).
Cancer drug resistance: an evolving paradigm. Nat. Rev. Cancer 13, 714–726.
Hong, S.P., Chan, T.E., Lombardo, Y., Corleone, G., Rotmensz, N., Bravaccini,
S., Rocca, A., Pruneri, G., McEwen, K.R., Coombes, R.C., et al. (2019). Single-
cell transcriptomics reveals multi-step adaptations to endocrine therapy. Nat.
Commun. 10, 3840.
Jacobsen, B.M., Harrell, J.C., Jedlicka, P., Borges, V.F., Varella-Garcia, M.,
and Horwitz, K.B. (2006). Spontaneous fusion with, and transformation of
mouse stroma by, malignant human breast cancer epithelium. Cancer Res.
66, 8274–8279.
Jain, R.K. (2014). Antiangiogenesis strategies revisited: from starving tumors
to alleviating hypoxia. Cancer Cell 26, 605–622.
Jamal-Hanjani, M., Wilson, G.A., McGranahan, N., Birkbak, N.J., Watkins,
T.B.K., Veeriah, S., Shafi, S., Johnson, D.H., Mitter, R., Rosenthal, R., et al.
(2017). Tracking the evolution of non-small-cell lung cancer. N. Engl. J. Med.
376, 2109–2121.
Janiszewska, M., Liu, L., Almendro, V., Kuang, Y., Paweletz, C., Sakr, R.A.,
Weigelt, B., Hanker, A.B., Chandarlapaty, S., King, T.A., et al. (2015). In situ sin-
gle-cell analysis identifies heterogeneity for PIK3CA mutation and HER2 ampli-
fication in HER2-positive breast cancer. Nat. Genet. 47, 1212–1219.
Jenkinson, G., Pujadas, E., Goutsias, J., and Feinberg, A.P. (2017). Potential
energy landscapes identify the information-theoretic nature of the epigenome.
Nat. Genet. 49, 719–729.
Jimenez-Sanchez, A., Memon, D., Pourpe, S., Veeraraghavan, H., Li, Y., Var-
gas, H.A., Gill, M.B., Park, K.J., Zivanovic, O., Konner, J., et al. (2017). Hetero-
geneous tumor-immune microenvironments among differentially growing me-
tastases in an ovarian cancer patient. Cell 170, 927–938 e920.
Joshi, K., Robert de Massy, M., Ismail, M., Reading, J.L., Uddin, I., Woolston,
A., Hatipoglu, E., Oakes, T., Rosenthal, R., Peacock, T., et al. (2019). Spatial
heterogeneity of the T cell receptor repertoire reflects the mutational land-
scape in lung cancer. Nat. Med. 25, 1549–1559.
Kalbasi, A., and Ribas, A. (2019). Tumour-intrinsic resistance to immune
checkpoint blockade. Nat. Rev. Immunol. 20, 25–39.
Cancer Cell 37, April 13, 2020 481

Kandoth, C., McLellan, M.D., Vandin, F., Ye, K., Niu, B., Lu, C., Xie, M., Zhang,
Q., McMichael, J.F., Wyczalkowski, M.A., et al. (2013). Mutational landscape
and significance across 12 major cancer types. Nature 502, 333–339.
Karnoub, A.E., Dash, A.B., Vo, A.P., Sullivan, A., Brooks, M.W., Bell, G.W., Ri-
chardson, A.L., Polyak, K., Tubo, R., and Weinberg, R.A. (2007). Mesenchymal
stem cells within tumour stroma promote breast cancer metastasis. Nature
449, 557–563.
Keren, L., Bosse, M., Marquez, D., Angoshtari, R., Jain, S., Varma, S., Yang,
S.R., Kurian, A., Van Valen, D., West, R., et al. (2018). A structured tumor-im-
mune microenvironment in triple negative breast cancer revealed by multi-
plexed ion beam imaging. Cell 174, 1373–1387.e1319.
Kersten, K., de Visser, K.E., van Miltenburg, M.H., and Jonkers, J. (2017).
Genetically engineered mouse models in oncology research and cancer med-
icine. EMBO Mol. Med. 9, 137–153.
Khan, K.H., Cunningham, D., Werner, B., Vlachogiannis, G., Spiteri, I., Heide,
T., Mateos, J.F., Vatsiou, A., Lampis, A., Damavandi, M.D., et al. (2018). Lon-
gitudinal liquid biopsy and mathematical modeling of clonal evolution forecast
time to treatment failure in the PROSPECT-C phase II colorectal cancer clinical
trial. Cancer Discov. 8, 1270–1285.
Knoechel, B., Roderick, J.E., Williamson, K.E., Zhu, J., Lohr, J.G., Cotton,
M.J., Gillespie, S.M., Fernandez, D., Ku, M., Wang, H., et al. (2014). An epige-
netic mechanism of resistance to targeted therapy in T cell acute lympho-
blastic leukemia. Nat. Genet. 46, 364–370.
Korenchan, D.E., and Flavell, R.R. (2019). Spatiotemporal pH heterogeneity as
a promoter of cancer progression and therapeutic resistance. Cancers (Basel)
11, https://doi.org/10.3390/cancers11071026.
Korkaya, H., Liu, S., and Wicha, M.S. (2011). Regulation of cancer stem cells
by cytokine networks: attacking cancer’s inflammatory roots. Clin. Cancer
Res. 17, 6125–6129.
LaBerge, G.S., Duvall, E., Grasmick, Z., Haedicke, K., and Pawelek, J. (2017).
A melanoma lymph node metastasis with a donor-patient hybrid genome
following bone marrow transplantation: a second case of leucocyte-tumor
cell hybridization in cancer metastasis. PLoS One 12, e0168581.
Landau, D.A., Clement, K., Ziller, M.J., Boyle, P., Fan, J., Gu, H., Stevenson, K.,
Sougnez, C., Wang, L., Li, S., et al. (2014). Locally disordered methylation
forms the basis of intratumor methylome variation in chronic lymphocytic leu-
kemia. Cancer Cell 26, 813–825.
Lazova, R., Laberge, G.S., Duvall, E., Spoelstra, N., Klump, V., Sznol, M.,
Cooper, D., Spritz, R.A., Chang, J.T., and Pawelek, J.M. (2013). A melanoma
brain metastasis with a donor-patient hybrid genome following bone marrow
transplantation: first evidence for fusion in human cancer. PLoS One 8,
e66731.
Leder, K., Pitter, K., Laplant, Q., Hambardzumyan, D., Ross, B.D., Chan, T.A.,
Holland, E.C., and Michor, F. (2014). Mathematical modeling of PDGF-driven
glioblastoma reveals optimized radiation dosing schedules. Cell 156, 603–616.
Lengauer, C., Kinzler, K.W., and Vogelstein, B. (1997). Genetic instability in
colorectal cancers. Nature 386, 623–627.
Li, L., McCormack, A.A., Nicholson, J.M., Fabarius, A., Hehlmann, R., Sachs,
R.K., and Duesberg, P.H. (2009). Cancer-causing karyotypes: chromosomal
equilibria between destabilizing aneuploidy and stabilizing selection for onco-
genic function. Cancer Genet. Cytogenet. 188, 1–25.
Liau, B.B., Sievers, C., Donohue, L.K., Gillespie, S.M., Flavahan, W.A., Miller,
T.E., Venteicher, A.S., Hebert, C.H., Carey, C.D., Rodig, S.J., et al. (2017).
Adaptive chromatin remodeling drives glioblastoma stem cell plasticity and
drug tolerance. Cell Stem Cell 20, 233–246 e237.
Lin, J.R., Fallahi-Sichani, M., and Sorger, P.K. (2015). Highly multiplexed imag-
ing of single cells using a high-throughput cyclic immunofluorescence method.
Nat. Commun. 6, 8390.
Lloyd, M.C., Cunningham, J.J., Bui, M.M., Gillies, R.J., Brown, J.S., and Gate-
nby, R.A. (2016). Darwinian dynamics of intratumoral heterogeneity: not solely
random mutations but also variable environmental selection forces. Cancer
Res. 76, 3136–3144.
Lovly, C.M., and Shaw, A.T. (2014). Molecular pathways: resistance to kinase
inhibitors and implications for therapeutic strategies. Clin. Cancer Res. 20,
2249–2256.
482 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Lu, X., and Kang, Y. (2009). Efficient acquisition of dual metastasis organotrop-
ism to bone and lung through stable spontaneous fusion between MDA-MB-
231 variants. Proc. Natl. Acad. Sci. U S A 106, 9385–9390.
Mann, K.M., Newberg, J.Y., Black, M.A., Jones, D.J., Amaya-Manzanares, F.,
Guzman-Rojas, L., Kodama, T., Ward, J.M., Rust, A.G., van der Weyden, L.,
et al. (2016). Analyzing tumor heterogeneity and driver genes in single myeloid
leukemia cells with SBCapSeq. Nat. Biotechnol. 34, 962–972.
Mariathasan, S., Turley, S.J., Nickles, D., Castiglioni, A., Yuen, K., Wang, Y.,
Kadel, E.E., III, Koeppen, H., Astarita, J.L., Cubas, R., et al. (2018). TGFbeta
attenuates tumour response to PD-L1 blockade by contributing to exclusion
of T cells. Nature 554, 544–548.
Marusyk, A., Almendro, V., and Polyak, K. (2012). Intra-tumour heterogeneity:
a looking glass for cancer? Nat. Rev. Cancer 12, 323–334.
Marusyk, A., Tabassum, D.P., Janiszewska, M., Place, A.E., Trinh, A., Rozhok,
A.I., Pyne, S., Guerriero, J.L., Shu, S., Ekram, M., et al. (2016). Spatial proximity
to fibroblasts impacts molecular features and therapeutic sensitivity of breast
cancer cells influencing clinical outcomes. Cancer Res. 76, 6495–6506.
Mazor, T., Pankov, A., Johnson, B.E., Hong, C., Hamilton, E.G., Bell, R.J.A.,
Smirnov, I.V., Reis, G.F., Phillips, J.J., Barnes, M.J., et al. (2015). DNA methyl-
ation and somatic mutations converge on the cell cycle and define similar
evolutionary histories in brain tumors. Cancer Cell 28, 307–317.
Mazor, T., Pankov, A., Song, J.S., and Costello, J.F. (2016). Intratumoral het-
erogeneity of the epigenome. Cancer Cell 29, 440–451.
McGranahan, N., Furness, A.J., Rosenthal, R., Ramskov, S., Lyngaa, R., Saini,
S.K., Jamal-Hanjani, M., Wilson, G.A., Birkbak, N.J., Hiley, C.T., et al. (2016).
Clonal neoantigens elicit T cell immunoreactivity and sensitivity to immune
checkpoint blockade. Science 351, 1463–1469.
McGranahan, N., and Swanton, C. (2017). Clonal heterogeneity and tumor
evolution: past, present, and the future. Cell 168, 613–628.
Meads, M.B., Gatenby, R.A., and Dalton, W.S. (2009). Environment-mediated
drug resistance: a major contributor to minimal residual disease. Nat. Rev.
Cancer 9, 665–674.
Merlo, L.M., Pepper, J.W., Reid, B.J., and Maley, C.C. (2006). Cancer as an
evolutionary and ecological process. Nat. Rev. Cancer 6, 924–935.
Merlo, L.M.F., Shah, N.A., Li, X., Blount, P.L., Vaughan, T.L., Reid, B.J., and
Maley, C.C. (2010). A comprehensive survey of clonal diversity measures in
Barrett’s esophagus as biomarkers of progression to esophageal adenocarci-
noma. Cancer Prev. Res. 3, 1388–1397.
Miller, A.M., Shah, R.H., Pentsova, E.I., Pourmaleki, M., Briggs, S., Distefano,
N., Zheng, Y., Skakodub, A., Mehta, S.A., Campos, C., et al. (2019). Tracking
tumour evolution in glioma through liquid biopsies of cerebrospinal fluid. Na-
ture 565, 654–658.
Milo, I., Bedora-Faure, M., Garcia, Z., Thibaut, R., Perie, L., Shakhar, G., De-
riano, L., and Bousso, P. (2018). The immune system profoundly restricts intra-
tumor genetic heterogeneity. Sci. Immunol. 3, https://doi.org/10.1126/sciim-
munol.aat1435.
Miroshnychenko, D., Baratchart, E., Ferrall-Fairbanks, M.C., Vander Velde, R.,
Laurie, M.A., Bui, M.M., Altrock, P.M., Basanta, D., and Marusyk, A. (2020).
Spontaneous cell fusions as a mechanism of parasexual recombination in tu-
mor cell populations. BioRxiv. https://doi.org/10.1101/2020.03.09.984419.
Misale, S., Yaeger, R., Hobor, S., Scala, E., Janakiraman, M., Liska, D., Val-
torta, E., Schiavo, R., Buscarino, M., Siravegna, G., et al. (2012). Emergence
of KRAS mutations and acquired resistance to anti-EGFR therapy in colorectal
cancer. Nature 486, 532–536.
Modi, S., Saura, C., Yamashita, T., Park, Y.H., Kim, S.B., Tamura, K., Andre, F.,
Iwata, H., Ito, Y., Tsurutani, J., et al. (2019). Trastuzumab deruxtecan in previ-
ously treated HER2-positive breast cancer. N. Engl. J. Med. 382, 610–621.
Mohlin, S., Wigerup, C., Jogi, A., and Pahlman, S. (2017). Hypoxia, pseudohy-
poxia and cellular differentiation. Exp. Cell Res. 356, 192–196.
Morris, L.G., Riaz, N., Desrichard, A., Senbabaoglu, Y., Hakimi, A.A., Makarov,
V., Reis-Filho, J.S., and Chan, T.A. (2016). Pan-cancer analysis of intratumor
heterogeneity as a prognostic determinant of survival. Oncotarget 7,
10051–10063.
Cancer Cell
Review
Naba, A., Clauser, K.R., Hoersch, S., Liu, H., Carr, S.A., and Hynes, R.O.
(2012). The matrisome: in silico definition and in vivo characterization by pro-
teomics of normal and tumor extracellular matrices. Mol. Cell Proteomics 11,
M111 014647.
Nathanson, D.A., Gini, B., Mottahedeh, J., Visnyei, K., Koga, T., Gomez, G.,
Eskin, A., Hwang, K., Wang, J., Masui, K., et al. (2014). Targeted therapy resis-
tance mediated by dynamic regulation of extrachromosomal mutant EGFR
DNA. Science 343, 72–76.
Neftel, C., Laffy, J., Filbin, M.G., Hara, T., Shore, M.E., Rahme, G.J., Richman,
A.R., Silverbush, D., Shaw, M.L., Hebert, C.M., et al. (2019). An integrative
model of cellular states, plasticity, and genetics for glioblastoma. Cell 178,
835–849 e821.
Niknafs, N., Zhong, Y., Moral, J.A., Zhang, L., Shao, M.X., Lo, A., Makohon-
Moore, A., Iacobuzio-Donahue, C.A., and Karchin, R. (2019). Characterization
of genetic subclonal evolution in pancreatic cancer mouse models. Nat. Com-
mun. 10, 5435.
Noman, M.Z., Hasmim, M., Messai, Y., Terry, S., Kieda, C., Janji, B., and
Chouaib, S. (2015). Hypoxia: a key player in antitumor immune response. A re-
view in the theme: cellular responses to hypoxia. Am. J. Physiol. Cell Physiol.
309, C569–C579.
Patel, A.P., Tirosh, I., Trombetta, J.J., Shalek, A.K., Gillespie, S.M., Wakimoto,
H., Cahill, D.P., Nahed, B.V., Curry, W.T., Martuza, R.L., et al. (2014). Single-
cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma.
Science 344, 1396–1401.
Pilon-Thomas, S., Kodumudi, K.N., El-Kenawi, A.E., Russell, S., Weber, A.M.,
Luddy, K., Damaghi, M., Wojtkowiak, J.W., Mule, J.J., Ibrahim-Hashim, A., and
Gillies, R.J. (2016). Neutralization of tumor acidity improves antitumor re-
sponses to immunotherapy. Cancer Res. 76, 1381–1390.
Piotrowska, Z., Niederst, M.J., Karlovich, C.A., Wakelee, H.A., Neal, J.W.,
Mino-Kenudson, M., Fulton, L., Hata, A.N., Lockerman, E.L., Kalsy, A., et al.
(2015). Heterogeneity underlies the emergence of EGFRT790 wild-type clones
following treatment of T790M-positive cancers with a third-generation EGFR
inhibitor. Cancer Discov. 5, 713–722.
Pisco, A.O., and Huang, S. (2015). Non-genetic cancer cell plasticity and ther-
apy-induced stemness in tumour relapse: ’what does not kill me strengthens
me’. Br. J. Cancer 112, 1725–1732.
Ramirez, M., Rajaram, S., Steininger, R.J., Osipchuk, D., Roth, M.A., Morinishi,
L.S., Evans, L., Ji, W., Hsu, C.H., Thurley, K., et al. (2016). Diverse drug-resis-
tance mechanisms can emerge from drug-tolerant cancer persister cells. Nat.
Commun. 7, 10690.
Reuben, A., Gittelman, R., Gao, J., Zhang, J., Yusko, E.C., Wu, C.J., Emerson,
R., Zhang, J., Tipton, C., Li, J., et al. (2017). TCR repertoire intratumor hetero-
geneity in localized lung adenocarcinomas: an association with predicted neo-
antigen heterogeneity and postsurgical recurrence. Cancer Discov. 7,
1088–1097.
Ribas, A., and Wolchok, J.D. (2018). Cancer immunotherapy using checkpoint
blockade. Science 359, 1350–1355.
Risom, T., Langer, E.M., Chapman, M.P., Rantala, J., Fields, A.J., Boniface, C.,
Alvarez, M.J., Kendsersky, N.D., Pelz, C.R., Johnson-Camacho, K., et al.
(2018). Differentiation-state plasticity is a targetable resistance mechanism
in basal-like breast cancer. Nat. Commun. 9, 3815.
Robertson-Tessi, M., Gillies, R.J., Gatenby, R.A., and Anderson, A.R. (2015).
Impact of metabolic heterogeneity on tumor growth, invasion, and treatment
outcomes. Cancer Res. 75, 1567–1579.
Rockne, R.C., and Scott, J.G. (2019). Introduction to mathematical oncology.
JCO Clin Cancer Inform 3, 1–4.
Roe, J.S., Hwang, C.I., Somerville, T.D.D., Milazzo, J.P., Lee, E.J., Da Silva, B.,
Maiorino, L., Tiriac, H., Young, C.M., Miyabayashi, K., et al. (2017). Enhancer
reprogramming promotes pancreatic cancer metastasis. Cell 170, 875–
888 e820.
Roesch, A., Vultur, A., Bogeski, I., Wang, H., Zimmermann, K.M., Speicher, D.,
Korbel, C., Laschke, M.W., Gimotty, P.A., Philipp, S.E., et al. (2013). Over-
coming intrinsic multidrug resistance in melanoma by blocking the mitochon-
drial respiratory chain of slow-cycling JARID1B(high) cells. Cancer Cell 23,
811–825.
Rosenthal, R., Cadieux, E.L., Salgado, R., Bakir, M.A., Moore, D.A., Hiley, C.T.,
Lund, T., Tanic, M., Reading, J.L., Joshi, K., et al. (2019). Neoantigen-directed
immune escape in lung cancer evolution. Nature 567, 479–485.
Rossi, G., and Ignatiadis, M. (2019). Promises and pitfalls of using liquid biopsy
for precision medicine. Cancer Res. 79, 2798–2804.
Russo, M., Crisafulli, G., Sogari, A., Reilly, N.M., Arena, S., Lamba, S., Bartolini,
A., Amodio, V., Magri, A., Novara, L., et al. (2019). Adaptive mutability of colo-
rectal cancers in response to targeted therapies. Science 366, 1473–1480.
Rye, I.H., Trinh, A., Saetersdal, A.B., Nebdal, D., Lingjaerde, O.C., Almendro,
V., Polyak, K., Borresen-Dale, A.L., Helland, A., Markowetz, F., and Russnes,
H.G. (2018). Intratumor heterogeneity defines treatment-resistant HER2+
breast tumors. Mol. Oncol. 12, 1838–1855.
Senovilla, L., Vitale, I., Martins, I., Tailler, M., Pailleret, C., Michaud, M., Gal-
luzzi, L., Adjemian, S., Kepp, O., Niso-Santano, M., et al. (2012). An immuno-
surveillance mechanism controls cancer cell ploidy. Science 337, 1678–1684.
Seth, S., Li, C.Y., Ho, I.L., Corti, D., Loponte, S., Sapio, L., Del Poggetto, E.,
Yen, E.Y., Robinson, F.S., Peoples, M., et al. (2019). Pre-existing functional
heterogeneity of tumorigenic compartment as the origin of chemoresistance
in pancreatic tumors. Cell Rep. 26, 1518–1532 e1519.
Shaffer, S.M., Dunagin, M.C., Torborg, S.R., Torre, E.A., Emert, B., Krepler, C.,
Beqiri, M., Sproesser, K., Brafford, P.A., Xiao, M., et al. (2017). Rare cell vari-
ability and drug-induced reprogramming as a mode of cancer drug resistance.
Nature 546, 431–435.
Sharma, S.V., Lee, D.Y., Li, B., Quinlan, M.P., Takahashi, F., Maheswaran, S.,
McDermott, U., Azizian, N., Zou, L., Fischbach, M.A., et al. (2010). A chro-
matin-mediated reversible drug-tolerant state in cancer cell subpopulations.
Cell 141, 69–80.
Shipitsin, M., Campbell, L.L., Argani, P., Weremowicz, S., Bloushtain-Qimron,
N., Yao, J., Nikolskaya, T., Serebryiskaya, T., Beroukhim, R., Hu, M., et al.
(2007). Molecular definition of breast tumor heterogeneity. Cancer Cell 11,
259–273.
Shu, S., Lin, C.Y., He, H.H., Witwicki, R.M., Tabassum, D.P., Roberts, J.M.,
Janiszewska, M., Huh, S.J., Liang, Y., Ryan, J., et al. (2016). Response and
resistance to BET bromodomain inhibitors in triple-negative breast cancer.
Nature 529, 413–417.
Singh, A., and Settleman, J. (2010). EMT, cancer stem cells and drug resis-
tance: an emerging axis of evil in the war on cancer. Oncogene 29, 4741–4751.
Siravegna, G., Mussolin, B., Buscarino, M., Corti, G., Cassingena, A., Crisafulli,
G., Ponzetti, A., Cremolini, C., Amatu, A., Lauricella, C., et al. (2015). Clonal
evolution and resistance to EGFR blockade in the blood of colorectal cancer
patients. Nat. Med. 21, 827.
Spitzer, M.H., and Nolan, G.P. (2016). Mass cytometry: single cells, many fea-
tures. Cell 165, 780–791.
Stacker, S.A., Williams, S.P., Karnezis, T., Shayan, R., Fox, S.B., and Achen,
M.G. (2014). Lymphangiogenesis and lymphatic vessel remodelling in cancer.
Nat. Rev. Cancer 14, 159–172.
Storchova, Z., and Pellman, D. (2004). From polyploidy to aneuploidy, genome
instability and cancer. Nat. Rev. Mol. Cell Biol. 5, 45–54.
Tang, Y.C., and Amon, A. (2013). Gene copy-number alterations: a cost-
benefit analysis. Cell 152, 394–405.
Thrane, K., Eriksson, H., Maaskola, J., Hansson, J., and Lundeberg, J. (2018).
Spatially resolved transcriptomics enables dissection of genetic heterogeneity
in stage III cutaneous malignant melanoma. Cancer Res. 78, 5970–5979.
Tomasetti, C., and Vogelstein, B. (2015). Cancer etiology. Variation in cancer
risk among tissues can be explained by the number of stem cell divisions. Sci-
ence 347, 78–81.
Tomlinson, I.P., Novelli, M.R., and Bodmer, W.F. (1996). The mutation rate and
cancer. Proc. Natl. Acad. Sci. U S A 93, 14800–14803.
Turner, K.M., Deshpande, V., Beyter, D., Koga, T., Rusert, J., Lee, C., Li, B.,
Arden, K., Ren, B., Nathanson, D.A., et al. (2017). Extrachromosomal onco-
gene amplification drives tumour evolution and genetic heterogeneity. Nature
543, 122–125.
Cancer Cell 37, April 13, 2020 483

Vander Velde, R., Yoon, N., Marusyk, V., Durmaz, A., Dhawan, A., Myroshny-
chenko, D., Lozano-Peral, D., Desai, B., Balynska, O., Poleszhuk, J., et al.
(2019). Resistance to targeted therapies as a multifactorial, gradual adaptation
to inhibitor specific selective pressures. bioRxiv, 504837.
Verhaak, R.G.W., Bafna, V., and Mischel, P.S. (2019). Extrachromosomal
oncogene amplification in tumour pathogenesis and evolution. Nat. Rev. Can-
cer 19, 283–288.
Vogelstein, B., and Kinzler, K.W. (2004). Cancer genes and the pathways they
control. Nat. Med. 10, 789–799.
Wilson, T.R., Fridlyand, J., Yan, Y., Penuel, E., Burton, L., Chan, E., Peng, J.,
Lin, E., Wang, Y., Sosman, J., et al. (2012). Widespread potential for growth-
factor-driven resistance to anticancer kinase inhibitors. Nature 487, 505–509.
Wodarz, D., and Komarova, N.L. (2005). Emergence and prevention of resis-
tance against small molecule inhibitors. Semin. Cancer Biol. 15, 506–514.
Wolf, Y., Bartok, O., Patkar, S., Eli, G.B., Cohen, S., Litchfield, K., Levy, R., Ji-
menez-Sanchez, A., Trabish, S., Lee, J.S., et al. (2019). UVB-induced tumor
heterogeneity diminishes immune response in melanoma. Cell 179, 219–
235.e21.
Yates, L.R., Gerstung, M., Knappskog, S., Desmedt, C., Gundem, G., Van Loo,
P., Aas, T., Alexandrov, L.B., Larsimont, D., Davies, H., et al. (2015). Subclonal
diversification of primary breast cancer revealed by multiregion sequencing.
Nat. Med. 21, 751–759.
484 Cancer Cell 37, April 13, 2020
Cancer Cell
Review
Yuan, Y. (2016). Spatial heterogeneity in the tumor microenvironment. Cold
Spring Harb Perspect. Med. 6, https://doi.org/10.1101/cshperspect.a026583.
Zhang, A.W., McPherson, A., Milne, K., Kroeger, D.R., Hamilton, P.T., Miranda,
A., Funnell, T., Little, N., de Souza, C.P.E., Laan, S., et al. (2018). Interfaces of
malignant and immunologic clonal dynamics in ovarian cancer. Cell 173,
1755–1769 e1722.
Zhang, J., Cunningham, J.J., Brown, J.S., and Gatenby, R.A. (2017). Inte-
grating evolutionary dynamics into treatment of metastatic castrate-resistant
prostate cancer. Nat Commun. 8, 1816.
Zhang, J., Fujimoto, J., Zhang, J., Wedge, D.C., Song, X., Zhang, J., Seth, S.,
Chow, C.W., Cao, Y., Gumbs, C., et al. (2014). Intratumor heterogeneity in
localized lung adenocarcinomas delineated by multiregion sequencing. Sci-
ence 346, 256–259.
Zhang, J., Yang, P.L., and Gray, N.S. (2009). Targeting cancer with small mole-
cule kinase inhibitors. Nat. Rev. Cancer 9, 28–39.
Zhang, L., and Vijg, J. (2018). Somatic mutagenesis in mammals and its impli-
cations for human disease and aging. Annu. Rev. Genet. 52, 397–419.
Zhang, N., Magee, B.B., Magee, P.T., Holland, B.R., Rodrigues, E., Holmes,
A.R., Cannon, R.D., and Schmid, J. (2015). Selective advantages of a para-
sexual cycle for the yeast Candida albicans. Genetics 200, 1117–1132.
 ll

Review
Engineering CAR-T Cells
for Next-Generation Cancer Therapy
Mihe Hong,1,4 Justin D. Clubb,1,4 and Yvonne Y. Chen1,2,3,*
1Department of Chemical and Biomolecular Engineering, University of California–Los Angeles, Los Angeles, CA 90095, USA
2Department of Microbiology, Immunology, and Molecular Genetics, University of California–Los Angeles, Los Angeles, CA 90095, USA
3Parker Institute for Cancer Immunotherapy Center at UCLA, Los Angeles, CA 90095, USA
4These authors contributed equally

*Correspondence: yvonne.chen@ucla.edu
https://doi.org/10.1016/j.ccell.2020.07.005

SUMMARY

T cells engineered to express chimeric antigen receptors (CARs) with tumor specificity have shown remark-
able success in treating patients with hematologic malignancies and revitalized the field of adoptive cell
therapy. However, realizing broader therapeutic applications of CAR-T cells necessitates engineering
approaches on multiple levels to enhance efficacy and safety. Particularly, solid tumors present unique chal-
lenges due to the biological complexity of the solid-tumor microenvironment (TME). In this review, we high-
light recent strategies to improve CAR-T cell therapy by engineering (1) the CAR protein, (2) T cells, and (3) the
interaction between T cells and other components in the TME.

INTRODUCTION
Chimeric antigen receptors (CARs) are synthetic receptors that
enable T cells to recognize tumor-associated antigens (TAAs)
in a major histocompatibility complex (MHC)-independent
manner. CAR-T cells targeting the pan–B-cell marker CD19
have shown unprecedented response rates in treating refractory
B cell malignancies (Maude et al., 2014; Neelapu et al., 2017) and
became the first genetically modified cell-based therapy to
receive US Food and Drug Administration approval (Bouchkouj
et al., 2019; O’Leary et al., 2019). However, the development
of effective CAR-T cell therapy for non–B-cell malignancies has
required more sophisticated engineering approaches to over-
come tumor-defense mechanisms such as immunosuppression,
antigen escape, and physical barriers to entry into solid tumors.
In this review, we examine current and prospective strategies to
engineer CARs, T cells that express CARs or tumor-specific
T cell receptors (TCRs), and the interaction between engineered
T cells and the tumor microenvironment (TME), with particular
focus on improving the efficacy and safety of adoptive T cell ther-
apy for the treatment of solid tumors (Figure 1).
ENGINEERING THE CAR PROTEIN
Evolution of CAR Designs
Kuwana et al. (1987) reported the first proof of principle of
combining antibody-type antigen specificity with T-cell signaling
by fusing the TCR constant region to the variable regions of a bac-
terial antigen-recognizing antibody. Single-chain variable frag-
ments (scFvs), composed of the variable heavy (VH) and light (VL)
chains of a monoclonal antibody (mAb) separated by a flexible
linker, are still commonly used as the extracellular antigen-sensing
domain of CARs. The first reports of tumor-targeting CARs demon-
strated that an scFv recognizing antigens such as human epidermal
growth factor receptor 2 (HER2) fused to the CD3z signaling domain
can elicit tumor-specific cytotoxicity (Eshhar et al., 1993; Moritz
et al., 1994; Stancovski et al., 1993), but T cells expressing these
‘‘first-generation’’ CARs that included only the CD3z chain for T-
cell signaling generally failed to elicit potent antitumor effects.
In the following years, second- and third-generation CARs
emerged that included one or two costimulatory domains,
respectively, drawing from the biological understanding that
the endogenous TCR requires association with other costimula-
tory or accessory molecules for robust signaling (Chen and Flies,
2013). Most commonly derived from CD28 or 4-1BB, these cos-
timulatory domains conferred more potent antitumor cytotox-
icity, increased cytokine production, and improved proliferation
and persistence of CAR-T cells (Haynes et al., 2002; Imai et al.,
2004). The choice of costimulatory domain has an impact on a
wide range of properties, including metabolic pathways (Kawa-
lekar et al., 2016), T-cell memory development (Kalos et al.,
2011; Kawalekar et al., 2016), and antigen-independent tonic
signaling (Long et al., 2015), prompting further research into
other costimulatory domains. For example, a third-generation
CAR with OX40 and CD28 costimulatory domains repressed
CD28-induced secretion of interleukin (IL)-10, an anti-inflamma-
tory cytokine that compromises T-cell activity (Hombach et al.,
2012). In addition, the inducible T-cell costimulator (ICOS) costi-
mulatory domain in combination with either CD28 or 4-1BB
costimulation increased in vivo persistence, and MyD88/CD40
costimulation improved in vivo proliferation of CAR-T cells (Col-
linson-Pautz et al., 2019; Guedan et al., 2018).
More recently, fourth-generation CARs that incorporate addi-
tional stimulatory domains, commonly referred to as ‘‘armored’’
CARs, have been reported. In one example, Chmielewski et al.
(2014) engineered armored CAR-T cells termed ‘‘T cells redir-
ected for universal cytokine-mediated killing’’ (TRUCK) to
secrete the proinflammatory cytokine IL-12 to stimulate innate

Cancer Cell 38, October 12, 2020 ª 2020 Elsevier Inc. 473
 ll
immune cells against the tumor and resist inhibitory elements of
the TME, including regulatory T (Treg) cells and myeloid-derived
suppressor cells (MDSCs) (Pegram et al., 2012). The secretion of
other soluble factors has been studied, including IL-15 or IL-18
to enhance T cell proliferation, as well as the combination of
CCL19 and IL-7 to recruit endogenous immune cells and estab-
lish a memory response against tumors (Adachi et al., 2018;
Hoyos et al., 2010; Hu et al., 2017).
In addition to the evolution of CAR designs outlined above, the
modularity of the four major components of a CAR—extracellular
antigen-sensing domain, extracellular hinge or spacer domain,
transmembrane domain, and intracellular signaling domain—
has enabled further optimization of each of these components
to improve the efficacy of CAR-T cell therapy. These engineering
efforts are well-summarized in other reviews (Labanieh et al.,
2018; Rafiq et al., 2020). Here, we focus our attention on
strategies that enable T cells to expand beyond the hard-wired,
single-input, single-output signaling capability programmed by
conventional CAR designs (Figure 2).
Combinatorial Antigen Sensing for Logic-Gated T Cell
Activation
Boolean logic gates have been utilized for the combinatorial
detection of multiple antigens by CAR-T cells to improve their
safety and antitumor efficacy (Figure 2A). AND-gate logic requires
the copresence of two different antigens to activate the CAR-T
cell, and this increased specification reduces the risk of either
off-target recognition or ‘‘on-target, off-tumor’’ toxicities, in which
healthy tissues that express the same antigen as tumor cells suffer
collateral damage. The synthetic Notch (synNotch) receptor—
which triggers inducible target-gene expression upon recognition
of a cell surface-bound ligand—was engineered to recognize a
TAA and induce the expression of a CAR, which can subsequently
trigger T-cell activation upon recognizing a second TAA (Roybal
et al., 2016). This strategy has been shown to reduce systemic
toxicity compared with constitutive CAR expression, provided
that the off-tumor target is not spatially proximal to the tumor cells
(Srivastava et al., 2019). Since there is a temporal delay between
the recognition of TAA #1 by synNotch and the recognition of
TAA #2 by CAR, a given T cell could have its synNotch receptor
triggered by TAA #1 from a tumor cell but subsequently attack a
healthy cell expressing TAA #2. An alternative AND-gate approach
474 Cancer Cell 38, October 12, 2020
Review
Figure 1. CAR-T Cell Engineering
Approaches
Strategies to engineer CAR-T cells for improved
function in solid tumors include a focus on CAR
engineering, T cell engineering, and TME interac-
tion optimization.
separates the CD3z chain and costimula-
tory domain into two constitutively ex-
pressed receptors each recognizing
different antigens, such that the CAR-T
cell is optimally activated only in the simul-
taneous presence of both antigens (Kloss
et al., 2013; Wilkie et al., 2012). However,
this approach often suffers from ‘‘leaki-
ness’’ due to the fact that first-generation
CARs containing only the CD3z chain are already signaling-
competent. Yet another strategy programs T cells to deliver a
conditionally active cytotoxic protein upon CAR- or TCR-medi-
ated detection of TAA #1 on the cell surface; the engineered pro-
tein becomes cytotoxic if and only if it detects TAA #2 inside the
target cell, thus requiring both antigens to be expressed by the
same target cell to trigger robust killing (Ho et al., 2017).
CAR-T cells programmed to execute AND-NOT logic can also
help prevent toxicities against healthy cells. This strategy utilizes
an inhibitory CAR (iCAR) that targets an antigen found on healthy
tissue, and pairs it with an activating CAR that targets a TAA. In a
proof-of-principle study, a prostate-specific membrane antigen
(PSMA)-targeting iCAR that incorporates the programmed cell
death protein 1 (PD-1) inhibitory signaling domain was coexpressed
with a second-generation CD19 CAR, and the iCAR inhibited CAR-
T cell activation in the presence of PSMA (Fedorov et al., 2013).
While AND and AND-NOT logic can improve the safety of CAR-T
cells by increasing specificity, OR-gate logic has been utilized to
increase antitumor efficacy by circumventing antigen escape, or
loss of the targeted epitope by tumor cells. An OR-gate CAR can
recognize two different TAAs, and the binding of either antigen in-
duces T-cell activation. One OR-gate strategy utilizes the pooled
mixture of two populations of CAR-T cells (CARpool), each ex-
pressing a monospecific CAR. A variation on this theme is to
sequentially administer two different CAR-T cell products (Shah
et al., 2020; Shalabi et al., 2018). Another strategy is the coexpres-
sion of two separate CARs in each T cell (dual CAR) (Ruella et al.,
2016). Yet another approach uses tandem bispecific CARs (Tan-
CAR) that comprise two scFv domains separated by a linker on
one receptor chain, and this strategy was shown to be functionally
superior to both the CARpool and dual-CAR approaches (Hegde
et al., 2016; Zah et al., 2020). In particular, CD19/CD20 and
CD19/CD22 bispecific CARs have been characterized for the
treatment of B-cell malignancies (Fry et al., 2018; Qin et al.,
2018; Zah et al., 2016) and are in clinical trials for lymphoma and
acute lymphocytic leukemia, respectively (NCT04007029,
NCT04215016, NCT03919526, NCT04303520).
ON/OFF Switches for Controllability and Safety
CAR modifications to improve safety and controllability have
also taken the form of externally inducible or self-regulating
ON/OFF switches. Conventional CAR-T cells are ‘‘always on,’’

Review
meaning their CARs are constitutively expressed and are always
capable of signaling upon antigen stimulation. However, this is
not always desirable when CAR-derived toxicity is an anticipated
risk. In addition to the off-target and ‘‘on-target, off-tumor’’ tox-
icities discussed previously, various systemic toxicities have
also been observed in patients treated with CAR-T cells. Com-
mon examples include cytokine release syndrome (CRS) (Fitz-
gerald et al., 2017; Grupp et al., 2013; Schuster et al., 2017),
neurotoxicity or ‘‘CAR-related encephalopathy syndrome’’
(CRES) (Grupp et al., 2013; Lee et al., 2019a; Schuster et al.,
2017), tumor lysis syndrome, and anaphylaxis (Maus et al.,
2013). It has been hypothesized that modulation of CAR-T cell
activity in space and time can prevent or significantly lessen
the severity of toxicities associated with CAR-T cell therapy while
maintaining its antitumor efficacy.
One approach to modulate CAR-T cell activity is to regulate the
presence of functional CARs on the surface of engineered T cells
by adjusting either the stability or the conformation of the CAR
protein itself (Figure 2B). As examples of the former strategy,
CARs have been fused to degradation tags that can be inactivated
either by small-molecule binding (Weber et al., 2020) or under hyp-
oxic conditions (Juillerat et al., 2017). The default state in such sys-
tems is ‘‘off,’’ and the removal or inactivation of the degradation
tag is required to stabilize the CAR protein, thus enabling CAR-
mediated T-cell activation upon antigen stimulation. An alternative
design in which the CAR protein is present by default but turned
off through the administration of a small-molecule drug has also
been demonstrated (Juillerat et al., 2019).
Instead of modulating protein half-life, controlling the confor-
mation of the CAR protein can also regulate the availability of
functional CARs, such that the receptor is signaling-competent
only under specified conditions. For example, the antigen-bind-
ing domain of CARs can be ‘‘masked’’ by a built-in inhibitory
ll
Figure 2. Protein Engineering Strategies to
Improve the Programmability, Safety, and
Efficacy of CAR-T Cells
(A) Combinatorial antigen recognition by AND and
AND-NOT logic using a synNotch receptor and
iCAR, respectively, can increase antigen speci-
ficity and safety. Tandem bispecific OR-gate CARs
can circumvent antigen escape and increase effi-
cacy.
(B) Engineered ON and OFF switches can easily
and efficiently alter CAR-T cell activity.
(C) Programming CARs to activate only in the
presence of an adaptor or by leucine-zipper-
mediated reconstitution can increase controlla-
bility over CAR-T cell activity.
peptide, such that the CAR’s functional
conformation is acquired only after the
inhibitory peptide has been cleaved by
proteases commonly active in the TME
(Han et al., 2017).
Adapter-Dependent CARs
Alternatively, T cells can be engineered to
express a receptor that must be comple-
mented by an additional protein compo-
nent before the resulting complex is
capable of translating antigen recognition to T-cell activation
(Figure 2C). Urbanska et al. (2012) reported a ‘‘universal’’ recep-
tor comprising a biotin-binding domain fused to an intracellular
T-cell signaling domain. T cells expressing this biotin-specific re-
ceptor can, in principle, be directed against any target cell that
has been labeled with a biotinylated antibody. Two advantages
of this strategy include the abilities to (1) control the ON/OFF
state of the T cell through administering or withholding the bio-
tinylated antibody and (2) use the same biotin-specific receptor
to target a wide variety of TAAs by changing the specificity of
the biotinylated antibody. The concept of using a universal
CAR coupled with one or more antigen-targeting adaptor pro-
teins to overcome tumor heterogeneity and mitigate toxicity
soon opened the floodgates for other adapter-dependent CAR
designs (Bachmann, 2019; Cartellieri et al., 2016; Herzig et al.,
2019; Lee et al., 2019c; Qi et al., 2020; Raj et al., 2019; Rodgers
et al., 2016), including a small-molecule adapter, which binds to
both fluorescein isothiocyanate (FITC) and folate, that alleviated
CRS-like toxicity in an NSG mouse model (Lee et al., 2019b).
Expanding on the concept of constitutively expressing a uni-
versal receptor on the T-cell surface combined with an externally
administered adaptor protein to dictate antigen specificity, Cho
et al. (2018) reported a SUPRA CAR system that can program
a variety of Boolean logic gates in engineered T cells by express-
ing multiple base receptors, each with multiple potential adaptor
protein partners reconstituted by leucine-zipper dimerization.
Such a system supports the possibility of simultaneously
increasing specificity through the use of AND or AND-NOT gates
while addressing tumor heterogeneity by targeting multiple anti-
gens with different adaptor proteins. However, the versatility of
multi-component systems comes at the cost of an increased
number of parameters that must be optimized, including the
half-life, biodistribution, and interaction dynamics among the
Cancer Cell 38, October 12, 2020 475
 ll
adaptor protein, base receptor, and the engineered T cell itself.
Therefore, it remains to be seen whether the complex signal pro-
cessing achievable through adaptor-dependent CAR designs
will translate to robust therapeutic candidates.
ENGINEERING THE CAR-EXPRESSING CELL
Safety Controls on CAR-T Cell Activity
In addition to engineering the CAR protein itself, engineering ap-
proaches applied at a cellular level to modulate T-cell activity can
also significantly affect therapeutic efficacy and safety. For
example, transient CAR expression resulting from mRNA elec-
troporation instead of viral integration can mitigate toxicities
induced by CARs that cross-recognize healthy tissue (Beatty
et al., 2014; Tchou et al., 2017; Wiesinger et al., 2019). Another
safeguard against severe toxicity is the implementation of sui-
cide genes that enable the depletion of engineered T cells by
the administration of small-molecule drugs (Figure 3A) (Marin
et al., 2012). For example, CAR-T cells targeting the CD44v6 an-
tigen for the treatment of leukemia and myeloma have been en-
gineered to express herpes simplex virus thymidine kinase (HSV-
TK), which enabled effective elimination of the CAR-T cells upon
exposure to ganciclovir in vitro (Casucci et al., 2018). However,
expression of viral HSV-TK raises concerns of immunogenicity
and requires 3 days of ganciclovir exposure to achieve effective
T-cell elimination (Marin et al., 2012). An alternative approach us-
ing the inducible caspase 9 (iCasp9) suicide gene has been
shown to eliminate >90% of engineered cells within 30 min of
drug administration in human patients (Di Stasi et al., 2011).
iCasp9 contains the intracellular domain of the pro-apoptotic
protein caspase 9 fused to FK506-binding protein (FKBP). The
small molecule AP1930 facilitates the dimerization of FKBP
and activation of the fused caspase 9, inducing apoptosis in cells
expressing the iCasp9 protein (Straathof et al., 2005). In addition,
the expression of transgenes for the surface proteins CD20 or
truncated epidermal growth factor receptor (tEGFR) can also
facilitate suicide mechanisms. The US Food and Drug Adminis-
tration-approved mAbs rituximab and cetuximab bind to CD20
and tEGFR, respectively, and the resulting antibody-dependent
cellular cytotoxicity can be used to eliminate T cells engineered
to express these antigens (Griffioen et al., 2009; Serafini et al.,
2004; Wang et al., 2011).
While suicide genes can efficiently deplete CAR-T cells to
counter toxicities, the activation of a suicide gene also results
in the irreversible termination of the therapy. An alternative
reversible strategy utilizes dasatinib, a tyrosine kinase inhibitor
476 Cancer Cell 38, October 12, 2020
Review
Figure 3. Engineering Strategies to Improve
CAR-T Cell Safety
(A) Coexpression of suicide genes such as HSK-
TV, iCasp9, CD20, and tEGFR enables induction of
T-cell death to abort the therapy in the case of
adverse events.
(B) Tet-ON and -OFF systems allow the control of
the CAR expression on the transcriptional level.
that interferes with the lymphocyte-spe-
cific protein kinase (LCK), thus inhibiting
CD3z phosphorylation and CAR activa-
tion. Dasatinib has been shown to function as a reversible ON/
OFF switch of CAR-T cell activity, where the cessation of dasa-
tinib administration rapidly reversed its inhibitory effects, high-
lighting its potential as an emergency drug for potentially lethal
toxicities such as CRS and CRES (Mestermann et al., 2019;
Weber et al., 2019).
Regulated CAR Expression to Improve CAR-T Cell Safety
As an alternative to post-translational regulation of CAR stability
and function, transcriptional regulation can provide another
tunable handle to improve CAR-T cell safety (Figure 3B). For
example, in the Tet-ON system, the small molecule doxycycline
(Dox) acts as an ‘‘ON switch,’’ whereby CAR expression is
induced by the reverse tetracycline transactivator (rtTA) protein
only in the presence of Dox (Sakemura et al., 2016). Conversely,
in the Tet-OFF system, Dox acts as an ‘‘OFF switch’’ by abolish-
ing the ability of tetracycline transactivator (tTA) to activate CAR
transcription; this system was used to reversibly inhibit delete-
rious CAR signaling and T-cell fratricide in CD5 CAR-T cells (Ma-
monkin et al., 2018). In addition, response to a tumor-associated
environmental cue could be achieved through inducible pro-
moters responsive to hypoxia-inducible factor (HIF)-1a, acti-
vating CAR expression only in hypoxic environments (Ede
et al., 2016). Finally, as previously described, the synNotch re-
ceptor enables inducible CAR transcription upon binding to a
membrane-bound ligand (Roybal et al., 2016; Srivastava
et al., 2019).
Site-Specific CAR Transgene Insertion and Allogeneic
Compatibility Engineering
The examples of regulated gene expression provided above uti-
lize synthetic inducible promoters, and the entire gene-expres-
sion cassette is typically integrated into the T-cell genome using
retroviral or lentiviral vectors, leading to variable integration sites
and copy numbers. An alternative method to achieve dynamic
CAR expression profiles is to integrate the transgene into
specific genetic loci regulated by endogenous transcriptional
machinery. A variety of gene-editing technologies, including
clustered regularly interspaced short palindromic repeats
(CRISPR) and CRISPR-associated protein 9 (Cas9), transcription
activator-like effector nucleases (TALENs), and zinc-finger nu-
cleases (ZFNs), have made this a feasible approach in T-cell en-
gineering (Chen, 2015). For example, Eyquem et al. (2017)
demonstrated that CRISPR-Cas9-mediated insertion of the
CD19 CAR transgene into the TCRa constant (TRAC) locus re-
sults in CAR-T cells with superior in vivo function compared
 ll
Review

Table 1. Targeted Genome-Editing Strategies to Enhance T-Cell Table 1. Continued

Function Target Tech
Target Tech locusa Motivationb nologyc Reference
locusa Motivationb nologyc Reference FAS abolish pro- CRISPR- (Ren
TRAC ablate TCRab ZFN (Torikai apoptotic Cas9 et al., 2017b)
expression to et al., 2012) signaling
reduce TALEN (Qasim GM- inhibit CRS- CRISPR- (Sterner
alloreactivity et al., 2017; CSF related Cas9 et al., 2019)
Valton toxicities
et al., 2015) a
TRAC, T-cell receptor alpha constant; TRBC, T-cell receptor beta con-
CRISPR- (Georgiadis stant; B2M, Beta-2 microglobulin; HLA, human leukocyte antigen; dCK,
Cas9 et al., 2018) deoxycytidine kinase; PD-1, programmed cell death protein 1; DGK, di-
CRISPR- (Eyquem acylglycerol kinase; LAG3, lymphocyte-activation gene 3; GM-CSF,
Cas9 et al., 2017; granulocyte-macrophage colony-stimulating factor.
b
CAR MacLeod TCR, T-cell receptor; CRS, cytokine release syndrome.
c
knockin et al., 2017) ZFN, zinc-finger nuclease; TALEN, transcription activator-like effector
nuclease; CRISPR, clustered regularly interspaced short palindromic re-
TRAC, enhance CRISPR- (Stadtmauer
peats; Cas9, CRISPR-associated protein 9.
TRBC transgenic Cas9 et al., 2020)
TCR
expression with retrovirus-mediated random CAR-transgene insertion,
B2M ablate HLA CRISPR- (Ren although more recent evidence suggests that whether site-spe-
expression to Cas9 et al., 2017b) cific CAR integration into the TRAC locus improves T-cell func-
reduce tion may depend on the specific CAR construct used (Zah
alloreactivity et al., 2020). A compelling example of the transgene insertion
HLA-A ablate HLA ZFN (Torikai site influencing CAR-T cell function comes from the case of a
expression to et al., 2013) chronic lymphocytic leukemia patient, whose disease regression
reduce
was observed to coincide with the clonal expansion of T cells
alloreactivity
carrying the CD19 CAR transgene inserted into the methylcyto-
CD52 confer TALEN (Qasim sine dioxygenase ten-eleven translocation 2 (TET2) locus
resistance to et al., 2017)
(Fraietta et al., 2018). Combined with a pre-existing hypomorphic
lympho
mutation in the patient’s other TET2 allele, this CAR insertion re-
depletion
sulted in the loss of wild-type TET2, a gene that encodes for a
dCK confer TALEN (Valton
regulator of DNA methylation and blood cell formation. This
resistance to et al., 2015)
lympho
fortuitous CAR insertion/TET2 ablation event led to an altered
depletion epigenetic landscape that conferred profound proliferative
capability and a central-memory phenotype to the engineered
PD-1 inhibit TALEN (Menger
immune- et al., 2016) T cells (Carty et al., 2018; Fraietta et al., 2018).
checkpoint Gene-editing technologies have also enabled the elimination
CRISPR- (Liu
signaling Cas9 et al., 2017; of endogenous genes in support of allogeneic T-cell therapy (Ta-
Ren ble 1). T-cell products derived from healthy donors can over-
et al., 2017a; come manufacturing challenges associated with autologous
Rupp cell therapy, such as the difficulty in obtaining enough high-qual-
et al., 2017; Stadtmauer ity T cells from heavily pretreated patients with advanced dis-
et al., 2020) ease. However, allogeneic T-cell transfer requires the elimination
REG disrupt a CRISPR- (Wei of the endogenous TCR to prevent graft-versus-host disease
NASE-1 negative Cas9 et al., 2019) (GvHD), and the removal of class-I major histocompatibility com-
regulator of plex (MHC-I) has been proposed to minimize allograft rejection
T-cell activity (Liu et al., 2017). To this end, Torikai et al. (2012, 2013) demon-
DGKa, disrupt a CRISPR- (Jung strated ZFN-mediated abrogation of TCRab or human leukocyte
DGKz negative Cas9 et al., 2018) antigen (HLA)-A expression in CD19 CAR-T cells. In addition to
regulator of preventing GvHD, gene editing has been utilized to protect
T-cell activity
CAR-T cells from lymphodepletion, which is a common precon-
LAG3 disrupt a CRISPR- (Zhang ditioning treatment administered prior to CAR-T cell infusion to
negative Cas9 et al., 2017)
improve the efficacy of the transferred cells. For example,
regulator of
TALEN-mediated simultaneous disruptions of TCRab/CD52 or
T-cell activity
TCRab/deoxycytidine kinase (dCK) have been shown to confer
CD19 CAR-T cells with resistance to anti-CD52 or dCK phos-
phorylation-dependent lymphodepleting regimens, respectively
(Qasim et al., 2017; Valton et al., 2015).

Cancer Cell 38, October 12, 2020 477

 ll
In the studies referenced above, endogenous gene disruption
and CAR transgene integration were independently executed,
yielding heterogeneous CAR-T cell populations. Georgiadis
et al. (2018) coupled gene-editing to CAR integration by incorpo-
rating a single-guide RNA (sgRNA) element into the U3 region of
the 30 long terminal repeat sequence of the CAR-encoding lenti-
viral vector. After electroporation of Cas9 mRNA, magnetic
bead-based selection for edited cells resulted in highly enriched
CAR+ TCR populations. In another strategy, Ren et al. (2017b)
used a ‘‘one-shot’’ CRISPR system with multiple sgRNA
expression cassettes in one CAR-encoding lentiviral vector to
simultaneously knock out the endogenous TCR and Beta-2 mi-
croglobulin (b2M), an essential subunit of the HLA-I molecule,
to generate CD19 CAR-T cells suitable for allogeneic therapy.
Knockout of Negative Regulators
Gene-editing technologies can also be used to abrogate the
expression of negative regulators of T-cell activity (Table 1). Tu-
mor cells frequently upregulate ligands to immune-checkpoint
receptors, such as cytotoxic T-lymphocyte-associated antigen
4 (CTLA-4) and PD-1 expressed on T cells, leading to inhibition
of T-cell activity in the TME (Buchbinder and Desai, 2016). Im-
mune-checkpoint blockade using antibodies targeting CTLA-4,
PD-1, or PD-1 ligand (PD-L1) has revolutionized the field of im-
muno-oncology in recent years (Wei et al., 2018). An alternative
approach to checkpoint blockade is the ablation of checkpoint-
receptor expression on engineered T cells. Menger et al. (2016)
demonstrated that PD-1 knockout through TALEN-mediated ed-
iting in melanoma-reactive CD8+ T cells and fibrosarcoma-reac-
tive polyclonal T cells enhanced the persistence of the modified
T cells, augmenting their antitumor activity against syngeneic tu-
mor models and establishing long-term antitumor memory. In
addition, CRISPR-Cas9-mediated disruption of PD-1 in CD19
CAR-T cells enhanced CAR-T cell cytotoxicity toward PD-L1+ tu-
mor xenografts in vivo (Rupp et al., 2017). The first CRISPR-
Cas9-edited cell therapy trial conducted in the United States
evaluated the adoptive transfer of autologous T cells genetically
modified to express a New York esophageal squamous cell car-
cinoma 1 (NY-ESO-1)-targeting TCR while eliminating endoge-
nous TCRab and PD-1 expression; recent results from the trial
confirmed the safety and feasibility of CRISPR-edited cell ther-
apy for cancer (Stadtmauer et al., 2020).
In addition to PD-1 knockout, elimination of the metabolic
regulator REGNASE-1 and CD3-signaling regulator diacylgly-
cerol kinase (DGK) has also been shown to enhance T-cell func-
tion in vitro and in vivo (Jung et al., 2018; Riese et al., 2013; Wei
et al., 2019). Similar to PD-1, lymphocyte-activation gene 3
(LAG3) is known as a T-cell exhaustion marker, and it has been
found to inhibit the effector function of T cells and enhance the
suppressive function of Tregs (Zhang et al., 2017). However,
CRISPR-Cas9-mediated deletion of LAG3 in CD19 CAR-T cells
did not result in a discernable phenotype, perhaps due to
compensatory effects from PD-1 (Woo et al., 2012; Zhang
et al., 2017).
Several studies have reported the feasibility of multiple knock-
outs using CRISPR-Cas9 in CAR-T cells to achieve the dual
goals of reducing alloreactivity and improving T-cell function. Tri-
ple-targeting of the TRAC, B2M, and PD-1 loci has been demon-
strated by electroporation of Cas9:sgRNA ribonucleoprotein
478 Cancer Cell 38, October 12, 2020
Review
(RNP) complex (Liu et al., 2017) or by electroporation of mRNA
encoding Cas9 and sgRNAs into CAR-T cells (Ren et al.,
2017a). Furthermore, triple-targeting of the TRAC, B2M, and
FAS loci using a lentiviral vector encoding both the sgRNAs
and a CD19-targeting CAR and electroporation of Cas9 mRNA
resulted in reduced alloreactivity and prolonged T-cell survival
by abrogating pro-apoptotic Fas/FasL signaling (Ren et al.,
2017b). Finally, CRISPR-Cas9-mediated deletion has been
used to improve the safety of CAR-T cells by disrupting the
expression of CRS-associated cytokines, such as granulocyte-
macrophage colony-stimulating factor (GM-CSF) (Sterner
et al., 2019).
Receptors that Rewire an Inhibitory Input to a
Stimulatory Output
While CAR-T cells have shown promising therapeutic efficacy
against hematologic malignancies, their targeting of solid tumors
has been stymied by a number of challenges. In addition to
antigen heterogeneity, tumor cells produce tumorigenic and
immunosuppressive factors in the TME, and this inhibitory envi-
ronment is further compounded by immunosuppressive cell
types such as MDSCs and Tregs (Labanieh et al., 2018). Among
the immunosuppressive soluble factors commonly found in the
TME, transforming growth factor b (TGF-b) is a particularly
potent cytokine that drives T-cell differentiation into Tregs, as
well as macrophage polarization to the immunosuppressive
M2 phenotype (Tormoen et al., 2018). The engineering of a
TGF-b-responsive CAR demonstrated that CARs can be used
to (1) sense a soluble factor and (2) rewire T cells to convert an
inhibitory signal to a trigger of antitumor activity (Figure 4) (Chang
et al., 2018; Hou et al., 2018). TGF-b-responsive CAR-T cells
were demonstrated to proliferate and produce T helper type 1
(Th1)-associated cytokines in the presence of soluble TGF-b,
as well as protect nearby cells from the inhibitory effects of
TGF-b, likely through the combined effects of TGF-b sequestra-
tion and paracrine Th1 cytokine signaling (Chang et al., 2018;
Hou et al., 2018).
Signal rewiring can also be achieved with ‘‘switch receptors,’’
which are chimeras comprising an ectodomain that binds a sup-
pressive molecule fused to an endodomain that drives a stimula-
tory pathway (Figure 4). IL-4 is a cytokine with complex roles in
the TME, such as inducing M2 polarization, promoting tumor
growth, and suppressing tumor-specific effector T cells (Tor-
moen et al., 2018; Wilkie et al., 2010). IL-4 switch receptors con-
sisting of the IL-4 receptor a (IL-4Ra) ectodomain fused to either
the IL-7Ra endodomain or the bc receptor subunit common to IL-
2 and IL-15 signaling have been shown to paradoxically enhance
T-cell proliferation in the presence of IL-4 (Leen et al., 2014;
Wilkie et al., 2010). Consequently, the coexpression of the IL-
4Ra:bC switch receptor in CAR-T cells augmented their anti-
tumor capacity (Wilkie et al., 2010). More recently, Roth et al.
(2020) used a pooled knockin screen using CRISPR-Cas9
coupled with single-cell RNA sequencing (scRNA-seq) to eval-
uate a panel of transgenes integrated into the TRAC locus.
Through this analysis, a novel TGFbR2:4-1BB switch receptor
was identified as the lead candidate for improved T-cell fitness
and solid-tumor clearance.
Switch receptors can also mitigate the suppressive effects
of immune checkpoints. For example, CTLA-4:CD28 and

Review
PD-1:CD28 switch receptors were shown to enhance the activity
of tumor-specific T cells (Liu et al., 2016; Shin et al., 2012). As
another checkpoint receptor, T-cell immunoreceptor with Ig
and ITIM domains (TIGIT) binds to CD155 or CD112 on tumor
cells, and these interactions hinder cytokine production and
effector function of T cells. To blunt these effects, Hoogi et al.
(2019) designed a TIGIT:CD28 switch receptor that enhanced
cytokine production and activation of T cells, as well as delayed
tumor growth in vivo when combined with a melanoma-spe-
cific TCR.
Yet another set of engineering strategies focuses on address-
ing tumorigenic factors that are not directly expressed on T cells
or tumor cells (Figure 4). For example, vascular endothelial
growth factor-A (VEGF-A) is a tumor-derived soluble factor that
promotes tumor growth and metastasis by facilitating angiogen-
esis. Chinnasamy et al. (2012) cotransduced T cells with a CAR
targeting VEGF receptor 2 (VEGFR2) and an inducible transgene
encoding IL-12. These CAR-T cells demonstrated trafficking to
the tumor vasculature, elimination of VEGFR+ MDSC subtypes
that participate in tumor angiogenesis, and IL-12-mediated
solid-tumor regression. In addition, T cells expressing a CAR
that targets fibroblast activation protein (FAP) eliminated FAP+
tumor stromal fibroblasts that support tumor growth, augment-
ing T-cell immunotherapy (Wang et al., 2014). However, strate-
gies targeting non-tumor tissues can cause significant toxicity.
For example, Tran et al. (2013) reported that CAR-T cells
cross-reactive to human and mouse FAP cause off-target bone
marrow toxicities and cachexia.
Transgene Expression to Promote T-Cell Function
In addition to abolishing or rewiring inhibitory signals, the overex-
pression of stimulatory signals can also improve the activity of
tumor-specific T cells. The constitutive expression of costimula-
tory ligands CD80 and 4-1BBL in PSMA-targeting CAR-T cells
showed superior activity against prostate tumor cells by
engaging costimulatory receptors in cis (autocostimulation)
and in trans (transcostimulation of bystander cells) (Stephan
et al., 2007). Furthermore, Zhao et al. (2015) coexpressed the
4-1BBL costimulatory ligand with a second-generation CD19-
targeting CAR with CD28 costimulation, and this combined cos-
timulation led to improved antitumor functions driven by the
continuous activation of the interferon regulatory factor 7
(IRF7)/interferon beta (IFNb) pathway. Yet another strategy har-
nesses the endogenous IL-7 signaling mechanism that confers
improved persistence in tumor-specific T cells. Shum et al.
(2017) engineered a constitutively active IL-7 receptor and coex-
pressed it with a GD2-targeting CAR, which resulted in
enhanced survival under repeated tumor challenges in neuro-
blastoma and glioblastoma xenograft models.
ll
Figure 4. Rewiring T-Cell Signaling with
Synthetic Receptors
Switch receptors rewire T cell responses by trig-
gering costimulatory signaling in the presence of
normally inhibitory ligands. CARs responsive to
environmental cues such as soluble TGF-b or
surface antigens present on tumor-supportive tis-
sues can enhance antitumor function by removing
and converting immunosuppressive factors.
The aberrant expression of the colony-stimulating factor 1
(CSF-1) in the TME drives macrophages to the M2 phenotype
and promotes tumor growth. Several clinical trials of small mol-
ecules and mAbs targeting the CSF-1/CSF-1R axis in combina-
tion with other mAbs and/or chemotherapy are under way for the
treatment of solid tumors (NCT01525602, NCT02777710,
NCT02323191, NCT02760797, NCT02923739). In the context
of CAR-T cells, the coexpression of CSF-1R, which T cells do
not naturally express, was shown to confer CSF-1 responsive-
ness and activated the RAS/MEK/Erk kinase pathway to
enhance T-cell proliferation, cytokine production, and CSF-1-
driven chemotaxis (Lo et al., 2008).
Metabolic Reprogramming of T Cells
T-cell metabolism, or the manner in which T cells utilize nutrient
sources, has consequential effects on their differentiation state
and effector function. The architecture of the CAR protein can
have an impact on the metabolic profiles of CAR-T cells. For
example, T cells expressing CARs with 4-1BB costimulation
favor the oxidative breakdown of fatty acids characteristic of
the central-memory phenotype, accompanied by enhanced pro-
liferation and persistence, while T cells expressing CARs with
CD28 costimulation favor aerobic glycolysis characteristic of
the effector-memory phenotype (Kawalekar et al., 2016).
Due to the intimate relationship between T-cell metabolism
and function, reprogramming the metabolic profile of CAR-T
cells can potentially increase their clinical efficacy. The TME of
solid tumors has an overabundance of potassium (K+) released
by necrotic tumor cells, which increases the intracellular K+ con-
centration in infiltrating T cells, downregulating Protein kinase B
(Akt)/mammalian target of rapamycin (mTOR) signaling and im-
pairing T-cell activation after TCR ligation. As a counterstrategy,
the overexpression of K+ channels was shown to increase Akt/
mTOR activity and rescue T-cell effector functions by facilitating
K+ efflux and lowering intracellular K+ levels (Eil et al., 2016). In
addition, elevated K+ in the TME and the resulting perturbation
of the transmembrane electrochemical gradient limit nutrient up-
take by T cells. Interestingly, ex vivo conditioning and activation
of tumor-specific CD8+ T cells in elevated K+ to mimic such func-
tional starvation in the TME led to epigenetic and metabolic re-
programming that maintained T-cell stemness—demonstrated
by improved persistence, engraftment, self-renewal, and multi-
potency—thereby enhancing their antitumor function in vivo
(Vodnala et al., 2019). Further, Geiger et al. (2016) showed sup-
plementing L-arginine to balance increased arginine metabolism
in activated T cells promotes the central-memory phenotype and
improves antitumor activity.
Altering the expression levels of metabolic genes can also pro-
mote an advantageous metabolic profile. Phosphoenolpyruvate
Cancer Cell 38, October 12, 2020 479
 ll
carboxykinase 1 (PKC1) increases the production of phospho-
enolpyruvate (PEP), which sustains TCR-mediated, Ca2+-
induced nuclear factor of activated T cells (NFAT) signaling and
effector functions, and the overexpression of PKC1 in T cells
has been shown to restrict tumor growth in melanoma-bearing
mice (Ho et al., 2015). Leukemic cells drive T-cell dysfunction
by causing suppressed Akt/mTORC1 signaling, decreased
expression of the glucose transporter Glut1, and reduced
glucose uptake. Accordingly, the overexpression of Akt or
Glut1 was demonstrated to partially rescue T-cell activity (Siska
et al., 2016). In addition, Yang et al. (2016) knocked out Acetyl-
CoA acetyltransferase (ACAT1), a cholesterol esterification
enzyme, in CD8+ T cells, and the resulting increase in the plasma
membrane cholesterol concentration enhanced TCR clustering
and signaling. Lastly, PPAR-gamma coactivator 1a (PGC1a) is
a metabolic regulator downregulated in tumor-infiltrating
T cells. It facilitates mitochondrial biogenesis by transcriptional
coactivation, and its overexpression in CD8+ T cells rescued
their mitochondrial function and protected their metabolic and
effector activities in the TME (Scharping et al., 2016).
Interplay of T-Cell Phenotypes, Function, and Versatility
The differentiation state of T cells influences their longevity and
efficacy, motivating the isolation or enrichment of specific T-
cell subtypes in CAR-T cell manufacturing. Generally, the selec-
tion of less differentiated phenotypes—naive (TN), memory stem
(TSCM), and central memory (TCM)—imparts greater engraftment
and efficacy than the more differentiated counterparts—effector
(TE) and effector memory (TEM) (Sadelain et al., 2017). Different
costimulatory domains in the CAR protein can affect T-cell sub-
type distribution: CD28 costimulation tends to promote the
short-lived, potent TEM phenotype, while 4-1BB results in enrich-
ment of the longer-lived, self-renewing TCM phenotype (Kawale-
kar et al., 2016). It has also been shown that TN, TCM, and TEM
CD4+ and CD8+ T cells can all be transduced and expanded
as CD19 CAR-T cells, but the combination of CD8+ TCM and
CD4+ TN subsets yields synergistic antitumor activity in vivo
(Sommermeyer et al., 2016). TSCM cells comprise only 2%–3%
of peripheral blood mononuclear cells, but this memory subset
possesses the highest self-renewal capacity and superior
persistence, and they are suggested to be the primary precur-
480 Cancer Cell 38, October 12, 2020
Review
Figure 5. Strategies in Optimizing CAR-T
Cell and Tumor Interactions
CAR-T cells have been engineered to utilize,
reverse, or circumvent tumor-driven immunosup-
pressive factors and axes through a variety of
mechanisms.
sors of T-cell memory establishment (Hur-
ton et al., 2016). IL-15 is a pro-survival
cytokine fundamental to T-cell memory,
and it can preserve a TSCM-like phenotype
by inhibiting mTORC1 activity, reducing
glycolysis, and improving mitochondrial
fitness (Alizadeh et al., 2019). Hurton
et al. (2016) incorporated IL-15 costimula-
tion in CAR-T cells by coexpressing a
membrane-bound chimeric IL-15, which
led to a TSCM-like molecular profile with improved T-cell persis-
tence regardless of CAR stimulation. In addition, the miR-17-
92 microRNA cluster was found to be upregulated in IFNg-pro-
ducing Th1 cells compared with T helper type 2 (Th2) cells,
and it was found to be downregulated in T cells derived from glio-
blastoma patients (Sasaki et al., 2010). To induce Th1-like
phenotype in glioblastoma-targeting CAR-T cells, Ohno et al.
(2013) cotransduced miR-17-92 with a third-generation EGFR-
vIII-specific CAR; in combination with the chemotherapeutic
agent temozolomide, this strategy led to improved cytolytic ac-
tivity and protection against tumor re-challenge in vivo.
Stem cells are a versatile starting material for adoptively trans-
ferred cellular products due to their abilities to self-renew and
differentiate into various cell types. Schmitt and Zúñiga-Pflu€cker
(2002) showed that OP9-DL1, a bone marrow stromal cell line
that ectopically expresses the Notch ligand Delta-like-1, can
induce the differentiation of hematopoietic progenitor cells
(HPCs) into T lymphocytes. In a subsequent study, functional
CD8+ T cells were generated from human umbilical cord blood
hematopoietic stem cells, which possess greater self-renewal
and potency than HPCs, in OP9-DL1 cocultures (Awong et al.,
2011). More recently, an artificial thymic organoid system has
been shown to facilitate in vitro differentiation of human embry-
onic stem cells (hESCs) and induced pluripotent stem cells
(iPSCs) into mature TN-like cells with potent antitumor efficacy
and a similar transcriptional profile to that of primary CD8+ TN
cells (Montel-Hagen et al., 2019).
Expansion of CAR Effectors beyond T Cells
Exosomes derived from CAR-T cells, rather than the T cells
themselves, have been studied as an alternative effector for
CAR-mediated antigen specificity and cytotoxicity. Exosomes
released by CAR-T cells can carry the CAR and a high level of
cytotoxic molecules; these tumor-specific cytotoxic packages
can traffic into solid tumors due to their nanoscale size, while
incurring a lower risk of CRS-related toxicities and conferring
protection against the immune-checkpoint molecule PD-L1
due to their lack of PD-1 expression (Fu et al., 2019).
Programming CARs into cell types other than T cells can
further expand the versatility of the therapy by realizing new
functions unachievable by CAR-T cells. Klinchinsky et al.

Review
(2020) recently demonstrated the feasibility of adenoviral trans-
duction of a CAR into primary macrophages. The resulting
CAR-M cells exhibited tumor-specific phagocytosis, inflamma-
tory cytokine production, polarization of bystander macro-
phages to the immunostimulatory M1 phenotype, and cross-pre-
sentation of the TAA to bystander T cells. Although a comparison
between CAR-T and CAR-M cells was not evaluated, the estab-
lished role of macrophages as professional antigen-presenting
cells (APCs) warrants the potential of CAR-M cells to more effec-
tively stimulate an adaptive antitumor immune response.
CAR-natural killer (NK) cells can be generated from cord blood
or iPSCs (Li et al., 2018; Liu et al., 2020), making them an attrac-
tive candidate for allogeneic, off-the-shelf products. Moreover,
CD19-targeting CAR-NK cells have achieved robust clinical effi-
cacy without inducing CRS, neurotoxicity, or GvHD in patients
with B-cell lymphoid tumors, highlighting their relative safety
compared with their T-cell counterparts (Liu et al., 2020). CAR-
NK cells have been shown to exert potent and specific cytotox-
icity toward a variety of tumor models, including leukemia, mul-
tiple myeloma, ovarian cancer, and glioblastoma (Chu et al.,
2014; Genbler et al., 2016; Li et al., 2018; Quintarelli et al.,
2020), as well as toward immunosuppressive cell types such
as MDSCs and follicular helper T cells (TFH) (Parihar et al.,
2019; Reighard et al., 2020). Lastly, natural killer T (NKT) cells
possess antitumor and tumor-homing capabilities, and GD2-tar-
geting CAR-NKT cells that harness these inherent advantages
exhibited effective localization to and lysis of neuroblastoma
cells without significant toxicity (Xu et al., 2019). Taken together,
these developments highlight both the potential for cell-based
immunotherapy to expand beyond T cells and the applicability
of the CAR technology across a variety of immune cell types.
ENGINEERING CAR-T CELL INTERACTIONS WITH THE
TUMOR AND TME
The immune-evasive and immunosuppressive nature of the TME
contributes to the poor therapeutic efficacy of CAR-T cells
observed in solid tumors. Hallmarks of the TME, which have
been extensively reviewed elsewhere (Gajewski et al., 2013;
Whiteside, 2008), include (1) physical barriers to tumor penetra-
tion by immune cells, (2) upregulated checkpoint ligands, (3) a
pro-tumor stromal niche, (4) abundant immunosuppressive and
pro-metastatic soluble factors, and (5) modulated expression
of chemokines to preferentially recruit leukocytes with an immu-
nosuppressive phenotype. These factors have in turn driven the
design of CAR-T cells that respond to TME elements to enhance
CAR-T cell efficacy (Figure 5).
Tumor Homing and Penetration
The efficacy of CAR-T cell therapy in solid tumors is significantly
hindered by poor immune-cell infiltration (Newick et al., 2017). T-
cell migration is regulated through chemokine axes. Tumor cells
can upregulate or downregulate chemokines, as well as modu-
late chemokine expression by tumor-associated cells, contrib-
uting to the poor recruitment of CAR-T cells (Oelkrug and Ram-
age, 2014). Engineering CAR-T cells to overexpress receptors
for chemokines that are overexpressed in the TME can turn a tu-
mor’s defense mechanism against itself. For example, GD2- and
mesothelin-targeting CAR-T cells have been engineered to co-
ll
express CCR2b, the dominant isoform of the chemokine recep-
tor of CCL2, resulting in enhanced T-cell homing to CCL2-ex-
pressing neuroblastoma and malignant pleural mesothelioma
xenografts, respectively (Craddock et al., 2010; Moon et al.,
2011). Similarly, CAR-T cells that coexpress CCR4 showed
improved migration toward tumors expressing CCL17 and
CCL22 in vivo, while those expressing CXCR1 or CXCR2 ex-
hibited enhanced homing toward tumor-derived IL-8 (Di Stasi
et al., 2009; Jin et al., 2019). Once CAR-T cells reach the tumor
site, their infiltration is hindered by the high-density structural
extracellular matrix (ECM) associated with solid-tumor nodules.
Accordingly, CAR-T cells engineered to express heparinase, an
enzyme that degrades ECM, have been shown to improve tumor
infiltration and overall survival in multiple xenograft models (Car-
uana et al., 2015).
CAR-T cells that successfully reach solid tumors are next
faced with a multitude of suppressive and evasive features that
induce CAR-T cell dysfunction (Newick et al., 2017). To improve
therapeutic efficacy in this immunosuppressive environment,
CAR-T cells have been engineered to produce proteins that (1)
improve CAR-T cell function in an autocrine fashion; (2) disrupt
immunosuppressive elements; and/or (3) induce TME remodel-
ing to enhance the endogenous antitumor immune response
(Figure 5). Each of these strategies is discussed in detail below.
Autocrine Stimulation of CAR-T Cells in the TME
Cytokines are signaling proteins with the ability to drastically
augment or abrogate CAR-T cell function. Coexpressing the
CAR with immunostimulatory cytokines could significantly
enhance CAR-T cell proliferation, survival, and effector function
in the immunosuppressive TME. For example, T cells that consti-
tutively coexpress a CD19-targeting CAR plus IL-2, IL-7, IL-15,
or IL-21 have been shown to achieve greater in vivo tumor con-
trol compared with T cells expressing the CAR alone (Markley
and Sadelain, 2010). Interestingly, although the receptor com-
plexes of these four cytokines contain the common gamma
chain (gc), each cytokine differentially affected the proliferation,
subtype differentiation, and function of the engineered T cells,
underscoring the complexity of T-cell biology and variety of
potential outcomes achievable through different engineering
strategies (Markley and Sadelain, 2010). T cells coexpressing
IL-12, IL-15, IL-18, and/or IL-21 plus a CAR targeting a variety
of antigens have also been described, resulting in improved effi-
cacy, proliferation, and/or persistence in vivo (Batra et al., 2020;
Hoyos et al., 2010; Hu et al., 2017; Koneru et al., 2015; Krenciute
et al., 2017; Pegram et al., 2012). However, constitutive overex-
pression of immunostimulatory cytokines can also increase
toxicity (Zhang et al., 2011). Regulatory strategies previously dis-
cussed in this review, such as inducible promoters, can be im-
plemented to modulate cytokine production and associated
toxicity (Liu et al., 2019).
Disruption of Immunosuppressive Axes
The expression of immune-checkpoint receptors and ligands
such as PD-1 and PD-L1 are prevalent in the TME, and they
can potently inhibit CAR-T cell cytotoxicity and induce anergy
(Drake et al., 2006). Thus, immune-checkpoint blockade has
strong synergistic potential with CAR-T cell therapy, and several
ongoing clinical trials are evaluating combination therapy with
Cancer Cell 38, October 12, 2020 481
 ll
CAR-T cells and exogenously administered checkpoint inhibi-
tors (Grosser et al., 2019). Furthermore, CAR-T cells have been
engineered to secrete immune-checkpoint inhibitors, including
anti-PD-1 scFvs and antiPD-L1 antibodies, or to express PD-1
dominant-negative receptors (DNRs) (Chen et al., 2017; Cher-
kassky et al., 2016; Li et al., 2017; Rafiq et al., 2018; Suarez
et al., 2016). In addition to enhancing efficacy, this approach
may also avoid toxicities associated with systemic immune-
checkpoint blockade by restricting checkpoint inhibitor distribu-
tion to the immediate environment of the producer T cells. For
example, it has been shown that anti-PD-1 scFvs secreted by
intraperitoneally (IP) injected CAR-T cells remained localized at
the injection site. However, when an equal number of conven-
tional CAR-T cells were administered IP with exogenous anti-
PD-1 antibody, the antibody was detected systemically within
3 h (Rafiq et al., 2018).
The solid-tumor milieu also houses a diverse collection of sol-
uble factors that promote tumorigenesis and inhibit CAR-T cell
function. For example, prostaglandin E2 (PGE2) is a bioactive
lipid often upregulated in tumors, where it contributes to tumor
survival through regulation of cell proliferation, migration,
apoptosis, and angiogenesis (Ricciotti and FitzGerald, 2011;
Wang and Dubois, 2006). In the context of CAR-T cell therapy,
PGE2, along with adenosine, inhibits T-cell signaling and activa-
tion though the activation of protein kinase A (PKA), thereby
reducing T-cell proliferation and effector function (Newick
et al., 2016). In two solid-tumor models that highly express
PGE2, CAR-T cells engineered to express a peptide inhibitor of
ezrin-mediated PKA translocation to the immune synapse ex-
hibited improved tumor infiltration and killing (Newick et al.,
2016). Similarly, elevated concentrations of bioreactive chemi-
cals such as reactive oxygen species (ROS) in the TME play an
important role in tumorigenesis (Weinberg et al., 2019). Catalase
is an enzyme that facilitates the decomposition of hydrogen
peroxide (H2O2), an ROS that impairs T-cell activity in the TME.
Increasing intracellular levels of catalase by coexpressing the
catalase gene in HER2-and carcinoembryonic antigen (CEA)-
specific CAR-T cells has been shown to enable CAR-T cells to
metabolize the suppressive H2O2, improving their tumor cytolytic
capacity (Ligtenberg et al., 2016).
The aberrant expression of cytokines in the TME plays a crit-
ical role in tumor progression and resistance to CAR-T cell ther-
apy. In particular, TGF-b plays a multiplexed role in cancer pro-
gression through interactions with tumor cells, stroma, and both
innate and adaptive immune cells to induce (1) the secretion of
immunosuppressive chemokines, cytokines, and growth fac-
tors; (2) ECM remodeling and matrix deposition; (3) immunosup-
pressive reprogramming of macrophages, neutrophils, and
T cells; and (4) inhibited maturation or proliferation of T cells
and NK cells (Pickup et al., 2013). To ablate these powerful ef-
fects, CAR-T cells have been engineered to express a TGF-b
DNR that potently inhibits endogenous TGF-b signaling, result-
ing in T cells with enhanced proliferation and antitumor efficacy
in a prostate cancer xenograft model (Kloss et al., 2018). Based
on these results, a phase 1 clinical trial has been initiated to
assess T cells coexpressing a PSMA CAR and the DNR for the
treatment of relapsed and refractory metastatic prostate cancer
(NCT03089203). The DNR is distinct from the TGF-b-targeting
CAR and TGF-b switch receptor discussed in a previous section
482 Cancer Cell 38, October 12, 2020
Review
in that the DNR does not transduce any signals that can stimulate
the engineered T cell. It remains to be seen whether the stimula-
tory effects of the CAR and switch receptors will confer addi-
tional clinical benefits compared with the DNR.
In the TME, IL-6 is often overexpressed by tumor cells, tumor-
associated macrophages (TAMs), and other resident cells (Ku-
mari et al., 2016). IL-6 supports tumorigenesis through a number
of mechanisms and plays a central role in the induction of CRS
after CAR-T cell infusion (Lee et al., 2014). Systemic administra-
tion of tocilizumab, an mAb targeting IL-6 receptor alpha (IL-
6Ra), has become standard treatment for CRS after CAR-T cell
therapy (Kotch et al., 2019). More recently, CD19-targeting
CAR-T cells that coexpress a non-signaling, membrane-bound
IL-6 receptor (mbaIL6) were shown to sequester IL-6 while
retaining in vivo antitumor efficacy (Tan et al., 2020). However,
it remains to be seen if CAR-T cells engineered in this fashion
can prevent CRS.
TME Remodeling to Promote the Endogenous Immune
Response
Tumors are adept at selectively attracting or evading subsets of
leukocytes, including CAR-T cells, to promote immune regulation
or suppression (Rabinovich et al., 2007). In addition, tumors are
often capable of inducing an immunosuppressive or pro-metasta-
tic phenotype on the local stroma, as well as an anti-inflammatory
or dysfunctional phenotype on resident leukocytes (Morgan and
Schambach, 2018). Another approach to enhancing the efficacy
of CAR-T cell therapy is to reverse this immunosuppressive cell
niche through remodeling the tumor-cellular composition and
phenotype. To realize this, CAR-T cells have been engineered to
secrete cytokines or other soluble factors that induce TME remod-
eling in a paracrine or endocrine fashion.
In germinal-center lymphomas, loss of herpesvirus entry
mediator (HVEM) expression induces the secretion of non-
redundant stroma-activating factors, resulting in acute
lymphoid-stroma activation. The hyperactivated stroma recruits
TFH cells, which support malignant B cells through CD40/CD40L
interactions and cytokine stimulation. As a counterstrategy,
CD19 CAR-T cells engineered to secrete a soluble form of
HVEM were shown to enhance tumor control in vivo (Boice
et al., 2016).
CAR-T cells engineered to secrete IL-12 have been shown to
remodel the TME by reprogramming TAMs to an M1 phenotype
and decreasing the presence of MDSCs and Tregs in syngeneic
mouse models (Chinnasamy et al., 2012; Liu et al., 2019; Yeku
et al., 2017). Similarly, CAR-T cells that constitutively secrete
IL-18 can alter the TME makeup by increasing intratumoral M1
macrophage, activated dendritic cell (DC), and activated NK
cell numbers, while decreasing M2 macrophage and Treg levels.
A direct comparison of CAR-T cells expressing IL-12- to IL-18
indicated that IL-18 is more effective at remodeling the immuno-
suppressive TME in a syngeneic murine pancreatic cancer
model (Chmielewski and Abken, 2017). Furthermore, CD19
CAR-T cells expressing IL-18 induced the expansion of endoge-
nous CD8+ T cells, NK cells, NKT cells, and DCs in the bone
marrow, potentially contributing to the control of tumors with
heterogeneous CD19 expression in a syngeneic mouse model
(Avanzi et al., 2018).

Review
Among DCs, conventional type 1 DCs (cDC1s) in particular
excel at inducing immunity against tumors via their ability to
cross-present cellular antigens and prime Th1 cells. Recently,
it has been shown that T cells engineered to secrete Fms-like
tyrosine kinase 3 ligand (Flt3L), a hematopoietic cell growth fac-
tor, promote intratumoral cDC1 and DC-precursor proliferation
(Lai et al., 2020). Furthermore, when T cells were cotransduced
to express Flt3L and an anti-HER2 CAR, a combined treatment
with these CAR-T cells and adjuvants induced an enhanced anti-
tumor response and endogenous T-cell epitope spreading in vivo
(Lai et al., 2020).
CAR-T cells have also been engineered to coexpress multiple
immune-modulatory proteins. In one example, CAR-T cells were
programmed to coexpress CCL19 and IL-7 to induce endoge-
nous immune-cell recruitment and stimulate the recruited cells,
respectively. In a syngeneic hCD20-expressing mastocytoma
mouse model, these CAR-T cells induced robust recruitment of
endogenous T cells and DCs, resulting in enhanced and durable
tumor clearance (Adachi et al., 2018).
CAR-T cells have also been designed to modulate the TME
through the expression of surface-bound proinflammatory ligands.
For example, CD40L is normally transiently expressed on T cells
after TCR stimulation, and its interaction with the CD40 receptor
on different immune cell types can lead to activation of APCs,
licensing of DCs, as well as apoptosis of CD40+ tumor cells (Cella
et al., 1996; Eliopoulos et al., 2000; Ridge et al., 1998). Constitutive
CD40L expression on CD19 CAR-T cells resulted in elevated sur-
face expression of costimulatory molecules, adhesion molecules,
HLA molecules, and the Fas death receptor on CD40+ tumor cells,
thus increasing their immunogenicity (Curran et al., 2015). These
T cells also induced the secretion of proinflammatory IL-12 by
monocyte-derived DCs in vitro and showed enhanced antitumor
efficacy in vitro and in vivo (Curran et al., 2015). It was subsequently
demonstrated that CD40L-expressing CAR-T cells can license
APCs in lymphatic tissues in a syngeneic immunocompetent
mouse model, and this licensing was found to be dependent on
the CD40L/CD40 interaction (Kuhn et al., 2019). Furthermore,
increased recruitment of macrophages, DCs, and endogenous
CD4+ and CD8+ T cells to lymphatic tissues was observed, along
with the recruitment of DCs, CD4+ and CD8+ T cells to the tumor.
Treg levels were also observed to slightly increase in the tumor,
but the ratio of CD8+ T cells to Tregs was unchanged. Thus,
CD40L-expressing CAR-T cells capable of remodeling the TME
and lymphatic tissue activated endogenous T cells to suppress an-
tigen-negative tumor re-challenge, strongly suggesting induced
epitope spreading (Kuhn et al., 2019). Similarly, the surface expres-
sion of 4-1BBL on CAR-T cells is proposed to remodel the TME
through autocrine-induced secretion of type I IFNs, which may
improve DC cross-priming, Treg inhibition, and angiogenesis sup-
pression (Zhao et al., 2015).
Finally, CAR-T cells can be engineered to facilitate the engage-
ment of tumor cells by endogenous, non-engineered T cells
through the secretion of bispecific T-cell engagers (BiTEs), which
are composed of two fused scFvs. Choi et al. (2019) engineered
BiTEs with one scFv targeting EGFR, which is overexpressed in
glioblastoma cells, and the other targeting CD3 on T cells. EGFR-
vIII-targeting CAR-T cells engineered to secrete EGFR/CD3 BiTEs
have been shown to eliminate orthotopic tumor xenografts with
heterogeneous EGFRvIII expression.
ll
Conclusion
CAR-T cell therapy has shown great promise in treating hemato-
logic malignancies. However, solid tumors pose unique
challenges that require further engineering and tuning of the
technology to successfully treat these intractable malignancies.
Recent protein- and cell-engineering strategies have made great
strides in boosting the intrinsic fitness and antitumor function of
T cells, increasing tumor-targeting specificity, preventing tumor
escape and relapse, as well as modifying the TME to enhance
immunotherapeutic outcomes. Although most engineering stra-
tegies reported to date have focused on delivering individual
desirable features, advancements in genome-editing methodol-
ogies and genetic circuitry design offer the possibility of multi-
layered approaches that can simultaneously address multiple
critical needs in T-cell therapeutics development.
At the same time, the biological complexity of and potential
crosstalk among different engineered features within the T cell,
as well as among engineered and endogenous immune cells, tu-
mor cells, and other tumor-associated factors, must be carefully
balanced when advancing the clinical translation of CAR-T cells
for the treatment of solid tumors. The decreasing cost and
increasing capacity of next-generation and single-cell sequencing
methods, as well as proteomic and metabolomic analyses, could
significantly enhance our ability to understand and rationally
manipulate these complex interactions while engineering the
next generation of CAR-T cell therapy for solid malignancies. The
growing toolbox of T-cell engineering strategies that can be syner-
gistically implemented and modularly calibrated for maximum
safety and efficacy will continue to enable innovations that aim to
generate new treatment options for currently intractable diseases.
ACKNOWLEDGMENTS
M.H. and J.D.C. are supported by the Parker Institute for Cancer Immuno-
therapy Center at UCLA and the Cancer Research Institute (grants to Y.Y.C.).
DECLARATION OF INTERESTS
Y.Y.C. is the inventor of multiple patents related to CAR-T cell technologies,
some of which are mentioned in this review article.
REFERENCES
Adachi, K., Kano, Y., Nagai, T., Okuyama, N., Sakoda, Y., and Tamada, K.
(2018). IL-7 and CCL19 expression in CAR-T cells improves immune cell infil-
tration and CAR-T cell survival in the tumor. Nat. Biotechnol. 36, 346–351.
Alizadeh, D., Wong, R.A., Yang, X., Wang, D., Pecoraro, J.R., Kuo, C.F., Agui-
lar, B., Qi, Y., Ann, D.K., Starr, R., et al. (2019). IL15 enhances CAR-T cell anti-
tumor activity by reducing mTORC1 activity and preserving their stem cell
memory phenotype. Cancer Immunol. Res. 7, 759–772.
Avanzi, M.P., Yeku, O., Li, X., Wijewarnasuriya, D.P., van Leeuwen, D.G., Cheung,
K., Park, H., Purdon, T.J., Daniyan, A.F., et al. (2018). Engineered tumor-targeted
T cells mediate enhanced anti-tumor efficacy both directly and through activation
of the endogenous immune system. Cell Rep. 23, 2130–2141.
€cker, J.C. (2011).
Awong, G., Herer, E., La Motte-Mohs, R.N., and Zúñiga-Pflu
Human CD8 T cells generated in vitro from hematopoietic stem cells are func-
tionally mature. BMC Immunol. 12, 22.
Bachmann, M. (2019). The UniCAR system: a modular CAR T cell approach to
improve the safety of CAR T cells. Immunol. Lett. 211, 13–22.
Batra, S.A., Rathi, P., Guo, L., Courtney, A.N., Fleurence, J., Balzeau, J., Shaik,
R.S., Nguyen, T.P., Wu, M.-F., Bulsara, S., et al. (2020). Glypican-3-specific
CAR T cells co-expressing IL15 and IL21 have superior expansion and
Cancer Cell 38, October 12, 2020 483
 ll
antitumor activity against hepatocellular carcinoma. Cancer Immunol. Res.
https://doi.org/10.1158/2326-6066.CIR-19-0293.
Beatty, G.L., Haas, A.R., Maus, M.V., Torigian, D.A., Soulen, M.C., Plesa, G.,
Chew, A., Zhao, Y., Levine, B.L., Albelda, S.M., et al. (2014). Mesothelin-spe-
cific chimeric antigen receptor mRNA-engineered T cells induce anti-tumor
activity in solid malignancies. Cancer Immunol. Res. 2, 112–120.
Boice, M., Salloum, D., Mourcin, F., Sanghvi, V., Amin, R., Oricchio, E., Jiang,
M., Mottok, A., Denis-Lagache, N., Ciriello, G., et al. (2016). Loss of the HVEM
tumor suppressor in lymphoma and restoration by modified CAR-T cells. Cell
167, 405–418.e13.
Bouchkouj, N., Kasamon, Y.L., de Claro, R.A., George, B., Lin, X., Lee, S., Blu-
menthal, G.M., Bryan, W., McKee, A.E., and Pazdur, R. (2019). FDA approval
summary: axicabtagene ciloleucel for relapsed or refractory large B-cell lym-
phoma. Clin. Cancer Res. 25, 1702–1708.
Buchbinder, E.I., and Desai, A. (2016). CTLA-4 and PD-1 pathways: similar-
ities, differences, and implications of their inhibition. Am. J. Clin. Oncol.
39, 98–106.
Cartellieri, M., Feldmann, A., Koristka, S., Arndt, C., Loff, S., Ehninger, A., von
Bonin, M., Bejestani, E.P., Ehninger, G., and Bachmann, M.P. (2016). Switch-
ing CAR T cells on and off: a novel modular platform for retargeting of T cells to
AML blasts. Blood Cancer J. 6, e458.
Carty, S.A., Gohil, M., Banks, L.B., Cotton, R.M., Johnson, M.E., Stelekati, E.,
Wells, A.D., Wherry, E.J., Koretzky, G.A., and Jordan, M.S. (2018). The loss of
TET2 promotes CD8+ T cell memory differentiation. J. Immunol. 200, 82–91.
Caruana, I., Savoldo, B., Hoyos, V., Weber, G., Liu, H., Kim, E.S., Ittmann,
M.M., Marchetti, D., and Dotti, G. (2015). Heparanase promotes tumor infiltra-
tion and antitumor activity of CAR-redirected T lymphocytes. Nat. Med. 21,
524–529.
Casucci, M., Falcone, L., Camisa, B., Norelli, M., Porcellini, S., Stornaiuolo, A.,
Ciceri, F., Traversari, C., Bordignon, C., Bonini, C., and Bondanza, A. (2018).
Extracellular NGFR spacers allow efficient tracking and enrichment of fully
functional CAR-T cells co-expressing a suicide gene. Front. Immunol. 9, 507.
Cella, M., Scheidegger, D., Palmer-Lehmann, K., Lane, P., Lanzavecchia, A.,
and Alber, G. (1996). Ligation of CD40 on dendritic cells triggers production
of high levels of interleukin-12 and enhances T cell stimulatory capacity: T-T
help via APC activation. J. Exp. Med. 184, 747–752.
Chang, Z.L., Lorenzini, M.H., Chen, X., Tran, U., Bangayan, N.J., and Chen,
Y.Y. (2018). Rewiring T-cell responses to soluble factors with chimeric antigen
receptors. Nat. Chem. Biol. 14, 317–324.
Chen, L., and Flies, D.B. (2013). Molecular mechanisms of T cell co-stimulation
and co-inhibition. Nat. Rev. Immunol. 13, 227–242.
Chen, N., Morello, A., Tano, Z., and Adusumilli, P.S. (2017). CAR T-cell intrinsic
PD-1 checkpoint blockade: a two-in-one approach for solid tumor immuno-
therapy. Oncoimmunology 6, e1273302.
Chen, Y.Y. (2015). Efficient gene editing in primary human T cells. Trends Im-
munol. 36, 667–669.
Cherkassky, L., Morello, A., Villena-Vargas, J., Feng, Y., Dimitrov, D.S., Jones,
D.R., Sadelain, M., and Adusumilli, P.S. (2016). Human CAR T cells with cell-
intrinsic PD-1 checkpoint blockade resist tumor-mediated inhibition. J. Clin.
Invest. 126, 3130–3144.
Chinnasamy, D., Yu, Z., Kerkar, S.P., Zhang, L., Morgan, R.A., Restifo, N.P.,
and Rosenberg, S.A. (2012). Local delivery of interleukin-12 using T cells tar-
geting VEGF receptor-2 eradicates multiple vascularized tumors in mice.
Clin. Cancer Res. 18, 1672–1683.
Chmielewski, M., and Abken, H. (2017). CAR T cells releasing IL-18 convert to
T-Bet(high) FoxO1(low) effectors that exhibit augmented activity against
advanced solid tumors. Cell Rep. 21, 3205–3219.
Chmielewski, M., Hombach, A.A., and Abken, H. (2014). Of CARs and
TRUCKs: chimeric antigen receptor (CAR) T cells engineered with an inducible
cytokine to modulate the tumor stroma. Immunol. Rev. 257, 83–90.
Cho, J.H., Collins, J.J., and Wong, W.W. (2018). Universal chimeric antigen re-
ceptors for multiplexed and logical control of T cell responses. Cell 173, 1426–
1438.e11.
484 Cancer Cell 38, October 12, 2020
Review
Choi, B.D., Yu, X., Castano, A.P., Bouffard, A.A., Schmidts, A., Larson, R.C.,
Bailey, S.R., Boroughs, A.C., Frigault, M.J., Leick, M.B., et al. (2019). CAR-T
cells secreting BiTEs circumvent antigen escape without detectable toxicity.
Nat. Biotechnol. 37, 1049–1058.
Chu, J., Deng, Y., Benson, D.M., He, S., Hughes, T., Zhang, J., Peng, Y., Mao,
H., Yi, L., Ghoshal, K., et al. (2014). CS1-specific chimeric antigen receptor
(CAR)-engineered natural killer cells enhance in vitro and in vivo antitumor ac-
tivity against human multiple myeloma. Leukemia 28, 917–927.
Collinson-Pautz, M.R., Chang, W.C., Lu, A., Khalil, M., Crisostomo, J.W., Lin,
P.Y., Mahendravada, A., Shinners, N.P., Brandt, M.E., Zhang, M., et al. (2019).
Constitutively active MyD88/CD40 costimulation enhances expansion and ef-
ficacy of chimeric antigen receptor T cells targeting hematological malig-
nancies. Leukemia 33, 2195–2207.
Craddock, J.A., Lu, A., Bear, A., Pule, M., Brenner, M.K., Rooney, C.M., and
Foster, A.E. (2010). Enhanced tumor trafficking of GD2 chimeric antigen recep-
tor T cells by expression of the chemokine receptor CCR2b. J. Immunother.
33, 780–788.
Curran, K.J., Seinstra, B.A., Nikhamin, Y., Yeh, R., Usachenko, Y., van Leeu-
wen, D.G., Purdon, T., Pegram, H.J., and Brentjens, R.J. (2015). Enhancing
antitumor efficacy of chimeric antigen receptor T cells through constitutive
CD40L expression. Mol. Ther. 23, 769–778.
Di Stasi, A., De Angelis, B., Rooney, C.M., Zhang, L., Mahendravada, A., Fos-
ter, A.E., Heslop, H.E., Brenner, M.K., Dotti, G., and Savoldo, B. (2009). T lym-
phocytes coexpressing CCR4 and a chimeric antigen receptor targeting CD30
have improved homing and antitumor activity in a Hodgkin tumor model. Blood
113, 6392–6402.
Di Stasi, A., Tey, S.K., Dotti, G., Fujita, Y., Kennedy-Nasser, A., Martinez, C.,
Straathof, K., Liu, E., Durett, A.G., Grilley, B., et al. (2011). Inducible apoptosis
as a safety switch for adoptive cell therapy. N. Engl. J. Med. 365, 1673–1683.
Drake, C.G., Jaffee, E., and Pardoll, D.M. (2006). Mechanisms of immune
evasion by tumors. Adv. Immunol. 90, 51–81.
Ede, C., Chen, X., Lin, M.Y., and Chen, Y.Y. (2016). Quantitative analyses of
core promoters enable precise engineering of regulated gene expression in
mammalian cells. ACS Synth. Biol. 5, 395–404.
Eil, R., Vodnala, S.K., Clever, D., Klebanoff, C.A., Sukumar, M., Pan, J.H.,
Palmer, D.C., Gros, A., Yamamoto, T.N., Patel, S.J., et al. (2016). Ionic immune
suppression within the tumour microenvironment limits T cell effector function.
Nature 537, 539–543.
Eliopoulos, A.G., Davies, C., Knox, P.G., Gallagher, N.J., Afford, S.C., Adams,
D.H., and Young, L.S. (2000). CD40 induces apoptosis in carcinoma cells
through activation of cytotoxic ligands of the tumor necrosis factor superfam-
ily. Mol. Cell Biol. 20, 5503–5515.
Eshhar, Z., Waks, T., Gross, G., and Schindler, D.G. (1993). Specific activation
and targeting of cytotoxic lymphocytes through chimeric single chains con-
sisting of antibody-binding domains and the gamma or zeta subunits of the
immunoglobulin and T-cell receptors. Proc. Natl. Acad. Sci. U S A 90, 720–724.
Eyquem, J., Mansilla-Soto, J., Giavridis, T., van der Stegen, S.J., Hamieh, M.,
Cunanan, K.M., Odak, A., Gonen, M., and Sadelain, M. (2017). Targeting a
CAR to the TRAC locus with CRISPR/Cas9 enhances tumour rejection. Nature
543, 113–117.
Fedorov, V.D., Themeli, M., and Sadelain, M. (2013). PD-1– and CTLA-4–
based inhibitory chimeric antigen receptors (iCARs) divert off-target immuno-
therapy responses. Sci. Transl. Med. 5, 2–15ra172.
Fitzgerald, J.C., Weiss, S.L., Maude, S.L., Barrett, D.M., Lacey, S.F., Melen-
horst, J.J., Shaw, P., Berg, R.A., June, C.H., Porter, D.L., et al. (2017). Cytokine
release syndrome after chimeric antigen receptor T cell therapy for acute
lymphoblastic leukemia. Crit. Care Med. 45, e124–e131.
Fraietta, J.A., Nobles, C.L., Sammons, M.A., Lundh, S., Carty, S.A., Reich, T.J.,
Cogdill, A.P., Morrissette, J.J.D., DeNizio, J.E., Reddy, S., et al. (2018). Disrup-
tion of TET2 promotes the therapeutic efficacy of CD19-targeted T cells. Na-
ture 558, 307–312.
Fry, T.J., Shah, N.N., Orentas, R.J., Stetler-Stevenson, M., Yuan, C.M., Ram-
akrishna, S., Wolters, P., Martin, S., Delbrook, C., Yates, B., et al. (2018).
CD22-targeted CAR T cells induce remission in B-ALL that is naive or resistant
to CD19-targeted CAR immunotherapy. Nat. Med. 24, 20–28.

Review
Fu, W., Lei, C., Liu, S., Cui, Y., Wang, C., Qian, K., Li, T., Shen, Y., Fan, X., Lin,
F., et al. (2019). CAR exosomes derived from effector CAR-T cells have potent
antitumour effects and low toxicity. Nat. Commun. 10, 4355.
Gajewski, T.F., Schreiber, H., and Fu, Y.X. (2013). Innate and adaptive immune
cells in the tumor microenvironment. Nat. Immunol. 14, 1014–1022.
Geiger, R., Rieckmann, J.C., Wolf, T., Basso, C., Feng, Y., Fuhrer, T., Koga-
deeva, M., Picotti, P., Meissner, F., Mann, M., et al. (2016). L-arginine modu-
lates T cell metabolism and enhances survival and anti-tumor activity. Cell
167, 829–842.
Genßler, S., Burger, M.C., Zhang, C., Oelsner, S., Mildenberger, I., Wagner,
M., Steinbach, J.P., and Wels, W.S. (2016). Dual targeting of glioblastoma
with chimeric antigen receptor-engineered natural killer cells overcomes het-
erogeneity of target antigen expression and enhances antitumor activity and
survival. Oncoimmunology 5, e1119354.
Georgiadis, C., Preece, R., Nickolay, L., Etuk, A., Petrova, A., Ladon, D., Danyi,
A., Humphryes-Kirilov, N., Ajetunmobi, A., Kim, D., et al. (2018). Long terminal
repeat CRISPR-CAR-coupled ‘‘universal" T cells mediate potent anti-leukemic
effects. Mol. Ther. 26, 1215–1227.
Griffioen, M., van Egmond, E.H., Kester, M.G., Willemze, R., Falkenburg, J.H.,
and Heemskerk, M.H. (2009). Retroviral transfer of human CD20 as a suicide
gene for adoptive T-cell therapy. Haematologica 94, 1316–1320.
Grosser, R., Cherkassky, L., Chintala, N., and Adusumilli, P.S. (2019). Combi-
nation immunotherapy with CAR T cells and checkpoint blockade for the treat-
ment of solid tumors. Cancer Cell 36, 471–482.
Grupp, S.A., Kalos, M., Barrett, D., Aplenc, R., Porter, D.L., Rheingold, S.R.,
Teachey, D.T., Chew, A., Hauck, B., Wright, J.F., et al. (2013). Chimeric antigen
receptor-modified T cells for acute lymphoid leukemia. N. Engl. J. Med. 368,
1509–1518.
Guedan, S., Posey, A.D., Jr., Shaw, C., Wing, A., Da, T., Patel, P.R., McGetti-
gan, S.E., Casado-Medrano, V., Kawalekar, O.U., Uribe-Herranz, M., et al.
(2018). Enhancing CAR T cell persistence through ICOS and 4-1BB costimula-
tion. JCI Insight 3, e96976.
Han, X., Bryson, P.D., Zhao, Y., Cinay, G.E., Li, S., Guo, Y., Siriwon, N., and
Wang, P. (2017). Masked chimeric antigen receptor for tumor-specific activa-
tion. Mol. Ther. 25, 274–284.
Haynes, N.M., Trapani, J.A., Teng, M.W., Jackson, J.T., Cerruti, L., Jane, S.M.,
Kershaw, M.H., Smyth, M.J., and Darcy, P.K. (2002). Single-chain antigen
recognition receptors that costimulate potent rejection of established experi-
mental tumors. Blood 100, 3155–3163.
Hegde, M., Mukherjee, M., Grada, Z., Pignata, A., Landi, D., Navai, S.A., Wake-
field, A., Fousek, K., Bielamowicz, K., Chow, K.K., et al. (2016). Tandem CAR
T cells targeting HER2 and IL13Ra2 mitigate tumor antigen escape. J. Clin.
Invest. 126, 3036–3052.
Herzig, E., Kim, K.C., Packard, T.A., Vardi, N., Schwarzer, R., Gramatica, A.,
Deeks, S.G., Williams, S.R., Landgraf, K., Killeen, N., et al. (2019). Attacking
latent HIV with convertibleCAR-T cells, a highly adaptable killing platform.
Cell 179, 880–894.e10.
Ho, P., Ede, C., and Chen, Y.Y. (2017). Modularly constructed synthetic gran-
zyme B molecule enables interrogation of intracellular proteases for targeted
cytotoxicity. ACS Synth. Biol. 6, 1484–1495.
Ho, P.C., Bihuniak, J.D., Macintyre, A.N., Staron, M., Liu, X., Amezquita, R.,
Tsui, Y.C., Cui, G., Micevic, G., Perales, J.C., et al. (2015). Phosphoenolpyr-
uvate is a metabolic checkpoint of anti-tumor T cell responses. Cell 162,
1217–1228.
Hombach, A.A., Heiders, J., Foppe, M., Chmielewski, M., and Abken, H.
(2012). OX40 costimulation by a chimeric antigen receptor abrogates CD28
and IL-2 induced IL-10 secretion by redirected CD4+ T cells. Oncoimmunology
1, 458–466.
Hoogi, S., Eisenberg, V., Mayer, S., Shamul, A., Barliya, T., and Cohen, C.J.
(2019). A TIGIT-based chimeric co-stimulatory switch receptor improves T-
cell anti-tumor function. J Immunother Cancer 7, 243.
Hou, A.J., Chang, Z.L., Lorenzini, M.H., Zah, E., and Chen, Y.Y. (2018). TGF-b–
responsive CAR-T cells promote anti-tumor immune function. Bioeng. Transl.
Med. 3, 75–86.
ll
Hoyos, V., Savoldo, B., Quintarelli, C., Mahendravada, A., Zhang, M., Vera, J.,
Heslop, H.E., Rooney, C.M., Brenner, M.K., and Dotti, G. (2010). Engineering
CD19-specific T lymphocytes with interleukin-15 and a suicide gene to
enhance their anti-lymphoma/leukemia effects and safety. Leukemia 24,
1160–1170.
Hu, B., Ren, J., Luo, Y., Keith, B., Young, R.M., Scholler, J., Zhao, Y., and June,
C.H. (2017). Augmentation of antitumor immunity by human and mouse CAR
T cells secreting IL-18. Cell Rep. 20, 3025–3033.
Hurton, L.V., Singh, H., Najjar, A.M., Switzer, K.C., Mi, T., Maiti, S., Olivares, S.,
Rabinovich, B., Huls, H., Forget, M.A., et al. (2016). Tethered IL-15 augments
antitumor activity and promotes a stem-cell memory subset in tumor-specific
T cells. Proc. Natl. Acad. Sci. U S A 113, E7788–E7797.
Imai, C., Mihara, K., Andreansky, M., Nicholson, I.C., Pui, C.H., Geiger, T.L.,
and Campana, D. (2004). Chimeric receptors with 4-1BB signaling capacity
provoke potent cytotoxicity against acute lymphoblastic leukemia. Leukemia
18, 676–684.
Jin, L., Tao, H., Karachi, A., Long, Y., Hou, A.Y., Na, M., Dyson, K.A., Grippin,
A.J., Deleyrolle, L.P., Zhang, W., et al. (2019). CXCR1- or CXCR2-modified
CAR T cells co-opt IL-8 for maximal antitumor efficacy in solid tumors. Nat.
Commun. 10, 4016.
Juillerat, A., Marechal, A., Filhol, J.M., Valogne, Y., Valton, J., Duclert, A.,
Duchateau, P., and Poirot, L. (2017). An oxygen sensitive self-decision making
engineered CAR T-cell. Sci. Rep. 7, 39833.
Juillerat, A., Tkach, D., Busser, B.W., Temburni, S., Valton, J., Duclert, A.,
Poirot, L., Depil, S., and Duchateau, P. (2019). Modulation of chimeric antigen
receptor surface expression by a small molecule switch. BMC Biotechnol.
19, 44.
Jung, I.Y., Kim, Y.Y., Yu, H.S., Lee, M., Kim, S., and Lee, J. (2018). CRISPR/
Cas9-mediated knockout of DGK improves antitumor activities of human
T cells. Cancer Res. 78, 4692–4703.
Kalos, M., Levine, B.L., Porter, D.L., Katz, S., Grupp, S.A., Bagg, A., and June,
C.H. (2011). T cells with chimeric antigen receptors have potent antitumor ef-
fects and can establish memory in patients with advanced leukemia. Sci.
Transl Med. 3, 95ra73.
Kawalekar, O.U., O’Connor, R.S., Fraietta, J.A., Guo, L., McGettigan, S.E.,
Posey, A.D., Jr., Patel, P.R., Guedan, S., Scholler, J., Keith, B., et al. (2016).
Distinct signaling of coreceptors regulates specific metabolism pathways
and impacts memory development in CAR T cells. Immunity 44, 380–390.
Klinchinsky, M., Ruella, M., Shestova, O., Lu, X.M., Best, A., Zeeman, M.,
Schmierer, M., Gabrusiewicz, K., Anderson, N.R., Petty, N.E., et al. (2020). Hu-
man chimeric antigen receptor macrophages for cancer immunotherapy. Nat.
Biotechnol. https://doi.org/10.1038/s41587-020-0462-y.
Kloss, C.C., Condomines, M., Cartellieri, M., Bachmann, M., and Sadelain, M.
(2013). Combinatorial antigen recognition with balanced signaling promotes
selective tumor eradication by engineered T cells. Nat. Biotechnol. 31, 71–75.
Kloss, C.C., Lee, J., Zhang, A., Chen, F., Melenhorst, J.J., Lacey, S.F., Maus,
M.V., Fraietta, J.A., Zhao, Y., and June, C.H. (2018). Dominant-negative TGF-
beta receptor enhances PSMA-targeted human CAR T cell proliferation and
augments prostate cancer eradication. Mol. Ther. 26, 1855–1866.
Koneru, M., Purdon, T.J., Spriggs, D., Koneru, S., and Brentjens, R.J. (2015).
IL-12 secreting tumor-targeted chimeric antigen receptor T cells eradicate
ovarian tumors in vivo. Oncoimmunology 4, e994446.
Kotch, C., Barrett, D., and Teachey, D.T. (2019). Tocilizumab for the treatment
of chimeric antigen receptor T cell-induced cytokine release syndrome. Expert
Rev. Clin. Immunol. 15, 813–822.
Krenciute, G., Prinzing, B.L., Yi, Z., Wu, M.F., Liu, H., Dotti, G., Balyasnikova,
I.V., and Gottschalk, S. (2017). Transgenic expression of IL15 improves anti-
glioma activity of IL13Ralpha2-CAR T cells but results in antigen loss variants.
Cancer Immunol. Res. 5, 571–581.
Kuhn, N.F., Purdon, T.J., van Leeuwen, D.G., Lopez, A.V., Curran, K.J., Dan-
iyan, A.F., and Brentjens, R.J. (2019). CD40 ligand-modified chimeric antigen
receptor T cells enhance antitumor function by eliciting an endogenous anti-
tumor response. Cancer Cell 35, 473–488.e6.
Kumari, N., Dwarakanath, B.S., Das, A., and Bhatt, A.N. (2016). Role of inter-
leukin-6 in cancer progression and therapeutic resistance. Tumour Biol. 37,
11553–11572.
Cancer Cell 38, October 12, 2020 485
 ll
Kuwana, Y., Asakura, Y., Utsunomiya, N., Nakanishi, M., Arata, Y., Itoh, S., Na-
gase, F., and Kurosawa, Y. (1987). Expression of chimeric receptor composed
of immunoglobulin-derived V regions and T-cell receptor-derived C regions.
Biochem. Biophys. Res. Commun. 149, 960–968.
Labanieh, L., Majzner, R.G., and Mackall, C.L. (2018). Programming CAR-T
cells to kill cancer. Nat. Biomed. Eng. 2, 377–391.
Lai, J., Mardiana, S., House, I.G., Sek, K., Henderson, M.A., Giuffrida, L., Chen,
A.X.Y., Todd, K.L., Petley, E.V., Chan, J.D., et al. (2020). Adoptive cellular ther-
apy with T cells expressing the dendritic cell growth factor Flt3L drives epitope
spreading and antitumor immunity. Nat. Immunol. 21, 914–926.
Lee, D.W., Gardner, R., Porter, D.L., Louis, C.U., Ahmed, N., Jensen, M.,
Grupp, S.A., and Mackall, C.L. (2014). Current concepts in the diagnosis and
management of cytokine release syndrome. Blood 124, 188–195.
Lee, D.W., Santomasso, B.D., Locke, F.L., Ghobadi, A., Turtle, C.J., Brudno,
J.N., Maus, M.V., Park, J.H., Mead, E., Pavletic, S., et al. (2019a). ASTCT
consensus grading for cytokine release syndrome and neurologic toxicity asso-
ciated with immune effector cells. Biol. Blood Marrow Transpl. 25, 625–638.
Lee, Y.G., Chu, H., Lu, Y., Leamon, C.P., Srinivasarao, M., Putt, K.S., and Low,
P.S. (2019b). Regulation of CAR T cell-mediated cytokine release syndrome-
like toxicity using low molecular weight adapters. Nat. Commun. 10, 2681.
Lee, Y.G., Marks, I., Srinivasarao, M., Kanduluru, A.K., Mahalingam, S.M., Liu,
X., Chu, H., and Low, P.S. (2019c). Use of a single CAR T cell and several bis-
pecific adapters facilitates eradication of multiple antigenically different solid
tumors. Cancer Res. 79, 387–396.
Leen, A.M., Sukumaran, S., Watanabe, N., Mohammed, S., Keirnan, J., Yana-
gisawa, R., Anurathapan, U., Rendon, D., Heslop, H.E., Rooney, C.M., et al.
(2014). Reversal of tumor immune inhibition using a chimeric cytokine recep-
tor. Mol. Ther. 22, 1211–1220.
Li, S., Siriwon, N., Zhang, X., Yang, S., Jin, T., He, F., Kim, Y.J., Mac, J., Lu, Z.,
Wang, S., et al. (2017). Enhanced cancer immunotherapy by chimeric antigen
receptor-modified T cells engineered to secrete checkpoint inhibitors. Clin.
Cancer Res. 23, 6982–6992.
Li, Y., Hermanson, D.L., Moriarity, B.S., and Kaufman, D.S. (2018). Human
iPSC-derived natural killer cells engineered with chimeric antigen receptors
enhance anti-tumor activity. Cell Stem Cell 23, 181–192.
Ligtenberg, M.A., Mougiakakos, D., Mukhopadhyay, M., Witt, K., Lladser, A.,
Chmielewski, M., Riet, T., Abken, H., and Kiessling, R. (2016). Coexpressed
catalase protects chimeric antigen receptor-redirected T cells as well as
bystander cells from oxidative stress-induced loss of antitumor activity.
J. Immunol. 196, 759–766.
Liu, E., Marin, D., Banerjee, P., Macapinlac, H.A., Thompson, P., Basar, R.,
Nassif Kerbauy, L., Overman, B., Thall, P., Kaplan, M., et al. (2020). Use of
CAR-transduced natural killer cells in CD19-positive lymphoid tumors.
N. Engl. J. Med. 382, 545–553.
Liu, X., Ranganathan, R., Jiang, S., Fang, C., Sun, J., Kim, S., Newick, K., Lo,
A., June, C.H., Zhao, Y., and Moon, E.K. (2016). A chimeric switch-receptor
targeting PD1 augments the efficacy of second-generation CAR T cells in
advanced solid tumors. Cancer Res. 76, 1578–1590.
Liu, X., Zhang, Y., Cheng, C., Cheng, A.W., Zhang, X., Li, N., Xia, C., Wei, X.,
Liu, X., and Wang, H. (2017). CRISPR-Cas9-mediated multiplex gene editing
in CAR-T cells. Cell Res. 27, 154–157.
Liu, Y., Di, S., Shi, B., Zhang, H., Wang, Y., Wu, X., Luo, H., Wang, H., Li, Z., and
Jiang, H. (2019). Armored inducible expression of IL-12 enhances antitumor
activity of glypican-3-targeted chimeric antigen receptor-engineered T cells
in hepatocellular carcinoma. J. Immunol. 203, 198–207.
Lo, A.S., Taylor, J.R., Farzaneh, F., Kemeny, D.M., Dibb, N.J., and Maher, J.
(2008). Harnessing the tumour-derived cytokine, CSF-1, to co-stimulate T-
cell growth and activation. Mol. Immunol. 45, 1276–1287.
Long, A.H., Haso, W.M., Shern, J.F., Wanhainen, K.M., Murgai, M., Ingaramo,
M., Smith, J.P., Walker, A.J., Kohler, M.E., Venkateshwara, V.R., et al. (2015).
4-1BB costimulation ameliorates T cell exhaustion induced by tonic signaling
of chimeric antigen receptors. Nat. Med. 21, 581–590.
MacLeod, D.T., Antony, J., Martin, A.J., Moser, R.J., Hekele, A., Wetzel, K.J.,
Brown, A.E., Triggiano, M.A., Hux, J.A., Pham, C.D., et al. (2017). Integration of
a CD19 CAR into the TCR alpha chain locus streamlines production of alloge-
neic gene-edited CAR T cells. Mol. Ther. 25, 949–961.
486 Cancer Cell 38, October 12, 2020
Review
Mamonkin, M., Mukherjee, M., Srinivasan, M., Sharma, S., Gomes-Silva, D.,
Mo, F., Krenciute, G., Orange, J.S., and Brenner, M.K. (2018). Reversible trans-
gene expression reduces fratricide and permits 4-1BB costimulation of CAR
T cells directed to T-cell malignancies. Cancer Immunol. Res. 6, 47–58.
Marin, V., Cribioli, E., Philip, B., Tettamanti, S., Pizzitola, I., Biondi, A., Biagi, E.,
and Pule, M. (2012). Comparison of different suicide-gene strategies for the
safety improvement of genetically manipulated T cells. Hum. Gene Ther.
Methods 23, 376–386.
Markley, J.C., and Sadelain, M. (2010). IL-7 and IL-21 are superior to IL-2 and
IL-15 in promoting human T cell-mediated rejection of systemic lymphoma in
immunodeficient mice. Blood 115, 3508–3519.
Maude, S.L., Frey, N., Shaw, P.A., Aplenc, R., Barrett, D.M., Bunin, N.J., Chew,
A., Gonzalez, V.E., Zheng, Z., Lacey, S.F., et al. (2014). Chimeric antigen re-
ceptor T cells for sustained remissions in leukemia. N. Engl. J. Med. 371,
1507–1517.
Maus, M., Haas, A., Beatty, G., Albelda, S., Levine, B., Liu, X., Zhao, Y., Kalos,
M., and June, C. (2013). T cells expressing chimeric antigen receptors can
cause anaphylaxis in humans. Cancer Immunol. Res. 1, 26–31.
Menger, L., Sledzinska, A., Bergerhoff, K., Vargas, F.A., Smith, J., Poirot, L.,
Pule, M., Hererro, J., Peggs, K.S., and Quezada, S.A. (2016). TALEN-mediated
inactivation of PD-1 in tumor-reactive lymphocytes promotes intratumoral T-
cell persistence and rejection of established tumors. Cancer Res. 76,
2087–2093.
Mestermann, K., Giavridis, T., Weber, J., Rydzek, J., Frenz, S., Nerreter, T.,
Mades, A., Sadelain, M., Einsele, H., and Hudecek, M. (2019). The tyrosine ki-
nase inhibitor dasatinib acts as a pharmacologic on/off switch for CAR T cells.
Sci. Transl. Med. 11, eaau5907.
Montel-Hagen, A., Seet, C.S., Li, S., Chick, B., Zhu, Y., Chang, P., Tsai, S.,
Sun, V., Lopez, S., Chen, H.C., et al. (2019). Organoid-induced differentiation
of conventional T cells from human pluripotent stem cells. Cell Stem Cell 24,
376–389.
Moon, E.K., Carpenito, C., Sun, J., Wang, L.C., Kapoor, V., Predina, J., Powell,
D.J., Jr., Riley, J.L., June, C.H., and Albelda, S.M. (2011). Expression of a func-
tional CCR2 receptor enhances tumor localization and tumor eradication by
retargeted human T cells expressing a mesothelin-specific chimeric antibody
receptor. Clin. Cancer Res. 17, 4719–4730.
Morgan, M.A., and Schambach, A. (2018). Engineering CAR-T cells for
improved function against solid tumors. Front. Immunol. 9, 2493.
Moritz, D., Wels, W., Mattern, J., and Groner, B. (1994). Cytotoxic T lympho-
cytes with a grafted recognition specificity for ERBB2-expressing tumor cells.
Proc. Natl. Acad. Sci. U S A 91, 4318–4322.
Neelapu, S.S., Locke, F.L., Bartlett, N.L., Lekakis, L.J., Miklos, D.B., Jacob-
son, C.A., Braunschweig, I., Oluwole, O.O., Siddiqi, T., Lin, Y., et al. (2017). Ax-
icabtagene ciloleucel CAR T-cell therapy in refractory large B-cell lymphoma.
N. Engl. J. Med. 377, 2531–2544.
Newick, K., O’Brien, S., Moon, E., and Albelda, S.M. (2017). CAR T cell therapy
for solid tumors. Annu. Rev. Med. 68, 139–152.
Newick, K., O’Brien, S., Sun, J., Kapoor, V., Maceyko, S., Lo, A., Pure, E.,
Moon, E., and Albelda, S.M. (2016). Augmentation of CAR T-cell trafficking
and antitumor efficacy by blocking protein kinase A localization. Cancer Immu-
nol. Res. 4, 541–551.
O’Leary, M.C., Lu, X., Huang, Y., Lin, X., Mahmood, I., Przepiorka, D., Gavin,
D., Lee, S., Liu, K., George, B., et al. (2019). FDA approval summary: tisagen-
lecleucel for treatment of patients with relapsed or refractory B-cell precursor
acute lymphoblastic leukemia. Clin. Cancer Res. 25, 1142–1146.
Oelkrug, C., and Ramage, J.M. (2014). Enhancement of T cell recruitment and
infiltration into tumours. Clin. Exp. Immunol. 178, 1–8.
Ohno, M., Ohkuri, T., Kosaka, A., Tanahashi, K., June, C.H., Natsume, A., and
Okada, H. (2013). Expression of miR-17-92 enhances anti-tumor activity of T-
cells transduced with the anti-EGFRvIII chimeric antigen receptor in mice
bearing human GBM xenografts. J. Immunother. Cancer 1, 21.
Parihar, R., Rivas, C., Huynh, M., Omer, B., Lapteva, N., Metelitsa, L.S., Gott-
schalk, S.M., and Rooney, C.M. (2019). NK cells expressing a chimeric acti-
vating receptor eliminate MDSCs and rescue impaired CAR-T cell activity
against solid tumors. Cancer Immunol. Res. 7, 363–375.

Review
Pegram, H.J., Lee, J.C., Hayman, E.G., Imperato, G.H., Tedder, T.F., Sadelain,
M., and Brentjens, R.J. (2012). Tumor-targeted T cells modified to secrete IL-
12 eradicate systemic tumors without need for prior conditioning. Blood 119,
4133–4141.
Pickup, M., Novitskiy, S., and Moses, H.L. (2013). The roles of TGFbeta in the
tumour microenvironment. Nat. Rev. Cancer 13, 788–799.
Qasim, W., Zhan, H., Samarasinghe, S., Adams, S., Amrolia, P., Stafford, S.,
Butler, K., Rivat, C., Wright, G., Somana, K., et al. (2017). Molecular remission
of infant B-ALL after infusion of universal TALEN gene-edited CAR T cells. Sci.
Transl. Med. 9, eaaj2013.
Qi, J., Tsuji, K., Hymel, D., Burke, T.R., Hudecek, M., Rader, C., and Peng, H.
(2020). Chemically programmable and switchable CAR-T therapy. Angew.
Chem. Int. Ed. 59, 12178–12185.
Qin, H., Ramakrishna, S., Nguyen, S., Fountaine, T.J., Ponduri, A., Stetler-Ste-
venson, M., Yuan, C.M., Haso, W., Shern, J.F., Shah, N.N., and Fry, T.J. (2018).
Preclinical development of bivalent chimeric antigen receptors targeting both
CD19 and CD22. Mol. Ther. Oncolytics 11, 127–137.
Quintarelli, C., Sivori, S., Caruso, S., Carlomagno, S., Falco, M., Boffa, I., Or-
lando, D., Guercio, M., Abbaszadeh, Z., Sinibaldi, M., et al. (2020). Efficacy
of third-party chimeric antigen receptor modified peripheral blood natural killer
cells for adoptive cell therapy of B-cell precursor acute lymphoblastic leuke-
mia. Leukemia 34, 1102–1115.
Rabinovich, G.A., Gabrilovich, D., and Sotomayor, E.M. (2007). Immunosup-
pressive strategies that are mediated by tumor cells. Annu. Rev. Immunol.
25, 267–296.
Rafiq, S., Hackett, C.S., and Brentjens, R.J. (2020). Engineering strategies to
overcome the current roadblocks in CAR T cell therapy. Nat. Rev. Clin. Oncol.
17, 147–167.
Rafiq, S., Yeku, O.O., Jackson, H.J., Purdon, T.J., van Leeuwen, D.G., Drakes,
D.J., Song, M., Miele, M.M., Li, Z., et al. (2018). Targeted delivery of a PD-1-
blocking scFv by CAR-T cells enhances anti-tumor efficacy in vivo. Nat. Bio-
technol. 36, 847–856.
Raj, D., Yang, M.H., Rodgers, D., Hampton, E.N., Begum, J., Mustafa, A., Lor-
izio, D., Garces, I., Propper, D., Kench, J.G., et al. (2019). Switchable CAR-T
cells mediate remission in metastatic pancreatic ductal adenocarcinoma.
Gut 68, 1052–1064.
Reighard, S.D., Cranert, S.A., Rangel, K.M., Ali, A., Gyurova, I.E., de la Cruz-
Lynch, A.T., Tuazon, J.A., Khodoun, M.V., Kottyan, L.C., Smith, D.F., et al.
(2020). Therapeutic targeting of follicular T Cells with chimeric antigen recep-
tor-expressing natural killer cells. Cell Rep. Med. 1, 100003.
Ren, J., Liu, X., Fang, C., Jiang, S., June, C.H., and Zhao, Y. (2017a). Multiplex
genome editing to generate universal CAR T cells resistant to PD1 Inhibition.
Clin. Cancer Res. 23, 2255–2266.
Ren, J., Zhang, X., Liu, X., Fang, C., Jiang, S., June, C.H., and Zhao, Y. (2017b).
A versatile system for rapid multiplex genome-edited CAR T cell generation.
Oncotarget 8, 17002–17011.
Ricciotti, E., and FitzGerald, G.A. (2011). Prostaglandins and inflammation. Ar-
terioscler Thromb. Vasc. Biol. 31, 986–1000.
Ridge, J.P., Di Rosa, F., and Matzinger, P. (1998). A conditioned dendritic cell
can be a temporal bridge between a CD4+ T-helper and a T-killer cell. Nature
393, 474–478.
Riese, M.J., Wang, L.C., Moon, E.K., Joshi, R.P., Ranganathan, A., June, C.H.,
Koretzky, G.A., and Albelda, S.M. (2013). Enhanced effector responses in acti-
vated CD8+ T cells deficient in diacylglycerol kinases. Cancer Res. 73,
3566–3577.
Rodgers, D.T., Mazagova, M., Hampton, E.N., Cao, Y., Ramadoss, N.S.,
Hardy, I.R., Schulman, A., Du, J., Wang, F., Singer, O., et al. (2016). Switch-
mediated activation and retargeting of CAR-T cells for B-cell malignancies.
Proc. Natl. Acad. Sci. U S A 113, E459–E468.
Roth, T.L., Li, P.J., Blaeschke, F., Nies, J.F., Apathy, R., Mowery, C., Yu, R.,
Nguyen, M.L.T., Lee, Y., Truong, A., et al. (2020). Pooled knockin targeting
for genome engineering of cellular immunotherapies. Cell 181, 1–17.
Roybal, K.T., Rupp, L.J., Morsut, L., Walker, W.J., McNally, K.A., Park, J.S.,
and Lim, W.A. (2016). Precision tumor recognition by T cells with combinatorial
antigen-sensing circuits. Cell 164, 770–779.
ll
Ruella, M., Barrett, D.M., Kenderian, S.S., Shestova, O., Hofmann, T.J., Peraz-
zelli, J., Klichinsky, M., Aikawa, V., Nazimuddin, F., Kozlowski, M., et al. (2016).
Dual CD19 and CD123 targeting prevents antigen-loss relapses after CD19-
directed immunotherapies. J. Clin. Invest. 126, 3814–3826.
Rupp, L.J., Schumann, K., Roybal, K.T., Gate, R.E., Ye, C.J., Lim, W.A., and
Marson, A. (2017). CRISPR/Cas9-mediated PD-1 disruption enhances anti-tu-
mor efficacy of human chimeric antigen receptor T cells. Sci. Rep. 7, 737.
Sadelain, M., Riviere, I., and Riddell, S. (2017). Therapeutic T cell engineering.
Nature 545, 423–431.
Sakemura, R., Terakura, S., Watanabe, K., Julamanee, J., Takagi, E., Miyao,
K., Koyama, D., Goto, T., Hanajiri, R., Nishida, T., et al. (2016). A Tet-On induc-
ible system for controlling CD19-chimeric antigen receptor expression upon
drug administration. Cancer Immunol. Res. 4, 658–668.
Sasaki, K., Kohanbash, G., Hoji, A., Ueda, R., McDonald, H.A., Reinhart, T.A.,
Martinson, J., Lotze, M.T., Marincola, F.M., Wang, E., et al. (2010). miR-17-92
expression in differentiated T cells - implications for cancer immunotherapy.
J. Transl Med. 8, 17.
Scharping, N.E., Menk, A.V., Moreci, R.S., Whetstone, R.D., Dadey, R.E., Wat-
kins, S.C., Ferris, R.L., and Delgoffe, G.M. (2016). The tumor microenvironment
represses T cell mitochondrial biogenesis to drive intratumoral T cell metabolic
insufficiency and dysfunction. Immunity 45, 374–388.
€cker, J.C. (2002). Induction of T cell develop-
Schmitt, T.M., and Zúñiga-Pflu
ment from hematopoietic progenitor cells by delta-like-1 in vitro. Immunity
17, 749–756.
Schuster, S.J., Svoboda, J., Chong, E.A., Nasta, S.D., Mato, A.R., Anak, O.,
Brogdon, J.L., Pruteanu-Malinici, I., Bhoj, V., Landsburg, D., et al. (2017).
Chimeric antigen receptor T cells in refractory B-cell lymphomas. N. Engl. J.
Med. 377, 2545–2554.
Serafini, M., Manganini, M., Borleri, G., Bonamino, M., Imberti, L., Biondi, A.,
Golay, J., Rambaldi, A., and Introna, M. (2004). Characterization of CD20-
transduced T lymphocytes as an alternative suicide gene therapy approach
for the treatment of graft-versus-host disease. Hum. Gene Ther. 15, 63–76.
Shah, N.N., Highfill, S.L., Shalabi, H., Yates, B., Jin, J., Wolters, P.L., Ombrello,
A., Steinberg, S.M., Martin, S., Delbrook, C., et al. (2020). CD4/CD8 T-cell selec-
tion affects chimeric antigen receptor (CAR) T-cell potency and toxicity: updated
results from a phase I anti-CD22 CAR T-cell trial. J. Clin. Oncol. 38, 1938–1950.
Shalabi, H., Kraft, I.L., Wang, H.W., Yuan, C.M., Yates, B., Delbrook, C., Zimbel-
man, J.D., Giller, R., Stetler-Stevenson, M., Jaffe, E.S., et al. (2018). Sequential
loss of tumor surface antigens following chimeric antigen receptor T-cell thera-
pies in diffuse large B-cell lymphoma. Haematologica 103, e215–e218.
Shin, J.H., Park, H.B., Oh, Y.M., Lim, D.P., Lee, J.E., Seo, H.H., Lee, S.J., Eom,
H.S., Kim, I.H., Lee, S.H., and Choi, K. (2012). Positive conversion of negative
signaling of CTLA4 potentiates antitumor efficacy of adoptive T-cell therapy in
murine tumor models. Blood 119, 5678–5687.
Shum, T., Omer, B., Tashiro, H., Kruse, R.L., Wagner, D.L., Parikh, K., Yi, Z.,
Sauer, T., Liu, D., Parihar, R., et al. (2017). Constitutive signaling from an engi-
neered IL7 receptor promotes durable tumor elimination by tumor-redirected
T cells. Cancer Discov. 7, 1238–1247.
Siska, P.J., van der Windt, G.J., Kishton, R.J., Cohen, S., Eisner, W., MacIver,
N.J., Kater, A.P., Weinberg, J.B., and Rathmell, J.C. (2016). Suppression of
Glut1 and glucose metabolism by decreased Akt/mTORC1 signaling drives
T cell impairment in B cell leukemia. J. Immunol. 197, 2532–2540.
Sommermeyer, D., Hudecek, M., Kosasih, P.L., Gogishvili, T., Maloney, D.G.,
Turtle, C.J., and Riddell, S.R. (2016). Chimeric antigen receptor-modified
T cells derived from defined CD8+ and CD4+ subsets confer superior antitumor
reactivity in vivo. Leukemia 30, 492–500.
Srivastava, S., Salter, A.I., Liggitt, D., Yechan-Gunja, S., Sarvothama, M.,
Cooper, K., Smythe, K.S., Dudakov, J.A., Pierce, R.H., Rader, C., and Riddell,
S.R. (2019). Logic-gated ROR1 chimeric antigen receptor expression rescues
T cell-mediated toxicity to normal tissues and enables selective tumor target-
ing. Cancer Cell 35, 489–503.
Stadtmauer, E.A., Fraietta, J.A., Davis, M.M., Cohen, A.D., Weber, K.L., Lan-
caster, E., Mangan, P.A., Kulikovskaya, I., Gupta, M., Chen, F., et al. (2020).
CRISPR-engineered T cells in patients with refractory cancer. Science 367,
eaba7365.
Cancer Cell 38, October 12, 2020 487
 ll
Stancovski, I., Schindler, D.G., Waks, T., Yarden, Y., Sela, M., and Eshhar, Z.
(1993). Targeting of T lymphocytes to Neu/HER2-expressing cells using
chimeric single chain Fv receptors. J. Immunol. 151, 6577–6582.
Stephan, M.T., Ponomarev, V., Brentjens, R.J., Chang, A.H., Dobrenkov, K.V.,
Heller, G., and Sadelain, M. (2007). T cell-encoded CD80 and 4-1BBL induce
auto- and transcostimulation, resulting in potent tumor rejection. Nat. Med. 13,
1440–1449.
Sterner, R.M., Sakemura, R., Cox, M.J., Yang, N., Khadka, R.H., Forsman,
C.L., Hansen, M.J., Jin, F., Ayasoufi, K., Hefazi, M., et al. (2019). GM-CSF in-
hibition reduces cytokine release syndrome and neuroinflammation but en-
hances CAR-T cell function in xenografts. Blood 133, 697–709.
Straathof, K.C., Pule, M.A., Yotnda, P., Dotti, G., Vanin, E.F., Brenner, M.K.,
Heslop, H.E., Spencer, D.M., and Rooney, C.M. (2005). An inducible caspase
9 safety switch for T-cell therapy. Blood 105, 4247–4254.
Suarez, E.R., Chang, D.K., Sun, J., Sui, J., Freeman, G.J., Signoretti, S., Zhu,
Q., and Marasco, W.A. (2016). Chimeric antigen receptor T cells secreting anti-
PD-L1 antibodies more effectively regress renal cell carcinoma in a humanized
mouse model. Oncotarget 7, 34341–34355.
Tan, A.H.J., Vinanica, N., and Campana, D. (2020). Chimeric antigen receptor-
T cells with cytokine neutralizing capacity. Blood Adv. 4, 1419–1431.
Tchou, J., Zhao, Y., Levine, B.L., Zhang, P.J., Davis, M.M., Melenhorst, J.J.,
Kulikovskaya, I., Brennan, A.L., Liu, X., Lacey, S.F., et al. (2017). Safety and ef-
ficacy of intratumoral injections of chimeric antigen receptor (CAR) T cells in
metastatic breast cancer. Cancer Immunol. Res. 5, 1152–1161.
Torikai, H., Reik, A., Liu, P.Q., Zhou, Y., Zhang, L., Maiti, S., Huls, H., Miller,
J.C., Kebriaei, P., Rabinovich, B., et al. (2012). A foundation for universal T-
cell based immunotherapy: T cells engineered to express a CD19-specific
chimeric-antigen-receptor and eliminate expression of endogenous TCR.
Blood 119, 5697–5705.
Torikai, H., Reik, A., Soldner, F., Warren, E.H., Yuen, C., Zhou, Y., Crossland,
D.L., Huls, H., Littman, N., Zhang, Z., et al. (2013). Toward eliminating HLA
class I expression to generate universal cells from allogeneic donors. Blood
122, 1341–1349.
Tormoen, G.W., Crittenden, M.R., and Gough, M.J. (2018). Role of the immu-
nosuppressive microenvironment in immunotherapy. Adv. Radiat. Oncol. 3,
520–526.
Tran, E., Chinnasamy, D., Yu, Z., Morgan, R.A., Lee, C.C., Restifo, N.P., and
Rosenberg, S.A. (2013). Immune targeting of fibroblast activation protein trig-
gers recognition of multipotent bone marrow stromal cells and cachexia.
J. Exp. Med. 210, 1125–1135.
Urbanska, K., Lanitis, E., Poussin, M., Lynn, R.C., Gavin, B.P., Kelderman, S.,
Yu, J., Scholler, N., and Powell, D.J., Jr. (2012). A universal strategy for adop-
tive immunotherapy of cancer through use of a novel T-cell antigen receptor.
Cancer Res. 72, 1844–1852.
Valton, J., Guyot, V., Marechal, A., Filhol, J.M., Juillerat, A., Duclert, A., Duch-
ateau, P., and Poirot, L. (2015). A multidrug-resistant engineered CAR T cell for
allogeneic combination immunotherapy. Mol. Ther. 23, 1507–1518.
Vodnala, S.K., Eil, R., Kishton, R.J., Sukumar, M., Yamamoto, T.N., Ha, N.H.,
Lee, P.H., Shin, M., Patel, S.J., Yu, Z., et al. (2019). T cell stemness and
dysfunction in tumors are triggered by a common mechanism. Science 363,
eaau0135.
Wang, D., and Dubois, R.N. (2006). Prostaglandins and cancer. Gut 55,
115–122.
Wang, L.C., Lo, A., Scholler, J., Sun, J., Majumdar, R.S., Kapoor, V., Antzis, M.,
Cotner, C.E., Johnson, L.A., Durham, A.C., et al. (2014). Targeting fibroblast
activation protein in tumor stroma with chimeric antigen receptor T cells can
inhibit tumor growth and augment host immunity without severe toxicity. Can-
cer Immunol. Res. 2, 154–166.
Wang, X., Chang, W.C., Wong, C.W., Colcher, D., Sherman, M., Ostberg, J.R.,
Forman, S.J., Riddell, S.R., and Jensen, M.C. (2011). A transgene-encoded
cell surface polypeptide for selection, in vivo tracking, and ablation of engi-
neered cells. Blood 118, 1255–1263.
488 Cancer Cell 38, October 12, 2020
Review
Weber, E.W., Lynn, R.C., Parker, K.R., Anbunathan, H., Lattin, J., Sotillo, E.,
Good, Z., Malipatlolla, M., Xu, P., Vandris, P., et al. (2020). Transient ‘‘rest’’ in-
duces functional reinvigoration and epigenetic remodeling in exhausted CAR-
T cells. bioRxiv. https://doi.org/10.1101/2020.01.26.920496.
Weber, E.W., Lynn, R.C., Sotillo, E., Lattin, J., Xu, P., and Mackall, C.L. (2019).
Pharmacologic control of CAR-T cell function using dasatinib. Blood Adv. 3,
711–717.
Wei, J., Long, L., Zheng, W., Dhungana, Y., Lim, S.A., Guy, C., Wang, Y.,
Wang, Y.D., Qian, C., Xu, B., et al. (2019). Targeting REGNASE-1 programs
long-lived effector T cells for cancer therapy. Nature 576, 471–476.
Wei, S.C., Duffy, C.R., and Allison, J.P. (2018). Fundamental mechanisms of
immune checkpoint blockade therapy. Cancer Discov. 8, 1069–1086.
Weinberg, F., Ramnath, N., and Nagrath, D. (2019). Reactive oxygen species in
the tumor microenvironment: an overview. Cancers (Basel) 11, 1191.
Whiteside, T.L. (2008). The tumor microenvironment and its role in promoting
tumor growth. Oncogene 27, 5904–5912.
Wiesinger, M., Marz, J., Kummer, M., Schuler, G., Dorrie, J., Schuler-Thurner,
B., and Schaft, N. (2019). Clinical-scale production of CAR-T cells for the treat-
ment of melanoma patients by mRNA transfection of a CSPG4-specific CAR
under full GMP compliance. Cancers (Basel) 11, 1198.
Wilkie, S., Burbridge, S.E., Chiapero-Stanke, L., Pereira, A.C., Cleary, S., van
der Stegen, S.J., Spicer, J.F., Davies, D.M., and Maher, J. (2010). Selective
expansion of chimeric antigen receptor-targeted T-cells with potent effector
function using interleukin-4. J. Biol. Chem. 285, 25538–25544.
Wilkie, S., van Schalkwyk, M.C., Hobbs, S., Davies, D.M., van der Stegen, S.J.,
Pereira, A.C., Burbridge, S.E., Box, C., Eccles, S.A., and Maher, J. (2012). Dual
targeting of ErbB2 and MUC1 in breast cancer using chimeric antigen recep-
tors engineered to provide complementary signaling. J. Clin. Immunol. 32,
1059–1070.
Woo, S.R., Turnis, M.E., Goldberg, M.V., Bankoti, J., Selby, M., Nirschl, C.J.,
Bettini, M.L., Gravano, D.M., Vogel, P., Liu, C.L., et al. (2012). Immune inhibi-
tory molecules LAG-3 and PD-1 synergistically regulate T-cell function to pro-
mote tumoral immune escape. Cancer Res. 72, 917–927.
Xu, X., Huang, W., Heczey, A., Liu, D., Guo, L., Wood, M., Jin, J., Courtney,
A.N., Liu, B., Di Pierro, E.J., et al. (2019). NKT cells coexpressing a GD2-spe-
cific chimeric antigen receptor and IL15 show enhanced in vivo persistence
and antitumor activity against neuroblastoma. Clin. Cancer Res. 25,
7126–7138.
Yang, W., Bai, Y., Xiong, Y., Zhang, J., Chen, S., Zheng, X., Meng, X., Li, L.,
Wang, J., Xu, C., et al. (2016). Potentiating the antitumour response of CD8+
T cells by modulating cholesterol metabolism. Nature 531, 651–655.
Yeku, O.O., Purdon, T.J., Koneru, M., Spriggs, D., and Brentjens, R.J. (2017).
Armored CAR T cells enhance antitumor efficacy and overcome the tumor
microenvironment. Sci. Rep. 7, 10541.
Zah, E., Lin, M.Y., Silva-Benedict, A., Jensen, M.C., and Chen, Y.Y. (2016). T
cells expressing CD19/CD20 bispecific chimeric antigen receptors prevent an-
tigen escape by malignant B cells. Cancer Immunol. Res. 4, 498–508.
Zah, E., Nam, E., Bhuvan, V., Tran, U., Ji, B.Y., Gosliner, S.B., Wang, X.,
Brown, C.E., and Chen, Y.Y. (2020). Systematically optimized BCMA/CS1 bis-
pecific CAR-T cells robustly control heterogeneous multiple myeloma. Nat.
Commun. 11, 2283.
Zhang, L., Kerkar, S.P., Yu, Z., Zheng, Z., Yang, S., Restifo, N.P., Rosenberg,
S.A., and Morgan, R.A. (2011). Improving adoptive T cell therapy by targeting
and controlling IL-12 expression to the tumor environment. Mol. Ther. 19,
751–759.
Zhang, Y., Zhang, X., Cheng, C., Mu, W., Liu, X., Li, N., Wei, X., Liu, X., Xia, C.,
and Wang, H. (2017). CRISPR-Cas9 mediated LAG-3 disruption in CAR-T
cells. Front. Med. 11, 554–562.
Zhao, Z., Condomines, M., van der Stegen, S.J.C., Perna, F., Kloss, C.C., Gun-
set, G., Plotkin, J., and Sadelain, M. (2015). Structural design of engineered
costimulation determines tumor rejection kinetics and persistence of CAR
T cells. Cancer Cell 28, 415–428.
Cancer Cell

Article

Integrated Molecular and Clinical Analysis

of 1,000 Pediatric Low-Grade Gliomas
Scott Ryall,1,2,27 Michal Zapotocky,1,3,4,27 Kohei Fukuoka,1,3 Liana Nobre,1,3 Ana Guerreiro Stucklin,1,3,5 Julie Bennett,1,3
Robert Siddaway,1 Christopher Li,1,2 Sanja Pajovic,1 Anthony Arnoldo,6 Paul E. Kowalski,6 Monique Johnson,6
Javal Sheth,1,6 Alvaro Lassaletta,1,3,7 Ruth G. Tatevossian,8 Wilda Orisme,8 Ibrahim Qaddoumi,9 Lea F. Surrey,10,11
Marilyn M. Li,10 Angela J. Waanders,11,12,13,14 Stephen Gilheeney,15 Marc Rosenblum,16 Tejus Bale,16 Derek S. Tsang,17,18
Normand Laperriere,17,18 Abhaya Kulkarni,19,20 George M. Ibrahim,19,20 James Drake,19,20

(Author list continued on next page)

1Arthur and Sonia Labatt Brain Tumour Research Centre, The Hospital for Sick Children, 555 University Avenue, Toronto, ON M5G 1X8,
Canada
2Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, ON, Canada
3Department of Haematology/Oncology, The Hospital for Sick Children, Toronto, ON, Canada
4Second Faculty of Medicine, Charles University and University Hospital Motol, Prague, Czech Republic
5Children’s Research Center, University Children’s Hospital Zurich, Zurich, Switzerland
6Department of Paediatric Laboratory Medicine, The Hospital for Sick Children, Toronto, ON, Canada
7Department of Pediatric Hematology and Oncology, Hospital Universitario Niño Jesús, Madrid, Spain
8Department of Pathology, St. Jude Children’s Research Hospital, Memphis, TN, USA
9Department of Oncology, St. Jude Children’s Research Hospital, Memphis, TN, USA
10Department of Pathology and Laboratory Medicine, Children’s Hospital of Philadelphia, Philadelphia, PA, USA
11Department of Genomic Diagnostics, Children’s Hospital of Philadelphia, Philadelphia, PA, USA
12Center for Data Driven Discovery in Biomedicine, Children’s Hospital of Philadelphia, Philadelphia, PA, USA
13Department of Hematology, Oncology, and Stem Cell Transplant, Ann & Robert H Lurie Children’s Hospital of Chicago, Chicago, IL, USA
14Department of Pediatrics, Feinberg School of Medicine Northwestern University, Chicago, IL, USA
15Department of Pediatrics, Memorial Sloan Kettering Cancer Center, New York, NY, USA

(Affiliations continued on next page)

SUMMARY

Pediatric low-grade gliomas (pLGG) are frequently driven by genetic alterations in the RAS-mitogen-acti-
vated protein kinase (RAS/MAPK) pathway yet show unexplained variability in their clinical outcome. To
address this, we characterized a cohort of >1,000 clinically annotated pLGG. Eighty-four percent of cases
harbored a driver alteration, while those without an identified alteration also often exhibited upregulation
of the RAS/MAPK pathway. pLGG could be broadly classified based on their alteration type. Rearrange-
ment-driven tumors were diagnosed at a younger age, enriched for WHO grade I histology, infrequently pro-
gressed, and rarely resulted in death as compared with SNV-driven tumors. Further sub-classification of clin-
ical-molecular correlates stratified pLGG into risk categories. These data highlight the biological and clinical
differences between pLGG subtypes and opens avenues for future treatment refinement.

INTRODUCTION (Ostrom et al., 2015). pLGG encompass a broad range of glial,

neuronal, and mixed glioneuronal entities in the World Health Or-
Pediatric low-grade gliomas (pLGG) are the most frequent brain ganization (WHO) classification of central nervous system (CNS)
tumors in children, accounting for approximately 30% of all cases tumors (Louis et al., 2016). Unlike lower-grade gliomas in adults,

Significance

Pediatric low-grade gliomas (pLGG) are the most common brain tumors affecting children. This integrated clinicopathologic
and genomic analysis of >1,000 pLGG defines molecular subgroups with distinct biological drivers and clinical features.
RAS/MAPK pathway is activated near universally in pLGG, regardless of the presence of a clear genomic activator. Further,
although many alterations converge on the RAS/MAPK pathway, clinical presentation and outcome are highly variable de-
pending on the type of underlying alteration. This information helped define clinical risk groups of pLGG with different pro-
gression-free and overall survival which likely require different treatment strategies. As modernized treatment regimens uti-
lize alteration-specific agents, we provide the framework for molecular classification of pLGG reflecting unique biological
mechanisms driving the disease that likely promote different therapeutic susceptibilities.

Cancer Cell 37, 569–583, April 13, 2020 ª 2020 Elsevier Inc. 569
Peter Dirks,1,2,20 Michael D. Taylor,1,2,20 James T. Rutka,1,2,20 Suzanne Laughlin,21,22 Manohar Shroff,21,22 Mary Shago,2,6
Lili-Naz Hazrati,2,23 Colleen D’Arcy,2,23,24 Vijay Ramaswamy,1,3,25 Ute Bartels,3,25 Annie Huang,1,2,3 Eric Bouffet,3,25
Matthias A. Karajannis,15 Mariarita Santi,10,11 David W. Ellison,8 Uri Tabori,1,3,26,28 and Cynthia Hawkins1,2,23,29,*
16Department of Pathology, Memorial Sloan Kettering Cancer Center, New York, NY, USA
17Radiation Medicine Program, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada
18Department of Radiation Oncology, Faculty of Medicine, University of Toronto, Toronto, ON, Canada
19Department of Surgery, University of Toronto, Toronto, ON, Canada
20Department of Neurosurgery, The Hospital for Sick Children, Toronto ON, Canada
21Department of Radiology, The Hospital for Sick Children, Toronto ON, Canada
22Department of Medical Imaging, University of Toronto, Toronto, ON, Canada
23Department of Pathology, The Hospital for Sick Children, Toronto, ON, Canada
24Department of Anatomical Pathology, The Alfred Hospital, Prahran, VIC, Australia
25Department of Paediatrics, University of Toronto, Toronto, ON, Canada
26Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada
27These authors contributed equally
28Co-senior authors
29Lead Contact

*Correspondence: cynthia.hawkins@sickkids.ca
https://doi.org/10.1016/j.ccell.2020.03.011

which primarily arise in the cerebral hemispheres and inevitably
transform to higher-grade glioma, pLGG can occur throughout
the CNS and rarely transform (Broniscer et al., 2007; Mistry
et al., 2015). Nevertheless, outcome and response to therapy is
highly variable. If complete surgical resection is successful,
10-year progression-free survival (PFS) exceeds 85%, but drops
below 50% if there is radiologically visible residual tumor (Wisoff
et al., 2011). In deep-seated midline locations, gross total resec-
tion (GTR) is not often achievable and adjuvant therapy may be
required, often with unsatisfactory tumor control and/or long-
term morbidity (Krishnatry et al., 2016; Nageswara and Packer,
2014). Which patients require these therapies and who will benefit
from them is not yet well understood.
Over the last decade, molecular profiling studies have incre-
mentally identified key genetic events in pLGG that converge on
the RAS-mitogen-activated protein kinase (RAS/MAPK) pathway.
Most commonly, these are somatic events involving BRAF or
germline NF1 alterations (Collins et al., 2015; Jones et al., 2008;
Schindler et al., 2011; Dougherty et al., 2010; Lassaletta et al.,
2017; Listernick et al., 1999; Uusitalo et al., 2016; Seminog and
Goldacre, 2013). In addition to these common pLGG alterations,
rarer alterations affecting RAS/MAPK signaling, including those
involving FGFR1/2/3, NTRK2, RAF1, ALK, and ROS1 (Zhang
et al., 2013; Jones et al., 2013; Qaddoumi et al., 2016; Guerreiro
Stucklin et al., 2019), as well as non-RAS/MAPK alterations,
such as MYB, MYBL1, IDH1, and H3F3A (Qaddoumi et al.,
2016; Tatevossian et al., 2010; Ramkissoon et al., 2013; Bando-
padhayay et al., 2016; Ryall et al., 2016; Hartmann et al., 2009)
have been identified in small numbers of cases. However, several
key issues remain undefined: (1) Are all NF1, BRAF fused and
BRAF mutant tumors the same? (2) What is the mechanism of
tumorigenesis in pLGG without an identifiable genetic alteration?
(3) What are the clinical features of the rare alterations in pLGG
and is their outcome unique? (4) Can molecular alterations help
provide biological insights for disease stratification?
To answer these questions and provide a population-based
resource for the pediatric neuro-oncology community, we
molecularly characterized >1,000 pLGG with comprehensive
clinical data. This enabled us to provide a statistically robust,
annotated resource which includes representation of the rarest
pLGG molecular entities and their clinical features.
RESULTS
Patient Cohort
Our population-based cohort of pLGG consisted of 976 patients
(<19 years) followed and treated at the Hospital for Sick Children
(Toronto, ON, Canada) from 1986 to 2017 (Table S1). For each
patient we collected demographic, treatment and outcome data.
At the population level, tumors were distributed equally between
the diencephalon (n = 313, 32%; n = 124, 13% were from neurofi-
bromatosis type 1 [NF1] patients), cerebral hemispheres (n = 265,
27%), and the cerebellum (n = 252, 26%), whereas pure brainstem
(n = 92, 9.4%), spinal cord (n = 41, 4.2%), and extensively dissem-
inated tumors (n = 13, 1.3%) were less frequent (Figure 1A). Upon
pathologic review of non-NF1 cases (n = 843), pilocytic astrocy-
toma (n = 303, 36%) was the most common diagnosis (excluding
LGG, not otherwise specified [NOS]) across tumor locations (Fig-
ure 1B) with the exception of the cerebral hemispheres, which
was histologically diverse (Figure 1C). The median age of diagnosis
was 7.6 years (range 0–18.7 years), with pLGG in the cerebral
hemispheres being diagnosed at a later age (median = 10.7 years)
as compared with other tumor locations (p < 0.0001, ANOVA; all
pairwise comparisons adjusted p < 0.0001, t test) (Figure 1D).
There was a significant association between tumor location,
PFS, and overall survival (OS) (p < 0.0001, log rank test) (Figures
1E and S1), with 10-year PFS and OS best for patients with tumors
in the cerebellum (89% and 99%, respectively), and worst for those
with extensively disseminated disease (0% and 67%, respec-
tively). Importantly, only 7.5% of patients succumbed to their dis-
ease (median time to death = 3.9 years, median OS follow-up =
15.9 years) despite 33% experiencing tumor progression (median
time to progression = 2.3 years, median PFS follow-up = 5.9 years)
(Figure 1F).
Characteristics of NF1-Driven pLGG
Patients with the genetic pre-disposition disorder NF1 are diag-
nosed using a series of clinical observations and tests indicative

570 Cancer Cell 37, 569–583, April 13, 2020

Figure 1. pLGG Cohort Details
(A) Anatomical location of all pLGG within the cohort (n = 976).
(B) The histological spectrum of all non-NF1 pLGG (n = 843). PA, pilocytic astrocytoma; LGG, NOS, low-grade glioma, not otherwise specified; GG, ganglio-
glioma; DNET, dysembryoplastic neuroepithelial tumor; PXA, pleomorphic xanthoastrocytoma; GNT, glioneuronal tumor; DA, diffuse astrocytoma; AG, angio-
centric glioma; ODG, oligodendroglioma; DIA/DIG, desmoplastic infantile astrocytoma/ganglioglioma.
(C) Histological distribution of samples based on tumor location of all non-NF1 pLGG (n = 843).
(D) Boxplot showing age at diagnosis separated by tumor location of the entire pLGG cohort (n = 976). The thick line within the box represents the median, the
lower and upper limits of the box represent the first and third quartiles and the whiskers the minimum and maximum values. Adjusted p value for all pairwise
comparisons, t test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001; n.s., not significant.
(E) Progression-free survival of the pLGG cohort segregated by tumor location. p Value calculated via the log rank test.
(F) Progression-free and overall survival of the entire pLGG cohort (n = 976). p value calculated via the log rank test.
See also Figure S1 and Table S1.

Cancer Cell 37, 569–583, April 13, 2020 571

of a germline NF1 mutation (Gutmann et al., 2017). Therefore,
pLGG arising in these patients are primarily diagnosed via imag-
ing and clinical observation, rather than their genetics. Therefore,
we examined this group of tumors separately, before our exten-
sive molecular profiling of somatic pLGG (Table S2).
Although NF1-driven pLGG are generally thought to have a
favorable clinical course, analysis of a large cohort revealed
important risk groups. In our study, NF1-driven pLGG accounted
for 14% of cases (n = 133). Although most NF1 pLGG occur as
optic pathway glioma (OPG), 19% (n = 25) arose outside of this
location. Patients with NF1 tumors arising outside the optic
pathway had significantly worse OS and PFS compared with
those arising in the optic pathway (p = 0.0011, p = 0.0029,
respectively, log rank test) (Figures S2A and S2B). Furthermore,
of the high risk, recurrent, and biopsied NF1 pLGG, 20% were
found to harbor mutations in other molecular drivers, including
BRAF p.V600E, FGFR1, and/or H3F3A (H3.3) p.K27M. All of
these patients experienced multiple progressions with two suc-
cumbing to their disease post-chemoradiation after 15.5 and
13.7 years, respectively. These observations, although prelimi-
nary, suggest that non-OPG NF1 tumors may require specialized
management, including an early biopsy and molecular profiling
in agreement with recent reports (D’Angelo et al., 2019).
The Molecular Landscape of Non-NF1 pLGG
To determine the true frequency of molecular alterations in pLGG
we analyzed 540 tumors from 2000 to 2017 where material qual-
ity was sufficient to utilize our tiered profiling approach (Fig-
ure S3A). In total, 88% (n = 477/540) had sufficient material for
molecular profiling.
Together, KIAA1549-BRAF (n = 166), BRAF p.V600E (n = 79),
and germline NF1 mutations (n = 79) accounted for 68% (n = 324)
of tumors (Figures 2A–2C). Rare alterations accounted for an
additional 17% of cases. These included non-canonical BRAF
alterations, such as fusions partnered with genes others than
KIAA1549 (n = 4), 2 insertion events at position 600 (p.V600ins)
and 1 SNV at position 594 (p.D594N) (Figures 2A–2C and 2E).
A further 1.3% (n = 6) of cases contained alterations in other
direct members of the RAS/MAPK pathway, including three
RAF1 fusions, two KRAS mutations, and one patient with a short
deletion in MAP2K1 (Figures 2A–2C and 2E). The next most
frequent alterations were those affecting receptor tyrosine ki-
nases (RTKs) (n = 45, 9.4%) (Figures 2A–2C) and included two
categories: FGFR and other RTKs. Alterations in FGFR most
frequently involved FGFR1/2 (n = 29, 6.1%) and included
FGFR1-TACC1 fusions (n = 7, 1.5%), FGFR1 tyrosine kinase
domain (TKD) duplications (n = 10, 2.1%), FGFR2 fusions (n =
5, 1.0%), and hotspot mutations in FGFR1 (n = 7, 1.5%) (Figures
2A–2C). Alterations in other RTKs (n = 16, 3.4%) included muta-
tions in MET (n = 5, 1.0%) or PDGFRA (n = 1, 0.2%), as well as
fusions involving ALK (n = 2, 0.4%), ROS1 (n = 2, 0.4%), and
NTRK2 (n = 2, 0.4%) (Figures 2A–2C and 2E). Finally, 4.6%
(n = 22) of cases contained alterations in genes with seemingly
no direct impact on the RAS/MAPK pathway (Figures 2A–2C).
These included mutations in H3F3A (n = 4, 0.8%), IDH1 (n = 4,
0.8%), and rearrangements involving MYB (n = 6, 1.3%) or
MYBL1 (n = 5, 1.0%) (Figures 2A–2C). Altogether, we identified
a driver mutation in 84% of pLGG. Incidences of molecular alter-
ations excluding NF1 patients (n = 397) are seen in Figures 2B
572 Cancer Cell 37, 569–583, April 13, 2020
and 2D and enrichment across tumor locations and pathologies
in Figures S3B and S3C.
The RAS/MAPK Pathway Is Upregulated across pLGG
Regardless of Underlying Mutation
The predilection of NF1 patients for developing pLGG together
with the identification of KIAA1549-BRAF and BRAF p.V600E
as molecular drivers in pLGG led to the hypothesis that upregu-
lation of the RAS/MAPK pathway may be the primary driver for
tumor formation (Collins et al., 2015; Northcott et al., 2015; Jones
et al., 2012; Zhang et al., 2013). However, how broadly this ap-
plies to pLGG has not been tested and is of major importance
due to the increasing use of the RAS/MAPK pathway-targeting
agents in the clinic.
We therefore asked whether the 16% of pLGG without identi-
fied mutations nonetheless resulted in upregulation of the RAS/
MAPK pathway. To test this hypothesis, we first analyzed a se-
ries of pLGG with non-BRAF alterations, including FGFR alter-
ations, rare RTKs, RAF1 fusions, KRAS mutations, and MYB or
MYBL1 rearrangements and compared their phosphorylated
ERK (ppERK) levels with KIAA1549-BRAF and BRAF p.V600E tu-
mors. Interestingly, all tumors had significantly increased ppERK
as compared with normal brain controls (p < 0.0001, ANOVA; all
pairwise comparisons adjusted p < 0.0001, t test) (Figure 3A).
Importantly, increased ppERK was also seen in MYB- and
MYBL1-altered tumors, which would not themselves be ex-
pected to directly signal via the RAS/MAPK pathway. To further
explore whether RAS/MAPK pathway upregulation is a unifying
event in pLGG even in the absence of an activating genetic
event, we examined a subset of pLGG in which a molecular
driver was not identified. Utilizing RNA sequencing, we per-
formed single sample gene set enrichment analysis (ssGSEA)
focusing on the RAS/MAPK pathway. We observed that genes
known to be up- or downregulated by RAS/MAPK activation
were significantly enriched in these tumors as compared with
normal brain controls (Figures 3B and 3C). Furthermore, when
compared against samples with known RAS/MAPK pathway al-
terations, the activation signature was indiscernible between the
two (p = 0.4103, ANOVA; all pairwise comparisons adjusted p <
0.0001, t test) (Figure 3D), indicating similar levels of RAS/MAPK
upregulation in pLGG lacking a clear molecular driver.
Alteration Type Predicts pLGG Outcome
To further interrogate the impact of rare pLGG alterations, we
collected data from an additional 61 patients in whom a rare
molecular alteration had been previously identified from St.
Jude Children’s Hospital (Memphis, TN, USA), Children’s Hos-
pital of Philadelphia (Philadelphia, PA, USA), and Memorial
Sloan Kettering Cancer Center (New York, NY, USA) (Table
S3). This yielded a total of 1,037 pLGG. With these additional
cases and detailed clinical and molecular information, we
asked which features were predictive of clinical outcome in
pLGG. Interestingly, patient outcome was significantly associ-
ated with the type of alteration (rearrangement versus SNV),
and not exclusively on which particular gene was altered (Fig-
ures 4A and 4B; Table 1). Patients with rearrangement-driven
pLGG had good long-term outcome with very few deaths
(n = 7, 2.6%) and fewer progressions (n = 67, 27%) (Figures
4B–4D; Table 1). In contrast, patients with SNV-driven
Figure 2. The Molecular Landscape of pLGG
(A) Oncoprint representation of the molecular alterations and their associated categories in 610 pLGG. Samples are arranged in columns with genes and gene
categories labeled along the row. *Denotes that these BRAF SNVs and fusions are not the canonical KIAA1549-BRAF or p.V600E.
(B) Bar graph of all recurrent somatic mutations across all 477 cases diagnosed from 2000 to 2017 for which sufficient material for molecular testing was available,
in order of frequency and colored based on the inclusion (blue) or exclusion (red) of NF1 patients.
(C) Pie chart depicting the frequency of alterations per molecular category in a population-based cohort of pLGG diagnosed from 2000 to 2017 (n = 477).
(D) Pie chart depicting the frequency of alterations per molecular category in non-NF1 pLGG diagnosed from 2000 to 2017 (n = 397).
(E) Schematic representation of the rare and novel fusions identified in this study. Figures were derived using the Protein Paint feature of the St. Jude PeCan
website (https://pecan.stjude.cloud/proteinpaint).
See also Figures S2 and S3 and Table S2.

Cancer Cell 37, 569–583, April 13, 2020 573

Figure 3. RAS/MAPK Pathway Upregulation in Non-canonical and Molecularly Undetermined pLGG
(A) Boxplot showing the ppERK/ERK protein levels, separated by molecular alteration. The thick line within the box represents the median, the lower and upper
limits of the box represent the first and third quartiles and the whiskers the minimum and maximum values. Adjusted p value for all pairwise comparisons,
t test.*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001; n.s., not significant.
(B) Pre-ranked gene set enrichment analysis (GSEA) of the RAS/MAPK pathway activation signature in molecularly undetermined pLGG. NES, normalized
enrichment score; FDR, false-discovery rate.
(C) Single sample gene set enrichment analysis (ssGSEA) of RAS/MAPK activation for normal brain controls and molecularly undetermined pLGG. The thick line
within the box represents the median, the lower and upper limits of the box represent the first and third quartiles and the whiskers the minimum and maximum
values. Adjusted p value for all pairwise comparisons, Mann-Whitney test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001; n.s., not significant.
(D) RAS/MAPK ssGSEA scores for known RAS/MAPK mutant and molecularly undetermined pLGG compared with normal brain. The thick line within the box
represents the median, the lower and upper limits of the box represent the first and third quartiles and the whiskers the minimum and maximum values. Adjusted
p value for all pairwise comparisons, t test.*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001; n.s., not significant.

pLGG were significantly more likely to succumb to their dis-
ease (n = 24, 13%, p < 0.0001, Fisher’s exact test versus rear-
rangement-driven) and/or progress (n = 80, 44%, p < 0.0001,
Fisher’s exact test versus fusion-driven) (Table 1; Figures 4B–
4D). Furthermore, rearrangement-driven pLGG were diagnosed
at a significantly younger age (median 6.6 versus 10.9 years,
p < 0.0001, t test) and were enriched for WHO grade I histology
(p < 0.0001, Fisher’s exact test) (Figure 4B; Table 1). This
pattern is evident in BRAF, where tumors with KIAA1549-
BRAF have superior outcome to BRAF p.V600E (5-year PFS
of 69% for KIAA1549-BRAF versus 52% for BRAF p.V600E,
p = 0.0058, log rank test). When investigating BRAF in grade
I tumors alone, the pattern remained (5-year PFS of 72% and
56% for KIAA1549-BRAF and BRAF p.V600E, respectively,
p = 0.0176). Although the numbers are too small to allow statis-
tical comparisons between SNVs and fusions for other genes,
the same trend is evident. For example, for FGFR1, patients
with FGFR1-TACC1 or FGFR1 TKD tumors had similar
outcome to those with KIAA1549-BRAF fusions (5-year PFS
of 69% for all). Patients with FGFR1 SNVs, on the other
hand, were more similar to BRAF p.V600E (5-year PFS of
53% and 52% for FGFR1 SNV and BRAF p.V600E,

574 Cancer Cell 37, 569–583, April 13, 2020

Figure 4. Rearrangement versus SNV-Driven pLGG
(A) Pie charts depicting the molecular alteration breakdown of rearrangement- (top) (n = 265) and SNV -driven (bottom) (n = 182) pLGG.
(B) Rearrangement- versus SNV-driven pLGG as compared across several clinical features. *Adjusted p < 0.05, Fisher’s exact test. GTR, gross total resection.
(C) Kaplan-Meier plot of overall survival of cases separated by rearrangement- or SNV-driven status. p value calculated via the log rank test.
(D) Kaplan-Meier plot of progression-free survival of cases separated by rearrangement- or SNV-driven status. p value calculated via the log rank test.
See also Table S3.

respectively). Importantly, where FGFR1-TACC1 and FGFR1-
TKD tumors did not contain additional alterations, FGFR1
SNVs often did, sometimes resulting in late deaths.
Characteristics of Fusion-Driven pLGG
BRAF Fusions
KIAA1549-BRAF was the most frequent alteration in pLGG
(35%) and was almost exclusively a single-event driver
(n = 175/180, 97%), with four cases also having a CDKN2A dele-
tion and one present in a patient with NF1 (Figure 2A). These five
rare cases are still alive (median follow-up = 7.5 years).
KIAA1549-BRAF was significantly enriched in pilocytic astrocy-
toma (n = 150/180, 83%, p < 0.0001, Fisher’s exact test), and in
cerebellar tumors (n = 100/180, 56%, p = 0.0002, Fisher’s exact
test) (Figure 5A).
Because of the large number of tumors, we could sub-stratify
the BRAF fusions into subgroups based on their breakpoints
(Figure S4A). The most common KIAA1549-BRAF fusion
involved exon 16 in KIAA1549 and exon 9 in BRAF (16:09).
Like all KIAA1549-BRAF fusions, 16:09 was significantly en-
riched in pilocytic astrocytoma (n = 73/83, 88%, p < 0.0001, Fish-
er’s exact test) and in cerebellar tumors (n = 60/83, 72%, p <
0.0001, Fisher’s exact test) (Figures S4B and S4C). Interestingly,
the only KIAA1549-BRAF fusion seen in hemispheric tumors was
15:09 (Figure S4B). 15:09 was also the primary fusion seen in tu-
mors with extensive dissemination (n = 5, 83%, p < 0.0001, Fish-
er’s exact test). 15:09 fusions were associated with a worse PFS
(p = 0.0003, log rank test, Figure S4D), with a 5-year PFS of 59%
as compared with 77%–100% for other fusion subtypes. BRAF
fusions not involving KIAA1549 occurred exclusively in adoles-
cents with no progression events (median follow-up = 3.7 years),
while 15:11 was only observed in two infants who rapidly expe-
rienced tumor progression and died (Figures S4D and S4E).
Identifying the specific fusion breakpoints of KIAA1549-BRAF
will be important to properly ascertain their propensity for certain
clinical features and impact on outcome.
FGFR1/2 Fusions and FGFR1 TKD
FGFR fusions/TKD were observed in 6.1% of the cohort (Fig-
ure 2B). FGFR1-TACC1 pLGG were often cystic lesions, most
commonly pilocytic astrocytoma (n = 7/14, 50%) and occurred

Cancer Cell 37, 569–583, April 13, 2020 575

Table 1. Clinicopathologic Features of Rearrangement versus 2017; Guerreiro Stucklin et al., 2019). These included
SNV-Driven pLGG CCDC88A-ALK, PPP1CB-ALK, GOPC-ROS1, and NTRK2-
Characteristic pLGG Subtype MID1 in addition to novel NTRK2-SF3B1 and PDGFB-LRP1 fu-
sions (Figure 2E). These fusions were restricted to the cerebral
Rearrangement SNV p Value
hemispheres with the exception of the ROS1 fusions, which
Number 265 182
were both seen in the intraventricular space (Table S4). Interest-
Grade ingly, ALK fusions were exclusively observed in infants (0.9 and
I 216 (88)* 97 (66)* <0.0001 1.1 years). No patients harboring these alterations succumbed
II 30 (12)* 50 (34)* to their disease after a median follow-up of 4.9 years and only
Sex a single ALK and single ROS1 fused patient progressed. A clin-
Male 130 (49) 98 (54) 0.337 ical summary of these rare fusions is included in Table S4.
Female 135 (51) 84 (46) MYB and MYBL1 Rearrangements
MYB and MYBL1 alterations were histologically enriched for an-
Age at Diagnosis
giocentric glioma (n = 14, 100%, p < 0.0001, Fisher’s exact test)
Under 10 years 185 (70) 83 (46) <0.0001
and diffuse astrocytoma (n = 5, 83%, p < 0.0001, Fisher’s exact
Over 10 years 80 (30) 99 (54) test), respectively, with both primarily arising in the cerebral
Mean 7.6 ± 4.8 10.1 ± 5.1 hemispheres (n = 13/14, 92% and n = 5/6, 83%, respectively,
Median 6.6 (0.5–18.9) 10.9 (0.2–18.9) p < 0.0001, Fisher’s exact test) (Figures 5E and 5F). All patients
Extent of Resection harboring either MYB or MYBL1 rearrangements are alive with
GTR 137 (52) 76 (44) 0.078 median follow-up of 8.1 and 5.3 years, respectively. However,
No GTR 127 (48) 96 (56) while progressions were rare in MYB-altered tumors (n = 2/10,
20%), they were more frequent in those with MYBL1 (n = 2/6,
Location
33%), resulting in a 5-year PFS of 90% and 67%, respectively
Cerebral hemisphere 86 (32) 97 (53) <0.0001
(Figures 5E and 5F). Our data suggest that the clinical differences
Midline 75 (28) 72 (40) between MYB- and MYBL1-altered tumors merit further
Cerebellum 104 (40) 13 (7) investigation.
Progression
Progressed 67 (27) 80 (46) <0.0001 Characteristics of SNV-Driven pLGG
Stable 184 (73) 94 (54) BRAF p.V600E
5-year PFS 70.6 51.4 BRAF p.V600E was the second most common alteration in
pLGG (17%) and was frequently associated with additional alter-
10-year PFS 59.8 30.0
ations, most commonly deletion of CDKN2A (n = 13, 9.6%) (Fig-
Outcome
ure 2A). BRAF p.V600E also co-occurred with several other
Dead 7 (3) 24 (13) <0.0001 SNVs, including those in NF1, FGFR1, KRAS, and H3F3A, but
Alive 249 (97) 155 (87) never with a fusion event (Figure 2A). Unlike KIAA1549-BRAF, tu-
5-year OS 97.8 91.2 mors with BRAF p.V600E were histologically diverse and
10-year OS 97.8 88.1 included ganglioglioma (n = 36, 31%), diffuse astrocytoma
Values are displayed as raw counts and the percentage of the group. (n = 16, 14%), and pleomorphic xanthoastrocytoma (n = 12,
*Denotes categories with omitted samples (LGG, NOS was not assigned 10%) (Figure 6A). Both ganglioglioma and pleomorphic xan-
a histological grade). GTR, gross total resection; PFS, progression-free thoastrocytoma were more likely to harbor BRAF p.V600E than
survival; OS, overall survival. other tumor types (p = 0.0028, p = 0.0048, respectively, Fisher’s
See also Tables S1 and S3. exact test, Figure S3C). BRAF p.V600E cases occurred most
frequently in the cerebral hemispheres (n = 64, 56%) but were
also common in the diencephalon (n = 33, 29%) and, in contrast
throughout the CNS, most commonly in the cerebral hemi- to KIAA1549-BRAF, were rare in the cerebellum (n = 6, 5.2%)
spheres (n = 6/14, 43%) (Figure 5B). FGFR1 TKD and FGFR2 (Figure 6A). BRAF p.V600E pLGG had worse OS and PFS than
fused pLGG were primarily glioneuronal or oligodendroglial in those with KIAA1549-BRAF (10-year OS 97% versus 89%,
origin, respectively and were restricted to the cerebral hemi- p = 0.0416 and 10-year PFS of 64% versus 30%, p = 0.0058,
spheres (Figures 5C and 5D). FGFR2 fusions included FGFR2- respectively, log rank test, Figures S6A and S6B). BRAF
INA, FGFR2-CTNNA3, and FGFR2-ERC1 (Figure S5). While p.V600E tumors with pleomorphic xanthoastrocytoma histology
progressions were seen in some cases (5-year PFS of 69%, had a worse outcome than those without (5-year PFS of 14%
69%, and 88% for FGFR1-TACC1, FGFR1 TKD, and FGFR2 versus 58%, respectively, p = 0.0328, log rank test, Figure S6C),
fused cases, respectively), none of these tumors resulted in pa- although OS was not significantly different (p = 0.1892, Fig-
tient death with a median follow-up of 11.3, 11.7, and 7.1 years, ure S6D). The same was not the case for BRAF p.V600E tumors
respectively (Figures 5B–5D). with co-occurring CDKN2A deletions (5-year PFS of 34% versus
ALK, ROS1, NTRK, and PDGFB Fusions 55%, respectively, p = 0.1157, log rank test, Figure S6E),
Fusions in other RTKs were rare in pLGG (3.4%) and included although OS was significantly different (p = 0.0100, log rank
novel events as well as some previously described in pediatric test, Figure S6F). However, both non-pleomorphic xanthoastro-
gliomas (Wu et al., 2014; Aghajan et al., 2016; Kiehna et al., cytoma (5-year PFS of 55%) and CDKN2A balanced (5-year PFS

576 Cancer Cell 37, 569–583, April 13, 2020

Figure 5. Clinicopathologic Features of Rearrangement-Driven pLGG
Schematic representation of key clinical features and outcomes for (A) KIAA1549-BRAF, (B) FGFR1-TACC1, (C) FGFR1 TKD, (D) FGFR2 fusions, (E) MYB, and (F)
MYBL1. PA, pilocytic astrocytoma; LGG, NOS, low-grade glioma, not otherwise specified; GG, ganglioglioma; DNET, dysembryoplastic neuroepithelial tumor;
PXA, pleomorphic xanthoastrocytoma; GNT, glioneuronal tumor; DA, diffuse astrocytoma; AG, angiocentric glioma; ODG, oligodendroglioma; DIA/DIG, des-
moplastic infantile astrocytoma/ganglioglioma; Dx, diagnosis; GTR, gross total resection. See also Figures S4 and S5 and Table S4.

of 55%) BRAF p.V600E tumors had inferior PFS to KIAA1549-
BRAF (5-year PFS of 69%) fused tumors (p = 0.0139 and
p = 0.0356, respectively, log rank test, Figures S7A and S7B),
despite their OS not being significantly different (p = 0.1169
and 0.1888, respectively, log rank test, Figures S7C and S7D).
FGFR1 Mutations
FGFR1 point mutations were observed in 1.5% of pLGG and pri-
marily consisted of p.N546K and p.K656E. Histologically, these
tumors were most frequently dysembryoplastic neuroepithelial tu-
mors (n = 13, 41%) or pilocytic astrocytoma (n = 9, 28%), diag-
nosed in older children (p = 0.032, Fisher’s exact test), and often
(n = 16, 50%) co-occurred with multiple genetic alterations,
including NF1 (n = 7, 22%) or additional RAS/MAPK pathway mu-
tations (n = 11, 34%) (Figures 2A and 6B). Interestingly, in some
cases, multiple point mutations in FGFR1 were seen (n = 6,
19%). Of those not lost to follow-up, 43% (n = 12) progressed
rapidly (median of 2.2 years, 5-year PFS of 53%). Despite this,
only two cases had late deaths after 13.7 and 15.5 years, respec-
tively, both of whom had additional alterations.
IDH1 p.R132H
IDH1 mutations are common in adult lower-grade gliomas,
arising in approximately 70% of grade II and III tumors (Parsons
et al., 2008; Yan et al., 2009, Balss et al., 2008). In pLGG, IDH1
p.R132H mutations were extremely rare, accounting for only
0.8% of cases (Figure 2B). Most IDH1 p.R132H patients pre-
sented with a prolonged history of seizures, sometimes years
before the biopsy was performed. All tumors were restricted to
the cerebral hemispheres and were either oligodendroglioma
or diffuse astrocytoma (Figure 6C). Patients harboring IDH1
p.R132H were diagnosed in late childhood (median: 15.7 years),

Cancer Cell 37, 569–583, April 13, 2020 577

Figure 6. Clinicopathologic Features of SNV-Driven pLGG
Schematic representation of key clinical features and outcomes for (A) BRAF p.V600E, (B) FGFR1 SNVs, (C) IDH1 p.R132H, and (D) H3.3 p.K27M. PA, pilocytic
astrocytoma; LGG, NOS, low-grade glioma, not otherwise specified; GG, ganglioglioma; DNET, dysembryoplastic neuroepithelial tumor; PXA, pleomorphic
xanthoastrocytoma; GNT, glioneuronal tumor; DA, diffuse astrocytoma; AG, angiocentric glioma; ODG, oligodendroglioma; DIA/DIG, desmoplastic infantile
astrocytoma/ganglioglioma; Dx, diagnosis; GTR, gross total resection. See also Figures S6 and S7.

with the youngest patient diagnosed at 8.9 years (Figure 6C).
Fifty percent of IDH1 p.R132H pLGG progressed within a median
of 5.1 years (5-year PFS of 56%) despite only one succumbing to
their disease (Figure 6C).
H3.3 p.K27M
H3F3A mutations are common in childhood high-grade gliomas
and DIPG and confer a dismal outcome (Khuong-Quang et al.,
2012; Buczkowicz et al., 2014). In our pLGG series, all H3F3A mu-
tations were p.K27M, with no p.G34R/V alterations identified
(Table S1). These cases were restricted to the midline (dienceph-
alon [n = 8] and brainstem [n = 4]), enriched for diffuse astrocy-
tomas (n = 8, 67%, p = 0.0011, Fisher’s exact test), and like other
SNVs, often co-occurred with other alterations (25%), most often
with BRAF p.V600E (Figures 2A and 6D). Although morphologically
and clinically different than midline HGG, H3.3 p.K27M patients
progressed early (median time to progression = 0.8 years), with
all patients eventually succumbing to their disease (Figure 6D).
These data support the role of H3.3 p.K27M as a marker of aggres-
sive behavior regardless of the initial morphology and
presentation.
Molecular-Based Risk Stratification for pLGG
Based on the above data, we define a risk stratification for children
with pLGG (Figure 7A). pLGG harboring gene fusions or germline

578 Cancer Cell 37, 569–583, April 13, 2020


NF1 mutations comprise the low-risk group. These tumors prog-
ress less frequently and often eventually stop growing with very
few progressions seen after 10 years and almost no deaths at
20 years follow-up (10-year PFS of 67% and OS of 98%, 20-
year PFS and OS of 58% and 96%, respectively) (Figures 7B
and 7C). These tumors require conservative management as ther-
apy may carry higher long-term morbidity than the tumor itself.
The intermediate-risk group of pLGG includes tumors with
BRAF p.V600E without CDKN2A deletion, FGFR1 SNV, IDH1
p.R132H, or MET mutations. Intermediate-risk tumors had a
10-year PFS and OS of 35% and 90%, respectively (Figures
7B and 7C). However, in contrast to the low-risk tumors, these
tumors continue to progress with a 20-year PFS of 27% and
20-year OS of 81%. Furthermore, they have a propensity for
acquiring additional alterations, which may result in the need to
refine treatment over time. These patients may therefore require
multiple treatment courses and longer-term follow-up than the
low-risk patients due to the risk of late death.
High-risk pLGG include those with H3.3 p.K27M, or BRAF
p.V600E with CDKN2A deletion. These tumors invariably prog-
ress (10-year PFS of 0%) and these patients often succumb to
their disease (10-year OS of 41%) (Figures 7B and 7C). Patients
with H3.3 p.K27M do worse than those with BRAF p.V600E and
CDKN2A deletion (10-year PFS and OS of 0% and 35% and 0%
and 60%, respectively), but both do far worse than low- and in-
termediate-risk patients. Although H3.3 p.K27M tumors are
more likely to result in patient death, both H3.3 p.K27M and
BRAF p.V600E and CDKN2A deletions result in rapid progres-
sion, indicating a need for immediate, aggressive treatment
and the introduction of novel, targeted agents.
Finally, pLGG with an undetermined molecular alteration (and
hence risk category) showed PFS and OS trends consistent with
Figure 7. Risk Stratification of pLGG
(A) Donut plot representing assigned risk portfolio
of pLGG and their associated biomarkers. Risk
assignment is based on the incidence of pro-
gression and/or death. In samples harboring mul-
tiple alterations, the highest potential risk group
was assigned. Alterations appearing in less than
five samples are not assigned a risk group.
(B) Kaplan-Meier plot of overall survival of cases
separated by risk. p value calculated via the log
rank test.
(C) Kaplan-Meier plot of progression-free of cases
separated by risk. p value calculated via the log
rank test.
See also Figure S8 Table S5.
representation of both low and intermedi-
ate risk (10-year PFS and OS of 51% and
92%, respectively, and 20-year PFS and
OS of 34% and 89%) (Figures S8A
and S8B).
In multivariate analysis, including tumor
location (midline), age at diagnosis, sex,
extent of resection (GTR), and histologi-
cal grade, risk group was determined to
be the most significant predictor of both
progression (hazard ratio = 4.030
[2.030–7.998], p < 0.0001, Cox proportional hazards model)
and death (hazard ratio = 16.547 [4.556–59.958], p < 0.0001,
Cox proportional hazards model) (Table S5).
DISCUSSION
Molecular studies of pLGG over the last decade have uncovered
oncogenic drivers shown to activate the RAS/MAPK pathway
(Collins et al., 2015; Jones et al., 2008; Schindler et al., 2011;
Dougherty et al., 2010; Lassaletta et al., 2017; Listernick et al.,
1999; Uusitalo et al., 2016; Seminog and Goldacre, 2013). How-
ever, despite these advances, the extent of molecular diversity,
the frequency of alterations in a population-based setting, and
the clinical significance of these diverse alterations are poorly
understood. In this study we perform combined morphological,
clinical, and molecular profiling of pLGG on a population-based
cohort with extensive clinical follow-up. This allowed us to
comprehensively investigate the molecular underpinnings and
provide comprehensive clinical insights for some of the rarest
of pLGG molecular subtypes. Furthermore, we introduce a
robust risk stratification system for pLGG, which has the poten-
tial to significantly influence the management of these tumors.
RAS/MAPK activation and pLGG were first linked due to the
appearance of OPGs in patients with NF1 (Listernick et al.,
1999) and this was further supported upon the discovery of
KIAA1549-BRAF and BRAF p.V600E in these tumors (Jones
et al., 2008; Schindler et al., 2011). Here, 378 tumors from
2000 to 2017 had an identifiable alteration affecting the RAS/
MAPK pathway with an additional ten showing upregulation via
ssGSEA analysis of RNA sequencing data. Therefore, of those
we could exhaustively profile (including RNA sequencing), 95%
(388/410) showed upregulation of the RAS/MAPK pathway.
Cancer Cell 37, 569–583, April 13, 2020 579
Thus, pLGG patients may benefit from RAS/MAPK pathway in-
hibitors even if no genomic alteration in the pathway is identified.
Indeed, this is supported by recent work showing favorable re-
sponses to MEK inhibitors in pLGG (Fangusaro et al., 2019).
Future work will inevitably look to investigate the alternative
mechanisms of RAS/MAPK activation in molecularly silent cases
which may include alternative splicing (Siegfried et al., 2013),
epigenetic changes, or miRNA alterations (Paroo et al., 2009).
However, understanding the specifics of these mechanisms
need not hinder the adoption of updated treatment protocols
that exploit the RAS/MAPK dependence of these tumors.
Our comprehensive approach to profiling these tumors
included morphologic, clinical, and molecular interrogation. In
utilizing these approaches, we were able to provide insights
into the clinical features of the rarest molecular entities of
pLGG (Figures 5B–5F and 6B–6D) and provide disease stratifica-
tion based on the type of molecular alteration driving the tumor
(Figures 4C, 4D, 7B, and 7C). In our dataset, rearrangement-
driven pLGG had a younger age of onset, were enriched for
WHO grade I histology, and had a less-aggressive clinical course
(Figures 4B–4D). This suggests that these oncogenic alterations
may occur early in development, promoting tumor initiation in a
developmental context permissive for one-hit tumorigenesis.
Previous work identified that BRAF fusions promoted gliogene-
sis in region-specific neural stem cell populations, while having
little effect in differentiated astrocytes (Kaul et al., 2012). When
originally identified, KIAA1549-BRAF was shown to have higher
kinase activity than BRAF p.V600E (Jones et al., 2008). This may
help explain why many KIAA1549-BRAF pLGG undergo sponta-
neous growth arrest; an environment where too much RAS/
MAPK upregulation promotes senescence and too little fails to
initiate tumor growth (Jacob et al., 2011; Raabe et al., 2011). In
comparison, SNV-driven pLGG were more commonly diag-
nosed at a later age, consistent with these alterations being ac-
quired later in development. Furthermore, SNV-driven pLGG co-
occurred with additional SNVs, but never co-harbored fusion
events; displaying a pattern of mutual exclusivity as seen across
many cancer types (Gao et al., 2018). Finally, SNV-driven pLGG
were associated with poorer outcome. Future work will need to
compare the mechanisms behind rearrangements and SNVs in
pLGG to elucidate why the observed clinical differences occur.
These results enabled us to develop a risk-based stratification
system for pLGG (Figure 7A). Genetic rearrangements, including
all fusions and duplications, as well as germline NF1 inactivation,
are considered low risk. As these tumors are rarely fatal, we pro-
pose they should be managed expectantly, with careful consid-
eration of additional treatment after surgery, and radiation
excluded from all post-operative treatment. For example, for
asymptomatic NF1-driven pLGG, surveillance is justified. How-
ever, if patients display progressive symptoms, most often as
vision loss, treatment with chemotherapy (Mahoney et al.,
2000; Packer et al., 1997) or targeted inhibitors (Banerjee et al.,
2017; Fangusaro et al., 2019) can be beneficial. Beyond 10 years
these tumors are much less likely to recur and the frequency of
follow-up may potentially be reduced. Intermediate-risk pLGG
are SNV-driven tumors, including those with BRAF or FGFR1
SNVs. These frequently harbor more than one alteration and
have a higher risk of recurrence which extends beyond 10 years.
These patients may require multiple treatments and longer
580 Cancer Cell 37, 569–583, April 13, 2020
follow-up than low-risk patients. Importantly, compared with
high-risk patients, intermediate-risk tumors rarely die of their dis-
ease, so efforts should focus on mitigating clinical progression.
Finally, high-risk tumors harboring H3.3 p.K27M or BRAF
p.V600E and CDKN2A deletion may require new approaches
to improve survival, including the development of novel agents
as well as combination therapies to promote synthetic lethality.
Importantly, as the era of novel targeted therapies such as
BRAF and MEK inhibitors inevitably arrives, the risk stratifica-
tions of these high- and intermediate-risk pLGG may change.
Nevertheless, significant long-term follow-up is required to
determine if the above is indeed true. The large number of pa-
tients in this cohort, long-term clinical follow-up data, and the
similarity between subgroups suggest that these findings are
robust and provide reliable information of critical importance to
clinicians today.
In conclusion, this comprehensive molecular landscape of the
clinical and molecular features of pLGG provides clinicians with
an invaluable resource for the management of common and rare
molecular pLGG subtypes. These data can guide diagnostic pro-
tocols and treatment approaches while aiding in expediting clin-
ical trials for new, better-targeted therapies for these children in
the near future.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d LEAD CONTACT AND MATERIALS AVAILABILITY
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Patient Samples
d METHOD DETAILS
B Nucleic Acid Extraction
B Droplet Digital PCR
B NanoString nCounter Analysis
B NanoString nCounter Vantage 3D for Protein Analysis
B Fluorescent In Situ Hybridization
B Immunohistochemistry
B SNP Array
B Targeted DNA Sequencing
B Whole Transcriptome Sequencing
B Gene Set Enrichment Analysis
B Genetic Analysis at Collaborating Institutions
d QUANTIFICATION AND STATISTICAL ANALYSIS
d DATA AND CODE AVAILABILITY
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2020.03.011.
ACKNOWLEDGMENTS
This work was supported by endowed funds from the Government of Canada
through Genome Canada and the Ontario Genomics Institute (OGI-121); A
Kid’s Brain Tumor Cure; Brain Tumor Research Assistance and Information
Network; The Pediatric Low-Grade Astrocytoma Foundation; Meagan’s
Walk; B.r.a.i.n.child Canada; Canadian Cancer Society (grant no. 702296); Ca-
nadian Institutes of Health Research (grant no. 159805); Restracomp
Scholarship and Fellowship funds from the Garron Family Chair in Childhood
Cancer Research at The Hospital for Sick Children (to S.R., M.Z., A.L., and
K.F.); Canadian Institutes of Health Research (CGS-M) scholarship (to S.R.);
Ontario Graduate Scholarship (to S.R.); Tokyo Children’s Cancer Study Group
(TCCSG) scholarship of the Gold Ribbons Network of Japan (to K.F.); The
Marie-Josée and Henry R. Kravis Center for Molecular Oncology; The National
Cancer Institute Cancer Center Core grant no. P30-CA008748; The American
Lebanese Syrian Associated Charities (ALSAC) of St. Jude Children’s
Research Hospital.
We thank colleagues Sergio Pereira and Jo-Anne Herbrick (both at the Cen-
ter for Applied Genomics at SickKids, Toronto, ON, Canada) for facilitating our
work in addition to Paula Marrano, Famida Spatare, and Monte Borden
(Department of Pathology, the Hospital for Sick Children, Toronto) and Cindy
Zhang (Brain Tumor Research Center, The Hospital for Sick Children, Toronto).
We further acknowledge the Members of the Molecular Diagnostics Service in
the Department of Pathology (MSK, New York, USA).
AUTHOR CONTRIBUTIONS
S.R., M.Z., U.T., and C.H. conceived the project and led the study. Molecular
characterization of samples from HSC were completed by S.R., M.Z., K.F.,
L.N., A.G.-S., J.B., A.L., M.J., and J.S. under the supervision of A.A., P.E.K.,
S.J., M.S., U.T., and C.H. Molecular characterization of samples from partner-
ing institutions were completed by R.G.T., W.O., I.Q., L.F.S., M.I.L., A.J.W.,
S.G., M.R., T.B., D.S.T., N.L., and C.D’A. under supervision from M.A.K., M.
Santi., and D.W.E. R.G.T., W.O., I.Q., L.F.S., M.I.L., A.J.W., S.G., M.R., T.B.,
D.S.T., N.L., C.D’A., A.K., G.M.I., J.D., P.D., M.D.T., J.T.R., S.L., M. Shroff.,
M. Shago., L.-N.H., V.R., E.B., U.B., A.H., M.A.K., M. Santi., and D.W.E. pro-
vided patient-related materials and/or clinical data used in this study. Histo-
pathological analysis was completed by A.A., M. Shago., C.D’A., and C.H.
Array and NGS data were processed and analyzed by S.R., R.S., and C.L. un-
der supervision from U.T. and C.H. Statistical analysis was done by S.R., M.Z.,
R.S., and C.L. U.T. and C.H provided overall supervision for the project and
wrote the manuscript with S.R. and M.Z. and with input from M.A.K., M. Santi.,
and D.W.E. All authors reviewed the manuscript and figures before
submission.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: January 3, 2020
Revised: January 27, 2020
Accepted: March 12, 2020
Published: April 13, 2020
REFERENCES
Aghajan, Y., Levy, M.L., Malicki, D.M., and Crawford, J.R. (2016). Novel
PPP1CB-ALK fusion protein in a high-grade glioma of infancy. BMJ Case
Rep. 16, 2016.
Anders, S., Pyl, P.T., and Huber, W. (2015). HTSeq-A Python framework to
work with high-throughput sequencing data. Bioinformatics 31, 166–169.
Balss, J., Meyer, J., Mueller, W., Korshunov, A., Hartmann, C., and von
Deimling, A. (2008). Analysis of the IDH1 codon 132 mutation in brain tumors.
Acta Neuropathol. 116, 597–602.
Bandopadhayay, P., Ramkissoon, L.A., Jain, P., Bergthold, G., Wala, J., Zeid,
R., Schumacher, S.E., Urbanski, L., O’Rourke, R., Gibson, W.J., et al. (2016).
MYB-QKI rearrangements in angiocentric glioma drive tumorigenicity through
a tripartite mechanism. Nat. Genet. 48, 273–282.
Banerjee, A., Jakacki, R.I., Onar-Thomas, A., Wu, S., Nicolaides, T., Young
Poussaint, T., Fangusaro, J., Phillips, J., Perry, A., Turner, D., et al. (2017). A
phase I trial of the MEK inhibitor selumetinib (AZD6244) in pediatric patients
with recurrent or refractory low-grade glioma: a Pediatric Brain Tumor
Consortium (PBTC) study. Neuro Oncol. 19, 1135–1144.
Benelli, M., Pescucci, C., Marseglia, G., Severgnini, M., Torricelli, F., and Magi,
A. (2012). Discovering chimeric transcripts in paired-end RNA-seq data by us-
ing EricScript. Bioinformatics 28, 3232–3239.
Broniscer, A., Baker, S.J., West, A.N., Fraser, M.M., Proko, E., Kocak, M.,
Dalton, J., Zambetti, G.P., Ellison, D.W., Kun, L.E., et al. (2007). Clinical and
molecular characteristics of malignant transformation of low-grade glioma in
children. J. Clin. Oncol. 25, 682–689.
Buczkowicz, P., Bartels, U., Bouffet, E., Becher, O., and Hawkins, C. (2014).
Histopathological spectrum of paediatric diffuse intrinsic pontine glioma: diag-
nostic and therapeutic implications. Acta Neuropathol. 128, 573–581.
Cheng, D.T., Mitchell, T.N., Zehir, A., Shah, R.H., Benayed, R., Syed, A.,
Chandramohan, R., Liu, Z.Y., Won, H.H., Scott, S.N., et al. (2015). Memorial
Sloan Kettering-integrated mutation profiling of actionable cancer targets
(MSK-IMPACT): a hybridization capture-based next-generation sequencing
clinical assay for solid tumor molecular oncology. J. Mol. Diagn. 17, 251–264.
Collins, V.P., Jones, D.T., and Giannini, C. (2015). Pilocytic astrocytoma: pa-
thology, molecular mechanisms and markers. Acta Neuropathol. 129,
775–788.
D’Angelo, F., Ceccarelli, M., Tala, Garofano, L., Zhang, J., Frattini, V., Caruso,
F.P., Lewis, G., Alfaro, K.D., Bauchet, L., et al. (2019). The molecular landscape
of glioma in patients with neurofibromatosis 1. Nat. Med. 25, 176–187.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Dougherty, M.J., Santi, M., Brose, M.S., Ma, C., Resnick, A.C., Sievert, A.J.,
Storm, P.B., and Biegel, J.A. (2010). Activating mutations in BRAF characterize
a spectrum of pediatric low-grade gliomas. Neuro Oncol. 12, 621–630.
Fangusaro, J., Onar-Thomas, A., Young Poussaint, T., Wu, S., Ligon, A.H.,
Lindeman, N., Banerjee, A., Packer, R.J., Kilburn, L.B., Goldman, S., et al.
(2019). Selumetinib in paediatric patients with BRAF-aberrant or neurofibro-
matosis type 1-associated recurrent, refractory, or progressive low-grade gli-
oma: a multicentre, phase 2 trial. Lancet Oncol. 20, 1011–1022.
Fina, F., Barets, D., Colin, C., Bouvier, C., Padovani, L., Nanni-Metellus, I.,
Ouafik, L., Scavarda, D., Korshunov, A., Jones, D.T., et al. (2017). Droplet dig-
ital PCR is a powerful technique to demonstrate frequent FGFR1 duplication in
dysembryoplastic neuroepithelial tumors. Oncotarget 8, 2104–2113.
Gao, Q., Liang, W.W., Foltz, S.M., Mutharasu, G., Jayasinghe, R.G., Cao, S.,
Liao, W.W., Reynolds, S.M., Wyczalkowski, M.A., Yao, L., et al. (2018).
Driver fusions and their implications in the development and treatment of hu-
man cancers. Cell Rep. 23, 227–238.
Ge, H., Liu, K., Juan, T., Fang, F., Newman, M., and Hoeck, W. (2011).
FusionMap: detecting fusion genes from next-generation sequencing data at
base-pair resolution. Bioinformatics 27, 1922–1928.
Greenberg, M.L., Barr, R.D., DiMonte, B., McLaughlin, E., and Greenberg, C.
(2003). Childhood cancer registries in Ontario, Canada: lessons learned from
a comparison of two registries. Int. J. Cancer 105, 88–91.
Guerreiro Stucklin, A.S., Ryall, S., Fukuoka, K., Zapotocky, M., Lassaletta, A.,
Li, C., Bridge, T., Kim, B.5, Arnoldo, A., Kowalski, P.E., et al. (2019). Alterations
in ALK/ROS1/NTRK/MET drive a group of infantile hemispheric gliomas. Nat.
Commun. 10, 4343.
Gutmann, D.H., Ferner, R.E., Listernick, R.H., Korf, B.R., Wolters, P.L., and
Johnson, K.J. (2017). Neurofibromatosis type 1. Nat. Rev. Dis. Primers
3, 17004.
Hartmann, C., Meyer, J., Balss, J., Capper, D., Mueller, W., Christians, A.,
Felsberg, J., Wolter, M., Mawrin, C., Wick, W., et al. (2009). Type and fre-
quency of IDH1 and IDH2 mutations are related to astrocytic and oligoden-
droglial differentiation and age: a study of 1,010 diffuse gliomas. Acta
Neuropathol. 118, 469–474.
Jacob, K., Quang-Khuong, D.A., Jones, D.T., Witt, H., Lambert, S., Albrecht,
S., Witt, O., Vezina, C., Shirinian, M., Faury, D., et al. (2011). Genetic aberra-
tions leading to MAPK pathway activation mediate oncogene-induced senes-
cence in sporadic pilocytic astrocytomas. Clin. Cancer Res. 17, 4650–4660.
Jones, D.T., Kocialkowski, S., Liu, L., Pearson, D.M., Ba €cklund, L.M.,
Ichimura, K., and Collins, V.P. (2008). Tandem duplication producing a novel
Cancer Cell 37, 569–583, April 13, 2020 581
oncogenic BRAF fusion gene defines the majority of pilocytic astrocytomas.
Cancer Res. 68, 8673–8677.
Jones, D.T., Gronych, J., Lichter, P., Witt, O., and Pfister, S.M. (2012). MAPK
pathway activation in pilocytic astrocytoma. Cell Mol Life Sci. 69, 1799–1811.
€ger, N., Korshunov, A., Kool, M., Warnatz, H.J.,
Jones, D.T., Hutter, B., Ja
Zichner, T., Lambert, S.R., Ryzhova, M., Quang, D.A., et al. (2013).
Recurrent somatic alterations of FGFR1 and NTRK2 in pilocytic astrocytoma.
Nat. Genet. 45, 927–932.
Kaul, A., Chen, Y.H., Emnett, R.J., Dahiya, S., and Gutmann, D.H. (2012).
Pediatric glioma-associated KIAA1549:BRAF expression regulates neuroglial
cell growth in a cell type-specific and mTOR-dependent manner. Genes
Dev. 26, 2561–2566.
Khuong-Quang, D.A., Buczkowicz, P., Rakopoulos, P., Liu, X.Y., Fontebasso,
A.M., Bouffet, E., Bartels, U., Albrecht, S., Schwartzentruber, J., Letourneau,
L., et al. (2012). K27M mutation in histone H3.3 defines clinically and biologi-
cally distinct subgroups of pediatric diffuse intrinsic pontine gliomas. Acta
Neuropathol. 124, 439–447.
Kiehna, E.N., Arnush, M.R., Tamrazi, B., Cotter, J.A., Hawes, D., Robison, N.J.,
Fong, C.Y., Estrine, D.B., Han, J.H., and Biegel, J.A. (2017). Novel GOPC(FIG)-
ROS1 fusion in a pediatric high-grade glioma survivor. J. Neurosurg. Pediatr.
20, 51–55.
Kim, D., and Salzberg, S.L. (2011). TopHat-Fusion: an algorithm for discovery
of novel fusion transcripts. Genome Biol. 12, R72.
Krishnatry, R., Zhukova, N., Guerreiro Stucklin, A.S., Pole, J.D., Mistry, M.,
Fried, I., Ramaswamy, V., Bartels, U., Huang, A.8, Laperriere, N., et al.
(2016). Clinical and treatment factors determining long-term outcomes for
adult survivors of childhood low-grade gliomas. Cancer 122, 1261–1269.
Lassaletta, A., Zapotocky, M., Mistry, M., Ramaswamy, V., Honnorat, M.,
Krishnatry, R., Guerreiro Stucklin, A., Zhukova, N., Arnoldo, A., Ryall, S.,
et al. (2017). Therapeutic and prognostic implications of BRAF V600E in pedi-
atric low-grade gliomas. J. Clin. Oncol. 35, 2934–2941.
Li, B., and Dewey, C.N. (2011). RSEM: accurate transcript quantification from
RNA-seq data with or without a reference genome. BMC Bioinformatics
12, 323.
Listernick, R., Charrow, J., and Gutmann, D.H. (1999). Intracranial gliomas in
neurofibromatosis type 1. Am. J. Med. Genet. 89, 38–44.
Louis, D.N., Perry, A., Reifenberger, G., von Deimling, A., Figarella-Branger, D.,
Cavenee, W.K., Ohgaki, H., Wiestler, O.D., Kleihues, P., and Ellison, D.W.
(2016). The 2016 World Health Organization classification of tumors of the cen-
tral nervous system: a summary. Acta Neuropathol. 131, 803–820.
Mahoney, D.H., Jr., Cohen, M.E., Friedman, H.S., Kepner, J.L., Gemer, L.,
Langston, J.W., James, H.E., Duffner, P.K., and Kun, L.E. (2000).
Carboplatin is effective therapy for young children with progressive optic
pathway tumors: a Pediatric Oncology Group phase II study. Neuro Oncol.
2, 213–220.
McPherson, A., Hormozdiari, F., Zayed, A., Giuliany, R., Ha, G., Sun, M.G.,
Griffith, M., Heravi Moussavi, A., Senz, J., Melnyk, N., et al. (2011). deFuse:
an algorithm for gene fusion discovery in tumor RNA-seq data. PLoS
Comput. Biol. 7, e1001138.
Mistry, M., Zhukova, N., Merico, D., Rakopoulos, P., Krishnatry, R., Shago, M.,
Stavropoulos, J., Alon, N., Pole, J.D., Ray, P.N., et al. (2015). BRAF mutation
and CDKN2A deletion define a clinically distinct subgroup of childhood sec-
ondary high-grade glioma. J. Clin. Oncol. 33, 1015–1022.
Nageswara Rao, A.A., and Packer, R.J. (2014). Advances in the management
of low-grade gliomas. Curr. Oncol. Rep. 16, 398.
Northcott, P.A., Pfister, S.M., and Jones, D.T. (2015). Next-generation
(epi)genetic drivers of childhood brain tumors and the outlook for targeted
therapies. Lancet Oncol. 16, e293–e302.
Ostrom, Q.T., de Blank, P.M., Kruchko, C., Petersen, C.M., Liao, P., Finlay,
J.L., Stearns, D.S., Wolff, J.E., Wolinsky, Y., Letterio, J.J., and Barnholtz-
Sloan, J.S. (2015). Alex’s Lemonade Stand Foundation infant and childhood
primary brain and central nervous system tumors diagnosed in the United
States in 2007–2011. Neuro Oncol. 16 (Suppl 10 ), x1–x36.
582 Cancer Cell 37, 569–583, April 13, 2020
Packer, R.J., Ater, J., Allen, J., Phillips, P., Geyer, R., Nicholson, H.S., Jakacki,
R., Kurczynski, E., Needle, M., Finlay, J., et al. (1997). Carboplatin and vincris-
tine chemotherapy for children with newly diagnosed progressive low-grade
gliomas. J. Neurosurg. 86, 747–754.
Paroo, Z., Ye, X., Chen, S., and Liu, Q. (2009). Phosphorylation of the human
microRNA-generating complex mediates MAPK/Erk signaling. Cell 139,
112–122.
Parsons, D.W., Jones, S., Zhang, X., Lin, J.C., Leary, R.J., Angenendt, P.,
Mankoo, P., Carter, H., Siu, I.M., Gallia, G.L., et al. (2008). An integrated
genomic analysis of human glioblastoma multiforme. Science 321,
1807–1812.
Qaddoumi, I., Orisme, W., Wen, J., Santiago, T., Gupta, K., Dalton, J.D., Tang,
B., Haupfear, K., Punchihewa, C., Easton, J., et al. (2016). Genetic alterations
in uncommon low-grade neuroepithelial tumors: BRAF, FGFR1, and MYB mu-
tations occur at high frequency and align with morphology. Acta Neuropathol.
131, 833–845.
Raabe, E.H., Lim, K.S., Kim, J.M., Meeker, A., Mao, X.G., Nikkhah, G.,
Maciaczyk, J., Kahlert, U., Jain, D., Bar, E., et al. (2011). BRAF activation in-
duces transformation and then senescence in human neural stem cells: a pilo-
cytic astrocytoma model. Clin. Cancer Res. 17, 3590–3599.
Ramkissoon, L.A., Horowitz, P.M., Craig, J.M., Ramkissoon, S.H., Rich, B.E.,
Schumacher, S.E., McKenna, A., Lawrence, M.S., Bergthold, G., Brastianos,
P.K., et al. (2013). Genomic analysis of diffuse pediatric low-grade gliomas
identifies recurrent oncogenic truncating rearrangements in the transcription
factor MYBL1. Proc. Natl. Acad. Sci. U S A 110, 8188–8193.
Robinson, M.D., McCarthy, D.J., and Smyth, G.K. (2010). edgeR: a
Bioconductor package for differential expression analysis of digital gene
expression data. Bioinformatics 26, 139–140.
Rusch, M., Nakitandwe, J., Shurtleff, S., Newman, S., Zhang, Z., Edmonson,
M.N., Parker, M., Jiao, Y., Ma, X., Liu, Y., et al. (2018). Clinical cancer genomic
profiling by three-platform sequencing of whole genome, whole exome and
transcriptome. Nat. Commun. 9, 3962.
Ryall, S., Krishnatry, R., Arnoldo, A., Buczkowicz, P., Mistry, M., Siddaway, R.,
Ling, C., Pajovic, S., Yu, M., Rubin, J.B., et al. (2016). Targeted detection of ge-
netic alterations reveal the prognostic impact of H3K27M and MAPK pathway
aberrations in paediatric thalamic glioma. Acta Neuropathol. Commun. 4, 93.
Ryall, S., Arnoldo, A., Krishnatry, R., Mistry, M., Khor, K., Sheth, J., Ling, C.,
Leung, S., Zapotocky, M., Guerreiro Stucklin, A., et al. (2017). Multiplex detec-
tion of pediatric low-grade glioma signature fusion transcripts and duplications
using the NanoString nCounter system. J. Neuropathol. Exp. Neurol. 76,
562–570.
Schindler, G., Capper, D., Meyer, J., Janzarik, W., Omran, H., Herold-Mende,
C., Schmieder, K., Wesseling, P., Mawrin, C., Hasselblatt, M., et al. (2011).
Analysis of BRAF V600E mutation in 1,320 nervous system tumors reveals
high mutation frequencies in pleomorphic xanthoastrocytoma, ganglioglioma
and extra-cerebellar pilocytic astrocytoma. Acta Neuropathol. 121, 397–405.
Seminog, O.O., and Goldacre, M.J. (2013). Risk of benign tumours of nervous
system, and of malignant neoplasms, in people with neurofibromatosis: pop-
ulation-based record-linkage study. Br. J. Cancer 108, 193–198.
Siegfried, Z.1, Bonomi, S., Ghigna, C., and Karni, R. (2013). Regulation of the
Ras-MAPK and PI3K-mTOR signalling pathways by alternative splicing in can-
cer. Int. J. Cell Biol. 2013, 568931.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L.,
Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., et al.
(2005). Gene set enrichment analysis: a knowledge-based approach for inter-
preting genome-wide expression profiles. Proc. Natl. Acad. Sci. U S A 102,
15545–15550.
Surrey, L.F., MacFarland, S.P., Chang, F., Cao, K., Rathi, K.S., Akgumus, G.T.,
Gallo, D., Lin, F., Gleason, A., Raman, P., et al. (2019). Clinical utility of custom-
designed NGS panel testing in pediatric tumors. Genome Med. 11, 32.
Tatevossian, R.G., Tang, B., Dalton, J., Forshew, T., Lawson, A.R., Ma, J.,
Neale, G., Shurtleff, S.A., Bailey, S., Gajjar, A., et al. (2010). MYB upregulation
and genetic aberrations in a subset of pediatric low-grade gliomas. Acta
Neuropathol. 120, 731–743.
 €a
Uusitalo, E., Rantanen, M., Kallionpa €, R.A., Pöyhönen, M., Leppa€virta, J., Yla
€-
Outinen, H., Riccardi, V.M., Pukkala, E., Pitka €niemi, J., Peltonen, S., et al.
(2016). Distinctive cancer associations in patients with neurofibromatosis
type 1. J. Clin. Oncol. 34, 1978–1986.
Wang, K., Li, M., and Hakonarson, H. (2010). ANNOVAR: functional annotation
of genetic variants from high-throughput sequencing data. Nucleic Acids Res.
38, e164.
Wisoff, J.H., Sanford, R.A., Heier, L.A., Sposto, R., Burger, P.C., Yates, A.J.,
Holmes, E.J., and Kun, L.E. (2011). Primary neurosurgery for pediatric low-
grade gliomas: a prospective multi-institutional study from the Children’s
Oncology Group. Neurosurgery 68, 1548–1554.
Wu, G., Diaz, A.K., Paugh, B.S., Rankin, S.L., Ju, B., Li, Y., Zhu, X., Qu, C.,
Chen, X., Zhang, J., et al. (2014). The genomic landscape of diffuse intrinsic
pontine glioma and pediatric non-brainstem high-grade glioma. Nat. Genet.
46, 444–450.
Yan, H., Parsons, D.W., Jin, G., McLendon, R., Rasheed, B.A., Yuan, W., Kos,
I., Batinic-Haberle, I., Jones, S., Riggins, G.J., et al. (2009). IDH1 and IDH2 mu-
tations in gliomas. N. Engl. J. Med. 360, 765–773.
Zhang, J., Wu, G., Miller, C.P., Tatevossian, R.G., Dalton, J.D., Tang, B.,
Orisme, W., Punchihewa, C., Parker, M., Qaddoumi, I., et al. (2013). Whole-
genome sequencing identifies genetic alterations in pediatric low-grade gli-
omas. Nat. Genet. 45, 602–612.
Cancer Cell 37, 569–583, April 13, 2020 583
 ll

Article
MYC Drives Temporal Evolution of Small
Cell Lung Cancer Subtypes
by Reprogramming Neuroendocrine Fate
Abbie S. Ireland,1 Alexi M. Micinski,1 David W. Kastner,1 Bingqian Guo,1 Sarah J. Wait,1 Kyle B. Spainhower,1
Christopher C. Conley,2 Opal S. Chen,3 Matthew R. Guthrie,1 Danny Soltero,1 Yi Qiao,4 Xiaomeng Huang,4
Szabolcs Tarapcsák,4 Siddhartha Devarakonda,5 Milind D. Chalishazar,1 Jason Gertz,1 Justin C. Moser,6 Gabor Marth,4
Sonam Puri,7 Benjamin L. Witt,8,9 Benjamin T. Spike,1 and Trudy G. Oliver1,10,*
1Department of Oncological Sciences, Huntsman Cancer Institute, University of Utah, Salt Lake City, UT 84112, USA
2Huntsman Cancer Institute Bioinformatic Analysis Shared Resource, Huntsman Cancer Institute, University of Utah, Salt Lake City, UT
84112, USA
3Huntsman Cancer Institute High-Throughput Genomics Shared Resource, Huntsman Cancer Institute, University of Utah, Salt Lake City, UT

84112, USA
4Utah Center for Genetic Discovery, Eccles Institute of Human Genetics, University of Utah, Salt Lake City, UT 84112, USA
5Division of Medical Oncology, Department of Medicine, Washington University School of Medicine, St. Louis, MO 63110, USA
6HonorHealth Research Institute, Scottsdale, AZ 85254, USA
7Department of Internal Medicine, University of Utah, Salt Lake City, UT 84112, USA
8Department of Pathology, University of Utah, Salt Lake City, UT 84112, USA
9ARUP Laboratories at University of Utah, Salt Lake City, UT 84108, USA
10Lead Contact

*Correspondence: trudy.oliver@hci.utah.edu
https://doi.org/10.1016/j.ccell.2020.05.001

SUMMARY

Small cell lung cancer (SCLC) is a neuroendocrine tumor treated clinically as a single disease with poor out-
comes. Distinct SCLC molecular subtypes have been defined based on expression of ASCL1, NEUROD1,
POU2F3, or YAP1. Here, we use mouse and human models with a time-series single-cell transcriptome anal-
ysis to reveal that MYC drives dynamic evolution of SCLC subtypes. In neuroendocrine cells, MYC activates
Notch to dedifferentiate tumor cells, promoting a temporal shift in SCLC from ASCL1+ to NEUROD1+ to YAP1+
states. MYC alternatively promotes POU2F3+ tumors from a distinct cell type. Human SCLC exhibits intratu-
moral subtype heterogeneity, suggesting that this dynamic evolution occurs in patient tumors. These
findings suggest that genetics, cell of origin, and tumor cell plasticity determine SCLC subtype.

INTRODUCTION terations in genes, such as EGFR, KRAS, or ALK, and targeting

Understanding molecular heterogeneity in cancer is critical for
precision medicine to tailor cancer treatments to the specific fea-
tures of a patient’s disease. For example, lung adenocarcinoma
comprises genetic subtypes with distinct, mutually exclusive al-
EGFR- or ALK-mutant tumors with targeted inhibitors prolongs
survival and improves patient outcome (Collisson et al., 2014;
Lin et al., 2016; Remon et al., 2019). In mouse models of other
tumor types, such as prostate cancer, the same oncogenes
can promote different tumor subtypes (i.e., luminal, basal) based

Significance

SCLC has historically been treated as a single disease, but is now recognized to comprise multiple molecular subtypes. We
find that MYC directly activates NOTCH signaling to reprogram neuroendocrine SCLC from ASCL1+ to NEUROD1+ to YAP1+
non-neuroendocrine states. Therefore, SCLC molecular subtypes are not distinct, but rather represent dynamic stages of
MYC-driven tumor evolution. We find that individual human SCLC tumors are composed of multiple molecular subtypes.
Given the reported unique therapeutic vulnerabilities of each subtype, we postulate that SCLC tumors represent a ‘‘moving
therapeutic target’’ that may require more general, combinatorial, or plasticity-directed therapeutic approaches to combat
this transcriptional flexibility. We speculate that molecular subsets of other cancer types may also represent dynamic stages
of tumor evolution.

60 Cancer Cell 38, 60–78, July 13, 2020 ª 2020 Elsevier Inc.
 ll
Article

A B

C D

F G H

Figure 1. MYC Drives Multiple SCLC Molecular Subtypes In Vivo

(A) Representative immunohistochemistry (IHC) for NE markers in early-stage (in situ) or invasive tumors in indicated GEMMs infected with Ad-Cgrp-Cre.
(B) IHC quantification from (A).
(legend continued on next page)
Cancer Cell 38, 60–78, July 13, 2020 61
 ll
Article

on the initiating cell of origin (Park et al., 2016; Wang et al., 2013).
The childhood cerebellar tumor medulloblastoma comprises ge-
netic subtypes (i.e., WNT, SHH) and subtypes that appear to pro-
ceed along a developmental trajectory or continuum (i.e., group
C and D) (Hovestadt et al., 2019). For most cancers, the relative
contribution of oncogenic pathway alterations, cell of origin, and
biological plasticity to the overall tumor phenotype (i.e., subtype)
is unknown.
Small cell lung cancer (SCLC) has historically been treated as a
single disease without patient stratification. SCLC exhibits ge-
netic loss of both tumor suppressors RB1 and TP53, along
with mutually exclusive expression of MYCL, MYC, or MYCN
(Bragelmann et al., 2017; Dammert et al., 2019; George et al.,
2015; Peifer et al., 2012; Poirier et al., 2015; Rudin et al., 2012).
Large-scale gene expression analyses of human tumors and
cell lines suggest that SCLC comprises four distinct molecular
subtypes based on expression of lineage-defining transcription
factors: ASCL1 (SCLC-A), NEUROD1 (SCLC-N), POU2F3
(SCLC-P), or YAP1 (SCLC-Y) (Rudin et al., 2019). Studies sug-
gest that SCLC subtypes have unique therapeutic vulnerabilities
(Cardnell et al., 2017; Chalishazar et al., 2019; Dammert et al.,
2019; Huang et al., 2018a, 2018b; Mollaoglu et al., 2017; Owoni-
koko et al., 2019), emphasizing the clinical importance of under-
standing these subtypes. The origins and relationships among
SCLC molecular subtypes are currently unknown (Poirier et al.,
2020; Rudin et al., 2019).
The SCLC-ASCL1 subtype comprises
70% of human SCLC.
ASCL1 is a master regulator of neuroendocrine (NE) fate that is
required for pulmonary NE cell (PNEC) development, and labels
adult PNECs (Ito et al., 2001). The adult PNEC is a cell of origin
for ASCL1+ tumors in multiple Rb1/Trp53 (RP)-null genetically en-
gineered mouse models (GEMMs) of SCLC (Meuwissen et al.,
2003; Park et al., 2011; Sutherland et al., 2011). In these GEMMs,
ASCL1 is essential for tumor development (Borromeo et al., 2016;
Kim et al., 2016). MYCL is amplified or highly expressed in the
SCLC-A subtype and is necessary for SCLC-A development (Bor-
romeo et al., 2016; Dooley et al., 2011; George et al., 2015; Huijb-
ers et al., 2014; Kim et al., 2016; McFadden et al., 2014). In
contrast, the other three SCLC subtypes representing
30% of
tumors (SCLC-N, SCLC-P, SCLC-Y) tend to exhibit amplification
or overexpression of MYC, and exhibit a low or non-NE cell fate
(George et al., 2015; Rudin et al., 2019). We showed that Myc
expression drives a non-NE SCLC phenotype in RP GEMMs
and tumors express NEUROD1 (Mollaoglu et al., 2017). However,
the relationship between SCLC-A and SCLC-N, and whether MYC
drives SCLC-P or SCLC-Y subtypes is unknown.
Notch signaling regulates PNEC fate and SCLC tumorigenesis
with dichotomous functions. Notch acts as a tumor suppressor
in SCLC with
25% of tumors harboring loss-of-function alter-
ations in Notch receptors (George et al., 2015). During lung injury,
Notch2 marks an NE-stem cell population (NEstem) that undergoes
self-renewal in the absence of Notch activation (Ouadah et al.,
2019). Loss of Notch function in the context of Rb1 and Trp53
loss is postulated to lock cells in a self-renewing NEstem-like state
and thereby contribute to SCLC. However, Notch is active in a
subset of human SCLC, and Notch activation can promote non-
NE SCLC fate (Lim et al., 2017). Moreover, an unknown signal
activates Notch signaling during lung injury to promote transit
amplification and deprogramming of NEstem cells to non-NE fates
(Ouadah et al., 2019). While MYC and NOTCH have both been
implicated in non-NE cell fate in SCLC, a functional relationship
between MYC and NOTCH in this setting has not been defined.
Here, we investigate the function of MYC in the origins and re-
lationships among SCLC molecular subtypes.
RESULTS
MYC Drives Multiple SCLC Molecular Subtypes In Vivo
To determine which SCLC molecular subtypes are promoted by
MYC, we analyzed tumors in Myc-driven (Rb1fl/fl;Trp53fl/fl;
Lox-Stop-Lox [LSL]-MycT58A, RPM) and Mycl-driven
(Rb1fl/fl;Trp53fl/fl;Rbl2fl/fl, RPR2) SCLC GEMMs (Figure S1A) at
time points with similar tumor burden (Kim et al., 2016; Schaffer
et al., 2010). We used adenoviral-Cgrp promoter-Cre viruses
(Ad-Cgrp-Cre) to specifically transform PNECs (Sutherland
et al., 2011). Early in situ lesions in RPM and RPR2 models exhibit
high expression of NE markers, including ASCL1, SYP, CGRP,
UCHL1, and DLL3 (Figures 1A, 1B, and S1B). Invasive RPR2
tumors retained an NE-high identity, whereas invasive RPM tu-
mors displayed significantly reduced NE marker expression (Fig-
ures 1A, 1B, and S1B). Bulk gene expression data from human
SCLC cell lines has shown that Myc, Notch/Rest, Hippo/Yap1,
and epithelial-mesenchymal transition (EMT) pathways correlate
with non-NE fate in SCLC (Zhang et al., 2018). We analyzed
expression of SCLC subtype markers and a subset of the non-
NE fate markers in RPM and RPR2 tumors. Invasive RPM tumors
gained expression of non-NE markers, including NEUROD1,
YAP1, HES1, ZEB1, and CD44 (Figures 1C, 1D, and S1B).
MYC expression was high in both in situ and invasive RPM tu-
mors, and not detectable in RPR2 tumors (Figure S1B). Interest-
ingly, POU2F3 was rarely detected in RPM tumors derived from
PNECs (Figure 1E). However, tumors initiated in RPM mice with a
general promoter (Ad-CMV-Cre), but not a club or alveolar type II
(AT2) promoter, were enriched for POU2F3 expression (Figures
1E, 1F, and S1C). POU2F3 was not detected in Mycl-associated
RPR2 tumors initiated from a general or NE promoter (Figure 1E).

(C) Representative IHC for non-NE markers in in situ or invasive tumors in indicated GEMMs infected with Ad-Cgrp-Cre.
(D) IHC quantification from (C).
(E) Representative IHC for POU2F3 in in situ or invasive tumors in indicated GEMMs infected with cell-type-specific Cre viruses. Positive control (+) is adult mouse
skin. IHC quantification in the right panel.
(F) Percentage of RPM tumors per animal that are POU2F3+ by IHC after cell-type-specific Cre viral infection. Indicated p values relative to Cmv.
(G) Left panel: percent of POU2F3+ (P) tumors expressing subtype markers analyzed by IHC from serial sections (n = 27 total tumors). Right panel: average
percentage of positive cells for each marker within individual POU2F3+ tumors.
(H) Representative IHC from serial sections of POU2F3+ tumors analyzed for SCLC subtype markers. Left color label indicates classification as in (G).
Data from n = 6–9 mice per genotype, except for Ccsp- and Spc-Cre mice, n = 5. Number of tumors scored by manual H-Score method is indicated within bar
graphs. Scale bars, 25 mm. Mean ± SEM, Student’s two-tailed unpaired t test, *p < 0.045, **p < 0.009, ***p = 0.002, ****p < 0.0001; n.s., not significant. See also
Figure S1.

62 Cancer Cell 38, 60–78, July 13, 2020

 ll
Article

A C
H&E

Classic
A

RPM
(Rb1, Trp53, MycT58A)

Variant
In situ

- -
ASCL1+ NEUROD1 YAP1
4 11 21

Days
Single-cell suspension

B
Day 4 7 11 14 17 21
RPM

Day 1 5 11 13 22 35
RPR2

D E
Day 5 7 10 12 14 17 19 21 24
ASCL1
MYC+

Vector MycT58A

MYC+
Day 22 25 30 22 25 30 Vector MycT58A
EPCAM
Day 28 32 35 37 28 32 35 37
INSM1 MYC

NEUROD1 ASCL1 MYC

REST INSM1
ASCL1
NOTCH2
YAP1
N2ICD
N2ICD
NOTCH2
N2ICD ZEB1 YAP1
HES1 YAP1
NEUROD1
MYC NEUROD1
HSP90
NKX2-1 HSP90
H1963
HSP90 H889

F
Circularity: 0.59 ± 0.06 Circularity: 0.64 ± 0.07
Roundness: 0.67 ± 0.03 Roundness: 0.65 ± 0.05
Vector

Vector

Cluster size: 79.4 microns2 ± 35.3 Cluster size: 137.3 microns2 ± 29.5

Circularity: 0.27 ± 0.03**** Circularity: 0.68 ± 0.04

MycT58A

MycT58A

Roundness: 0.65 ± 0.05 Roundness: 0.69 ± 0.04

Cluster size: 17.0 microns2 ± 3.5** Cluster size: 21.9 microns2 ± 8.3

H889 H1963

(legend on next page)

Cancer Cell 38, 60–78, July 13, 2020 63

 ll
Article

Analysis of serial sections of POU2F3+ tumors revealed that

44% lacked expression of other SCLC subtype markers,
whereas the majority of the remaining tumors expressed
POU2F3 and NEUROD1 (Figures 1G and 1H). The fraction of
cells expressing other subtype markers within POU2F3+ tumors
was relatively minor (<16% of cells) (Figures 1G and 1H). These
data suggest that MYC-driven SCLC-P tumors predominantly
arise from an unknown cell of origin that is not a PNEC, club,
or AT2 cell, and that MYC (as opposed to MYCL) promotes
SCLC-P development. Since POU2F3 is a master driver of the
tuft cell lineage (Huang et al., 2018b), future studies will be
required to determine whether MYC-driven SCLC-P tumors arise
from tuft cells.
Together, these data suggest that in the context of Rb1 and
Trp53 loss, MYC can promote SCLC-N and SCLC-Y molecular
subtypes from an ASCL1+ PNEC cell of origin. Furthermore,
MYC drives SCLC progression in PNECs from an NE-high to
non-NE phenotype. These findings are consistent with the
enrichment of MYC expression in human SCLC-N, -Y, and -P
subtypes (Rudin et al., 2019), and demonstrate that MYC is a
driver of these SCLC molecular subtypes in vivo.
MYC Drives SCLC Subtype Evolution In Vitro
Static observations of tumor histology provide limited insight into
the potential temporal connections between SCLC subtypes. To
address this point, we macro-dissected central portions of the
lung of Ad-Cgrp-Cre-infected RPM mice before detection of
macroscopic lesions at a time point with predominantly
ASCL1+ in situ tumors (Figure 2A). We cultured dissociated cells
and within 3–4 days, large clusters of tumor cells grew in suspen-
sion (Figure 2B). The tumor cells initially form tight, round, spher-
ical aggregates that resemble classic, NE-high human SCLC cell
lines (Figures 2B and 2C), but eventually form amorphous clus-
ters with a ‘‘chain-link’’ morphology that resembles variant,
non-NE human SCLC cells (Figures 2B, 2C, and S2A). This strik-
ing transition occurred reproducibly over a period of
20 days,
consistent with the time frame from early-to-invasive SCLC pro-
gression in the RPM GEMM (Mollaoglu et al., 2017). There was a
minor level of cell death during culture that was relatively consis-
tent throughout the time course (Figure S2B). Importantly, early-
stage RPR2 tumor cells maintained a classic morphology in
culture (Figures 2B and S2A).
To characterize the RPM tumor cell transition, we harvested
RPM tumor cells at multiple time points during culture and as-
sessed markers of SCLC subtypes. Cells at day 5 displayed an
NE-high identity, marked by ASCL1 and NE marker expression
(Figure 2D). NEUROD1 was transiently expressed from days 5
to 12, followed by expression of the non-NE markers REST,
YAP1, NOTCH2, and NOTCH2’s active cleaved intracellular
domain (N2ICD), and HES1 at days 10–24. In contrast, control
RPR2 tumor cells expressed ASCL1, and did not induce non-
NE markers (Figure S2C). Ectopic MycT58A-Ires-Gfp expression
in RPR2 cells triggered a variant morphology of looser clusters
or chains, and induced non-NE markers (Figures S2C and
S2D). This finding indicates that the NE-high to non-NE fate tran-
sition is not simply a consequence of prolonged culture.
To test if MYC is sufficient to alter tumor cell fate in established
human SCLC cells, we expressed MycT58A-Ires-Gfp in NE-high,
MYCL-associated classic lines. Ectopic MYCT58A expression
altered cell morphology from classic to variant-like, and induced
non-NE marker expression (Figures 2E and 2F). The levels of
ectopic MYC overexpression were consistently less than that
of MYC-high SCLC cell lines and RPM tumors (Figure S2E), sug-
gesting this phenotype is not due to supraphysiological MYC
levels. NEUROD1 expression was not detected in these experi-
ments (Figures 2E and S2C), suggesting that NEUROD1 expres-
sion is either early and transient (as in Figure 2D) or that YAP1
expression can be induced by MYC without NEUROD1 expres-
sion. Despite the gain of non-NE marker expression and variant
morphology induced by MYC, mouse and human MYC-express-
ing cells retained expression of some NE markers, including
ASCL1 (Figures 2E and S2C), suggesting either a
heterogeneous population, and/or the presence of hybrid NE/
non-NE cells. Overall, these data demonstrate MYC’s ability to
promote non-NE tumor cell fate and subtype evolution from
SCLC-A to SCLC-Y in tumor cells in vivo and in human
SCLC cells.
Human SCLC Subtypes Correspond with MYC-Driven
Evolution
Next, we sought to investigate the transcriptional states of MYC-
driven tumor evolution in human SCLC tumors. We performed
bulk RNA sequencing (RNA-seq) of RPM tumor cells from Ad-
Cgrp-Cre-infected mice at eight time points during the transition
spanning 21 days in culture. Analysis of NE and non-NE pathway
genes confirmed a temporal loss of NE identity, and a subse-
quent gain of non-NE signaling pathways, including Notch/
Rest, Hippo/Yap1, and EMT (Figures 3A and S3A) (Zhang

Figure 2. MYC Drives SCLC Subtype Evolution In Vitro

(A) Schematic of early RPM tumor cell isolation and culture. From left to right: whole slide H&E with bottom lung lobe (dashed box) provided in higher-magni-
fication inset. Scale bars, 500 mm, 100 mm (with successive magnification of airway lesions), and 10 mm (with in situ tumors [red and blue boxes] labeled in right top
panels). IHC of serial sections of the same tumors. Scale bars, 10 mm. Lungs were dissociated and single-cell suspensions placed in culture.
(B) Representative bright-field images of tumor cells in culture at indicated days after plating. Scale bars, 200 mm (top row), 100 mm (middle and bottom rows), and
50 mm (red inset). Representative of >60 independent assays for RPM cells and five assays for RPR2 cells.
(C) Representative bright-field images of human classic (H1092) and variant (H82) cell lines. Scale bar, 100 mm.
(D) Representative immunoblot on specified days following culture of early-stage RPM tumor cells. Representative of n = 5 independent assays. Bands
normalized to HSP90 values, with fold change relative to day 5 (red and black graphs) or day 24 (blue graphs) and averaged across three to five experiments.
(E) Representative immunoblot on specified days following puromycin selection from H889 (left) or H1963 (right) cells infected with vector control or MYCT58A
constructs. MYC-expressing human SCLC cells (GLC1) are used for positive controls (MYC+).
(F) Representative bright-field and GFP fluorescent images of H889 (left) and H1963 (right) human SCLC cells infected with retroviral control or MycT58A-Ires-Gfp
viruses. Scale bars, 100 mm (except H889 MycT58A, 50 mm). Average circularity, roundness and cluster size indicated ± SEM. Student’s unpaired two-tailed t test
for MYC versus vector control, ****p < 0.0001, **p < 0.009.
HSP90 serves as loading control. See also Figure S2.

64 Cancer Cell 38, 60–78, July 13, 2020

 ll
Article

A B
NE markers EMT NE genes Non-NE genes
10 Neurod1 6 Twist2

RPM early vs late

Enrichment score

Enrichment score
Chga Vim 0.8
0.0
8 Ascl1 Tgfb2 0.6
-0.2
0.4
Syp 4 Snai1 -0.4
6 Twist1
0.2
Uchl1 0.0 -0.6
4 Dll3 Snai2
Insm1 2 Tgfbr2
log2(RPKM fold change+1)

2 Nkx2-1 Zeb1 RPM early RPM late RPM early RPM late
Epcam
0 0 NES 5.15 p<0.0001 NES -3.75 p<0.0001
0 5 10 15 20 25 0 5 10 15 20 25

Notch signaling Hippo signaling

8 Dll4 8 Cy6r1 C
Dll3 Tead2 8
6 Hes6 6 Ajuba Ascl1
Neurod1

log2(RPKM+1)
Hes1 Wwtr1 6
4 Rest 4 Yap1 Yap1
Notch2 Tead3 4 Pou2f3
2 Jag2 2 Ctgf
Rbpj 2
0 0
0 5 10 15 20 25 0 5 10 15 20 25
0
Days in culture 0 5 10 15 20 25
Days in culture

D
Day 3-5 Day 7-10 Day 14-21
Enrichment score

Enrichment score

Enrichment score

0.20 0.05
0.25
0.20 0.15 0.00
0.15 0.10 -0.05
hSCLC-A

0.10 0.05 -0.10

0.05 0.00 -0.15
0.00 -0.05 -0.20

Day 3-5 Day >7 Day 7-10 Day <7, >10 Day 14-21 Day <14

NES 6.09 p<0.0001 NES 4.32 p<0.0001 NES -4.33 p<0.0001

Enrichment score

Enrichment score

Enrichment score

0.15 0.05
0.25 0.00
0.20 0.10 -0.05
0.15 -0.10
hSCLC-N

0.10 0.05 -0.15

0.05 0.00 -0.20
0.00 -0.25
-0.05 -0.05 -0.30

Day 3-5 Day >7 Day 7-10 Day <7, >10 Day 14-21 Day <14

NES 7.09 p<0.0001 NES 3.97 p<0.0001 NES -7.52 p<0.0001

Enrichment score

Enrichment score

Enrichment score

0.05 0.05 0.25

0.00 0.00 0.20
-0.05 0.15
-0.05 0.10
hSCLC-Y

-0.10 -0.10
0.05
-0.15 0.00
-0.15
-0.20 -0.05
-0.20

Day 3-5 Day >7 Day 7-10 Day <7, >10 Day 14-21 Day <14

NES -4.15 p<0.0001 NES -4.69 p<0.0001 NES 5.60 p<0.0001

Figure 3. Human SCLC Subtypes Correspond with MYC-Driven Evolution

(A) Log2 fold change of indicated NE (relative to the last time point) and non-NE pathway genes (relative to the first time point) from bulk RNA-seq of primary RPM
tumor cells at specific days in culture. Dashed lines in Notch signaling panel indicate genes predicted to be Notch-inhibitory.
(B) GSEA of 50 gene NE and non-NE gene signatures from Zhang et al. (2018) applied to early (days 3–7) versus late (days 14–21) time points of RPM transition.
(C) Log2 expression of SCLC-subtype defining transcription factor genes at indicated time points from bulk RNA-seq data of RPM transition.
(D) GSEA for human SCLC-ASCL1 (SCLC-A), SCLC-NEUROD1 (SCLC-N), and SCLC-YAP1 (SCLC-Y) gene signatures applied to bulk RNA-seq data grouped in
day increments.
See also Figure S3 and Table S1.

et al., 2018). Gene set enrichment analyses (GSEA) of an estab- scription factor genes Ascl1, Neurod1, and Yap1 during the tran-
lished 50 gene signature comprising 25 NE and 25 non-NE- sition (Figure 3C). Consistent with a lack of POU2F3 in RPM tu-
related genes from human SCLC cell lines (Zhang et al., 2018) mors derived from PNECs (Figure 1E), we observed extremely
demonstrated that MYC promotes a shift from NE to non-NE low counts of Pou2f3 mRNA during the transition (Figure 3C).
transcriptional states (Figures 3B and S3A). We observed To determine whether the expression of the subtype-defining
dynamic and sequential expression of the subtype-defining tran- transcription factors correlate with gene expression patterns in

Cancer Cell 38, 60–78, July 13, 2020 65

 ll
Article

A B
Days Day 4 11 17
25
4 7 11 14 17 21 7 14 21

tSNE 2
0

-25
(n = 15,434)
-50
-25 0 25
tSNE 1

Single cell RNA-seq

C
3 NE non-NE
1 Ascl1 Chga Rest Yap1
7
2 4 Min

Expression
2 2
5
6 2
Insm1 Epcam Hes1 Cd44
Day 4 (8174) 14 (5093) 1 Tumor cells 5 B/T cells Max
7 (4281) 17 (5086) 2 Stromal/myeloid 6 NK cells
11 (4371) 21 (4514) 3 AT2/club cells 7 Low quality cells
4 Neutrophils

D F
Pseudotime
Early Late
Epcam
log2 expression

3
Ascl1
Calca
0 Tcf4
Dll3
Pseudotime Day 4 7 11 14 17 21
Epha5
-3
Pou3f4
Early Late Syt11
Hdac9
Mycl
Sox11
Hes6
4% 26% 39% 30% Fez1
Myt1l
Sox1
Ncam
Uchl1
Bex2
RPM1 RPM2 Bex4
Chga
RPM3 RPM4 Myt1
4 11 17 Chgb
Day

7 14 21 Syp
Bex1
Insm1
RPM1 RPM2 RPM3 RPM4
(n = 2107) (n = 809) (n = 248) (n = 772) Runx1
Vim
Wwtr1
Anxa1
E Tgfb1
Tgfbr2
Neat1
Lgals1
Tead2
Twist2
Ascl1 Neurod1 Yap1 Pou2f3 Bmp4
Tuba1c
Cd44
Rest
Hes1
Tgfb2

Min
Expression

Mycl Myc

Max

Figure 4. MYC-Driven SCLC Subtypes Progress along a Single Evolutionary Trajectory

For a Figure360 author presentation of this figure, see https://doi.org/10.1016/j.ccell.2020.05.001.
(A) Schematic of primary RPM tumor cell transition analyzed by scRNA-seq with tumor cell populations colored by day, and number of cells analyzed per time
point in the legend (bottom left). Predicted cell types based on gene expression labeled on the same cells by number in the bottom right panel.
(legend continued on next page)
66 Cancer Cell 38, 60–78, July 13, 2020

Article
human tumors, we clustered
81 human SCLC tumors and
51
human SCLC cell lines according to molecular subtype using
bulk RNA-seq data (Figure S3B). We then created human
SCLC-subtype-specific gene signatures using the most highly
expressed genes per subtype (Table S1), and applied GSEA to
determine whether each signature was enriched or depleted dur-
ing the RPM transition (Figure 3D). Consistent with the patterns
of transcription factor gene expression (Figure 3C), days 3–5
and days 7–10 cells were enriched for human SCLC-A and
SCLC-N signatures, and depleted for the SCLC-Y signature. In
contrast, day 14–21 cells were enriched for the SCLC-Y
signature and depleted for SCLC-A and SCLC-N signatures,
demonstrating that MYC-driven tumor cell evolution in the
mouse corresponds with subtype-defining transcriptional signa-
tures in human tumors. We confirmed that Rb1 and Trp53 were
recombined in RPM tumor cells (Figures S3C and S3D). Whole-
genome sequencing of early and late time point RPM tumor cells
also confirmed complete loss of expected regions of Rb1 and
Trp53 with no detectable copy number variations (Figures S3E
and S3F). Together, MYC-driven tumor cell evolution in vitro re-
flects the temporal phenotypic changes observed in the RPM
GEMM, and places three of four human SCLC subtypes along
a defined MYC-driven trajectory from SCLC-A to SCLC-N to
SCLC-Y.
MYC-Driven SCLC Subtypes Progress along a Single
Evolutionary Trajectory
To better understand the transcriptional changes during MYC-
driven SCLC progression, we performed single-cell RNA-seq
(scRNA-seq). We isolated unsorted early-stage tumor cells
from RPM-Rosa26-LSL-Cas9-Ires-Gfp (RPM-Cas9) mice as
they were transitioning from days 4 to 21 in culture at six distinct
time points (Figure 4A). Approximately 4–8,000 cells were
captured per time point (n = 31,519 total cells). Days 4 and 7 cells
comprised both tumor and non-tumor cell populations and the
non-tumor populations were depleted in culture by day 11 (Fig-
ures 4A and S4A). For downstream analyses, non-tumor and
low-quality cells were filtered out of the time course (Figures
4B and S4A). Minor variation in gene expression due to cell-cycle
genes was regressed out for downstream analyses (Figure S4B).
Unsupervised tSNE clustering of tumor cells based on top highly
expressed genes revealed at least three distinct clusters corre-
sponding with the vast majority of days 4–7, 11, and 14–21 cells,
respectively (Figure 4B). Cells in the days 4–7 cluster expressed
high levels of Ascl1 and NE markers, which were largely absent in
days 11–21 cells that instead expressed high levels of non-NE
markers (Figure 4C). Although we captured only a few Neu-
rod1-expressing cells in the time course, these cells clustered
with NE-high cells (Figure S4C).
(B) Following removal of non-tumor and low-quality cells, RPM tumor cells labeled by day in tSNE space using Monocle 2.
(C) Expression of individual NE and non-NE marker genes in tSNE space from cells in (B).
(D) Pseudotime trajectory by Monocle 2 from early to late time points of primary RPM transition and four RPM-Cas9 tumors from Ad-Cgrp-Cre-infected mice.
Faceted pseudotime plots indicated on the top right and bottom right panels. Dashed insets in RPM tumors highlight percentage of cells in late stages of
progression; total post-QC number of tumor cells indicated below RPM1-4 labels.
(E) Expression of indicated genes projected onto pseudotime space as in (D).
(F) Heatmap of top 500 differentially expressed genes over pseudotime from the primary RPM tumor cell transition and 4 RPM tumors.
See also Figure S4 and Table S2.
ll
Next, we performed scRNA-seq on four invasive RPM-Cas9
tumors, one predominantly ASCL1-high (RPM1) and three that
had varying levels of NEUROD1 and YAP1 expression (RPM2-
4) (Figures S4D and S4E). Combining the RPM transition time
points and tumor samples, we constructed pseudotime trajec-
tories to identify predicted transcriptional relationships among
the tumor cells. Unsupervised pseudotime ordering of combined
tumor cells predicted a single lineage trajectory that corre-
sponds with early to late transition time points (Figures 4D–4F).
Faceted pseudotime trajectories reveal that cells of the RPM tu-
mors progress along the same trajectory as transitioning cells
in vitro, with the bulk of tumor cells in the earliest stages of pseu-
dotime and fewer cells progressing to later stages (Figure 4D).
RPM2-4 tumors had more cells in the later time points than
RPM1, consistent with their lower ASCL1 levels and higher
YAP1 protein expression (Figures 4D and S4D).
Expression of Ascl1, Neurod1, and Yap1 were consistent with
MYC-driven temporal evolution, whereas Pou2f3 was rarely de-
tected (Figure 4E). Mycl was associated with Ascl1 expression in
early time points and absent during tumor cell progression (Fig-
ure 4E), consistent with studies showing that ASCL1 and MYCL
are coordinately expressed in chemo-naive SCLC, and reduced
in chemotherapy-relapsed SCLC (Wagner et al., 2018). Expres-
sion of the top 500 differentially expressed genes across pseu-
dotime demonstrate broad loss of NE genes followed by gain
of non-NE genes (Figure 4F and Table S2).
We detected a small number of Ascl1-high cells at day 4 (n = 8
cells) that distributed to the late end of pseudotime (Figures 4D
and S4F); only one of these cells clustered with late-stage tumor
cells in tSNE space (Figures S4F and S4G), suggesting that these
cells are transcriptionally dissimilar from both early- and late-
stage tumor cells. As an orthogonal approach to predicting
cellular trajectories, we performed diffusion mapping (Angerer
et al., 2016; Coifman et al., 2005; Giraddi et al., 2018). Diffusion
mapping revealed the same cellular trajectory from early, NE-
high to late, non-NE cell states in transitioning cells in culture
and in tumors (Figure S4H). Diffusion components (DCs) that
best distribute time points in culture were compounded to define
a principal curve predicting diffusion pseudotime (i.e., cellular
trajectory) (Figure S4I). Diffusion pseudotime placed a small
number of day 4 cells late in pseudotime (Figure S4J), but these
cells were transcriptionally distinct from other day 4 cells and the
bulk of late cells from other time points and tumors (Figure S4K).
Diffusion analysis also revealed some features distinguishing pri-
mary tumor cells and transitioning cultured cells (e.g., DC6, DC4)
(Figure S4L). Monocle and diffusion map pseudotime coordi-
nates were significantly, positively correlated (Figure S4M).
Together, pseudotime analyses are consistent with our data sug-
gesting that MYC drives the temporal evolution of SCLC fate
from NE to non-NE states in culture and in tumors in vivo.
Cancer Cell 38, 60–78, July 13, 2020 67
 ll
Article

A B
ASCL1 NEUROD1 YAP1

100 2.6% ASCL1 only

10.5% 13% NEUROD1 only
27.1% YAP1 only (0%)
In situ tumors

75 1.7% ASCL1 + NEUROD1

% positive
83%
3.4% ASCL1 + YAP1
50 NEUROD1 + YAP1
86.8% 50.8% 87% ASCL1 + NEUROD1 + YAP1
25 1 subtype
3.4% >1 subtypes
13.6% 17%
0
In situ Invasive In situ Invasive

C
100 4.1%
Day 3-5
Invasive tumors

20.3%
80 Day 7-10

% expression
16.1% Day 12-14
60 Day 19-21
40 59.5%
20
0
RPM tumors (n = 10) Avg

D E
100 100
Ascl1 only Ascl1 only
35%
75 Neurod1 only 75 Neurod1 only
Yap1 only Yap1 only
% cells
% cells

10% Ascl1 + Neurod1

7% 50 Ascl1 + Neurod1
50 Ascl1 + Yap1 (<1%) Ascl1 + Yap1
Neurod1 + Yap1 (<1%) Neurod1 + Yap1 (0%)
25 48% Ascl1 + Neurod1 + Yap1 (<1%) 25 Ascl1 + Neurod1 + Yap1 (0%)

0
0
Day 4 7 11 14 17 21
1

g
PM

PM

PM

PM

Av
R

RPM1 RPM2 RPM3 RPM4

High
Expression

Ascl1
Low
88% 40% 59% 32%

High
Expression

Neurod1
Low
33% 16% 10% 9%

High
Expression

Yap1
Low
2% 44% 35% 62%

Figure 5. MYC-Driven Murine Tumors Exhibit Intratumoral SCLC Subtype Heterogeneity

(A) Representative IHC from in situ (n = 38) or invasive (n = 59) RPM tumors analyzed in serial sections. Left color panel indicates subtype classification matching
(B). Scale bar, 25 mm.
(legend continued on next page)
68 Cancer Cell 38, 60–78, July 13, 2020

Article
MYC-Driven Murine Tumors Exhibit Intratumoral
Molecular Subtype Heterogeneity
Pseudotime analyses suggest that individual RPM tumors are
composed of cells representing multiple SCLC molecular sub-
types (Figures 4D and 4E). To test this more comprehensively,
we analyzed serial sections of individual RPM tumors at the in
situ or invasive stage for subtype markers (Figures 5A and 5B).
In situ tumors were predominantly ASCL1+ and harbored a single
molecular subtype marker (
87%). In contrast, invasive tumors
predominantly harbored two or more subtypes, either ASCL1
and NEUROD1, or all three molecular subtypes (ASCL1, NEU-
ROD1, and YAP1). CIBERSORT analyses from an additional
ten RPM tumors also predicted intratumoral heterogeneity with
cells throughout the stages of MYC-driven progression (Fig-
ure 5C). Analyses of the RPM tumors by scRNA-seq suggest
that there is minimal co-expression of these markers within indi-
vidual cells, with the exception of Ascl1 and Neurod1 (Figure 5D),
consistent with their close connection in pseudotime. The
average abundance of Ascl1, Neurod1, and Yap1 in these tu-
mors is also relatively consistent with bulk RNA-seq and protein
analyses (Figures 5B–5D). scRNA-seq analyses of the RPM
transition in culture also shows little overlap between Ascl1, Neu-
rod1, and Yap1 in individual cells (Figure 5E), with day 11–21
cells resembling later stages of tumor progression. Thus, inde-
pendent methods suggest that MYC drives the evolution of mul-
tiple SCLC subtypes in vivo.
MYC Activates Notch Signaling during NE
Dedifferentiation
To uncover transcriptional patterns associated with MYC-driven
tumor cell evolution, we performed differential gene expression
analysis using the scRNA-seq data (Figure 6A). ENRICHR anal-
ysis of top differentially expressed genes between the major
clusters of cells (days 4–7 versus day 11 versus days 14–21)
identified transcription factors predicted to be important for
SCLC fate (Figure 6B) (Wooten et al., 2019), including REST,
SUZ12, TCF3, NEUROD1, NHLH1, MYC, and SOX2. As Notch/
Rest signaling can promote non-NE fate in SCLC (Lim et al.,
2017), and REST is a top predicted regulator of the earliest
changes promoted by MYC, we focused on the NE phenotypic
switch. Using the human-derived NE score vectors from (Zhang
et al., 2018) (Table S3), we assigned an NE score to every cell in
the RPM time-series and bulk tumors (Figure 6C). The NE score
accurately predicted the decrease in NE identity over time with a
dramatic reduction in NE score between days 7 and 11
(Figure 6D).
To determine how MYC promotes this transition at a mecha-
nistic level, we performed chromatin-immunoprecipitation
sequencing (ChIP-seq) for MYC in invasive RPM tumors (n = 4)
representing a spectrum of tumor cell states (Figure S5A). The
top 50 upregulated genes bound by MYC in RPM tumors whose
expression was enriched in MYC-high versus -low human SCLC
(B) IHC quantification from serial sections in (A) where individual tumors have detection of one or more than one (>1) subtypes.
(C) CIBERSORT analyses of bulk RNA-seq data from RPM tumors with average (Avg) percentage similarity to gene expression signatures of RPM transition cells
in culture.
(D) Top panel: percentage of cells per tumor expressing subtype-defining genes with average (Avg) across tumors. Bottom panel: individual RPM 1–4 trajectories
with localization of positive cells in pseudotime. Percentage of total tumor cells expressing indicated genes shown below trajectories.
(E) Percentage of cells expressing subtype-defining genes in the RPM transition experiment from Figure 4D.
ll
tumors and in mouse RPM versus RPR2 tumors were used to
define a conserved MYC-ChIP score (Table S3). Application of
the MYC-ChIP score to the time-series data revealed a small,
but significant increase in MYC activity from day 11 onward (Fig-
ure 6E), validating the ENRICHR predictions (Figure 6B). These
analyses suggest that days 7–11 represent a key transition state,
characterized by loss of NE identity and high MYC activity.
Because we observed Notch/Rest pathway induction during
MYC-driven SCLC progression, we analyzed the expression of
Notch-related machinery and target genes during the transition.
Expression of Notch-inhibitory factors were increased at early
time points (Figures S5B and S5C), including the inhibitory Notch
ligand gene Dll3, Hes6 (HES6 is a repressor of HES1), Fbxw7
(FBXW7 promotes NOTCH degradation), and Ncor2 (NCOR2
functions in a Notch-corepressor complex) (Gratton et al.,
2003; Matsumoto et al., 2011; Saunders et al., 2015). In contrast,
pro-Notch signaling genes increased in expression over time,
including the Notch target genes Hes1 and Rest, and the Notch
receptor gene, Notch2 (Figures S5B and S5C). GSEA of Notch
signaling and REST-transcriptional targets by bulk RNA-seq
reveals increased Notch pathway activity and repressive REST
activity in late compared with early transition time points (Fig-
ure S5D). Importantly, multiple Notch signaling components
were identified as MYC target genes, including Notch2, Hes1,
Hes6, and Jag2 (Figure 6F). Consistently, NOTCH2 and HES1
protein levels were induced by MYC in multiple cell types (Fig-
ures 2D, 2E, and S2C). Pro-Notch signaling genes NOTCH2,
HES1, and REST are also preferentially enriched in MYC-high
human SCLC (Figure 6G), while Notch-inhibitory HES6 is
reduced, suggesting broad positive regulation of Notch signaling
by MYC in SCLC. Together, these data suggest that MYC in-
creases NOTCH/REST activity to destabilize NE identity during
MYC-driven SCLC evolution.
Notch Activation Is Required for MYC-Driven Tumor
Evolution
To determine whether Notch signaling is required for MYC-
driven SCLC evolution, early RPM tumor cells were cultured in
vehicle control or the gamma-secretase inhibitor (GSI-IX) (e.g.,
DAPT) to block Notch signaling. In contrast to control cells that
began to convert to variant morphology on day 10, DAPT-treated
RPM cells did not switch to variant morphology until approxi-
mately day 20 (Figure 7A). Notch blockade increased and pro-
longed ASCL1, INSM1, EPCAM, and NEUROD1 expression
compared with control cells (Figures 7A, S6A, and S6B). In
contrast, DAPT treatment delayed or blocked expression of
non-NE markers (Figures 7B, S6A, and S6B). Interestingly,
when DAPT is added to Notch-active fully progressed RPM tu-
mor cells rather than cells in the classic/early state, we observe
equivalent expression of non-NE markers and no reversion to NE
marker expression, despite efficient Notch signaling blockade
(Figure S6C). Therefore, it is possible that MYC-driven tumor
Cancer Cell 38, 60–78, July 13, 2020 69
 ll
Article

A B
47 11 14 17 21 Day ENCODE and ChEA consensus TFs TF-Gene Coocurrence
REST ENCODE ISL1
Ascl1 REST CHEA INSM1

log2 expression
2 SOX2 CHEA MYT1
Epcam 1 TCF3 CHEA SOX11
SUZ12 CHEA NEUROD1
Insm1 0 SUZ12 ENCODE TOX3
Bex1 -1 FOSL2 ENCODE ST18
-2 RUNX1 CHEA NEUROD1
Uchl1 TCF3 ENCODE NEUROD4
GATA2 ENCODE NHLH1
Hes6
ENCODE and ChEA consensus TFs ENCODE and ChEA consensus TFs
MAX ENCODE MYC ENCODE
Hes1 MYC ENCODE SIN3A ENCODE
MYC CHEA TAF1 ENCODE
Odc1 TAF1 ENCODE MAX ENCODE
BRCA1 ENCODE BRCA1 ENCODE
Lgals1 SIN3A CHEA MYC CHEA
ATF2 ENCODE E2F1 CHEA
Lgals3 E2F6 CHEA PML ENCODE
ZMIZ1 CHEA NEF2L2 CHEA
Jun TCF3 ENCODE E2F6 ENCODE
Eno1
ENCODE and ChEA consensus TFs ChEA consensus TFs 2016
Vim TAF7 ENCODE E2F1 MESCs
NELFE ENCODE XRN2 HELA
Gstm1 TAF1 CHEA RUNX2 MC3T3
MYC ENCODE MYC MESCs
TCF3 ENCODE EKLF Erythrocyte
PML CHEA MYC MESCs
Fos BRCA1 ENCODE OCT4 MEFs
Fosb GATA1 CHEA SOX2 MEFs
E2F1 CHEA MYC MESCs
Egr1 SOX2 ENCODE SOX2 MESCs
Eef1a1
Up in days 4-7 v day 11 Up in day 11 v days 14-21
Up in day 11 v days 4-7 Up in days 14-21 v day 11

C D tumor
E
4 11 17 RPM1 RPM3

tumor
4 11 17 RPM1 RPM3
Day

Day
Expression 7 14 21 RPM2 RPM4 7 14 21 RPM2 RPM4
NE score Early ns
Min Max ***
pseudotime **** **** **** ns
***
0.8 **** 1.0 ****

MYC ChIP score

NE score **** **
25 ****
NE score

0.4 0.5
tSNE 2

Min Max
0
0.0 0.0
-25
Late -0.4 -0.5
-50 pseudotime
4 7 11 14 17 21
RPM1
RPM2
RPM3
RPM4

4 7 11 14 17 21

RPM1
RPM2
RPM3
RPM4
-25 0 25
tSNE 1 Day Day

F G
MYC H3k27Ac E-box motif
[0-3.32] [0-5.64]
RPM-1
15 **** 40 * MYC-high
NOTCH2 FPKM

RPM-2
MYC-low
HES1 FPKM

RPM-3
RPM-4
30
[0-1.96] [0-2.47] 10
RPM-2
RPM-3
20
RPM-4 5
10
Notch2-> Hes1-> 0 0
4 *** 400 ns
HES6 FPKM
REST FPKM

[0-7.13] [0-8.53]
3 300

2 200
[0-2.22] [0-4.41] 1 100

0 0

Jag2-> Hes6->

Figure 6. MYC Activates Notch Signaling during NE Reprogramming

(A) Heatmap of top 30 (or fewer, if <30 significant) differentially expressed genes for each time point of RPM transition using Seurat.
(B) ENRICHR analysis of top differentially expressed genes (0.25 log fold change) in RPM tumor cell transition compared by day as color-coded in the bottom
legend.
(C) Fifty gene NE score from (Zhang et al., 2018) applied to RPM tumor cell transition in tSNE (left) as in Figure 4B and in pseudotime space (right) as in Figure 4D.
(D) Violin plots of NE score from (C) applied to every cell of the RPM transition time points and individual RPM tumors. Student’s two-tailed unpaired t test, ****p <
2.22 3 1016.
(E) Violin plots of MYC-ChIP score applied to every cell of the RPM transition time points and individual RPM tumors. Student’s two-tailed unpaired t test, ****p <
2.22 3 1016, ***p < 0.0004, **p < 0.002; n.s., not significant.
(F) ChIP-seq analysis of MYC (red) and H3K27Ac (blue) genomic binding at indicated gene loci from n = 3–4 independent RPM tumor samples. Blue rectangles
below plots indicate gene exons with directionality of gene (->) near gene name. Black up-arrows indicate canonical E-Box 5’-CACGTG-3’ or non-canonical 5’-
CANNTG-30 motifs (‘‘E-box motif’’) selected in the vicinity of observed MYC binding.
(legend continued on next page)
70 Cancer Cell 38, 60–78, July 13, 2020

Article
evolution may be irreversible, but it remains to be determined
whether blockade of other non-NE-related pathways could
revert SCLC to an NE-high state. Next, we treated RPM mice
with early-stage tumors with DAPT to block Notch signaling
in vivo and monitored tumor growth by microCT imaging for
10 days, since control animals live an average of
12 days after
tumor detection by microCT imaging. DAPT treatment led to a
significant decrease in tumor burden in DAPT-treated versus
control mice (Figure 7C). Quantification of H&E-stained tissue re-
vealed that tumor burden was significantly reduced by DAPT
treatment (Figure 7D). DAPT-treated RPM tumors exhibited an
increase in classic morphology indicated by smaller cells with
high nuclear:cytoplasmic ratios compared with control tumors
and a significant increase in NE identity, measured by ASCL1
and DLL3 levels (Figures 7E and 7F). Moreover, DAPT-treated tu-
mors exhibited a decrease in tumor progression compared with
control tumors evident by reduced expression of NEUROD1 and
YAP1. We did not detect differences in HES1 expression, likely
because HES1 levels were already low in control tumors at these
time points (Figures 7E and 7F). MYC levels were high in both
treatment groups with >90% of cells positive for MYC, suggest-
ing that the impact of Notch inhibition on cell state is not due to
reducing MYC levels. Treatment of tumor-bearing RPM mice
with a second Notch inhibitor (dibenzazepine) also significantly
reduced tumor burden compared with control animals, but was
highly toxic limiting analysis to day 7 (Figures S6D and S6E).
Together, these findings suggest that Notch inhibition signifi-
cantly inhibits MYC-driven tumor progression.
Loss-of-function NOTCH mutations occur in
25% of human
SCLC, with mutually exclusive alterations in NOTCH1, -2, -3, and
-4 (George et al., 2015). We hypothesized that MYC requires
intact NOTCH to drive SCLC subtype evolution. Using published
functional classifications of NOTCH mutations, we grouped hu-
man SCLC tumors and cell lines as either NOTCH wild-type
(WT) or silent, non-damaging, or damaging. Consistent with
our hypothesis, MYC expression is significantly increased in
NOTCH WT or silent mutant compared with NOTCH-damaging
mutant samples (Figure 7G). All MYC-high SCLC samples (n =
30) are predicted to have intact NOTCH (Figure 7G, left panel).
Many MYC-expressing tumors with intact NOTCH exhibit low
NE scores, whereas all of the NOTCH-damaging mutant tumors
are NE-high (Figure 7G, right panel). We mined a recent cell line
genomics database (SCLC_CellMiner) (Tlemsani et al., 2020),
which allowed a similar analysis with 50 additional human
SCLC cell lines (Figure 7H). There was a significant difference
in the quantity and proportion of MYC-high NOTCH WT versus
MYC-high NOTCH mutant samples that were NE-low, consistent
with our model that MYC promotes non-NE progression by acti-
vating NOTCH signaling. GSEA suggests that Notch signaling
and MYC activity are enriched in NOTCH WT human SCLC
compared with samples with damaging mutations (Figure 7I; Ta-
ble S4). Together, these results suggest that MYC depends on
NOTCH to promote NE dedifferentiation, and that Notch
blockade can inhibit MYC-driven tumor progression.
(G) Gene expression from George et al. (2015) analyzed according to MYC status (n = 59 MYC-low and n = 11 MYC-high tumors). Mean ± SEM, Student’s two-
tailed unpaired t test, ****p < 0.0001, ***p < 0.0007, *p < 0.03; n.s., not significant.
For boxplots, the median and interquartile range are shown (lower bar is the 25th percentile, upper bar is the 75th percentile, and endpoints indicate minimum and
maximum values). See also Figure S5 and Table S3.
ll
Human SCLC Exhibits Intratumoral Molecular Subtype
Heterogeneity
We sought to determine the abundance of RPM time point gene
signatures in bulk RNA-seq data from human SCLC tumors and
cell lines (George et al., 2015; Newman et al., 2015). CIBERSORT
analyses revealed that most human tumors are predicted to have
cells resembling multiple stages of MYC-driven progression,
regardless of their classified SCLC subtype (Figure 8A), reminis-
cent of a similar approach predicting varying proportions of
SCLC phenotypes in human tumors (Wooten et al., 2019). The
majority of tumors harbor cells that resemble the NE-high
ASCL1+ early time point RPM cells (Figures 8A and 8B). The hu-
man SCLC-N subtype was enriched for the day 7–10 time point
signature and the human SCLC-Y samples exhibit a significantly
higher percentage of non-NE late time point signatures (Figures
8A and 8B). This suggests that individual human tumors have
cells in multiple stages of tumor progression, with MYC-high tu-
mors more likely to have cells at the latest stages of progression.
Interestingly, some SCLC-A tumors were predicted to have a
higher percentage of late-stage cells; these tumors had moder-
ately higher MYC levels and significantly higher levels of
NOTCH2, HES1, and REST, consistent with a reduction in NE
identity (Figures 8A and 8C).
To verify these predictions in human tissue, we obtained 21
chemo-naive human SCLC biopsies (n = 6 limited and 15 exten-
sive stage), since surgical specimens in SCLC are rare and diffi-
cult to obtain due to metastatic disease. We performed immuno-
histochemistry for ASCL1, NEUROD1, YAP1, and MYC on serial
sections. Many tumors harbored cells with more than one sub-
type, with the majority of tumors having some frequency of
ASCL1 and/or NEUROD1 protein (Figures 8D and 8E), consistent
with human gene expression data (Figure 8A). Approximately
14% of samples harbored detectable MYC protein and, of these,
none of them were in the ASCL1-only group. Finally, we obtained
one fresh human SCLC biopsy from a patient who briefly re-
sponded to carboplatin and etoposide and then progressed on
carboplatin-irinotecan for scRNA-seq analyses. The biopsy
harbored distinct ASCL1+ and NEUROD1+ populations, and
high MYC expression overlapped specifically with the NEU-
ROD1-high population (Figures 8F, 8G, S7A, and S7B). While
these data warrant a more comprehensive analysis in human tis-
sue, these findings are consistent with bioinformatic predictions
suggesting that tumors frequently harbor cells representing mul-
tiple SCLC subtypes.
DISCUSSION
While SCLC has historically been treated as a single disease,
recent studies have converged on the concept that SCLC is
composed of at least four molecular subsets with unique thera-
peutic vulnerabilities. Recent bioinformatic approaches suggest
that SCLC subtypes may represent dynamic states of transition
(Wooten et al., 2019). Our functional data here support that hy-
pothesis and suggest that MYC has the capacity to shift SCLC
Cancer Cell 38, 60–78, July 13, 2020 71
 ll
Article

A B
DMSO DAPT
Day 4 10 13 15 20 Day 4 7 10 13 15 4 7 10 13 15
**
ASCL1
DMSO

**
INSM1
*
ZEB1
YAP1
DAPT

N2ICD

REST
DMSO
DAPT HSP90

C Day 1 Day 10 D
DAPT Control DAPT **
Control 40
30 Control

% tumor volume by mCT

Control

* 21.0%
**
% tumor volume

30 DAPT
20
6.7%
ns 20
10
10
DAPT

0
Day 1 7 10 0

E F G
MYC-high ns
Control DAPT ASCL1 DLL3 15 ns
* 1.0
** **** ns
MYC log2(TPM+1)
% pos tumor cells
H&E

100 100 NOTCH status

10 0.5
WT or silent

NE Score
non-damaging
50 50 0.0
damaging
5
-0.5
0 0
ASCL1

0 -1.0
NEUROD1 YAP1
****
H
% pos tumor cells
DLL3

100 100 ***

1.0
MYC-high
** MYC-high MYC-high
50 50 0.5 MYC-low
NEUROD1

NOTCH WT NOTCH mutant

NE score

NE-high 28 14
0 0 0.0
NE-low 25 1
-0.5
p = 0.0055
YAP1

HES1 MYC
**** -1.0
ns NOTCH WT Mutant
100
% pos tumor cells

100
HES1

50 50 I
Notch signaling MYC ChIP score
NOTCH WT vs mutant
Enrichment score

Enrichment score

0.20 0.20
0 0 0.15 0.15
MYC

0.10 0.10

Control DAPT 0.05 0.05

0.00
0.00

NOTCH WT NOTCH mutant NOTCH WT NOTCH mutant

NES 1.78 p<0.02 NES 1.80 p<0.002

Figure 7. Notch Activation Is Required for MYC-Driven Tumor Evolution

(A) Representative bright-field images of primary RPM tumor cells cultured in DMSO or 10 mM DAPT treated every 3 days and visualized at indicated days. Red
box on each row indicates the day that variant morphology was first observed. Scale bar, 100 mm. n = 4 biological experiments.
(B) Immunoblot from cells in (A). HSP90 serves as loading control. Bar graphs on the left represent fold change in expression, summed across all time points
where protein was detected, relative to HSP90. Error bars represent mean ± SEM for n = 4 biological replicates. Student’s two-tailed unpaired t test, **p < 0.004,
*p < 0.03, n.s., not significant.

(legend continued on next page)

72 Cancer Cell 38, 60–78, July 13, 2020
 ll
Article

molecular subtypes. We find, in the context of Rb1 and Trp53 loss,
MYC can promote three of the four molecular subsets from a
PNEC cell of origin that proceed in a temporal evolution from
SCLC-A to SCLC-N to SCLC-Y in vivo. Studies in a limited number
of human SCLC cell lines and patient-derived xenograft models
also suggest that MYC can convert ASCL1+ SCLC to a variant
morphology with NEUROD1 expression (Johnson et al., 1986; Pa-
tel et al., 2019; Simpson et al., 2020). Interestingly, we observed
SCLC-P in RPM mice when an unknown cell type was targeted
that we speculate could be the tuft cell (Huang et al., 2018b; Rudin
et al., 2019). These data demonstrate that cell of origin, genetics,
and tumor cell plasticity can determine SCLC subtype.
Our data suggest that MYC requires Notch pathway activity to
promote tumor progression. Notch activity can promote non-NE
fate in Mycl-associated SCLC models in the absence of MYC
expression (George et al., 2015; Lim et al., 2017). In the RPR2
model, tumors with MYC-negative Notch-active cells do not
develop variant morphology or express NEUROD1 or YAP1, in
contrast to MYC-expressing Notch-active cells in our study.
Together, these findings suggest that Notch activity alone is
not sufficient to drive SCLC-N and SCLC-Y subtypes, and that
MYC and NOTCH likely cooperate with one another to drive
SCLC progression.
Recently identified NEstem cells are enriched for Notch2 and
Hes1 expression (Ouadah et al., 2019). Hes1 is particularly en-
riched in RPM time points where MYC-driven tumor cells are
transitioning to a non-NE fate. HES1 is also enriched in
ASCL1+ human SCLC tumors that exhibit a more non-NE fate
in the absence of high MYC, consistent with a model whereby
Notch can promote non-NE fate, but may be limited in its ability
to promote progression without MYC. We identify Notch2 and
Hes1 as MYC target genes, suggesting that MYC could be the
unidentified signal that induces Notch signaling to deprogram
NEstem cells to other cell fates during lung injury (Ouadah et al.,
2019). We propose a model whereby NEstem-like tumors with
defective Notch signaling are locked in an NE-high state,
whereas tumors with intact Notch can be reprogrammed by
MYC to non-NE fates—explaining the dichotomous nature of
Notch signaling in SCLC (Figure 8H). While our pharmacological
studies suggest that Notch activity is critical for MYC-driven tu-
mor progression, genetic Notch knockout studies are warranted
to fully test this model. These data warrant evaluation of pharma-
cological approaches that activate Notch in SCLC (Augert et al.,
2019; Oser et al., 2019) to determine whether they can promote
tumor progression, particularly in collaboration with MYC.
The RPM primary cell culture model should be a valuable tool
to better understand the role of additional signaling pathways in
MYC-driven tumor evolution. Lineage-tracing approaches in this
model in vitro and in vivo would be a powerful complement to the
studies here. As some of our assays suggest that SCLC can
progress from SCLC-A to SCLC-Y without evidence of the
SCLC-N subtype, further studies are warranted to determine
whether NEUROD1 is required or dispensable for SCLC progres-
sion. The functions of Hippo/Yap1 and EMT in SCLC (also
induced by MYC) are less well understood than Notch signaling
(Horie et al., 2016; Jia et al., 2018; McColl et al., 2017). Previous
studies suggest that SCLC-Y human tumors are often RB1 WT
(McColl et al., 2017). We observe MYC-high YAP1+ tumor cells
that appear to lack Rb1, suggesting there may be more than
one mechanism to induce YAP1 in SCLC.
Our work builds on an emerging concept that MYC and MYCL
are not functionally redundant in SCLC. MYC and MYCL corre-
late with distinct gene expression and methylation profiles, and
localize to distinct super enhancers (Borromeo et al., 2016;
Christensen et al., 2014; Poirier et al., 2015). Functional studies
modulating MYC reveal MYC’s capacity to change cell fate,
morphology, drug sensitivity, and molecular subtype (Dammert
et al., 2019; Mollaoglu et al., 2017; Patel et al., 2019). In other tu-
mor types, such as medulloblastoma, MYC and MYCN also
appear to have distinct roles (Vo et al., 2016), illustrating diver-
gent functions for MYC family members in cancer.
We find that many SCLC tumors are composed of cells repre-
senting multiple molecular subtypes at different frequencies,
suggesting that bulk analysis methods may only identify the
most abundant state in a given tumor. Studies are increasingly
identifying unique therapeutic vulnerabilities for SCLC subtypes.
For example, MYC-high tumors are preferentially sensitive to in-
hibition of AURKA/B, CHK1, IMPDH1/2, and arginine deprivation
(Cardnell et al., 2017; Chalishazar et al., 2019; Dammert et al.,
2019; Huang et al., 2018a; Mollaoglu et al., 2017; Sen et al.,
2017). POU2F3+ SCLC cells are sensitive to IGF1R inhibitors
(Huang et al., 2018b), while ASCL1+ tumors express more
DLL3 and are sensitive to DLL3-targeting drugs (Cardnell et al.,
2017; Saunders et al., 2015). Considering data that SCLC

(C) Representative microCT (mCT) imaging from vehicle control- (corn oil) (n = 6) and 20 mg/kg DAPT-treated (n = 8) RPM mice. Lung tumors pseudocolored in
yellow and heart outlined in red. Quantification of microCT imaging data for total tumor burden (percentage of lung tumor volume/total lung volume) in the graph
on the right at the indicated days. Mean ± SEM, Student’s two-tailed unpaired t test, **p < 0.004, *p < 0.03; n.s., not significant.
(D) Representative H&E from vehicle control- (n = 7) or DAPT-treated (n = 6) RPM mice at day 10 following treatment. Scale bar, 2,000 mm. Quantification of
average tumor burden (% tumor area/total lung area) in right panel with boxplots where each dot represents one animal, Student’s two-tailed unpaired t test,
**p < 0.009.
(E) Representative H&E and IHC in vehicle control- (n = 7) or DAPT-treated (n = 6) RPM mice at day 10 after treatment. Scale bars, 20 mm.
(F) Digital IHC quantification from lung tumor tissue (percentage of positive tumor cells) in (E) where each dot represents a tumor. Student’s two-tailed unpaired t
test, ****p < 0.0001, **p < 0.03; n.s., not significant.
(G) Left panel: MYC expression in normalized transcripts per million (TPM) + 1 grouped by NOTCH status with MYC-high samples in blue; RNA-seq data from n =
70 human tumors (George et al., 2015) and n = 48 human cell lines (CCLE), excluding POU2F3+ samples. Right panel: NE score applied to the same samples.
‘‘Non-damaging’’ indicates missense or in-frame deletions, and ‘‘damaging’’ indicates frameshift deletions, nonsense, or splice variant mutations. Mean ± SEM,
Student’s two-tailed unpaired t test with Welch’s correction, *p < 0.01; n.s., not significant.
(H) NE score applied to RNA-seq data from n = 70 human tumors (George et al., 2015) and n = 98 human cell lines (SCLC_CellMiner), excluding POU2F3+
samples; Student’s two-tailed unpaired t test with Welch’s correction, ***p = 0.0009. Contingency table analyzed by Fisher’s exact test.
(I) GSEA comparing hSCLCs with NOTCH WT or silent (‘‘WT’’) versus ‘‘damaging’’ mutations from samples in (G). Normalized enrichment scores (NES) indicated.
For boxplots, the median and interquartile range are shown (the lower bar is the 25th percentile, the upper bar is the 75th percentile, and endpoints indicate
minimum and maximum values). See also Figure S6 and Table S4.

Cancer Cell 38, 60–78, July 13, 2020 73

 E
C
A

G
% expression log2(TPM+1)
HES1 MYC
YAP1 NEUROD1 ASCL1 MYC

0
2
4
6
8
10

0
2
4
6
8
0
5

100

20
40
60
80
ll

0
10

0
S02243
S01524
NCIH1092
NCIH69
S01578
NCIH2196
CORL47
NCIH1436

20
NCIH510

20
SHP77
S02382
S00356
S02290
COLO668
NCIH1618

HCI-4
NCIH209
NCIH146

40
NCIH1930

40
CORL95
DMS153
DMS79
NCIH1836

**
ns
DMS454
S02241

ASCL1
S02284

60
NCIH2081
S02297

60

NEUROD1
S02234
CORL88

p = 0.0014
NCIH1963

p = 0.0652
NCIH1105
S01366
S01297
S02360

80
S02248

80
S02065
S01861
S00050
NCIH1184
Predicted NOTCH status:

NCIH2029
S00472
S02342
NOTCH2 REST S00022
Expression Expression S02397
NCIH889

0
2
4
6
8
0
1
2
3
4
5
S02354
NCIH2066

HCI-12
DMS53

0
0
S00832
S02242

ASCL1
S01512
S02294
NCIH1876
S01248

% late in SCLC-A
S02350

Low
Low
S00838

High
High
S02352

20
S01453

20
S02295
S02120
S02293
S00836

74 Cancer Cell 38, 60–78, July 13, 2020

S02289
S02376
S00501

40
S02353
S02351

40
S01728
wild-type or silent

S02285
S02378

****
S00827

****
S00035
S02163

60
S00825
S02287
S02322

60

p < 0.0001

HCI-3
S02244
S00830

p < 0.0001
S01792
S01242
S02328

80
S00831
S02249
S02194

80
S01698
S01864
S02347

MYC
S01563

MYCL
S00213
S02093
S02299
S02291
CORL24
non-damaging

S02139
HCC33
D

SCLC21H
SCLC22H
CORL279
NCIH2171
NCIH2227
NCIH1694
LU135
NCIH524
NCIH82
Expression Expression S01494
DMS273
NCIH446

HCI-20
S01873
NCIH1341
NCIH196
SW1271
HCI-27
damaging

NCIH841
ASCL1+

DMS114

HCI-30*
SBC5
HCI-11

S02298

Low
Low
NCIH1339

High
High
NCIH2286
NEUROD1 YAP1

S02246
HCI-4*
HCI-21

HCI-3*
HCI-39
MYC

HES1
REST
MYCL

NOTCH2

Day 3-5
Day 7-10

red = cell line

Day 19-21
Day 12-14

black = tumor

HCI-37
HCI-23

YAP1+
HCI-10
HCI-13* HCI-16 HCI-31
HCI-5* HCI-20
B

% Day 3-5 expression

H
HCI-14
HCI-6
0
20
40
60
80

HCI-12

tSNE 2 AS
HCI-36

N C
EU L1

-20
20
R
HCI-34*

0
HCI-26

O
D
1
****
**

YA
NEUROD1+

P1

-20
% Day 7-10 expression

n = 1021
0
10
20
30

AS
3/21

N C
1 subtype

EU L1
*

3 subtypes
2 subtypes

R
MYC+

O
D
1

0
YA

tSNE 1
P1
0%
14%

% Day 12-14 expression

0
10
20
30
40

AS
N C

20
EU L1
*

18/21

4
1 O
33% 1 subtype

D
**

3 subtypes
67% 2 subtypes

1
Cluster
MYC-
ns

5 YA
2
P1
0%
39%
61%
86%

6
3

% Day 19-21 expression

0
20
40
60
80

AS
N C
EU L1
R
O
D
1
****
****

YA
P1

(legend on next page)

Article
 ll
Article

subtypes have distinct therapeutic vulnerabilities, this suggests d EXPERIMENTAL MODEL AND SUBJECT DETAILS
that dynamically evolving tumors represent a ‘‘moving therapeu- B Human SCLC Samples
tic target,’’ adopting unique therapeutic vulnerabilities as they B Clinical Details
progress. Given many failed clinical trials with targeted therapies B Cell Lines
in SCLC, we speculate that the reason why chemotherapy (a B Mice
‘‘blunt instrument’’) has remained the most effective therapeutic d METHOD DETAILS
option is likely due to its non-specific cytotoxicity in the multiple B Mouse Lung Tumor Initiation
subtypes of SCLC that evolve during progression. B In Vivo DAPT and DBZ
As clinical trials begin to assess biomarkers of SCLC subtype B MicroCT Analysis and Quantification
and potentially enroll patients based on these subtypes (Owoni- B Immunohistochemistry
koko et al., 2019; Poirier et al., 2020), it will be critical to assess B Primary Tumor Cell Isolation
subtype heterogeneity and anticipate tumor evolution to other B Quantification of Tumor Cell Morphology
subtypes. Multiple studies have implicated an increase in MYC B Live/Dead Assays
and a decrease in MYCL and ASCL1 in chemotherapy-resistant B Immunoblotting
mouse and human SCLC (Chalishazar et al., 2019; Farago et al., B PCR for Recombination Efficiency
2019; Huang et al., 2018a; Mollaoglu et al., 2017; Stewart et al., B MYC Overexpression and Virus Production
2020; Wagner et al., 2018) and have correlated high MYC with B DAPT Treatments In Vitro
shorter patient survival and more aggressive, drug-resistant B ChIP-seq in RPM Tumors
phenotypes (Carney et al., 1985; Gazdar et al., 1985; Johnson B Whole Genome Sequencing (WGS)
et al., 1986). Recent studies also demonstrate an increase in in- B Mouse Tumor and Timepoint Bulk RNA-seq
tratumoral transcriptional heterogeneity in chemo-resistant pa- B Bulk RNA-seq Data Analysis
tient samples (Stewart et al., 2020). Together, these findings sug- B Human Genomics Data Analysis
gest the provocative notion that therapy selects for and/or B CIBERSORT
promotes the latest stages of tumor cell progression identified B Single Cell RNA-seq Sample and Library Prep
here. We speculate that cancers in other tissues may also harbor B 10X Single Cell RNA-seq Data Processing
cells in transcriptionally dynamic states of progression, and d QUANTIFICATION AND STATISTICAL ANALYSIS
would potentially benefit from either more blunt, combinatorial,
or plasticity-directed therapies. SUPPLEMENTAL INFORMATION

Supplemental Information can be found online at https://doi.org/10.1016/j.

STAR+METHODS ccell.2020.05.001.

Detailed methods are provided in the online version of this paper ACKNOWLEDGMENTS
and include the following:
We thank members of the Oliver Lab for technical assistance and mouse col-
d KEY RESOURCES TABLE ony management, R. Olsen, R. Dahlgren, and L. Houston for administrative
support, A. Andersen for editorial guidance, and B. Dalley, K. Rondem, and
d RESOURCE AVAILABILITY
C. Stubben for bioinformatics support. We are grateful to our late colleague
B Lead Contact Dr. Adi Gazdar for insightful discussions and inspiration for this work. We
B Materials Availability acknowledge support from the National Cancer Institute (NCI) of the NIH under
B Data and Code Availability award P30CA042014 to Huntsman Cancer Institute for the use of core facilities

Figure 8. Human SCLC Exhibits Intratumoral Molecular Subtype Heterogeneity

(A) CIBERSORT analysis of RPM transition signatures in 70 human SCLC tumors (George et al., 2015) (black ID) and 48 human SCLC cell lines from CCLE (red ID),
excluding POU2F3+ samples, grouped according to SCLC molecular subtype. Gene expression values for indicated MYC family member or Notch-related genes
overlaid above the stacked bar graphs. Predicted NOTCH-status marked by top color bar.
(B) Percentage of RPM time point signatures (y axis) within hSCLC molecular subtypes derived from (A). Mean ± SEM, Student’s two-tailed unpaired t test, ****p <
0.0001, **p < 0.007, *p < 0.02; n.s., not significant.
(C) Correlation of indicated genes (y axis) with percentage of late (days 19–21) signature determined by CIBERSORT within the human ASCL1+ subset only, with
Pearson correlation p values indicated.
(D) Venn diagram of human SCLC tissue IHC results with deidentified HCI patient number for positive samples for ASCL1 (red circle), NEUROD1 (yellow circle), or
YAP1 (blue circle) (n = 21 total). Samples in bold text harbor at least some cells with MYC+ IHC (HCI-12, -20, and -14). Table on the right summarizes the
percentage of samples with the indicated number of subtypes present at any frequency organized by MYC expression. * indicates tissue from limited stage as
opposed to extensive stage.
(E) Representative IHC in indicated patient biopsies. Serial sections were stained, but multiple tumor regions are shown to illustrate tumor heterogeneity. Scale
bar, 20 mm.
(F) tSNE unbiased clustering of tumor cell populations derived from chemotherapy-relapsed human SCLC liver biopsy analyzed by scRNA-seq. Number of post-
QC tumor cells indicated.
(G) Relative expression of the indicated genes in tSNE as in (F).
(H) We propose a hypothetical model whereby NOTCH-deficient SCLC tumors are locked in an NEstem-like (ASCL1+) state (right circle) similar to Ouadah et al.
(2019). In contrast, MYC reprograms NOTCH WT tumors to non-NE SCLC fates that proceed either from NEUROD1 to YAP1 (left circle, top dashed line), or
directly from ASCL1 to YAP1 (left circle, bottom dashed line).
See also Figure S7.

Cancer Cell 38, 60–78, July 13, 2020 75

 ll
and shared resources, including Biorepository and Molecular Pathology, High-
Throughput Genomics and Bioinformatic Analysis, and Flow Cytometry
Shared Resources; we acknowledge support and resources from the
Center for High Performance Computing at the University of Utah. T.G.O.
was supported in part by NIH NCI (U01-CA231844, U24-CA213274 and
R21-CA216504-01A1).
AUTHOR CONTRIBUTIONS
Study Design, A.S.I. and T.G.O.; Data Analysis and Acquisition, A.S.I., A.M.M.,
D.W.K., O.S.C., S.J.W., K.B.S., M.R.G., D.S., B.G., M.D.C., and J.G.; Bio-
informatic Analyses, A.S.I., C.C.C., Y.Q., X.H., S.T., S.D., G.M., B.T.S., and
J.G.; Pathology, B.L.W.; Clinical Annotation, S.P. and J.C.M.; Study Supervi-
sion and Funding, T.G.O.; Manuscript Preparation, A.S.I., B.T.S., and T.G.O.
DECLARATION OF INTERESTS
T.G.O. has pending patent applications related to subtype stratification of
SCLC: US16/335368, JP2019522392, and EP2017865057.
Received: December 22, 2019
Revised: March 23, 2020
Accepted: April 30, 2020
Published: May 28, 2020
REFERENCES
Angerer, P., Haghverdi, L., Buttner, M., Theis, F.J., Marr, C., and Buettner, F.
(2016). destiny: diffusion maps for large-scale single-cell data in R.
Bioinformatics 32, 1241–1243.
Augert, A., Eastwood, E., Ibrahim, A.H., Wu, N., Grunblatt, E., Basom, R.,
Liggitt, D., Eaton, K.D., Martins, R., Poirier, J.T., et al. (2019). Targeting
NOTCH activation in small cell lung cancer through LSD1 inhibition. Sci.
Signal. 12, https://doi.org/10.1126/scisignal.aau2922.
Borromeo, M.D., Savage, T.K., Kollipara, R.K., He, M., Augustyn, A., Osborne,
J.K., Girard, L., Minna, J.D., Gazdar, A.F., Cobb, M.H., and Johnson, J.E.
(2016). ASCL1 and NEUROD1 reveal heterogeneity in pulmonary neuroendo-
crine tumors and regulate distinct genetic programs. Cell Rep. 16, 1259–1272.
Bragelmann, J., Bohm, S., Guthrie, M.R., Mollaoglu, G., Oliver, T.G., and Sos,
M.L. (2017). Family matters: how MYC family oncogenes impact small cell lung
cancer. Cell Cycle 16, 1489–1498.
Butler, A., Hoffman, P., Smibert, P., Papalexi, E., and Satija, R. (2018).
Integrating single-cell transcriptomic data across different conditions, technol-
ogies, and species. Nat. Biotechnol. 36, 411–420.
Cardnell, R.J., Li, L., Sen, T., Bara, R., Tong, P., Fujimoto, J., Ireland, A.S.,
Guthrie, M.R., Bheddah, S., Banerjee, U., et al. (2017). Protein expression of
TTF1 and cMYC define distinct molecular subgroups of small cell lung cancer
with unique vulnerabilities to aurora kinase inhibition, DLL3 targeting, and
other targeted therapies. Oncotarget 8, 73419–73432.
Carney, D.N., Gazdar, A.F., Bepler, G., Guccion, J.G., Marangos, P.J., Moody,
T.W., Zweig, M.H., and Minna, J.D. (1985). Establishment and identification of
small cell lung cancer cell lines having classic and variant features. Cancer
Res. 45, 2913–2923.
Chalishazar, M.D., Wait, S.J., Huang, F., Ireland, A.S., Mukhopadhyay, A., Lee,
Y., Schuman, S.S., Guthrie, M.R., Berrett, K.C., Vahrenkamp, J.M., et al.
(2019). MYC-driven small-cell lung cancer is metabolically distinct and vulner-
able to arginine depletion. Clin. Cancer Res. 25, 5107–5121.
Christensen, C.L., Kwiatkowski, N., Abraham, B.J., Carretero, J., Al-Shahrour,
F., Zhang, T., Chipumuro, E., Herter-Sprie, G.S., Akbay, E.A., Altabef, A., et al.
(2014). Targeting transcriptional addictions in small cell lung cancer with a co-
valent CDK7 inhibitor. Cancer Cell 26, 909–922.
Cingolani, P., Platts, A., Wang le, L., Coon, M., Nguyen, T., Wang, L., Land,
S.J., Lu, X., and Ruden, D.M. (2012). A program for annotating and predicting
the effects of single nucleotide polymorphisms, SnpEff: SNPs in the genome of
Drosophila melanogaster strain w1118; iso-2; iso-3. Fly (Austin) 6, 80–92.
76 Cancer Cell 38, 60–78, July 13, 2020
Article
Coifman, R.R., Lafon, S., Lee, A.B., Maggioni, M., Nadler, B., Warner, F., and
Zucker, S.W. (2005). Geometric diffusions as a tool for harmonic analysis and
structure definition of data: multiscale methods. Proc. Natl. Acad. Sci. U S A
102, 7432–7437.
Collisson, E.A., Campbell, J.D., Brooks, A.N., Berger, A.H., Lee, W.,
Chmielecki, J., Beer, D.G., Cope, L., Creighton, C.J., Danilova, L., et al.
(2014). Comprehensive molecular profiling of lung adenocarcinoma. Nature
511, 543–550.
Dammert, M.A., Bragelmann, J., Olsen, R.R., Bohm, S., Monhasery, N.,
Whitney, C.P., Chalishazar, M.D., Tumbrink, H.L., Guthrie, M.R., Klein, S.,
et al. (2019). MYC paralog-dependent apoptotic priming orchestrates a spec-
trum of vulnerabilities in small cell lung cancer. Nat. Commun. 10, 3485.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Dooley, A.L., Winslow, M.M., Chiang, D.Y., Banerji, S., Stransky, N., Dayton,
T.L., Snyder, E.L., Senna, S., Whittaker, C.A., Bronson, R.T., et al. (2011).
Nuclear factor I/B is an oncogene in small cell lung cancer. Genes Dev. 25,
1470–1475.
DuPage, M., Dooley, A.L., and Jacks, T. (2009). Conditional mouse lung cancer
models using adenoviral or lentiviral delivery of Cre recombinase. Nat. Protoc.
4, 1064–1072.
Ewels, P., Magnusson, M., Lundin, S., and Kaller, M. (2016). MultiQC: summa-
rize analysis results for multiple tools and samples in a single report.
Bioinformatics 32, 3047–3048.
Farago, A.F., Yeap, B.Y., Stanzione, M., Hung, Y.P., Heist, R.S., Marcoux, J.P.,
Zhong, J., Rangachari, D., Barbie, D.A., Phat, S., et al. (2019). Combination
olaparib and temozolomide in relapsed small-cell lung cancer. Cancer
Discov. 9, 1372–1387.
Flowers, J.L., Burton, G.V., Cox, E.B., McCarty, K.S., Sr, ., Dent, G.A.,
Geisinger, K.R., and McCarty, K.S., Jr. (1986). Use of monoclonal antiestrogen
receptor antibody to evaluate estrogen receptor content in fine needle aspira-
tion breast biopsies. Ann. Surg. 203, 250–254.
Gazdar, A.F., Carney, D.N., Nau, M.M., and Minna, J.D. (1985).
Characterization of variant subclasses of cell lines derived from small cell
lung cancer having distinctive biochemical, morphological, and growth prop-
erties. Cancer Res. 45, 2924–2930.
George, J., Lim, J.S., Jang, S.J., Cun, Y., Ozretic, L., Kong, G., Leenders, F.,
Lu, X., Fernandez-Cuesta, L., Bosco, G., et al. (2015). Comprehensive
genomic profiles of small cell lung cancer. Nature 524, 47–53.
Giraddi, R.R., Chung, C.-Y., Heinz, R.E., Balcioglu, O., Novotny, M., Trejo,
C.L., Dravis, C., Hagos, B.M., Mehrabad, E.M., Rodewald, L.W., et al.
(2018). Single-cell transcriptomes distinguish stem cell state changes and line-
age specification programs in early mammary gland development. Cell Rep.
24, 1653–1666.e7.
Gratton, M.O., Torban, E., Jasmin, S.B., Theriault, F.M., German, M.S., and
Stifani, S. (2003). Hes6 promotes cortical neurogenesis and inhibits Hes1 tran-
scription repression activity by multiple mechanisms. Mol. Cell. Biol. 23,
6922–6935.
Horie, M., Saito, A., Ohshima, M., Suzuki, H.I., and Nagase, T. (2016). YAP and
TAZ modulate cell phenotype in a subset of small cell lung cancer. Cancer Sci.
107, 1755–1766.
Hovestadt, V., Smith, K.S., Bihannic, L., Filbin, M.G., Shaw, M.L.,
Baumgartner, A., DeWitt, J.C., Groves, A., Mayr, L., Weisman, H.R., et al.
(2019). Resolving medulloblastoma cellular architecture by single-cell geno-
mics. Nature 572, 74–79.
Huang, F., Ni, M., Chalishazar, M.D., Huffman, K.E., Kim, J., Cai, L., Shi, X.,
Cai, F., Zacharias, L.G., Ireland, A.S., et al. (2018a). Inosine monophosphate
dehydrogenase dependence in a subset of small cell lung cancers. Cell
Metab. 28, 369–382.e5.
Huang, Y.H., Klingbeil, O., He, X.Y., Wu, X.S., Arun, G., Lu, B., Somerville,
T.D.D., Milazzo, J.P., Wilkinson, J.E., Demerdash, O.E., et al. (2018b).
POU2F3 is a master regulator of a tuft cell-like variant of small cell lung cancer.
Genes Dev. 32, 915–928.

Article
Huijbers, I.J., Bin Ali, R., Pritchard, C., Cozijnsen, M., Kwon, M.C., Proost, N.,
Song, J.Y., de Vries, H., Badhai, J., Sutherland, K., et al. (2014). Rapid target
gene validation in complex cancer mouse models using re-derived embryonic
stem cells. EMBO Mol. Med. 6, 212–225.
Ito, T., Udaka, N., Ikeda, M., Yazawa, T., Kageyama, R., and Kitamura, H.
(2001). Significance of proneural basic helix-loop-helix transcription factors
in neuroendocrine differentiation of fetal lung epithelial cells and lung carci-
noma cells. Histol. Histopathol. 16, 335–343.
Jia, D., Augert, A., Kim, D.W., Eastwood, E., Wu, N., Ibrahim, A.H., Kim, K.B.,
Dunn, C.T., Pillai, S.P.S., Gazdar, A.F., et al. (2018). Crebbp loss drives small
cell lung cancer and increases sensitivity to HDAC inhibition. Cancer Discov.
8, 1422–1437.
Johnson, B.E., Battey, J., Linnoila, I., Becker, K.L., Makuch, R.W., Snider,
R.H., Carney, D.N., and Minna, J.D. (1986). Changes in the phenotype of hu-
man small cell lung cancer cell lines after transfection and expression of the
c-myc proto-oncogene. J. Clin. Invest. 78, 525–532.
Kim, D.W., Wu, N., Kim, Y.C., Cheng, P.F., Basom, R., Kim, D., Dunn, C.T.,
Lee, A.Y., Kim, K., Lee, C.S., et al. (2016). Genetic requirement for Mycl and
efficacy of RNA Pol I inhibition in mouse models of small cell lung cancer.
Genes Dev. 30, 1289–1299.
Kuleshov, M.V., Jones, M.R., Rouillard, A.D., Fernandez, N.F., Duan, Q.,
Wang, Z., Koplev, S., Jenkins, S.L., Jagodnik, K.M., Lachmann, A., et al.
(2016). Enrichr: a comprehensive gene set enrichment analysis web server
2016 update. Nucleic Acids Res. 44, W90–W97.
Li, H., and Durbin, R. (2009). Fast and accurate short read alignment with
Burrows-Wheeler transform. Bioinformatics 25, 1754–1760.
Liao, Y., Smyth, G.K., and Shi, W. (2014). featureCounts: an efficient general
purpose program for assigning sequence reads to genomic features.
Bioinformatics 30, 923–930.
Lim, J.S., Ibaseta, A., Fischer, M.M., Cancilla, B., O’Young, G., Cristea, S.,
Luca, V.C., Yang, D., Jahchan, N.S., Hamard, C., et al. (2017). Intratumoural
heterogeneity generated by Notch signalling promotes small-cell lung cancer.
Nature 545, 360–364.
Lin, J.J., Cardarella, S., Lydon, C.A., Dahlberg, S.E., Jackman, D.M., Janne,
P.A., and Johnson, B.E. (2016). Five-year survival in EGFR-mutant metastatic
lung adenocarcinoma treated with EGFR-TKIs. J. Thorac. Oncol. 11, 556–565.
Love, M.I., Huber, W., and Anders, S. (2014). Moderated estimation of fold
change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
Lun, A.T., Bach, K., and Marioni, J.C. (2016). Pooling across cells to normalize
single-cell RNA sequencing data with many zero counts. Genome Biol. 17, 75.
Martin, M. (2011). Cutadapt removes adapter sequences from high-
throughput sequencing reads. EMBnet J. 17.
Matsumoto, A., Onoyama, I., Sunabori, T., Kageyama, R., Okano, H., and
Nakayama, K.I. (2011). Fbxw7-dependent degradation of Notch is required
for control of "stemness" and neuronal-glial differentiation in neural stem cells.
J. Biol. Chem. 286, 13754–13764.
Mayr, C., and Bartel, D.P. (2009). Widespread shortening of 3’UTRs by alterna-
tive cleavage and polyadenylation activates oncogenes in cancer cells. Cell
138, 673–684.
McCarthy, D.J., Campbell, K.R., Lun, A.T., and Wills, Q.F. (2017). Scater: pre-
processing, quality control, normalization and visualization of single-cell RNA-
seq data in R. Bioinformatics 33, 1179–1186.
McColl, K., Wildey, G., Sakre, N., Lipka, M.B., Behtaj, M., Kresak, A., Chen, Y.,
Yang, M., Velcheti, V., Fu, P., and Dowlati, A. (2017). Reciprocal expression of
INSM1 and YAP1 defines subgroups in small cell lung cancer. Oncotarget 8,
73745–73756.
McFadden, D.G., Papagiannakopoulos, T., Taylor-Weiner, A., Stewart, C.,
Carter, S.L., Cibulskis, K., Bhutkar, A., McKenna, A., Dooley, A., Vernon, A.,
et al. (2014). Genetic and clonal dissection of murine small cell lung carcinoma
progression by genome sequencing. Cell 156, 1298–1311.
Meuwissen, R., Linn, S.C., Linnoila, R.I., Zevenhoven, J., Mooi, W.J., and
Berns, A. (2003). Induction of small cell lung cancer by somatic inactivation
of both Trp53 and Rb1 in a conditional mouse model. Cancer Cell 4, 181–189.
ll
Mollaoglu, G., Guthrie, M.R., Bohm, S., Bragelmann, J., Can, I., Ballieu, P.M.,
Marx, A., George, J., Heinen, C., Chalishazar, M.D., et al. (2017). MYC drives
progression of small cell lung cancer to a variant neuroendocrine subtype
with vulnerability to aurora kinase inhibition. Cancer Cell 31, 270–285.
Newman, A.M., Liu, C.L., Green, M.R., Gentles, A.J., Feng, W., Xu, Y., Hoang,
C.D., Diehn, M., and Alizadeh, A.A. (2015). Robust enumeration of cell subsets
from tissue expression profiles. Nat. Methods 12, 453–457.
Oliver, T.G., Meylan, E., Chang, G.P., Xue, W., Burke, J.R., Humpton, T.J.,
Hubbard, D., Bhutkar, A., and Jacks, T. (2011). Caspase-2-mediated cleavage
of Mdm2 creates a p53-induced positive feedback loop. Mol. Cell 43, 57–71.
Oser, M.G., Sabet, A.H., Gao, W., Chakraborty, A.A., Schinzel, A.C., Jennings,
R.B., Fonseca, R., Bonal, D.M., Booker, M.A., Flaifel, A., et al. (2019). The
KDM5A/RBP2 histone demethylase represses NOTCH signaling to sustain
neuroendocrine differentiation and promote small cell lung cancer tumorigen-
esis. Genes Dev. 33, 1718–1738.
Ouadah, Y., Rojas, E.R., Riordan, D.P., Capostagno, S., Kuo, C.S., and
Krasnow, M.A. (2019). Rare pulmonary neuroendocrine cells are stem cells
regulated by Rb, p53, and Notch. Cell 179, 403–416.e3.
Owonikoko, T.K., Niu, H., Nackaerts, K., Csoszi, T., Ostoros, G., Mark, Z., Baik,
C., Joy, A.A., Chouaid, C., Jaime, J.C., et al. (2019). Randomized phase II study
of paclitaxel plus alisertib versus paclitaxel plus placebo as second-line ther-
apy for small-cell lung cancer: primary and correlative biomarker analyses.
J. Thorac. Oncol. 19, 33572–33575.
Park, J.W., Lee, J.K., Phillips, J.W., Huang, P., Cheng, D., Huang, J., and Witte,
O.N. (2016). Prostate epithelial cell of origin determines cancer differentiation
state in an organoid transformation assay. Proc. Natl. Acad. Sci. U S A 113,
4482–4487.
Park, K.S., Liang, M.C., Raiser, D.M., Zamponi, R., Roach, R.R., Curtis, S.J.,
Walton, Z., Schaffer, B.E., Roake, C.M., Zmoos, A.F., et al. (2011).
Characterization of the cell of origin for small cell lung cancer. Cell Cycle 10,
2806–2815.
Patel, A.S., Yoo, S., Kongaelin, R.J., Sato, T., Fridrikh, M., Sinha, A.,
Nudelman, G., Powell, C., Beasley, M., Zhu, J., and Watanabe, H. (2019).
Integrative genomic and epigenomic analyses identify a distinct role of c-
Myc and L-Myc for lineage determination in small cell lung cancer. bioRxiv.
https://doi.org/10.1101/852939.
Peifer, M., Fernandez-Cuesta, L., Sos, M.L., George, J., Seidel, D., Kasper,
L.H., Plenker, D., Leenders, F., Sun, R., Zander, T., et al. (2012). Integrative
genome analyses identify key somatic driver mutations of small-cell lung can-
cer. Nat. Genet. 44, 1104–1110.
Platt, R.J., Chen, S., Zhou, Y., Yim, M.J., Swiech, L., Kempton, H.R., Dahlman,
J.E., Parnas, O., Eisenhaure, T.M., Jovanovic, M., et al. (2014). CRISPR-Cas9
knockin mice for genome editing and cancer modeling. Cell 159, 440–455.
Poirier, J.T., Gardner, E.E., Connis, N., Moreira, A.L., de Stanchina, E., Hann,
C.L., and Rudin, C.M. (2015). DNA methylation in small cell lung cancer defines
distinct disease subtypes and correlates with high expression of EZH2.
Oncogene 34, 5869–5878.
Poirier, J.T., George, J., Owonikoko, T.K., Berns, A., Brambilla, E., Byers, L.A.,
Carbone, D., Chen, H.J., Christensen, C.L., Dive, C., et al. (2020). New ap-
proaches to SCLC therapy: from the laboratory to the clinic. J. Thorac.
Oncol. 15, 520–540.
Qiu, X., Hill, A., Packer, J., Lin, D., Ma, Y.A., and Trapnell, C. (2017). Single-cell
mRNA quantification and differential analysis with Census. Nat. Methods 14,
309–315.
Remon, J., Ahn, M.J., Girard, N., Johnson, M., Kim, D.W., Lopes, G., Pillai,
R.N., Solomon, B., Villacampa, G., and Zhou, Q. (2019). Advanced-stage
non-small cell lung cancer: advances in thoracic oncology 2018. J. Thorac.
Oncol. 14, 1134–1155.
Rudin, C.M., Durinck, S., Stawiski, E.W., Poirier, J.T., Modrusan, Z., Shames,
D.S., Bergbower, E.A., Guan, Y., Shin, J., Guillory, J., et al. (2012).
Comprehensive genomic analysis identifies SOX2 as a frequently amplified
gene in small-cell lung cancer. Nat. Genet. 44, 1111–1116.
Rudin, C.M., Poirier, J.T., Byers, L.A., Dive, C., Dowlati, A., George, J.,
Heymach, J.V., Johnson, J.E., Lehman, J.M., MacPherson, D., et al. (2019).
Cancer Cell 38, 60–78, July 13, 2020 77
 ll
Molecular subtypes of small cell lung cancer: a synthesis of human and mouse
model data. Nat. Rev. Cancer 19, 289–297.
Saunders, L.R., Bankovich, A.J., Anderson, W.C., Aujay, M.A., Bheddah, S.,
Black, K., Desai, R., Escarpe, P.A., Hampl, J., Laysang, A., et al. (2015). A
DLL3-targeted antibody-drug conjugate eradicates high-grade pulmonary
neuroendocrine tumor-initiating cells in vivo. Sci. Transl. Med. 7, 302ra136.
Schaffer, B.E., Park, K.S., Yiu, G., Conklin, J.F., Lin, C., Burkhart, D.L.,
Karnezis, A.N., Sweet-Cordero, E.A., and Sage, J. (2010). Loss of p130 accel-
erates tumor development in a mouse model for human small-cell lung carci-
noma. Cancer Res. 70, 3877–3883.
Scialdone, A., Natarajan, K.N., Saraiva, L.R., Proserpio, V., Teichmann, S.A.,
Stegle, O., Marioni, J.C., and Buettner, F. (2015). Computational assignment
of cell-cycle stage from single-cell transcriptome data. Methods 85, 54–61.
Sen, T., Tong, P., Stewart, C.A., Cristea, S., Valliani, A., Shames, D.S.,
Redwood, A.B., Fan, Y.H., Li, L., Glisson, B.S., et al. (2017). CHK1 inhibition
in small-cell lung cancer produces single-agent activity in biomarker-defined
disease subsets and combination activity with cisplatin or olaparib. Cancer
Res. 77, 3870–3884.
Shen, R., and Seshan, V.E. (2016). FACETS: allele-specific copy number and
clonal heterogeneity analysis tool for high-throughput DNA sequencing.
Nucleic Acids Res. 44, e131.
Simpson, K.L., Stoney, R., Frese, K.K., Simms, N., Rowe, W., Pearce, S.P.,
Humphrey, S., Booth, L., Morgan, D., Dynowski, M., et al. (2020). A biobank
of small cell lung cancer CDX models elucidates inter- and intratumoral pheno-
typic heterogeneity. Nat. Cancer 1, 437–451.
Stewart, C.A., Gay, C.M., Xi, Y., Sivajothi, S., Sivakamasundari, V., Fujimoto,
J., Bolisetty, M., Hartsfield, P.M., Balasubramaniyan, V., Chalishazar, M.D.,
et al. (2020). Single-cell analyses reveal increased intratumoral heterogeneity
after the onset of therapy resistance in small-cell lung cancer. Nat. Cancer
1, 423–436.
Stewart, S.A., Dykxhoorn, D.M., Palliser, D., Mizuno, H., Yu, E.Y., An, D.S.,
Sabatini, D.M., Chen, I.S., Hahn, W.C., Sharp, P.A., et al. (2003). Lentivirus-
delivered stable gene silencing by RNAi in primary cells. RNA 9, 493–501.
Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck, W.M.,
3rd, Hao, Y., Stoeckius, M., Smibert, P., and Satija, R. (2019). Comprehensive
integration of single-cell data. Cell 177, 1888–1902.e1.
Sutherland, K.D., Proost, N., Brouns, I., Adriaensen, D., Song, J.Y., and Berns,
A. (2011). Cell of origin of small cell lung cancer: inactivation of Trp53 and Rb1
in distinct cell types of adult mouse lung. Cancer Cell 19, 754–764.
78 Cancer Cell 38, 60–78, July 13, 2020
Article
Tlemsani, C., Pongor, L., Girard, L., Roper, N., Elloumi, F., Varma, S., Luna, A.,
Rajapakse, V.N., Sebastian, R., Kohn, K.W., et al. (2020). SCLC_CellMiner: in-
tegrated genomics and therapeutics predictors of small cell lung cancer cell
lines based on their genomic signatures. bioRxiv. https://doi.org/10.1101/
2020.03.09.980623.
Trapnell, C., Cacchiarelli, D., Grimsby, J., Pokharel, P., Li, S., Morse, M.,
Lennon, N.J., Livak, K.J., Mikkelsen, T.S., and Rinn, J.L. (2014). The dynamics
and regulators of cell fate decisions are revealed by pseudotemporal ordering
of single cells. Nat. Biotechnol. 32, 381–386.
Treutlein, B., Brownfield, D.G., Wu, A.R., Neff, N.F., Mantalas, G.L., Espinoza,
F.H., Desai, T.J., Krasnow, M.A., and Quake, S.R. (2014). Reconstructing line-
age hierarchies of the distal lung epithelium using single-cell RNA-seq. Nature
509, 371–375.
Vo, B.T., Wolf, E., Kawauchi, D., Gebhardt, A., Rehg, J.E., Finkelstein, D., Walz,
S., Murphy, B.L., Youn, Y.H., Han, Y.G., et al. (2016). The interaction of myc
with Miz1 defines medulloblastoma subgroup identity. Cancer Cell 29, 5–16.
Wagner, A.H., Devarakonda, S., Skidmore, Z.L., Krysiak, K., Ramu, A., Trani,
L., Kunisaki, J., Masood, A., Waqar, S.N., Spies, N.C., et al. (2018).
Recurrent WNT pathway alterations are frequent in relapsed small cell lung
cancer. Nat. Commun. 9, 3787.
Wang, Z.A., Mitrofanova, A., Bergren, S.K., Abate-Shen, C., Cardiff, R.D.,
Califano, A., and Shen, M.M. (2013). Lineage analysis of basal epithelial cells
reveals their unexpected plasticity and supports a cell-of-origin model for
prostate cancer heterogeneity. Nat. Cell Biol. 15, 274–283.
Wooten, D.J., Groves, S.M., Tyson, D.R., Liu, Q., Lim, J.S., Albert, R., Lopez,
C.F., Sage, J., and Quaranta, V. (2019). Systems-level network modeling of
small cell lung cancer subtypes identifies master regulators and destabilizers.
PLoS Comput. Biol. 15, e1007343.
Xie, T., Wang, Y., Deng, N., Huang, G., Taghavifar, F., Geng, Y., Liu, N., Kulur,
V., Yao, C., Chen, P., et al. (2018). Single-cell deconvolution of fibroblast het-
erogeneity in mouse pulmonary fibrosis. Cell Rep. 22, 3625–3640.
Zhang, W., Girard, L., Zhang, Y.A., Haruki, T., Papari-Zareei, M., Stastny, V.,
Ghayee, H.K., Pacak, K., Oliver, T.G., Minna, J.D., and Gazdar, A.F. (2018).
Small cell lung cancer tumors and preclinical models display heterogeneity
of neuroendocrine phenotypes. Transl. Lung Cancer Res. 7, 32–49.
Zilionis, R., Engblom, C., Pfirschke, C., Savova, V., Zemmour, D., Saatcioglu,
H.D., Krishnan, I., Maroni, G., Meyerovitz, C.V., Kerwin, C.M., et al. (2019).
Single-cell transcriptomics of human and mouse lung cancers reveals
conserved myeloid populations across individuals and species. Immunity
50, 1317–1334.e10.
Cancer Cell

Article

A Probabilistic Classification Tool

for Genetic Subtypes of Diffuse Large B Cell
Lymphoma with Therapeutic Implications
George W. Wright,1 Da Wei Huang,2 James D. Phelan,2 Zana A. Coulibaly,2 Sandrine Roulland,2 Ryan M. Young,2
James Q. Wang,2 Roland Schmitz,2 Ryan D. Morin,3 Jeffrey Tang,3 Aixiang Jiang,3 Aleksander Bagaev,4 Olga Plotnikova,4
Nikita Kotlov,4 Calvin A. Johnson,5 Wyndham H. Wilson,2 David W. Scott,6 and Louis M. Staudt2,7,*
1Biometric Research Branch, Division of Cancer Diagnosis and Treatment, National Cancer Institute, National Institutes of Health, Bethesda,

MD, USA
2Lymphoid Malignancies Branch, National Cancer Institute, National Institutes of Health, Bethesda, MD 20892, USA
3Department of Molecular Biology and Biochemistry, Simon Fraser University, Burnaby, BC V5A 1S6, Canada
4BostonGene, Waltham, MA 02453, USA
5Office of Intramural Research, Center for Information Technology, National Institutes of Health, Bethesda, MD 20892, USA
6British Columbia Cancer, Vancouver, BC V5Z 4E6, Canada
7Lead Contact

*Correspondence: lstaudt@mail.nih.gov
https://doi.org/10.1016/j.ccell.2020.03.015

SUMMARY

The development of precision medicine approaches for diffuse large B cell lymphoma (DLBCL) is confounded
by its pronounced genetic, phenotypic, and clinical heterogeneity. Recent multiplatform genomic studies re-
vealed the existence of genetic subtypes of DLBCL using clustering methodologies. Here, we describe an
algorithm that determines the probability that a patient’s lymphoma belongs to one of seven genetic
subtypes based on its genetic features. This classification reveals genetic similarities between these DLBCL
subtypes and various indolent and extranodal lymphoma types, suggesting a shared pathogenesis. These
genetic subtypes also have distinct gene expression profiles, immune microenvironments, and outcomes
following immunochemotherapy. Functional analysis of genetic subtype models highlights distinct vulnera-
bilities to targeted therapy, supporting the use of this classification in precision medicine trials.

INTRODUCTION
Initial progress toward a molecular diagnosis of DLBCL subtypes
came with the advent of gene expression profiling, which was
used to define two prominent ‘‘cell-of-origin’’ (COO) subtypes
comprising 80%–85% of cases, termed germinal center B cell-
like (GCB) and activated B cell-like (ABC), with the remaining
cases declared ‘‘unclassified’’ (Alizadeh et al., 2000; Rosenwald
et al., 2002). This classification accounted for some of the het-
erogeneity in the clinical outcome following R-CHOP (rituximab
plus cyclophosphamide, doxorubicin, vincristine, and predni-
sone) chemotherapy (Alizadeh et al., 2000; Lenz et al., 2008a;
Rosenwald et al., 2002). This COO methodology has proved use-
ful in understanding the varied responses of patients with diffuse
large B cell lymphoma (DLBCL) to targeted therapies such as
ibrutinib, an inhibitor of B cell receptor (BCR) signaling (Wilson
et al., 2015b). Nonetheless, the COO distinction does not fully
account for the heterogeneous responses and outcomes
following either R-CHOP therapy or targeted therapy. This is
likely because gene expression profiling provides a phenotypic
description of cancers rather than a genetic description that en-
compasses tumor pathogenesis more directly.
While recurrent genetic aberrations in individual genes have
elucidated oncogenic mechanisms in DLBCL, progress toward

Significance

We describe a taxonomy of diffuse large B cell lymphoma (DLBCL) consisting of seven genetic subtypes and provide a prob-
abilistic method to classify a patient’s tumor within this taxonomy. Several DLBCL subtypes are genetically related to
distinct indolent lymphoma types, suggesting that these subtypes may arise from clinically occult precursors. We infer
oncogenic pathway activity and therapeutic vulnerabilities of the DLBCL subtypes using their genetic and gene expression
profiles supplemented by functional identification of essential genes using loss-of-function CRISPR/Cas9 screens of cell-
line models of the subtypes. To foster precision medicine studies of DLBCL, we have implemented our probabilistic algo-
rithm on a publicly accessible server.

Cancer Cell 37, 551–568, April 13, 2020 Published by Elsevier Inc. 551
a genetic classification of DLBCL tumors required the integration
of genomic data from multiple analytic platforms to identify
genes that were recurrently altered by mutations, translocations,
and/or copy-number alterations (Chapuy et al., 2018; Schmitz
et al., 2018). Mathematically distinct clustering methods were
used to assort DLBCL tumors into genetic subtypes that are
characterized by genomic aberrations in subtype-specific hall-
mark genes. The potential clinical utility of this genetic classifica-
tion was evident by the association of the subtypes with
outcome following R-CHOP therapy.
Many clustering methods are limited by the necessity to place
a tumor sample into no more than one subtype and by the fact
that the subtype assignment of a particular tumor can vary
when different tumors are included during the clustering pro-
cess. Such methods are therefore not appropriate in clinical
settings where molecular diagnoses are required in real time
for individual tumors. We have therefore developed an algorithm
to classify an individual patient’s tumor based on the probability
of belonging to a particular genetic subtype, allowing for the pos-
sibility that the tumor may have acquired more than one genetic
program during its evolution.
RESULTS
Development of the LymphGen Genetic Subtype
Classifier
We created the LymphGen algorithm to provide a probabilistic
classification of a tumor from an individual patient into a genetic
subtype. We define a genetic subtype as a group of tumors that
is enriched for genetic aberrations in a set of subtype predictor
genes. These subtype predictor genes are identified by consid-
ering each possible combination of genetic aberrations (i.e., mu-
tations, copy-number alterations, or fusions) as a separate
genetic ‘‘feature’’ and scoring a tumor as positive for a feature
if one or more of its genetic aberrations is observed. LymphGen
uses the presence or absence of each subtype predictor feature
to provide a probability that a tumor belongs to the subtype.
To implement LymphGen in DLBCL, we first needed to define
the genetic subtypes to which a tumor could be assigned. For
this subtype discovery effort, we used the GenClass algorithm
(Schmitz et al., 2018), which begins with several ‘‘seed’’ tumor
subsets and iteratively sorts tumors into and out of these seeds,
converging on a classification that maximizes the genetic
distinctiveness of the resulting subtypes (Figure 1A). First, we
chose seeds representing the four previously identified genetic
subtypes: MCD (including MYD88L265P and CD79B mutations),
BN2 (including BCL6 translocations and NOTCH2 mutations),
N1 (including NOTCH1 mutations), and EZB (including EZH2 mu-
tations and BCL2 translocations). Among the remaining cases in
our cohort (hereafter ‘‘NCI cohort’’ Schmitz et al., 2018), TP53
was the most frequently mutated gene (25.2%) that was not
also significantly enriched in one of the previous subtypes.
TP53 inactivation has been previously associated with aneu-
ploidy in DLBCL (Bea et al., 2004; Chapuy et al., 2018; Monti
et al., 2012). In the NCI cohort, tumors with a homozygous
TP53 deletion (5.9%) or the combination of a heterozygous
TP53 deletion and a TP53 mutation (8.7%) had the most aneu-
ploidy, as assessed by the number of gains and losses of chro-
mosomal segments. We therefore formed a seed class of cases
552 Cancer Cell 37, 551–568, April 13, 2020
with these TP53 features, which we term ‘‘A53’’ (aneuploid with
TP53 inactivation). We also observed that mutations in TET2,
P2RY8, and SGK1 were recurrently mutated among the geneti-
cally unassigned cases (10.1%, 6.9%, and 6.9% of cases,
respectively), with the majority (54%) of SGK1 mutations trun-
cating the protein. TET2 mutations were significantly associated
with truncating SGK1 mutations (p = 0.001) and with P2RY8 mu-
tations (p = 0.033), leading us to create a second new seed class
based on these features, which we term ‘‘ST2’’ (SGK1 and TET2
mutated). Using the six seed classes defined above, the
GenClass algorithm assigned 54% of the cases.
LymphGen next develops a separate Bayesian predictor
model for each GenClass subtype, which determines the prob-
ability that a tumor belongs to the subtype based on its genetic
features (Figure 1A). The algorithm defines subtype predictor
features that distinguish the subtype from all other cases
(p % 0.001, Fisher’s exact test, prevalence >0.2) and uses
the prevalence of the feature in the subtype and its prevalence
in other cases to estimate the likelihood that a tumor with that
feature belongs to the subtype. These likelihood estimates are
then used in Bayes formula to calculate the probability that an
individual tumor belongs to a subtype based on its constella-
tion of genetic features. Thus, for each DLBCL tumor, Lymph-
Gen calculates six probabilities, one for each GenClass-defined
subtype. We defined tumors with subtype probabilities
of >90% or 50%–90% as ‘‘core’’ or ‘‘extended’’ subtype mem-
bers, respectively. Tumors that were core members of more
than one subtype were termed ‘‘genetically composite’’ (Fig-
ure S1). In the NCI cohort, the LymphGen algorithm identified
47.6% core cases, 9.8% extended cases, and 5.7% genetically
composite cases (Figure 1B). Altogether 329 (63.1%) of the 574
cases in the NCI cohort were classified, which is substantially
greater than the 46.6% classified previously (Schmitz et al.,
2018) (Figures 1B and 1C). The inability of LymphGen to clas-
sify the remaining cases stemmed from three issues: some tu-
mors had a few features characteristic of one or more subtype
but not enough to be classified, some had unique features that
were not recurrent in DLBCL, and others had very few genetic
features at all.
In the resulting genetic taxonomy, each of the DLBCL COO
gene expression subgroups included multiple genetic subtypes,
with ABC tumors enriched for MCD, GCB tumors enriched for
EZB and ST2, and unclassified tumors enriched for BN2 (Fig-
ure 1D). Conversely, some genetic subtypes were largely
composed of tumors belonging to the same gene expression
subgroup (MCD, N1, EZB), while others comprised different
gene expression subgroups, with BN2, A53, and ST2 being the
most phenotypically diverse.
Genetic Attributes of DLBCL Subtypes
To display the genetic composition of the subtypes, we selected
a set of genetic features that were significantly associated with a
subtype (p % 0.01) and were present in >10% of the subtype
(Figures 2 and S2A). Many subtype-defining mutations are likely
due to activation-induced deaminase-dependent somatic hy-
permutation (Schmitz et al., 2018), which in many cases
produced truncating mutations in subtype-specific tumor-sup-
pressor genes (e.g., PRDM1, ETV6, TOX, HLA-A, HLA-B,
HLA-C in MCD, TNFRSF14 in EZB, and NFKBIA in ST2).
 A

B C D

Figure 1. Development of the LymphGen Classifier

(A) Cancer subtype discovery and prediction using the LymphGen algorithm. Shown at left is the discovery of cancer subtypes starting with ‘‘seed’’ sets of cases
using the GenClass algorithm (Schmitz et al., 2018). The LymphGen algorithm uses prevalences of genetic features to estimate the likelihood that a feature is
associated with a subtype and combines these likelihoods to calculate a probability that a tumor belongs to a genetic subtype. The example shows features
associated with the EZB subtype as present (‘‘1’’) or absent (‘‘0’’) in an individual tumor sample, and the likelihoods that the tumor is EZB or non-EZB based on
each feature. The final panel illustrates how LymphGen assigns a tumor using the subtype probabilities.
(B) Frequency of cases with high probability (‘‘Core’’) or moderate probability (‘‘Extended’’) subtype assignments, genetically composite cases, and unassigned
(Other) cases.
(C) Prevalence of various genetic subtypes.
(D) Top: prevalence of subtypes in DLBCL COO subgroups. Bottom: prevalence of COO subgroups within each genetic subtype.
See also Figure S1.

MYD88L265P and CD79B mutations, the genetic hallmarks of
MCD, cooperatively activate nuclear factor kB (NF-kB) via the
My-T-BCR supercomplex involving MYD88, TLR9, and the
BCR (Phelan et al., 2018). MCD tumors also frequently delete
the CDKN2A tumor-suppressor locus, encoding the cell-cycle
inhibitors p16INK4A and p15INK4B as well as p19ARF, which stabi-
lizes p53. The viability of MCD cells is likely sustained by BCL2,
which is upregulated epigenetically and by copy number gain/
amplification (Figure S2B). Another prominent theme in MCD tu-
mors is immune evasion (see below).
BN2 is characterized by mutations that activate NOTCH2 or
inactivate the NOTCH antagonist SPEN in 50% of tumors,
72% of which also have a BCL6 translocation. Mutations target-
ing components of the BCR-dependent NF-kB pathway
(PRKCB, BCL10, TNFAIP3, TNIP1) are also prominent in BN2,
suggesting that these tumors rely on this pathway for survival. In-
teractions with T and natural killer (NK) cells are potentially
compromised in BN2 by CD70 deletions. CCND3 mutations
likely foster vigorous proliferation in BN2, as in Burkitt’s lym-
phoma (BL) (Schmitz et al., 2012).
Epigenetic dysregulation is a defining attribute of EZB due to
inactivation of several epigenetic regulators (KMT2D, CREBBP,
EP300, ARID1A, IRF8, MEF2B, EBF1) and mutational activation
of EZH2 (Mlynarczyk et al., 2019; Pasqualucci and Dalla-Favera,

Cancer Cell 37, 551–568, April 13, 2020 553

 Prevalence (%) Prevalence (%)
Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 MCD Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 ST2
MYD88 L265P Mut -31.0 66.2 6.5 0 0 0 21.1 TET2 Mut -8.5 3.8 6.9 0 1.3 48.1 0
CD79B Mut -17.5 50.0 10.9 0 0 0 18.4 SGK1 Trunc -11.6 0 1.1 0 1.3 37.0 2.6
PIM1 Mut -34.4 92.5 26.1 18.8 13.2 40.7 23.7 DUSP2 Mut -5.2 18.8 8.7 6.2 5.3 44.4 7.9
HLA-B Mut HL HD -19.5 73.8 33.3 6.2 10.0 38.5 18.4 ZFP36L1 Mut -6.1 15.0 7.6 18.8 3.9 40.7 0
BTG1 Mut HL HD -16.8 70.0 24.4 31.2 11.4 38.5 28.9 ACTG1 Mut -6.1 6.3 6.9 0 6.6 37.0 0
CDKN2A Mut HD -12.6 62.0 28.9 25 11.4 7.7 15.8 ACTB Mut -4.0 8.9 10.3 6.2 10.5 33.3 2.7
ETV6 Mut HL HD -14.9 55.0 5.6 37.5 4.3 7.7 34.2 ITPKB Mut -4.9 2.5 8.7 0 3.9 33.3 2.6
SPIB Gain Amp -12.8 51.9 14.4 12.5 0 3.8 26.3 NFKBIA Mut -7.0 0 0 0 6.6 33.3 0
OSBPL10 Mut -15.5 51.2 21.7 12.5 6.6 25.9 5.3 STAT3 Mut -2.7 8.8 10.9 12.5 14.5 29.6 7.9
TOX Trunc HL HD -14.1 48.1 10.0 6.2 12.9 7.7 10.5 EIF4A2 Mut -6.1 6.2 2.2 6.2 0 29.6 0
BCL2 Gain Amp -3.0 48.1 16.7 43.8 11.4 7.7 47.4 JUNB Mut -7.5 2.5 0 0 0 29.6 0
BTG2 Mut -5.4 43.8 33.7 0 10.5 37.0 18.4 BCL2L1 Mut -7.1 2.5 1.1 0 0 25.9 0
MPEG1 Mut -15.3 43.8 6.5 0 1.3 11.1 5.3 DDX3X Mut -3.9 0 3.3 0 9.2 25.9 0
HLA-A Mut HL HD -6.9 43.0 23.3 25 7.0 23.1 18.4 SOCS1 Trunc -4.4 1.2 6.5 0 6.6 25.9 0
HLA-C Mut HL HD -8.3 42.5 19.5 6.2 8.6 23.1 13.2 CD83 Mut -5.8 2.5 0 0 1.3 25.9 0
SETD1B Mut HL HD -10.1 41.8 8.8 6.2 7.1 7.7 21.1 P2RY8 Mut -2.1 1.2 5.5 0 6.6 22.2 7.9
KLHL14 Mut -17.3 41.2 6.5 0 5.3 11.1 2.6 RFTN1 Mut -3.1 5.0 2.2 0 6.6 22.2 0
TBL1XR1 Mut -8.6 35.0 5.4 43.8 7.9 3.7 7.9 RAC2 Mut -3.8 2.5 0 6.2 2.6 18.5 0
GRHPR Mut -15.6 33.8 2.2 0 0 11.1 5.3 XBP1 Mut -3.9 3.8 1.1 0 1.3 18.5 2.7
PRDM1 Mut -7.8 32.5 8.7 0 2.6 7.4 10.5 SEC24C Mut -4.7 1.3 1.1 0 0 18.5 0
CD58 Trunc HD -6.7 31.6 10.0 12.5 4.2 15.4 7.9 MED16 Mut -4.1 1.3 0 0 3.9 18.5 0
TAP1 Mut HL HD -3.8 26.6 11.6 18.8 4.3 11.5 18.4 PRRC2C Mut -3.4 1.3 2.3 0 2.6 18.5 5.4
PIM2 Mut -7.2 25.0 3.3 0 1.3 11.1 10.5 EDRF1 Mut -5.0 1.3 0 0 1.3 18.5 0
FOXC1 Mut -7.6 21.2 1.1 0 3.9 7.4 0 DOCK8 Mut -3.7 0 0 6.2 1.3 18.5 2.7
IRF4 Mut -2.9 20.0 1.1 18.8 2.6 18.5 13.2 CLTC Mut -3.3 0 1.1 0 2.6 14.8 2.7
VMP1 Mut -2.3 16.2 6.5 12.5 2.6 18.5 7.9 ZNF516 Mut -3.5 1.3 0 0 0 14.8 0
SLC1A5 Mut Amp -5.3 15.4 2.3 6.2 0 0 2.7 WDR24 Mut -3.0 1.2 0 0 1.3 14.8 0 Genetic alteration
DAZAP1 Mut Amp -2.5 15.2 5.5 6.2 2.9 0 7.9 ZC3H12D Mut -3.1 0 1.1 0 2.6 14.8 0
BCL11A Mut -5.2 15.0 1.1 0 2.6 7.4 0 Genetic subtype Missense
PPP1R9B Mut -3.6 12.7 2.3 0 2.6 11.1 0 mutation
IL10RA Mut -4.4 12.5 0 6.2 0 0 2.6 Gene expression subgroup
IL16 Mut -5.1 11.4 1.1 0 1.3 0 0 Truncating
CHST2 Mut -6.1 11.4 0 0 0 3.7 0 mutation
ARID5B Mut -3.3 11.4 4.6 0 2.6 3.7 0
PDCD1LG2 / CD274 Fusion -4.5 11.3 1.1 0 3.9 0 0 Prevalence (%) Inframe
WEE1
KLHL42
Mut
Mut
-2.9
-4.3
11.3
10.1
4.3
1.1
0
0
2.6
0
14.8
0
0
0 Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 A53 mutation
TNRC18 Mut -2.1 10.1 5.7 0 3.9 3.7 0 TP53 Mut HL HD -11.8 22.8 24.4 50.0 37.5 26.9 86.8 Fusion
Chrom. 6q HL HD -10.3 29.1 22.2 0 11.4 7.7 65.8
Genetic subtype Chrom. 17p HL HD -10.0 16.5 4.4 18.8 17.1 3.8 60.5 Amplification
Gene expression subgroup Chrom. 8p HL HD -9.8 6.3 6.7 6.2 4.3 19.2 50.0
Chrom. 7p Gain Amp -2.9 17.7 12.2 12.5 45.7 19.2 47.4
Chrom. 3q Gain Amp -2.1 29.1 12.2 31.2 4.3 0 42.1 Gain
Prevalence (%)
Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 BN2 Chrom. 1p
Chrom. 7q
Chrom. 24p
HL HD
Gain Amp
HL HD
-8.1
-2.2
-2.6
3.8
24.1
15.2
2.2
12.2
14.4
6.2
0
0
5.7
32.9
15.7
15.4
15.4
19.2
39.5
39.5
36.8
Homozygous
deletion
BCL6 Fusion -39.4 8.8 72.8 0 7.9 3.7 7.9 B2M Trunc HD -2.8 3.8 20.0 6.2 14.3 26.9 34.2
NOTCH2 Trunc Amp -25.7 0 41.8 0 1.4 3.8 2.6 Chrom. 24q HL HD -2.2 15.2 13.3 0 15.7 23.1 34.2 Heterozygous
TNFAIP3
DTX1
Trunc HD
Mut
-15.7
-11.8
5.1
28.8
51.6
50.0
43.8 15.5 26.9 5.3
25.0 10.5 14.8 2.6
Chrom. 2q HL HD -13.5 0 0 0 0 0 31.6 deletion
Chrom. 4p HL HD -6.5 1.3 2.2 0 14.3 3.8 31.6
CD70 Mut -23.9 2.5 41.3 0 2.6 7.4 2.6 Chrom. 11q Gain Amp -2.2 13.9 8.9 6.2 20.0 3.8 31.6 Genetic subtype
BCL10 Mut Gain Amp -15.1 2.5 39.6 6.2 2.9 3.8 5.3 Chrom. 15q HL HD -3.9 1.3 6.7 0 18.6 7.7 28.9
UBE2A Mut -13.7 5.0 30.4 6.2 1.3 3.7 5.3
TMEM30A Trunc HL HD -5.3 10.1 26.7 0 4.3 0 7.9
Chrom. 16q
TP53BP1
HL HD
Mut HD
-6.2
-5.2
1.3
2.6
3.3
2.3
0
0
4.3
5.6
0
19.2
28.9
27.0
Core Extended
KLF2 Mut -3.8 12.5 21.7 0 2.6 18.5 0 Chrom. 2p HL HD -10.3 0 0 0 1.4 0 26.3 MCD
SPEN Trunc -8.9 2.5 21.7 0 0 7.4 0 Chrom. 4q HL HD -5.0 0 3.3 0 14.3 3.8 26.3
CCND3 Mut -2.2 7.5 18.5 18.8 7.9 7.4 5.3 Chrom. 22q HL HD -7.6 2.5 1.1 0 4.3 3.8 26.3 BN2
NOL9 Mut -5.0 5.1 18.4 6.2 0 11.1 2.7
TP63 Mut HL HD -3.4 7.6 17.6 6.2 1.4 19.2 2.6
CNPY3
Chrom. 8q
Amp
HL HD
-7.2
-6.3
1.3
1.3
2.2
1.1
0
0
1.4
2.9
0
3.8
23.7
23.7
N1
ETS1
HIST1H1D
Mut
Mut
-4.5
-3.2
10.0
10.0
17.4
16.3
6.2 6.6 7.4
0 7.9 7.4 2.6
0 Chrom. 19p HL HD -6.3 1.3 1.1 0 5.7 0 23.7 EZB
Chrom. 20p HL HD -10.8 0 0 0 0 0 23.7
PRKCB Mut Gain Amp -2.7 2.5 15.6 12.5 4.3 11.1 10.5 Chrom. 6p HL HD -6.9 1.3 0 0 1.4 3.8 21.1 ST2
HIST1H2BK Mut -3.2 12.5 14.1 0 2.6 0 2.6
TRIP12 Mut -3.3 3.8 11.5 0 2.6 7.4 2.7
Chrom. 9q
Chrom. 12p
HL HD
HL HD
-5.6
-7.4
2.5
1.3
2.2
0
0
0
2.9
0
0
0
21.1
21.1
A53
KLHL21 Mut -3.7 3.8 11.5 0 1.3 0 2.7 Chrom. 3p HL HD -4.5 0 1.1 0 5.7 3.8 18.4 Gene expression
TRRAP Mut -2.4 2.5 10.3 0 1.3 7.4 2.7 Chrom. 11p HL HD -4.9 3.8 1.1 0 2.9 0 18.4
PABPC1 Mut -2.6 0 10.3 6.2 3.9 3.7 5.4 Chrom. 12q HL HD -8.3 0 0 0 0 0 18.4 subgroup
Genetic subtype Chrom. 14q HL HD -4.5 2.5 1.1 0 2.9 7.7 18.4
Gene expression subgroup
Chrom. 21p
ING1
HL HD
HL HD
-2.4
-2.0
8.9
2.5
2.2
2.2
18.8
0
1.4
14.3
0
7.7
18.4
15.8
ABC
NFKBIZ Amp -4.0 2.5 0 0 0 0 15.8
Prevalence (%) Chrom. 1q HL HD -5.2 1.3 0 0 1.4 0 15.8 GCB
Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 EZB Chrom. 9p
Chrom. 10p
HL HD
HL HD
-3.8
-2.5
2.5
3.8
0
2.2
0
12.5
5.7
2.9
3.8
7.7
15.8
15.8 Unclassified
BCL2 Fusion -51.4 0 0 0 68.4 0 0 Chrom. 10q HL HD -3.6 1.3 0 0 1.4 7.7 15.8
EZH2 Mut -24.5 0 3.3 0 44.7 0 2.6 Chrom. 13q HL HD -3.1 2.5 0 0 10.0 3.8 15.8
TNFRSF14 Mut HD -27.9 0 16.7 6.2 66.2 19.2 5.3 Chrom. 21q HL HD -3.1 3.8 1.1 12.5 0 0 15.8
KMT2D Mut -5.2 31.2 27.2 18.8 53.9 14.8 34.2 TP73 Trunc HD -2.2 1.3 3.3 0 7.1 0 13.2
CREBBP Mut HL HD -13.0 7.6 22.2 0 52.7 15.4 13.2 Chrom. 16p HL HD -5.1 0 0 0 0 0 13.2
REL Amp -11.7 0 3.3 6.2 34.3 11.5 7.9 Chrom. 18q HL HD -3.9 1.3 0 0 0 0 13.2
FAS Mut HL HD -2.7 12.7 20.0 6.2 30.1 11.5 10.5 Chrom. 20q HL HD -4.2 2.5 0 0 0 0 13.2
IRF8 Mut -5.8 8.8 9.8 6.2 28.9 14.8 2.6 Chrom. 23p Amp -2.2 3.8 3.3 0 1.4 7.7 13.2
EP300 Mut HD -7.4 7.6 5.5 0 27.8 3.8 7.9 Chrom. 19q HL HD -2.7 0 0 0 4.3 3.8 10.5
Chrom. 12p Gain Amp -2.4 13.9 15.6 12.5 27.1 15.4 5.3
MEF2B Mut -4.5 12.5 14.1 6.2 26.3 7.4 2.6 Genetic subtype
CIITA Mut HL HD Fusion -3.7 7.5 12.0 0 25.0 22.2 2.6 Gene expression subgroup
ARID1A Trunc HD -5.9 3.8 8.8 0 22.9 0 0
GNA13 Mut HL HD -3.0 5.1 10.0 0 22.5 19.2 7.9
STAT6 Mut Amp -9.0 0 1.1 0 21.1 3.8 0
PTEN HL HD -3.1 8.9 4.4 6.2 20.0 3.8 7.9 Prevalence (%)
Chrom. 21
EBF1
Gain Amp
Trunc HL HD
-2.0
-2.2
7.6
2.5
7.8 25.0 20.0 7.7 2.6
10.0 0 18.6 19.2 15.8 Gene
log
Genetic feature pval MCD BN2 N1 EZB ST2 A53 N1
GNAI2 Mut -3.6 1.2 6.5 0 17.1 7.4 2.6 NOTCH1 Mut -30.7 0 0 100 0 0 0
C10orf12 Trunc HL HD -6.6 1.3 0 0 17.1 0 5.4 IRF2BP2 Mut -2.7 21.2 7.6 43.8 5.3 33.3 15.8
BCL7A Mut -3.0 5.1 10.3 0 15.8 0 0 Chrom. 4p Gain Amp -4.3 2.5 1.1 31.2 1.4 3.8 10.5
HLA-DMB Mut -4.1 0 6.9 6.2 15.8 3.7 2.7 ID3 Mut -2.7 1.2 3.3 25.0 5.3 11.1 0
S1PR2 Mut -8.1 0 0 0 14.5 0 2.6 BCOR Trunc -2.9 3.8 6.5 25.0 0 0 2.6
MAP2K1 Mut HL HD -3.3 0 2.3 0 11.4 0 2.7 EPB41 Mut -2.8 0 2.3 18.8 3.9 0 0
FBXO11 Mut HL HD -2.1 0 2.2 0 10.0 0 5.3 IKBKB Mut -2.4 1.2 0 18.8 6.6 3.7 0
MIR17HG Amp -2.1 2.5 1.1 0 10.0 0 2.6 ALDH18A1 Mut -3.5 3.8 0 18.8 0 0 0
Genetic subtype Genetic subtype
Gene expression subgroup Gene expression subgroup

Figure 2. Genetic Features of DLBCL Genetic Subtypes

Shown is the prevalence of the indicated genetic features in each DLBCL subtype. Log10 p value (pval) is based on the difference in prevalence in the indicated
subtype versus all other samples. HL, heterozygous loss; HD, homozygous deletion; Gain, single-copy gain; Amp, amplification; Mut, mutation; Trunc, protein-
truncating mutation: Fusion, chromosomal translocation. See also Figure S2.

2018). Inactivation of the S1PR2/GNA13 pathway in EZB alters
GC B cell migration and signaling (Muppidi et al., 2014). Phos-
phatidylinositol 3-kinase (PI3K) signaling in EZB is promoted by
inactivating mutations and deletions of PTEN and MIR17HG
amplification (Pfeifer et al., 2013). Recurrent REL amplification
may deregulate EZB metabolism and growth, as in normal GC
B cells (Heise et al., 2014).
Several genetic lesions in EZB potential perturb their interac-
tion with T follicular helper (TFH) cells. Major histocompatibility
complex (MHC) class II expression and function in EZB may be
compromised by EZH2 activation (Ennishi et al., 2019b) and
inactivation of CIITA (Steidl et al., 2011) and HLA-DMB (Denzin
and Cresswell, 1995), potentially modulating interactions be-
tween lymphoma cells and TFH cells. The survival of EZB lym-
phoma cells that inactivate herpesvirus entry mediator
(TNFRSF14) may be enhanced by TFH-mediated CD40 signaling
(Mintz et al., 2019). STAT6 mutations may modulate the ability of
TFH-derived interleukin-4 (IL-4) to promote plasma cell differenti-
ation (Weinstein et al., 2016).
ST2 is named for its recurrent SGK1 and TET2 mutations. ST2
tumors acquire TET2 truncating mutations suggesting a tumor-
suppressor function, as in mouse GC B cell lymphomagenesis
(Dominguez et al., 2018). The majority of SGK1 mutations are
truncating, suggesting a tumor-suppressor function that may
involve PI3K signaling, since SGK1 is an AKT-family kinase
(Di Cristofano, 2017). JAK/STAT signaling is likely promoted in
ST2 by inactivation of SOCS1 (Linossi and Nicholson, 2015),
inactivation of DUSP2 (Lu et al., 2015), and by known STAT3-
activating mutations (Y640F, D661Y) (Crescenzo et al., 2015). In-
activating mutations in ST2 targeting P2RY8 and its signaling
mediator GNA13 prevent responses to S-geranylgeranyl-L-
glutathione, which spatially confines normal GC B cells and

554 Cancer Cell 37, 551–568, April 13, 2020

 A C D

B E F G

Figure 3. Similarities of DLBCL Genetic Subtypes to Other Lymphoid Malignancies

(A) Prevalence of CD79B and MYD88L265P mutations in the indicated nodal and extranodal forms of DLBCL, shown according to the color code above. The
percent prevalence of tumors with the indicated genotypes in each of the indicated lymphoma types is shown according to the color code.
(B) Prevalence of MCD-defining mutations in primary CNS lymphoma and primary cutaneous lymphoma. Other NHL, other non-Hodgkin lymphomas (see STAR
Methods).
(C) Secondary extranodal involvement in genetic subtypes of DLBCL. p Value is based on Fisher’s exact test,
(D) Genetic aberrations favoring immune escape in MCD DLBCL.
(E) Prevalence of BN2-defining mutations in the indicated types of marginal-zone lymphoma (MZL) and in other NHLs.
(F) Prevalence of EZB-defining mutations in follicular lymphoma (FL), transformed FL, and other NHLs.
(G) Prevalence of ST2-defining mutations in nodular lymphocyte-predominant Hodgkin lymphoma (NLPHL), T cell/histiocyte-rich large B cell lymphoma
(THRLBCL), and other NHLs.
See also Figure S3.

inhibits their AKT activity (Lu et al., 2019). Finally, some ST2 tu-
mors apparently activate NF-kB by inactivating IkBa (NFKBIA)
(Baeuerle and Baltimore, 1988).
A53 is characterized by TP53 mutations and deletions. A53
tumors also acquire homozygous deletions and mutations tar-
geting 53BP1 (TP53BP1), a DNA-damage sensor that pre-
vents aneuploidy (Celeste et al., 2002), consistent with the
recurrent gains and losses of chromosome arms in A53.
Some A53 abnormalities have been previously associated
with ABC DLBCL, including: deletion of 6q, harboring the
tumor suppressors TNFAIP3 and PRDM1; gain/amplification
of 3q (Lenz et al., 2008b); focal amplification of NFKBIZ (Nogai
et al., 2013); amplification of CNPY3 (Phelan et al., 2018); and
BCL2 amplification. Additional focal deletions target the
tumor suppressors p73 and ING1 (Tallen and Riabowol,
2014). Finally, A53 tumors frequently delete or mutationally
inactivate b2-microglobulin (B2M), providing a mechanism
of escape from immune surveillance (Challa-Malladi
et al., 2011).
N1 is characterized by gain-of-function NOTCH1 mutations,
similar to those in chronic lymphocytic leukemia and mantle
cell lymphoma. These tumors additionally acquire mutations tar-
geting B cell differentiation regulators (ID3, BCOR) and IkB ki-
nase b (IKBKB), including the V203I isoform that constitutively
activates NF-kB (Cardinez et al., 2018).
Relationship of DLBCL Genetic Subtypes to Other
Lymphoid Malignancies
While DLBCLs typically present clinically in lymph nodes and other
immune tissues, primary extranodal lymphomas present as tu-
mors involving various non-lymphoid organs. Primary extranodal
lymphomas frequently acquire MYD88L265P and/or CD79B muta-
tions as well as other MCD-defining mutations (Figures 3A, 3B,
and S3B). Notably, in the NCI cohort, MCD DLBCL tumors spread
secondarily to extranodal sites in 30% of cases, and 46% of these
occurred at sites that can give rise to primary extranodal lym-
phomas—the central nervous system (CNS), vitreo-retina, testis,
and breast—whereas other DLBCL subtypes spread to these sites
significantly less often (Figure 3C). Primary extranodal lymphomas
often arise in the CNS, ocular vitreo-retina, and testis, all consid-
ered ‘‘immune-privileged’’ sites because they tolerate allografts
and permit only selective access by immune cells (Shechter
et al., 2013). In this regard it is notable that 72.5% of MCD tumors
acquire homozygous deletions, truncating mutations, or translo-
cations that could allow them to evade immune surveillance by
several mechanisms including (1) reduced antigen presentation
due to MHC class I or TAP1 inactivation, (2) decreased T cell acti-
vation due to gene fusions that elevate expression of CD274 and
PDCD1LG2, encoding PD-L1 and PD-L2, respectively, and (3)
diminished NK activation due to CD58 inactivation (Challa-Malladi
et al., 2011) (Figures 3D and S3A).

Cancer Cell 37, 551–568, April 13, 2020 555

 Genetic aberrations of several DLBCL subtypes reveal poten-
tial pathogenetic relationships with more indolent lymphomas.
Mutations characteristic of BN2 link this subtype to marginal-
zone lymphomas (MZLs) (Figure 3E), befitting the essential role
of NOTCH2 in the differentiation of follicular B cells to marginal
zone B cells (Saito et al., 2003). BCL6 translocations, which char-
acterize BN2, are rare in indolent MZLs but common in MZLs that
have transformed into aggressive large cell variants (Flossbach
et al., 2011; Ye et al., 2008). Follicular lymphoma (FL) shares
many genetic lesions with EZB (Figure 3F), as does transformed
FL, which can histologically resemble DLBCL. The genetic
signature of ST2 betrayed an intriguing similarity with two histo-
logically distinct lymphomas, nodular lymphocyte-predominant
Hodgkin lymphoma (NLPHL) and T cell histiocyte-rich large
B cell lymphoma (THRLBCL) (Figure 3G). NLPHL is an indolent
Hodgkin lymphoma variant that retains expression of GC B cell
genes and can transform into an aggressive large cell form
(Timens et al., 1986). Morphological similarities between NLPHL
and THRLBCL led pathologists to suspect a link between these
entities, which was reinforced by shared mutations in SOCS1,
DUSP2, SGK1, and JUNB, all characteristic of ST2 (Hartmann
et al., 2016; Schuhmacher et al., 2019).
Validation of the LymphGen Classification
To evaluate the reproducibility of the LymphGen algorithm in
identifying genetic subtypes of DLBCL, we used it to assign tu-
mors from two validation cohorts to genetic subtypes (Figure S4A
and Table S1). The first cohort (n = 304) was used previously to
identify DLBCL subtypes (denoted ‘‘Harvard’’) and was analyzed
for exomic mutations, copy number in selected genomic regions,
and BCL2/BCL6 rearrangements (Chapuy et al., 2018). A second
cohort (n = 332) was used previously to identify signatures of
poor prognosis in DLBCL (denoted ‘‘BCC’’) and was analyzed
here for mutations in 82 lymphoma-associated genes as well
as for whole-genome copy-number aberrations (Table S2) (En-
nishi et al., 2019a).
Because each of the DLBCL cohorts had different data types
available, we designed LymphGen to function using various
combinations of mutational data (whole-exome or gene panel re-
sequencing), copy-number data (regional or whole genome), and
rearrangement data for BCL2 and BCL6. Given the robust per-
formance of LymphGen with varying genetic inputs (Figure S4B),
we have implemented LymphGen for general research use at
https://llmpp.nih.gov/lymphgen/index.php.
To compare the LymphGen-assigned subtypes between the
cohorts, we first normalized the cohorts to be equivalent to a
population-based DLBCL cohort with respect to overall COO
composition (Scott et al., 2015), given the relationship between
COO and the genetic subtypes. In these normalized cohorts,
the prevalences of the gene subtypes were roughly comparable
(Figure 4A), as were their COO compositions (Figure 4B). More-
over, in the Harvard cohort, each LymphGen subtype was drawn
predominantly from a single genetic ‘‘cluster,’’ as defined previ-
ously (Chapuy et al., 2018), with a 75% overall agreement be-
tween the analytic methodologies (Figure S4C).
The genetic features associated with each subtype in the three
cohorts were generally comparable in prevalence (Figure 4C). To
evaluate this genetic coherence quantitatively, we iteratively
computed LymphGen subtype scores based on gene sets in
556 Cancer Cell 37, 551–568, April 13, 2020
which one subtype-defining genetic feature was left out, and sta-
tistically compared the scores between tumors in which the
omitted genetic feature was present or absent. By this metric,
we observed significant similarity in the genetic coherence of
features defining the MCD, BN2, EZB, and ST2 subtypes in the
two validation cohorts (p % 6.3 3 10 7; Table S3); N1 could
not be evaluated by this method due to the statistical dominance
of NOTCH1 mutations. Since A53 is defined primarily by copy-
number alterations and TP53 inactivation, we instead statistically
evaluated the relationship between the number of copy-number
alterations in each case with TP53 mutations and/or deletion,
which again revealed significant genetic similarity between the
cohorts (p % 4.4 3 10 9; Table S3).
We next evaluated the survival of patients assigned to Lymph-
Gen subtypes in each cohort. The overall survival characteristics
of the cohorts were distinct, most likely reflecting accrual bias (Fig-
ure S5A, Table S1). Nonetheless, the genetic subtypes defined in
each cohort had similar associations with overall survival, as
judged by the Kaplan-Meier method (Figure 4D). Within each
cohort, MCD had an inferior survival, especially when compared
with ST2 and BN2. Among ABC cases in each cohort, BN2 was
favorable, especially when compared with MCD and A53. Among
GCB cases, EZB had an inferior survival compared with ST2. Given
these consistent survival trends, we used data from all three co-
horts to estimate joint hazard ratios (Figure 4E). In this model, the
survival of MCD was inferior to ST2, BN2, and all non-MCD pa-
tients (p < 0.001); the survival of BN2 was favorable compared
with MCD, A53, and all non-BN2 patients within ABC (p < 0.01);
and the survival of EZB was inferior to ST2 within GCB (p = 0.032).
While the genetic subtypes clearly subdivided the outcomes
within the ABC and GCB gene expression subgroups, the
reverse was also true. Within BN2 and A53, the COO subgroups
had significantly disparate survival characteristics (Figure S5B),
demonstrating that tumor genotype and phenotype must both
be considered when attempting to understand the response to
therapy.
Subtypes of EZB DLBCL with Distinct Genetic,
Phenotypic, and Clinical Attributes
Given the recent demonstration that GCB DLBCL can be subdi-
vided into prognostic subtypes by two gene expression signa-
tures (MHG and DHIT), we investigated how this phenotypic
distinction was related to the DLBCL genetic subtypes (Ennishi
et al., 2019a; Sha et al., 2019). Since these signatures were
correlated in the NCI cohort (p = 1.5 3 10 14), we focused on
DHIT for simplicity. This signature was initially identified using
the subset of GCB tumors that have BCL2 and MYC rearrange-
ments (‘‘double hit’’), which are known to have a poor prognosis,
but could also identify a larger subset of GCB tumors with an
inferior prognosis (Ennishi et al., 2019a).
Using the NCI cohort, we investigated the relationship of DHIT
with other gene expression signatures and genetic features.
Among GCB cases, EZB had significantly higher DHIT scores
than other subtypes (p = 0.002; Figure 5A). Likewise, among
30 GCB tumors classified as DHIT+, the majority (70%) were
either EZB or genetically composite cases with features of EZB
and A53 (Figure 5B).
The gene expression signatures most correlated with DHIT
were two that distinguish BL from DLBCL: the MHG signature
A B

D E

Figure 4. Validation of the LymphGen Classification

(A) Prevalence of DLBCL subtypes classified by LymphGen.
(B) Prevalence of COO subgroups within genetic subtypes.
(C) Prevalence of the indicated genetic features within the genetic subtypes defined in the Harvard and BCC cohorts in comparison with the NCI cohort.
(D) Kaplan-Meier plots of overall survival within the indicated DLBCL cohorts, in all cases, ABC cases, or GCB cases, as indicated.
(E) Hazard ratios ( log2 transformed) for the indicated comparisons between LymphGen subtypes in the indicated DLBCL cohorts. Error bars denote SEM.
Significance: ****p % 0.0001; ***p % 0.001; **p % 0.01; *p % 0.05. See also Figures S4 and S5; Tables S1, S2, and S3.

and GCB-4, defined as the subset of genes characteristically ex-
pressed in normal GC B cells that are expressed more highly in
BL than in DLBCL (Dave et al., 2006) (Table S4). Signatures of in-
termediate- and dark-zone GC B cells were also correlated
(GCB-9 and GCB-10, respectively), suggesting that DHIT re-
flects dynamic changes in GC B cell differentiation. DHIT was
also correlated with signatures of MYC activity, notably
MYCUp-4, consisting of genes that are induced by MYC and
are direct MYC binding targets (Zeller et al., 2006). DHIT also
correlated with a signature of adverse survival in DLBCL
(Prolif-6) that includes MYC and its target genes GNL3 and
NPM3 (Rosenwald et al., 2002).
We therefore hypothesized that DHIT is a composite signature
that reflects both GC B cell differentiation and MYC activity. We
used GCB-4 and MycUp-4 to represent these phenotypes, given
their strong association with DHIT by gene set enrichment anal-
ysis (Figure 5C). A linear model combining these signatures was
significantly correlated with DHIT (Figure 5D) and accounted for
60.2% of DHIT variance among EZB cases within GCB.
GCB patients with DHIT– tumors had a better survival than
those with DHIT+ tumors (Figure S6A), in part reflecting the
enrichment among DHIT– cases of the prognostically favorable
ST2, BN2, and A53 subtypes. Within the EZB subset of GCB,
the survival of DHIT+ was significantly worse than DHIT–, which
was not true in the non-EZB subset (Figure 5E), leading us to
confine our investigation of DHIT to EZB.
We next explored the association of genetic features with the
DHIT+ and DHIT– subsets of EZB GCB cases (Figure 5F). The
majority of EZB-defining genetic features were observed
comparably in these two subsets with the exception of GNA13

Cancer Cell 37, 551–568, April 13, 2020 557

A B C D

E F

Figure 5. Genetic Analysis of the DHIT Signature

(A) Relative expression of DHIT in the indicated subtypes within GCB DLBCL. Error bars indicate SEM.
(B) Prevalence of subtypes within DHIT+ GCB DLBCL.
(C) Gene set enrichment analysis of DHIT versus the GCB-4 and MYCUp-4 signatures. Cases are ranked by T statistic, with high expression of the DHIT signature
to the left. Kolmogorov-Smirnov p values are shown.
(D) Correlation between the DHIT score and a linear model score derived using GCB-4 and MYCUp-4 signature averages. Each dot is an EZB case. A F-test
p value with two degrees of freedom is shown.
(E) Kaplan-Meier plots of survival for DHIT+ and DHIT– cases among EZB (left) and non-EZB (right) GCB cases. p Values are calculated using a log rank test.
(F) Genetic features that distinguish EZB-MYC+ (DHIT+) from EZB-MYC– (DHIT–) GCB DLBCL (top two panels), and features shared by EZB-MYC+ and EZB-MYC–
(bottom panel). Log10 p value (pval) is based on the difference in prevalence between EZB-MYC+ and EZB-MYC– cases. ns, not significant.
(G) Prevalence of genetic features that distinguish EZB-MYC+ from EZB-MYC– in BL.
See also Figure S6 and Table S4.

mutations, which were more prevalent among DHIT+ than DHIT–
cases (p = 0.025). In keeping with a role for MYC in DHIT+ cases,
MYC rearrangements, amplifications, and mutations were signif-
icantly enriched in these tumors (p = 0.0079). Mutations or homo-
zygous deletions of TP53 were more than twice as prevalent in
DHIT+ than DHIT– cases, while the tumor suppressor DDX3X
was mutated in one-third of DHIT+ tumors but never in DHIT– tu-
mors. FOXO1, a transcription factor (TF) that is inactivated by
PI3K signaling, was targeted by mutations more than 3 times
as often in DHIT+ cases. Conversely, DHIT– cases were signifi-
cantly enriched in mutations targeting the NF-kB regulators
A20 (TNFAIP3) and CARD11, as well as deletions of the tumor
suppressor TP73. We were able to use the genetic data to create
a probabilistic model of EZB-MYC+ versus EZB-MYC– that could
distinguish these subtypes effectively (Figure S6B; permutation
p value = 0.004, see STAR Methods). Given the association of
MYC abnormalities with DHIT+ EZB cases, we hereafter refer
to DHIT+ EZB as ‘‘EZB-MYC+’’ and DHIT– EZB as ‘‘EZB-MYC–’’.
Genes preferentially mutated in EZB-MYC+ were also
frequently mutated in BL whereas those preferentially mutated
in EZB-MYC– were not (Figure 5G). However, some genetic ab-
errations that define BL, such as ID3 and TCF3 mutations
(Schmitz et al., 2012), were not observed in EZB-MYC+. Thus,
the genetic program adopted by EZB-MYC+ is shared by BL,
but these lymphomas are genetically distinct.
Phenotypic Distinctions between Genetic Subtypes
Gene expression signatures offer glimpses into tumor pheno-
types that are differentially manifested in DLBCL genetic sub-
types (Schmitz et al., 2018). We identified signatures whose

558 Cancer Cell 37, 551–568, April 13, 2020

 Malignant Processes B Cell Differentiation B cell Transcription Factors
1.4 1.4 1.4
1.2 1.2 1.2
1.0 1.0 1.0
Normalized Signature Average

0.8 0.8 0.8

0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
** * ** ** ** ** ** ** ** ** ** ** ** ** ** ** * * ** **
0 * * * ** * **
** ** *
0
** ** * * * * * ** ** *** 0 * * * * ** * * ** ** ** * ** ** ** *
-0.2 * * * -0.2 * * * ** ** -0.2 ** ** * ** * * *
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
-1.0 -1.0 -1.0
-1.2 -1.2 -1.2
-1.4 -1.4 -1.4
-2.0
-3.0
Signature Prolif-11 MycUp-4 Ribo-1 Quiesce-2 Glycol-1 Lipid-1 GCB-1 GCB-8 GCB-9 GCB-10 PC-1 Bcell-6 IRF4Up-7 IRF4Dn-1 OCT2Up-1 BCL6Dn-1 TCF3Up-1
Phenotype Proliferation MYC Ribosomal Proliferation Glycolytic Lipid GC GC light GC intermediate GC dark Plasma Memory IRF4 IRF4 OCT2 BCL6 TCF3
high induced proteins low pathway synthesis B cells zone cells zone cells zone cells cells B cells activated repressed induced repressed activated

Oncogenic Signaling Pathways Tumor Microenvironment

0.8 0.8

0.6 0.6
DLBCL
Normalized Signature Average

0.4 0.4 genetic

subtype
MCD
0.2 0.2
BN2
*** ** ** ** ** ** ** ** N1
EZB-MYC+
** * ** ** * * ** * * * * * * * * *
0 0
** * * * *** ** ** **
* ** * * ** ** ** ** **
*
** **
*
** **
*
** **
*
** ** **
* * EZB-MYC–
ST2
-0.2 -0.2
A53

-0.4 -0.4 ** p < 0.001

* p < 0.01
-0.6 -0.6 **

-0.8 -0.8
-1.0 -1.0
-2.0 -2.0
-3.0

Signature NFkB-10 p53Up-1 NotchUp-6 PI3KUp-1 JAK2Up-2 JAKUp-1 GCThUp-1 CD4T-2 CD8T-7 Treg-1 NK-3 MPhage-1 MPhage-2 DC-9 Stromal-1
Phenotype NF-κB p53 Notch PI3 kinase JAK2 kinase JAK1 kinase GC TFH CD4 CD8 Regulatory Natural killer Type 1 Type 2 Activated Fibrosis
activated activated activated actvated activated activated cells T cells T cells T cells cells macrophages macrophages dendritic cells (good prognosis)

Figure 6. Gene Expression Signature Expression in DLBCL Subtypes

Shown is average log2 expression of signature genes in each subtype versus other DLBCL samples in the NCI cohort. Error bars denote SEM. See also Table S5.

average expression was significantly associated with one or 10 4) and their respective target genes. Although both IRF4
more genetic subtypes (Figure 6 and Table S5). The subtypes and OCT2 promote GC B cell differentiation to plasma cells
differed with respect to various malignant attributes, with MCD (Hodson et al., 2016; Sciammas et al., 2006), MCD had low
and EZB-MYC+ highly expressing signatures of proliferation expression of a plasma cell signature, likely due to inactivation
and MYC activity, while N1 instead expressed a signature of of Blimp-1 (PRDM1), which is required for plasma cell
quiescence. MYC induces ribosome biogenesis (Dang, 2013) differentiation.
and, accordingly, EZB-MYC+ tumors highly expressed a ribo- NF-kB target genes were highly expressed in MCD but not
somal protein signature. Metabolic distinctions between the EZB, as expected (Davis et al., 2001), but were also high in
subtypes included high expression of glycolytic pathway en- BN2 and ST2. As expected, p53 target genes were lowest in
zymes in ST2, consistent with a Warburg effect, and high expres- A53 tumors, and NOTCH target genes were significantly upregu-
sion of lipid synthetic enzymes in EZB-MYC+. lated in N1 and BN2. ST2 expressed a signature of PI3K signaling
The subtypes appeared to derive from distinct B cell differen- highly, potentially due to SGK1 inactivation, as well as a signa-
tiation stages, with a signature of GC B cells characterizing EZB ture of JAK2 signaling, consistent with SOCS1 and DUSP2
and ST2. EZB-MYC– tumors resembled GC light-zone cells inactivation.
whereas EZB-MYC+ tumors resembled intermediate-zone cells The tumor microenvironments of the subtypes were strikingly
(Milpied et al., 2018), which are likely generated from light-zone discordant: N1 expressed signatures of multiple immune cell
cells by MYC expression (Dominguez-Sola et al., 2012). Genes types while A53 and EZB-MYC+ tumors were relatively low for
repressed by BCL6 in the GC were lowest in EZB-MYC+ tumors, all immune signatures. T cell signatures were selectively low in
and genes transactivated by another GC TF, TCF3, were highest MCD, potentially stemming from defective antigen presentation
in EZB and ST2 tumors. due to genetic abnormalities or to their high IL10 expression rela-
MCD and N1 lacked GC signature expression, with N1 instead tive to other DLBCLs (p = 1.7 3 10 11) (Mittal and Roche, 2015).
expressing a memory B cell signature (Suan et al., 2017), sug- A GC TFH signature was highest in EZB-MYC– tumors, consistent
gesting a post-GC origin. MCD tumors had high expression of with their similarity to GC light-zone cells, but was lower in EZB-
the TFs IRF4 (p = 5.1 3 10 7) and OCT2 (POU2F2) (p = 1.3 3 MYC+ and ST2, despite their GC derivation. The stromal-1

Cancer Cell 37, 551–568, April 13, 2020 559

signature, which is prognostically favorable and reflects a
fibrotic, macrophage-rich microenvironment (Lenz et al.,
2008a), was upregulated in EZB-MYC– and ST2 tumors, befitting
their relatively favorable outcomes.
Functional Genomics of DLBCL Genetic Subtypes
We next considered whether the DLBCL genetic subtypes might
offer insights into the response to targeted therapy. Many
lymphoid malignancies, including DLBCL, respond to BTK inhib-
itors, which block the transmission of signals from the BCR to
NF-kB (Davis et al., 2010). Genetic lesions targeting the BCR
pathway are prevalent in DLBCL but are differentially enriched
in the genetic subtypes (Figures 7A and 7B). Mutations that acti-
vate the BCR subunit CD79B were confined to MCD, BN2, and
A53, whereas mutations targeting the CD79A BCR subunit
were enriched in EZB, suggesting qualitatively distinct roles of
CD79A and CD79B mutations in lymphomagenesis. MCD tu-
mors were enriched in the MYD88L265P mutation, a hallmark of
tumors in which the My-T-BCR protein supercomplex activates
NF-kB (Phelan et al., 2018). By contrast, BN2 tumors acquired
PRKCB, BCL10, and TRAF6 mutations that may promote the for-
mation or function of the CARD11-BCL10-MALT1 (CBM) protein
complex, and also frequently acquired mutations inactivating
TNFAIP3 (A20) and TNIP, thereby promoting IkB kinase activity.
In aggregate, BN2 and MCD had the highest prevalence of ge-
netic lesions altering BCR-dependent NF-kB signaling, while
N1 and EZB had the lowest prevalence (Figure 7B). Finally, it is
notable that each genetic subtype acquired lesions targeting
known negative regulators of proximal BCR signaling, suggest-
ing that BCR signaling has a pervasive influence on lymphoma-
genesis (Figure 7B).
The survival of ABC cell lines relies on engagement of autor-
eactive BCRs by self-antigens, whereas GCB models rely on
an antigen-independent, ‘‘toncogenic’’ form of BCR signaling
(Havranek et al., 2017; Young et al., 2015, 2019). Consistent
with this model, self-reactive BCRs with the VH4-34 immuno-
globulin (Ig) heavy-chain variable (VH) region are enriched among
ABC tumors (Young et al., 2015). We assembled the expressed
IgVH regions in tumors in the NCI cohort using RNA-sequencing
data and observed that VH4-34 was the dominant IgVH region in
MCD, BN2, and A53, suggesting that these subtypes may rely
upon self-antigen-dependent chronic active BCR signaling (Fig-
ure 7C). These subtypes were also distinctive in their use of IgM
BCRs, which in normal B cells promote proliferation rather than
differentiation (Dogan et al., 2009).
To functionally evaluate the BCR pathway in DLBCL, we used
cell lines that bear genetic hallmarks of the genetic subtypes
(Table S6). We first investigated whether the negative regulators
of proximal BCR signaling that are genetically inactivated in
DLBCL modulate chronic active BCR signaling in DLBCL
models. We assayed the relative ability of cells to survive in the
presence of submaximal concentrations of the BTK inhibitor
ibrutinib as an effective proxy for BCR signaling strength (Wilson
et al., 2015b). In Cas9-expressing models of MCD and BN2, we
knocked out various BCR-negative regulators by expressing
short guide RNAs (sgRNAs) together with GFP and quantified
the relative numbers of live, GFP+/sgRNA+ cells in the presence
of ibrutinib compared with DMSO-treated cultures. Knockout of
each BCR-negative regulator promoted survival in ibrutinib, as
560 Cancer Cell 37, 551–568, April 13, 2020
did knockout of the NF-kB negative regulators A20 (TNFAIP3)
and TNIP1, whereas a control sgRNA had no effect (Figure 7D).
To investigate essential pathways in the genetic subtypes, we
performed whole-genome loss-of-function CRISPR screens in
Cas9-expressing models of MCD (TMD8, HBL1, OCI-Ly10),
BN2 (Riva), and EZB (OCI-Ly1, SUDHL4, WSU-DLCL2) (Phelan
et al., 2018). For each gene targeted by the sgRNA library, we
calculated a CRISPR screen score (see STAR Methods), with
negative scores indicating an essential gene (Table S7). Based
on this metric, all three subtypes depended on the BCR subunits
CD79A and CD79B (Figure 7E), whereas only MCD and BN2
models depended on signaling proteins downstream of the
BCR that activate NF-kB. Of particular note, BTK was essential
in the MCD and BN2 models but not in the EZB models (Fig-
ure 7E). Previous studies have focused on the intense addiction
of MCD models to BCR-dependent NF-kB signaling (Davis et al.,
2010; Phelan et al., 2018), but the contribution of BCR signaling
in BN2 was not anticipated. Constitutive BCR signaling in the
BN2 model was confirmed by knockdown of IgM or CD79A,
which decreased phosphorylation of LYN, SYK, and BTK (Fig-
ure 7F). Accordingly, the growth of Riva xenografts was strongly
suppressed by low doses of ibrutinib (Figure 7G), whereas
similar doses of ibrutinib only modestly suppressed the growth
of MCD xenografts (Figure S7).
We further used the whole-genome CRISPR data to predict
the dependency of MCD, BN2, and EZB on other signaling and
regulatory pathways that can by targeted by clinically available
drugs (Figure 7E). Two master regulatory TFs in ABC DLBCL,
IRF4 and SPIB, were selectively essential in the MCD and BN2
but not EZB models, which is notable since they can be downre-
gulated by lenalidomide (Yang et al., 2012). MCD models de-
pended on the IL-10 receptor a and b subunits, JAK1 and
STAT3, consistent with autocrine IL-10 signaling in this subtype
(Lam et al., 2008; Rui et al., 2016). The PRC2 chromatin
repressor complex was especially essential in the EZB models,
all of which had gain-of-function EZH2 mutations. The PI3K
pathway, which can be activated by BCR signaling in ABC and
GCB subtypes (Young et al., 2019), was essential in models of
all three subtypes. BCL2 was also pan-essential, whereas
BCL-XL (BCL2L1) was selectively required in the MCD models.
DISCUSSION
The extreme genetic and phenotypic heterogeneity of DLBCL
presents a challenge to the development of precision therapies.
Here, we provide a genetic framework from which to understand
therapeutic responses in subsets of DLBCL tumors defined by
shared pathogenesis. The DLBCL taxonomy defined by the
LymphGen algorithm unifies two recent genetic profiling studies
(Chapuy et al., 2018; Schmitz et al., 2018) and was also evident in
the independent BCC cohort. This classification breaks DLBCL
into seven genetic subtypes that differ with respect to oncogenic
pathway engagement, gene expression phenotype, tumor
microenvironment, survival rates, and potential therapeutic tar-
gets (Figure 8A). As such, this taxonomy provides a roadmap
for understanding the biological diversity encompassed within
the pathological diagnosis of DLBCL and will likely shed light
on the heterogeneous responses of DLBCL to cytotoxic and
molecularly targeted therapies.
 A E

F G

Figure 7. Functional Genomics Using Models of DLBCL Genetic Subtypes

(A) Contribution of each genetic subtype to the indicated genetic aberrations in the BCR-dependent NF-kB pathway. The color bar associated with each gene
illustrates the prevalences of each subtype, as indicated, estimated using the NCI cohort and adjusting for a population-based distribution of COO subgroups
(see STAR Methods).
(B) Fraction of DLBCL subtype cases with genetic alterations targeting the BCR-dependent NF-kB pathway or negative regulators of proximal BCR signaling.
(C) Top: fraction of cases expressing the IgVH4-34 variable region or other IgVH regions. Bottom: fraction of cases expressing the indicated Ig heavy chain (IgH)
constant regions.
(D) CRISPR-mediated knockout of BCR and NF-kB-negative regulators promotes survival in models of MCD and BN2 DLBCL. Cas9+ cells expressing the
indicated sgRNAs with GFP were cocultured with parental (GFP–) cells for the indicated times in ibrutinib. Increasing values indicate relative ibrutinib resistance of
the sgRNA+ cells.
(E) Genome-wide CRISPR loss-of-function screens. The indicated Cas9+ models of MCD, BN2, and EZB were transduced with a genome-wide sgRNA library,
and the sgRNA abundance was quantified before and after 3 weeks in culture. Asterisk: targeted by approved or investigational drugs.
(F) Effect of BCR knockdown on signaling in a BN2 model. Riva cells were transduced with the indicated small hairpin RNAs (shRNA) and the effect on BCR
signaling was assessed by immunoblotting for the indicated proteins. Ctrl, control.
(G) Effect of ibrutinib on Riva xenograft growth. Following tumor establishment, mice (n = 5/group) were treated with the indicated ibrutinib doses or vehicle
control.
See also Figure S7; Tables S6 and S7.

Our investigation provides insight into the mechanisms by that endow the B cell precursor with the hallmarks of cancer.
which tumors in a DLBCL genetic subtype acquire a shared ge- Support for this model comes from the observation that the ge-
netic program (Figure 8B). In one model, the epigenetic nature of netic subtypes differ in the expression of B cell differentiation sig-
a subtype’s cell of origin necessitates certain oncogenic events natures. Alternatively, a precursor B cell may randomly acquire a

Cancer Cell 37, 551–568, April 13, 2020 561

 A

B C D

Figure 8. Implications of the DLBCL Genetic Subtypes for Pathogenesis and Therapy
(A) Summary of the relationship between DLBCL COO subgroups and the genetic subtypes (left). The genetic themes, phenotypic attributes, clinical correlates,
and treatment implications of each subtype are shown at right. Prevalences were estimated using the NCI cohort, adjusting for a population-based distribution of
COO subgroups (see STAR Methods). dep., dependent; FDC, follicular dendritic cell; LZ, light zone; IZ, intermediate zone.
(B) Models of selection for shared genetic features in DLBCL subtypes.
(C) Models accounting for genetic attributes shared by DLBCL genetic subtypes and indolent NHLs.
(D) Model of EZB-MYC+ and EZB-MYC– evolution.

‘‘founder’’ genetic lesion, the nature of which dictates the subse-
quent selection of secondary genetic lesions. For example, MYC
overexpression kills normal cells unless they also have lesions
that prevent cell death, such as the BCL2 translocations that
occur in the EZB-MYC+ subtype (Evan et al., 1992). Our proba-
bilistic approach raises a third, hybrid possibility. A substantial
subset of DLBCL tumors (5.7%) had a high probability of
belonging to more than one genetic subtype. This suggests a
model in which one genetic program is adopted by a tumor
initially and a second is subsequently acquired because it con-
fers an additional selective advantage (Figure 8B). Future work
will be needed to understand whether therapeutic sensitivity or
resistance of such genetically composite lymphomas is dictated
largely by one of its genetic programs or is influenced by each
program.
Several of the DLBCL genetic subtypes have intriguing similar-
ities to more indolent lymphoma types: BN2 resembles MZLs,
EZB resembles FL, and ST2 resembles both NLPHL and
THRLBCL. Three models could account for these genetic rela-
tionships (Figure 8C). A ‘‘direct evolution’’ model suggests that
some DLBCL patients have a concurrent but undiagnosed low-
grade malignancy that acquires additional genetic lesions, trans-
forming it into DLBCL. Consistent with this model, pathologists
recognize histologically ‘‘composite lymphomas’’ that have, at

562 Cancer Cell 37, 551–568, April 13, 2020

diagnosis of DLBCL, evidence of a concurrent low-grade lym-
phoma in the same biopsy (Kuppers et al., 2014). For example,
in composite lymphomas with both marginal zone and large
cell components, the large cells frequently acquire BCL6 trans-
locations, as is typical of BN2 (Flossbach et al., 2011). A
‘‘branched evolution’’ model posits the existence of a premalig-
nant B cell clone that can become an indolent lymphoma or a
DLBCL, depending on the nature of additional genetic alterations
it acquires. In some cases of transformed FL, for example, the
transformed lymphoma shares some of the genetic features
with the antecedent FL, but each lymphoma type has genetic at-
tributes not shared by the other (Green et al., 2013). Recent
studies of patients with autoimmune diseases uncovered B cell
clones that expand pathogenically with the acquisition of muta-
tions characteristic of DLBCL genetic subtypes, suggesting that
such cells could serve as a reservoir that can readily evolve into
either an indolent or aggressive lymphoma (Malecka et al., 2018;
Singh et al., 2020). In a final ‘‘convergent evolution’’ model, it is
formally possible that indolent lymphomas and DLBCL subtypes
separately select the same genetic programs to acquire a partic-
ular oncogenic phenotype while differing in other attributes.
Future genetic studies of composite and transformed lym-
phomas may shed light on these evolutionary models.
The sequential tumor evolution model most likely accounts for
the genetic relationship between EZB-MYC+ and EZB-MYC–,
which emerged from our study of the DHIT signature (Ennishi
et al., 2019a). The relatively inferior outcome of DHIT+ DLBCL
is not only due to the association of EZB-MYC+ with adverse sur-
vival but also to the enrichment of DHIT– cases in ST2, BN2, and
A53, all of which have good outcomes among GCB cases. While
EZB-MYC– and EZB-MYC+ share a substantial genetic program,
EZB-MYC+ is enriched in aberrations in MYC and four other
genes that are frequently mutated in BL. These tumors also ex-
pressed the subset of genes expressed by normal GC B cells
that are at higher levels in BL than DLBCL (Dave et al., 2006).
These genetic and phenotypic associations suggest that EZB-
MYC+ should be considered a genetic subtype of DLBCL that
arises from EZB-MYC– tumors with the acquisition of these ge-
netic lesions (Figures 8A and 8D).
‘‘Double-hit’’ lymphomas harboring translocations of MYC
and BCL2 have been associated with inferior survival in most se-
ries. Importantly, not all EZB-MYC+ cases are ‘‘double hit:’’ only
38% of these cases had a MYC abnormality while 78% had a
BCL2 translocation. Nonetheless, EZB-MYC+ cases expressed
genes that are direct targets of MYC (Zeller et al., 2006), suggest-
ing either that they have cryptic genetic abnormalities that
deregulate MYC, as described by Hilton et al. (2019), or have
other mechanisms to enhance MYC function. The EZB-MYC+
subtype thus expands the concept of ‘‘double-hit’’ lymphoma
while still identifying DLBCL patients with relatively adverse out-
comes. Of note, among non-EZB GCB cases, the DHIT signature
was not associated with adverse outcome. Hence, the EZB-
MYC+ and EZB-MYC– distinction refines the DHIT signature to
focus on GCB cases with an inferior prognosis.
Another intriguing genetic relationship links the MCD subtype
to primary extranodal lymphomas, including those involving im-
mune-privileged sites. Mutations in MCD-defining genes are
also characteristic of primary skin, breast, uterus, adrenal, and
intravascular lymphomas, supporting the hypothesis that these
tissues confer ‘‘relative’’ immune privilege (Shechter et al.,
2013). Notably, MCD tumors arising primarily in lymph nodes
often secondarily involved immune-privileged sites. Moreover,
MCD genomes are extensively modified by immunoediting (Mat-
sushita et al., 2012), characterized by one or more lesions impair-
ing MHC class I antigen presentation or the activation of T and
NK cells. Primary central nervous system lymphomas (PCNSLs)
also genetically abrogate immune responsiveness despite
arising in an immune-privileged site (Chapuy et al., 2016). These
observations suggest a quantitative, not qualitative, model for
immune evasion by MCD-like aggressive lymphomas in which
these tumors must acquire multiple lesions affecting immune
recognition and/or grow in relatively (but not absolutely) im-
mune-privileged sites to become ‘‘invisible’’ to immune
surveillance.
Our combined genetic, phenotypic, functional, and clinical
data demonstrate that the DLBCL genetic subtypes differ strik-
ingly in their response to standard immunochemotherapy and
may also respond differentially to targeted therapies (Figure 8A).
Genetic lesions targeting the BCR-dependent NF-kB pathway
were most frequent in the BN2, MCD, and A53 subtypes as
was the autoreactive VH4-34 variable region, suggesting that
these subtypes may be sensitive to BTK inhibitors. Indeed,
MCD-like tumors with MYD88L265P and CD79B mutations have
been associated with a high rate of response to ibrutinib
(R80%) in relapsed DLBCL and in PCNSL (Grommes et al.,
2017; Lionakis et al., 2017; Wilson et al., 2015b). Among the ge-
netic subtypes, BN2 had the highest prevalence of lesions
affecting the BCR-dependent NF-kB pathway. Moreover, a
BN2 model was highly sensitive to ibrutinib. These consider-
ations support the clinical evaluation of BTK inhibitors in BN2.
The PI3K pathway was essential in MCD, BN2, and EZB
models, likely for different mechanistic reasons. Other molecular
targets in MCD and BN2 include the master regulatory TFs IRF4
and SPIB, which are both downregulated in expression by lena-
lidomide, a drug that has shown promise in combination with
other agents in DLBCL (Goy et al., 2019; Wilson et al., 2015a;
Yang et al., 2012). IkB kinase activity, which is required in MCD
and BN2 models, can be attenuated by BET inhibitors targeting
BRD4 (Ceribelli et al., 2014). MCD models rely on autocrine IL-10
receptor signaling to activate JAK1 and STAT3 (Rui et al., 2016),
and a selective JAK1 inhibitor, INCB040093, has shown activity
in combination with a PI3Kd inhibitor in non-GCB DLBCL (Phillips
et al., 2018). The MYD88L265P isoform in MCD models spontane-
ously coordinates a signaling complex involving IRAK1 and
IRAK4 (Ngo et al., 2011) supporting the evaluation of IRAK4 in-
hibitors in MCD, especially in combination with a BTK inhibitor
(Kelly et al., 2015). EZB models bearing an EZH2 mutation
were preferentially reliant on the PRC2 repressor complex and
thus may respond preferentially to EZH2 inhibitors. BCL2 was
required in MCD, BN2, and EZB models while BCL-XL was addi-
tionally essential in MCD, suggesting that agents such as vene-
toclax or navitoclax may provide benefit (Mathews Griner
et al., 2014).
Given the above evidence that R-CHOP chemotherapy and
targeted therapies may be differentially active in particular ge-
netic subtypes, we feel that the LymphGen algorithm will be a
useful tool in DLBCL clinical trials that extends the utility of
gene expression-based assays. We speculate that the
Cancer Cell 37, 551–568, April 13, 2020 563
LymphGen classification will find initial utility in the retrospective
analysis of clinical trials. Faced with the genetic complexity of
DLBCL, it is challenging to identify and statistically verify the as-
sociation of individual genetic alterations with clinical outcome,
given the problem of multiple hypothesis testing. This problem
is mitigated by the fact that there are only seven DLBCL genetic
subtypes. Because these subtypes differentially acquire muta-
tions in particular signaling and regulatory pathways and have
distinct microenvironmental compositions, we anticipate that
they may differ in response to therapies targeting oncogenic
signaling pathways as well as immunotherapies. Ultimately, if a
DLBCL genetic subtype is enriched for therapeutic responses,
it could be used as a selection criterion for an expansion cohort
in a subsequent clinical trial. We have made the LymphGen algo-
rithm publicly accessible at https://llmpp.nih.gov/lymphgen/
index.php in order facilitate its use in DLBCL clinical trials and
accelerate the development of improved therapies for these
aggressive cancers.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d LEAD CONTACT AND MATERIALS AVAILABILITY
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Cell Lines
B Mice
d METHOD DETAILS
B LymphGen Algorithm Development
B Revision of Genclass Procedure
B Mutation Features
B Copy Number Features
B Combination Features
B LymphGen Methodology
B Measures of Feature Significance
B Gene List Selection
B Feature Selection within a Gene
B Hierarchical Feature Selection within a Gene
B Example 1: ETV6 in MCD
B Example 2: IRF4 in MCD
B Single-Class Sample Prediction
B Example 1b: ETV6 in MCD
d EXAMPLE 2B: IRF4 IN MCD
B Combining Models to Generate Final Sample Call
B Application of LymphGen to Imperfect Data
B Evaluation of Model Performance on Subclasses of
Features
B Prediction on Validation Sets
B Model Verification via Gene Cross-Validation
+ –
B Genetic Prediction of EZB-MYC and EZB-MYC
B Phosphoprotein Analysis of BCR Signaling
B CRISPR Screens
B Analysis of Tumor Suppressor Genes
B Mouse Xenograft Experiments
B Publicly Available Data Used in Study
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Estimation of DLBCL Genetic Subtype Prevalence
564 Cancer Cell 37, 551–568, April 13, 2020
B RNA-seq Analysis
B CRISPR Screen Data Analysis
B Prevalence of Genetic Alterations in Other Lymphomas
B General Statistical Methods
d DATA AND CODE AVAILABILITY
d ADDITIONAL RESOURCES
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2020.03.015.
ACKNOWLEDGMENTS
Supported by the Intramural Research Program of the NIH, National Cancer
Institute, Center for Cancer Research, grant 1P01CA229100 from the National
Cancer Institute, and by grants from the Terry Fox Research Institute (1061,
1043). We wish to acknowledge the NIH Biowulf high-performance computing
system for help with data analysis and the Frederick National Laboratory for
Cancer Research Sequencing Facility managed by Bao Tran.
AUTHOR CONTRIBUTIONS
Conceptualization, G.W.W., D.W.H., and L.M.S.; Methodology, G.W.W.,
D.W.H., and L.M.S.; Software, G.W.W., D.W.H.M and Z.A.C.; Formal Analysis,
G.W.W., D.W.H., O.P., N.K., R.D.M., J.T., R.M.Y., A.B., and L.M.S.; Investiga-
tion, J.D.P., S.R., R.M.Y., J.Q.W., and R.S.; Data Curation, D.W.H., R.D.M.,
D.W.S., and J.T.; Writing – Original Draft, G.W.W., D.W.H., J.D.P., R.M.Y.,
J.Q.W., and L.M.S.; Writing – Review & Editing, all authors; Visualization,
G.W.W., D.W.H., C.A.J., and L.M.S.; Supervision, L.M.S.
DECLARATION OF INTERESTS
G.W.W., D.W.H., and L.M.S. are inventors on an NIH patent application that is
based on the work presented herein. D.W.S. was a consultant for Abbvie, Cel-
gene, and Janssen; received research funding from NanoString Technologies,
Janssen, and Roche/Genentech. Authors are included on additional patents,
some of which are licensed by NanoString Technologies.
Received: October 19, 2019
Revised: January 3, 2020
Accepted: March 16, 2020
Published: April 13, 2020
REFERENCES
Agarwal, R., Chan, Y.C., Tam, C.S., Hunter, T., Vassiliadis, D., Teh, C.E.,
Thijssen, R., Yeh, P., Wong, S.Q., Ftouni, S., et al. (2019). Dynamic molecular
monitoring reveals that SWI-SNF mutations mediate resistance to ibrutinib
plus venetoclax in mantle cell lymphoma. Nat. Med. 25, 119–129.
Alizadeh, A.A., Eisen, M.B., Davis, R.E., Ma, C., Lossos, I.S., Rosenwald, A.,
Boldrick, J.C., Sabet, H., Tran, T., Yu, X., et al. (2000). Distinct types of diffuse
large B cell lymphoma identified by gene expression profiling. Nature 403,
503–511.
Amin, N.A., Seymour, E., Saiya-Cork, K., Parkin, B., Shedden, K., and Malek,
S.N. (2016). A quantitative analysis of subclonal and clonal gene mutations
before and after therapy in chronic lymphocytic leukemia. Clin. Cancer Res.
22, 4525–4535.
Baeuerle, P.A., and Baltimore, D. (1988). I kappa B: a specific inhibitor of the
NF-kappa B transcription factor. Science 242, 540–546.
Bea, S., Colomo, L., Lopez-Guillermo, A., Salaverria, I., Puig, X., Pinyol, M.,
Rives, S., Montserrat, E., and Campo, E. (2004). Clinicopathologic significance
and prognostic value of chromosomal imbalances in diffuse large B cell lym-
phomas. J. Clin. Oncol. 22, 3498–3506.
Bea, S., Valdes-Mas, R., Navarro, A., Salaverria, I., Martin-Garcia, D., Jares,
P., Gine, E., Pinyol, M., Royo, C., Nadeu, F., et al. (2013). Landscape of somatic
mutations and clonal evolution in mantle cell lymphoma. Proc. Natl. Acad. Sci.
U S A 110, 18250–18255.
Bolotin, D.A., Poslavsky, S., Davydov, A.N., Frenkel, F.E., Fanchi, L.,
Zolotareva, O.I., Hemmers, S., Putintseva, E.V., Obraztsova, A.S., Shugay,
M., et al. (2017). Antigen receptor repertoire profiling from RNA-seq data.
Nat. Biotechnol. 35, 908–911.
Bouska, A., Bi, C., Lone, W., Zhang, W., Kedwaii, A., Heavican, T., Lachel,
C.M., Yu, J., Ferro, R., Eldorghamy, N., et al. (2017a). Adult high-grade B
cell lymphoma with Burkitt lymphoma signature: genomic features and poten-
tial therapeutic targets. Blood 130, 1819–1831.
Bouska, A., Zhang, W., Gong, Q., Iqbal, J., Scuto, A., Vose, J., Ludvigsen, M.,
Fu, K., Weisenburger, D.D., Greiner, T.C., et al. (2017b). Combined copy num-
ber and mutation analysis identifies oncogenic pathways associated with
transformation of follicular lymphoma. Leukemia 31, 83–91.
Braggio, E., Van Wier, S., Ojha, J., McPhail, E., Asmann, Y.W., Egan, J., da
Silva, J.A., Schiff, D., Lopes, M.B., Decker, P.A., et al. (2015). Genome-wide
analysis uncovers novel recurrent alterations in primary central nervous sys-
tem lymphomas. Clin. Cancer Res. 21, 3986–3994.
Bruno, A., Boisselier, B., Labreche, K., Marie, Y., Polivka, M., Jouvet, A.,
Adam, C., Figarella-Branger, D., Miquel, C., Eimer, S., et al. (2014).
Mutational analysis of primary central nervous system lymphoma.
Oncotarget 5, 5065–5075.
Cao, X.X., Li, J., Cai, H., Zhang, W., Duan, M.H., and Zhou, D.B. (2017).
Patients with primary breast and primary female genital tract diffuse large B
cell lymphoma have a high frequency of MYD88 and CD79B mutations. Ann.
Hematol. 96, 1867–1871.
Cardinez, C., Miraghazadeh, B., Tanita, K., da Silva, E., Hoshino, A., Okada, S.,
Chand, R., Asano, T., Tsumura, M., Yoshida, K., et al. (2018). Gain-of-function
IKBKB mutation causes human combined immune deficiency. J. Exp. Med.
215, 2715–2724.
Celeste, A., Petersen, S., Romanienko, P.J., Fernandez-Capetillo, O., Chen,
H.T., Sedelnikova, O.A., Reina-San-Martin, B., Coppola, V., Meffre, E.,
Difilippantonio, M.J., et al. (2002). Genomic instability in mice lacking histone
H2AX. Science 296, 922–927.
Ceribelli, M., Kelly, P.N., Shaffer, A.L., Wright, G.W., Xiao, W., Yang, Y.,
Mathews Griner, L.A., Guha, R., Shinn, P., Keller, J.M., et al. (2014).
Blockade of oncogenic IkappaB kinase activity in diffuse large B cell lym-
phoma by bromodomain and extraterminal domain protein inhibitors. Proc.
Natl. Acad. Sci. U S A 111, 11365–11370.
Challa-Malladi, M., Lieu, Y.K., Califano, O., Holmes, A.B., Bhagat, G., Murty,
V.V., Dominguez-Sola, D., Pasqualucci, L., and Dalla-Favera, R. (2011).
Combined genetic inactivation of beta2-Microglobulin and CD58 reveals
frequent escape from immune recognition in diffuse large B cell lymphoma.
Cancer Cell 20, 728–740.
Chapuy, B., Roemer, M.G., Stewart, C., Tan, Y., Abo, R.P., Zhang, L., Dunford,
A.J., Meredith, D.M., Thorner, A.R., Jordanova, E.S., et al. (2016). Targetable
genetic features of primary testicular and primary central nervous system lym-
phomas. Blood 127, 869–881.
Chapuy, B., Stewart, C., Dunford, A.J., Kim, J., Kamburov, A., Redd, R.A.,
Lawrence, M.S., Roemer, M.G.M., Li, A.J., Ziepert, M., et al. (2018).
Molecular subtypes of diffuse large B cell lymphoma are associated with
distinct pathogenic mechanisms and outcomes. Nat. Med. 24, 679–690.
Clipson, A., Wang, M., de Leval, L., Ashton-Key, M., Wotherspoon, A.,
Vassiliou, G., Bolli, N., Grove, C., Moody, S., Escudero-Ibarz, L., et al.
(2015). KLF2 mutation is the most frequent somatic change in splenic marginal
zone lymphoma and identifies a subset with distinct genotype. Leukemia 29,
1177–1185.
Crescenzo, R., Abate, F., Lasorsa, E., Tabbo, F., Gaudiano, M., Chiesa, N., Di
Giacomo, F., Spaccarotella, E., Barbarossa, L., Ercole, E., et al. (2015).
Convergent mutations and kinase fusions lead to oncogenic STAT3 activation
in anaplastic large cell lymphoma. Cancer Cell 27, 516–532.
Dang, C.V. (2013). MYC, metabolism, cell growth, and tumorigenesis. Cold
Spring Harbor Perspect. Med. 3, https://doi.org/10.1101/cshperspect.
a014217.
Dave, S.S., Fu, K., Wright, G.W., Lam, L.T., Kluin, P., Boerma, E.J., Greiner,
T.C., Weisenburger, D.D., Rosenwald, A., Ott, G., et al. (2006). Molecular diag-
nosis of Burkitt’s lymphoma. N. Engl. J. Med. 354, 2431–2442.
Davis, R.E., Brown, K.D., Siebenlist, U., and Staudt, L.M. (2001). Constitutive
nuclear factor kappa B activity is required for survival of activated B cell-like
diffuse large B cell lymphoma cells. J. Exp. Med. 194, 1861–1874.
Davis, R.E., Ngo, V.N., Lenz, G., Tolar, P., Young, R.M., Romesser, P.B.,
Kohlhammer, H., Lamy, L., Zhao, H., Yang, Y., et al. (2010). Chronic active
B cell-receptor signalling in diffuse large B cell lymphoma. Nature 463, 88–92.
Denzin, L.K., and Cresswell, P. (1995). HLA-DM induces CLIP dissociation
from MHC class II alpha beta dimers and facilitates peptide loading. Cell 82,
155–165.
Di Cristofano, A. (2017). SGK1: the dark side of PI3K signaling. Curr. Top. Dev.
Biol. 123, 49–71.
Dogan, I., Bertocci, B., Vilmont, V., Delbos, F., Megret, J., Storck, S., Reynaud,
C.A., and Weill, J.C. (2009). Multiple layers of B cell memory with different
effector functions. Nat. Immunol. 10, 1292–1299.
Dominguez, P.M., Ghamlouch, H., Rosikiewicz, W., Kumar, P., Beguelin, W.,
Fontan, L., Rivas, M.A., Pawlikowska, P., Armand, M., Mouly, E., et al.
(2018). TET2 deficiency causes germinal center hyperplasia, impairs plasma
cell differentiation, and promotes B cell lymphomagenesis. Cancer Discov.
8, 1632–1653.
Dominguez-Sola, D., Victora, G.D., Ying, C.Y., Phan, R.T., Saito, M.,
Nussenzweig, M.C., and Dalla-Favera, R. (2012). The proto-oncogene MYC
is required for selection in the germinal center and cyclic reentry. Nat.
Immunol. 13, 1083–1091.
Ducharme, O., Beylot-Barry, M., Pham-Ledard, A., Bohers, E., Viailly, P.J.,
Bandres, T., Faur, N., Frison, E., Vergier, B., Jardin, F., et al. (2019).
Mutations of the B cell receptor pathway confer chemoresistance in primary
cutaneous diffuse large B cell lymphoma leg-type. J. Invest. Dermatol. 139,
2334–2342.e8.
Ennishi, D., Healy, S., Bashashati, A., Saberi, S., Hother, C., Mottok, A., Chan,
F.C., Chong, L., Abraham, L., Kridel, R., et al. (2020). TMEM30A loss-of-func-
tion mutations drive lymphomagenesis and confer therapeutically exploitable
vulnerability in B cell lymphoma. Nat. Med. https://doi.org/10.1038/s41591-
020-0757-z.
Ennishi, D., Jiang, A., Boyle, M., Collinge, B., Grande, B.M., Ben-Neriah, S.,
Rushton, C., Tang, J., Thomas, N., Slack, G.W., et al. (2019a). Double-hit
gene expression signature defines a distinct subgroup of germinal center B
cell-like diffuse large B cell lymphoma. J. Clin. Oncol. 37, 190–201.
Ennishi, D., Takata, K., Beguelin, W., Duns, G., Mottok, A., Farinha, P.,
Bashashati, A., Saberi, S., Boyle, M., Meissner, B., et al. (2019b). Molecular
and genetic characterization of MHC deficiency identifies EZH2 as therapeutic
target for enhancing immune recognition. Cancer Discov. 9, 546–563.
Evan, G.I., Wyllie, A.H., Gilbert, C.S., Littlewood, T.D., Land, H., Brooks, M.,
Waters, C.M., Penn, L.Z., and Hancock, D.C. (1992). Induction of apoptosis
in fibroblasts by c-myc protein. Cell 69, 119–128.
Flossbach, L., Antoneag, E., Buck, M., Siebert, R., Mattfeldt, T., Moller, P., and
Barth, T.F. (2011). BCL6 gene rearrangement and protein expression are asso-
ciated with large cell presentation of extranodal marginal zone B cell lym-
phoma of mucosa-associated lymphoid tissue. Int. J. Cancer 129, 70–77.
Fontanilles, M., Marguet, F., Bohers, E., Viailly, P.J., Dubois, S., Bertrand, P.,
Camus, V., Mareschal, S., Ruminy, P., Maingonnat, C., et al. (2017). Non-inva-
sive detection of somatic mutations using next-generation sequencing in pri-
mary central nervous system lymphoma. Oncotarget 8, 48157–48168.
Franco, F., Gonzalez-Rincon, J., Lavernia, J., Garcia, J.F., Martin, P., Bellas,
C., Piris, M.A., Pedrosa, L., Miramon, J., Gomez-Codina, J., et al. (2017).
Mutational profile of primary breast diffuse large B cell lymphoma.
Oncotarget 8, 102888–102897.
Fukumura, K., Kawazu, M., Kojima, S., Ueno, T., Sai, E., Soda, M., Ueda, H.,
Yasuda, T., Yamaguchi, H., Lee, J., et al. (2016). Genomic characterization
of primary central nervous system lymphoma. Acta Neuropathol. 131,
865–875.
Cancer Cell 37, 551–568, April 13, 2020 565
Ganapathi, K.A., Jobanputra, V., Iwamoto, F., Jain, P., Chen, J., Cascione, L.,
Nahum, O., Levy, B., Xie, Y., Khattar, P., et al. (2016). The genetic landscape of
dural marginal zone lymphomas. Oncotarget 7, 43052–43061.
Goy, A., Ramchandren, R., Ghosh, N., Munoz, J., Morgan, D.S., Dang, N.H.,
Knapp, M., Delioukina, M., Kingsley, E., Ping, J., et al. (2019). Ibrutinib plus le-
nalidomide and rituximab has promising activity in relapsed/refractory non-
germinal center B cell-like DLBCL. Blood 134, 1024–1036.
Green, M.R., Gentles, A.J., Nair, R.V., Irish, J.M., Kihira, S., Liu, C.L., Kela, I.,
Hopmans, E.S., Myklebust, J.H., Ji, H., et al. (2013). Hierarchy in somatic mu-
tations arising during genomic evolution and progression of follicular lym-
phoma. Blood 121, 1604–1611.
Green, M.R., Kihira, S., Liu, C.L., Nair, R.V., Salari, R., Gentles, A.J., Irish, J.,
Stehr, H., Vicente-Duenas, C., Romero-Camarero, I., et al. (2015). Mutations
in early follicular lymphoma progenitors are associated with suppressed anti-
gen presentation. Proc. Natl. Acad. Sci. U S A 112, E1116–E1125.
Grommes, C., Pastore, A., Palaskas, N., Tang, S.S., Campos, C., Schartz, D.,
Codega, P., Nichol, D., Clark, O., Hsieh, W.Y., et al. (2017). Ibrutinib unmasks
critical role of bruton tyrosine kinase in primary CNS lymphoma. Cancer
Discov. 7, 1018–1029.
Hartmann, S., Schuhmacher, B., Rausch, T., Fuller, L., Doring, C., Weniger, M.,
Lollies, A., Weiser, C., Thurner, L., Rengstl, B., et al. (2016). Highly recurrent
mutations of SGK1, DUSP2 and JUNB in nodular lymphocyte predominant
Hodgkin lymphoma. Leukemia 30, 844–853.
Hattori, K., Sakata-Yanagimoto, M., Kusakabe, M., Nanmoku, T., Suehara, Y.,
Matsuoka, R., Noguchi, M., Yokoyama, Y., Kato, T., Kurita, N., et al. (2019).
Genetic evidence implies that primary and relapsed tumors arise from com-
mon precursor cells in primary central nervous system lymphoma. Cancer
Sci. 110, 401–407.
Hattori, K., Sakata-Yanagimoto, M., Okoshi, Y., Goshima, Y., Yanagimoto, S.,
Nakamoto-Matsubara, R., Sato, T., Noguchi, M., Takano, S., Ishikawa, E.,
et al. (2017). MYD88 (L265P) mutation is associated with an unfavourable
outcome of primary central nervous system lymphoma. Br. J. Haematol.
177, 492–494.
Havranek, O., Xu, J., Kohrer, S., Wang, Z., Becker, L., Comer, J.M.,
Henderson, J., Ma, W., Man Chun Ma, J., Westin, J.R., et al. (2017). Tonic B
cell receptor signaling in diffuse large B cell lymphoma. Blood 130, 995–1006.
Heise, N., De Silva, N.S., Silva, K., Carette, A., Simonetti, G., Pasparakis, M.,
and Klein, U. (2014). Germinal center B cell maintenance and differentiation
are controlled by distinct NF-kappaB transcription factor subunits. J. Exp.
Med. 211, 2103–2118.
Hellmuth, J.C., Louissaint, A., Jr., Szczepanowski, M., Haebe, S., Pastore, A.,
Alig, S., Staiger, A.M., Hartmann, S., Kridel, R., Ducar, M.D., et al. (2018).
Duodenal-type and nodal follicular lymphomas differ by their immune microen-
vironment rather than their mutation profiles. Blood 132, 1695–1702.
Hickmann, A.K., Frick, M., Hadaschik, D., Battke, F., Bittl, M., Ganslandt, O.,
Biskup, S., and Docker, D. (2019). Molecular tumor analysis and liquid biopsy:
a feasibility investigation analyzing circulating tumor DNA in patients with cen-
tral nervous system lymphomas. BMC Cancer 19, 192.
Hilton, L., Tang, J., Ben-Neriah, S., Alcaide, M., Jiang, A., Grande, B.M.,
Rushton, C., Boyle, M., Meissner, B., Scott, D., and Morin, R.D. (2019). The
double hit signature identifies double-hit diffuse large B cell lymphoma with
genetic events cryptic to FISH. Blood 134, 1528–1532.
Hodson, D.J., Shaffer, A.L., Xiao, W., Wright, G.W., Schmitz, R., Phelan, J.D.,
Yang, Y., Webster, D.E., Rui, L., Kohlhammer, H., et al. (2016). Regulation of
normal B cell differentiation and malignant B cell survival by OCT2. Proc.
Natl. Acad. Sci. U S A 113, E2039–E2046.
Hyeon, J., Lee, B., Shin, S.H., Yoo, H.Y., Kim, S.J., Kim, W.S., Park, W.Y., and
Ko, Y.H. (2018). Targeted deep sequencing of gastric marginal zone lym-
phoma identified alterations of TRAF3 and TNFAIP3 that were mutually exclu-
sive for MALT1 rearrangement. Mod. Pathol. 31, 1418–1428.
Johansson, P., Klein-Hitpass, L., Grabellus, F., Arnold, G., Klapper, W.,
Pfortner, R., Duhrsen, U., Eckstein, A., Durig, J., and Kuppers, R. (2016).
Recurrent mutations in NF-kappaB pathway components, KMT2D, and
NOTCH1/2 in ocular adnexal MALT-type marginal zone lymphomas.
Oncotarget 7, 62627–62639.
566 Cancer Cell 37, 551–568, April 13, 2020
Kelly, P.N., Romero, D.L., Yang, Y., Shaffer, A.L., 3rd, Chaudhary, D.,
Robinson, S., Miao, W., Rui, L., Westlin, W.F., Kapeller, R., and Staudt, L.M.
(2015). Selective interleukin-1 receptor-associated kinase 4 inhibitors for the
treatment of autoimmune disorders and lymphoid malignancy. J. Exp. Med.
212, 2189–2201.
Kiel, M.J., Velusamy, T., Betz, B.L., Zhao, L., Weigelin, H.G., Chiang, M.Y.,
Huebner-Chan, D.R., Bailey, N.G., Yang, D.T., Bhagat, G., et al. (2012).
Whole-genome sequencing identifies recurrent somatic NOTCH2 mutations
in splenic marginal zone lymphoma. J. Exp. Med. 209, 1553–1565.
Kraan, W., van Keimpema, M., Horlings, H.M., Schilder-Tol, E.J., Oud, M.E.,
Noorduyn, L.A., Kluin, P.M., Kersten, M.J., Spaargaren, M., and Pals, S.T.
(2014). High prevalence of oncogenic MYD88 and CD79B mutations in primary
testicular diffuse large B cell lymphoma. Leukemia 28, 719–720.
Krysiak, K., Gomez, F., White, B.S., Matlock, M., Miller, C.A., Trani, L., Fronick,
C.C., Fulton, R.S., Kreisel, F., Cashen, A.F., et al. (2017). Recurrent somatic
mutations affecting B cell receptor signaling pathway genes in follicular lym-
phoma. Blood 129, 473–483.
Kuppers, R., Duhrsen, U., and Hansmann, M.L. (2014). Pathogenesis, diag-
nosis, and treatment of composite lymphomas. Lancet Oncol. 15, e435–e446.
Lam, L.T., Wright, G., Davis, R.E., Lenz, G., Farinha, P., Dang, L., Chan, J.W.,
Rosenwald, A., Gascoyne, R.D., and Staudt, L.M. (2008). Cooperative
signaling through the signal transducer and activator of transcription 3 and nu-
clear factor-{kappa}B pathways in subtypes of diffuse large B cell lymphoma.
Blood 111, 3701–3713.
Landau, D.A., Sun, C., Rosebrock, D., Herman, S.E.M., Fein, J., Sivina, M.,
Underbayev, C., Liu, D., Hoellenriegel, J., Ravichandran, S., et al. (2017).
The evolutionary landscape of chronic lymphocytic leukemia treated with ibru-
tinib targeted therapy. Nat. Commun. 8, 2185.
Landau, D.A., Tausch, E., Taylor-Weiner, A.N., Stewart, C., Reiter, J.G., Bahlo,
J., Kluth, S., Bozic, I., Lawrence, M., Bottcher, S., et al. (2015). Mutations
driving CLL and their evolution in progression and relapse. Nature 526,
525–530.
Lenz, G., Wright, G., Dave, S.S., Xiao, W., Powell, J., Zhao, H., Xu, W., Tan, B.,
Goldschmidt, N., Iqbal, J., et al. (2008a). Stromal gene signatures in large-B
cell lymphomas. N. Engl. J. Med. 359, 2313–2323.
Lenz, G., Wright, G.W., Emre, N.C., Kohlhammer, H., Dave, S.S., Davis, R.E.,
Carty, S., Lam, L.T., Shaffer, A.L., Xiao, W., et al. (2008b). Molecular subtypes
of diffuse large B cell lymphoma arise by distinct genetic pathways. Proc. Natl.
Acad. Sci. U S A 105, 13520–13525.
Linossi, E.M., and Nicholson, S.E. (2015). Kinase inhibition, competitive bind-
ing and proteasomal degradation: resolving the molecular function of the sup-
pressor of cytokine signaling (SOCS) proteins. Immunol. Rev. 266, 123–133.
Lionakis, M.S., Dunleavy, K., Roschewski, M., Widemann, B.C., Butman, J.A.,
Schmitz, R., Yang, Y., Cole, D.E., Melani, C., Higham, C.S., et al. (2017).
Inhibition of B cell receptor signaling by ibrutinib in primary CNS lymphoma.
Cancer Cell 31, 833–843 e5.
Ljungstrom, V., Cortese, D., Young, E., Pandzic, T., Mansouri, L., Plevova, K.,
Ntoufa, S., Baliakas, P., Clifford, R., Sutton, L.A., et al. (2016). Whole-exome
sequencing in relapsing chronic lymphocytic leukemia: clinical impact of
recurrent RPS15 mutations. Blood 127, 1007–1016.
Love, C., Sun, Z., Jima, D., Li, G., Zhang, J., Miles, R., Richards, K.L., Dunphy,
C.H., Choi, W.W., Srivastava, G., et al. (2012). The genetic landscape of muta-
tions in Burkitt lymphoma. Nat. Genet. 44, 1321–1325.
Lu, D., Liu, L., Ji, X., Gao, Y., Chen, X., Liu, Y., Liu, Y., Zhao, X., Li, Y., Li, Y.,
et al. (2015). The phosphatase DUSP2 controls the activity of the transcription
activator STAT3 and regulates TH17 differentiation. Nat. Immunol. 16,
1263–1273.
Lu, E., Wolfreys, F.D., Muppidi, J.R., Xu, Y., and Cyster, J.G. (2019).
S-Geranylgeranyl-L-glutathione is a ligand for human B cell-confinement re-
ceptor P2RY8. Nature 567, 244–248.
Malecka, A., Troen, G., Tierens, A., Ostlie, I., Malecki, J., Randen, U., Wang, J.,
Berentsen, S., Tjonnfjord, G.E., and Delabie, J.M.A. (2018). Frequent somatic
mutations of KMT2D (MLL2) and CARD11 genes in primary cold agglutinin dis-
ease. Br. J. Haematol. 183, 838–842.
Mareschal, S., Pham-Ledard, A., Viailly, P.J., Dubois, S., Bertrand, P.,
Maingonnat, C., Fontanilles, M., Bohers, E., Ruminy, P., Tournier, I., et al.
(2017). Identification of somatic mutations in primary cutaneous diffuse large
B cell lymphoma, leg type by massive parallel sequencing. J. Invest.
Dermatol. 137, 1984–1994.
Martinez, N., Almaraz, C., Vaque, J.P., Varela, I., Derdak, S., Beltran, S.,
Mollejo, M., Campos-Martin, Y., Agueda, L., Rinaldi, A., et al. (2014). Whole-
exome sequencing in splenic marginal zone lymphoma reveals mutations in
genes involved in marginal zone differentiation. Leukemia 28, 1334–1340.
Mathews Griner, L.A., Guha, R., Shinn, P., Young, R.M., Keller, J.M., Liu, D.,
Goldlust, I.S., Yasgar, A., McKnight, C., Boxer, M.B., et al. (2014). High-
throughput combinatorial screening identifies drugs that cooperate with ibru-
tinib to kill activated B cell-like diffuse large B cell lymphoma cells. Proc. Natl.
Acad. Sci. U S A 111, 2349–2354.
Matsushita, H., Vesely, M.D., Koboldt, D.C., Rickert, C.G., Uppaluri, R.,
Magrini, V.J., Arthur, C.D., White, J.M., Chen, Y.S., Shea, L.K., et al. (2012).
Cancer exome analysis reveals a T cell-dependent mechanism of cancer im-
munoediting. Nature 482, 400–404.
Milpied, P., Cervera-Marzal, I., Mollichella, M.L., Tesson, B., Brisou, G.,
Traverse-Glehen, A., Salles, G., Spinelli, L., and Nadel, B. (2018). Human
germinal center transcriptional programs are de-synchronized in B cell lym-
phoma. Nat. Immunol. 19, 1013–1024.
Mintz, M.A., Felce, J.H., Chou, M.Y., Mayya, V., Xu, Y., Shui, J.W., An, J., Li, Z.,
Marson, A., Okada, T., et al. (2019). The HVEM-BTLA axis restrains T cell help
to germinal center B cells and functions as a cell-extrinsic suppressor in lym-
phomagenesis. Immunity 51, 310–323.e7.
Mittal, S.K., and Roche, P.A. (2015). Suppression of antigen presentation by
IL-10. Curr. Opin. Immunol. 34, 22–27.
Mlynarczyk, C., Fontan, L., and Melnick, A. (2019). Germinal center-derived
lymphomas: the darkest side of humoral immunity. Immunol. Rev. 288,
214–239.
Monti, S., Chapuy, B., Takeyama, K., Rodig, S.J., Hao, Y., Yeda, K.T.,
Inguilizian, H., Mermel, C., Currie, T., Dogan, A., et al. (2012). Integrative anal-
ysis reveals an outcome-associated and targetable pattern of p53 and cell cy-
cle deregulation in diffuse large B cell lymphoma. Cancer Cell 22, 359–372.
Muppidi, J.R., Schmitz, R., Green, J.A., Xiao, W., Larsen, A.B., Braun, S.E., An,
J., Xu, Y., Rosenwald, A., Ott, G., et al. (2014). Loss of signalling via Galpha13
in germinal centre B cell-derived lymphoma. Nature 516, 254–258.
Nakamura, T., Tateishi, K., Niwa, T., Matsushita, Y., Tamura, K., Kinoshita, M.,
Tanaka, K., Fukushima, S., Takami, H., Arita, H., et al. (2016). Recurrent muta-
tions of CD79B and MYD88 are the hallmark of primary central nervous system
lymphomas. Neuropathol. Appl. Neurobiol. 42, 279–290.
Ngo, V.N., Young, R.M., Schmitz, R., Jhavar, S., Xiao, W., Lim, K.H.,
Kohlhammer, H., Xu, W., Yang, Y., Zhao, H., et al. (2011). Oncogenically active
MYD88 mutations in human lymphoma. Nature 470, 115–119.
Nogai, H., Wenzel, S.S., Hailfinger, S., Grau, M., Kaergel, E., Seitz, V., Wollert-
Wulf, B., Pfeifer, M., Wolf, A., Frick, M., et al. (2013). IkappaB-zeta controls the
constitutive NF-kappaB target gene network and survival of ABC DLBCL.
Blood 122, 2242–2250.
Okosun, J., Bodor, C., Wang, J., Araf, S., Yang, C.Y., Pan, C., Boller, S.,
Cittaro, D., Bozek, M., Iqbal, S., et al. (2014). Integrated genomic analysis iden-
tifies recurrent mutations and evolution patterns driving the initiation and pro-
gression of follicular lymphoma. Nat. Genet. 46, 176–181.
Okosun, J., Wolfson, R.L., Wang, J., Araf, S., Wilkins, L., Castellano, B.M.,
Escudero-Ibarz, L., Al Seraihi, A.F., Richter, J., Bernhart, S.H., et al. (2016).
Recurrent mTORC1-activating RRAGC mutations in follicular lymphoma.
Nat. Genet. 48, 183–188.
Parry, M., Rose-Zerilli, M.J., Gibson, J., Ennis, S., Walewska, R., Forster, J.,
Parker, H., Davis, Z., Gardiner, A., Collins, A., et al. (2013). Whole exome
sequencing identifies novel recurrently mutated genes in patients with splenic
marginal zone lymphoma. PLoS One 8, e83244.
Parry, M., Rose-Zerilli, M.J., Ljungstrom, V., Gibson, J., Wang, J., Walewska,
R., Parker, H., Parker, A., Davis, Z., Gardiner, A., et al. (2015). Genetics and
prognostication in splenic marginal zone lymphoma: revelations from deep
sequencing. Clin. Cancer Res. 21, 4174–4183.
Pasqualucci, L., and Dalla-Favera, R. (2018). Genetics of diffuse large B cell
lymphoma. Blood 131, 2307–2319.
Pasqualucci, L., Khiabanian, H., Fangazio, M., Vasishtha, M., Messina, M.,
Holmes, A.B., Ouillette, P., Trifonov, V., Rossi, D., Tabbo, F., et al. (2014).
Genetics of follicular lymphoma transformation. Cell Rep. 6, 130–140.
Pfeifer, M., Grau, M., Lenze, D., Wenzel, S.S., Wolf, A., Wollert-Wulf, B., Dietze,
K., Nogai, H., Storek, B., Madle, H., et al. (2013). PTEN loss defines a PI3K/AKT
pathway-dependent germinal center subtype of diffuse large B cell lymphoma.
Proc. Natl. Acad. Sci. U S A 110, 12420–12425.
Phelan, J.D., Young, R.M., Webster, D.E., Roulland, S., Wright, G.W.,
Kasbekar, M., Shaffer, A.L., 3rd, Ceribelli, M., Wang, J.Q., Schmitz, R., et al.
(2018). A multiprotein supercomplex controlling oncogenic signalling in lym-
phoma. Nature 560, 387–391.
Phillips, T.J., Forero-Torres, A., Sher, T., Diefenbach, C.S., Johnston, P.,
Talpaz, M., Pulini, J., Zhou, L., Scherle, P., Chen, X., and Barr, P.M. (2018).
Phase 1 study of the PI3Kdelta inhibitor INCB040093 +/- JAK1 inhibitor itaci-
tinib in relapsed/refractory B cell lymphoma. Blood 132, 293–306.
Pillonel, V., Juskevicius, D., Ng, C.K.Y., Bodmer, A., Zettl, A., Jucker, D.,
Dirnhofer, S., and Tzankov, A. (2018). High-throughput sequencing of nodal
marginal zone lymphomas identifies recurrent BRAF mutations. Leukemia
32, 2412–2426.
Puente, X.S., Bea, S., Valdes-Mas, R., Villamor, N., Gutierrez-Abril, J., Martin-
Subero, J.I., Munar, M., Rubio-Perez, C., Jares, P., Aymerich, M., et al. (2015).
Non-coding recurrent mutations in chronic lymphocytic leukaemia. Nature
526, 519–524.
Quesada, V., Conde, L., Villamor, N., Ordonez, G.R., Jares, P., Bassaganyas,
L., Ramsay, A.J., Bea, S., Pinyol, M., Martinez-Trillos, A., et al. (2011). Exome
sequencing identifies recurrent mutations of the splicing factor SF3B1 gene in
chronic lymphocytic leukemia. Nat. Genet. 44, 47–52.
Richter, J., Schlesner, M., Hoffmann, S., Kreuz, M., Leich, E., Burkhardt, B.,
Rosolowski, M., Ammerpohl, O., Wagener, R., Bernhart, S.H., et al. (2012).
Recurrent mutation of the ID3 gene in Burkitt lymphoma identified by inte-
grated genome, exome and transcriptome sequencing. Nat. Genet. 44,
1316–1320.
Rosenwald, A., Wright, G., Chan, W.C., Connors, J.M., Campo, E., Fisher, R.I.,
Gascoyne, R.D., Muller-Hermelink, H.K., Smeland, E.B., Giltnane, J.M., et al.
(2002). The use of molecular profiling to predict survival after chemotherapy
for diffuse large-B cell lymphoma. N. Engl. J. Med. 346, 1937–1947.
Rossi, D., Trifonov, V., Fangazio, M., Bruscaggin, A., Rasi, S., Spina, V., Monti,
S., Vaisitti, T., Arruga, F., Fama, R., et al. (2012). The coding genome of splenic
marginal zone lymphoma: activation of NOTCH2 and other pathways regu-
lating marginal zone development. J. Exp. Med. 209, 1537–1551.
Rui, L., Drennan, A.C., Ceribelli, M., Zhu, F., Wright, G.W., Huang, D.W., Xiao,
W., Li, Y., Grindle, K.M., Lu, L., et al. (2016). Epigenetic gene regulation by
Janus kinase 1 in diffuse large B cell lymphoma. Proc. Natl. Acad. Sci. U S A
113, E7260–E7267.
Saito, T., Chiba, S., Ichikawa, M., Kunisato, A., Asai, T., Shimizu, K.,
Yamaguchi, T., Yamamoto, G., Seo, S., Kumano, K., et al. (2003). Notch2 is
preferentially expressed in mature B cells and indispensable for marginal
zone B lineage development. Immunity 18, 675–685.
Scherer, F., Kurtz, D.M., Newman, A.M., Stehr, H., Craig, A.F., Esfahani, M.S.,
Lovejoy, A.F., Chabon, J.J., Klass, D.M., Liu, C.L., et al. (2016). Distinct biolog-
ical subtypes and patterns of genome evolution in lymphoma revealed by
circulating tumor DNA. Sci. Transl. Med. 8, 364ra155.
Schmitz, R., Wright, G.W., Huang, D.W., Johnson, C.A., Phelan, J.D., Wang,
J.Q., Roulland, S., Kasbekar, M., Young, R.M., Shaffer, A.L., et al. (2018).
Genetics and pathogenesis of diffuse large B cell lymphoma. N. Engl. J.
Med. 378, 1396–1407.
Schmitz, R., Young, R.M., Ceribelli, M., Jhavar, S., Xiao, W., Zhang, M., Wright,
G., Shaffer, A.L., Hodson, D.J., Buras, E., et al. (2012). Burkitt lymphoma path-
ogenesis and therapeutic targets from structural and functional genomics.
Nature 490, 116–120.
Cancer Cell 37, 551–568, April 13, 2020 567
Schrader, A.M.R., Jansen, P.M., Willemze, R., Vermeer, M.H., Cleton-Jansen,
A.M., Somers, S.F., Veelken, H., van Eijk, R., Kraan, W., Kersten, M.J., et al.
(2018). High prevalence of MYD88 and CD79B mutations in intravascular large
B cell lymphoma. Blood 131, 2086–2089.
Schuhmacher, B., Bein, J., Rausch, T., Benes, V., Tousseyn, T., Vornanen, M.,
Ponzoni, M., Thurner, L., Gascoyne, R., Steidl, C., et al. (2019). JUNB, DUSP2,
SGK1, SOCS1 and CREBBP are frequently mutated in T cell/histiocyte-rich
large B cell lymphoma. Haematologica 104, 330–337.
Sciammas, R., Shaffer, A.L., Schatz, J.H., Zhao, H., Staudt, L.M., and Singh, H.
(2006). Graded expression of interferon regulatory factor-4 coordinates iso-
type switching with plasma cell differentiation. Immunity 25, 225–236.
Scott, D.W., Mottok, A., Ennishi, D., Wright, G.W., Farinha, P., Ben-Neriah, S.,
Kridel, R., Barry, G.S., Hother, C., Abrisqueta, P., et al. (2015). Prognostic sig-
nificance of diffuse large B cell lymphoma cell of origin determined by digital
gene expression in formalin-fixed paraffin-embedded tissue biopsies.
J. Clin. Oncol. 33, 2848–2856.
Sha, C., Barrans, S., Cucco, F., Bentley, M.A., Care, M.A., Cummin, T.,
Kennedy, H., Thompson, J.S., Uddin, R., Worrillow, L., et al. (2019).
Molecular high-grade B cell lymphoma: defining a poor-risk group that re-
quires different approaches to therapy. J. Clin. Oncol. 37, 202–212.
Shaffer, A.L., Wright, G., Yang, L., Powell, J., Ngo, V., Lamy, L., Lam, L.T.,
Davis, R.E., and Staudt, L.M. (2006). A library of gene expression signatures
to illuminate normal and pathological lymphoid biology. Immunological Rev.
210, 67–85.
Shechter, R., London, A., and Schwartz, M. (2013). Orchestrated leukocyte
recruitment to immune-privileged sites: absolute barriers versus educational
gates. Nat. Rev. Immunol. 13, 206–218.
Shin, S.H., Kim, Y.J., Lee, D., Cho, D., Ko, Y.H., Cho, J., Park, W.Y., Park, D.,
Kim, S.J., and Kim, W.S. (2019). Analysis of circulating tumor DNA by targeted
ultra-deep sequencing across various non-Hodgkin lymphoma subtypes.
Leuk. Lymphoma 60, 2237–2246.
Singh, M., Jackson, K.J.L., Wang, J.J., Schofield, P., Field, M.A., Koppstein,
D., Peters, T.J., Burnett, D.L., Rizzetto, S., Nevoltris, D., et al. (2020).
Lymphoma driver mutations in the pathogenic evolution of an iconic human
autoantibody. Cell 180, 878–894.e19.
Spina, V., Khiabanian, H., Messina, M., Monti, S., Cascione, L., Bruscaggin, A.,
Spaccarotella, E., Holmes, A.B., Arcaini, L., Lucioni, M., et al. (2016). The ge-
netics of nodal marginal zone lymphoma. Blood 128, 1362–1373.
Steidl, C., Shah, S.P., Woolcock, B.W., Rui, L., Kawahara, M., Farinha, P.,
Johnson, N.A., Zhao, Y., Telenius, A., Neriah, S.B., et al. (2011). MHC class
II transactivator CIITA is a recurrent gene fusion partner in lymphoid cancers.
Nature 471, 377–381.
Suan, D., Krautler, N.J., Maag, J.L.V., Butt, D., Bourne, K., Hermes, J.R.,
Avery, D.T., Young, C., Statham, A., Elliott, M., et al. (2017). CCR6 defines
memory B cell precursors in mouse and human germinal centers, revealing
light-zone location and predominant low antigen affinity. Immunity 47, 1142–
1153.e4.
Suehara, Y., Sakata-Yanagimoto, M., Hattori, K., Nanmoku, T., Itoh, T., Kaji,
D., Yamamoto, G., Abe, Y., Narita, K., Takeuchi, M., et al. (2018). Liquid biopsy
for the identification of intravascular large B cell lymphoma. Haematologica
103, e241–e244.
Tallen, G., and Riabowol, K. (2014). Keep-ING balance: tumor suppression by
epigenetic regulation. FEBS Lett. 588, 2728–2742.
Timens, W., Visser, L., and Poppema, S. (1986). Nodular lymphocyte predom-
inance type of Hodgkin’s disease is a germinal center lymphoma. Lab. Invest.
54, 457–461.
Tsukamoto, T., Nakano, M., Sato, R., Adachi, H., Kiyota, M., Kawata, E.,
Uoshima, N., Yasukawa, S., Chinen, Y., Mizutani, S., et al. (2017). High-risk
follicular lymphomas harbour more somatic mutations including those in the
AID-motif. Sci. Rep. 7, 14039.
Vater, I., Montesinos-Rongen, M., Schlesner, M., Haake, A., Purschke, F.,
Sprute, R., Mettenmeyer, N., Nazzal, I., Nagel, I., Gutwein, J., et al. (2015).
568 Cancer Cell 37, 551–568, April 13, 2020
The mutational pattern of primary lymphoma of the central nervous system
determined by whole-exome sequencing. Leukemia 29, 677–685.
Wang, L., Lawrence, M.S., Wan, Y., Stojanov, P., Sougnez, C., Stevenson, K.,
Werner, L., Sivachenko, A., DeLuca, D.S., Zhang, L., et al. (2011). SF3B1 and
other novel cancer genes in chronic lymphocytic leukemia. N. Engl. J. Med.
365, 2497–2506.
Webster, D.E., Roulland, S., and Phelan, J.D. (2019). Protocols for CRISPR-
cas9 screening in lymphoma cell lines. Methods Mol. Biol. 1956, 337–350.
Weinstein, J.S., Herman, E.I., Lainez, B., Licona-Limon, P., Esplugues, E.,
Flavell, R., and Craft, J. (2016). TFH cells progressively differentiate to regulate
the germinal center response. Nat. Immunol. 17, 1197–1205.
Wilson, W.H., Popplewell, L.L., Phillips, T., Kimball, A.S., Chhabra, S., Ping, J.,
Neuenburg, J., Cavazos, N., Staudt, L.M., and de Vos, S. (2015a). Multicenter
phase 1b dose-escalation study of ibrutinib and lenalidomide combined with
dose-adjusted EPOCH-R in patients with relapsed/refractory DLBCL. Blood
126, 1527.
Wilson, W.H., Young, R.M., Schmitz, R., Yang, Y., Pittaluga, S., Wright, G., Lih,
C.J., Williams, P.M., Shaffer, A.L., Gerecitano, J., et al. (2015b). Targeting
B cell receptor signaling with ibrutinib in diffuse large B cell lymphoma. Nat.
Med. 21, 922–926.
Wu, C., de Miranda, N.F., Chen, L., Wasik, A.M., Mansouri, L., Jurczak, W.,
Galazka, K., Dlugosz-Danecka, M., Machaczka, M., Zhang, H., et al. (2016).
Genetic heterogeneity in primary and relapsed mantle cell lymphomas: impact
of recurrent CARD11 mutations. Oncotarget 7, 38180–38190.
Yang, Y., Shaffer, A.L., 3rd, Emre, N.C., Ceribelli, M., Zhang, M., Wright, G.,
Xiao, W., Powell, J., Platig, J., Kohlhammer, H., et al. (2012). Exploiting syn-
thetic lethality for the therapy of ABC diffuse large B cell lymphoma. Cancer
Cell 21, 723–737.
Ye, H., Remstein, E.D., Bacon, C.M., Nicholson, A.G., Dogan, A., and Du, M.Q.
(2008). Chromosomal translocations involving BCL6 in MALT lymphoma.
Haematologica 93, 145–146.
Yonese, I., Takase, H., Yoshimori, M., Onozawa, E., Tsuzura, A., Miki, T.,
Mochizuki, M., Miura, O., and Arai, A. (2019). CD79B mutations in primary vit-
reoretinal lymphoma: diagnostic and prognostic potential. Eur. J. Haematol.
102, 191–196.
Young, R.M., Phelan, J.D., Wilson, W.H., and Staudt, L.M. (2019). Pathogenic
B cell receptor signaling in lymphoid malignancies: new insights to improve
treatment. Immunological Rev. 291, 190–213.
Young, R.M., Wu, T., Schmitz, R., Dawood, M., Xiao, W., Phelan, J.D., Xu, W.,
Menard, L., Meffre, E., Chan, W.C., et al. (2015). Survival of human lymphoma
cells requires B cell receptor engagement by self-antigens. Proc. Natl. Acad.
Sci. U S A 112, 13447–13454.
Zamo, A., Pischimarov, J., Schlesner, M., Rosenstiel, P., Bomben, R., Horn,
H., Grieb, T., Nedeva, T., Lopez, C., Haake, A., et al. (2018). Differences be-
tween BCL2-break positive and negative follicular lymphoma unraveled by
whole-exome sequencing. Leukemia 32, 685–693.
Zeller, K.I., Zhao, X., Lee, C.W., Chiu, K.P., Yao, F., Yustein, J.T., Ooi, H.S.,
Orlov, Y.L., Shahab, A., Yong, H.C., et al. (2006). Global mapping of c-Myc
binding sites and target gene networks in human B cells. Proc. Natl. Acad.
Sci. U S A 103, 17834–17839.
Zhang, J., Jima, D., Moffitt, A.B., Liu, Q., Czader, M., Hsi, E.D., Fedoriw, Y.,
Dunphy, C.H., Richards, K.L., Gill, J.I., et al. (2014). The genomic landscape
of mantle cell lymphoma is related to the epigenetically determined chromatin
state of normal B cells. Blood 123, 2988–2996.
Zhou, X.A., Louissaint, A., Jr., Wenzel, A., Yang, J., Martinez-Escala, M.E.,
Moy, A.P., Morgan, E.A., Paxton, C.N., Hong, B., Andersen, E.F., et al.
(2018a). Genomic analyses identify recurrent alterations in immune evasion
genes in diffuse large B cell lymphoma, leg type. J. Invest. Dermatol. 138,
2365–2376.
Zhou, Y., Liu, W., Xu, Z., Zhu, H., Xiao, D., Su, W., Zeng, R., Feng, Y., Duan, Y.,
Zhou, J., and Zhong, M. (2018b). Analysis of genomic alteration in primary cen-
tral nervous system lymphoma and the expression of some related genes.
Neoplasia 20, 1059–1069.
Cancer Cell

Article

Dendritic Cell Paucity Leads

to Dysfunctional Immune
Surveillance in Pancreatic Cancer
Samarth Hegde,1 Varintra E. Krisnawan,1 Brett H. Herzog,1 Chong Zuo,1 Marcus A. Breden,1 Brett L. Knolhoff,1
Graham D. Hogg,1 Jack P. Tang,3 John M. Baer,1 Cedric Mpoy,3 Kyung Bae Lee,1 Katherine A. Alexander,1
Buck E. Rogers,3,7 Kenneth M. Murphy,4,5 William G. Hawkins,6,7 Ryan C. Fields,6,7 Carl J. DeSelm,3,7 Julie K. Schwarz,2,3,7
and David G. DeNardo1,4,7,8,*
1Department of Medicine, Washington University School of Medicine, St. Louis, MO 63110, USA
2Department of Cell Biology and Physiology, Washington University School of Medicine, St. Louis, MO 63110, USA
3Department of Radiation Oncology, Washington University School of Medicine, St. Louis, MO 63110, USA
4Department of Pathology and Immunology, Washington University School of Medicine, St. Louis, MO 63110, USA
5Howard Hughes Medical Institute, Washington University School of Medicine, St. Louis, MO 63110, USA
6Department of Surgery, Barnes—Jewish Hospital, St. Louis, MO 63110, USA
7Alvin J. Siteman Comprehensive Cancer Center, St. Louis, MO 63110, USA
8Lead Contact

*Correspondence: ddenardo@wustl.edu
https://doi.org/10.1016/j.ccell.2020.02.008

SUMMARY

Here, we utilized spontaneous models of pancreatic and lung cancer to examine how neoantigenicity shapes
tumor immunity and progression. As expected, neoantigen expression during lung adenocarcinoma devel-
opment leads to T cell-mediated immunity and disease restraint. By contrast, neoantigen expression in
pancreatic ductal adenocarcinoma (PDAC) results in exacerbation of a fibro-inflammatory microenvironment
that drives disease progression and metastasis. Pathogenic TH17 responses are responsible for this neoan-
tigen-induced tumor progression in PDAC. Underlying these divergent T cell responses in pancreas and lung
cancer are differences in infiltrating conventional dendritic cells (cDCs). Overcoming cDC deficiency in early-
stage PDAC leads to disease restraint, while restoration of cDC function in advanced PDAC restores tumor-
restraining immunity and enhances responsiveness to radiation therapy.

INTRODUCTION
Pancreatic ductal adenocarcinoma (PDAC) is notoriously resis-
tant to immunotherapy, including cytokine therapy, adoptive
T cell therapy and checkpoint blockade strategies (Brahmer
et al., 2012; Kunk et al., 2016; Royal et al., 2010). Failure of these
therapies has been attributed to CD8+ T cell scarcity and pro-
found immunosuppression in the PDAC microenvironment
(Beatty et al., 2015; Stromnes et al., 2014; Zhang et al., 2016).
However, recent studies have challenged this paradigm and
have revealed that many PDAC patients indeed harbor intratu-
moral T cells and potentially actionable neoantigens that can
elicit T cell responses (Bailey et al., 2016a, 2016b; Balachandran
et al., 2017; Cristescu et al., 2018; Poschke et al., 2016). To
develop therapies to revive T cell responses to neoantigens in
PDAC, it is critical to understand how this endogenous T cell
response becomes ineffectual.
The magnitude and persistence of a T cell response against a
tumor is dependent on initial priming by antigen-presenting cells.
Conventional dendritic cells (cDCs) have been recognized as

Significance

T cell-directed immunotherapies have not been effective for the majority of pancreatic cancer patients, in part due to our
limited understanding of how T cell immunity is subverted in this disease. We sought to identify mechanisms for this failure
using spontaneous mouse models. We report that endogenous antigen-specific responses in pancreatic ductal adenocar-
cinoma are aberrant due to a scarcity of dendritic cells, which favors the expansion of tumor-promoting TH17 immunity.
Restoring conventional dendritic cells (cDCs) in pancreatic cancer can enhance CD8+ T cell and TH1 activity to ultimately
help control disease. These findings expand our understanding of T cell ineffectiveness in pancreatic cancer, and propose
combinatorial strategies to modulate cDCs in conjunction with existing therapies for pancreatic cancer and similar solid
malignancies.

Cancer Cell 37, 289–307, March 16, 2020 ª 2020 Elsevier Inc. 289
critical mediators of antigen-priming and T cell activity, with
Batf3/Irf8-dependent CD103+ CD24+ cDC1s being responsible
for CD8+ cytotoxic T lymphocyte (CTL) cross-priming and Irf4-
dependent CD11b+ CD172a+ cDC2s being implicated in helper
CD4+ T cell (TH) priming (Gardner and Ruffell, 2016). In addition
to initial T cell priming, cDCs have been implicated in T cell-
dependent tumor killing and response to immunotherapies (Bin-
newies et al., 2019; de Mingo Pulido et al., 2018; Roberts et al.,
2016; Salmon et al., 2016; Spranger et al., 2017). Studies on
antigen-presenting cells (APCs) in PDAC models have focused
on tolerogenic subsets (Barilla et al., 2019; Bellone et al., 2014;
Jang et al., 2017; Ochi et al., 2012). Presently, more granular
studies of DC subsets in the PDAC context are needed to distin-
guish the differential impact of cDCs from monocytic and other
inflammatory APC subsets.
Studies in cancers of different etiologies have shown that neo-
antigen-directed immunity can be subverted by diverse mecha-
nisms (DuPage et al., 2011, 2012; Gubin et al., 2014; Schietinger
et al., 2016); mechanisms of immune evasion in PDAC are thus
worth elucidating. However, existing genetic models for PDAC
have not been amenable to study the heterogeneous interac-
tions between developing tumors and the host adaptive immu-
nity due to a dearth of tumor-specific neoepitopes (Evans
et al., 2016). Transplanted PDAC models are constrained by
their lack of stroma and a very distinct inflammatory/immunized
milieu upon tumor grafting, which can mask de novo immune
responses (Spear et al., 2019). In this study, we sought to deter-
mine how antigen-specific anti-tumor immunity becomes dysre-
gulated during progression of autochthonous PDAC.
RESULTS
Neoantigen Expression during Pancreas Cancer
Development Elicits Antigen-Specific Responses
The ‘‘KPC’’ genetic mouse model of pancreas cancer has been
widely used because of its fidelity to human PDAC, notably acti-
vating mutations in Kras(G12D) and loss of Trp53, associated
desmoplasia, and inflammation (Hingorani et al., 2003; Morton
et al., 2010). The model also mirrors human disease in its resis-
tance to both cytotoxic and immunotherapies (Beatty et al.,
2011; Gopinathan et al., 2015). However, KPC mice seldom
develop additional genetic alterations that drive prominent neo-
antigens for studying immune surveillance and evasion (Evans
et al., 2016; Li et al., 2018). Studies that have assessed the
impact of antigenicity have utilized heterotopic or orthotopic
tumor grafts that do not recapitulate de novo pancreas cancer
progression and thus may have very divergent immune contex-
ture (Spear et al., 2019). To study antigen-specific responses
in the context of de novo pancreas cancer development, we
engineered a mouse designed to express a model neoantigen
chicken ovalbumin (OVA) bicistronically with green fluorescent
protein (GFP) under the control of both Cre activation and tetra-
cycline repression (R26tm1(LSL-OG) or OG, Figure 1A). The
presence of neoepitopes for CD8+ T cell, CD4+ T cells, and B
cells allows us to study OVA-specific cellular and humoral immu-
nity raised during the course of tumor progression. These ‘‘OG’’
mice were crossed to KPC mice to create ‘‘KPC-OG’’ mice.
We first sought to test several key parameters of this model,
including antigen expression and presentation, CD8+ T cell
290 Cancer Cell 37, 289–307, March 16, 2020
recognition, and central tolerance. Using KPC-OG mice and
cell lines derived from tumors (KP-OG cells), we found that
PDAC cells express OVA and GFP, which can be repressed
by administration of doxycycline (Figure 1A). Notably, similar
to other lineage-tracing studies in this model (Rhim et al.,
2012), GFP and OVA were expressed concomitant with early
transformation and induction of metaplasia. As such, >95%
of cytokeratin 19+ (CK19) ductal cells co-expressed GFP (Fig-
ure 1B). Tumor lines derived from KPC-OG mice express major
histocompatibility complex 1 (MHCI) at levels equivalent to
those of traditional KPC mice (Figure S1A). To assess functional
antigen presentation by KPC-OG tumor cells, we performed
T cell killing assays ex vivo using OVA-specific OT-I CD8+
T cells and found that T cells both recognized and killed
KPC-OG-derived cells (Figures 1C and S1B). To verify that
endogenous antigen-specific T cells generated in OG+ mice
were not subjected to central tolerance prior to tumorigenesis,
we vaccinated tumor-free p48-Cre;Trp53fl/fl;OG (PC-OG) mice
with OVA and observed a clear enrichment of dextramer+
OVA-specific CD8+ T cells in peripheral blood and draining
lymph nodes (Figure S1C). Furthermore, we implanted KPC-
OG-derived tumor cells into PC-OG or PC (control) littermates
in the presence or absence of doxycycline. We observed that
grafted antigen-positive KP-OG cells grew equally slowly in
PC-OG and PC mice and that doxycycline repression of OVA
expression led to tumor progression (Figure S1D). Together,
these data suggest that p48-Cre-driven OVA neoantigen in
developing pancreatic tumors is presented and recognized by
T cells not subject to thymic deletion. Notably, we found that
untreated KPC-OG mice had OVA-specific T cell density equiv-
alent to that of mice with doxycycline withdrawal at birth, so we
did not treat with doxycycline for the remainder of the studies.
To determine the impact of neoantigen expression during
tumor initiation, we analyzed immune infiltrates in pre-
cancerous lesions of KPC-OG or KPC mice. At the early stage
of tumorigenesis (6 weeks), we observed increased infiltration
of CD8+ and CD4+ T cells and B cells in KPC-OG mice
compared with KPC littermates (Figures 1D and 1E). Assess-
ing antigen-specific responses, we observed increased
numbers of OVA-dextramer+ CD8+ T cells in pre-malignant
pancreas, draining lymph nodes (dLNs), and spleens of
early-stage KPC-OG compared with non-tumor-bearing PC-
OG mice (Figure 1F). Interestingly, compared with control an-
imals, the dextramer+ CD8+ T cells in KPC-OG pancreas had
higher Ki67+ frequency, but >30% of these cells were PD1hi/
TIM3hi, suggesting an early exhausted/dysfunctional pheno-
type (Figures S1E and S1F). This recapitulates observations
in liver cancer models that found a reversibly dysfunctional
phenotype (Schietinger et al., 2016). To ascertain whether
there was a systemic response toward tumor neoantigens,
we measured OVA immunoglobulin G (IgG) levels in serum.
We found total IgG1 to be similar between KPC-OG and
KPC littermates, but OVA-specific IgG1 titers in KPC-OG
serum were markedly higher than in age-matched controls
(Figure S1G). Together, these data suggest that there is an an-
tigen-directed immune response in KPC-OG pancreas during
initial stages of tumorigenesis. These observations emphasize
that early pancreatic lesions do not grow in an immune-privi-
leged environment.
A B C
p53
KPC

p48/Cre 8

OFF DOX
STOP Kras G12D

fluorescence flux (a.u.)

ROSA26 locus
6 *
OG

Ex1 STOP tTA TRE OVA IRES GFP Ex2

KPC-OG
**
– + 4 **
Doxycyline

OVA (45 kDa)

ON DOX
0
GFP (27 kDa) CTL : tumor

1
:1

:1
0
5:
10

20
ratio
β-actin (42 kDa) Ag non-specific T cells
OVA-specific OT-I
GFP CK19 DAPI

D CD8+ T cells CD4+ T cells B220+ B cells

* * 2.5
*
1.0 2.0
cells per cm2 tissue (x 103)

cells per cm2 tissue (x 103)

cells per cm2 tissue (x 104)
0.8 1.6
0.6 1.2 2.0
KPC-OG

0.9
0.4 1.5
0.6
1.0
0.2
0.3 0.5
CD8a PAN-K DAPI CD4 PAN-K DAPI 0 B220 DAPI 0
0

KPC (n=10) KPC-OG (n=10)

E CD8+ T CD4+ TH FOXP3+ TREG CD19+CD22+ B cells F OVA-specific T cells

* * ns * * * *
1.0 1.40 60 0.8 4 1.5 2.00
Treg (% of CD4+ T cells)

OVA-specific CD8+ T cells

OVA-specific CD8+ T cells

OVA-specific CD8+ T cells

cells per g tissue (x 106)

cells per g tissue (x 106)

0.8
cells per g tissue (x 106)

0.6 1.90
0.5 1.05 0.6 3
per dLN (x 103)

1.40
per g (x 104)

per g (x 105)
40 1.0
0.4 1.05
0.70 0.4 2
0.3
0.70
0.2 20 0.5
0.35 0.2 1 0.35
0.1
0 0 0 0 0 0 0
pancreas panc-dLN spleen

KPC (n=5) KPC-OG (n=8) PC-OG (n=3) KPC-OG (n=8)

Figure 1. Neoantigen Expression during Pancreas Cancer Development Elicits Antigen-Specific Responses
(A) Genetic loci for KPC-OG model and immunoblot for OVA and GFP expression in KPC-OG-derived cell line 72 h after doxycycline withdrawal. Representative of
three independent cell lines.
(B) Gross images (left) of pancreatic tissue at 6 weeks in KPC-OG mice on or off doxycycline, and immunofluorescence images (right) of pancreatic tumors at
36 weeks in KPC-OG mice on or off doxycycline. Scale bar, 100 mm.
(C) KPC-OG tumor-derived cell line depicting GFP fluorescence after 24-h co-culture with antigen-specific (OT-I TCR) or non-specific (C57Bl/6) activated CD8+
T cells (CTL) consistent across three independent cell lines. n = 3/group.
(D) Representative images and quantification of CD8+ T cells, CD4+ T cells, and B220+ B cells in 6-week-old KPC-OG and KPC mice. n = 10 mice/group. Scale
bar, 100 mm.
(E) Density of CD8+ T cells, CD4+ TH, CD4+ TREG, and CD19+CD22+ B cells measured by flow cytometry in early-stage KPC-OG and KPC mice. n = 5–8
mice/group.
(F) Density of OVA-specific CD8+ T cells in pancreas, pancreas dLNs, and spleen of early-stage KPC-OG and PC-OG mice. n = 3–8 mice/group.
Data were consistent across two independent experiments. Data are presented as mean ± SEM. For comparisons between two groups, Student’s two-tailed
t test was used. ns, not significant; *p < 0.05. See also Figure S1.

Neoantigen Expression Accelerates PDAC Progression We utilized the KPC-OG mouse (p53fl/+) and validated our find-
but Restrains Lung Adenocarcinomas ings in the KPPC-OG model (p53fl/fl), which exhibits faster
To evaluate the impact of neoantigen expression on pancreatic progression. Surprisingly, in both models we found that OG
cancer progression, we employed three distinct PDAC models. expression accelerated tumor progression at every stage of

Cancer Cell 37, 289–307, March 16, 2020 291

A B C D

50 * *

KPC
KPC
100 80
KPC

12

sirius red+ (% of tissue)

*

α-SMA+ (% of tissue)
lesions (% of tissue)

percent of lesions
80 40 9
60 5
60 ns 30 4
* 40
40 * 20 3

KPC-OG
KPC-OG

KPC-OG
20 2
20 10
1
0 0 0 0

KPC
KPC-OG
N3

oc
N2
N1

KPC
KPC-OG
en
nI
nI
nI

Pa
Pa
Pa

ad
H&E KPC (n=12) KPC (n=12) Sirius Red α-SMA
KPC-OG (n=12) KPC-OG (n=12)

G H
E F *
15 * 1.5 * 1.5 ns
2.5 * 100
14

KPC
KPC

high-grade (% of total)
100
per g tissue (x 106)

2.0 80
overall survival (%)

80
p48-Cre model

10 1.0 1.0
1.5 60 25% met
60
incidence
1.0 40
5 0.5 0.5 40

KPC-OG

KPC-OG
0.5 20
* 20

0 0 0 0 0 0 14
TAMs Granul- Mono- Eosino- Time (d) 40 60 80 100

KPC
KPC-OG
ocytes cytes phils
KPPC littermates (n=14)
62.5% met
*
KPC (n=5) KPC-OG (n=6) KPPC-OG (n=20) H&E H&E incidence

I J K
8 * 14.0 * 7.2 * 100
cells per g tissue (x 106)

lesion area (x 106 μm2)

100 8 lesions (% of tissue)
KPL
overall survival (%)

percent of lesions
80
Pdx-CreER model

80 6 T 10.5 5.4
40% met
60 *
60 4 7.0 3.6
incidence 40
40 *
2 3.5 1.8 20 ns
20
KPL-OG

0 8 0 0 0 0
Time (d post-tam) 150 200 250

gr 1

gr 2
3
e

e
KPL (n=5) KPL (n=6)

ad

ad

ad
iKPC (n=10) **

gr
100% met KPL-OG (n=5) KPL-OG (n=6)
iKPC-OG (n=8)
incidence
H&E

L M N
GFP CK19 DAPI

100 OT-I + IL-2 100 *

*
(% of total tumor cells)
KPC-OG

(% of total tumor cells)

untreated

100
GFP+ tumor cells

GFP+ tumor cells

overall survival (%)

80 80 80
KPPC-OG

60
60 * 60
40
40 20 40
T*
KPL-OG

OT-I Rx

0
GFP DAPI

0 15 30 45 60 0
Time (d since Rx )
T
KPC-OG (n=8) untreated (n=18) untreated (n=6)
KPL-OG (n=6) OT-I + IL-2 Rx (n=10) GFP CK19 DAPI OT-I Rx (n=6)

Figure 2. Neoantigen Expression Accelerates PDAC Progression but Restrains Lung Adenocarcinomas
(A) Representative hematoxylin-eosin (H&E) images with quantification of lesions in early-stage KPC-OG and KPC mice. n = 12 mice/group.
(B) Lesion grades for early-stage KPC-OG and KPC pancreata. n = 12 mice/group.
(C) Sirius Red staining with quantification in KPC-OG and KPC mice. n = 12 mice/group.
(D) a-SMA staining with quantification in KPC-OG and KPC mice. n = 12 mice/group.
(E) Flow-cytometric quantification of various myeloid infiltrates in KPC-OG and KPC mice. n = 5–6 mice/group.
(F) Kaplan-Meier survival curve for KPPC-OG mice compared with KPPC littermates. n = 14–20 mice/group.
(G) Representative histology of late-stage KPC-OG and KPC tumors with quantification of high-grade tumors. n = 14–16 mice/group.
(H) Representative H&E images of late-stage KPC-OG and KPC livers with quantification of metastases. n = 14 mice/group.
(I) Kaplan-Meier survival curve for Pdx1-Cre-ER-driven iKPC-OG mice compared with iKPC littermates, with quantification of liver metastases. n = 8–10
mice/group.

(legend continued on next page)

292 Cancer Cell 37, 289–307, March 16, 2020
disease. In early-stage KPC-OG mice at 6 weeks, we found that
OG expression led to a marked increase in intraepithelial
neoplasia (PANIN) area, higher-grade PANIN lesions, and
increased tumor cell proliferation (Figures 2A, 2B, and S2A).
Associated with this early disease progression was an increased
collagen deposition and a-smooth muscle actin (a-SMA)-posi-
tive fibroblast density (Figures 2C, 2D, and S2B). Analysis of
the inflammatory infiltrates indicated an increased infiltration of
neutrophils, eosinophils, and macrophages, but not of natural
killer (NK) cells, NK T (NKT) cells, or gdT cells (Figures 2E, S2C,
and S2D). Correspondingly, OG+ mice had reduced overall sur-
vival and tumors were of higher grade with markedly more liver
metastases (Figures 2F–2H). To understand the mechanisms un-
derpinning enhanced disease progression, we conducted RNA
sequencing of matched KPC-OG and KPC tissue. We observed
enrichment of mitogenic pathways (including mitogen-activated
protein kinase [MAPK], epidermal growth factor receptor
[EGFR], and tumor necrosis factor a [TNF-a] signaling) and in-
flammatory pathway activation, along with a robust upregulation
of EGFR ligands (Ray et al., 2014) and pro-inflammatory myeloid
chemokines (Figures S2E; Tables S1 and S2). Correspondingly,
we observed increased phosphorylated ERK, STAT3, and EGFR
staining in transformed cells of KPC-OG mice (Figure S2F),
suggesting the neoantigen results in changes in the tumor micro-
environment (TME) that support key pathways of transformation
and progression.
To address the issue that p48-Cre recombination occurs early
in pancreas development and leads to recombination in the ma-
jority of acinar cells, we employed an inducible Pdx1-Cre/Esr1*
driver (Gu et al., 2002). We generated Pdx1-Cre/Esr1*;
KrasLSL-G12D;Trp53fl/fl;OG mice, denoted iKPC-OG. We induced
sporadic recombination in iKPC-OG mice by tamoxifen adminis-
tration at 5 weeks of age, which led to mosaic activation in a field
of normal acini (Figure S2G). Thus, oncogenic mutations were
activated in parallel with neoantigen expression but only in a sub-
set of pancreas cells. Nevertheless, iKPC-OG mice ultimately
developed higher-grade tumors, and had reduced overall
survival and increased liver metastases compared with iKPC lit-
termates (Figure 2I).
Previous work in analogous KP lung models showed that
neoantigen recognition leads to productive immunity and re-
strains tumor progression (DuPage et al., 2011). To mirror these
studies in our model system, we intratracheally administered
Cre to KrasLSL-G12D;Trp53fl/fl;R26tm1(LSL-OG) (KPL-OG) or OG-
negative littermate mice (KPL) to generate lung adenocarci-
nomas and assess the impact of antigenicity (Figure S2H). In
agreement with previous studies, we observed increased
numbers of CD8+ and CD4+ T cells, lower tumor burden, and
decreased disease grade in KPL-OG mice compared with con-
trols (Figures 2J, 2K, and S2I). These data suggest that OG
expression in the lung and pancreas elicit different tumor pro-
gression outcomes.
One of the pathways by which tumors escape immune surveil-
lance is through loss of expression of prominent neoantigens via
immune editing (O’Donnell et al., 2019; Schumacher and
Schreiber, 2015). To test the propensity for antigen loss as an
evasion mechanism in our model, we analyzed end-stage
KPC-OG and KPL-OG tumors for persistence of neoantigen
expression. KPL mice exhibited substantial loss of GFP expres-
sion as tumors advanced (Figure 2L). However, in all three
models of pancreatic cancer we found no evidence of
antigen loss either at the primary site or in liver metastases (Fig-
ures 2L and S2J). To further verify these contrasting results, we
administered Ad-Cre (adenovirus-encoded Cre recombinase)
intramuscularly to create KP-OG+ sarcomas (DuPage et al.,
2012). Mirroring observations in the lung, we observed substan-
tial loss of GFP in advanced sarcomas (Figure S2K). Taken
together, these data indicate that immunogenic lung tumors
and sarcomas elicit an early immune response that delays tumor
progression, although antigenicity is lost or silenced with pro-
gression. In contrast, this was not observed during pancreatic
disease progression.
Neoantigen-Expressing PDAC Tumors Are Poorly
Responsive to Checkpoint Immunotherapy
Lack of high mutational and/or neoantigen burden has been pro-
posed to explain the poor responsiveness to immunotherapy in
PDAC patients. An alternative hypothesis is that the PDAC
TME enforces this lack of responsiveness to immunotherapy,
even when tumor antigens are present (Clark et al., 2009; Kieler
et al., 2018; Salmon et al., 2019; Sharma et al., 2017). To deter-
mine whether OVA expression leads to improved responsive-
ness in PDAC, we tested the efficacy of checkpoint and adoptive
T cell therapy. We treated established KPPC-OG tumors with
anti-PD1 and anti-CTLA4 IgGs. Despite neoantigen expression,
checkpoint therapy did not affect survival (Figure S2L). To model
the impact of adoptive T cell-therapy, we treated KPPC-OG
tumors with three rounds of OT-I adoptive transfer supple-
mented with interleukin-2 (IL-2) and observed a modest survival
benefit associated with loss of neoantigen at end stage (Figures
2M and 2N). These data suggest that a key bottleneck for treat-
ment efficacy in PDAC is priming sufficient antigen-specific
T cells, and not checkpoint activation on existing T cells
(Stromnes et al., 2015).

(J) Density of CD8+ T cells in early-stage KPL-OG and KPL lung lesions. n = 5 mice/group.
(K) Representative H&E images of early-stage KPL-OG and KPL lung with quantification of lesion area and grade. Lesions demarcated by yellow line. n = 5
mice/group.
(L) Representative immunofluorescence images and quantification of GFP (green) expression in CK19+ (red) tumors of late-stage KPC-OG or KPL-OG tumors.
GFP-negative lesions in KPL-OG tumors are demarcated by yellow arrowhead. n = 6–8 mice/group.
(M) Overall survival since start of treatment for KPPC-OG mice undergoing OT-I adoptive transfer therapy compared with untreated controls. n = 10–18
mice/group.
(N) Representative immunofluorescence images and quantification of GFP (green) expression in CK19+ (red) tumors of KPPC-OG mice subjected to OT-I
adoptive transfer compared with untreated controls. n = 6 mice/group.
Data were consistent across two independent experiments. Data are presented as mean ± SEM. For comparisons between two groups, Student’s two-tailed
t test was used. For survival analyses, log-rank (Mantel-Cox) test was used. ns, not significant; *p < 0.05, **p < 0.01. Scale bars, 500 mm (A, C, D, G, H, and K) and
100 mm (L and N). See also Figure S2; Tables S1 and S2.

Cancer Cell 37, 289–307, March 16, 2020 293

 100 ns B C
A
80 * ns

lesions (% of tissue)
* ns
60 42 24
control

40

control

control
18 *

sirius red+ (% of tissue)

35

α-SMA+ (% of tissue)
16 12
8 28

0 21
14 6
80 *
percent of lesions

60 ns 7

α-CD4

α-CD4
α-CD4

* * * 0 0
40 * * KPC KPC-OG KPC KPC-OG

control

α-CD19/α-B220
α-CD4

control

α-CD19/α-B220
α-CD4
20

H&E 0 PanIN1 PanIN2 PanIN3 adenoc Sirius Red α-SMA

KPC (n=12) KPC-OG + α-CD4 (n=10)
KPC-OG (n=12) KPC-OG + α-CD19/α-B220 (n=8)

D CD4+ TH cells
E CD4+ TH cells IL17A+ TH TNFα+ TH
RORγt+ GATA3+
KPC KPC-OG 2.8 * 2.20 * KPC KPC-OG 0.5
* 0.3 *
2.4% 12.6% 10% <1% 12% 4.5%
cells per g tissue (x 105)

cells per g tissue (x 105)

cells per g tissue (x 106)

cells per g tissue (x 105)

2.1 0.4
1.65
PE-A :: GATA3

FITC :: TNF-α 0.2

0.3
1.4 1.10
1% 5.5% 0.2
1.1% 9% 0.1
0.7 0.55
PE-Texas Red-A :: RORγt PE-Cy7 :: IL17A 0.1

KPC (n=3) KPC-OG (n=6) 0 KPC (n=3) KPC-OG (n=6) 0 0

F 60 G 25 H 5
* * *
sirius red+ (% of tissue)
lesions (% of tissue)

α-SMA+ (% of tissue)
20 4
α-IL17A,F

α-IL17A,F
α-IL17A,F

40
15 3

10 2
20
5 1
H&E Sirius Red α-SMA
KPC 0 KPC 0 KPC 0
KPC-OG KPC-OG KPC-OG KPC-OG KPC-OG KPC-OG KPC-OG KPC-OG KPC-OG
+ α-CD4 + α-IL17A/α-IL17F + α-CD4 + α-IL17A/α-IL17F + α-CD4 + α-IL17A/α-IL17F

I
10 * 14.0 * 24 *
ns
pERK+ cells (% of total)

pSTAT3+ cells (% of total)

pEGFR+ cells (% of total)

control
control

8
control

10.5 18
6
7.0 12
4
3.5 6
2

0 0 0
α-CD4
α-CD4

α-CD4

18 * 18 * 56 *
pERK+ cells (% of lesion)

pSTAT3+ cells (% of lesion)

pEGFR+ cells (% of lesion)

ns
48
40
12 12
32
α-IL17A,F
α-IL17A,F

α-IL17A,F

24
6 6
16
8
0 0 0
p-ERK1/2 p-STAT3 (Y705) p-EGFR (Y1068)
KPC (n=10) KPC-OG (n=10) KPC-OG + α-CD4 (n=10) KPC-OG + α-IL17A/α-IL17F (n=8)

Figure 3. Pro-inflammatory CD4+ T Cell Responses Drive PDAC Acceleration in Response to Neoantigen
(A) Representative H&E images with quantification of pancreatic lesion area (top) and grade (bottom) in early-stage KPC-OG mice subjected to depletion of CD4+
T cells or CD19+B220+ B cells. n = 8–12 mice/group.
(legend continued on next page)
294 Cancer Cell 37, 289–307, March 16, 2020
Pathogenic CD4+ T Cell Responses Drive PDAC
Acceleration in Response to Neoantigen
Our data suggest that neoantigen expression leads to adap-
tive immune responses that surprisingly drive tumor progres-
sion. Previous studies have shown that CD4+ T cells or acti-
vated B cells can drive pathogenic inflammation and
accelerate PDAC progression (Barilla et al., 2019; Gunderson
et al., 2016; McAllister et al., 2014; Pylayeva-Gupta et al.,
2016; Zhang et al., 2014). Thus, we evaluated whether CD4+
T or B cells were critical for the early-stage disease progres-
sion observed in OG+ mice. We found that while B cell deple-
tion did not alter tumor progression, CD4+ T cell depletion led
to a decrease in pre-malignant disease burden and PANIN
grade, reduced collagen density, and decreased a-SMA+
fibroblast accumulation (Figures 3A–3C, S3A, and S3B). These
data suggest CD4+ T cells accelerate pancreatic neoplasia in
response to neoantigen expression. We next evaluated the
polarization of CD4+ T cell responses in OG+ mice and
observed higher numbers of pancreas-infiltrating ROR-gt+
and GATA3+ CD4+ TH cells, and more TH cells producing IL-
17A, TNF-a, IL-4, and IL-10, consistent with dominant TH17
and TH2 responses (Figures 3D, 3E, and S3C–S3E). By
contrast, we did not see increased frequency of Tbet+ or inter-
feron-g (IFN-g)-producing TH cells in KPC-OG (Figure S3F).
Surprisingly, partial depletion of FOXP3+ TREGS did not affect
disease progression in KPC-OG tumors (Figures S3G and
S3H). To determine the function of this enhanced TH17 signa-
ture observed in early lesions, we depleted cytokines neces-
sary for their activity. Upon neutralizing pro-inflammatory IL-
17 signaling, we observed lower disease burden and patho-
logical fibrosis (Figures 3F–3H). Correspondingly, we found
the increased expression of phosphorylated (p)-ERK, p-
STAT3, and p-EGFR signaling observed in KPC-OG lesions
was attenuated upon CD4+ T cell depletion or IL-17 neutraliza-
tion (Figure 3I). We next compared CD4+ T cell polarization in
early-stage pancreatic and lung lesions and found higher fre-
quency of GATA3+ and ROR-gt+ TH cells in KPC-OG pancreas
compared with KPL-OG lung tumors. By contrast, KPL-OG
lung-infiltrating TH cells were more TH1-skewed, with
increased frequency of Tbet+ and IFN-g-producing cells (Fig-
ures S3I and S3J). Overall, these data indicate that, in contrast
to lung, pancreatic neoantigen expression results in enhanced
pathogenic TH17 responses that can facilitate progression
(Alam et al., 2015; McAllister et al., 2014; Zhang et al.,
2016, 2018).
cDCs Are Fewer and Less Functional in PDAC Compared
with Lung Cancer
We sought to determine the cellular origins for differences in
response to neoantigenicity in the lung and pancreas. We per-
formed immune profiling of major innate immune cell subsets
in both early- and late-stage lung and pancreas tumor tissues
and found that cDCs were among the most divergent (Figures
4A, S4A, and S4B). In early pre-malignant stages, we observed
that CD103+ CD24hi cDC1s were 10-fold fewer and CD172a+
CD11b+ cDC2s were 4-fold fewer in the pancreas when
compared with lung. The disparity in cDC1s was magnified at
later stages, with cDC1s in PDAC being 79-fold less than in
counterpart lung adenocarcinomas (Figure 4B). These observa-
tions were not affected by OG expression. We also observed
fewer migratory CD103+ cDC1s and CD11b+ cDC2s in pancreas
dLNs of KPC-OG mice when compared with lung-dLNs of KPL-
OG mice, but no major difference in resident DC populations
(Figures 4C and S4C). To determine possible differences in
cDC localization between the pancreas and lung TME, we trans-
planted irradiated KPC and KPL mice with Zbtb46GFP and
Snx22GFP bone marrow. The Zbtb46GFP reporter model marks
all cDCs (Satpathy et al., 2012) while the Snx22GFP model labels
Batf3-dependent cDC1s (Bra €hler et al., 2018) (Figure S4D). Using
immunohistochemistry to quantify GFP+ cells, we observed pat-
terns similar to our flow-cytometric results with markedly more
Zbtb46-GFP+ cDCs and Snx22-GFP+ cDC1s in lung when
compared with pancreas, both in pre-malignant tissues and
late-stage cancer (Figure 4D). We next analyzed cDC localization
and found that both Zbtb46-GFP+ cDCs and Snx22-GFP+
cDC1s localized close to lung tumor cells (>60% were within
5 mm or less), while in PDAC cDCs were more distant from tumor
nests (Figure 4E). Analysis of lung and pancreatic tumoral cDCs
found that cDC1s and cDC2s had lower co-stimulatory and
maturation markers in PDAC (Figure S4E), and pancreatic
cDC1s were less functional at antigen presentation in ex vivo as-
says (Figure S4F). The higher cDC1 density found in lung cancer
also paralleled higher OVA-specific CD8+ T cell density, suggest-
ing a critical role in antigen-specific T cell immunity (Figure 4F).
This was further supported by the observation that depleting
cDC1s prior to KPL-OG lung tumor initiation via Batf3 / bone
marrow transplant results in drastically reduced CD8+ T cell infil-
tration (Figure S4G).
To determine whether these observations in mouse models
held true in human pancreatic tumors, we analyzed cDC density
in human PDAC tissue by mass cytometry and publicly available

(B) Sirius red staining with quantification in early-stage KPC-OG mice subjected to indicated depletions. n = 8–12 mice/group.
(C) a-SMA staining with quantification in early-stage KPC-OG mice subjected to indicated depletions. n = 8–12 mice/group.
(D) Representative flow-cytometry plots of RORgt and GATA3 bias in TH cells of KPC-OG and KPC tumors, with cellular density of RORgt+ TH17 and GATA3+ TH2
cells quantified. n = 3–6 mice/group.
(E) Representative flow-cytometry plots of IL-17A and TNF-a expression in TH cells of KPC-OG and KPC tumors, with cellular density quantified. n = 3–6
mice/group.
(F) Representative H&E image with quantification of pancreatic lesion area of early-stage KPC-OG mice subjected to IL-17A and IL-17F neutralization. n = 8–10
mice/group.
(G) Representative Sirius red staining of KPC-OG mice subjected to IL-17A and IL-17F neutralization, with quantification. n = 8–10 mice/group.
(H) Representative a-SMA staining of KPC-OG mice subjected to IL-17A and IL-17F neutralization, with quantification. n = 8–10 mice/group.
(I) Representative p-ERK1/2, p-STAT3, and p-EGFR immunohistochemistry staining in KPC-OG mice subjected to indicated depletions, with quantification over
overall tissue area and lesion area. n = 8–10 mice/group.
Data were consistent across two independent experiments and pooled. Data are presented as mean ± SEM. For comparisons between two groups, Student’s
two-tailed t test was used. ns, not significant; *p < 0.05, **p < 0.01. Scale bars, 500 mm. See also Figure S3.

Cancer Cell 37, 289–307, March 16, 2020 295

 A B C
KPC-OG

KPL-OG

** * 0.70 ** 0.6 * 3 * 8 *
1.5 ns 3 ns

migr DC1s per dLN (x 104)

migr DC2s per dLN (x 104)

ns ns

cDC2s per g tissue (x 106)

KPC

cDC2s per g tissue (x 106)

cDC1s per g tissue (x 106)
cDC1s per g tissue (x 106)
KPL

0.53 6
1.0 2 0.4 2
TAMs
eosinophils 0.35 4
granulocytes ns ns
0.5 ns 1 0.2 1
monocytes * 0.17 2
CD103 cDC1
CD11b cDC2 0 0 0 0 0 0
DN cDC early stage late stage late stage
mo-DC KPC early KPC KPL
KPL early
mean (log10) cells/g KPC-OG early KPL-OG early KPC-OG KPL-OG KPC-OG KPL-OG
3.5 7.0

D E

2.0 60 80
4 ** **
KPC

relative frequency (%)

relative frequency (%)

Zbtb46-GFP+ cells
per cm2 tissue (x 104)

per cm2 tissue (x 104)

Snx22-GFP+ cells
60
Zbtb46-GFP+ cells

Snx22-GFP+ cells

3 1.5 40

2 1.0 ** 40 **
* ** 20
K-S: 0.648 20 K-S: 0.735
1 0.5
KPL

0 0
0 0 10 30 50 70 90 10 30 50 70 90
distance from lesion (μm) distance from lesion (μm)

WT KPC WT KPL KPC tumor KPL tumor

pancreas tumor normal lung tumor
CK19 GFP

F G *
H
6
** **
early KPL-OG 4.2 * 1.40 * 2 2
OVA-specific CD8+ T cells

per g (x 105) at early stage

OVA-specific CD8+ T cells

OVA-specific CD8+ T cells

per g (x 105) at late stage

log10 (% of CD45+ cells)

cDC1 gene expression

per g tissue (Log10)

late KPL-OG
t-SNE dim 2

5 1.05 1 1
2.8

Z-score
0.70 0
0
4 early KPC-OG 1.4
late KPC-OG 0.35 -1
macrophages
neutrophils

-1
3 0 CD45 immune cells; n=11
+
-2
3 4 5 6 7
cDC1
cDC2

t-SNE dim 1
PAAD

LUAD
cDC1 cells per g tissue (Log10)
(177)

(230)
KPC-OG KPL-OG

I CD103+ migr cDC1 J

7.00 *
32 * 80 *
OVA-specific CD8+ T cells
ZsGreen+ (% of population)
ZsGreen+ (% of population)
normalized to mode

in early stage tumors

early
in late stage tumors

5.25
per dLN (x 103)

1.8 KPC-Z 24 60
early
21.5 KPL-Z *
16 * 40 3.50
32.3 KPC-Z
8 20 1.75
57.1 KPL-Z
0.2 uninv LN
control
0 0 0
migr migr migr migr
0 ZsGreen CD103+ CD11b+ CD103+ CD11b+

KPC-ZsG KPL-ZsG KPC-OG KPL-OG

Figure 4. cDCs Are Fewer and Less Functional in PDAC Compared with Lung Cancer
(A) Heatmap depicting mean density (log scale) of major myeloid cell infiltrates in advanced KPC-OG pancreatic and KPL-OG lung tumors. n = 5–10 mice/group.
(B) CD103+ cDC1 and CD11b+ cDC2 density in pancreas and lung tumors, at (left) early stage and (right) late stage. n = 5–10 mice/group.
(C) Migratory cDC1 and cDC2 density in respective draining lymph nodes of late-stage KPC-OG pancreatic tumors and KPL-OG lung tumors. n = 7 mice/group.
(D) Immunohistochemistry for tumor cytokeratin (CK7/19) expression, and Zbtb46-GFP+ (pink) cDCs in late-stage KPC or KPL bone marrow chimeras. Right:
Zbtb46-GFP+ cDC density in non-tumor (WT) tissue and late-stage tumors. Far right: Snx22-GFP+ cDC1 density in WT and late-stage tumors. n = 4–6 mice/group.
Scale bar, 100 mm.

(legend continued on next page)

296 Cancer Cell 37, 289–307, March 16, 2020
datasets. Using mass cytometry, we found that cDC1s specif-
ically are extremely rare and 100-fold less abundant when
compared with tumoral macrophages (TAMs) or neutrophils in
human PDAC tissues (Figure 4G). Additionally, normalized
cDC1 gene-signature levels (Barry et al., 2018; Böttcher et al.,
2018; Spranger et al., 2017) are much lower in PDAC when
compared with lung adenocarcinoma (Figure 4H). These obser-
vations mirror our mouse models and indicate that cDCs, specif-
ically cDC1s, are particularly rare in human PDAC tissue.
A major anti-tumor function of cDCs involves antigen sampling
and migration to tumor dLNs to prime T cell responses. To
assess this priming function, we bred LSL-ZsGreen (ZsG) mice
into KPC or KPL mice. ZsGreen expressed by transformed tissue
is lysosome-stable (Roberts et al., 2016) and enables us to track
antigen uptake and trafficking by different APCs in the TME and
tumor dLNs. We observed ZsGreen throughout transformed
pancreatic and lung lesions; and TAMs, cDC1s, and cDC2s
robustly took up ZsGreen from malignant cells in both tissues
(Figure S4H). However, there were stark differences in the fre-
quency of ZsGreen+ migratory cDC1s and cDC2s in tumor
dLNs. Across multiple time points, significantly more migratory
cDCs were ZsGreen+ in KPL-ZsG mice when compared with
stage-matched KPC-ZsG mice (Figure 4I). Most striking was
the fact that in early-stage KPC-ZsG mice nearly no migratory
cDCs trafficked tumor-derived ZsGreen, despite clear antigen
expression in lesions and loading on intrapancreatic cDCs at
this stage. To determine whether poor antigen trafficking by
migratory cDCs at early stages of pancreatic tumorigenesis influ-
enced antigen-specific T cell priming, we analyzed OVA-specific
CD8+ T cells in the tumor dLNs of OG+ mice. We found that KPC-
OG pancreas dLNs had far fewer OVA-specific CD8+ T cells
compared with stage-matched KPL-OG dLNs (Figure 4J).
Collectively, these data suggest that T cell priming by cDCs
against neoantigens in developing PDAC is less functional
compared with lung adenocarcinomas.
Mobilizing cDCs into Early Pancreatic Lesions Can
Reverse Fibro-Inflammatory Responses
We next tested whether increasing cDCs in early stages of PDAC
could reassert TH1 and CTL-mediated disease control. To accom-
plish this, we stimulated hematopoietic mobilization of cDC
precursors using Fms-related tyrosine kinase 3 ligand (Flt3L)
treatment (Hammerich et al., 2019; Meyer et al., 2018; Salmon
et al., 2016). Flt3L administration at early stages of tumorigenesis
resulted in robust infiltration of cDCs in the pancreas, including a
10-fold increase in cDC1s (Figure 5A). These data suggest that
when mobilized, cDC precursors can successfully infiltrate early
pancreatic lesions. Additionally, we observed that Flt3L treatment
alone could revert disease acceleration and fibro-inflammatory
pathology of KPC-OG mice. Compared with untreated mice,
Flt3L-treated KPC-OG mice had a reduced lesion area and
lower-grade PANIN lesions as well as reduced collagen deposi-
tion and a-SMA+ fibroblast density (Figures 5B–5D and S5A).
Flt3L treatment also resulted in a reduced number of ROR-gt+
IL-17A-expressing TH cells and GATA3+ TNF-a-expressing TH
cells, and increases in IFN-g-producing TH1 cells (Figures 5E,
5F, and S5B–S5E). Notably, there was a sharp reduction in TNF-
a-expressing TH17 cells (Figure 5F). While the absolute number
of CTLs did not increase, Flt3L treatment increased CD8+
T cells, proximity to lesions, effector function as measured by
IFN-g+ and TNF-a+ production, and proliferation (Figures 5G–5I
and S5F). To determine whether these changes were functional,
we depleted CD8+ T cells or IFN-g and found that it abolished
the tumor control by Flt3L treatment (Figure 5J). Concurrently,
we found that CD8+ T cell depletion attenuated increases in the
tumor cell death and antigen editing observed in Flt3L-treated
KPC-OG mice (Figures S5G and S5H). Together, these data imply
that restoring cDC numbers in early stages of PDAC results in a
switch from pathogenic tumor-promoting TH17 to tumor-restrain-
ing TH1 and CD8+ CTL responses to neoantigens.
Enhancing cDC Infiltration and Activation in Established
PDAC Leads to Disease Stabilization
Our observations raised the possibility that Flt3L-based cDC
mobilization in established pancreatic tumors could benefit
anti-tumor immunity. Additionally, our previous work has shown
that established PDAC in human patients and KPC mouse
models can impair cDC1 development in the bone marrow
(Meyer et al., 2018), and therapeutic strategies might require
boosting cDC mobilization to overcome this disruption. Thus,
we treated KPPC-OG mice bearing established tumors with
Flt3L. Notably, increases in cDC infiltration upon Flt3L treatment
were more modest in established PDAC compared with pre-ma-
lignant pancreas (Figures 6A and 6B). Also, administering Flt3L
alone in this advanced setting did not change intratumoral CTL
or TH cells and led to an increase in TREG frequency (Figure 6C).
This finding is in line with previous work (Ager et al., 2017;
Salmon et al., 2016), and suggested that increasing mobilization
of cDC progenitors is insufficient in generating favorable T cell
responses in established PDAC.

(E) Frequency distribution of Zbtb46-GFP+ cDC and Snx22-GFP+ cDC1 proximity to nearest CK7/19+ tumor cell. n = 3–5 mice/group.
(F) Tumoral cDC1 density (log scale) plotted against OVA-specific CD8+ T cell density (log scale) across tissue and stage. Right: OVA-specific CD8+ T cell density
in early- and late-stage tumors. n = 5–8 mice/group.
(G) Phenograph of CD45+ immune infiltrates from human PDAC patient CyTOF samples (pooled), with quantification of individual cellular fractions (log scale).
n = 11.
(H) Z–normalized cDC1 infiltration score between pancreatic (PAAD, n = 177) and lung (LUAD, n = 230) adenocarcinoma based on conserved cDC1 gene
signature.
(I) Representative histogram indicating ZsGreen in migratory cDC1s and cDC2s from respective draining lymph nodes of KPC-Z or KPL-Z tumors at denoted time
points. Right: percentage of migratory cDC subsets that have ZsGreen antigen in respective lymph nodes at denoted time points. n = 3–4 mice/group.
(J) Density of OVA-specific CD8+ T cells in draining lymph nodes of early-stage tumors. n = 5–8 mice/group.
Data were consistent across two independent experiments, except in (A), (B), (F), (G), and (H), where they were pooled across multiple experiments. Data are
presented as mean ± SEM, except in (H) where box plot denotes 10th to 90th percentile, middle line indicates median, range lines indicate maximal values, and
data points beyond indicate outliers (>1.53 range). For comparisons between any two groups, Student’s two-tailed t test was used. Frequency distributions were
compared using non-parametric Kolmogorov-Smirnov test. ns, not significant; *p < 0.05, **p < 0.01. See also Figure S4.

Cancer Cell 37, 289–307, March 16, 2020 297

 A B control + Flt3L
1.5 * 3.0 * 60
* *

cDC2s per g tissue (x 106)

cDC1s per g tissue (x 106)
2.0
early stage KPC-OG 1.5

lesions (% of tissue)
Flt3L i.p. 1.0 40
1.0
0 9 12
0.5 20
0.5

H&E
0 0 0
KPC (n=5) KPC-OG (n=6) KPC-OG (+Flt3L) (n=6) KPC KPC-OG KPC-OG (+Flt3L)

C control + Flt3L 28
* *
D control + Flt3L 5
* *

sirius red+ (% of tissue)

4

α-SMA+ (% of tissue)
21
3
14
2
Sirius Red

α-SMA
1

0 0
KPC KPC-OG KPC-OG (+Flt3L) KPC KPC-OG KPC-OG (+Flt3L)

E RORγt+ GATA3+ F CD4+ TH cells IL17a+ TH TNFα+ IL17a+ TH

3.0 * * 0.4 * *
2.20 control +Flt3L 1.6
cells per g tissue (x 105)

2.0
cells per g tissue (x 105)

cells per g tissue (x 104)

cells per g tissue (x 105)
1.8 8.5% 2.5% 5.4% 0.1%
1.65 0.3 1.2
FITC :: TNF-α

1.2
1.10 0.2 0.8

0.6 2.8% 0.7% 0.1 0.4

0.55
PE-Cy7 :: IL17a
0 0 0 0

KPC (n=3) KPC-OG (n=6) KPC-OG (+Flt3L) (n=6) KPC (n=3) KPC-OG (n=6) KPC-OG (+Flt3L) (n=6)

G H I
control + Flt3L 1.6
ns 50 0.60 *
IFNγ+ TNFα+ CD8+ T cell
CD8+ T cells within 30 μm

0.55
relative frequency (%)

40
oer g tissue (x 104)
/ cm2 lesion ( x 103)

1.2 0.35
CD8+ T cell

30 0.26
0.8 **
CD8α CK19

20 K-S: 0.306 0.17

0.4 10 0.08

0 0 0
10 30 50 70 90
control
+Flt3L
control
+Flt3L

distance from lesion (μm)

control +Flt3L

J
+Flt3L +Flt3L + α-CD8α +Flt3L + α-IFNγ 90 * *
lesions (% of tissue)

60
all KPC-OG
control (n=12)
30
+Flt3L (n=9)
H&E

+Flt3L +α-CD8α (n=9)

+Flt3L +α-IFNγ (n=7)
0

Figure 5. Mobilizing cDCs into Early Pancreatic Lesions Can Reverse Fibro-Inflammatory Responses
(A) Schematic of Flt3L administration in KPC-OG mice (starting at P30), with quantification of cDC1 and cDC2 density in pancreata of KPC-OG mice either treated
or not with Flt3L and control KPC mice. n = 5–6 mice/group.
(legend continued on next page)
298 Cancer Cell 37, 289–307, March 16, 2020
 We speculated that the lack of impact of Flt3L on CTL and TH
cell infiltration could be due to ineffective licensing of incoming
cDCs, allowing for the immature DCs to become tolerogenic in
the TME. To overcome this barrier, we either intratumorally
injected an STING agonist (RR-S2-CDA) to influence IFN-depen-
dent DC maturation (Corrales et al., 2015; Sivick et al., 2018) or
administered CD40-agonist IgGs to improve licensing and
enhance APC function and survival (Beatty et al., 2011; Byrne
and Vonderheide, 2016). Unlike other tumor models (Kinkead
et al., 2018; Ma et al., 2019), we found that neither CD40 nor
STING agonist alone enhanced cDC abundance in the PDAC
TME. However, combination of either agent with Flt3L worked
synergistically to drive massive influx of cDC1s and cDC2s,
with CD40 agonist showing clear superiority and a >64-fold
increase in cDC1s (Figure 6D). Phenotypically, cDC1s and
cDC2s had higher MHCI and MHCII expression and modestly
higher CD80/CD86 expression upon CD40 agonist and Flt3L
treatment (Figure S6A). Notably, combination treatment with
CD40 agonist and Flt3L did not mobilize more pre-cDCs than
Flt3L treatment alone, suggesting that the observed synergism
was in the PDAC TME (Figures S6B and S6C).
While CD40 or STING agonist alone modestly enhanced CD8+
T cell infiltration, combination with Flt3L triggered markedly
enhanced intratumoral CD8+ CTL and CD4+ TH cell infiltration
without TREG induction (Figure 6E). Additionally, CD40 agonist
plus Flt3L combination treatment increased the abundance of
intratumoral OVA-specific CD8+ T cells (Figure 6F). Histological-
ly, we found more CD8+ CTLs were in close proximity and
frequently in contact with PDAC cells of mice treated with the
CD40 agonist plus Flt3L combination (Figure 6G). Notably, the
CD40 agonist plus Flt3L treatment elicited integrated anti-tumor
responses involving marked increases in infiltrating NK, NKT,
and gdT cells (Figures 6H and S6D) and resulted in significant
loss of GFP in OG tumors, suggesting immune evasion under
T cell pressure (Figure S6E). Due to the substantial changes
observed in cDC numbers upon anti-CD40 plus Flt3L combina-
tion treatment, we next tested priming capacity in PDAC dLNs
using KPPC-ZsG mice. We found a considerable increase in
ZsGreen+ migratory cDC1s trafficking tumor antigen to dLNs
of CD40 agonist plus Flt3L-treated mice (Figure 6I). To determine
whether this enhanced antigen trafficking results in better
peripheral cross-priming, we analyzed the enrichment of anti-
gen-specific CD8+ T cells in the dLNs of treated KPPC-OG
tumors. There was a 6.7-fold increase in OVA-specific CD8+
T cells in dLNs of the anti-CD40 plus Flt3L treatment cohort,
compared with a 3.5-fold increase in the anti-CD40-only cohort
(Figure 6J). This suggested that combination treatment was effi-
cacious in improving T cell priming in PDAC dLNs.
Combining Flt3L administration with either CD40 or STING ag-
onism resulted in improved disease control beginning a week
into either treatment (Figure 6K). Combining CD40 agonism
and Flt3L was also effective at slowing tumor progression in
the non-antigenic KPPC model (Figures S6F and S6G) and was
dependent on T cells for efficacy (Figures S6H and S6I). Histo-
logically, tumors receiving Flt3L and anti-CD40 had no major
changes in total PDPN+ fibroblast presence, but substantially
lower collagen deposition and a reduction in desmoplastic
a-SMA+ fibroblasts (Figures 6L, 6M, and S6J). These data sug-
gested that Flt3L plus anti-CD40 treatment might either result
in collagen degradation and/or indirectly switch fibroblast
phenotype. In agreement with potential myeloid-dependent
matrix remodeling, we observed higher matrix metallopepti-
dase-9 production in tumors treated with Flt3L plus anti-CD40
(Figure S6K) (Long et al., 2016).
cDC-Directed Therapy Renders PDAC Responsive to
Radiation Therapy
While we found that CD40 agonist plus Flt3L reshaped the im-
mune response, we did not observe tumor regression. A possible
explanation is that immune priming and ‘‘antigen spill’’ by tumor
cell death is very limited in this model (Vonderheide, 2018). One
effective modality to induce immunogenic cell death and boost
CTL priming by APCs is radiation therapy (RT) (Ngwa et al.,
2018; Twyman-Saint Victor et al., 2015). RT in PDAC patients
has limited benefit and is mostly palliative (Balaban et al.,
2016), possibly because the TME does not support induction
of tumor immunity. To test whether cDCs could provide the
necessary induction and synergize with RT, we combined anti-
CD40 agonist and Flt3L treatment with radiotherapy. We utilized
computed tomography-guided RT to provide three fractionated
8-Gy doses directed to KPPC-OG pancreata after cDCs were
mobilized. While RT alone only modestly affected tumor progres-
sion, the triple therapy resulted in tumor regression in the major-
ity of KPPC-OG animals (Figures 7A–7C). Similar results were

(B) Representative H&E images of early-stage KPC-OG mice either treated or not with Flt3L, with quantification of lesion area. n = 8–12 mice/group; mice from
Figure 2 are included in this and following analyses.
(C) Sirius Red staining with quantification in KPC-OG mice either treated or not with Flt3L. n = 8–12 mice/group.
(D) a-SMA staining with quantification in KPC-OG mice either treated or not with Flt3L. n = 8–12 mice/group.
(E) Density of RORgt+ TH17 and GATA3+ TH2 cells in early-stage KPC-OG mice treated as indicated. n = 3–6 mice/group.
(F) Representative flow-cytometry plots of IL-17A and TNF-a expression in TH cells of KPC-OG mice either treated or not with Flt3L, with IL-17A+ and TNF-a+
IL-17A+ TH cellular density quantified. n = 3–6 mice/group.
(G) Representative immunohistochemistry of CD8+ T cells (brown) and CK19+ tumor lesions (pink) in early-stage KPC-OG mice either treated or not with Flt3L.
n = 6 mice/group.
(H) Cumulative CD8+ T cell density within 30 mm of CK19+ lesions, and distribution of CD8+ T cell proximity to nearest tumor cell in KPC-OG mice treated as
indicated. n = 6 mice/group.
(I) Density of IFN-g+ TNF-a+ cytotoxic CD8+ T cells in KPC-OG mice treated as indicated. n = 6 mice/group.
(J) Representative H&E images of early-stage KPC-OG mice treated with Flt3L and anti-CD8 or anti-IFN-g depletion antibodies, with quantification of lesion area.
n = 7–12 mice/group.
Data were consistent across two independent experiments, except in (B), (C), (D), (H), (I), and (J), where they were pooled across multiple experiments. Data are
presented as mean ± SEM. For comparisons between any two groups, Student’s two-tailed t test was used. Frequency distributions were compared using non-
parametric Kolmogorov-Smirnov test. ns, not significant; *p < 0.05, **p < 0.01. Scale bars, 500 mm (B, C, D, J) and 100 mm (G). See also Figure S5.

Cancer Cell 37, 289–307, March 16, 2020 299

 A B C CD8+ TIL CD4+ TH CD4+ TREG
* * * * ns ns
4 8 6 12 3.2 2.0
*
42

cDC2s per g tissue (x 105)

cDC1s per g tissue (x 105)

migr cDC1s per dLN (x104)

migr cDC2s per dLN (x104)

0 Flt3L

cells per g tissue (x 106)

cells per g tissue (x 105)
35

% of CD4+ T cells
3 6 9 2.4 1.5
4 28
2 4 6 1.6 1.0 21
2 14
1 2 3 0.8 0.5
7
9
0 0 0 0 0
14 untreated Flt3L untreated Flt3L

D E CD8+ TIL CD4+ TH CD4+ TREG F

3.50 * 3 * 3.2 * 4.20 * 42 * 6 *
cDC1s per g tissue (x 106)

cDC2s per g tissue (x 106)

cells per g tissue (x 104)

cells per g tissue (x 106)
cells per g tissue (x 106)

OVA-specific CD8+ T
35

% of CD4+ T cells
2.62 2.4 ns
3.15
2 28 4

1.75 1.6 2.10 21 *

*
1 14 2
0.87 0.8 1.05
7

0 0 0 0 0 0
untreated Flt3L α-CD40 untreated Flt3L α-CD40 untreated Flt3L
STINGag + Flt3L α-CD40 + Flt3L STINGag + Flt3L α-CD40 + Flt3L α-CD40 α-CD40
+ Flt3L

G H NK cells NKT cells γδ T cells

untreated α-CD40 + Flt3L * 5 1.0 1.0
2.5 * * *
CD8+ TILs < 50 μm (x 105)

cells per g tissue (x 105)

cells per g tissue (x 105)

cells per g tissue (x 105)

2.0 4 0.8 0.8
per cm2 tissue

1.5 3 0.6 0.6

ns ns *
*
2 0.4 0.4
CD8α CK19

1.0

0.5 1 0.2 0.2

0 0 0 0

untreated Flt3L untreated Flt3L α-CD40

α-CD40 α-CD40 + Flt3L STINGag + Flt3L α-CD40 + Flt3L

I 4 * J 7.2 * K 0.3 0.4

untreated
ZsGreen+ cells in dLN (x 104)

0.3
volume measure (cm3)
volume measure (cm3)
cells per dLN (x 103)
normalized to mode

0.2
3 5.4 * 0.2 0.1

CD103+ DC1 2 3.6

ns 0.4
27.2% α-CD40
0.1 0.3 + Flt3L
1 1.8
* 0.2
CD11b+ DC2
0.1
19.1%
0 0
ZsGreen migr cDC1 migr cDC2 0 7 14 0 7 14
0
days since treatment start days since treatment
untreated Flt3L untreated Flt3L untreated Flt3L α-CD40
α-CD40 α-CD40 + Flt3L α-CD40 α-CD40 + Flt3L STINGag + Flt3L α-CD40 + Flt3L

L M
untreated α-CD40 + Flt3L untreated α-CD40 + Flt3L
12 * 10 *
8
% of tumor tissue
% of tumor tissue

9
Trichrome (collagen)

6
6
4
3
α-SMA

0 0
untreated Flt3L untreated Flt3L
α-CD40 α-CD40 + Flt3L α-CD40 α-CD40 + Flt3L

(legend on next page)

300 Cancer Cell 37, 289–307, March 16, 2020

observed in orthotopic Kras-Ink tumors that do not express OVA/
GFP (Figures 7D–7F and S7A), suggesting that these responses
were not specific to the genetically engineered mouse model
(GEMM) or expression of exogenous antigen. Notably, triple
therapy extended survival over RT alone (Figure 7G). A different
strategy in the KPPC GEMM involving RT at the induction step
also improved treatment efficacy and survival benefit (Figures
7H–7J). In conclusion, amplifying cDC density and function
might be a desirable strategy to augment the impact of RT and
similar treatments in PDAC.
DISCUSSION
In lieu of human tissue analyses, spontaneous mouse models are
invaluable for defining how immune surveillance gradually
becomes ineffective as tumors progress. In our unperturbed
model of PDAC, strong antigenicity is insufficient to drive
T cell-dependent tumor control and does not elicit immune edit-
ing. This is in contrast to recent work in transplant models (Evans
et al., 2016), underscoring how outcomes can differ depending
on the initial inflammatory context of the model system. Notably,
while the genetic approach in our study avoids the wounding
associated with orthotopic models, one limitation may be that
a large portion of acinar cells receives both oncogenic mutations
and neoantigen; this should be taken into consideration. None-
theless, it is plausible that the pancreas responds differently to
neoplastic cues when compared with mucosal/barrier organs
such as the lung (Salmon et al., 2019). Frequent environmental
insults and mitogens can entrain a rapid response to lesions in
the airway epithelia (Lelkes et al., 2014), and the lung-draining
lymphatics are amenable to efficient T cell priming (Cook and
Bottomly, 2007). In contrast, the pancreas and its lymphatic
drainage might not be designed to be at this heightened state,
tempering T cell immunity against antigenic lesions and allowing
fibro-inflammatory programs to dominate. Additionally, local mi-
crobial communities can influence immune responses to lung or
skin tumor antigens in a different context when compared with
the hepatobiliary/pancreatic tract, wherein the gut-proximal
commensal population is divergent (Jin et al., 2019; Pushalkar
et al., 2018; Routy et al., 2018). Overall, our comparative studies
suggest that anti-tumor surveillance in PDAC is heavily influ-
enced by a ‘‘hard-wired’’ program.
TH17 cells and their associated cytokines have been sepa-
rately implicated in tumor-promoting inflammation, fibrosis, neo-
vascularization, and recruitment of inflammatory myeloid cells
(Grivennikov et al., 2012; McAllister et al., 2014; Ochi et al.,
2012). Our observations suggest an interplay between these
TME compartments via immune-cell-derived TNF-a and
IL-17A, which can drive mitogenic signaling in PDAC and other
tumors through alternative p38MAPK activation (Alam et al.,
2015) and nuclear factor kB (NF-kB) signaling (Charles et al.,
2009; De Simone et al., 2015; Egberts et al., 2008). Despite their
plasticity, TH2 and TH17 infiltrates have been linked to worse out-
comes in PDAC patients (Bellone et al., 1999; De Monte et al.,
2011, 2016; Fukunaga et al., 2004; Wang et al., 2017). Work
from other groups have shown that TREGS also play an important
role in sustaining pancreatic tumorigenesis (Zhang et al., 2014),
although we did not observe this to be prominent in our study.
Human PDAC patients have low numbers of DCs that become
rarer with tumor progression (Dallal et al., 2002; Hiraoka et al.,
2011). Such an absence or dysfunction of DCs can magnify
unproductive TH cell responses (Furuhashi et al., 2012; Ibrahim
et al., 2012; Ochi et al., 2012). In agreement, cDC mobilization
at PANIN stages reduced pathogenic TH17 activity in the
pancreas and limited progression, supporting a protective TH1
role at this early stage (Bedrosian et al., 2011; Henning et al.,
2013). The influx of cDCs in our study was associated with
concomitant reduction in collagen deposition, which might
further benefit antigen sampling and improved trafficking to
dLNs (Hugues, 2010). Importantly, these observations insinuate
that PDAC TME retains its capacity for TH1 and CTL activity and
is held back by insufficient cDCs.
cDC2s represent a heterogeneous population that can have
tumor-suppressive roles based on the inflammatory context
(Binnewies et al., 2019; Brown et al., 2019; Laoui et al., 2016).
CD11b+ TAMs/DCs that skew immunity toward TH2 responses
have been described both in the PDAC TME (Ochi et al., 2012)
and metastases (Kenkel et al., 2017). Tumor-permissive roles
for CD11b+ DCs via FOXP3+ TREGs or FOXP3neg regulatory TR1
cells have also been reported (Barilla et al., 2019; Jang et al.,

Figure 6. Enhancing cDC Infiltration and Activation in Established PDAC Leads to Disease Stabilization
(A) Schematic of Flt3L administration in ultrasound-diagnosed KPPC-OG mice. n = 5–8 mice/group.
(B) Density of (left) CD103+ cDC1s and CD11b+ cDC2s in tumors, and (right) migratory cDC1 and cDC2 populations in respective dLNs of KPPC-OG mice treated
with Flt3L. n = 5–8 mice/group.
(C) Density of CD8+ T cells and CD4+ TH cells, and frequency of CD4+ TREG in tumors of KPPC-OG mice treated with Flt3L. n = 5–8 mice/group.
(D) Density of cDC1s and cDC2s in tumors of KPPC-OG mice treated as described. n = 7–8 mice/group.
(E) Density of CD8+ T cells and CD4+ TH cells, and frequency of CD4+ TREGS in tumors of KPPC-OG mice treated as described. n = 7–8 mice/group.
(F) Density of OVA-specific CD8+ T cells in tumors of treated KPPC-OG mice. n = 5–8 mice/group.
(G) Representative immunohistochemistry of CD8+ T cells (brown) and CK19+ tumor lesions (pink) in KPPC-OG mice treated as indicated. Right: cumulative CD8+
T cell density within 50 mm of CK19+ lesions. n = 5–8 mice/group.
(H) Density of tumor-infiltrating NK cells, NKT cells, and gd-T cells in treated KPPC-OG mice. n = 5–8 mice/group.
(I) Representative flow histogram indicating ZsGreen in migratory cDC1s from draining nodes of KPPC-Z treated as indicated. Right: absolute number of
migratory cDC subsets that have ZsGreen. n = 3–4 mice/group.
(J) Density of OVA-specific CD8+ T cells in draining lymph nodes of treated KPPC-OG mice. n = 5–8 mice/group.
(K) Tumor growth quantified by ultrasound measurements over 2 weeks of treatment. Right: individual traces of untreated and anti-CD40 plus Flt3L combination
cohorts. n = 5–8 mice/group.
(L) Representative Masson’s trichrome staining with quantification in KPPC-OG mice treated as denoted. n = 5–8 mice/group.
(M) Representative a-SMA staining with quantification in KPPC-OG mice treated as denoted. n = 5–8 mice/group.
Data were pooled across multiple independent experiments for all treatments. Data are presented as mean ± SEM. For comparisons between any two groups,
Student’s two-tailed t test was used. ns, not significant; *p < 0.05, **p < 0.01. Scale bars, 100 mm (G) and 500 mm (L, M). See also Figure S6.

Cancer Cell 37, 289–307, March 16, 2020 301

A KPPC-OG B C Dayy 7 Dayy 14
0.25 200
*

untreated
% change in volume
Flt3L 0.20 150

volume measure (cm3)

0 RT
α-CD40

(D7 to D14)
0.15 100

0.10 50

α-CD40 + Flt3L
RT
(8Gy x 3)

+ RT
0.05 0
*
9 -50
0 7 14
14 days since treatment start

untreated α-CD40 + Flt3L untreated α-CD40 + Flt3L

RT alone α-CD40 + Flt3L + RT RT alone α-CD40 + Flt3L + RT

D KI E F G
tumor implant 600
0.9 100
day –8 *
400
% change in volume
RT 80
Flt3L
volume measure (cm3)

overall survival (%)

0
(D6 to D13)
α-CD40 0.6
200 60 **
20 ns
40
0.3 0
RT
(8Gy x 3) 20
-20 RT
9
* -40 0
0 7 14 10 20 30 40 50
days since treatment start days since RT
14
untreated RT + Flt3L untreated RT + Flt3L untreated
RT alone RT + α-CD40 RT alone RT + α-CD40 RT alone
RT + α-CD40 + Flt3L RT + α-CD40 + Flt3L RT + α-CD40 + Flt3L

H KPPC I J
100
0.6
** 80
0 Flt3L 0.5
volume measure (cm3)

overall survival (%)

α-CD40
RT 0.4 60
(8Gy x 3) 0.3
RT 40 **
0.2
0.1 20
RT
9 0
0 7 14 21 28 35 20 40 60
14 days since treatment start days since treatment start

RT alone (n=9) RT alone (n=9)

α-CD40 + Flt3L + RT (n=13) α-CD40 + Flt3L + RT (n=13)

Figure 7. cDC-Directed Therapy Renders PDAC Responsive to Radiation Therapy

(A) Dosage schema for administration of radiation (RT) in KPPC-OG mice treated with Flt3L and anti-CD40 upon ultrasound-based tumor diagnosis at day 0.
(B) KPPC-OG tumor growth kinetics quantified by ultrasound measurements over 2 weeks of treatment. n = 8 mice/group.
(C) Percentage change in KPPC-OG tumor volume after RT (day 7 to day 14) with representative ultrasound images. n = 8 mice/group. Scale bar, 5 mm.
(D) Dosage schema for radiation (RT) in orthotopic Kras-Ink model treated with Flt3L and anti-CD40 upon tumor diagnosis.
(E) Kras-Ink tumor growth kinetics quantified by ultrasound measurements over 2 weeks of treatment. n = 8 mice/group.
(F) Percentage change in Kras-Ink tumor volume after RT (day 6 to day 13). n = 8 mice/group.
(G) Kaplan-Meier survival curve for Kras-Ink orthotopic tumor-bearing mice undergoing RT alone or RT in conjunction with Flt3L and anti-CD40. n = 9–14 mice/group.
(H) Dosage schema for administration of radiation (RT) in KPPC mice treated with Flt3L and anti-CD40 upon ultrasound-based tumor diagnosis at day 0.
(I) KPPC tumor growth kinetics quantified by ultrasound measurements over 5 weeks since starting treatment. n = 8 mice/group.
(J) Kaplan-Meier survival curve for KPPC mice undergoing RT alone or RT in conjunction with Flt3L and anti-CD40. n = 10–16 mice/group.
Data were pooled across multiple experiments for (A), (B), (C), (H), (I), and (J), and are representative of two independent experiments for (D) to (G). Data are presented as
mean ± SEM. For comparisons between any two groups, Student’s two-tailed t test was used. ns, not significant; *p < 0.05, **p < 0.01. See also Figure S7.

302 Cancer Cell 37, 289–307, March 16, 2020

2017). While we limited our scope to canonical cDC1 and cDC2 STAR+METHODS
subsets, recent work has demonstrated greater heterogeneity
and plasticity (Brown et al., 2019; Zilionis et al., 2019). It will Detailed methods are provided in the online version of this paper
become important to map cDC subset distribution and localiza- and include the following:
tion, and to study their impact on PDAC pathogenesis as well as
treatment response. d KEY RESOURCES TABLE
Prior trials involving Flt3L monotherapy have not shown d LEAD CONTACT AND MATERIAL AVAILABILITY
benefit due to a lack of appropriate DC activation and licensing d EXPERIMENTAL MODEL AND SUBJECT DETAILS
(Freedman et al., 2003; Morse et al., 2000). The paradigm has B Genetic Mice and Other Models
B Human Subjects
now shifted to include strategies for enhancing DC function in
the TME, and trials are now under way in lymphoma, squamous B Cell Lines

carcinoma, and non-small cell lung cancer (NCT03789097, d METHOD DETAILS

B Tissue Harvest
NCT02839265). The described therapeutic strategy targeting
B Orthotopic Implantations
anti-CD40 and Flt3L is also being tested for solid malignancies
(NCT03329950). This combination caused a dramatic increase B Perinatal and Immunotherapeutic Neutralizing Anti-

in tumoral cDCs and CD8+ T cells, despite mobilization from bodies

B Adoptive T Cell Transfer
bone marrow being similar to Flt3L monotherapy. This could
be due to enhanced DC survival in situ, as signaling downstream B DC-Modulatory Therapy and Radiation Therapy

of RANK and non-canonical NF-kB signaling upon CD40 ligation B Cytotoxicity and Degranulation Assays
B Bone Marrow Chimerism
is known to enhance DC survival (Hou and Van Parijs, 2004; Miga
et al., 2001; Ouaaz et al., 2002). Flt3L and CD40 agonism have B Vaccination Strategy

been shown to independently enhance IFN-g and IL-12 produc- B Immunohistochemical Staining
B Immunofluorescence Staining
tion (Borges et al., 1999; Chaudhry et al., 2006; Garris et al.,
B Multiparametric Flow Cytometry
2018), and their combined activity could impose TH1 immunity
in the otherwise suppressive TME. Even though CD40 agonism B Mass Cytometry
B ELISA
has been shown to rely on Batf3-dependent DCs in PDAC (Byrne
B Ex Vivo cDC Cross-Presentation Assay
and Vonderheide, 2016), it is also known to be mediated by
TAMs (Hoves et al., 2018; Stromnes et al., 2019). Multiple B Western Immunoblot
B RNA Isolation
myeloid cell types could be mediating CD40-dependent immu-
B RNA Sequencing and Analysis
nity in the PDAC TME; this redundancy could be of benefit to
patients with extremely low numbers of cDCs. B TCGA Patient Analysis

The described cDC-targeted strategy may have the added d QUANTIFICATION AND STATISTICAL ANALYSIS
B Image Analysis
benefit of altering the TME for fully integrated immune killing.
The increase in intratumoral NK and NKT cells via Flt3L can B Statistical Analysis

have a profound effect on sustaining DC-T cell interaction
and TH1 help (Barry et al., 2018; Böttcher et al., 2018; Nair
and Dhodapkar, 2017). Additionally, stromal remodeling and
resolution of fibrosis can facilitate drug delivery and enhance
CTL activity in the TME (Feig et al., 2013; Jiang et al., 2016;
Stromnes et al., 2015). Notably, CD40 agonism has been
shown to reduce pathogenic fibrosis by enhancing the matrix
remodeling properties of tumor-infiltrative monocytes (Long
et al., 2016). In addition to facilitating CTL activity and access,
stromal remodeling has a positive effect on cDC migration, an-
tigen sampling, and activity (Boissonnas et al., 2013; Hope
et al., 2017); this can be advantageous for sustaining DCs in
similarly ‘‘insulated’’ solid tumors.
Going forward, it becomes important to experimentally deter-
mine the best treatment window in which to administer inductive
RT to maximize priming capacity of mobilized cDCs and main-
tain their T cell-assistive function in the TME. It might also be
germane to combine this treatment with DC agonistic pathways
such as OX40 or 4-1BB to further benefit T cell co-stimulation.
Meaningful interventions in PDAC will likely necessitate such
combinations (Baird et al., 2016; Hammerich et al., 2019; Rech
et al., 2018). In conclusion, increasing the numbers and/or activ-
ity of scarce cDCs in PDAC is a distinct strategy that could
improve cytotoxic or immune-based therapies that by them-
selves are poorly effective in the treatment of this disease.
d DATA AND SOFTWARE AVAILABILITY
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2020.02.008.
ACKNOWLEDGMENTS
S.H. was supported by NCI predoctoral-to-postdoctoral fellowship
F99CA223043. J.K.S. was supported by NCI R01CA181745, and the Small An-
imal Radiation Research Platform was purchased with S10OD020136. D.G.D.
was supported by NCI P50CA196510, R01CA177670, R01CA203890,
P30CA09184215, and BJC Cancer Frontier Fund. C.J.D. was supported by
1DP5OD026427. We thank Brian Ruffell for his feedback while drafting the
manuscript. We thank the Genome Technology Access Center for help with
genomic analyses, supported by NCI P30CA91842 and by ICTS/CTSA
UL1RR024992 from the NCRR and the NIH. We also thank the Digestive Dis-
ease Research Core Center for histological help, supported by NIDDK
P30DK052574. Imaging experiments were performed through the use of
Washington University Center for Cellular Imaging supported by CDI-CORE-
2015-505 and the Foundation for Barnes-Jewish Hospital.
AUTHOR CONTRIBUTIONS
Conceptualization, S.H. and D.G.D.; Methodology, S.H., C.J.D., J.K.S., and
D.G.D.; Investigation, S.H., V.E.K., B.H.H., C.Z., M.A.B., B.L.K., J.P.T.,
G.D.H., J.M.B., C.M., K.B.L., and K.A.A.; Writing – Original Draft, S.H.; Writing –
Review & Editing, S.H., R.C.F., and D.G.D.; Funding, S.H. and D.G.D.;

Cancer Cell 37, 289–307, March 16, 2020 303

Resources, B.E.R., W.G.H., R.C.F., K.M.M., C.J.D., J.K.S., and D.G.D.; Super-
vision, D.G.D.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: June 5, 2019
Revised: December 4, 2019
Accepted: February 14, 2020
Published: March 16, 2020
REFERENCES
Ager, C.R., Reilley, M.J., Nicholas, C., Bartkowiak, T., Jaiswal, A.R., and
Curran, M.A. (2017). Intratumoral STING activation with T-cell checkpoint
modulation generates systemic antitumor immunity. Cancer Immunol. Res.
5, 676–684.
Alam, M.S., Gaida, M.M., Bergmann, F., Lasitschka, F., Giese, T., Giese, N.A.,
Hackert, T., Hinz, U., Hussain, S.P., Kozlov, S.V., and Ashwell, J.D. (2015).
Selective inhibition of the p38 alternative activation pathway in infiltrating
T cells inhibits pancreatic cancer progression. Nat. Med. 21, 1337–1343.
Bailey, P., Chang, D.K., Forget, M.A., Lucas, F.A.S., Alvarez, H.A., Haymaker,
C., Chattopadhyay, C., Kim, S.H., Ekmekcioglu, S., Grimm, E.A., et al. (2016a).
Exploiting the neoantigen landscape for immunotherapy of pancreatic ductal
adenocarcinoma. Sci. Rep. 6, 35848.
Bailey, P., Chang, D.K., Nones, K., Johns, A.L., Patch, A.M., Gingras, M.C.,
Miller, D.K., Christ, A.N., Bruxner, T.J., Quinn, M.C., et al. (2016b). Genomic
analyses identify molecular subtypes of pancreatic cancer. Nature 531, 47–52.
Baird, J.R., Friedman, D., Cottam, B., Dubensky, T.W., Kanne, D.B., Bambina,
S., Bahjat, K., Crittenden, M.R., and Gough, M.J. (2016). Radiotherapy com-
bined with novel STING-targeting oligonucleotides results in regression of
established tumors. Cancer Res. 76, 50–61.
Balaban, E.P., Mangu, P.B., Khorana, A.A., Shah, M.A., Mukherjee, S., Crane,
C.H., Javle, M.M., Eads, J.R., Allen, P., Ko, A.H., et al. (2016). Locally
advanced, unresectable pancreatic cancer: American Society of Clinical
Oncology clinical practice guideline. J. Clin. Oncol. 34, 2654–2668.
Balachandran, V.P., Luksza, M., Zhao, J.N., Makarov, V., Moral, J.A., Remark,
R., Herbst, B., Askan, G., Bhanot, U., Senbabaoglu, Y., et al. (2017).
Identification of unique neoantigen qualities in long-term survivors of pancre-
atic cancer. Nature 551, S12–S16.
Barilla, R.M., Diskin, B., Caso, R.C., Lee, K.B., Mohan, N., Buttar, C., Adam, S.,
Sekendiz, Z., Wang, J., Salas, R.D., et al. (2019). Specialized dendritic cells
induce tumor-promoting IL-10+IL-17+ FoxP3neg regulatory CD4+ T cells in
pancreatic carcinoma. Nat. Commun. 10, 1424.
Barry, K.C., Hsu, J., Broz, M.L., Cueto, F.J., Binnewies, M., Combes, A.J.,
Nelson, A.E., Loo, K., Kumar, R., Rosenblum, M.D., et al. (2018). A natural
killer-dendritic cell axis defines checkpoint therapy-responsive tumor micro-
environments. Nat. Med. 24, 1178–1191.
Beatty, G.L., Chiorean, E.G., Fishman, M.P., Saboury, B., Teitelbaum, U.R.,
Sun, W., Huhn, R.D., Song, W., Li, D., Sharp, L.L., et al. (2011). CD40 agonists
alter tumor stroma and show efficacy against pancreatic carcinoma in mice
and humans. Science 331, 1612–1616.
Beatty, G.L., Winograd, R., Evans, R.A., Long, K.B., Luque, S.L., Lee, J.W.,
Clendenin, C., Gladney, W.L., Knoblock, D.M., Guirnalda, P.D., and
Vonderheide, R.H. (2015). Exclusion of T cells from pancreatic carcinomas in
mice is regulated by Ly6Clow F4/80+ extratumoral macrophages.
Gastroenterology 149, 201–210.
Bedrosian, A.S., Nguyen, A.H., Hackman, M., Connolly, M.K., Malhotra, A.,
Ibrahim, J., Cieza-Rubio, N.E., Henning, J.R., Barilla, R., Rehman, A., et al.
(2011). Dendritic cells promote pancreatic viability in mice with acute pancre-
atitis. Gastroenterology 141, 1915–1926.e4.
Bellone, G., Carbone, A., Smirne, C., Scirelli, T., Buffolino, A., Novarino, A.,
Stacchini, A., Bertetto, O., Palestro, G., Sorio, C., et al. (2014). Cooperative in-
duction of a tolerogenic dendritic cell phenotype by cytokines secreted by
pancreatic carcinoma cells. J. Immunol. 177, 3448–3460.
304 Cancer Cell 37, 289–307, March 16, 2020
Bellone, G., Turletti, A., Artusio, E., Mareschi, K., Carbone, A., Tibaudi, D.,
Robecchi, A., Emanuelli, G., and Rodeck, U. (1999). Tumor-associated trans-
forming growth factor-beta and interleukin-10 contribute to a systemic Th2
immune phenotype in pancreatic carcinoma patients. Am. J. Pathol. 155,
537–547.
Binnewies, M., Mujal, A.M., Pollack, J.L., Combes, A.J., Hardison, E.A., Barry,
K.C., Tsui, J., Ruhland, M.K., Kersten, K., Abushawish, M.A., et al. (2019).
Unleashing type-2 dendritic cells to drive protective antitumor CD4+ T cell
immunity. Cell 177, 556–571.e16.
Boissonnas, A., Licata, F., Poupel, L., Jacquelin, S., Fetler, L., Krumeich, S.,
Théry, C., Amigorena, S., and Combadière, C. (2013). CD8+ tumor-infiltrating
T cells are trapped in the tumor-dendritic cell network. Neoplasia 15, 85–94.
Borges, L., Miller, R.E., Jones, J.S., Ariail, K., Whitmore, J., Fanslow, W., and
Lynch, D.H. (1999). Synergistic action of fms-like tyrosine kinase 3 ligand and
CD40 ligand in the induction of dendritic cells and generation of antitumor
immunity in vivo. J. Immunol. 163, 1289–1297.
Böttcher, J.P., Bonavita, E., Chakravarty, P., Blees, H., Cabeza-Cabrerizo, M.,
Sammicheli, S., Rogers, N.C., Sahai, E., Zelenay, S., and Reis e Sousa, C.
(2018). NK cells stimulate recruitment of cDC1 into the tumor microenviron-
ment promoting cancer immune control. Cell 172, 1022–1037.e14.
€hler, S., Zinselmeyer, B.H., Raju, S., Nitschke, M., Suleiman, H., Saunders,
Bra
B.T., Johnson, M.W., Böhner, A.M.C., Viehmann, S.F., Theisen, D.J., et al.
(2018). Opposing roles of dendritic cell subsets in experimental GN. J. Am.
Soc. Nephrol. 29, 138–154.
Brahmer, J.R., Tykodi, S.S., Chow, L.Q., Hwu, W.J., Topalian, S.L., Hwu, P.,
Drake, C.G., Camacho, L.H., Kauh, J., Odunsi, K., et al. (2012). Safety and
activity of anti-PD-L1 antibody in patients with advanced cancer. N. Engl. J.
Med. 366, 2455–2465.
Brown, C.C., Gudjonson, H., Pritykin, Y., Deep, D., Lavallee, V.P., Mendoza,
A., Fromme, R., Mazutis, L., Ariyan, C., Leslie, C., et al. (2019).
Transcriptional basis of mouse and human dendritic cell heterogeneity. Cell
179, 846–863 e824.
Byrne, K.T., and Vonderheide, R.H. (2016). CD40 stimulation obviates innate
sensors and drives T cell immunity in cancer. Cell Rep. 15, 2719–2732.
Charles, K.A., Kulbe, H., Soper, R., Escorcio-Correia, M., Lawrence, T.,
Schultheis, A., Chakravarty, P., Thompson, R.G., Kollias, G., Smyth, J.F.,
et al. (2009). The tumor-promoting actions of TNF-a involve TNFR1 and
IL-17 in ovarian cancer in mice and humans. J. Clin. Invest. 119, 3011–3023.
Chaudhry, U.I., Katz, S.C., Kingham, T.P., Pillarisetty, V.G., Raab, J.R., Shah,
A.B., and DeMatteo, R.P. (2006). In vivo overexpression of Flt3 ligand expands
and activates murine spleen natural killer dendritic cells. FASEB J. 20,
982–984.
Clark, C.E., Beatty, G.L., and Vonderheide, R.H. (2009). Immunosurveillance of
pancreatic adenocarcinoma: insights from genetically engineered mouse
models of cancer. Cancer Lett. 279, 1–7.
Cook, D.N., and Bottomly, K. (2007). Innate immune control of pulmonary
dendritic cell trafficking. Proc. Am. Thorac. Soc. 4, 234–239.
Corrales, L., Glickman, L.H., McWhirter, S.M., Kanne, D.B., Sivick, K.E.,
Katibah, G.E., Woo, S.R., Lemmens, E., Banda, T., Leong, J.J., et al. (2015).
Direct activation of STING in the tumor microenvironment leads to potent
and systemic tumor regression and immunity. Cell Rep. 11, 1018–1030.
Cristescu, R., Mogg, R., Ayers, M., Albright, A., Murphy, E., Yearley, J., Sher,
X., Liu, X.Q., Lu, H., Nebozhyn, M., et al. (2018). Pan-tumor genomic
biomarkers for PD-1 checkpoint blockade-based immunotherapy. Science
362, eaar3593.
Dallal, R.M., Christakos, P., Lee, K., Egawa, S., Son, Y.I., and Lotze, M.T.
(2002). Paucity of dendritic cells in pancreatic cancer. Surgery 131, 135–138.
de Mingo Pulido, Á., Gardner, A., Hiebler, S., Soliman, H., Rugo, H.S.,
Krummel, M.F., Coussens, L.M., and Ruffell, B. (2018). TIM-3 regulates
CD103+ dendritic cell function and response to chemotherapy in breast
cancer. Cancer Cell 33, 60–74.e66.
De Monte, L., Reni, M., Tassi, E., Clavenna, D., Papa, I., Recalde, H., Braga,
M., Di Carlo, V., Doglioni, C., and Protti, M.P. (2011). Intratumor T helper
type 2 cell infiltrate correlates with cancer-associated fibroblast thymic
stromal lymphopoietin production and reduced survival in pancreatic cancer.
J. Exp. Med. 208, 469–478.
De Monte, L., Woermann, S., Brunetto, E., Heltai, S., Magliacane, G., Reni, M.,
Paganoni, A.M., Recalde, H., Mondino, A., Falconi, M., et al. (2016). Basophil
recruitment into tumor draining lymph nodes correlates with Th2 inflammation
and reduced survival in pancreatic cancer patients. Cancer Res. 76,
1792–1803.
De Simone, V., Franzè, E., Ronchetti, G., Colantoni, A., Fantini, M.C., Di Fusco,
D., Sica, G.S., Sileri, P., MacDonald, T.T., Pallone, F., et al. (2015). Th17-type
cytokines, IL-6 and TNF-a synergistically activate STAT3 and NF-kB to
promote colorectal cancer cell growth. Oncogene 34, 3493–3503.
DuPage, M., Cheung, A.F., Mazumdar, C., Winslow, M.M., Bronson, R.,
Schmidt, L.M., Crowley, D., Chen, J., and Jacks, T. (2011). Endogenous
T cell responses to antigens expressed in lung adenocarcinomas delay malig-
nant tumor progression. Cancer Cell 19, 72–85.
DuPage, M., Mazumdar, C., Schmidt, L.M., Cheung, A.F., and Jacks, T. (2012).
Expression of tumour-specific antigens underlies cancer immunoediting.
Nature 482, 405–409.
Egberts, J.H., Cloosters, V., Noack, A., Schniewind, B., Thon, L., Klose, S.,
Kettler, B., von Forstner, C., Kneitz, C., Tepel, J., et al. (2008). Anti-tumor
necrosis factor therapy inhibits pancreatic tumor growth and metastasis.
Cancer Res. 68, 1443–1450.
Evans, R.A., Diamond, M.S., Rech, A.J., Chao, T., Richardson, M.W., Lin, J.H.,
Bajor, D.L., Byrne, K.T., Stanger, B.Z., Riley, J.L., et al. (2016). Lack of immu-
noediting in murine pancreatic cancer reversed with neoantigen. JCI Insight
1, 1–16.
Feig, C., Jones, J.O., Kraman, M., Wells, R.J.B., Deonarine, A., Chan, D.S.,
Connell, C.M., Roberts, E.W., Zhao, Q., Caballero, O.L., et al. (2013).
Targeting CXCL12 from FAP-expressing carcinoma-associated fibroblasts
synergizes with anti-PD-L1 immunotherapy in pancreatic cancer. Proc. Natl.
Acad. Sci. U S A 110, 20212–20217.
Freedman, R.S., Vadhan-Raj, S., Butts, C., Savary, C., Melichar, B.,
Verschraegen, C., Kavanagh, J.J., Hicks, M.E., Levy, L.B., Folloder, J.K.,
and Garcia, M.E. (2003). Pilot study of Flt3 ligand comparing intraperitoneal
with subcutaneous routes on hematologic and immunologic responses in
patients with peritoneal carcinomatosis and mesotheliomas. Clin. Cancer
Res. 9, 5228–5237.
Fukunaga, A., Miyamoto, M., Cho, Y., Murakami, S., Kawarada, Y., Oshikiri, T.,
Kato, K., Kurokawa, T., Suzuoki, M., Nakakubo, Y., et al. (2004). CD8+ tumor-
infiltrating lymphocytes together with CD4+ tumor-infiltrating lymphocytes and
dendritic cells improve the prognosis of patients with pancreatic adenocarci-
noma. Pancreas 28, e26–e31.
Furuhashi, K., Suda, T., Hasegawa, H., Suzuki, Y., Hashimoto, D., Enomoto,
N., Fujisawa, T., Nakamura, Y., Inui, N., Shibata, K., et al. (2012). Mouse
lung CD103+ and CD11b high dendritic cells preferentially induce distinct
CD4+ T-cell responses. Am. J. Respir. Cell Mol. Biol. 46, 165–172.
Gardner, A., and Ruffell, B. (2016). Dendritic cells and cancer immunity. Trends
Immunol. 37, 855–865.
Garris, C.S., Arlauckas, S.P., Kohler, R.H., Trefny, M.P., Garren, S., Piot, C.,
Engblom, C., Pfirschke, C., Siwicki, M., Gungabeesoon, J., et al. (2018).
Successful anti-PD-1 cancer immunotherapy requires T cell-dendritic cell
crosstalk involving the cytokines IFN-g and IL-12. Immunity 49, 1148–1161.e7.
Gopinathan, a., Morton, J.P., Jodrell, D.I., and Sansom, O.J. (2015). GEMMs
as preclinical models for testing pancreatic cancer therapies. Dis. Model.
Mech. 8, 1185–1200.
Grivennikov, S.I., Wang, K., Mucida, D., Stewart, C.A., Schnabl, B., Jauch, D.,
Taniguchi, K., Yu, G.Y., Österreicher, C.H., Hung, K.E., et al. (2012). Adenoma-
linked barrier defects and microbial products drive IL-23/IL-17-mediated
tumour growth. Nature 491, 254–258.
Gu, G., Dubauskaite, J., and Melton, D.A. (2002). Direct evidence for the
pancreatic lineage: NGN3+ cells are islet progenitors and are distinct from
duct progenitors. Development 129, 2447–2457.
Gubin, M.M., Zhang, X., Schuster, H., Caron, E., Ward, J.P., Noguchi, T.,
Ivanova, Y., Hundal, J., Arthur, C.D., Krebber, W.-j., et al. (2014). Checkpoint
blockade cancer immunotherapy targets tumour-specific mutant antigens.
Nature 515, 577–581.
Gunderson, A.J., Kaneda, M.M., Tsujikawa, T., Nguyen, A.V., Affara, N.I.,
Ruffell, B., Gorjestani, S., Liudahl, S.M., Truitt, M., Olson, P., et al. (2016).
Bruton tyrosine kinase-dependent immune cell cross-talk drives pancreas
cancer. Cancer Discov. 6, 270–285.
Hammerich, L., Marron, T.U., Upadhyay, R., Svensson-Arvelund, J., Dhainaut,
M., Hussein, S., Zhan, Y., Ostrowski, D., Yellin, M., Marsh, H., et al. (2019).
Systemic clinical tumor regressions and potentiation of PD1 blockade with
in situ vaccination. Nat. Med. 25, 814–824.
Henning, J.R., Graffeo, C.S., Rehman, A., Fallon, N.C., Zambirinis, C.P., Ochi,
A., Barilla, R., Jamal, M., Deutsch, M., Greco, S., et al. (2013). Dendritic cells
limit fibroinflammatory injury in nonalcoholic steatohepatitis in mice.
Hepatology 58, 589–602.
Hingorani, S.R., Petricoin, E.F., Maitra, A., Rajapakse, V., King, C., Jacobetz,
M.a., Ross, S., Conrads, T.P., Veenstra, T.D., Hitt, B.a., et al. (2003).
Preinvasive and invasive ductal pancreatic cancer and its early detection in
the mouse. Cancer Cell 4, 437–450.
Hiraoka, N., Yamazakiitoh, R., Ino, Y., Mizuguchi, Y., Yamada, T., Hirohashi,
S., and Kanai, Y. (2011). CXCL17 and ICAM2 are associated with a potential
anti-tumor immune response in early intraepithelial stages of human pancre-
atic carcinogenesis. Gastroenterology 140, 310–321.e4.
Hope, C., Emmerich, P.B., Papadas, A., Pagenkopf, A., Matkowskyj, K.A., Van
De Hey, D.R., Payne, S.N., Clipson, L., Callander, N.S., Hematti, P., et al.
(2017). Versican-derived matrikines regulate Batf3-dendritic cell differentiation
and promote T cell infiltration in colorectal cancer. J. Immunol. 199,
1933–1941.
Hou, W.S., and Van Parijs, L. (2004). A Bcl-2-dependent molecular timer
regulates the lifespan and immunogenicity of dendritic cells. Nat. Immunol.
5, 583–589.
Hoves, S., Ooi, C.H., Wolter, C., Sade, H., Bissinger, S., Schmittnaegel, M.,
Ast, O., Giusti, A.M., Wartha, K., Runza, V., et al. (2018). Rapid activation of
tumor-associated macrophages boosts preexisting tumor immunity. J. Exp.
Med. 215, 859–876.
Hugues, S. (2010). Dynamics of dendritic cell-T cell interactions: a role in T cell
outcome. Semin. Immunopathol. 32, 227–238.
Ibrahim, J., Nguyen, A.H., Rehman, A., Ochi, A., Jamal, M., Graffeo, C.S.,
Henning, J.R., Zambirinis, C.P., Fallon, N.C., Barilla, R., et al. (2012).
Dendritic cell populations with different concentrations of lipid regulate toler-
ance and immunity in mouse and human liver. Gastroenterology 143,
1061–1072.
Jang, J.E., Hajdu, C.H., Liot, C., Miller, G., Dustin, M.L., and Bar-Sagi, D.
(2017). Crosstalk between regulatory T cells and tumor-associated dendritic
cells negates anti-tumor immunity in pancreatic cancer. Cell Rep. 20, 558–571.
Jiang, H., Hegde, S., Knolhoff, B.L., Zhu, Y., Herndon, J.M., Meyer, M.A.,
Nywening, T.M., Hawkins, W.G., Shapiro, I.M., Weaver, D.T., et al. (2016).
Targeting focal adhesion kinase renders pancreatic cancers responsive to
checkpoint immunotherapy. Nat. Med. 22, 851–860.
Jin, C., Lagoudas, G.K., Zhao, C., Bullman, S., Bhutkar, A., Hu, B., Ameh, S.,
Sandel, D., Liang, X.S., Mazzilli, S., et al. (2019). Commensal microbiota
promote lung cancer development via gd T cells. Cell 176, 998–1013.e16.
Kenkel, J.A., Tseng, W.W., Davidson, M.G., Tolentino, L.L., Choi, O.,
Bhattacharya, N., Seeley, E.S., Winer, D.A., Reticker-Flynn, N.E., and
Engleman, E.G. (2017). An immunosuppressive dendritic cell subset accumu-
lates at secondary sites and promotes metastasis in pancreatic cancer.
Cancer Res. 77, 4158–4170.
Kieler, M., Unseld, M., Bianconi, D., and Prager, G. (2018). Challenges and
perspectives for immunotherapy in adenocarcinoma of the pancreas: the
cancer immunity cycle. Pancreas 47, 142–157.
Kim, M.P., Evans, D.B., Wang, H., Abbruzzese, J.L., Fleming, J.B., and Gallick,
G.E. (2009). Generation of orthotopic and heterotopic human pancreatic
cancer xenografts in immunodeficient mice. Nat. Protoc. 4, 1670–1680.
Kinkead, H.L., Hopkins, A., Lutz, E., Wu, A.A., Yarchoan, M., Cruz, K.,
Woolman, S., Vithayathil, T., Glickman, L.H., Ndubaku, C.O., et al. (2018).
Cancer Cell 37, 289–307, March 16, 2020 305
Combining STING-based neoantigen-targeted vaccine with checkpoint mod-
ulators enhances antitumor immunity in murine pancreatic cancer. JCI Insight
3, 122857.
Kunk, P.R., Bauer, T.W., Slingluff, C.L., and Rahma, O.E. (2016). From bench
to bedside a comprehensive review of pancreatic cancer immunotherapy.
J. Immunother. Cancer 4, 14.
Laoui, D., Keirsse, J., Morias, Y., Van Overmeire, E., Geeraerts, X., Elkrim, Y.,
Kiss, M., Bolli, E., Lahmar, Q., Sichien, D., et al. (2016). The tumour microen-
vironment harbours ontogenically distinct dendritic cell populations with
opposing effects on tumour immunity. Nat. Commun. 7, 13720.
Lelkes, E., Headley, M.B., Thornton, E.E., Looney, M.R., and Krummel, M.F.
(2014). The spatiotemporal cellular dynamics of lung immunity. Trends
Immunol. 35, 379–386.
Li, J., Byrne, K.T., Yan, F., Yamazoe, T., Chen, Z., Baslan, T., Richman, L.P.,
Lin, J.H., Sun, Y.H., Rech, A.J., et al. (2018). Tumor cell-intrinsic factors under-
lie heterogeneity of immune cell infiltration and response to immunotherapy.
Immunity 49, 178–193.e7.
Long, K.B., Gladney, W.L., Tooker, G.M., Graham, K., Fraietta, J.A., and
Beatty, G.L. (2016). IFNgamma and CCL2 cooperate to redirect tumor-infil-
trating monocytes to degrade fibrosis and enhance chemotherapy efficacy
in pancreatic carcinoma. Cancer Discov. 6, 400–413.
Ma, H.S., Poudel, B., Torres, E.R., Sidhom, J.-W., Robinson, T.M., Christmas,
B., Scott, B., Cruz, K., Woolman, S., Wall, V.Z., et al. (2019). A CD40 agonist
and PD-1 antagonist antibody reprogram the microenvironment of nonimmu-
nogenic tumors to allow T-cell–mediated anticancer activity. Cancer Immunol.
Res. 7, 428–442.
McAllister, F., Bailey, J.M., Alsina, J., Nirschl, C.J., Sharma, R., Fan, H.,
Rattigan, Y., Roeser, J.C., Lankapalli, R.H., Zhang, H., et al. (2014).
Oncogenic Kras activates a hematopoietic-to-epithelial IL-17 signaling axis
in preinvasive pancreatic neoplasia. Cancer Cell 25, 621–637.
Meyer, M.A., Baer, J.M., Knolhoff, B.L., Nywening, T.M., Panni, R.Z., Su, X.,
Weilbaecher, K.N., Hawkins, W.G., Ma, C., Fields, R.C., et al. (2018). Breast
and pancreatic cancer interrupt IRF8-dependent dendritic cell development
to overcome immune surveillance. Nat. Commun. 9, 1250.
Miga, A.J., Masters, S.R., Durell, B.G., Gonzalez, M., Jenkins, M.K.,
Maliszewski, C., Kikutani, H., Wade, W.F., and Noelle, R.J. (2001). Dendritic
cell longevity and T cell persistence is controlled by CD154-CD40 interactions.
Eur. J. Immunol. 31, 959–965.
Miyazaki, S., Miyazaki, T., Tashiro, F., Yamato, E., and Miyazaki, J.-i. (2005).
Development of a single-cassette system for spatiotemporal gene regulation
in mice. Biochem. Biophys. Res. Commun. 338, 1083–1088.
Morse, M.A., Nair, S., Fernandez-Casal, M., Deng, Y., St Peter, M., Williams,
R., Hobeika, A., Mosca, P., Clay, T., Cumming, R.I., et al. (2000).
Preoperative mobilization of circulating dendritic cells by Flt3 ligand adminis-
tration to patients with metastatic colon cancer. J. Clin. Oncol. 18, 3883–3893.
Morton, J.P., Timpson, P., Karim, S.a., Ridgway, R.a., Athineos, D., Doyle, B.,
Jamieson, N.B., Oien, K.a., Lowy, A.M., Brunton, V.G., et al. (2010). Mutant
p53 drives metastasis and overcomes growth arrest/senescence in pancreatic
cancer. Proc. Natl. Acad. Sci. U S A 107, 246–251.
Nair, S., and Dhodapkar, M.V. (2017). Natural killer T cells in cancer immuno-
therapy. Front. Immunol. 8, 1178.
Ngwa, W., Irabor, O.C., Schoenfeld, J.D., Hesser, J., Demaria, S., and
Formenti, S.C. (2018). Using immunotherapy to boost the abscopal effect.
Nat. Rev. Cancer 18, 313–322.
O’Donnell, J.S., Teng, M.W.L., and Smyth, M.J. (2019). Cancer immunoediting
and resistance to T cell-based immunotherapy. Nat. Rev. Clin. Oncol. 16,
151–167.
Ochi, A., Nguyen, A.H., Bedrosian, A.S., Mushlin, H.M., Zarbakhsh, S., Barilla,
R., Zambirinis, C.P., Fallon, N.C., Rehman, A., Pylayeva-Gupta, Y., et al.
(2012). MyD88 inhibition amplifies dendritic cell capacity to promote pancre-
atic carcinogenesis via Th2 cells. J. Exp. Med. 209, 1671–1687.
Ouaaz, F., Arron, J., Zheng, Y., Choi, Y., and Beg, A.A. (2002). Dendritic cell
development and survival require distinct NF-kappaB subunits. Immunity 16,
257–270.
306 Cancer Cell 37, 289–307, March 16, 2020
Poschke, I., Faryna, M., Bergmann, F., Flossdorf, M., Lauenstein, C., Hermes,
J., Hinz, U., Hank, T., Ehrenberg, R., Volkmar, M., et al. (2016). Identification
of a tumor-reactive T-cell repertoire in the immune infiltrate of patients
with resectable pancreatic ductal adenocarcinoma. OncoImmunology 5,
e1240859.
Pushalkar, S., Hundeyin, M., Daley, D., Zambirinis, C.P., Kurz, E., Mishra, A.,
Mohan, N., Aykut, B., Usyk, M., Torres, L.E., et al. (2018). The pancreatic
cancer microbiome promotes oncogenesis by induction of innate and adap-
tive immune suppression. Cancer Discov. 8, 403–416.
Pylayeva-Gupta, Y., Das, S., Handler, J.S., Hajdu, C.H., Coffre, M., Koralov,
S.B., and Bar-Sagi, D. (2016). IL35-producing B cells promote the develop-
ment of pancreatic neoplasia. Cancer Discov. 6, 247–255.
Ray, K.C., Moss, M.E., Franklin, J.L., Weaver, C.J., Higginbotham, J., Song, Y.,
Revetta, F.L., Blaine, S.A., Bridges, L.R., Guess, K.E., et al. (2014). Heparin-
binding epidermal growth factor-like growth factor eliminates constraints on
activated Kras to promote rapid onset of pancreatic neoplasia. Oncogene
33, 823–831.
Rech, A.J., Dada, H., Kotzin, J.J., Henao-Mejia, J., Minn, A.J., Victor, C.T.S.,
and Vonderheide, R.H. (2018). Radiotherapy and CD40 activation separately
augment immunity to checkpoint blockade in cancer. Cancer Res. 78,
4282–4291.
Rhim, A.D., Mirek, E.T., Aiello, N.M., Maitra, A., Bailey, J.M., McAllister, F.,
Reichert, M., Beatty, G.L., Rustgi, A.K., Vonderheide, R.H., et al. (2012).
EMT and dissemination precede pancreatic tumor formation. Cell 148,
349–361.
Roberts, E.W., Broz, M.L., Binnewies, M., Bogunovic, D., Bhardwaj, N.,
Krummel, M.F., Roberts, E.W., Broz, M.L., Binnewies, M., Headley, M.B.,
et al. (2016). Critical role for CD103+/CD141+ dendritic cells bearing CCR7
for tumor antigen trafficking and priming of T cell immunity in melanoma.
Cancer Cell 30, 324–336.
Routy, B., Le Chatelier, E., Derosa, L., Duong, C.P.M., Alou, M.T., Daillere, R.,
Fluckiger, A., Messaoudene, M., Rauber, C., Roberti, M.P., et al. (2018). Gut
microbiome influences efficacy of PD-1-based immunotherapy against
epithelial tumors. Science 359, 91–97.
Royal, R.E., Levy, C., Turner, K., Mathur, A., Hughes, M., Kammula, U.S.,
Sherry, R.M., Topalian, S.L., Yang, J.C., Lowy, I., and Rosenberg, S.a.
(2010). Phase 2 trial of single agent Ipilimumab (anti-CTLA-4) for locally
advanced or metastatic pancreatic adenocarcinoma. J. Immunother. 33,
828–833.
Salmon, H., Idoyaga, J., Rahman, A., Leboeuf, M., Remark, R., Jordan, S.,
Casanova-Acebes, M., Khudoynazarova, M., Agudo, J., Tung, N., et al.
(2016). Expansion and activation of CD103+ dendritic cell progenitors at the
tumor site enhances tumor responses to therapeutic PD-L1 and BRAF inhibi-
tion. Immunity 44, 924–938.
Salmon, H., Remark, R., Gnjatic, S., and Merad, M. (2019). Host tissue deter-
minants of tumour immunity. Nat. Rev. Cancer 19, 215–227.
Satpathy, A.T., KC, W., Albring, J.C., Edelson, B.T., Kretzer, N.M.,
Bhattacharya, D., Murphy, T.L., and Murphy, K.M. (2012). Zbtb46 expression
distinguishes classical dendritic cells and their committed progenitors from
other immune lineages the. J. Exp. Med. 209, 1135–1152.
Schietinger, A., Philip, M., Krisnawan, V.E., Chiu, E.Y., Delrow, J.J., Basom,
R.S., Lauer, P., Brockstedt, D.G., Knoblaugh, S.E., Ha €mmerling, G.J., et al.
(2016). Tumor-specific T cell dysfunction is a dynamic antigen-driven differen-
tiation program initiated early during tumorigenesis. Immunity 45, 389–401.
Schumacher, T.N., and Schreiber, R.D. (2015). Neoantigens in cancer immu-
notherapy. Science 348, 69–74.
Sharma, P., Hu-Lieskovan, S., Wargo, J.A., and Ribas, A. (2017). Primary,
adaptive, and acquired resistance to cancer immunotherapy. Cell 168,
707–723.
Sivick, K.E., Desbien, A.L., Glickman, L.H., Reiner, G.L., Corrales, L., Surh,
N.H., Hudson, T.E., Vu, U.T., Francica, B.J., Banda, T., et al. (2018).
Magnitude of therapeutic STING activation determines CD8+ T cell-mediated
anti-tumor immunity. Cell Rep. 25, 3074–3085.e5.
Spear, S., Candido, J.B., McDermott, J.R., Ghirelli, C., Maniati, E., Beers, S.A.,
Balkwill, F.R., Kocher, H.M., and Capasso, M. (2019). Discrepancies in the
tumor microenvironment of spontaneous and orthotopic murine models of
pancreatic cancer uncover a new immunostimulatory phenotype for B cells.
Front Immunol. 10, 542.
Spranger, S., Dai, D., Horton, B., and Gajewski, T.F. (2017). Tumor-residing
Batf3 dendritic cells are required for effector T cell trafficking and adoptive
T cell therapy. Cancer Cell 31, 711–723.e4.
Stromnes, I.M., Brockenbrough, J.S., Izeradjene, K., Carlson, M.a., Cuevas,
C., Simmons, R.M., Greenberg, P.D., and Hingorani, S.R. (2014). Targeted
depletion of an MDSC subset unmasks pancreatic ductal adenocarcinoma
to adaptive immunity. Gut 63, 1769–1781.
Stromnes, I.M., Burrack, A.L., Hulbert, A., Bonson, P., Black, C.,
Brockenbrough, J.S., Raynor, J.F., Spartz, E.J., Pierce, R.H., Greenberg,
P.D., and Hingorani, S.R. (2019). Differential effects of depleting versus
programming tumor-associated macrophages on engineered T cells in
pancreatic ductal adenocarcinoma. Cancer Immunol. Res. 7, 977–989.
Stromnes, I.M., Schmitt, T.M., Hulbert, A., Brockenbrough, J.S., Nguyen, H.N.,
Cuevas, C., Dotson, A.M., Tan, X., Hotes, J.L., Greenberg, P.D., and Hingorani,
S.R. (2015). T cells engineered against a native antigen can surmount immuno-
logic and physical barriers to treat pancreatic ductal adenocarcinoma. Cancer
Cell 28, 638–652.
Twyman-Saint Victor, C., Rech, A.J., Maity, A., Rengan, R., Pauken, K.E.,
Stelekati, E., Benci, J.L., Xu, B., Dada, H., Odorizzi, P.M., et al. (2015).
Radiation and dual checkpoint blockade activate non-redundant immune
mechanisms in cancer. Nature 520, 373–377.
Vonderheide, R.H. (2018). The immune revolution: a case for priming, not
checkpoint. Cancer Cell 33, 563–569.
Wang, S., Li, Z., and Hu, G. (2017). Prognostic role of intratumoral IL-17A
expression by immunohistochemistry in solid tumors: a meta-analysis.
Oncotarget 8, 66382–66391.
Zhang, Y., Velez-delgado, A., Mathew, E., Li, D., Mendez, F.M., Flannagan, K.,
Rhim, A.D., Simeone, D.M., Beatty, G.L., and Pasca, M. (2016). Myeloid cells
are required for PD-1/PD-L1 checkpoint activation and the establishment of an
immunosuppressive environment in pancreatic cancer. Gut 66, 124–136.
Zhang, Y., Yan, W., Mathew, E., Bednar, F., Wan, S., Collins, M.a., Evans, R.a.,
Welling, T.H., Vonderheide, R.H., and di Magliano, M.P. (2014). CD4+ T
lymphocyte ablation prevents pancreatic carcinogenesis in mice. Cancer
Immunol. Res. 2, 423–435.
Zhang, Y., Zoltan, M., Riquelme, E., Xu, H., Sahin, I., Castro-Pando, S.,
Montiel, M.F., Chang, K., Jiang, Z., Ling, J., et al. (2018). Immune cell produc-
tion of interleukin 17 induces stem cell features of pancreatic intraepithelial
neoplasia cells. Gastroenterology 155, 210–223.e3.
Zilionis, R., Engblom, C., Pfirschke, C., Savova, V., Zemmour, D., Saatcioglu,
H.D., Krishnan, I., Maroni, G., Meyerovitz, C.V., Kerwin, C.M., et al. (2019).
Single-cell transcriptomics of human and mouse lung cancers reveals
conserved myeloid populations across individuals and species. Immunity
50, 1317–1334.e10.
Cancer Cell 37, 289–307, March 16, 2020 307
 Cancer Cell

Article

Treatment-Induced Tumor Dormancy

through YAP-Mediated Transcriptional
Reprogramming of the Apoptotic Pathway
Kari J. Kurppa,1,23 Yao Liu,2,3,23 Ciric To,1 Tinghu Zhang,2,3 Mengyang Fan,2,3 Amir Vajdi,4 Erik H. Knelson,1 Yingtian Xie,1,5
Klothilda Lim,1,5 Paloma Cejas,1,5 Andrew Portell,1,6 Patrick H. Lizotte,1,6 Scott B. Ficarro,2,7,8,21 Shuai Li,9 Ting Chen,9
Heidi M. Haikala,1 Haiyun Wang,10 Magda Bahcall,1 Yang Gao,11,12 Sophia Shalhout,13,22 Steffen Boettcher,1,14
Bo Hee Shin,1 Tran Thai,1 Margaret K. Wilkens,15 Michelle L. Tillgren,15 Mierzhati Mushajiang,1 Man Xu,1,6 Jihyun Choi,1

(Author list continued on next page)

1Department of Medical Oncology, Dana-Farber Cancer Institute and Harvard Medical School, Boston, MA 02215, USA
2Department of Cancer Biology, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA 02215, USA
3Department of Biological Chemistry and Molecular Pharmacology, Harvard Medical School, Boston, MA 02215, USA
4Department of Informatics and Analytics, Dana-Farber Cancer Institute, Boston, MA 02215, USA
5Center for Functional Cancer Epigenetics, Dana-Farber Cancer Institute, Boston, MA 02215, USA
6Belfer Center for Applied Cancer Science, Dana-Farber Cancer Institute, Boston, MA 02215, USA
7Blais Proteomics Center, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA 02215, USA
8Department of Oncologic Pathology, Dana-Farber Cancer Institute, Boston, MA 02215, USA
9Division of Hematology & Medical Oncology, Laura and Isaac Perlmutter Cancer Center, New York University Langone Medical Center, New

York, NY, USA

10School of Life Science and Technology, Tongji University, 200092 Shanghai, China
11Department of Pediatric Oncology, Dana-Farber Cancer Institute, Boston, MA 02215, USA

(Affiliations continued on next page)

SUMMARY

Eradicating tumor dormancy that develops following epidermal growth factor receptor (EGFR) tyrosine
kinase inhibitor (TKI) treatment of EGFR-mutant non-small cell lung cancer, is an attractive therapeutic strat-
egy but the mechanisms governing this process are poorly understood. Blockade of ERK1/2 reactivation
following EGFR TKI treatment by combined EGFR/MEK inhibition uncovers cells that survive by entering a
senescence-like dormant state characterized by high YAP/TEAD activity. YAP/TEAD engage the epithelial-
to-mesenchymal transition transcription factor SLUG to directly repress pro-apoptotic BMF, limiting drug-
induced apoptosis. Pharmacological co-inhibition of YAP and TEAD, or genetic deletion of YAP1, all deplete
dormant cells by enhancing EGFR/MEK inhibition-induced apoptosis. Enhancing the initial efficacy of
targeted therapies could ultimately lead to prolonged treatment responses in cancer patients.

INTRODUCTION
Epidermal growth factor receptor (EGFR) tyrosine kinase inhibi-
tors (TKIs) are the standard of care for patients with advanced
EGFR-mutant non-small cell lung cancer (NSCLC) (Mok et al.,
2009; Rosell et al., 2012; Soria et al., 2018). However, acquired
resistance inevitably develops, limiting clinical efficacy (Cortot
and Ja€nne, 2014). In most cases, resistance arises after a dra-
matic initial response followed by a stable minimal residual dis-
ease (MRD), or dormant, state, with subsequent gradual growth

Significance

Genotype-directed therapy rarely, if ever, leads to complete tumor eradication. The residual tumors, following drug treat-
ment, can serve as reservoirs for the development of acquired drug resistance. Using EGFR-mutant lung cancer as a model
system, we identify YAP activation as a major counter regulatory mechanism promoting cell survival and dormancy
following sustained inhibition of EGFR and its downstream signaling, through transcriptional repression of the pro-
apoptotic protein BMF. Targeting YAP, indirectly using a tankyrase inhibitor or directly using a TEAD inhibitor, enhances
the apoptotic effects of EGFR/MEK inhibition. Our findings suggest that targeting YAP/TEAD following genotype-directed
therapy may be one strategy to enhance initial drug-induced apoptosis, reduce residual disease, and as such lead to
improved treatment outcomes.

104 Cancer Cell 37, 104–122, January 13, 2020 ª 2019 Elsevier Inc.
Arrien A. Bertram,1 Benjamin L. Ebert,1,14,16 Rameen Beroukhim,1,14,17 Pratiti Bandopadhayay,11,14,18 Mark M. Awad,1
Prafulla C. Gokhale,6,15 Paul T. Kirschmeier,1,6 Jarrod A. Marto,2,7,8,21 Fernando D. Camargo,13,22 Rizwan Haq,17,19
Cloud P. Paweletz,1,6 Kwok-Kin Wong,9 David A. Barbie,1 Henry W. Long,1,5 Nathanael S. Gray,2,3 and
€nne1,6,20,24,*
Pasi A. Ja
12Department of Pediatrics, Harvard Medical School, Boston, MA 02215, USA
13Stem Cell Program, Boston Children’s Hospital, Boston, MA 02215, USA
14Broad Institute of MIT and Harvard, Cambridge, MA 02142, USA
15Experimental Therapeutics Core, Dana-Farber Cancer Institute, Boston, MA 02210, USA
16Howard Hughes Medical Institute, Dana-Farber Cancer Institute, Boston, MA 02215, USA
17Department of Medicine, Harvard Medical School, Boston, MA 02115, USA
18Dana-Farber/Boston Children’s Cancer and Blood Disorders Center, Boston, MA 02115, USA
19Department of Molecular and Cellular Oncology, Dana-Farber Cancer Institute, Boston, MA 02215, USA
20Lowe Center for Thoracic Oncology, Dana-Farber Cancer Institute, 450 Brookline Avenue, LC4114, Boston, MA 02215, USA
21Department of Pathology, Brigham and Women’s Hospital and Harvard Medical School, Boston, MA 02215, USA
22Department of Stem Cell and Regenerative Biology, Harvard University, Cambridge, MA 02138, USA
23These authors contributed equally
24Lead Contact

*Correspondence: pasi_janne@dfci.harvard.edu
https://doi.org/10.1016/j.ccell.2019.12.006

of a drug-resistant tumor. Previous preclinical studies suggest
that, following EGFR TKI treatment, EGFR-mutant tumor cells
can enter a drug-tolerant state, reminiscent of dormancy in pa-
tients, allowing cells to evade apoptosis and survive (Hata
et al., 2016; Sharma et al., 2010). Over time, the drug-tolerant
cells can acquire resistance through mutational or non-muta-
tional mechanisms (Hata et al., 2016). While the establishment
of this state seems to be largely stochastic and dictated mostly
by epigenetic mechanisms (Guler et al., 2017; Sharma et al.,
2010), the mechanistic bases of how cancer cells evade the
initial apoptosis in response to drug treatment––the absolute
requirement to enter the drug-tolerant state––or maintain toler-
ance in the presence of drug, are poorly understood.
Our previous work demonstrated that despite sustained
EGFR inhibition following EGFR TKI treatment, reactivation
of ERK1/2 occurs within just a few days (Ercan et al., 2012;
Tricker et al., 2015). Concomitant inhibition of MEK effectively
prevents reactivation of ERK1/2, results in a greater initial
apoptotic response, and leads to a more durable tumor con-
trol in vitro and in vivo than single-agent EGFR inhibition
(Ercan et al., 2012; Tricker et al., 2015). The EGFR (osimertinib)
and MEK (selumetinib) inhibitor combination has been studied
in patients resistant to previous EGFR TKIs and is also under
evaluation as initial therapy for advanced EGFR-mutant
NSCLC in a phase II clinical trial (NCT03392246; Ramalingam
et al., 2019). However, even with this combination, acquired
resistance can still develop (Tricker et al., 2015). In this study
we set out to elucidate the mechanisms that allow cancer
cells to evade apoptosis and survive despite combined
EGFR/MEK inhibition.
RESULTS
Combined EGFR and MEK Inhibition Results in a Stable
but Reversible Dormant State
Combined EGFR/MEK inhibition prevents the reactivation of
ERK1/2 following EGFR inhibition and delays the onset of drug
resistance in vitro and in vivo (Tricker et al., 2015) (Figure S1A).
In PC-9 cells, treatment with single-agent osimertinib (O) leads
to re-colonization of wells within 8 weeks (Figure 1A). The
combination of O and the MEK inhibitor trametinib (T) prevents
any measurable regrowth (Figure 1A). However, few viable cells
can still be detected after 15 weeks of treatment (Figure 1B). We
used live-cell imaging and observed that the OT-treated cells
surviving the initial apoptosis remained in a largely non-prolifer-
ative, or dormant, state throughout the treatment period. How-
ever, within days following drug withdrawal the cells began to
proliferate and re-colonize the wells (Figure 1C, Video S1). This
phenomenon was consistent across EGFR-mutant NSCLC
cell lines (Figures 1C and S1B). These observations suggest
that while combined EGFR/MEK inhibition eliminates cells in
which reactivation of ERK signaling occurs following single-
agent EGFR inhibition, a separate population enters a dormant
state, surviving combined EGFR/MEK inhibition. There was no
evidence of reactivation of EGFR and/or ERK signaling in the
dormant cells during treatment (Figure 1D) and EGFR signaling
was restored in cells that grew following drug washout (Fig-
ure 1D). These cells were still sensitive to OT, and morphologi-
cally indistinguishable from the untreated cells (Figure S1C),
suggesting that we did not select out a subclone with a pre-ex-
isting resistance mutation (Hata et al., 2016). To formally address
whether the establishment of dormancy following OT treat-
ment is pre-determined or a stochastic process, we barcoded
PC-9 cells using the EvoSeq library (Feldman et al., 2019),
treated the cells with DMSO, gefitinib (G), O, or OT for 3 weeks,
sequenced DNA from the remaining cells and analyzed the
findings as described (Bhang et al., 2015) with some modifica-
tions. We observed a large fraction of shared barcodes within
the G-treated (data not shown) and O-treated (Figures 1E, 1F,
and S2A) cells, strongly suggesting selection of pre-existing
clones, consistent (for G) with previous studies (Hata et al.,
2016). In contrast, the vast majority of barcodes in the OT-
treated cells were unique (Figures 1E, 1F, and S2A). Comparison
of the shared barcodes between O and OT cells demonstrated
that 89% of the barcodes identified in the O group are not
present in the OT group (Figure S2B). These findings suggest
that, while resistance to O likely occurs through a selection
process of a pre-existing clone, the ability of cells to enter

Cancer Cell 37, 104–122, January 13, 2020 105

Figure 1. Combined EGFR/MEK Inhibition Promotes a Senescence-like Dormant State
(A) Proliferation of PC-9 cells treated with DMSO, 100 nM osimertinib (O) alone, or in combination with 30 nM trametinib (T).
(B) Images of control cells (at 1 week) or dormant PC-9 cells (at 15 weeks). Scale bars, 200 mm.
(C) Cells were treated as in (A) for 6 weeks followed by drug washout.
(D) Western blot analysis of EGFR downstream signaling following treatment with OT for the indicated times or 21 days followed by drug washout (rebound).
(E) Fraction of barcodes shared among replicates following indicated treatments in barcoded PC-9 cells.
(F) Relative abundance of individual barcodes. Shared and unique indicate barcodes shared by > 2 or %2 replicates, respectively.

(legend continued on next page)

106 Cancer Cell 37, 104–122, January 13, 2020
dormancy following OT is predominately driven by a stochastic
process.
Dormant State following EGFR/MEK Inhibition Shares
Characteristics with Cellular Senescence
To characterize the dormant state, we performed RNA
sequencing (RNA-seq) in PC-9, HCC827, and HCC4006 cells
following treatment with either DMSO or with OT for 2 weeks.
Gene set enrichment analysis (GSEA) revealed an upregulation
of gene expression signatures involved in inflammatory response,
epithelial-to-mesenchymal transition (EMT), and protein secretion,
while cell-cycle-associated gene expression signatures were
robustly downregulated (Figure 1G). These findings along with
the spread-out, flattened morphology of the dormant cells (Fig-
ure 1B) prompted us to query similarities between the dormant
state and cellular senescence. We stained DMSO-, O-, or OT-
treated PC-9, HCC827, and HCC4006 cells for senescence-asso-
ciated b-galactosidase activity (SA-b-Gal) (Debacq-Chainiaux
et al., 2009), and noted that for all three cell lines the majority of cells
surviving the combination treatment stained positive for SA-b-Gal
(Figures 1H and 1I). Further GSEA revealed a significant enrich-
ment of a senescence-associated gene expression signature
(Fridman and Tainsky, 2008) in the dormant cells (Figures 1J and
S3A). A significantly smaller proportion of O-treated cells demon-
strated SA-b-Gal activity compared with those treated with the OT
combination (Figures 1H and 1I). Senescent cells characteristically
exhibit increased secretion of pro-inflammatory factors (senes-
cence-associated secretory phenotype [SASP]) (Coppé et al.,
2010). By analyzing conditioned medium from dormant cells, we
observed an increased secretion of several cytokines and chemo-
kines, including the prominent SASP factor interleukin-6,
compared with untreated cells (Figure S3B). The expression of
several classical SASP factors (Coppé et al., 2008) in the cytokine,
chemokine, IGFBP, and MMP families was also upregulated in the
dormant cells (Figure S3C), analogous to SASP.
Using immunofluorescence following a 10-day treatment with
OT, we detected a robust increase in punctate, H3K9Me3-posi-
tive nuclear foci, another hallmark of senescence (Narita et al.,
2003) (Figure 1K). Senescence is also invariably associated
with the induction of p16INK4a, p21Cip1, and/or p27Kip (Campisi
and D’Adda Di Fagagna, 2007). Although p16INK4a or p21Cip1
were not consistently induced, we observed a robust induction
of p27Kip, which was subsequently downregulated in growing
cells following drug withdrawal (Figure S3D).
The Establishment of Dormancy following EGFR/MEK
Inhibition Is Critically Dependent on Activation of
YAP/TEAD
To explore potential epigenetic changes in the dormant cells, we
employed an assay for transposase-accessible chromatin com-
(G) GSEA of Hallmark gene sets comparing dormant cells versus DMSO-treated control cells. Normalized enrichment scores (NES) for gene sets with false
discovery rate (FDR) < 0.1 in at least two cell lines are shown.
(H) Senescence-associated b-galactosidase (SA-b-Gal) staining of cells treated as indicated for 10 days. Scale bars, 100 mm.
(I) Quantification of (H).
(J) GSEA of senescence signature comparing dormant, OT-treated PC-9 cells versus control cells.
(K) Immunofluorescence (IF) staining for H3K9Me3 in control cells or dormant cells treated with OT for 10 days. Scale bars, 20 mm.
Mean ± SEM are shown in all plots except (I) where mean ± SD are shown. ANOVA (I) or t test (K) were used for statistical analyses. ***p < 0.001, **p < 0.01. See
also Figures S1–S3.
bined with next-generation sequencing (ATAC-seq). We
observed a significant difference in the global epigenetic states
between the OT-induced dormant versus DMSO-treated cells
(Figure 2A), which reverted upon drug washout, demonstrating
that the changes acquired at dormancy are reversible (Figure 2A).
There was also a significant difference in epigenetic states
between single-agent O- and OT-induced dormant cells (Fig-
ure 2A). We performed a motif analysis to interrogate transcrip-
tion factor binding sites associated with the ATAC-seq peaks
with higher signal (more accessible chromatin) in OT-induced
dormant cells versus control cells (Figure 2B). The three most
significantly enriched motifs were the consensus sites for
TEAD family transcription factors (Figure 2C), suggesting that
the OT-induced epigenetic state is associated with increased
TEAD transcription factor binding. The TEAD transcription fac-
tors serve as canonical partners for the Hippo pathway effector
YAP, which has been associated with resistance to targeted
therapy in several contexts, including in resistance to EGFR
TKIs in EGFR-mutant NSCLC (Chaib et al., 2017; Hsu et al.,
2016). Indeed, we observed a significant enrichment of a YAP/
TEAD gene expression signature (Zhang et al., 2009) in dormant
EGFR-mutant cells (Figure 2D). TEAD-binding motifs also scored
as the most significant top hits separating the OT- and O-treat-
ment-induced states (Figure 2E). In accordance with these find-
ings, we observed significantly higher YAP/TEAD activity, as
measured by CTGF and ANKRD1 expression, in OT-induced
dormant PC-9 cells compared with O-treated cells (Figure 2F).
Consistently, we also detected increased chromatin accessi-
bility at putative distal enhancer sites upstream of CTGF TSS
in OT-induced dormant cells compared with cells treated with
O alone (Figure S4A). Taken together, these results demonstrate
that dormant cells induced by combined EGFR/MEK inhibition
adopt a distinct, reversible epigenetic state distinguished from
the untreated or the O-treated state by increased YAP/TEAD
activity.
To assess the role of YAP activity in the establishment of
OT-induced dormant state, we treated EGFR-mutant NSCLC
cell lines for 3 weeks with OT with or without the tankyrase inhib-
itor XAV939, an indirect inhibitor of YAP activity (Wang et al.,
2015), and assessed the regrowth of cells after drug washout.
Remarkably, the OT/XAV939 combination dramatically reduced
the number of dormant cells (Figure S4B), diminishing regrowth
(Figure 2G). Similar findings with three additional structurally
divergent tankyrase inhibitors were observed (Figure S4B). In
addition, we tested the effect of several inhibitors targeting
putative resistance pathways to EGFR TKIs, or the effect of
chemotherapy in PC-9 cells, but observed little to no effect on
the establishment of the dormant cell population (Figure S4C).
Notably, the combination of single-agent O and XAV939 was
significantly inferior to the OT/XAV939 combination in all
Cancer Cell 37, 104–122, January 13, 2020 107
Figure 2. The Establishment of Dormancy following EGFR/MEK Inhibition Is Critically Dependent on Activation of YAP/TEAD
(A) Principal component analysis of ATAC-seq data from cells treated as indicated for 2 weeks.
(B) ATAC-seq signal intensities centered on upregulated (UP) or downregulated (DOWN) peaks in dormant, OT-treated cells versus control cells.
(C and D) Analysis for enriched transcription factor motifs (C) and GSEA of YAP/TEAD signature (D) (Zhang et al., 2009).
(E) Left: ATAC-seq signal intensities centered on UP or DOWN peaks in OT-treated versus O-treated cells. Right: analysis for transcription factor motifs enriched
in upregulated peaks.
(F) qPCR analysis of YAP target gene expression.
(legend continued on next page)
108 Cancer Cell 37, 104–122, January 13, 2020
EGFR-mutant NCSLC cell lines tested (Figure 2G), suggesting
that the differences seen in YAP/TEAD activity between OT-
induced dormant cells and single-agent O-treated cells may
reflect different degrees of dependency on YAP.
To further validate the role of YAP in the establishment of the
dormant state, we knocked out YAP1 in three different EGFR-
mutant NSCLC cell lines, including a patient-derived cell line,
DFCI243, using the CRISPR/Cas9 system (Figure 2H). Strikingly,
YAP1 knockout (KO) completely abolished the establishment
of dormant cells in all cases, measured by lack of regrowth
following drug withdrawal after a 3-week treatment with OT
(Figure 2I). In contrast, two of three YAP1 KO cell lines treated
with O alone regrew following washout (Figure S4D).
In vivo, OT treatment of mice bearing xenograft tumors from
YAP1 KO PC-9, HCC4006, and DFCI243 cells, or from the corre-
sponding control cells, led to a durable response for the entire
4-week treatment period (Figure 2J). However, the control tu-
mors started to regrow soon after treatment cessation, consis-
tent with the presence of a dormant cell population in vivo. In
comparison, YAP1 KO tumors had an increased tumor regrowth
latency and significantly smaller tumors at the time of regrowth in
all models (Figure 2J), consistent with a reduction in the dormant
cell population. Collectively, these results demonstrate that the
establishment of a drug-tolerant state following EGFR/MEK
inhibition, but not single-agent EGFR inhibition, is critically
dependent on YAP/TEAD activity.
YAP Activation Is Necessary for Cancer Cell Viability
upon Combined EGFR/MEK Inhibition
To monitor YAP/TEAD activity following drug treatment, we
introduced a fluorescent YAP/Hippo pathway reporter (Mohseni
et al., 2014) into PC-9 cells (PC-9 YAP reporter cells) and used
live-cell imaging to track YAP activity over time. The OT treat-
ment robustly induced YAP activity (Figure 3A), which was
completely blocked by the addition of XAV939 (Figure 3A).
Consistently, we noted decreased phosphorylation of the main
YAP upstream negative regulator, LATS1, and a decrease in
YAP S127 phosphorylation, which regulates YAP cytosolic
retention (Figure S5). Consequently, YAP nuclear localization
was significantly increased in EGFR-mutant NCSLC cells
following a 10-day OT treatment, but not following single-agent
O treatment (Figure 3B), in agreement with the more prominent
increase in YAP activity observed in the combination-treated
cells (Figures 2F and 3A).
Treatment-induced activation of YAP suggested that active
YAP protects cells from the initial apoptosis. To evaluate this
possibility, we used non-invasive monitoring of caspase-3/7
activity over time in the PC-9 YAP reporter cells. An increase in
YAP activation occurred proportionally to apoptosis (Figures
3C and 3D), and cells with high YAP activity (YAPhigh cells)
were significantly less likely to undergo apoptosis (Figure 3D;
odds ratio 0.26 at 80 h, p < 0.0001 by Fisher’s exact test).
Consistently, XAV939/OT treatment increased apoptosis in
EGFR-mutant NSCLC cells compared with OT or to O/XAV939
alone (Figures 3E and 3F). Also, the YAP1 KO cells underwent
heightened and accelerated apoptosis in response to OT (Fig-
ure 3G). Importantly, this hypersensitive phenotype was rescued
by the re-expression of wild-type YAP1 in the YAP1 KO cells
(Figure 3H).
The OT combination triggered significantly higher YAP re-
porter activity than single-agent O (Figure 3A), suggesting that
the differences in YAP/TEAD activity seen in the drug-tolerant
state (Figures 2E and 2F) reflect the cells’ immediate responses
to the different drug treatments. As the main consequence of
concomitant MEK inhibition is the suppression of ERK1/2 reac-
tivation following EGFR inhibition, our results suggest that
ERK1/2 reactivation and YAP activation are two separate means
by which EGFR-mutant NSCLC cells evade apoptosis following
single-agent O treatment. We quantified the proportion of
YAPhigh cells in single-agent O- and OT-treated populations us-
ing the PC-9 YAP reporter cells. Following a 10-day treatment,
the O-treated cell population contained both YAPhigh (40%)
and YAP-negative (60%) cells, whereas the OT treatment largely
depleted YAP-negative cells, leaving behind mostly YAPhigh
cells (>80%) (Figure 3I). Taken together, these observations
demonstrate that chronic downregulation of EGFR and its
downstream signaling by concomitant EGFR and MEK inhibition
selects for cells that induce high YAP activity upon treatment,
creating a vulnerability that can be exploited to selectively pro-
mote apoptosis in these cells through simultaneous inhibition
of YAP (Figure 3J).
YAP-High, Senescence-like Dormant State Also Occurs
In Vivo
To study the dormant state in vivo, we performed single-cell
RNA-seq (scRNA-seq) in cells from PC-9 xenograft tumors
treated with OT until MRD (3 weeks) (Figures 4A, 4B, and S6A).
We also performed scRNA-seq in dormant PC-9 cells following
3 weeks of OT treatment in vitro. We detected a significant in-
crease in cells enriched for YAP, EMT, or senescence gene
expression signatures in the OT-treated PC-9 cells in vitro and
in vivo (Figure 4C). We also analyzed YAP expression and sub-
cellular localization from the same PC-9 MRD in vivo samples
using immunohistochemistry. Consistent with the scRNA-seq
studies, we detected an increase in active, nuclear YAP in
the MRD tumors, with more intense nuclear staining in the
combination-treated tumors (Figure 4D and 4F). We further eval-
uated MRD tumors from genetically engineered EGFRL858R/T790M
mice (Zhou et al., 2009) following 2 weeks of O treatment and
similarly noted an increase in YAP nuclear localization (Figures
4E and 4F). As these mice have an intact immune system,
we evaluated T cell infiltration. We observed an increase in

(G) Regrowth of EGFR-mutant NSCLC cells after washout following a 3-week treatment with the indicated drug combinations.
(H) Western blot analysis of YAP protein levels in YAP1 knockout (KO) and control (CTRL) cells.
(I) Proliferation of cells in (H) treated as indicated for 21 days, followed by drug washout.
(J) Mice bearing CTRL or YAP1 KO cell xenograft tumors were treated with vehicle or OT followed by treatment cessation and follow-up. Data with at least 6/8 live
mice per group are plotted. Right: tumor volumes at time of regrowth, indicated by an arrow.
Mean ± SEM are shown in all plots except (F), where mean ± SD are shown. ANOVA was used for statistical analyses in all but (J), where t test was used.
***p < 0.001, *p < 0.05. See also Figure S4.

Cancer Cell 37, 104–122, January 13, 2020 109

Figure 3. YAP Activation Is Necessary for Cancer Cell Viability upon Combined EGFR/MEK Inhibition
(A) YAP activity following indicated treatments in PC-9 cells transduced with a fluorescent YAP/Hippo pathway reporter (PC-9 YAP reporter cells).
(B) IF staining for YAP nuclear localization following the indicated treatments. Scale bars, 500 mm.
(C) YAP activity and apoptosis in PC-9 YAP reporter cells treated with OT.
(D) Analysis of overlap between YAPhigh cells (red) and apoptotic cells (green) after 80 h of treatment in PC-9 YAP reporter cells. Scale bars, 150 mm.
(E) Apoptosis in PC-9 cells treated with the indicated drugs or drug combinations.
(F) Apoptosis in EGFR-mutant NSCLC cells treated as indicated. Peak apoptosis values over 72 h are shown.
(G) Apoptosis in YAP1 KO or CTRL cells treated as indicated.
(legend continued on next page)

110 Cancer Cell 37, 104–122, January 13, 2020

infiltration of CD4+ and CD8+ T cells (Figures 4G and S6B),
suggesting that the MRD tumors elicit an immune response,
consistent with our findings of an increase in secreted inflamma-
tory factors (Figures S2B and S2C) and with previous studies
in lung cancer patients (Thress et al., 2017). However, despite
the T cell response, neither O- nor O/selumetinib-treated
EGFRL858R/T790M mice could be cured by the treatment (Fig-
ure S6C) suggesting that the immune response is insufficient
to eradicate the YAPhigh residual cells.
We further studied tumors from EGFRL858R/T790M mice that
had developed acquired resistance to WZ4002 (preclinical
third-generation EGFR inhibitor)/T combination from our
previous study (Tricker et al., 2015) and detected robust YAP
nuclear staining and a lack of pERK1/2 expression in the resis-
tant tumor nodules (Figure 4H). A significantly higher proportion
of nuclear YAP-positive cells in WZ4002/T-resistant nodules
was detected compared with single-agent WZ4002-resistant
nodules (Figure 4H), consistent with our in vitro observations
(Figures 2 and 3). Finally, we analyzed YAP nuclear staining
and pERK1/2 expression in a tumor from a patient with
advanced EGFR-mutant lung cancer treated with O/selumetinib
(NCT03392246) who had a sustained partial response. The pa-
tient underwent surgery following 11 months of treatment while
in a clinical MRD state. The residual tumor demonstrated intense
YAP staining and an absence of pERK1/2 staining (Figure 4I).
YAP Mediates the Evasion of Apoptosis by Repressing
the Induction of the Pro-apoptotic Protein BMF
We next sought to elucidate the mechanism by which YAP
protects EGFR-mutant NSCLC cells from apoptosis. YAP1 KO
had no effect on canonical EGFR signaling or on the induction
of BIM, a known mediator of apoptosis following EGFR inhibi-
tion (Costa et al., 2007; Cragg et al., 2007; Gong et al., 2007)
(Figure 5A). This suggests that YAP affects the apoptotic pro-
cess independently of EGFR signaling and downstream of
BIM. Unlike previous reports (Lin et al., 2015; Rosenbluh et al.,
2012), we did not observe any significant changes in the levels
of anti-apoptotic proteins BCL-XL, BCL2, BCL-w, or MCL-1 in
YAP1 KO cells compared with control cells at baseline or
following OT (Figure S7A). In contrast, we observed a substantial
increase in BAX activity (Figure S7B), cytochrome c release
(Figure S7C), and caspase activation (Figure 3G) in YAP1
KO cells in response to OT, suggesting that the increased
apoptosis seen in YAP1 KO cells is mediated by the intrinsic
apoptotic pathway.
To identify potential YAP target gene(s), we performed RNA-
seq on PC-9 and HCC4006 YAP1 KO or control cells with and
without OT treatment (Figure 5B). Focusing on genes that
mediate apoptosis through the activation of caspases (Hallmark
Apoptosis gene set, 161 genes) (Liberzon et al., 2015), we iden-
tified BMF as one of the top upregulated genes in drug-treated
YAP1 KO cells compared with drug-treated control cells in
(H) Left: western blot analysis of YAP protein levels in YAP1 KO cells transduced with wild-type YAP1. Right: cells were treated with OT and analyzed as in (G).
Only data from drug-treated cells is shown.
(I) Proportions of YAPhigh cells in PC-9 YAP reporter cell populations treated as indicated. Scale bars, 150 mm.
(J) Different means for EGFR-mutant NSCLC cells to avoid apoptosis following EGFR inhibition.
Mean ± SEM are shown in all plots except (I), where SD is shown. ANOVA was used for statistical analyses in all but (D), where Fisher’s exact test was used.
***p < 0.001. See also Figure S5.
both cell lines (Figure 5C). The BMF gene encodes a pro-
apoptotic BH3-only protein that can sequester anti-apoptotic
proteins, but, unlike BIM, cannot directly activate BAX or BAK
(Bhola and Letai, 2016; Kuwana et al., 2005). Together with
BIM induction upon EGFR inhibition (Figure 5A), an increase in
BMF in YAP1 KO cells would be expected to lead to increased
activation of BAX and thus to enhanced apoptosis, consistent
with our observations (Figures 3G, S7B, and S7C). We confirmed
the RNA-seq results using qPCR; YAP inhibition by XAV939 or
YAP1 KO significantly increased BMF expression in response
to OT in EGFR-mutant NSCLC cell lines in vitro and in vivo
(Figure 5D). Due to lack of high-affinity antibodies for BMF, we
used CRISPR/Cas9 technology to produce N-terminally HA-
tagged BMF under the endogenous promoter in PC9 (Figure 5E
and S7D). In these cells, YAP inhibition or YAP knockdown led
to a robust increase in BMF protein levels in response to OT,
while BIM levels remained unchanged (Figure 5F). Moreover,
re-introduction of wild-type YAP, but not a TEAD-binding-defi-
cient S94A mutant, to the YAP1 KO background completely
abolished the increase in BMF expression (Figure 5G). Ectopic
overexpression of BMF using a doxycycline-inducible vector in
EGFR-mutant NSCLC cells (Figure S7E) induced rapid
apoptosis, which was further increased by co-treatment with
OT (Figure S7F), demonstrating that induction of BMF alone,
without YAP activation, is sufficient to sensitize EGFR-mutant
NSCLC cells to apoptosis. Downregulation of BMF expression
using small interfering RNA (siRNA) significantly decreased
apoptosis in YAP1 KO cells in response to OT (Figures 5H and
S7G), demonstrating that the induction of BMF expression is
necessary for the increased apoptosis in YAP1 KO cells. Hence,
YAP facilitates evasion of apoptosis in EGFR-mutant NSCLC
cells by repressing the expression of BMF upon combined
EGFR/MEK inhibition, leading to the establishment of the
dormant cell population (Figure 5I).
YAP Represses BMF Induction by Engaging EMT
Transcription Factor SLUG
Next, we investigated the molecular mechanisms by which
YAP represses the expression of BMF. Although YAP is mostly
linked to transcriptional activation, it has also been shown to
complex with transcription factors and modulators to drive
transcriptional repression, often in association with TEAD (Beyer
et al., 2013; Kim et al., 2015; Zaidi et al., 2004). Thus, we hypoth-
esized that the YAP/TEAD complex is directly repressing BMF
by forming a tertiary complex with a transcriptional repressor.
In search for such transcriptional repressors, we noted that
an EMT gene expression signature was highly enriched in the
dormant cells induced by OT treatment (Figure 1G and 4C). In
addition to EMT being a known mechanism of drug resistance
in EGFR-mutant lung cancer (Byers et al., 2013; Sequist et al.,
2011; Shibue and Weinberg, 2017; Zhang et al., 2012), YAP
has been reported to mediate EMT and to directly bind to
Cancer Cell 37, 104–122, January 13, 2020 111
Figure 4. YAP-High, Senescence-like Dormant State Also Occurs In Vivo
(A) Growth curves of PC-9 xenograft tumors harvested for single-cell RNA-seq (scRNA-seq) and immunohistochemistry (IHC).
(B) Fluorescence-activated cell sorting scheme used to obtain scRNA-seq samples from the dissociated xenograft tumors.
(C) YAP, EMT, and senescence signature enrichments in single cells from the xenograft tumors.
(D and E) IHC staining for YAP in the xenograft tumors (D) or in residual tumors from EGFRL858/T790M mice (E) following 2-week treatment with vehicle or O. Scale
bars, 100 mm.
(F) Quantification of (D) and (F).
(G) Quantification of infiltrating T cells in the same tumors as in (E) based in CD4/CD8 IHC.
(H and I) IHC staining for YAP and pERK in WZ4002- or WZ4002/T-resistant tumors from EGFRL858/T790M mice (H) or in a residual tumor of an EGFR-mutant
NSCLC patient following treatment with O/selumetinib for 11 months (I). Scale bars, 800 mm (H), 100 mm (I).
Kolmogorov-Smirnov test (C), ANOVA (F) when more than two groups, (H) or t test (F) when two groups, (G) and (I) were used for statistical analyses. ***p < 0.001,
**p < 0.01; n.s., not significant. See also Figure S6.

112 Cancer Cell 37, 104–122, January 13, 2020

Figure 5. YAP Mediates the Evasion of Apoptosis by Repressing the Induction of Pro-apoptotic BMF
(A) Western blot analysis of EGFR downstream signaling following 24 h treatment as indicated.
(B) RNA-seq samples used in (C).
(C) Expression of genes regulating apoptosis in OT-treated YAP1 KO cells versus OT-treated CTRL cells. Colors indicate log2 fold change values with p < 0.001.
(D) qPCR analysis of BMF expression in CTRL or YAP1 KO cells treated as indicated for 24 h in vitro or 3 days in vivo.
(E) Schematic representation of the endogenous BMF locus in PC-9 HA-BMF cells.

(legend continued on next page)

Cancer Cell 37, 104–122, January 13, 2020 113

canonical EMT transcription factors, including SNAIL, SLUG,
and ZEB1 (Lehmann et al., 2016; Tang et al., 2016). We therefore
explored the possibility that EMT, the development of a dormant
state, and the requirement for YAP in mediating evasion of
apoptosis through repression of BMF following drug treatment,
were all linked. In PC-9 and HCC4006 YAP1 KO cells following
24 h OT treatment, compared with control cells, the EMT signa-
ture was negatively enriched, suggesting that YAP is triggering
the EMT program in EGFR-mutant NSCLC cells (Figure 6A).
qPCR analysis revealed that SNAI2, encoding SLUG, was the
dominantly expressed EMT transcription factor in EGFR-mutant
NCSLC cell lines (Figure 6B). We further observed that endoge-
nous YAP, TEAD, and SLUG proteins co-immunoprecipitate in
both PC-9 and HCC4006 cells upon 48 h OT treatment (Fig-
ure 6C). Knockdown of SLUG by siRNA in PC-9 and HCC4006
cells resulted in a significant increase in BMF expression
following treatment with OT, similar to that observed following
YAP knockdown (Figure 6D and 6E), and the increase in BMF
expression translated into a robust increase in apoptosis upon
treatment (Figure 6F). These results suggest that members of
the YAP/TEAD/SLUG complex co-operate to repress BMF
expression and thus suppress apoptosis in response to OT treat-
ment. To confirm that the YAP/TEAD/SLUG complex directly
binds the BMF locus to mediate repression, we performed chro-
matin immunoprecipitation followed by next-generation
sequencing (ChIP-seq) using antibodies against endogenous
YAP, TEAD4, and SLUG in PC-9 cells treated either with
DMSO or OT. We detected a robust increase in YAP and
SLUG binding to chromatin after 48 h of OT treatment, while
TEAD4 chromatin binding was less affected (Figure 6G). Specif-
ically, we observed overlapping YAP, TEAD, and SLUG peaks at
the BMF promoter region as well as at nearby H3K27Ac-positive
enhancer regions upon treatment (Figure 6H), demonstrating
that the YAP/TEAD/SLUG repressor complex directly binds to
the BMF locus. Taken together, these results provide a mecha-
nistic explanation for YAP-mediated suppression of pro-
apoptotic signaling through the engagement of TEAD and the
EMT program to directly repress the induction of BMF expres-
sion upon combined EGFR/MEK inhibition (Figure 6I).
Development of a Novel Covalent TEAD Inhibitor to
Target YAP Dependency upon Combined EGFR/MEK
Inhibition
The strict dependency of OT-treated cells on YAP presents an
attractive target for drug development. Although our results point
toward TEAD as the main mediator of YAP effects in this context
(Figures 2C, 2D, and 6H), we wanted to further confirm whether
other effector pathways downstream of YAP might also play a
role. The YAP protein has several functional domains, many of
which mediate protein-protein interactions (Piccolo et al.,
2014). We systematically mutated the key functional domains
(F) Western blot analysis of BMF, BIM, and YAP expression in PC-9 HA-BMF cells transfected with non-targeting (NT) or YAP siRNA and treated as indicated
for 24 h.
(G) qPCR analysis of BMF expression in CTRL or YAP1 KO cells transduced as indicated, and following treatment with either DMSO or OT for 24 h.
(H) Peak apoptosis over 72-h treatment in PC-9 and HCC4006 cells transfected with NT or BMF siRNA.
(I) The mechanism of YAP/TEAD-mediated suppression of apoptosis in EGFR-mutant NSCLC cells following EGFR/MEK inhibition.
Mean ± SD are shown in all plots except (H), where mean ± SEM is shown. ANOVA was used for statistical analyses. ***p < 0.001, **p < 0.01; n.s., not significant
(p > 0.05). See also Figure S7.
114 Cancer Cell 37, 104–122, January 13, 2020
in YAP (Figure 7A), and determined which of the mutants could
rescue the apoptotic phenotype imparted by YAP1 deficiency
following OT treatment in PC-9 cells. We observed that intro-
duction of YAP1 with a mutation in the TEAD-binding domain
(S94A) (Zhao et al., 2008), or with a deleted transactivation
domain (TAdel; Figure 7A), had the least ability to rescue YAP1
deficiency (Figure 7B), confirming that YAP-mediated evasion
of apoptosis is absolutely dependent on TEAD, as well as on
intact transactivation domain.
TEAD, as a transcription factor, is presumed undruggable.
However, recent studies revealed a hydrophobic pocket for
the post-translational palmitoylation of TEAD (Chan et al.,
2016; Noland et al., 2016), and flufenamic acid as a molecule
binding to this pocket (Pobbati et al., 2015). Flufenamic acid/
TEAD2 co-crystal revealed extensive hydrophobic interactions
as its main binding mode (Pobbati et al., 2015), providing
a structural basis for the rational design of a covalent TEAD in-
hibitor through an acrylamide as a covalent warhead to react
with the conserved Cys380 on TEAD2. We reasoned that the
trifluoromethylated phenyl ring forms extensive hydrophobic
interactions, whereas the carboxylic acid of flufenamic acid,
proximal to Cys380, might be replaced by an acrylamide
warhead to react with the cysteine. Hence, MYF-01-37 (Fig-
ure 7C) was developed as a covalent binder to TEAD, targeting
Cys380 when incubated with the TEAD2 protein (C359 in
TEAD1) (Figures S8A–S8C). Pretreating cells with MYF-01-37
led to loss of direct TEAD pull-down by biotin-MYF-01-037 (Fig-
ure S8D) from whole-cell lysate (Figure S8E), confirming MYF-
01-037 binding to TEAD in cells. This target engagement of
TEAD resulted in inhibition of direct YAP/TEAD interaction
(Figure 7D) in HEK293T cells and in the reduction in canonical
YAP target gene CTGF expression in PC-9 cells (Figure 7E).
This reduction was abrogated by the overexpression of a
TEAD1 C359S mutant that disrupts the covalent binding of
the drug to TEAD, but not by overexpression of wild-type
TEAD1 (Figure 7E), demonstrating that the observed inhibition
of YAP activity is due to on-target covalent binding of the
compound to TEAD. Importantly, XAV939, which inhibits YAP
activity via a TEAD-independent mechanism (Wang et al.,
2015), was still able to inhibit CTGF expression also in cells
expressing the TEAD1 C359S mutant (Figure 7E). As a single
agent, MYF-01-37 had minimal impact on cell viability of several
EGFR-mutant NCSLC cell lines (Figure S8F), which is consis-
tent with the apparent dispensability of YAP activity in EGFR-
mutant NSCLC cells at steady state (Figure 2I). When combined
with OT, MYF-01-37 completely suppressed the increased
YAP activity induced by OT treatment in PC-9 YAP reporter
cells (Figure 7F), led to a robust increase in BMF expression
(Figure 7G), and to subsequent increase in apoptosis in PC-9
and HCC4006 cells compared with OT alone (Figure 7H), thus
phenocopying the effects of tankyrase inhibition or YAP1 KO
Figure 6. YAP Represses BMF Induction by Engaging EMT Transcription Factor SLUG
(A) GSEA of EMT signature in YAP1 KO versus control cells treated with OT for 24 h.
(B) qPCR analysis of EMT transcription factor expression in untreated EGFR-mutant NSCLC cells.
(C) Co-immunoprecipitation analysis of the interaction between YAP, TEAD, and SLUG in PC-9 cells following treatment with DMSO or OT for 48 h.
(D) Western blot analysis of YAP and SLUG protein levels in PC-9 or HCC4006 cells transfected with non-targeting (NT), YAP, or SLUG siRNA.
(E) qPCR analysis of BMF expression in cells in (D) following 24 h treatment with DMSO or OT.

(legend continued on next page)

Cancer Cell 37, 104–122, January 13, 2020 115
(Figures 3A, 3E, 3G, and 5D). Importantly, a 10-day treatment
with MYF-01-37 in combination with OT led to a dramatic
decrease in dormant cells compared with OT alone (Figure 7I).
DISCUSSION
Genotype-directed therapy is the standard of care for many
cancers that harbor an activated oncogene (Blanke et al.,
2008; Drilon et al., 2018; Peters et al., 2017). While this treatment
approach has transformed cancer care for many genomic sub-
types of cancer, these therapies are rarely, if ever, curative.
One explanation for such observations is the inability of geno-
type-directed therapies to eradicate all tumors cells. In EGFR-
mutant NSCLC, a complete response is observed only in a
small minority (<5%) of patients following treatment with EGFR
TKI (Mok et al., 2017; Soria et al., 2018). As EGFR mutations
are truncal mutations (i.e., in every cell of a tumor) (Jamal-Hanjani
et al., 2017), it is not clear why a proportion of tumor cells can
survive initial EGFR inhibitor-induced apoptosis and subse-
quently persist in the presence of drug treatment.
The development of EMT, as a drug-resistant state following
EGFR inhibitor treatment, has been observed both in model
systems and in lung cancer patients (Byers et al., 2013; Sequist
et al., 2011; Zhang et al., 2012). In addition, hyperactivation of
the Hippo pathway effector YAP has been shown to dampen
the efficacy of targeted treatment in several contexts (Zanconato
et al., 2016), including the efficacy of EGFR TKIs in EGFR-mutant
NCSLC (Hsu et al., 2016; Chaib et al., 2017). However, the mech-
anistic basis for these observations remains largely unexplored.
Here we provide a mechanistic link for these two different obser-
vations and demonstrate that a critical transcription factor medi-
ating EMT, SLUG, and YAP together leads to transcriptional
repression of BMF following EGFR/MEK treatment and as such
limits the initial drug-induced apoptotic effect allowing the for-
mation of a dormant state.
Apoptosis in response to EGFR TKIs in EGFR-mutant NCSLC
is executed by the intrinsic apoptotic pathway and invariably
associated with the upregulation of BIM (Costa et al., 2007;
Cragg et al., 2007; Gong et al., 2007). BIM levels are suppressed
by the MAPK pathway, both transcriptionally and post-transcrip-
tionally (Ley et al., 2005), and thus mechanisms which uncouple
EGFR inhibition from ERK1/2 inhibition would be expected to
block EGFR inhibitor-mediated upregulation of BIM, promoting
cell survival (Ercan et al., 2012; Tricker et al., 2015). O can also
activate YAP, allowing drug-induced cell survival through a
completely different mechanism (Figures 3A, 3I, 3J, and 4F).
Thus, single-agent EGFR TKI treatment can lead to both ERK1/
2 reactivation and YAP activation, whereas, upon combined
EGFR/MEK inhibition, activation of YAP becomes the dominant
survival mechanism in EGFR-mutant NSCLC cells (Figures 3A,
3I, 3J, 4F, 4H, and 4I). These results suggest that YAP can main-
tain the viability of EGFR-mutant NSCLC cells in the chronic
absence of EGFR and its downstream signaling. Intriguingly,
(F) Apoptosis in cells in (D) following treatment with DMSO or OT.
(G) Number of peaks called by MACS2 (FDR < 0.01).
(H) ChIP-seq signal traces in BMF locus. H3K27Ac was used to identify enhancer regions.
(I) The mechanism by which YAP/TEAD/SLUG complex represses BMF expression upon combined EGFR/MEK inhibition.
Mean ± SD (E) or mean ± SEM (F) are shown. ANOVA was used for statistical analyses. ***p < 0.001, **p < 0.01.
116 Cancer Cell 37, 104–122, January 13, 2020
the ability of YAP to compensate for the loss of dominant onco-
gene in MAPK-dependent cancers has been previously shown in
the context of mutant KRAS-driven NSCLC and pancreatic
ductal adenocarcinoma (Kapoor et al., 2014; Shao et al., 2014).
In these studies, YAP1 overexpression (Shao et al., 2014) or
YAP1 amplification (Kapoor et al., 2014) rescued the loss of
KRAS in a MAPK pathway-independent mechanism.
Overexpression of YAP and its paralog TAZ has been shown to
induce EMT in a TEAD-dependent manner (Lei et al., 2008;
Zhang et al., 2009; Zhao et al., 2008). Considering that YAP
activation seems to be a specific adaptation mechanism in cells
that cannot re-activate EGFR downstream signaling, the YAP/
TEAD/SLUG interplay repressing BMF may be the immediate
response protecting cells undergoing a YAP-dictated global
change in cellular state. Analogously, Shao et al. (2014) also
found that the YAP-mediated bypass of KRAS loss was associ-
ated with the acquisition of a mesenchymal state, suggesting
that YAP may drive the EMT program as a mechanism to adapt
to loss of oncogene signaling in other cancer contexts as well.
Whether our observations on YAP-mediated suppression
of apoptosis through transcriptional repression of BMF also
extend to other genotype-directed cancer therapies remains un-
known and will need to be evaluated in future studies. Also, we
cannot rule out a possibility that, in addition to YAP/TEAD/
SLUG, other factors are involved in the long-term survival of
EGFR-mutant cancer cells treated with OT.
Interestingly, the dormant state resulting from YAP/TEAD
activation shared several characteristics with cellular senes-
cence. Unlike treatment-induced senescence in response to
DNA-damaging chemotherapeutic agents (Ewald et al., 2010),
EGFR-mutant NSCLC cells seem to only reversibly adopt the
senescence program upon EGFR/MEK inhibition to tolerate the
lethal drug exposure, and revert back to the normal steady
state upon drug withdrawal. Consequently, the senescent-like
population, at least in this context, can serve as a reservoir of
dormant cells that are later, upon acquisition of additional resis-
tance mechanisms, capable of re-establishing the tumor and
drive clinically observed drug resistance.
The therapeutic vulnerability of EGFR-mutant NSCLC to YAP/
TEAD antagonism identified in this study led us to develop a
novel covalent TEAD inhibitor, MYF-01-37. As YAP is widely
associated with resistance to cancer therapies, we also tested
the effect of TEAD inhibition and/or YAP1 KO in other genotype
or TKI combination contexts within the NSCLC space, including
in ALK-rearranged, MET-amplified, and EGFR-mutant MET-a-
mplified models, and observed increased apoptosis following
YAP/TEAD co-targeting in most models (Figures 8A–8C). These
data suggest wide potential for co-targeting YAP/TEAD with ge-
notype-directed therapy. In accordance with our observations
that YAP1 loss has negligible consequences in EGFR-mutant
NSCLC cells at steady state, MYF-01-37 did not demonstrate
single-agent toxicity (Figure S8F). This is in stark contrast to a
recently published, structurally similar covalent TEAD inhibitor,
Figure 7. Development of Novel Covalent TEAD Inhibitor to Target YAP Dependency upon Combined EGFR/MEK Inhibition
(A) YAP1 mutants used in the rescue experiment in (B). BD, binding domain.
(B) Viability (CellTiter-Glo) of CTRL cells or PC-9 YAP1 KO cells transduced with YAP1 mutants (A) following 72 h treatment with OT.
(C) Top: the structure of MYF-01-37. Bottom: MYF-01-37 binding to the palmitoylation pocket in TEAD1 based on molecular docking. The cysteine 359 targeted
by MYF-01-37 is indicated.
(D) Effect of MYF-01-37 or the corresponding reversible control on YAP/TEAD interaction measured using Split Gaussia Luciferase Assay.
(E) Left: western bot analysis of the expression of myc-tagged TEAD1 in PC-9 cells transduced as indicated. Right: qPCR analysis of CTGF expression after 24 h
treatment with XAV939 or MYF-01-37 in the transduced PC-9 cells.
(F) YAP activity in PC-9 YAP reporter cells after 72 h treatment with OT or OT in combination with XAV939 (XAV) or MYF-01-37 (MYF).
(legend continued on next page)
Cancer Cell 37, 104–122, January 13, 2020 117
TED-347 (Bum-Erdene et al., 2019), (Figure S8F), which is toxic
most likely due to its covalent warhead. Unlike the highly reactive
a-chloroketone covalent warhead in TED-347, the acrylamide
warhead in MYF-01-37 is more suitable for covalent targeting
of proteins in living cells, and thus most likely contributes to
the low non-specific toxicity of MYF-01-37 as a single agent
(De Cesco et al., 2017; Liu et al., 2013). Further development is
needed to optimize the pharmacological properties of MYF-01-
37 to enable preclinical testing of the compound using in vivo
models of EGFR-mutant NSCLC.
Ultimately, the strategy of co-targeting EGFR, MEK, and YAP/
TEAD to enhance the initial treatment efficacy in EGFR-mutant
NSCLC and limit the establishment of the dormant state, will
need to be tested in a clinical trial. Although EGFR and MEK in-
hibitors can be administered together (NCT03392246; Ramalin-
gam et al., 2019), three-drug combinations raise the concern
for toxicity. Auspiciously, YAP appears dispensable for normal
homeostasis in many adult organs, suggesting that targeting
YAP might be well tolerated (Zanconato et al., 2016). Also, as
our findings reveal, the main role of YAP1 loss is in enhancing
the initial apoptotic effect of EGFR/MEK inhibition. Thus, it is
plausible that a three-drug combination would be necessary
only transiently, followed by a two-drug treatment, thus reducing
potential toxicity. In support of this approach, we observed iden-
tical potency when we treated PC-9 cells for 1 week with OT/
XAV939 or MYF-01-37 followed by 2 weeks of OT compared
with a 3-week continuous OT/XAV939 or MYF-01-37 treatment
(Figure 8D). The potential different treatment approaches will
need to undergo clinical evaluation to determine both their effi-
cacy and toxicity.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d LEAD CONTACT AND MATERIALS AVAILABILITY
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Animal Models
B Cell Line Authentication
B Patient Specimen
d METHOD DETAILS
B Expression Vectors
B Cell Growth and Viability Assays
B Western Blotting and Antibodies
B Cellular Barcoding
B RNA Extraction and Quantitative PCR (QPCR)
B RNA-sequencing
B Senescence-Associated b-galactosidase Staining
B Cytokine Profiling
B Immunofluorescence Staining and Imaging
B ATAC-sequencing
(G) qPCR analysis of BMF expression in cells in (E) following 24 h treatment as indicated.
(H) Apoptosis in PC-9 and HCC4006 cells treated as indicated.
(I) Regrowth of PC-9 and HCC4006 cells after drug washout following a 2-week treatment as indicated.
Mean ± SEM are shown in all plots except (E), where mean ± SD is shown. ANOVA was used for statistical analyses. ***p < 0.001, **p < 0.01. See also Figure S8.
118 Cancer Cell 37, 104–122, January 13, 2020
B ChIP-sequencing
B ATAC-seq and ChIP-seq Analyses
B CRISPR/CAS9 Gene Editing
B Monitoring Caspase-3/7 Activity
B Determining YAP Activity and Apoptosis in PC-9 YAP/
Hippo Reporter Cells
B Viral Transductions
B Single-cell RNA Sequencing
B Immunohistochemistry
B Detection of Activated BAX
B Detection of Mitochondrial Cytochrome c Release
B Gene Knock-down by siRNA
B Co-immunoprecipitation
B Chemistry
B MYF-01-37 Docking to TEAD2
B MYF-01-37 Competition Pulldown
B Mass Spectrometry Analysis
B YAP-TEAD Split Gaussia Luciferase (SGL) Assay
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Statistical Analyses
d DATA AND CODE AVAILABILITY
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2019.12.006.
ACKNOWLEDGMENTS
The authors would like to thank the following for valuable help with this study:
David Feldman and Paul Blainey from the Massachusetts Institute of Technol-
ogy in Cambridge, MA, for providing research materials and expertise for the
barcoding experiment; Zachary Herbert and Maura Berkeley from the Molec-
ular Biology Core Facility at the Dana-Farber Cancer Institute for the
sequencing services; Dana-Farber/Harvard Cancer Center in Boston, MA,
for the use of the Specialized Histopathology Core, which provided histology
and immunohistochemistry service. Dana-Farber/Harvard Cancer Center is
supported in part by an NCI Cancer Center Support Grant no. NIH 5 P30
CA06516; the members of the NYU Experimental Pathology Research Labora-
tory, which is partially supported by the Cancer Center Support Grant
P30CA016087 at NYU Langone’s Laura and Isaac Perlmutter Cancer Center,
for their expertise and immunohistochemistry support.
This work was supported by the National Cancer Institute grants
R01CA135257 (to P.A.J.) and R35CA220497 (to P.A.J.), The American Cancer
Society (CRP-17-111-01-CDD) (to P.A.J.), the Claudia Adams Barr Program
for Innovative Cancer Research (to P.A.J.), the Ildiko Medve and Adria Sai-Ha-
lasz EGFR Lung Cancer Research Fund (to P.A.J.), Balassiano Family Fund for
Lung Cancer Research (to P.A.J.), the Jane and Aatos Erkko Foundation (to
K.J.K.), the Instrumentarium Science Foundation (to K.J.K.), and the Orion
Research Foundation (to K.J.K.). S.B. was supported by fellowships from
the Swiss Cancer League (KLS-3625-02 2015) and Swiss National Science
Foundation (P300PB_161026/1 and P400PM_183862), Swizerland. J.A.M.
was supported by NIH R01 CA222218 and CA233800.
AUTHOR CONTRIBUTIONS
Conceptualization, K.J.K., P.A.J., T.Z., N.S.G., and R.H.; Investigation, K.J.K.,
Y.L., C.T., M.F., E.H.K., K.L., A.P., P.H.L., S.B.F., S.L., T.C., H.M.H., M.B.,
Y.G., S.S., B.H.S., T.T., M.K.W., M.L.T., and M.M.; Resources, M.X., J.C.,
Figure 8. Co-targeting YAP/TEAD with Genotype-Directed Therapy
(A and B) Apoptosis in NSCLC cell lines treated as indicated.
(C) Left: western blot analysis of YAP expression in control (CTRL) and YAP1 KO H3122 and EBC-1 cells. Right: apoptosis in CTRL and YAP1 KO H3122 and EBC-
1 cells treated as indicated.
(D) PC-9 cells were treated as indicated in the scheme on the left, followed by drug washout. Regrowth of cells was monitored and quantified as in Figure 2G.
Mean ± SEM are shown. ANOVA was used for statistical analyses. ***p < 0.001.

Cancer Cell 37, 104–122, January 13, 2020 119

P.T.K., D.A.B., F.D.C., P.B., R.B., A.A.B., K.-K.W., and M.M.A.; Methodology,
S.B., B.L.E., P.C., H.W.L., and C.P.P.; Formal Analyses, Y.X., H.W., and A.V.;
Writing, K.J.K. and P.A.J.; Funding Acquisition, P.A.J. and N.S.G.; Supervi-
sion, P.A.J., N.S.G., T.Z., P.C.G., and J.A.M.
DECLARATION OF INTERESTS
P.A.J. has received consulting fees from AstraZeneca, Boehringer Ingelheim,
Pfizer, Roche/Genentech, Merrimack Pharmaceuticals, Chugai Pharmaceuti-
cals, Ariad Pharmaceuticals, Eli Lilly and Company, Araxes Pharama, Ignyta,
Novartis, Mirati Therapeutics, Takeda Oncology, Daiichi Sankyo, Biocartis,
Voronoi, SFJ Pharmaceuticals, and LOXO Oncology; receives post-marketing
royalties from DFCI-owned intellectual property on EGFR mutations licensed
to Lab Corp; has sponsored research agreements with AstraZeneca, Daichi
Sankyo, Boehringer Ingelheim, PUMA, Eli Lilly, Astellas Pharmaceuticals,
and Takeda Oncology; and has stock ownership in Gatekeeper Pharmaceuti-
cals and LOXO Oncology. N.S.G. is a founder, science advisory board member
(SAB), and equity holder in Gatekeeper, Syros, Petra, C4, B2S, and Soltego.
The Gray lab receives or has received research funding from Novartis, Takeda,
Astellas, Taiho, Janssen, Kinogen, Voronoi, Her2llc, Deerfield, and Sanofi.
R.H. has received research grants from Bristol-Myers Squibb and Novartis.
K.K.W. is a founder and equity holder of G1 Therapeutics and has Consul-
ting/Sponsored Research Agreements with AstraZeneca, Janssen, Pfizer,
Array, Novartis, Merck, Takeda, Ono, Targimmune, and BMS. C.P.P. has
received honoraria from Bio-Rad and AstraZeneca, is a co-founder of Xsphera
Biosciences, and is on the scientific advisory board of DropWorks and
XSphera Biosciences. P.B. receives grant funding from Novartis Institute of
Biomedical Research for an unrelated project. J.A.M. serves on the SAB of
908 Devices.
Received: May 17, 2019
Revised: October 11, 2019
Accepted: December 10, 2019
Published: January 13, 2020
REFERENCES
Aibar, S., González-Blas, C.B., Moerman, T., Huynh-Thu, V.A., Imrichova, H.,
Hulselmans, G., Rambow, F., Marine, J.C., Geurts, P., Aerts, J., et al. (2017).
SCENIC: single-cell regulatory network inference and clustering. Nat.
Methods 14, 1083–1086.
Alexander, W.M., Ficarro, S.B., Adelmant, G., and Marto, J.A. (2017).
Multiplierz v2.0: a Python-based ecosystem for shared access and analysis
of native mass spectrometry data. Proteomics 17, 15–16.
Bahcall, M., Sim, T., Paweletz, C.P., Patel, J.D., Alden, R.S., Kuang, Y.,
Sacher, A.G., Kim, N.D., Lydon, C.A., Awad, M.M., et al. (2016). Acquired
METD1228V mutation and resistance to MET inhibition in lung cancer.
Cancer Discov. 6, 1334–1341.
Bankhead, P., Loughrey, M.B., Fernández, J.A., Dombrowski, Y., McArt, D.G.,
Dunne, P.D., McQuaid, S., Gray, R.T., Murray, L.J., Coleman, H.G., et al.
(2017). QuPath: open source software for digital pathology image analysis.
Sci. Rep. 7, 1–7.
Beyer, T.A., Weiss, A., Khomchuk, Y., Huang, K., Ogunjimi, A.A., Varelas, X.,
and Wrana, J.L. (2013). Switch enhancers interpret TGF-b and hippo signaling
to control cell fate in human embryonic stem cells. Cell Rep. 5, 1611–1624.
Bhang, H.C., Ruddy, D.A., Krishnamurthy Radhakrishna, V., Caushi, J.X.,
Zhao, R., Hims, M.M., Singh, A.P., Kao, I., Rakiec, D., Shaw, P., et al.
(2015). Studying clonal dynamics in response to cancer therapy using high-
complexity barcoding. Nat. Med. 21, 440–448.
Bhola, P.D., and Letai, A. (2016). Mitochondria-judges and executioners of cell
death sentences. Mol. Cell 61, 695–704.
Blanke, C.D., Rankin, C., Demetri, G.D., Ryan, C.W., Von Mehren, M.,
Benjamin, R.S., Raymond, A.K., Bramwell, V.H.C., Baker, L.H., Maki, R.G.,
et al. (2008). Phase III randomized, intergroup trial assessing imatinib mesylate
at two dose levels in patients with unresectable or metastatic gastrointestinal
stromal tumors expressing the kit receptor tyrosine kinase: S0033. J. Clin.
Oncol. 26, 626–632.
120 Cancer Cell 37, 104–122, January 13, 2020
Buenrostro, J., Wu, B., Chang, H., and Greenleaf, W. (2015). ATAC-seq: a
method for assaying chromatin accessibility genome-wide. Curr. Protoc.
Mol. Biol. 109, 21.29.1–21.29.9.
Bum-Erdene, K., Zhou, D., Gonzalez-Gutierrez, G., Ghozayel, M.K., Si, Y., Xu,
D., Shannon, H.E., Bailey, B.J., Corson, T.W., Pollok, K.E., et al. (2019). Small-
molecule covalent modification of conserved cysteine leads to allosteric inhi-
bition of the TEAD,YAP protein-protein interaction. Cell Chem. Biol. 26, 378–
389.e13.
Butler, A., Hoffman, P., Smibert, P., Papalexi, E., and Satija, R. (2018).
Integrating single-cell transcriptomic data across different conditions, technol-
ogies, and species. Nat. Biotechnol. 36, 411–420.
Byers, L.A., Diao, L., Wang, J., Saintigny, P., Girard, L., Peyton, M., Shen, L.,
Fan, Y., Giri, U., Tumula, P.K., et al. (2013). An epithelial-mesenchymal transi-
tion gene signature predicts resistance to EGFR and PI3K inhibitors and iden-
tifies Axl as a therapeutic target for overcoming EGFR inhibitor resistance. Clin.
Cancer Res. 19, 279–290.
Campisi, J., and D’Adda Di Fagagna, F. (2007). Cellular senescence: when bad
things happen to good cells. Nat. Rev. Mol. Cell Biol. 8, 729–740.
De Cesco, S., Kurian, J., Dufresne, C., Mittermaier, A.K., and Moitessier, N.
(2017). Covalent inhibitors design and discovery. Eur. J. Med. Chem.
138, 96–114.
Chaib, I., Karachaliou, N., Pilotto, S., Codony Servat, J., Cai, X., Li, X.,
Drozdowskyj, A., Servat, C.C., Yang, J., Hu, C., et al. (2017). Co-activation
of STAT3 and YES-associated protein 1 (YAP1) pathway in EGFR-mutant
NSCLC. J. Natl. Cancer Inst. 109, 1–12.
Chan, P., Han, X., Zheng, B., Deran, M., Yu, J., Jarugumilli, G.K., Deng, H.,
Pan, D., Luo, X., and Wu, X. (2016). Autopalmitoylation of TEAD proteins reg-
ulates transcriptional output of the Hippo pathway. Nat. Chem. Biol. 12,
282–289.
Coppé, J.-P., Patil, C.K., Rodier, F., Sun, Y., Muñoz, D.P., Goldstein, J.,
Nelson, P.S., Desprez, P.-Y., and Campisi, J. (2008). Senescence-associated
secretory phenotypes reveal cell-nonautonomous functions of oncogenic RAS
and the p53 tumor suppressor. PLoS Biol. 6, e301.
Coppé, J.-P., Desprez, P.-Y., Krtolica, A., and Campisi, J. (2010). The senes-
cence-associated secretory phenotype: the dark side of tumor suppression.
Annu. Rev. Pathol. Mech. Dis. 5, 99–118.
Corces, M.R., Trevino, A.E., Hamilton, E.G., Greenside, P.G., Sinnott-
Armstrong, N.A., Vesuna, S., Satpathy, A.T., Rubin, A.J., Montine, K.S., Wu,
B., et al. (2017). An improved ATAC-seq protocol reduces background and en-
ables interrogation of frozen tissues. Nat. Methods 14, 959–962.
Cordenonsi, M., Zanconato, F., Azzolin, L., Forcato, M., Rosato, A., Frasson,
C., Inui, M., Montagner, M., Parenti, A.R., Poletti, A., et al. (2011). The hippo
transducer TAZ confers cancer stem cell-related traits on breast cancer cells.
Cell 147, 759–772.
Cornwell, M.I., Vangala, M., Taing, L., Herbert, Z., Köster, J., Li, B., Sun, H., Li,
T., Zhang, J., Qiu, X., et al. (2018). VIPER: visualization pipeline for RNA-seq, a
Snakemake workflow for efficient and complete RNA-seq analysis. BMC
Bioinformatics 19, 1–14.
€nne, P.A. (2014). Molecular mechanisms of resistance in
Cortot, A.B., and Ja
epidermal growth factor receptor-mutant lung adenocarcinomas. Eur.
Respir. Rev. 23, 356–366.
Costa, D.B., Halmos, B., Kumar, A., Schumer, S.T., Huberman, M.S., Boggon,
T.J., Tenen, D.G., and Kobayashi, S. (2007). BIM mediates EGFR tyrosine ki-
nase inhibitor-induced apoptosis in lung cancers with oncogenic EGFR muta-
tions. PLoS Med. 4, 1669–1680.
Cragg, M.S., Kuroda, J., Puthalakath, H., Huang, D.C.S., and Strasser, A.
(2007). Gefitinib-induced killing of NSCLC cell lines expressing mutant EGFR
requires BIM and can be enhanced by BH3 mimetics. PLoS Med. 4,
1681–1690.
Debacq-Chainiaux, F., Erusalimsky, J.D., Campisi, J., and Toussaint, O.
(2009). Protocols to detect senescence-associated beta-galactosidase (SA-
bgal) activity, a biomarker of senescent cells in culture and in vivo. Nat.
Protoc. 4, 1798–1806.
Drilon, A., Laetsch, T.W., Kummar, S., DuBois, S.G., Lassen, U.N., Demetri,
G.D., Nathenson, M., Doebele, R.C., Farago, A.F., Pappo, A.S., et al. (2018).
Efficacy of larotrectinib in TRK fusion-positive cancers in adults and children.
N. Engl. J. Med. 378, 731–739.
Dupont, S., Morsut, L., Aragona, M., Enzo, E., Giulitti, S., Cordenonsi, M.,
Zanconato, F., Le Digabel, J., Forcato, M., Bicciato, S., et al. (2011). Role of
YAP/TAZ in mechanotransduction. Nature 474, 179–183.
Engelman, J.A., Zejnullahu, K., Mitsudomi, T., Song, Y., Hyland, C., Park, J.O.,
Lindeman, N., Gale, C.-M., Zhao, X., Christensen, J., et al. (2007). MET ampli-
fication leads to gefitinib resistance in lung cancer by activating ERBB3
signaling. Science 316, 1039–1043.
Ercan, D., Xu, C., Yanagita, M., Monast, C.S., Pratilas, C.A., Montero, J.,
Butaney, M., Shimamura, T., Sholl, L., Ivanova, E.V., et al. (2012).
Reactivation of ERK signaling causes resistance to EGFR kinase inhibitors.
Cancer Discov. 2, 934–947.
Ewald, J.A., Desotelle, J.A., Wilding, G., and Jarrard, D.F. (2010). Therapy-
induced senescence in cancer. J. Natl. Cancer Inst. 102, 1536–1546.
Feldman, D., Tsai, F., Garrity, A.J., O’Rourke, R., Brenan, L., Ho, P., Gonzalez,
E., Konermann, S., Johannessen, C.M., Beroukhim, R., et al. (2019).
CloneRetriever: retrieval of rare clones from heterogeneous cell populations
2. bioRxiv. https://doi.org/10.1101/762708.
Ficarro, S., Alexander, W., and Marto, J. (2017). mzStudio: a dynamic digital
canvas for user-driven interrogation of mass spectrometry data. Proteomes
5, 20.
Ficarro, S.B., Browne, C.M., Card, J.D., Alexander, W.M., Zhang, T., Park, E.,
McNally, R., Dhe-Paganon, S., Seo, H.S., Lamberto, I., et al. (2016).
Leveraging gas-phase fragmentation pathways for improved identification
and selective detection of targets modified by covalent probes. Anal. Chem.
88, 12248–12254.
Fridman, A.L., and Tainsky, M.A. (2008). Critical pathways in cellular senes-
cence and immortalization revealed by gene expression profiling. Oncogene
27, 5975–5987.
Gong, Y., Somwar, R., Politi, K., Balak, M., Chmielecki, J., Jiang, X., and Pao,
W. (2007). Induction of BIM is essential for apoptosis triggered by EGFR kinase
inhibitors in mutant EGFR-dependent lung adenocarcinomas. PLoS Med. 4,
1655–1668.
Guler, G.D., Tindell, C.A., Pitti, R., Wilson, C., Nichols, K., KaiWai Cheung, T.,
Kim, H.J., Wongchenko, M., Yan, Y., Haley, B., et al. (2017). Repression of
stress-induced LINE-1 expression protects cancer cell subpopulations from
lethal drug exposure. Cancer Cell 32, 221–237.e13.
Hata, A.N., Niederst, M.J., Archibald, H.L., Gomez-Caraballo, M., Siddiqui,
F.M., Mulvey, H.E., Maruvka, Y.E., Ji, F., Bhang, H.C., Krishnamurthy
Radhakrishna, V., et al. (2016). Tumor cells can follow distinct evolutionary
paths to become resistant to epidermal growth factor receptor inhibition.
Nat. Med. 22, 262–269.
Heinz, S., Benner, C., Spann, N., Bertolino, E., Lin, Y.C., Laslo, P., Cheng, J.X.,
Murre, C., Singh, H., and Glass, C.K. (2010). Simple combinations of lineage-
determining transcription factors prime cis-regulatory elements required for
macrophage and B cell identities. Mol. Cell 38, 576–589.
Hsu, P.-C., You, B., Yang, Y.-L., Zhang, W.-Q., Wang, Y.-C., Xu, Z., Dai, Y., Liu,
S., Yang, C.-T., Li, H., et al. (2016). YAP promotes erlotinib resistance in human
non-small cell lung cancer cells. Oncotarget 7, 51922–51933.
Jamal-Hanjani, M., Wilson, G.A., McGranahan, N., Birkbak, N.J., Watkins,
T.B.K., Veeriah, S., Shafi, S., Johnson, D.H., Mitter, R., Rosenthal, R., et al.
(2017). Tracking the evolution of non-small-cell lung cancer. N. Engl. J. Med.
376, 2109–2121.
Kapoor, A., Yao, W., Ying, H., Hua, S., Liewen, A., Wang, Q., Zhong, Y., Wu,
C.J., Sadanandam, A., Hu, B., et al. (2014). Yap1 activation enables bypass
of oncogenic KRAS addiction in pancreatic cancer. Cell 158, 185–197.
Kim, M., Kim, T., Johnson, R.L., and Lim, D.S. (2015). Transcriptional co-
repressor function of the hippo pathway transducers YAP and TAZ. Cell
Rep. 11, 270–282.
Kuwana, T., Bouchier-Hayes, L., Chipuk, J.E., Bonzon, C., Sullivan, B.A.,
Green, D.R., and Newmeyer, D.D. (2005). BH3 domains of BH3-only proteins
differentially regulate Bax-mediated mitochondrial membrane permeabiliza-
tion both directly and indirectly. Mol. Cell 17, 525–535.
Lehmann, W., Mossmann, D., Kleemann, J., Mock, K., Meisinger, C.,
Brummer, T., Herr, R., Brabletz, S., Stemmler, M.P., and Brabletz, T. (2016).
ZEB1 turns into a transcriptional activator by interacting with YAP1 in aggres-
sive cancer types. Nat. Commun. 7, 1–15.
Lei, Q.-Y., Zhang, H., Zhao, B., Zha, Z.-Y., Bai, F., Pei, X.-H., Zhao, S., Xiong,
Y., and Guan, K.-L. (2008). TAZ promotes cell proliferation and epithelial-
mesenchymal transition and is inhibited by the hippo pathway. Mol. Cell.
Biol. 28, 2426–2436.
Ley, R., Ewings, K.E., Hadfield, K., and Cook, S.J. (2005). Regulatory phos-
phorylation of Bim: sorting out the ERK from the JNK. Cell Death Differ. 12,
1008–1014.
Li, H., and Durbin, R. (2010). Fast and accurate long-read alignment with
Burrows-Wheeler transform. Bioinformatics 26, 589–595.
Liberzon, A., Birger, C., Thorvaldsdóttir, H., Ghandi, M., Mesirov, J.P., and
Tamayo, P. (2015). The molecular signatures database hallmark gene set
collection. Cell Syst. 1, 417–425.
Lin, L., Sabnis, A.J., Chan, E., Olivas, V., Cade, L., Pazarentzos, E., Asthana,
S., Neel, D., Yan, J.J., Lu, X., et al. (2015). The Hippo effector YAP promotes
resistance to RAF- and MEK-targeted cancer therapies. Nat. Genet. 47,
250–256.
Liu, Q., Sabnis, Y., Zhao, Z., Zhang, T., Buhrlage, S.J., Jones, L.H., and Gray,
N.S. (2013). Developing irreversible inhibitors of the protein kinase cysteinome.
Chem. Biol. 20, 146–159.
Mohseni, M., Sun, J., Lau, A., Curtis, S., Goldsmith, J., Fox, V.L., Wei, C.,
Frazier, M., Samson, O., Wong, K.K., et al. (2014). A genetic screen identifies
an LKB1-MARK signalling axis controlling the Hippo-YAP pathway. Nat. Cell
Biol. 16, 108–117.
Mok, T.S., Wu, Y., Thongprasert, S., Yang, C., Saijo, N., Sunpaweravong, P.,
Han, B., Margono, B., Ichinose, Y., Nishiwaki, Y., et al. (2009). Gefitinib or car-
boplatin-paclitaxel in pulmonary adenocarcinoma. N. Engl. J. Med. 361,
947–957.
Mok, T.S., Wu, Y.-L., Ahn, M.-J., Garassino, M.C., Kim, H.R., Ramalingam,
S.S., Shepherd, F.A., He, Y., Akamatsu, H., Theelen, W.S.M.E., et al. (2017).
Osimertinib or platinum-pemetrexed in EGFR T790M-positive lung cancer.
N. Engl. J. Med. 376, 629–640.
Narita, M., Nun, S., Heard, E., Narita, M., Lin, A.W., Hearn, S.A., Spector, D.L.,
Hannon, G.J., Lowe, S.W., Brook, S., et al. (2003). Rb-mediated heterochro-
matin formation and silencing of E2F target genes during cellular senescence.
Cell 113, 703–716.
Noland, C.L., Gierke, S., Schnier, P.D., Murray, J., Sandoval, W.N., Sagolla,
M., Dey, A., Hannoush, R.N., Fairbrother, W.J., and Cunningham, C.N.
(2016). Palmitoylation of TEAD transcription factors is required for their stability
and function in hippo pathway signaling. Structure 24, 179–186.
Peters, S., Camidge, D.R., Shaw, A.T., Gadgeel, S., Ahn, J.S., Kim, D.-W., Ou,
S.-H.I., Pérol, M., Dziadziuszko, R., Rosell, R., et al. (2017). Alectinib versus cri-
zotinib in untreated ALK -positive non-small-cell lung cancer. N. Engl. J. Med.
377, 829–838.
Piccolo, S., Dupont, S., and Cordenonsi, M. (2014). The biology of YAP/TAZ:
hippo signaling and beyond. Physiol. Rev. 94, 1287–1312.
Pobbati, A.V., Han, X., Hung, A.W., Weiguang, S., Huda, N., Chen, G.Y., Kang,
C.B., Chia, C.S.B., Luo, X., Hong, W., et al. (2015). Targeting the central pocket
in human transcription factor TEAD as a potential cancer therapeutic strategy.
Structure 23, 2076–2086.
Qin, Q., Mei, S., Wu, Q., Sun, H., Li, L., Taing, L., Chen, S., Li, F., Liu, T., Zang,
C., et al. (2016). ChiLin: a comprehensive ChIP-seq and DNase-seq quality
control and analysis pipeline. BMC Bioinformatics 17, 1–13.
Ramalingam, S., Saka, H., Ahn, M.-J., Yu, H., Horn, L., Hida, T., Cantarini, M.,
Verheijen, R., Wessen, J., Oxnard, G., et al. (2019). Osimertinib plus selumeti-
nib for patients with EGFR-mutant (EGFRm) NSCLC following disease pro-
gression on an EGFR-TKI: results from the Phase Ib TATTON study. In
AACR Annual Meeting 2019, Atlanta (GA).
Cancer Cell 37, 104–122, January 13, 2020 121
Ramı́rez, F., Ryan, D.P., Gru€ning, B., Bhardwaj, V., Kilpert, F., Richter, A.S.,
Heyne, S., Du €ndar, F., and Manke, T. (2016). deepTools2: a next generation
web server for deep-sequencing data analysis. Nucleic Acids Res. 44,
W160–W165.
Richardson, C.D., Ray, G.J., DeWitt, M.A., Curie, G.L., and Corn, J.E. (2016).
Enhancing homology-directed genome editing by catalytically active and inac-
tive CRISPR-Cas9 using asymmetric donor DNA. Nat. Biotechnol. 34,
339–344.
Rosell, R., Carcereny, E., Gervais, R., Vergnenegre, A., Massuti, B., Felip, E.,
Palmero, R., Garcia-Gomez, R., Pallares, C., Sanchez, J.M., et al. (2012).
Erlotinib versus standard chemotherapy as first-line treatment for European
patients with advanced EGFR mutation-positive non-small-cell lung cancer
(EURTAC): a multicentre, open-label, randomised phase 3 trial. Lancet
Oncol. 13, 239–246.
Rosenbluh, J., Nijhawan, D., Cox, A.G., Li, X., Neal, J.T., Schafer, E.J., Zack,
T.I., Wang, X., Tsherniak, A., Schinzel, A.C., et al. (2012). b-Catenin-driven can-
cers require a YAP1 transcriptional complex for survival and tumorigenesis.
Cell 151, 1457–1473.
Sequist, L.V., Waltman, B.A., Dias-Santagata, D., Digumarthy, S., Turke, A.B.,
Fidias, P., Bergethon, K., Shaw, A.T., Gettinger, S., Cosper, A.K., et al. (2011).
Genotypic and histological evolution of lung cancers acquiring resistance to
EGFR inhibitors. Sci. Transl. Med. 3, 75ra26.
Shao, D.D., Xue, W., Krall, E.B., Bhutkar, A., Piccioni, F., Wang, X., Schinzel,
A.C., Sood, S., Rosenbluh, J., Kim, J.W., et al. (2014). KRAS and YAP1
converge to regulate EMT and tumor survival. Cell 158, 171–184.
Sharma, S.V., Lee, D.Y., Li, B., Quinlan, M.P., Takahashi, F., Maheswaran, S.,
McDermott, U., Azizian, N., Zou, L., Fischbach, M.A., et al. (2010). A chro-
matin-mediated reversible drug-tolerant state in cancer cell subpopulations.
Cell 141, 69–80.
Shibue, T., and Weinberg, R.A. (2017). EMT, CSCs, and drug resistance: the
mechanistic link and clinical implications. Nat. Rev. Clin. Oncol. 14, 611–629.
Soria, J.-C., Ohe, Y., Vansteenkiste, J., Reungwetwattana, T.,
Chewaskulyong, B., Lee, K.H., Dechaphunkul, A., Imamura, F., Nogami, N.,
Kurata, T., et al. (2018). Osimertinib in untreated EGFR-mutated advanced
non-small-cell lung cancer. N. Engl. J. Med. 378, 113–125.
Sudol, M. (2012). YAP1 oncogene and its eight isoforms. Oncogene 32, 3922.
Tang, Y., Feinberg, T., Keller, E.T., Li, X.Y., and Weiss, S.J. (2016). Snail/Slug
binding interactions with YAP/TAZ control skeletal stem cell self-renewal and
differentiation. Nat. Cell Biol. 18, 917–929.
Thress, K.S., Jacobs, V., Angell, H.K., Yang, J.C.H., Sequist, L.V., Blackhall, F.,
Su, W.C., Schuler, M., Wolf, J., Gold, K.A., et al. (2017). Modulation of
biomarker expression by osimertinib: results of the paired tumor biopsy co-
horts of the AURA phase I trial. J. Thorac. Oncol. 12, 1588–1594.
122 Cancer Cell 37, 104–122, January 13, 2020
Tricker, E.M., Xu, C., Uddin, S., Capelletti, M., Ercan, D., Ogino, A., Pratilas,
C.A., Rosen, N., Gray, N.S., Wong, K., et al. (2015). Combined EGFR/MEK in-
hibition prevents the emergence of resistance in EGFR-mutant lung cancer.
Cancer Discov. 5, 960–971.
Wang, W., Li, N., Li, X., Tran, M.K., Han, X., and Chen, J. (2015). Tankyrase in-
hibitors target YAP by stabilizing angiomotin family proteins. Cell Rep. 13,
524–532.
Wang, Y., Xu, X., Maglic, D., Dill, M.T., Mojumdar, K., Ng, P.K.-S., Jeong, K.J.,
Tsang, Y.H., Moreno, D., Bhavana, V.H., et al. (2018). Comprehensive molec-
ular characterization of the hippo signaling pathway in cancer. Cell Rep. 25,
1304–1317.e5.
Zaidi, S.K., Sullivan, A.J., Medina, R., Ito, Y., van Wijnen, A.J., Stein, J.L., Lian,
J.B., and Stein, G.S. (2004). Tyrosine phosphorylation controls Runx2-medi-
ated subnuclear targeting of YAP to repress transcription. EMBO J. 23,
790–799.
Zanconato, F., Cordenonsi, M., and Piccolo, S. (2016). YAP/TAZ at the roots of
cancer. Cancer Cell 29, 783–803.
Zhang, Z., and Marshall, A.G. (1998). A universal algorithm for fast and auto-
mated charge state deconvolution of electrospray mass-to-charge ratio
spectra. J. Am. Soc. Mass Spectrom. 9, 225–233.
Zhang, H., Liu, C.Y., Zha, Z.Y., Zhao, B., Yao, J., Zhao, S., Xiong, Y., Lei, Q.Y.,
and Guan, K.L. (2009). TEAD transcription factors mediate the function of TAZ
in cell growth and epithelial-mesenchymal transition. J. Biol. Chem. 284,
13355–13362.
Zhang, J., Smolen, G.A., and Haber, D.A. (2008a). Negative regulation of YAP
by LATS1 underscores evolutionary conservation of the Drosophila Hippo
pathway. Cancer Res. 68, 2789–2794.
Zhang, Y., Liu, T., Meyer, C.A., Eeckhoute, J., Johnson, D.S., Bernstein, B.E.,
Nussbaum, C., Myers, R.M., Brown, M., Li, W., et al. (2008b). Model-based
analysis of ChIP-seq (MACS). Genome Biol. 9, R137.
Zhang, Z., Lee, J.C., Lin, L., Olivas, V., Au, V., Laframboise, T., Abdel-Rahman,
M., Wang, X., Levine, A.D., Rho, J.K., et al. (2012). Activation of the AXL kinase
causes resistance to EGFR-targeted therapy in lung cancer. Nat. Genet. 44,
852–860.
Zhao, B., Ye, X., Yu, J., Li, L., Li, W., Li, S., Yu, J., Lin, J.D., Wang, C.-Y.,
Chinnaiyan, A.M., et al. (2008). TEAD mediates YAP-dependent gene induction
and growth control. Genes Dev. 22, 1962–1971.
Zhou, W., Ercan, D., Chen, L., Yun, C.-H., Li, D., Capelletti, M., Cortot, A.B.,
Chirieac, L., Iacob, R.E., Padera, R., et al. (2009). Novel mutant-selective
EGFR kinase inhibitors against EGFR T790M. Nature 462, 1070–1074.
Zorita, E., Cuscó, P., and Filion, G.J. (2015). Starcode: sequence clustering
based on all-pairs search. Bioinformatics 31, 1913–1919.
 Changing medicine.
Together.

Med, a new journal from Cell Press, publishes transformative, evidence-based science across
the clinical and translational research continuum – from large-scale clinical trials to translational
studies with demonstrable functional impact, offering novel insights in disease understanding.
We aim to elevate the global standard of medical research by accelerating translation of bench
research to the clinic, serving as a hub for engagement between all stakeholders, improving
reproducibility, and changing medical practice.
Elevate your research. Submit your paper today.
cell.com/med
 ll

Article
Predicting Drug Response and Synergy
Using a Deep Learning Model of Human Cancer Cells
Brent M. Kuenzi,1,5 Jisoo Park,1,5 Samson H. Fong,1,2 Kyle S. Sanchez,1 John Lee,1 Jason F. Kreisberg,1 Jianzhu Ma,4
and Trey Ideker1,2,3,6,*
1Division of Genetics, Department of Medicine, University of California San Diego, La Jolla, CA 92093, USA
2Department of Bioengineering, University of California San Diego, La Jolla, CA 92093, USA
3Department of Computer Science and Engineering, University of California San Diego, La Jolla, CA 92093, USA
4Department of Computer Science, Purdue University, West Lafayette, IN 47907, USA
5These authors contributed equally
6Lead Contact

*Correspondence: tideker@ucsd.edu
https://doi.org/10.1016/j.ccell.2020.09.014

SUMMARY

Most drugs entering clinical trials fail, often related to an incomplete understanding of the mechanisms gov-
erning drug response. Machine learning techniques hold immense promise for better drug response predic-
tions, but most have not reached clinical practice due to their lack of interpretability and their focus on mono-
therapies. We address these challenges by developing DrugCell, an interpretable deep learning model of
human cancer cells trained on the responses of 1,235 tumor cell lines to 684 drugs. Tumor genotypes induce
states in cellular subsystems that are integrated with drug structure to predict response to therapy and,
simultaneously, learn biological mechanisms underlying the drug response. DrugCell predictions are accu-
rate in cell lines and also stratify clinical outcomes. Analysis of DrugCell mechanisms leads directly to the
design of synergistic drug combinations, which we validate systematically by combinatorial CRISPR,
drug-drug screening in vitro, and patient-derived xenografts. DrugCell provides a blueprint for constructing
interpretable models for predictive medicine.

INTRODUCTION 2019). In a typical application (reviewed in Table S1), the model

Each year dozens of new therapies enter clinical trials for the po-
tential treatment of various types of cancer, but fewer than 4%
will ultimately gain approval by the US Food and Drug Adminis-
tration (Wong et al., 2019). Although many factors contribute to
this challenge, a major failure is in understanding how or why a
particular cancer responds to therapy. The problem becomes
particularly acute for cancers that are not associated with strong
targetable genetic drivers (e.g., BCR-ABL fusion, EGFR muta-
tion, or EML4-ALK translocation), since cancers without these
known drivers lack clear biomarkers with which to stratify drug
response. A better basic understanding of the molecular path-
ways governing drug sensitivity would help greatly in deter-
mining which patients should be treated and with which drugs.
There has recently been a great deal of interest in applying ad-
vances in artificial intelligence, including machine learning and
deep learning, to classic problems in biomedicine (Topol,
2019). Whereas popular applications include disease diagnosis
from biomedical images and interpretation of electronic medical
records (Esteva et al., 2019; Rajkomar et al., 2019; Wainberg
et al., 2018), machine learning models are also of high interest
in predicting drug responses (Barretina et al., 2012; Costello
et al., 2014; Garnett et al., 2012; Iorio et al., 2016; Zeng et al.,
uses the ’omics profile of a cell line or tissue sample as input
to predict the 50% inhibitory concentration (IC50) of a drug. For
example, Iorio et al. (2016) built elastic net models to predict
the drug IC50 of cancer cell lines given their profiles of gene mu-
tations and expression levels; a range of predictive accuracy is
observed, depending on the compound. Using the same data-
set, Cortés-Ciriano et al. (2016) showed that predictive perfor-
mance could in some cases be improved using a random forest
model linked to a measure of statistical confidence in each pre-
diction. Deep neural networks (Baptista et al., 2020; Chiu et al.,
2019; Menden et al., 2013; Sakellaropoulos et al., 2019) and vari-
ational autoencoders (Rampá
sek et al., 2019) have also been
applied to drug response prediction, with significant perfor-
mance gains noted depending on the drug and disease context.
Owing to the significant molecular heterogeneity observed
across tumors, there are often many different molecular features
and feature combinations that can lead a model to predict a partic-
ular drug response. What these features are, and whether they are
distinct or functionally interrelated, can be very difficult to interpret,
however. The reason is that most machine learning models are
‘‘black boxes,’’ optimized for prediction accuracy without knowl-
edge of or attention to the biological mechanisms underlying pre-
dicted outcomes (Ching et al., 2018). To address these difficulties,

672 Cancer Cell 38, 672–684, November 9, 2020 ª 2020 Elsevier Inc.

Article
A B
Genotype Drug
Increasing complexity & scale
Embedding Embedding
of genotype of chemical
via cell model structure
VNN ANN
In silico
treatment of
cell with drug
DrugCell
Response of cell to drug
Figure 1. DrugCell Design
(A) DrugCell uses a modular neural network design that combines conventional artificial neural networks (ANN) with a visible neural network (VNN) to make drug
response predictions.
(B) Binary encodings of individual genotypes are processed through a VNN with architecture guided by a hierarchy of cell subsystems, with multiple neurons
assigned per subsystem.
(C) Compound chemical structures are processed through an ANN using the Morgan fingerprint as input features.
model interpretation is now a rapidly growing subfield within ma-
chine learning, with a growing arsenal of approaches for achieving
models with not only high predictive accuracy, but also high
descriptive accuracy (Murdoch et al., 2019). One major strategy
has been to use prior knowledge or data to add structure to the
model, which can then be interpreted. Applied to genomics, such
a strategy has been used to recast the thousands of measured mo-
lecular features of a tumor as states on a much smaller number of
functional modules (Cortés-Ciriano et al., 2016; Yang et al., 2019).
For example, a recent study mapped raw molecular measurements
to a set of pre-defined metabolic pathways drawn from prior knowl-
edge bases; the states of these pathways predict antibiotic resis-
tance in Escherichia coli, with particular pathway features emerging
as candidate mechanisms of resistance (Yang et al., 2019). Organi-
zation of molecular features into predictive modules can also be
accomplished using prior data as opposed to literature-curated
knowledge. Such an approach was recently exemplified by Deep-
Profile, which analyzed a large collection of leukemia expression
profiles to extract a low-dimensional representation of these data
as a set of functional gene modules; these modules are then
used as interpretable features for drug response prediction (Dincer
et al., 2018). Apart from model-based approaches, a second major
strategy to increase model interpretability has been to perform post
hoc analysis of model features or feature weights to interpret the
underlying drug response mechanisms (Chiu et al., 2019; Iorio
et al., 2016; Murdoch et al., 2019). For example, the weights as-
signed to each input gene by a black-box neural network model
are subjected to gene set enrichment analysis (Subramanian
et al., 2005) to identify pathways regulating the predicted drug
response (Sakellaropoulos et al., 2019). These pathways, however,
were not used during modeling or validated experimentally.
To more explicitly link the structure of a machine learning
model to cellular functions, we recently developed a visible neu-
ral network (VNN) simulating a simple eukaryotic cell, Saccharo-
myces cerevisiae (Ma et al., 2018; Yu et al., 2018). This model,
called DCell, was made mechanistically interpretable, or
‘‘visible,’’ by directly mapping the neurons of a deep neural
network into a large hierarchy of known and putative molecular
components and pathways. DCell is able to accurately predict
ll
Binary mutations 2,086 subsystems C Chemical
6 neurons / subsystem structure
5 layers
Morgan fingerprint
Genes
Large complexes,
signaling pathways Neurons
100 ...
Organelles,
broad processes
50 ...
Cellular processes
6 ...
Genotype embedding Drug structure embedding
the impact of genetic mutations on cellular growth response
and, simultaneously, identify the most relevant molecular path-
ways driving those predictions. Building from this paradigm,
we now describe DrugCell, a VNN that simulates the response
of human cancer cells to therapeutic chemical compounds.
DrugCell couples the inner workings of the model to the hierar-
chical structure of human cell biology, allowing for response pre-
dictions for any drug in any cancer and intelligent design of effec-
tive combination therapies.
RESULTS
Design and Training of an Interpretable Neural Network
of Drug Response
The cellular drug response is a complex phenomenon that de-
pends on both biological and chemical factors (Turner et al.,
2015). Current black-box models of drug response that use
both these factors have begun to reach the limits of predictive
performance (Table S1). We therefore aimed to design a model
that maintains this high level of predictive capability while gaining
mechanistic interpretability of the model predictions. To capture
both determinants of drug response in an interpretable model,
we devised DrugCell as a neural network with two branches (Fig-
ure 1A, STAR Methods). The first branch was a VNN modeling
the hierarchical organization of molecular subsystems in a hu-
man cell, drawn from 2,086 biological processes documented
in the Gene Ontology (GO) database (Ashburner et al., 2000) (Fig-
ure S1A). Each of these subsystems, from those involving small
protein complexes (e.g., b-catenin destruction complex) to
larger signaling pathways (e.g., MAPK signaling pathway) to
overarching cellular functions (e.g., glycolysis), was assigned a
bank of artificial neurons to represent the state of that subsystem
(Figure 1B). Connectivity of neurons was set to mirror the biolog-
ical hierarchy, so that neurons accept inputs only from child sub-
systems and send outputs only to parent systems, with connec-
tion weights determined during training. The use of multiple
neurons per subsystem (here six, see STAR Methods) allowed
cellular subsystems to be multifunctional, with distinct states
able to adopt a range of values along multiple dimensions
Cancer Cell 38, 672–684, November 9, 2020 673
 ll
Article

A B C
rho = 0.8
1.0 1.0

Spearman correlation (DrugCell)

Spearman correlation (DrugCell)
1.0
Predicted drug response (AUC)

0.8 0.5 0.5

0.6
0.0 0.0

0.4
-0.5 -0.5
0.2
p < 0.0001 p > 0.05
-1.0 -0.5 0.5 1.0 -1.0 -0.5 0.0 0.5 1.0
0.0
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0

0.0
Inf

Spearman correlation (Elastic Net) Spearman correlation (matched black box)

Actual drug response (AUC)

Spearman rho
0.8
D E 0.6
1.0 0.4
Spearman correlation (DrugCell)

Spearman rho (Predicted vs. actual)

0.2
0.5
0.8

G ML tine
46 030

rb et 4
da el
le
11 lit 1

Ti 02 l
Te an 8
po ib
de
6 e
Pa Doc 136

C c 9

v 5
en ax

36 x
KX zo

ni tin
N Pa 2-3

si
a
ris
SK -
nc
Performance

Vi
0.0

ZI
0.4
-0.5

p < 0.0001 0.0

-1.0 -0.5 0.0 0.5 1.0 Drugs
Spearman correlation (tissue only black box)

High confidence drugs (rho > 0.5)

Figure 2. Predictive Performance

(A) Predicted versus actual drug responses across all (cell line, drug) pairs studied. Box plots show the 25th, 50th, and 75th percentiles of values in each bin;
whiskers show maximum and minimum values.
(B–D) Scatterplots of the predictive performance (Spearman rho between actual and predicted drug response across 684 drugs) of DrugCell versus three
alternative models: (B) elastic net, (C) matched black-box neural network, and (D) tissue-only black-box neural network. Points represent individual drugs; points
above the diagonal represent drugs better predicted by DrugCell.
(E) Waterfall plot of predictive performance for each drug in the dataset (y axis), ranked from highest to lowest (x axis). ‘‘High confidence’’ drugs are highlighted in
red (rho > 0.5). The inset shows the performance for the top 10 best predicted drugs.

(Copley, 2012). The input layer of the hierarchy mapped to the recording the mutational status (1 = mutated, 0 = non-mutated)
mutation status of genes. The six neurons at the VNN output, of the top 15% most frequently mutated genes in cancer (n =
corresponding to the root of the hierarchy, represented the 3,008; median mutated genes per cell line = 73; Figure S1E).
embedded state of the whole cell based on its genotype (Fig- Each drug’s chemical structure was represented by an average
ure 1B). In total, the VNN used 12,516 neurons distributed hier- of 81 activated bits in the Morgan fingerprint vector, with each bit
archically across six distinct layers (STAR Methods, Figure S1B). typically representing fewer than 10 molecular fragments (Fig-
The second branch of DrugCell was a conventional artificial neu- ures S1F and S1G). DrugCell was trained to associate each ge-
ral network (ANN) embedding the Morgan fingerprint of a drug, a notype-drug pair with its corresponding drug response,
canonical vector representation of chemical structure (Figure 1C, measured by the area under the dose-response curve (AUC,
STAR Methods) (Rogers and Hahn, 2010). Outputs from the two STAR Methods). The DrugCell model and its codebase are avail-
branches of the model, the VNN embedding cell genotype and able for public download on GitHub (https://github.com/
the ANN embedding drug structure, were combined in a single idekerlab/DrugCell).
layer of neurons, which were then integrated to generate
the response of a given genotype to a particular treatment Interpretable Modeling of Drug Response Has No
(Figure 1A). Performance Loss
To train the model, we harmonized data from two large cancer We first sought to assess the prediction accuracy of DrugCell us-
drug screening resources: the Cancer Therapeutics Response ing the Spearman correlation (rho) between predicted and
Portal (CTRP) v2 and the Genomics of Drug Sensitivity in Cancer observed AUC values in 5-fold cross validation (STAR Methods).
(GDSC) database (Seashore-Ludlow et al., 2015; Yang et al., The total accuracy over all cell line-drug pairs was rho = 0.80
2013). The combined dataset consisted of 509,294 cell line- (Figure 2A). Further insight was achieved by computing the pre-
drug pairs, covering 684 drugs and 1,235 cell lines (Figure S1C, diction accuracy for each drug individually, revealing a subpop-
STAR Methods). All major tissue types were represented, with ulation of drugs with very high prediction accuracy (30% of
hematopoietic and lung lineages the most prevalent (Figure S1D). drugs with rho > 0.5) amid a much wider general distribution
Each cell-line genotype was represented by a binary vector (range 0.29 to +0.83, median 0.37). These accuracies were

674 Cancer Cell 38, 672–684, November 9, 2020


Article
significantly higher than those achieved for elastic net (median
rho = 0.35), a state-of-the-art regression technique used in
many previous approaches to drug response prediction (Eskio-
cak et al., 2017; Iorio et al., 2016; Kuenzi et al., 2019; Potts
et al., 2015) (Figure 2B). DrugCell’s drug-by-drug predictive per-
formance was not significantly different from that of a conven-
tional black-box ANN with matching numbers of neurons, layers,
and connections (Figure 2C). It was also comparable to previous
efforts to incorporate chemical features of drugs into the
response prediction (e.g., structure and physiochemical proper-
ties such as solubility, lipophilicity, and molecular weight), and it
outperformed models that predict response using biological fea-
tures alone (e.g., expression of biomarkers, point mutation, copy
number variation, and microsatellites; Table S1). Finally, since
knowledge of tissue type can be predictive of drug response
even in the absence of other information (Iorio et al., 2016), we
considered that some of the performance of these models might
be due to their ability to recognize the tissue type of a cell line
from its input data (i.e., its mutational profile). Accordingly, we
compared DrugCell with an equivalent neural network model
trained on drug structure and a tissue label only (STAR Methods).
DrugCell vastly outperformed this tissue-only model (median
rho = 0.18; Figure 2D), indicating that the model had learned in-
formation from somatic mutations beyond the tissue of origin.
Compounds for which DrugCell predictions were most accu-
rate came from diverse target classes, including chemothera-
peutics (e.g., vincristine, teniposide) and targeted therapies
(e.g., GSK461364 targeting PLK1, KX2-391 targeting Src; Fig-
ure 2E). DrugCell maintained the specificity of the training data
in that its predictions were specific to individual classes of drugs
(e.g., MEK inhibitor predictions were highly specific) and did not
simply reflect general drug toxicity (Figure S2A). Predictive per-
formance for a drug did not strongly correlate with the number
of cell line-drug pairs used for training, nor with the structural
complexity of a compound (number of activated bits; Figures
S2B and S2C). We did find that compounds eliciting a larger
range of cell-line responses tended to be more predictable (Fig-
ure S2D). Similarly, individual cell lines (Figure S2E) and tissue
types (Figure S2F), which elicit a large range of responses,
were in general highly predictable.
DrugCell Learns Mechanisms that Mediate Specific
Drug Responses
Having evaluated predictive ability, we next turned to mecha-
nistic interpretation. This task was aided by the two model
branches, which dissect the effects of genotype on the configu-
ration of cell systems (genotype embedding) from the effects of
chemical structure on drug activity within the cell (drug embed-
ding, Figure 1A). We visually inspected these embeddings by
plotting the top two principal components (Figures 3A–3E). The
genotype embedding from the VNN revealed a separation of ge-
notypes according to mutations known to confer specific drug
sensitivities, such as activating mutations in BRAF (Figure 3A)
that promote sensitivity to the MEK inhibitor selumetinib (Fig-
ure 3B). The genotype embedding also distinguished mutations
leading to drug resistance, such as mutations in EGFR (Yin et al.,
2019), LKB1 (Shimamura et al., 2013), or BRAF (Ma et al., 2017)
(Figure 3C) that confer resistance to the BET-family inhibitor JQ-
1 (Figure 3D). We similarly inspected the DrugCell embeddings of
ll
individual subsystems within the VNN and found that many were
in agreement with subsystem activities measured experimentally
by an independent analysis of protein abundances and phos-
phorylation states using reverse-phase protein arrays (RPPAs;
Figure S3A; STAR Methods). For example, DrugCell accurately
captured MAPK pathway activity within the subsystem embed-
ding of Regulation of MAPK cascade (Figure S3B), which signif-
icantly correlated with ERK1/2 phosphorylation (Figure S3C).
Overall, the majority of DrugCell subsystems were well corre-
lated with the RPPA measurements of those subsystems (note
bimodal distribution of correlation in Figure S3A). Other accu-
rately captured subsystems included Proteolysis (Figure S3D),
Regulation of PI3K signaling (Figure S3E), and Cell-cycle arrest
(Figure S3F).
Inspection of the drug embedding from the ANN revealed a
stratification of drugs based on their mechanisms of action
within major drug target classes (Figure 3E). The distance be-
tween each pair of drugs in the chemical structure embedding
did not correlate with their overall chemical similarity (Fig-
ure S4A), consistent with previous studies of drug activity and
chemical structure (Breinig et al., 2015). Since the training data
consisted solely of drugs and drug-like molecules, the chemical
structural embedding did not stratify drugs on chemical features
such as membrane permeability (Figure S4B), solubility (Fig-
ure S4C), or pharmacodynamic properties (Lipinski; Figure S4D).
Together these results suggest that DrugCell is able to learn key
features of the genotype that govern drug sensitivity and resis-
tance, as well as features of chemical structure that govern
drug biological activity.
Since DrugCell’s VNN is structured according to the hierarchy
of biological subsystems comprising a human cell, its output (ge-
notype embedding) is the result of state changes in particular
subsystems within that hierarchy. To identify the most important
of these subsystems, we scored subsystems by the degree to
which their states were significantly more predictive of a drug
response than the states of their child subsystems using the rela-
tive local improvement in predictive power metric (RLIPP, STAR
Methods) (Ma et al., 2018). As an initial proof of concept, we used
RLIPP scoring to identify subsystems important for the cellular
response to taxol (paclitaxel), an agent that stabilizes microtu-
bules (Figures 2E and 3F, Table S2). Among the top scores for
paclitaxel, many subsystems were metabolic processes (hyper-
geometric p < 0.05; Figures 3G and 3H), including Response to
cAMP (top score) along with Insulin secretion in response to
glucose and Response to glucose. We confirmed by inspection
that the states of these subsystems had the ability to stratify
paclitaxel sensitive versus resistant cell lines (e.g., Response to
cAMP subsystem, Figure 3I). Given these underlying metabolic
pathways, we hypothesized that paclitaxel efficacy might be
modulated by metabolic perturbation. We therefore exposed
A427 cells to three different treatments – paclitaxel, the glycol-
ysis inhibitor 2-deoxyglucose (2-DG), or a combination of the
two – and found that the combination was substantially more
effective than either individual compound (Figure 3J).
A similar analysis was performed for the next (second-most)
important subsystem, Regulation of ubiquitin-protein trans-
ferase activity (Figure 3G, Table S2). We combined paclitaxel
with perturbation of ubiquitin-dependent protein degradation
via the proteasome inhibitor bortezomib (Figure S5A). We found
Cancer Cell 38, 672–684, November 9, 2020 675
 ll
Article

A B C D
Sensitive Resistant Sensitive Resistant
BRAF mutations EGFR, BRAF,
0.0 0.4 0.8 1.2 or LKB1 mutations 0.0 0.4 0.8 1.2
Selumetinib AUC JQ-1 AUC

Genotype embedding (PC2)

Genotype embedding (PC2)

Genotype embedding (PC2)

Genotype embedding (PC2)

n = 229 n = 382 n = 460 n = 860
Genotype embedding (PC1) Genotype embedding (PC1) Genotype embedding (PC1) Genotype embedding (PC1)

E H
Chemical structure embedding (PC2)

Insulin
secretion in
response to
glucose
Response to
cAMP
Response to
glucose

Target class
EGFR
MEK
CDK
BRAF
PARP
BRD4

Chemical structure
embedding (PC1)
Sensitive Resistant
F Sensitive Resistant G I 0.0 0.4 0.8 1.2 J
Metabolic pathway
Paclitaxel AUC
Response to cAMP embedding (PC3)

0.0 0.4 0.8 1.2 110 ***

Paclitaxel AUC Response ***
Genotype embedding (PC2)

to cAMP ***
Importance for paclitaxel response

100
Regulation

Cell viability (%)

of ubiquitin
transferase 90
100
Insulin secretion
(RLIPP score)

in response 80
to glucose
70
Response
to glucose 60
10
n = 239 n = 239 50
Top 5% of subsystems Paclitaxel (8nM)
Genotype embedding (PC1) Response to cAMP
2-DG (800µM)
embedding (PC1)

Figure 3. Characterization of Cancer Cell States Learned by DrugCell

(A–D) Genotype embeddings of each cell line, showing the first two principal components (PC). Points are cell lines, with colors indicating specific drug responses
or genetic markers according to the panel. (A and C) Green denotes cell lines harboring mutations in BRAF or in EGFR, BRAF, or LKB1, respectively. Gray denotes
cell lines without mutations in these genes. (B and D) Blue-to-red gradient represents the response to selumetinib or JQ-1, respectively. Gray denotes cell lines
not tested against that drug.
(E) Drug structure embedding. Points are drugs, with colors indicating drug target classes.
(F) Genotype embeddings of each cell line as in (A–D), but with blue-to-red gradient representing response to paclitaxel.
(G) Waterfall plot of top 5% of subsystems (x axis) important for paclitaxel response by RLIPP score (y axis). Subsystems capturing metabolic pathways are
highlighted in red.
(H) Visualization of select subsystems highlighted in (G), comprising a sub-hierarchy of the full DrugCell model. Red is used to trace the branches of the hierarchy
related specifically to regulation of glycolysis.
(I) Response to cAMP subsystem embedding. Points are cell lines, blue-to-red gradient represents response to paclitaxel.
(J) Boxplot of the relative cell viability of treatment with DMSO, paclitaxel, 2-deoxyglucose (2-DG), or the combination at the indicated concentrations in A427
cells. Data are representative of drug treatments performed in biological and technical triplicates. The boxes represent the interquartile range (IQR) bisected by
the median, whiskers represent the maximum and minimum range of the data that do not exceed 1.5 times the IQR. ***p < 0.0001 from a t test.

that these treatments were antagonistic, consistent with recent pathways were not identified by earlier analyses of genetic
findings showing that glycolysis is subject to negative physical mutations (Table S4) and were distinct from those identified
regulation by ubiquitin ligases at the cytoskeleton (Park et al., by differential mRNA expression analysis of paclitaxel sensitive
2020). Ubiquitin and subsystems were also identified for doce- versus resistant lines (Figures S5B and S5C). Unlike the
taxel, a sister compound (Table S3). Notably, these DrugCell glycolytic perturbations emerging from DrugCell analysis,

676 Cancer Cell 38, 672–684, November 9, 2020

 ll
Article

A B C D E

Trametinib RLIPP Score

2K

P1

Nutlin-3 RLIPP Score

Olaparib RLIPP Score
53
Performance

300

AP

R
-5 0 5

PA

TP
M
120
40
200
Area under 80
fitness curve 20
40 100
of double KO
Select 0 0 0

of lic en res ER

on via an se

bi eo n

im e
us
accurately

sy sfe hyl t
on

D e lo it y

t o lic t io
ic om
pl tion

lo R

al thy t

du n
n
n

pa en

en
in

rg on

ul
ns tio

io
su tra m me

N
iv
st ras ati

se sp iza
ag
A pm

s
ct
de d N

pm
de act

p
ou za

tra la
predicted

N n fi mu pled
9 op

st
ck

Po eg llar pt-c ani

e
l

em e
et
qu H eve

ve

im u
gn me

e
compounds

ot
A ed -co

re ing t o
n
sc t or
d

t
iti 3K
er

ia r ip
g. on tex
er ran en

AS N
n

g. sp m
. o Re lay

,
R fD

ed sc
ri
T m

la
si
co
bi of

-m an
o
e

f il

t
in

f u g.

Ig Tr

sp
al

Po mR cti
th

t in

re ati
e
in

Ac

re A
eb

of
Vi

ul
r
by
La

R
eg

s.

s.
176 cancer genes

R
F G H NS
*
* BLM
3 POLK
TDP2
2.0
Identify subsystems 1
mediating drug 2

Area under fitness curve

Area under fitness curve
0 1.5
Area under fitness curve

response

in combination with
in combination with
in combination with

PARP1 KO
MAP2K1 KO

TP53 KO
-1 1.0
CDK7
0
-2 0.5 KDM1A
TOP2A
-1
-3
PLK1 0.0 PRMT5
LIG1
-4 MPHOSPH8
-2
CDK7
CDK1 -0.5
Dual gene KOs to validate top subsystem

Gene KOs in
top nutlin-3
subsystems
Gene KOs in
top olaparib
subsystems

Gene KOs in
random
subsystems
Gene KOs in
random
subsystems
Gene KOs in
top trametinib
subsystems
Gene KOs in
random
subsystems

importance using CRISPR/Cas9

Figure 4. Systematic Validation of Identified Mechanisms of Sensitivity Using CRISPR/Cas9

(A) Workflow of systematic analysis using CRISPR/Cas9.
(B) Heatmap of the area under the fitness curves for 176 cancer genes in combination with MAP2K1, PARP1, and TP53.
(C–E) Bar plots of the RLIPP scores of the top five subsystems for (C) trametinib, (D) olaparib, and (E) nutlin-3.
(F–H) Boxplots of the area under the fitness curve following CRISPR/Cas9-mediated knockout of (F) MAP2K1, (G) PARP1, and (H) TP53 in combination with highly
weighted genes within the top five subsystems identified by DrugCell for each parent drug compared with random. Select genes are labeled. The boxes represent
the IQR bisected by the median, and whiskers represent the maximum and minimum range of the data that do not exceed 1.5 times the IQR. *p < 0.05 from a t test,
NS denotes not significant.

combination treatments suggested by differentially expressed which had broad representation of cancer signaling pathways
pathways were not successful at enhancing paclitaxel efficacy (MCF7 cells; Figure 4B). The top five important subsystems in
(Figure S5D). the response to each drug were identified (RLIPP analysis; Fig-
Moving beyond paclitaxel to examine the important subsys- ures 4C–4E), along with the genes in those subsystems covered
tems identified for other drugs, we found that some of these sub- by the CRISPR library. Combinatorial disruption of MAPK1 with
systems corresponded to previously identified mechanisms of genes in trametinib subsystems (Figure 4C) resulted in signifi-
drug sensitivity, while many others were novel pathways war- cantly more cell killing than observed for genes from random un-
ranting further investigation. In particular, we examined 60 drugs important subsystems (Figure 4F). A similar cell killing effect (Fig-
for which pan-cancer diagnostic gene mutations had been re- ure 4G) was observed for combinatorial disruption of PARP1 with
ported by an earlier analysis of the GDSC dataset using type II genes in olaparib subsystems (Figure 4D). In contrast, combina-
error ANOVA modeling (Iorio et al., 2016). For a number of drugs, torial disruption of TP53 with genes in nutlin-3 subsystems (Fig-
DrugCell recovered the previously reported diagnostic gene(s) ure 4E) had effects on cell growth that were not significantly
within the top subsystem (4 drugs) or top 10 subsystems (14 different from random (Figure 4H). This result was expected, as
drugs, upper 0.4th percentile of subsystems). For the vast major- TP53 knockout has the opposite effect compared with nutlin-3,
ity, however (56 drugs), DrugCell achieved better predictive per- which leads to p53 activation. These results, together with the
formance by consulting additional, or different, markers than had preliminary results from paclitaxel, provide systematic support
been previously reported (Table S4). for the importance of top response pathways identified by
Given the extent of novel drug response pathways, we sought DrugCell.
to systematically investigate the indicated mechanisms (Fig-
ure 4A; STAR Methods), focusing on trametinib, a MEK1 inhibi- Identified Subsystems Represent Synergistic Drug
tor; olaparib, a PARP1 inhibitor; and nutlin-3, an MDM2 antago- Combination Opportunities
nist that stabilizes and activates p53. CRISPR knockouts of each The parallel pathway inhibition theory of drug synergy (Yeh et al.,
of the three drug targets (MEK1, PARP1, TP53) were combined 2009) holds that two drugs will be synergistic if they inhibit sepa-
with knockouts of each gene in a custom CRISPR/Cas9 library, rate pathways that regulate a common essential function

Cancer Cell 38, 672–684, November 9, 2020 677

 ll
Article

A B C D 40 *** E
***
Pathway 1 Pathway 2 Pathway(s) MoA
Negative
A S X A S regulation of Regulation of
yn yn

DeepSynergy synergy score

er Drug A er Regulation of ERK1/ERK2 proteasomal protein
gy gy 25 compounds in PI3K activity cascade catabolic process
20
Drug B DeepSynergy database (PI3K) (ERK) (neg. control)
B Y B Y D1
26.1 77.5 0.86
Select
Drug A subsystems
0
C Z C RLIPP analysis

...

...
...
top 5 bottom 5
VNN ANN
Essential cellular Drug response Synergistic Non-synergistic
function D2 partner drugs partner drugs
D’2 -20
Cell response
Druggable target Active signaling
to etoposide

= y

| = rgy

= ed
Inactive signaling

)| rg
Gene in pathway

75

70

)
(n ict
76
D2 yne

D’ yne

d
re
Assess synergy

-s
2)

tp
|(D ted

|(D on
( D1, D2 ), ( D1, D’2 )

No
n
ic
1,

1,
ed
ed

ct
Pr

i
ed
Pr
0 = wild-type
1 = mutant Odds ratio
F G H FLT1 J 0 2 4 6 8 10 12 14

SRC PI3K DrugCell 51.6%

***
*** **
*** * Resistance to PI3K×ERK 76.9%
150 etoposide
*** 1 * PIK3R4
*** *** 0 = sensitive 17.9%
100 ERK
*** DUSP1 1 = resistant
Etoposide synergy score

*** 0 PI3K 18.8%

50 Logic Logic
PIN1 ERK OR AND
-1 PIN1 0.9%
Cell Fitness

0 RPS6KA6 Logic Logic

XOR Negation
RPS6KA6 4.6%
-50 -2
I Subsystem binary Prediction
approximation for of logic Percentage of observed Number DUSP1 2.1%
-100
-3 etoposide resistance circuit sensitive / resistant cells of cells
DrugCell
PI3K ERK PIK3R4 7.9%
-150 0 40 80
Subsystem pair
-4 0 0 Sensitive 30% 70% 783 4.4%
SRC
MK2206 PD325901 bortezomib NT Single subsystem
0 1 Sensitive 51% 49% 154
(AKTi) (MEKi) (proteasome TOP2 8.9% Single gene
1 0 Sensitive FLT1
inhibitor) MAP2K1 47% 53% 165
PIK3CA 1 1 Resistant 77% 23% 52
APC
fraction of resistant cells among all cells resistant sensitive

Figure 5. Discovery and Validation of Synergistic Mechanisms

(A) Parallel pathway theory of drug synergy, in which a pathway 2 is targeted by the mechanism of action (MoA) of drug A, and synergy is achieved by simul-
taneously targeting parallel pathway 1 with drug B.
(B) Logic learned by DrugCell for drug A, in which pathway 1 arises as a predicted mechanism of the VNN.
(C) Workflow demonstrating systematic design and assessment of pairwise combinations of drugs.
(D) Boxplots of DeepSynergy synergy scores for predicted drug combinations, predicted non-synergistic combinations, and random combinations. The boxes
represent the IQR bisected by the median, and whiskers represent the maximum and minimum range of the data that do not exceed 1.5 times the IQR.
***p < 0.0001.
(E) Representative subsystems used by DrugCell to simulate etoposide sensitivity (red nodes), along with a negative control branch (white node). RLIPP scores
are displayed inside each node. Subsystem names are abbreviated.
(F) Bee swarm plot of the Loewe synergy scores observed upon combination of etoposide with MK2206, PD325901, or bortezomib. Drug combinations were
chosen based on subsystems identified in (E). Red dotted line indicates the mean of all Loewe synergy scores in the dataset (Figure S6). ***p <0.0001. *** without
bars represent t test against the synergy score distribution of the full dataset (Figure S6), or bortezomib negative control, as indicated. Red points are cell lines for
which synergy is observed. Blue points are cell lines for which antagonism is observed.
(G) Boxplots of the relative cell growth of A549 cells following CRISPR/Cas9-mediated knockout of MAP2K1, PIK3CA, or APC (negative control) in combination
with TOP2 or a non-targeting control (NT). Data are reflective of two independent transductions. ***p <0.0001, *p <0.1, **p <0.01.
(H) Boolean logic circuit approximating how the mutational status of genes in the PI3K and ERK subsystems is translated to an etoposide response by DrugCell.
(I) Truth table showing translation of PI3K and ERK states to a binary drug response output. The percentage of observed sensitive versus resistant cells for each
state is shown. Dotted line indicates baseline percentage of etoposide-resistant samples among all cell lines.
(J) Odds ratios of etoposide response prediction for DrugCell, the ERK and PI3K logic functions from (H), and individual genes from (H). Percentages of cell lines
with an alteration to that biomarker are also shown. Odds ratios are against a background of cell lines that are wild type with respect to this circuit.

(Figure 5A). The branched architecture of the DrugCell model been tested across a panel of 39 cell lines (Figure 5C). We then
(Figure 1A) mirrors this parallel pathway structure, in that the bio- analyzed the top 5 and bottom 5 DrugCell subsystems for each
logical activity of a drug is learned by the drug embedding of these compounds to nominate synergistic and non-synergistic
branch, and the parallel pathways are learned by the genotype drug combinations. We found that drug combinations nominated
embedding branch (Figure 5B). Subsystems important for pre- by DrugCell were strongly and significantly enriched for synergis-
dicting a drug response may therefore represent synergistic tic cell killing outcomes, in contrast to combinations predicted to
drug combination opportunities. Exactly such parallelism was be non-synergistic or random combinations (Figure 5D).
used to nominate the combination treatments in the above anal- One such example was etoposide, a topoisomerase inhibitor
ysis (i.e., 2-DG as synergistic with paclitaxel). that leads to DNA damage (Table S5). Among the top etoposide
To further explore this concept, we used RLIPP scores to rank subsystems were the major kinase signaling pathways PI3K-
subsystems regulating sensitivity to 25 drugs in the DeepSynergy AKT (Regulation of PI3K activity, PI3K; Figure 5E) and RAF-
database (Preuer et al., 2018), in which all pairs of 25 drugs had MEK-ERK (Negative regulation of ERK1/ERK2 cascade, ERK;

678 Cancer Cell 38, 672–684, November 9, 2020


Article
Figure 5E). Indeed, etoposide synergized strongly with AKT and
MEK inhibition across the majority of cell lines tested in DeepSy-
nergy (Figure 5F). We further validated the observed synergy by
deleting the target of etoposide, TOP2, using CRISPR/Cas9
gene editing in A549 cells, either alone or in combination with
core genes in PI3K-AKT signaling (PIK3CA) or RAF-MEK-ERK
signaling (MAP2K1). We observed that deletion of TOP2 with
either PIK3CA or MAP2K1 demonstrated significant loss of cell
viability compared with single-gene knockout (Figure 5G).
APC, whose subsystem (b-catenin destruction complex) was
not identified by RLIPP (Table S5), did not show this same
pattern (Figure 5G). Similarly, etoposide did not synergize with
the proteasome inhibitor bortezomib (Figure 5F), consistent
with the proteasome subsystem not being identified by DrugCell
(Figure 5E).
Further inspection suggested that the relationship between
PI3K signaling, ERK signaling, and etoposide sensitivity
captured by DrugCell could be roughly approximated by a logic
function integrating the mutational status of six genes (Figures
5H and 5I; STAR Methods). Among these, FLT1 (Das et al.,
2005) and PIN1 (Mathur et al., 2011) had previously been shown
to regulate etoposide response, whereas DUSP1, PIK3R4, SRC,
and RPS6KA6 had not. Considered individually, any one of these
genes was mutated rarely in cancer cell lines, with limited power
to predict etoposide sensitivity versus resistance (mutation fre-
quencies 0.9%–8.9%; odds ratios <2; Figure 5J). Considered
as an integrated circuit, however, these gene mutations
converge on PI3K or ERK subsystems to create a powerful
network-based biomarker of drug response (odds ratio 7.8; Fig-
ure 5J). We also noted that these two pathways represent only a
portion of the full DrugCell model, which predicts etoposide
sensitivity with an odds ratio of 14.3.
DrugCell Improves Progression-Free Survival of
Patient-Derived Xenograft Models
We next wished to move beyond cell lines to predict and inter-
pret drug responses in the in vivo setting of patient-derived xeno-
graft models (PDX; Figure 6A, STAR Methods). To do so, we ac-
cessed the PDX Encyclopedia (Gao et al., 2015), in which 399
PDX tumors of varying tissue types had been screened against
a total of 40 different monotherapies and 27 combination thera-
pies. The genotypes of each PDX had also been established
(Gao et al., 2015), which were provided to DrugCell to make
response predictions to each monotherapy. We considered a
PDX tumor to be sensitive to a therapy (DrugCell (+)) if its pre-
dicted AUC was beneath the median predicted for all PDX-
drug pairs; otherwise this tumor was labeled as insensitive
(DrugCell ( )). DrugCell (+) tumors demonstrated significantly
longer progression-free survival (PFS) than DrugCell ( ) tumors
(2.19 versus 1.58 months, p = 9.4 3 10 10, log rank test). How-
ever, given the overall insensitivity of these PDX tumors to mono-
therapy, corresponding to the short observed PFS observed for
both DrugCell classes, we wished to evaluate how well DrugCell
is able to suggest effective drug combinations. We used RLIPP
scoring to rank subsystems by importance in mediating drug re-
sponses to six primary drugs, filtering this list to those that con-
tained secondary drug targets. The observed PFS of each of
these (primary, secondary) combinations was used to estimate
the prediction sensitivity and specificity along a receiver oper-
ll
ating characteristic curve (ROC; Figure 6A, STAR Methods).
We found that DrugCell was able to accurately identify subsys-
tems that correspond to effective drug combinations in PDX tu-
mors (auROC = 0.75; Figure 6B) with relatively few false positives
and negatives (Figure 6C).
For example, DrugCell analysis of BKM-120, a PI3K inhibitor,
identified Negative regulation of ERK1 + ERK2 cascade as an
important subsystem for BKM-120 response, suggesting a com-
bination of PI3K + MAPK pathway inhibitors (BKM-120 + encor-
afenib). This combination significantly increased PFS across the
PDX panel compared with monotherapy (Figure 6D). Similarly,
DrugCell identified DNA damage response, signal transduction
by p53 class mediator resulting in cell-cycle arrest as an impor-
tant subsystem for abraxane response, suggesting combination
chemotherapy with an agent inducing DNA damage and cell-cy-
cle arrest (abraxane + gemcitabine). This combination similarly
significantly improved PFS (Figure 6D). For the combinations
that were not prioritized by DrugCell (not in top 20% of subsys-
tems by RLIPP), these combinations indeed failed to significantly
improve PFS (Figures 6C and 6E). These results suggested that
DrugCell has utility in guiding design of combination therapies in
patient tumors.
DrugCell Predicts the Response of Estrogen Receptor-
Positive Metastatic Breast Cancer Patients to mTOR and
CDK4/6 Inhibitors
Last, we sought to evaluate whether DrugCell could be used
clinically to stratify cancer patients into responsive and non-
responsive patient populations. We obtained and analyzed
aggregated clinical trial data (Smyth et al., 2020) from 221 es-
trogen receptor (ER)-positive metastatic breast cancer patients
who had undergone multiple rounds of therapy, including an ER
antagonist (fulvestrant) in addition to treatment with an mTOR
inhibitor (everolimus) or CDK4/6 inhibitor (ribociclib). For this
analysis (STAR Methods), we predicted patient response to
either mTOR or CDK4/6 inhibition using our pre-trained Drug-
Cell model. We considered a patient to be DrugCell (+) if they
were predicted to be sensitive to either therapy and DrugCell
( ) if they were predicted to be insensitive to both therapies.
DrugCell (+) patients had significantly longer overall survival
than DrugCell ( ) patients (48.2 versus 33.6 months, p =
0.018; Figure 7A).
We next interrogated the mechanisms underlying the differ-
ential sensitivity between DrugCell (+) and DrugCell ( ) patients
by performing an RLIPP analysis for both the mTOR and the
CDK4/6 inhibitors. Notably, we found that both drug responses
were modulated by ER-related subsystems (Figures 7B and
7C), consistent with their use in ER-positive breast cancer
(Hare and Harvey, 2017; Pernas et al., 2018). We also found
that the major mechanisms of action of both drugs were among
the top pathways, with PI3K signaling being especially impor-
tant for response to mTOR inhibitors and CDK activity being
important for CDK4/6 inhibitor activity (Figures 7B and 7C).
Interestingly, TOR signaling was also identified for CDK4/6 in-
hibitors (Figure 7B), and CDK activity was identified for mTOR
inhibitors (Figure 7C), suggesting that these drugs could be
an effective combination therapy, a finding supported by recent
preclinical studies (Michaloglou et al., 2018; Occhipinti
et al., 2020).
Cancer Cell 38, 672–684, November 9, 2020 679
 ll
Article

6 small molecules
RLIPP

Evaluate ROC
combination analysis
efficacy
Identify subsystems mediating response
399 PDX
and design drug combinations
tumor samples
B D E
1.0 Predicted effective by DrugCell Predicted ineffective by DrugCell

0.8
True positive rate

100 100
Untreated

Surviving fraction (%)

0.6 Untreated
Surviving fraction (%)

Encorafenib Paclitaxel
80
80 NS Paclitaxel + LCL161
BKM120
0.4 *** BKM120 +
60 60
Encorafenib
0.2 40 40

auROC = 0.75 20 20
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0 0
0 5 10 15 20 25 30 35 0 2 4 6 8 10
False negative rate Progression free survival (months) Progression free survival (months)

C
100 100

Surviving fraction (%)

Surviving fraction (%)
predicted by DrugCell

Untreated Untreated
5 2
Yes
Drug combination

80 80
Abraxane Trastuzumab
60
*** Abraxane +
60 NS
INC280
Gemcitabine INC280 +
Trastuzumab
40 40

1 5 20 20
No

0 0
0 2 4 6 8 10 0 5 10 15 20 25

Yes No Progression free survival (months) Progression free survival (months)

Drug combination
effective in PDX tumor samples

Figure 6. Guiding Combination Therapy in Patient-Derived Xenograft Tumors

(A) Flowchart of analysis procedure.
(B) ROC curve of DrugCell performance in distinguishing effective from ineffective drug combinations.
(C) Error matrix for point indicated in (B) demonstrating best performance of DrugCell against the PDX dataset.
(D) Survival curves for drug combinations predicted to be effective by DrugCell (true positives) showing a significant improvement in progression-free survival.
(E) Survival curves for drug combinations predicted to be ineffective by DrugCell (true negatives) showing a lack of improvement in progression-free survival. p
values indicate significance by log rank test. ***p <0.0001, NS indicates not significant.

With respect to specific genetic alterations, we found that DISCUSSION

DrugCell (+) patients were much more likely to harbor AKT1
mutations than DrugCell ( ) patients (Figure 7D). In contrast, Here we have explored an interpretable deep learning model of
DrugCell ( ) patients had mutations in genes previously associ- the structure and function of a human cancer cell in response
ated with drug resistance, including ESR1 (Reinert et al., 2017), to treatment. This work advances predictive modeling toward a
RB1 (Condorelli et al., 2018), and PTEN (Costa et al., 2020) (Fig- systematic representation of the biological mechanisms underly-
ure 7D), suggesting that we had stratified patients based on a ing a drug response, a critical direction for precision medicine.
complex pattern of mutations leading to therapy resistance. Following a model prediction, access to a mechanistic interpre-
Strikingly, AKT1 mutation status alone was not predictive of tation engages the experimentalist or clinician in reasoning about
therapeutic response, with AKT1-mutant patients actually biological function. For example, analysis of DrugCell’s model of
trending toward shorter overall survival (35.8 versus etoposide identified a small set of subsystems important for the
43.1 months), although this difference was not statistically sig- cellular response and for which targeted drugs were available
nificant (Figure 7E). This analysis illustrates how DrugCell can (Figure 5). This analysis motivated us to perform subsequent ex-
be used to effectively guide clinical treatment decisions with periments to target both genetically and pharmacologically
significantly greater precision and insight than single-gene topoisomerase II with either MAPK or PI3K pathways; both of
marker studies. these combinations showed significant synergistic effects.

680 Cancer Cell 38, 672–684, November 9, 2020

 ll
Article

A B Figure 7. Guiding CDK4/6 and mTOR Inhibi-

Response
to estradiol
32.1 ... tor Therapy in ER-Positive Breast Cancer Pa-
100 Positive Patient tients
DrugCell (+) regulation 26.6
of PI3K
... response
(A–C) (A) Survival curves for DrugCell (+) and Drug-
DrugCell (–) to mTORi
signaling
Surviving Fraction (%)

80 Cell ( ) patients treated with CDK4/6 or mTOR in-

Positive
regulation 5.2 ... hibitors in any line of therapy. The p value indicates
of CDK
60 activity
significance by log rank test. (B, C) Important sub-
C systems used by DrugCell to simulate (B) mTOR or
40
Regulation of intracellular
estrogen receptor 17.2 ...
(C) CDK4/6 inhibitor sensitivity. Dotted line abbre-
14.6 months

signaling pathway
p = 0.018

20
Positive
regulation 15.3 ... Patient
response
viates parent subsystems at subsequent layers of
Median OS Median OS of CDK to CDK4/6i
33.6 months 48.2 months activity the hierarchy. RLIPP scores are displayed inside
n = 117 n = 104
0 each node.
Negative
regulation of 5.1 ...
0 20 40 60 80 100 120 TOR signaling (D) Scatterplot of the absolute (x axis) and per-
Overall survival (months) centage (y axis) difference in mutation frequencies
of genes between DrugCell (+) and DrugCell ( )
D Mutation E patients. Red points represent genes mutated more
Frequency frequently in DrugCell (+) patients. Blue points
0
Percent difference in mutation frequency

ERBB2 MAP2K4 represent genes mutated more frequently in Drug-

100 25 100
PRKDC PTEN CDH1 50 Cell ( ) patients. Point size is proportional to overall
75 AKT1E17K
RB1 AKT1 wild-type mutation frequency in the patient population.
Surviving Fraction (%)

80 80
ESR1
(E) Survival curves for AKT1-mutant and wild-type
AKT1
60 ATM 60 patients treated with CDK4/6 or mTOR inhibitors in
NTRK1
RUNX1
TP53 any line of therapy. The p value indicates signifi-
40 CBFB 40 cance by log rank test.
CREBBP
7.3 months

p = 0.877

20 PIK3CA 20
Median OS Median OS
DrugCell (+) 35.8 months 43.1 months
DrugCell (–) n = 68 n = 153
0 0
0 5 10 15 20 25 35 0 20 40 60 80 100 120
Absolute difference in mutation frequency Overall survival (months)

Such engagement of human reasoning and follow-up experi- combinatorial number of pairwise and higher order drug combi-
mentation helps greatly to increase accountability and trust in nations necessary for training.
the predictions of a machine learning model. In contrast, conven- If the favorable performance observed in PDX samples (Fig-
tional black-box predictive modeling yields only a model ure 6) and ER-positive breast cancer patients (Figure 7) con-
output—the drug response—without further information by tinues in further clinical studies, DrugCell and its successors
which to build trust in the process. have the potential to substantially expand the set of clinically
DrugCell is a flexible model that is amenable to both auto- meaningful mutations. DrugCell translates the mutational status
mated and semi-automated combinatorial drug design. First, of approximately 3,000 genes into treatment recommendations.
the importance of each cellular subsystem is scored by DrugCell Although we have not fully studied which of these genes are
during a response to monotherapy. These important subsystems absolutely required for DrugCell prediction accuracy, RLIPP
are then annotated with second points of intervention, such as analysis suggests that many of them are—1,467 of the 2,086
the PI3K or ERK pathway in the response to etoposide (Figures subsystems are assigned relatively high importance (RLIPP
5E–5G). To follow up on this analysis, drug combinations can >10) for at least one drug, collectively covering 2,855 genes.
be selected automatically based on the druggable targets pre- This breadth of information contrasts with the fewer genes
sent in top DrugCell subsystems. Alternatively, if DrugCell is be- included in current cancer mutation panels such as MSK-
ing used in a clinical context, its recommendations can be pro- IMPACT or FoundationOne CDx (468 and 324 genes, respec-
vided to physician-scientists (e.g., a molecular tumor board) tively), which were designed to be queried manually by a physi-
who consider the recommended combinations in light of other cian (Cheng et al., 2015; Harris, 2017). Moreover, since we
biological knowledge not explicitly used in modeling, such as po- currently do not understand the clinical implications of the major-
tential toxicities and specific information about the case. After ity of cancer mutations, there is little consensus on what genes
careful consideration of all relevant information, the ultimate should be included in these pan-cancer mutation panels
treatment decision remains in the hands of the physician and (Nguyen and Gocke, 2017) or on how physicians should act on
the patient. Such need for human accountability is not unique the results. An increase in the number of clinically meaningful
to drug response prediction but is a central tenet of high-stakes cancer mutations, facilitated by interpretable machine learning
applications of machine learning (Rudin, 2019). models such as DrugCell, could further motivate the case for
Notably, previous models trained on monotherapy responses complete genomic sequencing of cancer patients (Katsanis
(Ammad-ud-din et al., 2017; Cortés-Ciriano et al., 2016; Iorio and Katsanis, 2013; Kuenzi and Ideker, 2020).
et al., 2016; Zhang et al., 2015) have not attempted to suggest Future work may also elect to integrate mutations with addi-
combination therapies. Rather, drug combinations have been tional levels of molecular information such as epigenetic states,
predicted using models of synergy trained directly on data gene expression, or microenvironmental influences. This inte-
from pairwise drug treatments (Preuer et al., 2018). This brute gration could be accomplished by pre-processing multiple
force approach faces the challenge of scalability, given the layers of information to derive a profile of gene scores for each

Cancer Cell 38, 672–684, November 9, 2020 681

 ll
cell line or tumor, which would then be input to DrugCell. Extra
levels of information could also be integrated by adding new
visible or conventional neural network branches alongside exist-
ing ones. Alternatively, the effects of specific mutations on gene
functions could be incorporated by a metric such as the Com-
bined Annotation-Dependent Depletion score (Rentzsch et al.,
2019) or by including gene structural domains as an additional
layer of the hierarchy.
Another opportunity is to structure the DrugCell system hierar-
chy from ’omics data rather than literature curation (GO), as has
previously been done in budding yeast (Kramer et al., 2014; Ma
et al., 2018). A data-driven, rather than literature-curated, hierar-
chy has the potential to incorporate new gene-subsystem asso-
ciations as well as entirely new subsystems into the model. It also
has the potential to revise and tailor subsystem definitions in GO,
which are generic, to their particular contexts relevant to cancer.
For instance, we found that in its current form DrugCell contains
a number of subsystems that have misleading labels based on
GO naming conventions. For example, Labyrinthine develop-
ment was among the top subsystems for trametinib, which
was initially puzzling, but upon further inspection corresponds
to MAPK cascade genes with well-known involvement in cancer
proliferation (e.g., MAP2K1, MAPK1, GRB2, FGFR2). Incorpo-
rating data-driven hierarchies into DrugCell provides a route to
relabel such subsystems and revise their specific gene contents.
Finally, given that DrugCell inputs a full drug structure, it can
potentially be used to design compounds de novo. Leveraging
advancements in reinforcement learning for drug design (Zha-
voronkov et al., 2019), it may then be possible to design com-
pounds for maximal efficacy against any given genomic
background.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Data and Code Availability
B Materials Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Cell Culture and Reagents
d METHOD DETAILS
B Defining a Hierarchy of Genes and Cellular Sub-
systems
B Pharmacogenomics Data Processing and Morgan
Fingerprint Encoding
B Neural Network Configuration, Training and Evaluation
B Implementation of Alternative Models for Comparison
of Predictive Performance
B Ranking Important Subsystems in DrugCell
B Comparing Important DrugCell Subsystems to Predic-
tive Biomarkers Reported by Alternative Models
B Viability Assays
B Combinatorial CRISPR-Cas9 Gene Knockouts and
Systematic Evaluation
d QUANTIFICATION AND STATISTICAL ANALYSIS
682 Cancer Cell 38, 672–684, November 9, 2020
Article
B Assessing the Correspondence of Learned Subsystem
Embeddings to Measured Subsystem Activities
B Differential Expression Analysis
B Synergy Determinations
B Translation of Continuous Cell Response (AUC) to Bi-
nary Cell Response
B Identification of Boolean Logic Combinations
B PDX Tumor Analysis
B Breast Cancer Patient Analysis
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2020.09.014.
ACKNOWLEDGMENTS
We thank Dr. Gaudenz Danuser for his valuable insights into mechanisms of
taxol sensitivity. We gratefully acknowledge the support for this work provided
by grants from the National Institutes of Health to T.I. (GM103504, CA209891,
ES014811), J.F.K. (CA243885), and B.M.K. (CA212456).
AUTHOR CONTRIBUTIONS
B.M.K., J.P., S.H.F., J.M., J.F.K., and T.I. designed the study and developed
the conceptual ideas. B.M.K. collected all the input sources and additional
data. J.P., B.M.K., and J.M. implemented the main algorithm. B.M.K. and
J.P. implemented all other computational methods and analyses. B.M.K.,
S.H.F., J.L., and K.S. performed all experiments and B.M.K. analyzed the re-
sults. B.M.K., J.P., S.H.F., J.F.K., and T.I. wrote the manuscript with sugges-
tions from other authors.
DECLARATION OF INTERESTS
T.I. is a co-founder of Data4Cure, Inc., and has an equity interest. T.I. has an
equity interest in Ideaya BioSciences, Inc. The terms of this arrangement
have been reviewed and approved by the University of California San Diego
in accordance with its conflict of interest policies.
Received: April 8, 2020
Revised: August 7, 2020
Accepted: September 22, 2020
Published: October 22, 2020
REFERENCES
Ammad-ud-din, M., Khan, S.A., Wennerberg, K., and Aittokallio, T. (2017).
Systematic identification of feature combinations for predicting drug response
with Bayesian multi-view multi-task linear regression. Bioinformatics 33,
i359–i368.
Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H., Cherry, J.M.,
Davis, A.P., Dolinski, K., Dwight, S.S., Eppig, J.T., et al. (2000). Gene ontology:
tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet.
25, 25–29.
Bani, M.R., Nicoletti, M.I., Alkharouf, N.W., Ghilardi, C., Petersen, D., Erba, E.,
Sausville, E.A., Liu, E.T., and Giavazzi, R. (2004). Gene expression correlating
with response to paclitaxel in ovarian carcinoma xenografts. Mol. Cancer Ther.
3, 111–121.
Baptista, D., Ferreira, P.G., and Rocha, M. (2020). Deep learning for drug
response prediction in cancer. Brief. Bioinform bbz171, https://doi.org/10.
1093/bib/bbz171.
Barretina, J., Caponigro, G., Stransky, N., Venkatesan, K., Margolin, A.A., Kim,
S., Wilson, C.J., Lehár, J., Kryukov, G.V., Sonkin, D., et al. (2012). The Cancer
Cell Line Encyclopedia enables predictive modelling of anticancer drug sensi-
tivity. Nature 483, 603–607.

Article
Breinig, M., Klein, F.A., Huber, W., and Boutros, M. (2015). A chemical–genetic
interaction map of small molecules using high-throughput imaging in cancer
cells. Mol. Syst. Biol. 11, 846.
Chen, Y., Lun, A.T.L., and Smyth, G.K. (2016). From reads to genes to path-
ways: differential expression analysis of RNA-Seq experiments using
Rsubread and the edgeR quasi-likelihood pipeline. F1000Research 5, 1438.
Cheng, D.T., Mitchell, T.N., Zehir, A., Shah, R.H., Benayed, R., Syed, A.,
Chandramohan, R., Liu, Z.Y., Won, H.H., Scott, S.N., et al. (2015). Memorial
sloan kettering-integrated mutation profiling of actionable cancer targets
(MSK-IMPACT). J. Mol. Diagn. 17, 251–264.
Ching, T., Himmelstein, D.S., Beaulieu-Jones, B.K., Kalinin, A.A., Do, B.T.,
Way, G.P., Ferrero, E., Agapow, P.-M., Zietz, M., Hoffman, M.M., et al.
(2018). Opportunities and obstacles for deep learning in biology and medicine.
J. R. Soc. Interf. 15, 20170387.
Chiu, Y.-C., Chen, H.-I.H., Zhang, T., Zhang, S., Gorthi, A., Wang, L.-J., Huang,
Y., and Chen, Y. (2019). Predicting drug response of tumors from integrated
genomic profiles by deep neural networks. BMC Med. Genomics 12, 18.
Condorelli, R., Spring, L., O’Shaughnessy, J., Lacroix, L., Bailleux, C., Scott,
V., Dubois, J., Nagy, R.J., Lanman, R.B., Iafrate, A.J., et al. (2018).
Polyclonal RB1 mutations and acquired resistance to CDK 4/6 inhibitors in pa-
tients with metastatic breast cancer. Ann. Oncol. Off. J. Eur. Soc. Med. Oncol.
29, 640–645.
Copley, S.D. (2012). Moonlighting is mainstream: paradigm adjustment
required. BioEssays 34, 578–588.
Cortés-Ciriano, I., van Westen, G.J.P., Bouvier, G., Nilges, M., Overington,
J.P., Bender, A., and Malliavin, T.E. (2016). Improved large-scale prediction
of growth inhibition patterns using the NCI60 cancer cell line panel.
Bioinformatics 32, 85–95.
Costa, C., Wang, Y., Ly, A., Hosono, Y., Murchie, E., Walmsley, C.S., Huynh,
T., Healy, C., Peterson, R., Yanase, S., et al. (2020). PTEN loss mediates clin-
ical cross-resistance to CDK4/6 and PI3Ka inhibitors in breast cancer. Cancer
Discov. 10, 72–85.
Costello, J.C., Heiser, L.M., Georgii, E., Gönen, M., Menden, M.P., Wang, N.J.,
Bansal, M., Ammad-ud-din, M., Hintsanen, P., Khan, S.A., et al. (2014). A com-
munity effort to assess and improve drug sensitivity prediction algorithms. Nat.
Biotechnol. 32, 1202–1212.
Cover, T., and Hart, P. (1967). Nearest neighbor pattern classification. IEEE
Trans. Inf. Theor. 13, 21–27.
Das, B., Yeger, H., Tsuchida, R., Torkin, R., Gee, M.F.W., Thorner, P.S.,
Shibuya, M., Malkin, D., and Baruchel, S. (2005). A hypoxia-driven vascular
endothelial growth factor/flt1 autocrine loop interacts with hypoxia-inducible
factor-1a through mitogen-activated protein kinase/extracellular signal-regu-
lated kinase 1/2 pathway in neuroblastoma. Cancer Res. 65, 7267–7275.
Dincer, A.B., Celik, S., Hiranuma, N., and Lee, S.-I. (2018). DeepProfile: deep
learning of cancer molecular profiles for precision medicine. BioRxiv. https://
doi.org/10.1101/278739.
Eskiocak, B., McMillan, E.A., Mendiratta, S., Kollipara, R.K., Zhang, H.,
Humphries, C.G., Wang, C., Garcia-Rodriguez, J., Ding, M., Zaman, A., et al.
(2017). Biomarker accessible and chemically addressable mechanistic sub-
types of BRAF melanoma. Cancer Discov. 7, 832–851.
Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, M., Chou,
K., Cui, C., Corrado, G., Thrun, S., and Dean, J. (2019). A guide to deep
learning in healthcare. Nat. Med. 25, 24.
Gao, H., Korn, J.M., Ferretti, S., Monahan, J.E., Wang, Y., Singh, M., Zhang,
C., Schnell, C., Yang, G., Zhang, Y., et al. (2015). High-throughput screening
using patient-derived tumor xenografts to predict clinical trial drug response.
Nat. Med. 21, 1318–1325.
Garnett, M.J., Edelman, E.J., Heidorn, S.J., Greenman, C.D., Dastur, A., Lau,
K.W., Greninger, P., Thompson, I.R., Luo, X., Soares, J., et al. (2012).
Systematic identification of genomic markers of drug sensitivity in cancer cells.
Nature 483, 570–575.
Hare, S.H., and Harvey, A.J. (2017). mTOR function and therapeutic targeting
in breast cancer. Am. J. Cancer Res. 7, 383–404.
ll
Harris, J. (2017). FDA Approves FoundationOne CDx, CMS Agrees to Cover
(OncLive Novemb).
Hatzis, C., Bedard, P.L., Birkbak, N.J., Beck, A.H., Aerts, H.J.W.L., Stern, D.F.,
Shi, L., Clarke, R., Quackenbush, J., and Haibe-Kains, B. (2014). Enhancing
reproducibility in cancer drug screening: how do We move forward? Cancer
Res. 74, 4016–4023.
Huang, D.W., Sherman, B.T., Tan, Q., Kir, J., Liu, D., Bryant, D., Guo, Y.,
Stephens, R., Baseler, M.W., Lane, H.C., et al. (2007). DAVID Bioinformatics
Resources: expanded annotation database and novel algorithms to better
extract biology from large gene lists. Nucleic Acids Res. 35, W169–W175.
Hwang, S. (2012). Comparison and evaluation of pathway-level aggregation
methods of gene expression data. BMC Genomics 13, S26.
Iorio, F., Knijnenburg, T.A., Vis, D.J., Bignell, G.R., Menden, M.P., Schubert,
M., Aben, N., Gonçalves, E., Barthorpe, S., Lightfoot, H., et al. (2016). A land-
scape of pharmacogenomic interactions in cancer. Cell 166, 740–754.
Kang, H.C., Kim, I.-J., Park, J.-H., Shin, Y., Ku, J.-L., Jung, M.S., Yoo, B.C.,
Kim, H.K., and Park, J.-G. (2004). Identification of genes with differential
expression in acquired drug-resistant Gastric cancer cells using high-
Density Oligonucleotide microarrays. Clin. Cancer Res. 10, 272–284.
Karnaugh, M. (1953). The map method for synthesis of combinational logic cir-
cuits. Trans. Am. Inst. Electr. Eng. Part Commun. Electron. 72, 593–599.
Katsanis, S.H., and Katsanis, N. (2013). Molecular genetic testing and the
future of clinical genomics. Nat. Rev. Genet. 14, 415–426.
Kramer, M., Dutkowski, J., Yu, M., Bafna, V., and Ideker, T. (2014). Inferring
gene ontologies from pairwise similarity data. Bioinformatics 30, i34–i42.
Kuenzi, B.M., and Ideker, T. (2020). A census of pathway maps in cancer sys-
tems biology. Nat. Rev. Cancer 20, 1–14.
Kuenzi, B.M., Rix, L.L.R., Kinose, F., Kroeger, J.L., Lancet, J.E., Padron, E.,
and Rix, U. (2019). Off-target based drug repurposing opportunities for tivan-
tinib in acute myeloid leukemia. Sci. Rep. 9, 606.
Li, J., Zhao, W., Akbani, R., Liu, W., Ju, Z., Ling, S., Vellano, C.P., Roebuck, P.,
Yu, Q., Eterovic, A.K., et al. (2017). Characterization of human cancer cell lines
by reverse-phase protein arrays. Cancer Cell 31, 225–239.
Liu, T., Sun, H., Zhu, D., Dong, X., Liu, F., Liang, X., Chen, C., Shao, B., Wang,
M., Wang, Y., et al. (2017). TRA2A promoted paclitaxel resistance and tumor
progression in triple-negative breast cancers via regulating alternative
splicing. Mol. Cancer Ther. 16, 1377–1388.
Loewe, S. (1953). The problem of synergism and antagonism of combined
drugs. Arzneimittelforschung 3, 285–290.
Ma, J., Yu, M.K., Fong, S., Ono, K., Sage, E., Demchak, B., Sharan, R., and
Ideker, T. (2018). Using deep learning to model the hierarchical structure
and function of a cell. Nat. Methods. 15, 290–298.
Ma, Y., Wang, L., Neitzel, L.R., Loganathan, S.N., Tang, N., Qin, L., Crispi, E.E.,
Guo, Y., Knapp, S., Beauchamp, R.D., et al. (2017). The MAPK pathway regu-
lates intrinsic resistance to BET inhibitors in colorectal cancer. Clin. Cancer
Res. 23, 2027–2037.
Mathur, R., Chandna, S., N Kapoor, P., and S Dwarakanath, B. (2011). Peptidyl
prolyl isomerase, Pin1 is a potential target for enhancing the therapeutic effi-
cacy of etoposide. Curr. Cancer Drug Targets 11, 380–392.
Menden, M.P., Iorio, F., Garnett, M., McDermott, U., Benes, C.H., Ballester,
P.J., and Saez-Rodriguez, J. (2013). Machine learning prediction of cancer
cell sensitivity to drugs based on genomic and chemical properties. PLoS
One 8, https://doi.org/10.1371/journal.pone.0061318.
Michaloglou, C., Crafter, C., Siersbaek, R., Delpuech, O., Curwen, J.O.,
Carnevalli, L.S., Staniszewska, A.D., Polanska, U.M., Cheraghchi-Bashi, A.,
Lawson, M., et al. (2018). Combined inhibition of mTOR and CDK4/6 is
required for optimal blockade of E2F function and long-term growth inhibition
in estrogen receptor–positive breast cancer. Mol. Cancer Ther. 17, 908–920.
Moos, P.J., and Fitzpatrick, F.A. (1998). Taxane-mediated gene induction is in-
dependent of microtubule stabilization: induction of transcription regulators
and enzymes that modulate inflammation and apoptosis. Proc. Natl. Acad.
Sci. U S A 95, 3896–3901.
Cancer Cell 38, 672–684, November 9, 2020 683
 ll
Murdoch, W.J., Singh, C., Kumbier, K., Abbasi-Asl, R., and Yu, B. (2019).
Definitions, methods, and applications in interpretable machine learning.
Proc. Natl. Acad. Sci. U S A 116, 22071–22080.
Nguyen, D., and Gocke, C.D. (2017). Managing the genomic revolution in can-
cer diagnostics. Virchows Arch. 471, 175–194.
Nutt, C.L., Noble, M., Chambers, A.F., and Cairncross, J.G. (2000). Differential
expression of drug resistance genes and chemosensitivity in glial cell lineages
correlate with differential response of oligodendrogliomas and astrocytomas
to chemotherapy. Cancer Res. 60, 4812–4818.
Occhipinti, G., Romagnoli, E., Santoni, M., Cimadamore, A., Sorgentoni, G.,
Cecati, M., Giulietti, M., Battelli, N., Maccioni, A., Storti, N., et al. (2020).
Sequential or concomitant inhibition of cyclin-dependent kinase 4/6 before
mTOR pathway in hormone-positive HER2 negative breast cancer: biological
insights and clinical implications. Front. Genet. 11, 349.
Park, J.S., Burckhardt, C.J., Lazcano, R., Solis, L.M., Isogai, T., Li, L., Chen,
C.S., Gao, B., Minna, J.D., Bachoo, R., et al. (2020). Mechanical regulation
of glycolysis via cytoskeleton architecture. Nature 578, 621–626.
Pernas, S., Tolaney, S.M., Winer, E.P., and Goel, S. (2018). CDK4/6 inhibition in
breast cancer: current practice and future directions. Ther. Adv. Med.
Oncol. 10.
Potts, M.B., McMillan, E.A., Rosales, T.I., Kim, H.S., Ou, Y.-H., Toombs, J.E.,
Brekken, R.A., Minden, M.D., MacMillan, J.B., and White, M.A. (2015). Mode of
action and pharmacogenomic biomarkers for exceptional responders to di-
demnin B. Nat. Chem. Biol. 11, 401–408.
Pozdeyev, N., Yoo, M., Mackie, R., Schweppe, R.E., Tan, A.C., and Haugen,
B.R. (2016). Integrating heterogeneous drug sensitivity data from cancer phar-
macogenomic studies. Oncotarget 7, 51619–51625.
Pratt, D., Chen, J., Welker, D., Rivas, R., Pillich, R., Rynkov, V., Ono, K., Miello,
C., Hicks, L., Szalma, S., et al. (2015). NDEx, the network data Exchange. Cell
Syst. 1, 302–305.
Preuer, K., Lewis, R.P.I., Hochreiter, S., Bender, A., Bulusu, K.C., Klambauer,
G., and Wren, J. (2018). DeepSynergy: predicting anti-cancer drug synergy
with Deep Learning. Bioinformatics 34, 1538–1546.
Rajkomar, A., Dean, J., and Kohane, I. (2019). Machine learning in medicine.
N. Engl. J. Med. 380, 1347–1358.
Rampá
sek, L., Hidru, D., Smirnov, P., Haibe-Kains, B., and Goldenberg, A.
(2019). Dr.VAE: improving drug response prediction via modeling of drug
perturbation effects. Bioinformatics 35, 3743–3751.
Reinert, T., Saad, E.D., Barrios, C.H., and Bines, J. (2017). Clinical implications
of ESR1 mutations in hormone receptor-positive advanced breast cancer.
Front. Oncol. 7, 26.
Rentzsch, P., Witten, D., Cooper, G.M., Shendure, J., and Kircher, M. (2019).
CADD: predicting the deleteriousness of variants throughout the human
genome. Nucleic Acids Res. 47, D886–D894.
Ritchie, M.E., Phipson, B., Wu, D., Hu, Y., Law, C.W., Shi, W., and Smyth, G.K.
(2015). Limma powers differential expression analyses for RNA-sequencing
and microarray studies. Nucleic Acids Res. 43, e47.
Robinson, M.D., and Oshlack, A. (2010). A scaling normalization method for
differential expression analysis of RNA-seq data. Genome Biol. 11, R25.
Rogers, D., and Hahn, M. (2010). Extended-connectivity fingerprints. J. Chem.
Inf. Model. 50, 742–754.
Rudin, C. (2019). Stop explaining black box machine learning models for high
stakes decisions and use interpretable models instead. Nat. Mach. Intell. 1,
206–215.
Sakellaropoulos, T., Vougas, K., Narang, S., Koinis, F., Kotsinas, A., Polyzos,
A., Moss, T.J., Piha-Paul, S., Zhou, H., Kardala, E., et al. (2019). A deep
learning framework for predicting response to therapy in cancer. Cell Rep.
29, 3367–3373.e4.
Sanjana, N.E., Shalem, O., and Zhang, F. (2014). Improved vectors and
genome-wide libraries for CRISPR screening. Nat. Methods 11, 783–784.
Seashore-Ludlow, B., Rees, M.G., Cheah, J.H., Cokol, M., Price, E.V., Coletti,
M.E., Jones, V., Bodycombe, N.E., Soule, C.K., Gould, J., et al. (2015).
Harnessing connectivity in a large-scale small-molecule sensitivity dataset.
Cancer Discov. 5, 1210–1223.
684 Cancer Cell 38, 672–684, November 9, 2020
Article
Shen, J.P., Zhao, D., Sasik, R., Luebeck, J., Birmingham, A., Bojorquez-
Gomez, A., Licon, K., Klepper, K., Pekin, D., Beckett, A.N., et al. (2017).
Combinatorial CRISPR-Cas9 screens for de novo mapping of genetic interac-
tions. Nat. Methods 14, 573–576.
Shimamura, T., Chen, Z., Soucheray, M., Carretero, J., Kikuchi, E., Tchaicha,
J.H., Gao, Y., Cheng, K.A., Cohoon, T.J., Qi, J., et al. (2013). Efficacy of BET
bromodomain inhibition in Kras-mutant non-small cell lung cancer. Clin.
Cancer Res. Off. J. Am. Assoc. Cancer Res. 19, 4325.
Smyth, L.M., Zhou, Q., Nguyen, B., Yu, C., Lepisto, E.M., Arnedos, M., Hasset,
M.J., Lenoue-Newton, M.L., Blauvelt, N., Dogan, S., et al. (2020).
Characteristics and outcome of AKT1E17K-mutant breast cancer defined
through AACR Project GENIE, a clinicogenomic registry. Cancer Discov. 10,
526–535.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L.,
Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., et al.
(2005). Gene set enrichment analysis: a knowledge-based approach for inter-
preting genome-wide expression profiles. Proc. Natl. Acad. Sci. U S A 102,
15545–15550.
Suzuki, S., Horinouchi, T., and Furusawa, C. (2014). Prediction of antibiotic
resistance by gene expression profiles. Nat. Commun. 5, 1–12.
Topol, E.J. (2019). High-performance medicine: the convergence of human
and artificial intelligence. Nat. Med. 25, 44–56.
Turner, R.M., Park, B.K., and Pirmohamed, M. (2015). Parsing interindividual
drug variability: an emerging role for systems pharmacology. Wiley
Interdiscip. Rev. Syst. Biol. Med. 7, 221–241.
Wainberg, M., Merico, D., Delong, A., and Frey, B.J. (2018). Deep learning in
biomedicine. Nat. Biotechnol. 36, 829–838.
Wang, Y., Zhang, S., Li, F., Zhou, Y., Zhang, Y., Wang, Z., Zhang, R., Zhu, J.,
Ren, Y., Tan, Y., et al. (2020). Therapeutic target database 2020: enriched
resource for facilitating research and early development of targeted therapeu-
tics. Nucleic Acids Res. 48, D1031–D1041.
Wong, C.H., Siah, K.W., and Lo, A.W. (2019). Estimation of clinical trial success
rates and related parameters. Biostatistics 20, 273–286.
Yang, J.H., Wright, S.N., Hamblin, M., McCloskey, D., Alcantar, M.A.,
Schru€bbers, L., Lopatkin, A.J., Satish, S., Nili, A., Palsson, B.O., et al. (2019).
A white-box machine learning approach for revealing antibiotic mechanisms
of action. Cell 177, 1649–1661.e9.
Yang, L., Ainali, C., Tsoka, S., and Papageorgiou, L.G. (2014). Pathway activity
inference for multiclass disease classification through a mathematical pro-
gramming optimisation framework. BMC Bioinformatics 15, 390.
Yang, W., Soares, J., Greninger, P., Edelman, E.J., Lightfoot, H., Forbes, S.,
Bindal, N., Beare, D., Smith, J.A., Thompson, I.R., et al. (2013). Genomics of
Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker dis-
covery in cancer cells. Nucleic Acids Res. 41, D955–D961.
Yeh, P.J., Hegreness, M.J., Aiden, A.P., and Kishony, R. (2009). Drug interac-
tions and the evolution of antibiotic resistance. Nat. Rev. Microbiol. 7,
460–466.
Yin, Y., Sun, M., Zhan, X., Wu, C., Geng, P., Sun, X., Wu, Y., Zhang, S., Qin, J.,
Zhuang, Z., et al. (2019). EGFR signaling confers resistance to BET inhibition in
hepatocellular carcinoma through stabilizing oncogenic MYC. J. Exp. Clin.
Cancer Res. 38, 83.
Yu, M.K., Ma, J., Fisher, J., Kreisberg, J.F., Raphael, B.J., and Ideker, T. (2018).
Visible machine learning for biomedicine. Cell 173, 1562–1565.
Zeng, X., Zhu, S., Liu, X., Zhou, Y., Nussinov, R., and Cheng, F. (2019).
deepDR: a network-based deep learning approach to in silico drug reposition-
ing. Bioinformatics 35, 5191–5198.
Zhang, N., Wang, H., Fang, Y., Wang, J., Zheng, X., and Liu, X.S. (2015).
Predicting anticancer drug responses using a dual-layer integrated cell line-
drug network model. PLoS Comput. Biol. 11, e1004498.
Zhavoronkov, A., Ivanenkov, Y.A., Aliper, A., Veselov, M.S., Aladinskiy, V.A.,
Aladinskaya, A.V., Terentiev, V.A., Polykovskiy, D.A., Kuznetsov, M.D.,
Asadulaev, A., et al. (2019). Deep learning enables rapid identification of potent
DDR1 kinase inhibitors. Nat. Biotechnol. 37, 1038–1040.
See what’s coming.
Know what’s working.
Cell Reports Medicine—a new broad-scope, open access journal from Cell Press—publishes
original, thought-provoking research from exciting translational concepts in human biology, health,
and disease to all phases of clinical work.

We give authors the same level of editorial excellence and support that they expect from flagship
journals such as Cell and foster deeper engagement between scientists, clinicians, and industry
professionals, increasing the reach and visibility of research that can positively impact human health.

Help close the loop. Submit your paper today.

cell.com/cell-reports-medicine
 ll

Article
Conserved Interferon-g Signaling
Drives Clinical Response to Immune
Checkpoint Blockade Therapy in Melanoma
Catherine S. Grasso,1,2,* Jennifer Tsoi,1 Mykola Onyshchenko,1,17 Gabriel Abril-Rodriguez,1 Petra Ross-Macdonald,3
Megan Wind-Rotolo,3 Ameya Champhekar,1 Egmidio Medina,1 Davis Y. Torrejon,1 Daniel Sanghoon Shin,1 Phuong Tran,1
Yeon Joo Kim,1 Cristina Puig-Saus,1,5 Katie Campbell,1 Agustin Vega-Crespo,1 Michael Quist,1 Christophe Martignier,4
Jason J. Luke,6,18 Jedd D. Wolchok,5,7 Douglas B. Johnson,8 Bartosz Chmielowski,1 F. Stephen Hodi,5,9
Shailender Bhatia,10 William Sharfman,11 Walter J. Urba,12 Craig L. Slingluff, Jr.,13 Adi Diab,14 John B.A.G. Haanen,15
Salvador Martin Algarra,16 Drew M. Pardoll,11 Valsamo Anagnostou,11 Suzanne L. Topalian,11 Victor E. Velculescu,11
Daniel E. Speiser,4 Anusha Kalbasi,1 and Antoni Ribas1,5,19,*
1Jonsson Comprehensive Cancer Center at the University of California, Los Angeles (UCLA), Los Angeles, CA, USA
2Cedars-Sinai Medical Center, Los Angeles, CA, USA
3Translational Bioinformatics, Bristol-Myers Squibb, Hopewell, NJ, USA
4University of Lausanne, Lausanne, Switzerland
5Parker Institute for Cancer Immunotherapy, San Francisco, CA, USA
6University of Chicago, Chicago, IL, USA
7Memorial Sloan Kettering Cancer Center, New York, NY, USA
8Vanderbilt-Ingram Cancer Center, Nashville, TN, USA
9Dana Farber Cancer Institute, Boston, MA, USA
10University of Washington, Seattle, WA, USA
11Bloomberg-Kimmel Institute for Cancer Immunotherapy and Sidney Kimmel Comprehensive Cancer Center, Johns Hopkins University

School of Medicine, Baltimore, MD, USA

12Earle A. Chiles Research Institute, Providence Cancer Institute, Portland, OR, USA
13University of Virginia Health System, Charlottesville, VA, USA
14The University of Texas MD Anderson Cancer Center, Houston, TX, USA
15The Netherlands Cancer Institute, Amsterdam, the Netherlands
16Clı́nica Universitaria de Navarra, Pamplona, Spain
17Present address: LA Biomed, Harbor-UCLA Medical Center, Torrance, CA, USA
18Present address: University of Pittsburgh, Pittsburgh, PA, USA
19Lead Contact

*Correspondence: catherine.grasso@cshs.org (C.S.G.), aribas@mednet.ucla.edu (A.R.)

https://doi.org/10.1016/j.ccell.2020.08.005

SUMMARY

We analyze the transcriptome of baseline and on-therapy tumor biopsies from 101 patients with advanced
melanoma treated with nivolumab (anti-PD-1) alone or combined with ipilimumab (anti-CTLA-4). We find
that T cell infiltration and interferon-g (IFN-g) signaling signatures correspond most highly with clinical
response to therapy, with a reciprocal decrease in cell-cycle and WNT signaling pathways in responding bi-
opsies. We model the interaction in 58 human cell lines, where IFN-g in vitro exposure leads to a conserved
transcriptome response unless cells have IFN-g receptor alterations. This conserved IFN-g transcriptome
response in melanoma cells serves to amplify the antitumor immune response. Therefore, the magnitude
of the antitumor T cell response and the corresponding downstream IFN-g signaling are the main drivers
of clinical response or resistance to immune checkpoint blockade therapy.

INTRODUCTION
Immune checkpoint blockade (ICB) therapy with antibodies tar-
geting the cytotoxic T lymphocyte-associated protein 4 (CTLA-
4) or the programmed cell death-1 receptor (PD-1) blocks two
main negative regulators of antitumor immune responses
(Curran et al., 2010; Sharma and Allison, 2015a; Wei et al.,
2017) and induces durable responses in a subset of patients
with melanoma and other cancers (Ribas and Wolchok, 2018;
Sharma and Allison, 2015b). Pathological studies performed
in tumor biopsies from treated patients support that clinical re-
sponses induced by the use of ICB are mediated by tumor-infil-
trating T cells that have been re-activated by inhibiting these
immune checkpoints (Sharma et al., 2019; Tumeh et al.,
2014). Transcriptomic and genomic sequencing of baseline
and on-therapy biopsies from patients treated with ICB allow

500 Cancer Cell 38, 500–515, October 12, 2020 ª 2020 Elsevier Inc.

Article
for a comprehensive analysis of mechanisms underlying tumor
response and resistance. As the mechanism of action is based
on the interaction between immune effector cells with their can-
cer cell targets, these studies have to focus not only on the ge-
netic alterations and gene expression profiles of cancer cells
(Hugo et al., 2016; Liu et al., 2019; Riaz et al., 2017; Rodig
et al., 2018), but they also need to analyze the composition
of the immune infiltrate and its expression of immune activating
gene programs (Auslander et al., 2018; Ayers et al., 2017; Cab-
rita et al., 2020; Chen et al., 2016; Cristescu et al., 2018; Feh-
renbacher et al., 2016; Gide et al., 2019; Helmink et al., 2020;
Jerby-Arnon et al., 2018; Jiang et al., 2018; Liu et al., 2019; Pe-
titprez et al., 2020; Rodig et al., 2018; Roh et al., 2017; Sade-
Feldman et al., 2018). A major focus has been on the study
of biopsies from patients with advanced melanoma treated
with anti-PD-1 antibodies administered alone, in sequence or
in combination with anti-CTLA-4 antibodies. A study combining
immunohistochemistry analyses with RNA sequencing (RNA-
seq) in biopsies from patients treated sequentially with the
anti-CTLA-4 antibody ipilimumab before or after the anti-PD-1
antibody nivolumab concluded that primary response to anti-
CTLA-4 required high levels of major histocompatibility com-
plex (MHC) class I expression by cancer cells at baseline, while
response to anti-PD-1 was more associated with a pre-existing
interferon-g transcriptome signature (Rodig et al., 2018).
Another study combining immunohistochemistry analyses
with RNA-seq in tumor biopsies from patients treated with
anti-PD-1 monotherapy or combined with anti-CTLA-4
confirmed the association of response to PD-1 blockade ther-
apy with baseline evidence of activated T cells using morpho-
logical and transcriptome signatures, in particular of an effector
memory T cell phenotype (Gide et al., 2019). A third study with
a large cohort of baseline biopsies from patients treated with
anti-PD-1 therapy, with or without prior anti-CTLA-4 therapy,
revealed that response to PD-1 blockade was associated
with increased MHC class I and II expression (Liu et al.,
2019). In this study, whole-exome sequencing revealed occa-
sional genetic alterations in antigen presentation machinery
genes in biopsies of patients who did not respond to therapy
(Liu et al., 2019).
CheckMate 038 is a prospective, multicenter, international,
multi-cohort clinical trial designed to collect tumor biopsies
from patients with metastatic melanoma treated with the anti-
PD-1 antibody nivolumab as front-line therapy or after progress-
ing on-therapy with the anti-CTLA-4 antibody ipilimumab, or
receiving the combination of both antibodies. Biopsies of 68 pa-
tients receiving nivolumab monotherapy in part 1 of this study
have been reported previously (Riaz et al., 2017). These samples
were analyzed by whole-exome, transcriptome, and T cell re-
ceptor (TCR) sequencing. Data revealed increases in distinct im-
mune cell subsets and upregulation of immune activation gene
programs that was more pronounced in patients with a clinical
response to therapy (Riaz et al., 2017). In this study, we provide
information on the transcriptome analysis of tumor biopsies from
the complete set of 101 patients treated with single-agent nivo-
lumab or with the combination of nivolumab and ipilimumab,
which is correlated with the in vitro analysis of how a panel of
melanoma cell lines change gene expression upon exposure to
interferon-g.
ll
RESULTS
Patient Characteristics and Response to ICB Therapy
We analyzed tumor biopsies obtained from patients treated with
nivolumab or nivolumab plus ipilimumab within the CheckMate
038 study (NCT01621490), which required baseline and on-ther-
apy biopsies from all patients. The clinical trial had several parts
(Figure S1). Part 1 included two cohorts that received nivolumab
monotherapy: patients whose melanomas had previously pro-
gressed on anti-CTLA-4 monotherapy, and those who were
naive to prior anti-CTLA-4. In part 2, patients who were anti-
CTLA-4-naı̈ve received combination therapy with nivolumab
and ipilimumab. Part 3 was also restricted to patients who
were anti-CTLA-4 naı̈ve, and randomized patients 1:2 to receive
either nivolumab monotherapy or combination therapy with ipili-
mumab. Part 4 was similar to part 3 in being restricted to patients
naive to anti-CTLA-4, but included patients with active brain me-
tastases who received either nivolumab monotherapy or combi-
nation therapy with ipilimumab. Results of analysis of biopsies of
patients from part 1 have been reported previously by Riaz et al.
(2017), while analyses from parts 2–4 have not been reported
previously. Biopsies were planned at baseline and during treat-
ment (week 4 for part 1, and week 2 or 4 for parts 2–4) and
were processed centrally to obtain RNA for sequencing
analyses.
Among the 170 patients enrolled in the clinical trial (Figure 1A),
106 patients received nivolumab alone and 62 received nivolu-
mab and ipilimumab combined therapy. RNA was isolated
from 101 baseline and 99 on-therapy biopsies (Figure 1A).
Several biopsies were not included in this analysis for different
reasons, including specimens not meeting quality standards, pa-
tients who received a treatment that was not assigned, not hav-
ing adequate tumor response assessment, and a diagnosis of
primary choroidal melanoma (eight patients) due to the distinct
biology of this uncommon melanoma subtype (Figure 1A). At
the end, there were 84 baseline and 85 on-therapy biopsies (68
of which were paired), that contributed to transcriptome ana-
lyses for RNA-seq (Figure 1A). Table 1 shows the clinical charac-
teristics of the analysis cohort, including 101 total patients: me-
dian age was 56 years (range 22–89 years). The majority (71%) of
patients had cutaneous melanoma, with 11% mucosal and 4%
acral melanoma. The majority of patients had stage IV disease
(93%), and the adverse prognosis factor of increased lactate de-
hydrogenase was present in 28% of patients.
CheckMate 038 was not designed to assess comparisons be-
tween study groups as these were mostly non-randomized co-
horts (Figure S1). Therefore, the account of tumor responses
and time-to-event outcomes is descriptive and used for the inter-
pretation of the tumor biopsy analyses. Overall, there were more
patients with an objective response (complete response or par-
tial response [CRPR]) in the group receiving the combination of
nivolumab and ipilimumab than among patients treated with ni-
volumab monotherapy, whether or not naive to anti-CTLA-4 ther-
apy (Figure 1B). Analyses of overall survival (OS) and progres-
sion-free survival (PFS) suggested that survival outcomes were
similar among patients receiving nivolumab monotherapy,
whether or not they had received prior anti-CTLA-4; OS and
PFS among patients receiving combination nivolumab plus ipili-
mumab therapy appeared to exceed nivolumab monotherapy
Cancer Cell 38, 500–515, October 12, 2020 501
 ll
Article

A RNAseq analysis B

170 Patients Treated

• 106 Monotherapy
• 62 Combination therapy
Treatment Response
• 2 Unscheduled: CA209038-4-30003

7 7
Nivo
(Ipi- 10 13
99 experi CRPR
101 enced)
On-
Baseline 13 22
treatment
13
Nivo 10
Reasons for Sample Reasons for Sample 8
(Ipi- SD
Removal Removal naïve) 8
• 1 Biopsied at 24h • 1 Unscheduled 3
• 1 Unscheduled 16
• 0 Insufficient coverage
• 0 Insufficient coverage • 4 QC failure (Low
• 5 QC failure (No Pearson Correlation/ 13
Melanin expression/ Male with XIST
Low Pearson expression) 22 PD
Correlation/ Male with Nivo/
• 2 NE by RECIST Ipi 16
XIST expression) • 7 Ocular melanoma
• 2 NE by RECIST 3
• 8 Ocular melanoma 9 9

84 for analysis 85 for analysis

• 55 Mono • 57 Mono
• 29 Combo • 27 Combo

Figure 1. Outline of Sample Collection, Patient Treatments, and Response to Therapy

(A) Consort diagram describing data generation and selection of final transcriptome cohort.
(B) Alluvial plot showing the number of patients of each treatment subtype that resulted in each response. The numbers in the treatment subtype box represent the
number of patients of that treatment type who had that response (red, PD; green, SD; blue, CRPR).
See also Figures S1 and S2.

outcomes (Figure S2A). When analyzing OS and PFS according
to tumor response assessment, patients with CRPR unsurpris-
ingly had prolonged survival compared with those with stable
disease (SD) or progressive disease (PD) (Figure S2B). Based
on these observations, for the RNA-seq analyses we compared
patients receiving combination therapy to those receiving nivolu-
mab monotherapy (merging the groups that had or had not pre-
viously received ipilimumab), and clinical response was defined
as CRPR compared with SD or PD.
The Dominant Signal Associated with Response or
Resistance to ICB Therapy in Patient Biopsies Is
Mediated by T Cells
Consistently, it has been shown that high levels of CD8 T cell
infiltration of tumors at baseline and on-treatment are predictive
of clinical response to ICB in patients with advanced melanoma
(Chen et al., 2016; Gide et al., 2019; Tumeh et al., 2014). We
used the method denominated microenvironment cell popula-
tions-counter (MCP-counter) (Becht et al., 2016) for RNA-seq
data deconvolution to define immune cell types. We applied
an optimal pooled t test, since only a subset of the samples
were paired. Using these approaches, we documented that bi-
opsies of patients with a response to therapy had higher base-
line levels of T cells than those with PD (Figure 2A). They also
had significantly higher baseline levels of B lineage cells,
myeloid dendritic cells, and natural killer cells (Figure 2A). The
change from baseline to on-therapy was greater in biopsies of
patients with a response to therapy, and there were increases
in T cell infiltrates in biopsies of patients whether or not they
had an objective clinical response to therapy, and regardless
of receiving combination therapy or nivolumab alone (Figures
2A and 2B).
With the goal to define the main drivers explaining the overall
biopsy data, we then analyzed the whole-transcriptome RNA-
seq dataset using principal-component analysis. Analysis of
the first two principal components showed that the signature
for T cell score derived from MCP-counter RNA-seq data decon-
volution was the main factor organizing the data, and it segre-
gated with response to therapy (with the exception of 19 outlier
samples, 12 of which were matched from the same patients
with baseline and on-therapy biopsies, Figure 2C). Of note, in
previous work we had validated the MCP-counter T cell score
with pathological analysis of immune infiltrates in corresponding
patient biopsies (Grasso et al., 2018). In the CheckMate 038 bi-
opsy dataset, the best organization of the samples was accord-
ing to the degree of T cell infiltration regardless of whether this
was observed in baseline or on-therapy biopsies, with or without
prior treatment with ipilimumab, single or combination ICB ther-
apy, as well as clinical response or no response to therapy, with
the paired CRPR samples showing the most consistent
alignment along the T cell score vector (Figures 2C and 2D). Fig-
ure S3 shows the degree to which the T cell score dominates the

502 Cancer Cell 38, 500–515, October 12, 2020

 ll
Article

Table 1. Baseline Patient Characteristics data, with the large number of genes correlated (n = 2,047) and
anti-correlated (n = 1,087) with this score based on a Pearson
n (%)
N = 101 correlation cutoff of ±0.3. This supplemental figure also includes
information about the effect of the melanoma subtype (cuta-
Age (years)
neous, mucosal, or acral) on the tumor transcriptome. The coor-
Median 56
dinated alignment of the correlated and anti-correlated genes
Range 22–89 on-treatment is reflected in the significant upregulation of the
Sex, n (%) correlated genes and the significant downregulation from base-
Male 58 (57.4) line to on-treatment in CRPR biopsies (optimal pooled t test p <
Female 43 (42.6) 1 3 10 4 and p < 1 3 10 4, respectively, Figure 2C). The corre-
ECOG performance status, n (%) lated genes accounted not only for systematic changes in the tu-
0 70 (69.3) mor cell gene expression, but also in the tumor microenviron-
ment. Except for neutrophils, the MCP-counter immune cell
1 29 (28.7)
signatures were highly correlated with the T cell signature (R2
Not reported 2 (2.0)
R0.6, Figures S4A and S4B). Use of t-SNE embedding to cluster
Stage at study entry, n (%) genes with their closest MCP-counter immune cell signature re-
III 8 (7.9) sulted in only two main clusters: genes that correlated with the
IV 93 (92.1) T cell score and genes that were anti-correlated (Figure S4C).
Prior anti-CTLA-4 therapy, n (%) These data indicate a truly coordinated immune response,
Yes 30 (29.7) including large numbers of genes, which was dominated by
No 71 (70.3) changes in T cells. We acknowledge that this conclusion is
limited by being derived from bulk RNA-seq analyses as
BRAF mutation status, n (%)
opposed to single-cell RNA-seq. Based on these analyses, we
Positive 33 (32.7)
conclude that ICB therapy has the potential to increase T cell in-
Negative 61 (60.4) filtrates regardless of whether there are clinical responses to
Not reported 7 (6.9) therapy, and that a greater degree of T cell infiltrate increases
PD-L1 status at baseline, n (%) from baseline to on-therapy biopsies is associated with
PD-L1 negative 30 (29.7) improved response to therapy.
(TPS baseline = 0)
PD-L1 positive 48 (47.5) Interferon-g Response Genes Play an Integral Role in
(TPS baseline >0) Response and Resistance to Therapy as Mediators of
NA (TPS baseline = NA) 23 (22.8) the T Cell Program
Metastatic staging, n (%) We examined the expression of effector cytokines and toxic
M0 2 (2.0)
granules induced by TCR engagement with cognate antigen,
including the cytotoxic molecules perforin and granzyme B, tu-
M1A 17 (16.8)
mor necrosis factor alpha (TNF-a) family members FAS, TRAIL
M1B 15 (14.9)
(TNFSF10), and TNF-a, and interferon-g, applying an optimal
M1C 53 (52.5) pooled t test. The expression of perforin, granzyme B, TRAIL,
Not reported (stage III)/unknown 14 (13.9) and TNF-a followed the pattern of expression of interferon-g,
Brain metastases, n (%) and was higher in biopsies of patients with a clinical response
Yes (*) 13 (12.9) to therapy (Figure S5). As exposure to most of these molecules
No 80 (79.2) results in cytotoxic death of cancer cells, we reasoned that the
Not reported (stage III) 8 (7.9)
cancer cell expression of interferon response genes may be rele-
vant to attracting other immune cells into the tumor microenvi-
Lactate dehydrogenase, n (%)
ronment and amplifying the antitumor immune response once
Normal (%ULN) 72 (71.3)
it is initiated.
Elevated (>ULN, %2*ULN) 24 (23.8)
Highly elevated (>2*ULN) 4 (4.0) Melanoma Cells Have a Uniform Gene Set Response to
Not reported 1 (1.0) Interferon-g Provided That They Can Signal through the
Melanoma primary site, n (%) Interferon-g Receptor
Acral 4 (4.0) With the goal of addressing the question if differential responses
Cutaneous 71 (70.3) to ICB may be due to a different inherent ability of melanoma
cells from different patients to respond to interferon-g, we set
Mucosal 11 (10.9)
Other 15 (14.9)
ECOG, Eastern Cooperative Oncology Group; M1a, metastases to tration of lactate dehydrogenase; PD-L1, programmed death
skin, subcutaneous tissues, or distant lymph nodes; M1b, metasta- ligand 1.
ses to lung; M1c, metastases to all other visceral sites or distant *In addition to all patients enrolled in part 4, a number of patients in parts
metastases at any site combined with an increased serum concen- 1–3 also had reported brain metastases.

Cancer Cell 38, 500–515, October 12, 2020 503

 ll
Article

B C D

Figure 2. Immune Cell Infiltration in Patient Samples from Clinical Study CheckMate 038
(A) Boxplot the MCP-counter T cell score according to response to therapy combining all treatment groups using an optimal pooled t test since data are paired
and unpaired.
(B) Boxplots showing the average expression of the genes correlated with the T cells broken down by response and pre- and on-treatment using an optimal
pooled t test since data are paired and unpaired.

(legend continued on next page)

504 Cancer Cell 38, 500–515, October 12, 2020
 ll
Article

out to examine the transcriptome changes in a panel of ex-vivo-
cultured melanoma cells exposed to interferon-g. For this study,
we used 57 previously established and described cell lines from
cultures of human melanoma biopsies or surgical resections
(Atefi et al., 2014; Neubert et al., 2017; Shin et al., 2017; Tsoi
et al., 2018) and the primary melanocyte line, HeMa, as a non-
malignant pigmented cell control. This cell line panel included
three variants of corresponding parental melanoma cell lines
that had been rendered resistant to BRAF inhibitors by contin-
uous in vitro drug exposure, denoted as AR for acquired resis-
tance (Atefi et al., 2014; Nazarian et al., 2010; Tsoi et al., 2018);
three cell lines that had developed Janus kinase 1 (JAK1) or
JAK2 loss-of-function mutations in patients; as well as variants
of parental cell lines that had JAK1 (N = 5), JAK2 (N = 5), or
beta-2 microglobulin (B2M) (N = 2) loss of function generated
in vitro through CRISPR-Cas9 gene editing, as these are genetic
events known to be associated with resistance to anti-PD-1 ther-
apy (Garcia-Diaz et al., 2017; Gettinger et al., 2017; Sade-Feld-
man et al., 2017; Shin et al., 2017; Sucker et al., 2017; Torrejon
et al., 2020; Zaretsky et al., 2016). These cell lines were exposed
to 5 ng/mL of interferon-g for 6 h to assess immediate whole-
transcriptome changes by RNA-seq analysis.
We found that a cell line’s gene expression profile was most
correlated with itself at baseline and upon exposure to inter-
feron-g, indicating that cell line identity is a major contributor
to gene expression (Figure S6). In addition, we observed that
the baseline samples were not correlated with each other, nor
were the interferon-g-exposed samples correlated with each
other (Figure S6). However, for the 46 cell lines that responded
to interferon-g, the changes between on-treatment and baseline
for interferon-g were highly correlated as a result of consistent
large changes in interferon-g response genes (Figure S7). Fig-
ure S8A shows the high level of concordance among expression
changes of known interferon-g response genes, making it
possible to identify outlier gene expression changes, such as
the two samples whose JAK2 expression decreased signifi-
cantly compared with the rest of the samples. Analysis of all
increased and decreased gene expression showed that there
were two broad groups of cell lines: the great majority (n = 46)
that could signal through the interferon-g receptor and had large
changes in the expression of the same set of interferon-g
response genes; and a distinct set of 12 cell lines that had
JAK1 or JAK2 natural or induced loss-of-function mutations
and were largely unable to signal through the interferon-g recep-
tor (Figures 3 and S8B). Of note, there were no major differences
in gene expression profiles induced by interferon-g exposure in
cell lines with acquired resistance to BRAF inhibitors or in B2M
knockout cell lines. Therefore, this large panel of melanoma
cell lines, each of which had a unique gene expression signature,
responded consistently to interferon-g exposure.
To identify a set of genes up- and downregulated in response
to interferon-g in typical melanoma cell lines, we used all the
samples that were not experimentally modified (excluding cases
with JAK1, JAK2, and B2M loss generated by CRISPR-Cas9, as
well as the AR samples but including naturally occurring JAK and
B2M loss cases). We performed a paired t test between cell lines
before and after treatment with interferon-g and observed a sig-
nificant decrease in the expression of 1,176 genes and increase
in the expression of 549 genes after applying a false discovery
rate (FDR) cutoff of 0.01. These changes clustered with the
changes in the primary melanocyte line HeMa, included as a
normal control, indicating that normal interferon-g signaling
was intact in all of these cases (Figure S8B). This transcriptome
signature includes previously reported interferon-g response
genes, melanoma-specific responses, as well as a deeper look
into the genes whose expression decreases in response to inter-
feron-g (Figure S9).
Interferon-g Triggers Melanoma Cells to Increase
Expression of Interferon Pathway Genes, Antigen
Presenting Machinery, and T Cell-Attracting
Chemokines
We next focused the analysis on changes in known genes that
are related to interferon-g exposure, including interference with
cell proliferation (the first functional effect that led to the descrip-
tion of interferons), upregulation of antigen presentation machin-
ery, enhancement of immune cell-attracting chemokines, as well
as genes involved in positive and negative downstream inter-
feron signaling pathways (Bach et al., 1997). Figure 3A shows
the response as fold changes, while Figure S10 shows it in terms
of baseline and on-treatment gene expression. The overall
change in transcripts involved in cell proliferation was low after
6 h of in vitro interferon-g exposure (Figure 3A). There was evi-
dence of increased expression of multiple HLA class I and II
genes, with the most consistent increase being in HLA-E. The
two major clusters in the data were defined by whether or not
HLA class II genes are expressed at baseline. There was a
very strong and consistent increase in the expression of addi-
tional genes in the antigen presentation pathway, in particular
NLRC5 (also known as HLA class I transactivator) and CIITA
(HLA class II transactivator), TAP transporters and proteasome
subunits. Another set of genes with large increases were those
involved in the interferon signaling pathway itself, representing
an amplification of the interferon-g signal, including JAK2 (but
not JAK1), several signal transductors and activators of tran-
scription and interferon response factors (IRFs), in particular
the transcription factor IRF1, which serves as a well-defined pos-
itive control gene of interferon-g response (Garcia-Diaz et al.,
2017), going up in all the samples with JAKs intact (Figure 3B).
With this gene group, there was also an increase in negative

(C) Principal-component analysis showing all RNA-seq samples, paired and unpaired, plotted on the first two components. The size of the data point is pro-
portional to the T cell score. Open circles correspond to pre-treatment samples, while closed circles correspond to on-treatment samples. The data points are
colored by response (red, PD; green, SD; blue, CRPR). The vector for the T cell score is plotted as well. Outlier samples are labeled by case. The 12 paired outlier
samples were either due to quality control issues that had not been detected, or may be due to an actual different biology. For example, cases 48, 20038, and
30022 all had interferon-g signaling either going down or not changing on-treatment, while cases 30004 and 20001 have G2M cell-cycle genes that either increase
or did not respond.
(D) Boxplots of MCP-counter immune cell deconvolution types according to response to therapy.
Mixed t test for paired and unpaired samples using an optimal pooled t test since data are paired and unpaired (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).
See also Figures S3–S5.
Cancer Cell 38, 500–515, October 12, 2020 505
 ll
Article

Figure 3. Interferon-g-Induced Changes in Gene Expression in 58 Human Cell Lines

(A) Heatmap of changes in key interferon-g response genes after 6 h of treatment at 5 ng/mL expressed as fold changes pre-treatment to post-treatment. Genes
are organized by class: anti-proliferative (black), antigen presentation (white), chemoattractants (cyan), cytotoxic effectors (orange), feedback/signaling (purple).
(B) The pre- and post-treatment gene expression levels of IRF1, the transcription factor executing the downstream interferon-g program, shown as arrows (red,
up; blue, down).
(C) The pre- and post-treatment gene expression levels of CD274, the gene that codes for PD-L1, shown as arrows (red, up; blue, down).
See also Figures S6–S10.

506 Cancer Cell 38, 500–515, October 12, 2020


Article
regulators of the pathway including SOCS genes, as well as PD-
L1 (CD274), which are also well characterized interferon-g imme-
diate response genes (Garcia-Diaz et al., 2017; Shin et al., 2017)
(Figure 3C). Chemokines, in particular CXCL9, CXCL10, and
CXCL11, were increased more consistently than immune
effector molecules and physiologically serve to attract more
T cells in response to interferon-g exposure. These increases
in gene expression were very inconsistent or non-existent in
the 12 cell lines that had JAK1 or JAK2 natural or induced
loss-of-function mutations. Overall, melanoma cell lines have a
near uniform transcriptome response to interferon-g exposure,
with increases in transcripts for antigen presentation, interferon
pathway, and chemokine genes, that is lost if there are loss-of-
function mutations in interferon-g receptor pathway signaling
at the level of JAK1 or JAK2.
Analysis of Interferon-g Response Genes in Patient
Biopsies Reveals Increased HLA Expression as the
Major Difference between Response and Lack of
Response to ICB Therapy
We analyzed the CheckMate 038 patient biopsies for interferon-
g response genes before and during ICB therapy. For this anal-
ysis, we acknowledge that cells other than melanoma cells may
be responsible for the expression of interferon-g response genes
when analyzing bulk RNA-seq from biopsies as opposed to the
experiment using human cell lines or single-cell RNA-seq from
fresh tumor biopsies. Biopsies of patients with PD and SD
included samples with and without expression of interferon
response genes at baseline (Figure 4). There was evidence of
an increase in the frequency and intensity of gene expression
in on-therapy biopsies, in particular for chemokines and immune
effector molecules (Figure 4). Biopsies of patients with CRPR
had higher baseline expression of interferon response genes,
which increased substantially in the on-therapy biopsies (Fig-
ure 4). The largest difference between the PD and SD groups
compared with the CRPR group was in the marked increase in
antigen presentation pathway genes, in particular in HLA class
I and class II genes, B2M, TAP1, TAP2, NLRC5, and CIITA. After
normalizing the data by leukocyte common antigen ([LCA], CD45
gene) expression to account for changes in immune cell infil-
trates, the HLA class I gene expression did not increase. This
does not mean that it did not increase in the tumor, but rather
that the predicted increase in HLA class I expression due to inter-
feron-g signaling was not large enough to be detected against
the background of other changes in the tumor microenvironment
(Figures S11 and S12). Therefore, the main difference between
interferon-g genes in biopsies of patients with response and
resistance to ICB was an increase in antigen presentation genes
in biopsies obtained during therapy, a conclusion that can be
confounded by being derived from bulk RNA-seq analyses.
CRPR Cases Were Associated with Consistent Changes
in Hallmark Pathways, While PD Cases Involve Different
Pathways for Each Sample
We found a large set of genes that were significantly different be-
tween on-treatment CRPR and on-treatment PD, likely driven in
part by the large number of consistent changes in the on-treat-
ment CRPR relative to the pre-treatment CRPR (1,935 up and
3,503 down after applying an optimal pooled t test with an
ll
FDR cutoff of 0.05, Figure S13). The SD and PD samples did
not change consistently from pre-treatment to on-treatment,
except for four genes, ADI1, ARHGEF9, POU6F2, and RGPD4,
which were significantly different on-treatment relative to pre-
treatment SD cases (after applying an optimal pooled t test
with an FDR cutoff of 0.05). Only one gene in baseline samples
was associated with response, as expression of GPR31 was
significantly different between biopsies of patients with CRPR
or PD (p = 8.4 3 10 7, after applying an optimal pooled t test
with an FDR cutoff of 0.05).
To understand the major pathways driving response or lack of
response, we applied gene set enrichment analysis (GSEA) using
the hallmark gene signatures to identify enrichment of gene sets
changing on-treatment in CRPR (Figure 5A). We identified a set
of gene signatures that were also present when we applied
GSEA to genes significantly altered in cell lines after interferon-
g treatment (Figure 5B). When analyzed individually for the two
top-scoring pathways according to response to therapy, most
biopsies with CRPR had increases in interferon-g genes and de-
creases in G2M checkpoint genes, while biopsies from patients
with PD had increases or decreases of these gene sets with no
apparent directionality (Figure S14). This implies that inter-
feron-g response, presumably resulting from higher tumor anti-
gen-specific T cell infiltration, is the major driver of clinical
response in the CRPR samples. Despite the caveat of the mixed
cell content of the tumor microenvironment contributing to the
bulk RNA-seq transcriptome analysis, the CRPR and SD cases
had high increases in interferon-g response gene sets that
were consistent with the interferon-g 6 h in vitro human cell line
data, while there were PD cases for which all the interferon-g
genes were downregulated (Figures S14 and S15). Similarly,
for the HALLMARK_G2M_CHECKPOINT gene set, most of the
CRPR and SD cases were decreased, consistent with the inter-
feron-g 6-h in vitro exposure data (Figures S14 and S15). Howev-
er, there were biopsies from multiple patients with PD and one
with SD for which all the cell-cycle genes increased. Other indi-
vidual Hallmark gene sets similarly revealed fold changes
consistent with interferon-g treatment for CRPR cases, while
changes for the PD cases were the opposite of what would be
expected based on interferon-g biology. This is true both for
additional immune gene sets (Figures S16 and S17) and for addi-
tional cell-cycle gene sets for E2F targets and mitotic spindle
(Figure S18). The main observation was that the gene expression
profile of biopsies from patients with clinical response is well
defined and consistent with exposure to interferon-g, while there
was a large diversity of gene expression changes in SD and
PD cases.
Changes in WNT, MYC, and T Cell Exclusion Programs as
Tumor-Intrinsic Responses to Interferon-g
Several publications have recently shown that high WNT
signaling contributes to immune exclusion (Grasso et al., 2018;
Luke et al., 2019; Nsengimana et al., 2018; Spranger et al.,
2015; Spranger and Gajewski, 2018). We used a nine-gene
RNA-seq-based WNT score to assess the level of WNT signaling
in each biopsy (Nsengimana et al., 2018). The WNT score is the
geometric average of APC, APC2, CTNNB1, MYC, SOX11,
SOX2, TCF12, TCF7, and VEGFA. Analysis of WNT gene score
revealed that biopsies of patients with CRPR exhibited a
Cancer Cell 38, 500–515, October 12, 2020 507
 ll
Article

Figure 4. Interferon-g-Induced Changes in Gene Expression in Specimens in Clinical Study CheckMate 038
Heatmap of key interferon-g response genes in biopsies obtained at baseline and on-treatment with immune blockade therapy, separated by response to
therapy. Samples are ordered by T cell score and annotated by ipilimumab naive status, monotherapy versus combination therapy and melanoma subtype.
Genes are organized by class: anti-proliferative (black), antigen presentation (white), chemoattractants (cyan), cytotoxic effectors (orange), feedback/signaling
(purple).
See also Figures S11 and S12.

508 Cancer Cell 38, 500–515, October 12, 2020


Article
A B
Significantly Up and Down Genes in T-test for CRPR Pre to Post
HALLMARK_INTERFERON_GAMMA_RESPONSE
HALLMARK_ALLOGRAFT_REJECTION
HALLMARK_INFLAMMATORY_RESPONSE
HALLMARK_INTERFERON_ALPHA_RESPONSE
HALLMARK_COMPLEMENT
HALLMARK_TNFA_SIGNALING_VIA_NFKB
HALLMARK_IL2_STAT5_SIGNALING
HALLMARK_IL6_JAK_STAT3_SIGNALING
HALLMARK_MYC_TARGETS_V1
HALLMARK_E2F_TARGETS
HALLMARK_G2M_CHECKPOINT
HALLMARK_MYC_TARGETS_V2
HALLMARK_MTORC1_SIGNALING
HALLMARK_OXIDATIVE_PHOSPHORYLATION
HALLMARK_DNA_REPAIR
0 20 40 60 80 100 120 140
# Genes in Overlap # Remaining Genes in Gene Set
Figure 5. Global Changes in Gene Expression in Clinical Study CheckMate 038 Consistent with Interferon-g-Induced Changes in Biopsies
from Patients with Response but Not without Response to Therapy
Gene set enrichment analysis (GSEA) using hallmark gene sets for significantly up- or downregulated genes for CRPR post-treatment relative to pre-treatment for
clinical dataset CheckMate 038 (A) and significantly up- or downregulated genes for interferon-g 6-h treatment for the unmodified human cell lines (B). At most
2,000 of most significant genes by q value submitted to GSEA for each class (maximum set by GSEA).
See also Figures S13–S18.
consistent significant decrease in WNT score in the on-therapy
biopsies, while there was no change in the score in biopsies of
patients with SD or PD (after applying an optimal pooled t test)
(Figure 6A). To explore if this effect was correlated with the de-
gree of change in T cell infiltration, we plotted the WNT score
with T cell infiltration by RNA-seq deconvolution with baseline
and on-therapy biopsies tracked with arrows. Again, we
acknowledge that this analysis is limited by the bulk RNA-seq
that includes cellular changes in the tumor microenvironment
in responding biopsies, which may make it harder to interpret
the results. Despite this caveat, we noted that biopsies of pa-
tients with PD or SD had a rather random distribution in this
WNT/T cell space, while the majority of biopsies from patients
with CRPR had a trend of going from a high WNT/low T cell to
a low WNT/high T cell when comparing the baseline and on-ther-
apy biopsies (Figure 6B). A key advantage of the data derived
from the cell lines before and after exposure to interferon-g
in vitro was that the results were not confounded by tumor purity.
We observed the same significant downward trend in the WNT
gene score (p < 7.8 3 10 6 after applying a paired t test) in the
interferon-g responsive melanoma cell lines (Figure 6C), as six
of the genes comprising the nine gene WNT score were signifi-
cantly decreased in response to interferon-g exposure. MYC, a
WNT target gene used to monitor changes in WNT signaling,
was similarly downregulated in response to interferon-g expo-
sure across most samples (Figures 6D and S19). Upstream
WNT signaling genes also changed consistently in response to
interferon-g exposure, including DKK1 and the frizzled gene fam-
ily (Figure S20A). In fact, upregulation of FZD5 was one of the
highest increases across all samples (Figure S20B).
Expression of Genes that Drive Immune Exclusion Are
Selectively Decreased in Biopsies during Response to
Therapy
Recently, Jerby-Arnon et al. (2018) developed an immune exclu-
sion signature by single-cell RNA-seq of melanoma cells from bi-
opsies with low levels of immune infiltration. This work identified
a set of transcripts originating from tumor cells that were posi-
tively or negatively associated with immune exclusion by that tu-
ll
Significantly Up and Down Genes in T-test for IFNG 6h Treatment
HALLMARK_INTERFERON_GAMMA_RESPONSE
HALLMARK_INTERFERON_ALPHA_RESPONSE
HALLMARK_TNFA_SIGNALING_VIA_NFKB
HALLMARK_ALLOGRAFT_REJECTION
HALLMARK_COMPLEMENT
HALLMARK_INFLAMMATORY_RESPONSE
HALLMARK_IL6_JAK_STAT3_SIGNALING
HALLMARK_IL2_STAT5_SIGNALING
HALLMARK_G2M_CHECKPOINT
HALLMARK_MYC_TARGETS_V1
HALLMARK_E2F_TARGETS
HALLMARK_MITOTIC_SPINDLE
HALLMARK_UV_RESPONSE_DN
HALLMARK_ESTROGEN_RESPONSE_EARLY
HALLMARK_MTORC1_SIGNALING
160 180 200 0 20 40 60 80 100 120 140 160 180 200
# Genes in Overlap # Remaining Genes in Gene Set
mor. Figure S21 shows a heatmap of these two sets of genes for
the CheckMate 038 dataset. CRPR cases showed a significant
decrease in the immune excluded-high genes (p < 1.4 3 10 4,
after applying an optimal pooled t test, Figures 6F and 6G). These
genes did not change significantly in response to interferon-g in
our in vitro testing in melanoma cell lines (the immune excluded-
down genes do go up significantly, p < 0.006, after applying a
paired t test). We similarly observed significant downregulation
in CDK4 restricted to CRPR cases (after applying an optimal
pooled t test, Figure 6H), which was reported as the driver of
the immune exclusion signature (Jerby-Arnon et al., 2018).
Based on our melanoma cell line cohort, interferon-g does not
appear to significantly alter CDK4 expression. On the other
hand, MYC genes followed the same pattern (after applying an
optimal pooled t test, Figure 6I), consistent with the WNT gene
score (Figure 6A), while also going down significantly in response
to interferon-g. We previously reported, using a separate set of
biopsies, that PAK4 was overexpressed in biopsies of patients
with low T cell infiltration and lack of response to PD-1 blockade
therapy, and that in experimental mouse models the T cell exclu-
sion and anti-PD-1 resistance could be reversed by PAK4 inhibi-
tion (Abril-Rodrı́guez et al., 2020). In the CheckMate 038 biopsy
dataset, PAK4 expression did not differ significantly between
baseline and on-treatment for patients with PD or SD, but
expression was significantly (p < 2.7 3 10 4) downregulated in
on-treatment biopsies of patients with CRPR (after applying an
optimal pooled t test, Figure 6J). This change was not explained
by interferon-g signaling, since in our melanoma cell line cohort,
PAK4 did not change significantly in response to interferon-g.
Interestingly, all of these biomarkers of immune exclusion were
highly correlated, except for the T cell exclusion-up signature,
which, as expected, was anti-correlated (Pearson correlation,
Figure 6K). Together, these data indicate that ICB therapy results
in decreased expression of immune exclusion gene programs in
biopsies of patients who experience clinical response to therapy.
While interferon-g from T cells directly decreases WNT signaling
and MYC, decreases in the Jerby-Arnon immune exclusion
down signature, CDK4, or PAK4 expression are likely down-
stream effects in clinically responding tumors.
Cancer Cell 38, 500–515, October 12, 2020 509
 ll
Article

A B

F I

G J

H K
E

Figure 6. Immune Exclusion Signatures and WNT Signaling in Biopsies of Patients Receiving ICB Therapy
(A) Boxplot of the nine gene WNT score (the geometric average of APC, APC2, CTNNB1, MYC, SOX11, SOX2, TCF12, TCF7, and VEGFA), by RNA-seq according
to response to therapy in patient biopsies using an optimal pooled t test since data are paired and unpaired (*p < 0.05, **p < 0.01, ***p < 0.001).
(B) Plot of WNT score versus T cell score with arrows connecting pre-treatment and on-treatment and one figure for each response PD (red), SD (green), and
CRPR (blue).
(C–E) Pre- and post-treatment gene expression levels of the WNT score (C), MYC (D), and WNT5A (E), the WNT ligand initiating the downstream WNT signaling
programming (red, up; blue, down).

(legend continued on next page)

510 Cancer Cell 38, 500–515, October 12, 2020


Article
DISCUSSION
Results from the current study provide a possible scenario by
which response to ICB therapy, either PD-1 blockade alone or
in combination with CTLA-4 blockade, can be explained by the
development of a strong T cell response that can overcome im-
mune cell-intrinsic, tumor-intrinsic, and tumor microenvironment
limitations to a clinically effective antitumor immune response
(Kalbasi and Ribas, 2020). Key contributing factors are: (1) the
pre-existing level of T cell infiltration of the tumor (Herbst et al.,
2014; Taube et al., 2014; Tumeh et al., 2014), which reflects
both the immunogenicity of the cancer cells and the ability of
the host immune system to have made a serious attempt to
attack them specifically; (2) the baseline expression of immune
suppressive gene programs by cancer cells that are detrimental
to the initiation of an antitumor immune response, such as WNT
signaling (Grasso et al., 2018; Spranger et al., 2015; Spranger
and Gajewski, 2018); and (3) the strength of the antitumor im-
mune response that results from the release of the immune
checkpoint PD-1, alone or together with the checkpoint CTLA-
4, driving the expression of interferon-g response genes in the
tumor microenvironment (Ayers et al., 2017; Cristescu et al.,
2018). The release of interferon-g from T cells recognizing
cognate antigen on cancer cells serves to amplify the nascent
antitumor immune response, which our and previous data (Neu-
bert et al., 2017) suggest is mediated by the conserved ability of
the great majority of melanoma cells to signal through the inter-
feron-g receptor. Amplification of the immune response is a
result of an increase in antigen presentation machinery, positive
feedback resulting in increased interferon-g pathway signaling,
production of chemokines that attract other immune cells favor-
ably altering the tumor microenvironment and inhibition of im-
mune exclusion cancer signatures.
The strength of the antitumor immune response depends on the
interplay of these key contributing factors, which could be individ-
ually modulated with additional interventions as there does not
seem to be a dominant process that would always inhibit mounting
the antitumor response in the majority of cases without clinical
response to ICB therapy. In preclinical modeling (Curran et al.,
2010; Wei et al., 2017, 2019) and in clinical series (Larkin et al.,
2019), the combination of anti-PD-1 and anti-CTLA-4 is arguably
a stronger immune stimulation than either therapy alone, thereby
shifting the balance in favor of an antitumor immune response in
a greater number of cases. Another way to shift this balance would
be to induce a physiological process of intratumoral interferon pro-
duction by triggering pattern recognition receptors, such as the
use of intratumoral administration of oncolytic viruses or Toll-like
receptor agonists, two approaches already successfully reported
to improve response rates of PD-1 blockade therapy (Ribas et al.,
2018a, 2018b; Vanpouille-Box et al., 2019). Additional combinato-
rial approaches to enhance responses to ICB, not yet demon-
strated to be clinically active in patients, include the triggering of
the STING pathway (Li and Chen, 2018), inhibition of immune sup-
(F–J) Boxplots of immune exclusion genes/signatures by RNA-seq according to response to therapy: Jerby-Arnon et al. associated immune exclusion gene set 1
up in immune excluded tissues (F) and set 2 down in immune excluded samples (G), CDK4 (H), MYC (I), and PAK4 (J) using an optimal pooled t test since data are
paired and unpaired (*p < 0.05, **p < 0.01, ***p < 0.001).
(K) Heatmap showing the level of correlation between the immune exclusion genes and signatures.
See also Figures S19–S21.
ll
pressive factors, such as WNT signaling or the adenosine pathway
(Abril-Rodrı́guez et al., 2020; Grasso et al., 2018; Smyth et al.,
2016; Spranger and Gajewski, 2018), or the release of other im-
mune checkpoints, such as LAG-3, TIM-3, or TIGIT, and others
in T cells (Smyth et al., 2016). These observations provide hope
that the benefit of cancer immunotherapy can be expanded to
more patients and indications by using a combination therapy
that reaches the level where antitumor T cells are potent enough
to overcome the different nodes that restrict immune responses
to cancer (Smyth et al., 2016).
It was possible that cancer cells may have different inherent
abilities to respond to interferon-g, and this may lead to some
patients responding or not to ICB because some cancers may
not induce the full set of interferon-g response genes that are
needed to mount a productive antitumor immune response.
However, our data using a panel of melanoma cell lines suggest
that this is very unlikely, at least in patients with metastatic mel-
anoma. The data on interferon-g response in these cell lines was
dichotomous, clearly separating the cell lines with or without the
ability to signal from the interferon-g receptor. The cell lines un-
able to signal had natural or modeled homozygous loss-of-func-
tion mutations in JAK1 or JAK2. These mutations can develop
sporadically in patients with melanoma and other cancers before
receiving ICB immunotherapy, but they are infrequent, likely less
than 1% of the cases (Liu et al., 2019; Shin et al., 2017). Upon se-
lective immune pressure they may become more frequent and
lead to acquired resistance, but these seem to also be rather
infrequent cases (Sucker et al., 2017; Zaretsky et al., 2016).
Given the low baseline frequency of loss-of-function mutations
in the interferon-g signaling pathway, signaling through the inter-
feron-g receptor inducing the full set of interferon-g response
genes may be a positively selected feature of melanoma cells,
which is rather counter-intuitive. At baseline, it may be favorable
to the cancer cells to maintain this signaling and express immune
suppressive molecules, such as PD-L1, indoleamine 2,3-dioxy-
genase, colony-stimulating factor 1, vascular endothelial growth
factor, or interleukin-6, for example, to stop the antitumor im-
mune response, but this is at the expense of making the cancer
cells more vulnerable to a future immune response.
A common feature of biopsies from patients who respond to
ICB is the expression of immune activation genes with mostly
overlapping signatures (Ayers et al., 2017; Cristescu et al.,
2018; Fehrenbacher et al., 2016; Jerby-Arnon et al., 2018; Liu
et al., 2019; Rodig et al., 2018; Roh et al., 2017; Rooney et al.,
2015; Sade-Feldman et al., 2018). Our studies suggest that the
common denominator of these immune activation transcrip-
tomes is the expression of interferon-g response genes initiated
by the activation of tumor antigen-specific T cells, which in-
creases upon ICB. Key among these gene sets is the expression
of antigen presenting machinery, most notably MHC class I and
II, among biopsies of patients who go on to respond to ICB ther-
apy (Liu et al., 2019; Rodig et al., 2018). This was the most rele-
vant feature separating patients who responded or not to
Cancer Cell 38, 500–515, October 12, 2020 511
 ll
therapy in our series, in particular the increase in MHC genes in
on-therapy biopsies. However, a significant component of the in-
crease in expression of MHC genes had to be from the infiltration
with hematopoietic lineage cells into responding tumors as as-
sessed by LCA correction. Therefore, future analyses would
need to use techniques allowing determination of cancer cell-
intrinsic antigen processing and presentation pathway.
The baseline expression of several T cell immune exclusion sig-
natures has been associated with lack of response to ICB in
different series, most of which also have overlapping gene sets
(Auslander et al., 2018; Jerby-Arnon et al., 2018; Jiang et al.,
2018; Liu et al., 2019). In our series, baseline and on-therapy bi-
opsies of patients without a response to ICB therapy had a higher
expression of the Jerby-Arnon et al. (2018) signatures of T cell
exclusion, with the relevant feature that responding biopsies dis-
played decreases in expression of WNT and MYC genes on-ther-
apy, while non-responding biopsies continued to express the im-
mune exclusion genes. Overall, our data together with
transcriptome analyses of other biopsy series (Auslander et al.,
2018; Ayers et al., 2017; Cabrita et al., 2020; Chen et al., 2016; Cris-
tescu et al., 2018; Fehrenbacher et al., 2016; Gide et al., 2019; Hel-
mink et al., 2020; Hugo et al., 2016; Jerby-Arnon et al., 2018; Jiang
et al., 2018; Liu et al., 2019; Petitprez et al., 2020; Riaz et al., 2017;
Rodig et al., 2018; Roh et al., 2017; Sade-Feldman et al., 2018), im-
plies the presence of feedback loops between the tumor and CD8
T cells mediated by interferon-g that modulates immune exclusion
secondary to WNT signaling (Luke et al., 2019).
In conclusion, T cell infiltration and expression of interferon-g-
regulated genes increase in biopsies of patients receiving ICB
therapy regardless of clinical response, whereas the degree of
HLA upregulation on-therapy and the decrease in expression
of genes associated with immune exclusion are features of clin-
ically responding biopsies. The difference between responsive
and non-responsive melanomas does not seem to be due to
an inherent ability of some cancer cells to respond differently
to interferon-g, as the quality of the interferon-g response pro-
gram was similar in the great majority of melanoma cell lines
tested. As we did not have single cells for RNA-seq analyses,
we were not able to directly determine if the same was true in
the tumor biopsies. In most cases, ICB was able to change the
transcriptome of melanoma biopsies, but it did so to different de-
grees that correlate with favorable changes in the immune cell
infiltrate. The ability to provide a more powerful anti-melanoma
immune response with combination ICB explains a higher rate
of responses. This information also provides hope that additional
combinatorial strategies may dial up the antitumor immune
response to become clinically meaningful in more patients, in
particular when combining ICB therapy with treatments that in-
crease interferon signaling inside tumors to jump-start an anti-
tumor immune response when it is not already pre-existing,
such as the intratumoral administration of oncolytic viruses or
nucleotide sequences that trigger pattern recognition receptors
to produce interferons (Ribas et al., 2018a, 2018b; Torrejon et al.,
2020; Vanpouille-Box et al., 2019).
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
512 Cancer Cell 38, 500–515, October 12, 2020
Article
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Clinical Trial and Biopsy Collections
B Human Melanoma Cell Lines
d METHODS DETAILS
B RNA Sequencing
B RNA Immune Deconvolution Using MCP-Counter
B Interferon-Gamma In Vitro Exposure and Transcrip-
tome Analysis
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Statistics and Survival Analysis
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
ccell.2020.08.005.
ACKNOWLEDGMENTS
C.S.G. was funded by the Parker Institute for Cancer Immunotherapy. J.T. was
supported by the NIH Ruth L. Kirschstein Institutional National Research Ser-
vice Award no. T32-CA009120. G.A-R. was supported by the Isabel & Harvey
Kibel Fellowship award and the Alan Ghitis Fellowship Award for Melanoma
Research. D.Y.T. was supported by a Young Investigator Award from ASCO,
a grant from the Spanish Society of Medical Oncology for Translational
Research in Reference Centers and the V Foundation-Gil Nickel Family En-
dowed Fellowship in Melanoma Research. K.C. was supported by the NIH
Ruth L. Kirschstein Institutional National Research Service Award no. T32-
CA009120 and a Cancer Research Institute Irvington Postdoctoral Fellowship.
J.D.W. was funded by the Parker Institute for Cancer Immunotherapy and NIH
grant P30 CA008748. V.A. was funded by US NIH grants CA121113, the V
Foundation, Swim Across America, the Allegheny Health Network–Johns Hop-
kins Research Fund, and the LUNGevity Foundation. D.M.P. and S.L.T. were
funded by NCI R01 CA142779, the Bloomberg-Kimmel Institute for Cancer
Immunotherapy, and a Cancer Immunology Translational Cancer Research
Grant (SU2C-AACR-DT1012) from Cancer Research Institute–Stand Up 2
Cancer (to A.R., D.M.P., and S.L.T.). Stand Up 2 Cancer is a program of the
Entertainment Industry Foundation administered by the American Association
for Cancer Research. V.E.V. was funded in part by US NIH grants CA121113,
CA006973, and CA233259, the Commonwealth Foundation, the Bloomberg-
Kimmel Institute for Cancer Immunotherapy, the Dr. Miriam and Sheldon G.
Adelson Medical Research Foundation, the V Foundation, the Allegheny
Health Network–Johns Hopkins Research Fund, and the Mark Foundation
for Cancer Research. A.K. was supported by UCLA CTSI KL2 Award, Sarcoma
Alliance for Research through Collaboration Career Enhancement Program,
Tower Cancer Research Foundation Young Investigator Award, and Radiolog-
ical Society for North America Research Scholar Grant. A.R. was funded by the
Parker Institute for Cancer Immunotherapy, NIH grants R35 CA197633, P01
CA244118, and P30 CA016042, the Ressler Family Fund, Ken and Donna
Schultz Fund, and Cancer Immunology Translational Cancer Research Grant
(SU2C-AACR-DT1012) from Cancer Research Institute–Stand Up 2 Cancer
with a supplemental grant from the Melanoma Research Alliance.
AUTHOR CONTRIBUTIONS
C.S.G. and A.R. designed the experiments and wrote the paper. C.S.G., P.R.-
M., M.W.-R., D.M.P., V.A., S.L.T., V.E.V., D.E.S., A.K., and A.R. provided over-
all scientific oversight on the analyses. C.S.G., J.T., M.Q., G.A.-R., and E.M.
performed bioinformatics analyses. M.O. and P.T. performed experimental
studies. A.C., D.Y.T., D.S.S., Y.J.K., C.P.-S., K.C., A.V.-C., C.M., D.E.S.,
A.K., and A.R. provided specific expertise for bioinformatics data

Article
interpretation. J.J.L., J.D.W., D.B.J., B.C., F.S.H., S.B., W.S., W.J.U., C.L.S.,
A.D., J.B.A.G.H., S.M.A., and A.R. collected patient samples.
DECLARATION OF INTERESTS
C.S.G. reports a pending patent on the use of PAK4 inhibitors. J.T. is currently
an employee of Tango Therapeutics. M.O. reports no conflicts of interest.
G.A.R. reports a pending patent on the use of PAK4 inhibitors. P.R.-M. is an
employee and stockholder for Bristol-Myers Squibb. M.W.-R. is an employee
and stockholder for Bristol-Myers Squibb. A.C. reports no conflicts of interest.
E.M. reports no conflicts of interest. D.Y.T. reports no conflicts of interest.
D.S.S. received honoraria from Genentech (speakers bureau) and NovMeta-
Pharma (consultant). P.T. reports no conflicts of interest. Y.J.K. reports no
conflicts of interest. C.P.S. has received payment for licensing a patent on
non-viral T cell gene editing to Arsenal. K.M.C. is a shareholder in Geneoscopy
LLC. A.V-C. reports no conflicts of interest. M.Q. reports no conflicts of inter-
est. C.M. reports no conflicts of interest. J.J.L. discloses Data and Safety
Monitoring Board: TTC Oncology; Scientific Advisory Board: 7 Hills, Actym, Al-
phamab Oncology, Kanaph, Mavu (now part of AbbVie), Onc.AI, Pyxis, Spring-
bank, Tempest, Consultancy: Abbvie, Akrevia, Algios, Array, Astellas, Bayer,
Bristol-Myers Squibb, Eisai, EMD Serono, Ideaya, Incyte, Janssen, Merck,
Mersana, Novartis, PTx, RefleXion, Silicon, Tesaro, Vividion; Research Sup-
port: (all to institution for clinical trials unless noted) AbbVie, Agios (IIT), Array
(IIT), Astellas, Bristol-Myers Squibb, CheckMate (SRA), Compugen, Corvus,
EMD Serono, Evelo (SRA), Five Prime, FLX Bio, Genentech, Immatics, Immu-
nocore, Incyte, Leap, MedImmune, Macrogenics, Necktar, Novartis, Palleon
(SRA), Merck, Springbank, Tesaro, Tizona, Xencor; Travel: Akrevia, Bayer,
Bristol-Myers Squibb, EMD Serono, Incyte, Janssen, Merck, Mersana, Novar-
tis, Pyxis, RefleXion; Patents: (both provisional) serial no. 15/612,657 (Cancer
Immunotherapy), PCT/US18/36052 (Microbiome Biomarkers for Anti-PD-1/
PD-L1 Responsiveness: Diagnostic, Prognostic and Therapeutic Uses
Thereof). J.D.W. is consultant for: Adaptive Biotech; Amgen; Apricity; Ascent-
age Pharma; Astellas; AstraZeneca; Bayer; Beigene; Bristol-Myers Squibb;
Celgene; Eli Lilly; F Star; Imvaq; Kyowa Hakko Kirin; Linneaus; AstraZeneca;
Merck; Neon Therapuetics; Polynoma; Psioxus; Recepta; Takara Bio; Trieza;
Truvax; Serametrix; Surface Oncology; Syndax; Syntalogic. J.D.W. receives
research support from: Bristol-Myers Squibb; AstraZeneca; Sephora. J.D.W.
has Equity in: Tizona Pharmaceuticals; Adaptive Biotechnologies; Imvaq; Bei-
gene; Linneaus. D.B.J. serves on advisory boards for Array BioPharma, BMS,
Jansen, Merck, and Novartis, receives research funding from BMS and Incyte,
and has a patent pending on using MHC-II as a biomarker for immunotherapy
response. B.C. is a member of the speakers bureau for Regeneron and Sanofi.
F.S.H. reports funding from Bristol-Myers Squibb to institution, during the
conduct of the study; grants, personal fees and other from Bristol-Myers
Squibb, personal fees from Merck, personal fees from EMD Serono, grants,
personal fees and other from Novartis, personal fees from Takeda, personal
fees from Surface, personal fees from Genentech/Roche, personal fees from
Compass Therapeutics, personal fees from Apricity, personal fees from Bayer,
personal fees from Aduro, personal fees from Partners Therapeutics, personal
fees from Sanofi, personal fees from Pfizer, personal fees from Pionyr, per-
sonal fees from 7 Hills Pharma, personal fees from Verastem, personal fees
from Torque, personal fees from Rheos, personal fees from Kairos, personal
fees from Bicara, from Psioxus Therapeutics, personal fees from Amgen, other
from Pieris Pharmaceutical, from Boston Pharmaceuticals, from Zumutor,
outside the submitted work; In addition, Dr. Hodi has a patent for Methods
for Treating MICA-Related Disorders (no. 20100111973) with royalties paid,
a patent for Tumor Antigens and Uses Thereof (no. 7250291) issued, a patent
for Angiopoiten-2 Biomarkers Predictive of Anti-immune checkpoint response
(no. 20170248603) pending, a patent for Compositions and Methods for Iden-
tification, Assessment, Prevention, and Treatment of Melanoma using PD-L1
Isoforms (no. 20160340407) pending, a patent for Therapeutic Peptides (no.
20160046716) pending, a patent for Therapeutic Peptides (no.
20140004112) pending, a patent for Therapeutic Peptides (no.
20170022275) pending, a patent for Therapeutic Peptides (no.
20170008962) pending, a patent for Therapeutic Peptides patent no.
9402905 issued, a patent for Methods of Using Pembrolizumab and Trebana-
nib pending, a patent for vaccine compositions and methods for restoring
NKG2D pathway function against cancers patent no. 10279021 issued, a pat-
ll
ent for antibodies that bind to MHC class I polypeptide-related sequence A
patent no. 10106611 issued, and a patent for Anti-Galectin Antibody Bio-
markers Predictive of Anti-Immune Checkpoint and Anti-Angiogenesis Re-
sponses publication no. 20170343552 pending. S.B. reports advisory board
participation (with honorarium) from Genentech, EMD Serono, BMS and Sa-
nofi-Genzyme; and research funding to his institution (University of Washing-
ton) from BMS, Oncosec, EMD Serono, Merck, NantKwest, Novartis, Exicure,
Incyte, and Immune Design. W.S. discloses consulting fees from BMS, Merck,
Novartis, Regeneron and research grant support from BMS, Merck, Novartis.
W.J.U. serves on a compensated data and safety monitoring committee for
Astra Zeneca, and reports institutional grants from BMS and Merck. C.L.S. dis-
closes research funding to his University from Merck, Celldex, and GlaxoS-
mithKline; research support in kind to his University from Theraclion and 3M;
Scientific Advisory Board role with Immatics (prior), and Curvac (planned); PI
role for Polynoma with compensation to his University; and patent royalties
as co-inventor of peptides for use in cancer vaccines (patents held by the
UVA Licensing and Ventures Group). A.D. reports receiving honoraria from
Nektar, Idera, Novartis, and Array BioPharma. J.B.A.G.H. Advisory roles:
Aimm, Amgen, AZ, Bayer, BMS, Celsius, Gadeta, GSK, Immunocore, MSD,
Merck Serono, Neon, Neogene, Novartis, Pfizer, Roche, Sanofi, Seattle Ge-
netics, Vaximm. Grant support: Neon, BMS, MSD, Novartis. Stock options:
Neogene Therapeutics. S.M.A. participated in advisory boards with Amgen,
BMS, Merck Serono, MSD, Pierre-Fabre, Sanofi-Aventis and in educational
events organized by BMS, MSD, Pierre-Fabre, Roche, Sanofi-Aventis. V.A. re-
ceives research funding from Bristol-Myers Squibb. D.M.P. and S.L.T. report
stock and other ownership interests from Aduro Biotech, Compugen, DNAtrix,
Dragonfly Therapeutics, Ervaxx, Five Prime Therapeutics, FLX Bio, Jounce
Therapeutics, Potenza Therapeutics, Tizona Therapeutics, and WindMIL;
consulting or advisory roles with AbbVie, Amgen, Bayer, Compugen, DNAtrix,
Dragonfly Therapeutics, Dynavax, Ervaxx, Five Prime Therapeutics, FLX Bio,
Immunocore, lmmunomic Therapeutics, Janssen Oncology, Medlmmune,
Merck, Tizona Therapeutics, and Wind MIL; research funding from Bristol-
Myers Squibb, Compugen, and Potenza Therapeutics; patents, royalties,
and other intellectual property from Aduro Biotech, Bristol-Myers Squibb,
and lmmunonomic Therapeutics; and travel, accommodations, and expenses
from Bristol-Myers Squibb and Five Prime Therapeutics. V.E.V. is a founder of
Delfi Diagnostics and Personal Genome Diagnostics, serves on the Board of
Directors and as a consultant for both organizations, and owns Delfi Diagnos-
tics and Personal Genome Diagnostics stock, which are subject to certain re-
strictions under university policy. In addition, Johns Hopkins University owns
equity in Delfi Diagnostics and Personal Genome Diagnostics. V.E.V. is an
advisor to Bristol-Myers Squibb, Genentech, Merck, and Takeda Pharmaceu-
ticals. Within the last 5 years, V.E.V. has been an advisor to Daiichi Sankyo,
Janssen Diagnostics, and Ignyta. These arrangements have been reviewed
and approved by the Johns Hopkins University in accordance with its conflict
of interest policies. D.E.S. reports no conflicts of interest. A.K. reports no con-
flicts of interest. A.R. has received honoraria from consulting with Amgen, Bris-
tol-Myers Squibb, Chugai, Genentech, Merck, Novartis, Roche, and Sanofi, is
or has been a member of the scientific advisory board and holds stock in Ad-
vaxis, Apricity, Arcus Biosciences, Bioncotech Therapeutics, Compugen, Cy-
tomX, Five Prime, FLX Bio, ImaginAb, Isoplexis, Kite-Gilead, Lutris Pharma,
Merus, PACT Pharma, Rgenix, and Tango Therapeutics, has a reports a
pending patent on the use of PAK4 inhibitors, has received research funding
from Agilent and from Bristol-Myers Squibb through Stand Up to Cancer
(SU2C), and has received payment for licensing a patent on non-viral T cell
gene editing to Arsenal.
Received: March 10, 2020
Revised: June 17, 2020
Accepted: August 10, 2020
Published: September 10, 2020
REFERENCES
Abril-Rodrı́guez, G., Torrejon, D.Y., Liu, W., Zaretsky, J.M., Nowicki, T.S., Tsoi,
J., Puig-Saus, C., Baselga Carretero, I., Medina, E., Quist, M.J., et al. (2020).
PAK4 inhibition improves PD-1 blockade immunotherapy. Nat. Cancer 1, 46–58.
Cancer Cell 38, 500–515, October 12, 2020 513
 ll
Anders, S., Pyl, P.T., and Huber, W. (2015). HTSeq—a Python framework to
work with high-throughput sequencing data. Bioinformatics 31, 166–169.
Atefi, M., Avramis, E., Lassen, A., Wong, D.J., Robert, L., Foulad, D., Cerniglia,
M., Titz, B., Chodon, T., Graeber, T.G., et al. (2014). Effects of MAPK and PI3K
pathways on PD-L1 expression in melanoma. Clin. Cancer Res. 20,
3446–3457.
Auslander, N., Zhang, G., Lee, J.S., Frederick, D.T., Miao, B., Moll, T., Tian, T.,
Wei, Z., Madan, S., Sullivan, R.J., et al. (2018). Robust prediction of response
to immune checkpoint blockade therapy in metastatic melanoma. Nat. Med.
24, 1545–1549.
Ayers, M., Lunceford, J., Nebozhyn, M., Murphy, E., Loboda, A., Kaufman,
D.R., Albright, A., Cheng, J.D., Kang, S.P., Shankaran, V., et al. (2017). IFN-
gamma-related mRNA profile predicts clinical response to PD-1 blockade.
J. Clin. Invest. 127, 2930–2940.
Bach, E.A., Aguet, M., and Schreiber, R.D. (1997). The IFN gamma receptor: a
paradigm for cytokine receptor signaling. Annu. Rev. Immunol. 15, 563–591.
Becht, E., Giraldo, N.A., Lacroix, L., Buttard, B., Elarouci, N., Petitprez, F.,
Selves, J., Laurent-Puig, P., Sautes-Fridman, C., Fridman, W.H., et al.
(2016). Estimating the population abundance of tissue-infiltrating immune
and stromal cell populations using gene expression. Genome Biol. 17, 218.
Cabrita, R., Lauss, M., Sanna, A., Donia, M., Skaarup Larsen, M., Mitra, S.,
Johansson, I., Phung, B., Harbst, K., Vallon-Christersson, J., et al. (2020).
Tertiary lymphoid structures improve immunotherapy and survival in mela-
noma. Nature 577, 561–565.
Chen, P.L., Roh, W., Reuben, A., Cooper, Z.A., Spencer, C.N., Prieto, P.A.,
Miller, J.P., Bassett, R.L., Gopalakrishnan, V., Wani, K., et al. (2016).
Analysis of immune signatures in longitudinal tumor samples yields insight
into biomarkers of response and mechanisms of resistance to immune check-
point blockade. Cancer Discov. 6, 827–837.
Cristescu, R., Mogg, R., Ayers, M., Albright, A., Murphy, E., Yearley, J., Sher,
X., Liu, X.Q., Lu, H., Nebozhyn, M., et al. (2018). Pan-tumor genomic bio-
markers for PD-1 checkpoint blockade-based immunotherapy. Science 362,
eaar3593.
Curran, M.A., Montalvo, W., Yagita, H., and Allison, J.P. (2010). PD-1 and
CTLA-4 combination blockade expands infiltrating T cells and reduces regula-
tory T and myeloid cells within B16 melanoma tumors. Proc. Natl. Acad. Sci. U
S A 107, 4275–4280.
Fehrenbacher, L., Spira, A., Ballinger, M., Kowanetz, M., Vansteenkiste, J.,
Mazieres, J., Park, K., Smith, D., Artal-Cortes, A., Lewanski, C., et al. (2016).
Atezolizumab versus docetaxel for patients with previously treated non-
small-cell lung cancer (POPLAR): a multicentre, open-label, phase 2 rando-
mised controlled trial. Lancet 387, 1837–1846.
Garcia-Diaz, A., Shin, D.S., Moreno, B.H., Saco, J., Escuin-Ordinas, H.,
Rodriguez, G.A., Zaretsky, J.M., Sun, L., Hugo, W., Wang, X., et al. (2017).
Interferon receptor signaling pathways regulating PD-L1 and PD-L2 expres-
sion. Cell Rep. 19, 1189–1201.
Gettinger, S., Choi, J., Hastings, K., Truini, A., Datar, I., Sowell, R., Wurtz, A.,
Dong, W., Cai, G., Melnick, M.A., et al. (2017). Impaired HLA class I antigen
processing and presentation as a mechanism of acquired resistance to im-
mune checkpoint inhibitors in lung cancer. Cancer Discov. 7, 1420–1435.
Gide, T.N., Quek, C., Menzies, A.M., Tasker, A.T., Shang, P., Holst, J., Madore,
J., Lim, S.Y., Velickovic, R., Wongchenko, M., et al. (2019). Distinct immune
cell populations define response to anti-PD-1 monotherapy and anti-PD-1/
anti-CTLA-4 combined therapy. Cancer Cell 35, 238–255.e6.
Grasso, C.S., Giannakis, M., Wells, D.K., Hamada, T., Mu, X.J., Quist, M.,
Nowak, J.A., Nishihara, R., Qian, Z.R., Inamura, K., et al. (2018). Genetic mech-
anisms of immune evasion in colorectal cancer. Cancer Discov. 8, 730–749.
Guo, B., and Yuan, Y. (2017). A comparative review of methods for comparing
means using partially paired data. Stat. Methods Med. Res. 26, 1323–1340.
Helmink, B.A., Reddy, S.M., Gao, J., Zhang, S., Basar, R., Thakur, R., Yizhak,
K., Sade-Feldman, M., Blando, J., Han, G., et al. (2020). B cells and tertiary
lymphoid structures promote immunotherapy response. Nature 577, 549–555.
Herbst, R.S., Soria, J.C., Kowanetz, M., Fine, G.D., Hamid, O., Gordon, M.S.,
Sosman, J.A., McDermott, D.F., Powderly, J.D., Gettinger, S.N., et al. (2014).
514 Cancer Cell 38, 500–515, October 12, 2020
Article
Predictive correlates of response to the anti-PD-L1 antibody MPDL3280A in
cancer patients. Nature 515, 563–567.
Hugo, W., Zaretsky, J.M., Sun, L., Song, C., Moreno, B.H., Hu-Lieskovan, S.,
Berent-Maoz, B., Pang, J., Chmielowski, B., Cherry, G., et al. (2016). Genomic
and transcriptomic features of response to anti-PD-1 therapy in metastatic
melanoma. Cell 165, 35–44.
Jerby-Arnon, L., Shah, P., Cuoco, M.S., Rodman, C., Su, M.J., Melms, J.C.,
Leeson, R., Kanodia, A., Mei, S., Lin, J.R., et al. (2018). A cancer cell program
promotes T cell exclusion and resistance to checkpoint blockade. Cell 175,
984–997.e24.
Jiang, P., Gu, S., Pan, D., Fu, J., Sahu, A., Hu, X., Li, Z., Traugh, N., Bu, X., Li,
B., et al. (2018). Signatures of T cell dysfunction and exclusion predict cancer
immunotherapy response. Nat. Med. 24, 1550–1558.
Kalbasi, A., and Ribas, A. (2020). Tumour-intrinsic resistance to immune
checkpoint blockade. Nat. Rev. Immunol. 20, 25–39.
Kim, D., Paggi, J.M., Park, C., Bennett, C., and Salzberg, S.L. (2019). Graph-
based genome alignment and genotyping with HISAT2 and HISAT-genotype.
Nat. Biotechnol. 37, 907–915.
Larkin, J., Chiarion-Sileni, V., Gonzalez, R., Grob, J.J., Rutkowski, P., Lao,
C.D., Cowey, C.L., Schadendorf, D., Wagstaff, J., Dummer, R., et al. (2019).
Five-year survival with combined nivolumab and ipilimumab in advanced mel-
anoma. N. Engl. J. Med. 381, 1535–1546.
Li, T., and Chen, Z.J. (2018). The cGAS-cGAMP-STING pathway connects
DNA damage to inflammation, senescence, and cancer. J. Exp. Med. 215,
1287–1299.
Liu, D., Schilling, B., Liu, D., Sucker, A., Livingstone, E., Jerby-Amon, L.,
Zimmer, L., Gutzmer, R., Satzger, I., Loquai, C., et al. (2019). Integrative molec-
ular and clinical modeling of clinical outcomes to PD1 blockade in patients with
metastatic melanoma. Nat. Med. 25, 1916–1927.
Luke, J.J., Bao, R., Sweis, R.F., Spranger, S., and Gajewski, T.F. (2019). WNT/
beta-catenin pathway activation correlates with immune exclusion across hu-
man cancers. Clin. Cancer Res. 25, 3074–3083.
Nazarian, R., Shi, H., Wang, Q., Kong, X., Koya, R.C., Lee, H., Chen, Z., Lee,
M.K., Attar, N., Sazegar, H., et al. (2010). Melanomas acquire resistance to
B-RAF(V600E) inhibition by RTK or N-RAS upregulation. Nature 468, 973–977.
Neubert, N.J., Tille, L., Barras, D., Soneson, C., Baumgaertner, P., Rimoldi, D.,
Gfeller, D., Delorenzi, M., Fuertes Marraco, S.A., and Speiser, D.E. (2017).
Broad and conserved immune regulation by genetically heterogeneous mela-
noma cells. Cancer Res. 77, 1623–1636.
Nsengimana, J., Laye, J., Filia, A., O’Shea, S., Muralidhar, S., Pozniak, J.,
Droop, A., Chan, M., Walker, C., Parkinson, L., et al. (2018). beta-Catenin-
mediated immune evasion pathway frequently operates in primary cutaneous
melanomas. J. Clin. Invest. 128, 2048–2063.
Petitprez, F., de Reynies, A., Keung, E.Z., Chen, T.W., Sun, C.M., Calderaro, J.,
Jeng, Y.M., Hsiao, L.P., Lacroix, L., Bougouin, A., et al. (2020). B cells are asso-
ciated with survival and immunotherapy response in sarcoma. Nature 577,
556–560.
Riaz, N., Havel, J.J., Makarov, V., Desrichard, A., Urba, W.J., Sims, J.S., Hodi,
F.S., Martin-Algarra, S., Mandal, R., Sharfman, W.H., et al. (2017). Tumor and
microenvironment evolution during immunotherapy with nivolumab. Cell 171,
934–949.e916.
Ribas, A., Dummer, R., Puzanov, I., VanderWalde, A., Andtbacka, R.H.I.,
Michielin, O., Olszanski, A.J., Malvehy, J., Cebon, J., Fernandez, E., et al.
(2018a). Oncolytic virotherapy promotes intratumoral T cell infiltration and im-
proves anti-PD-1 immunotherapy. Cell 174, 1031–1032.
Ribas, A., Medina, T., Kummar, S., Amin, A., Kalbasi, A., Drabick, J.J., Barve,
M., Daniels, G.A., Wong, D.J., Schmidt, E.V., et al. (2018b). SD-101 in combi-
nation with pembrolizumab in advanced melanoma: results of a phase Ib,
multicenter study. Cancer Discov. 8, 1250–1257.
Ribas, A., and Wolchok, J.D. (2018). Cancer immunotherapy using checkpoint
blockade. Science 359, 1350–1355.
Rodig, S.J., Gusenleitner, D., Jackson, D.G., Gjini, E., Giobbie-Hurder, A., Jin,
C., Chang, H., Lovitch, S.B., Horak, C., Weber, J.S., et al. (2018). MHC proteins

Article
confer differential sensitivity to CTLA-4 and PD-1 blockade in untreated met-
astatic melanoma. Sci. Transl. Med. 10, eaar3342.
Roh, W., Chen, P.L., Reuben, A., Spencer, C.N., Prieto, P.A., Miller, J.P.,
Gopalakrishnan, V., Wang, F., Cooper, Z.A., Reddy, S.M., et al. (2017).
Integrated molecular analysis of tumor biopsies on sequential CTLA-4 and
PD-1 blockade reveals markers of response and resistance. Sci. Transl.
Med. 9, eaah3560.
Rooney, M.S., Shukla, S.A., Wu, C.J., Getz, G., and Hacohen, N. (2015).
Molecular and genetic properties of tumors associated with local immune
cytolytic activity. Cell 160, 48–61.
Sade-Feldman, M., Jiao, Y.J., Chen, J.H., Rooney, M.S., Barzily-Rokni, M.,
Eliane, J.P., Bjorgaard, S.L., Hammond, M.R., Vitzthum, H., Blackmon, S.M.,
et al. (2017). Resistance to checkpoint blockade therapy through inactivation
of antigen presentation. Nat. Commun. 8, 1136.
Sade-Feldman, M., Yizhak, K., Bjorgaard, S.L., Ray, J.P., de Boer, C.G.,
Jenkins, R.W., Lieb, D.J., Chen, J.H., Frederick, D.T., Barzily-Rokni, M.,
et al. (2018). Defining T cell states associated with response to checkpoint
immunotherapy in melanoma. Cell 175, 998–1013 e1020.
Sharma, A., Subudhi, S.K., Blando, J., Vence, L., Wargo, J., Allison, J.P.,
Ribas, A., and Sharma, P. (2019). Anti-CTLA-4 immunotherapy does not
deplete FOXP3(+) regulatory T cells (Tregs) in human cancers-response.
Clin. Cancer Res. 25, 3469–3470.
Sharma, P., and Allison, J.P. (2015a). The future of immune checkpoint ther-
apy. Science 348, 56–61.
Sharma, P., and Allison, J.P. (2015b). Immune checkpoint targeting in cancer
therapy: toward combination strategies with curative potential. Cell 161,
205–214.
Shin, D.S., Zaretsky, J.M., Escuin-Ordinas, H., Garcia-Diaz, A., Hu-Lieskovan,
S., Kalbasi, A., Grasso, C.S., Hugo, W., Sandoval, S., Torrejon, D.Y., et al.
(2017). Primary resistance to PD-1 blockade mediated by JAK1/2 mutations.
Cancer Discov. 7, 188–201.
Smyth, M.J., Ngiow, S.F., Ribas, A., and Teng, M.W. (2016). Combination can-
cer immunotherapies tailored to the tumour microenvironment. Nat. Rev. Clin.
Oncol. 13, 143–158.
Sondergaard, J.N., Nazarian, R., Wang, Q., Guo, D., Hsueh, T., Mok, S.,
Sazegar, H., Macconaill, L.E., Barretina, J.G., Kehoe, S.M., et al. (2010).
Differential sensitivity of melanoma cell lines with BRAFV600E mutation to
the specific raf inhibitor PLX4032. J. Transl Med. 8, 39.
Spranger, S., Bao, R., and Gajewski, T.F. (2015). Melanoma-intrinsic beta-cat-
enin signalling prevents anti-tumour immunity. Nature 523, 231–235.
ll
Spranger, S., and Gajewski, T.F. (2018). Impact of oncogenic pathways on
evasion of antitumour immune responses. Nat. Rev. Cancer 18, 139–147.
Sucker, A., Zhao, F., Pieper, N., Heeke, C., Maltaner, R., Stadtler, N., Real, B.,
Bielefeld, N., Howe, S., Weide, B., et al. (2017). Acquired IFNgamma resis-
tance impairs anti-tumor immunity and gives rise to T-cell-resistant melanoma
lesions. Nat. Commun. 8, 15440.
Taube, J.M., Klein, A.P., Brahmer, J.R., Xu, H., Pan, X., Kim, J.H., Chen, L.,
Pardoll, D.M., Topalian, S.L., and Anders, R.A. (2014). Association of PD-1,
PD-1 ligands, and other features of the tumor immune microenvironment
with response to anti-PD-1 therapy. Clin. Cancer Res. 20, 5064–5074.
Torrejon, D.Y., Abril-Rodriguez, G., Champhekar, A.S., Tsoi, J., Campbell,
K.M., Kalbasi, A., Parisi, G., Zarestsky, J.M., Garcia-Diaz, A., Puig-Saus, C.,
et al. (2020). Overcoming genetically-based resistance mechanisms to PD-1
blockade. Cancer Discov. 10, 1140–1157.
Tsoi, J., Robert, L., Paraiso, K., Galvan, C., Sheu, K., Lay, J., Wong, D.J.L.,
Atefi, M., Shirazi, R., Wang, X., et al. (2018). Multi-stage differentiation defines
melanoma subtypes with differential vulnerability to drug-induced iron-depen-
dent oxidative stress. Cancer Cell 33, 890–904.e5.
Tumeh, P.C., Harview, C.L., Yearley, J.H., Shintaku, I.P., Taylor, E.J., Robert,
L., Chmielowski, B., Spasic, M., Henry, G., Ciobanu, V., et al. (2014). PD-1
blockade induces responses by inhibiting adaptive immune resistance.
Nature 515, 568–571.
Vanpouille-Box, C., Hoffmann, J.A., and Galluzzi, L. (2019). Pharmacological
modulation of nucleic acid sensors—therapeutic potential and persisting ob-
stacles. Nat. Rev. Drug Discov. 18, 845–867.
Wei, S.C., Anang, N.A.S., Sharma, R., Andrews, M.C., Reuben, A., Levine,
J.H., Cogdill, A.P., Mancuso, J.J., Wargo, J.A., Pe’er, D., et al. (2019).
Combination anti-CTLA-4 plus anti-PD-1 checkpoint blockade utilizes cellular
mechanisms partially distinct from monotherapies. Proc. Natl. Acad. Sci. U S A
116, 22699–22709.
Wei, S.C., Levine, J.H., Cogdill, A.P., Zhao, Y., Anang, N.A.S., Andrews, M.C.,
Sharma, P., Wang, J., Wargo, J.A., Pe’er, D., et al. (2017). Distinct cellular
mechanisms underlie anti-CTLA-4 and anti-PD-1 checkpoint blockade. Cell
170, 1120–1133.e17.
Zaretsky, J.M., Garcia-Diaz, A., Shin, D.S., Escuin-Ordinas, H., Hugo, W., Hu-
Lieskovan, S., Torrejon, D.Y., Abril-Rodriguez, G., Sandoval, S., Barthly, L.,
et al. (2016). Mutations associated with acquired resistance to PD-1 blockade
in melanoma. N. Engl. J. Med. 375, 819–829.
Cancer Cell 38, 500–515, October 12, 2020 515
Cell Mentor
Looking for insights?
Advice? Techniques?
Cell Mentor, an online resource
for researchers, can help.

Cell Mentor—an online resource from Cell Press and Cell Signaling Technology—empowers
early-career researchers with career insights, publishing advice, and techniques on experimental
processes and procedures. Now it’s even easier to tap into the knowledge and experience of
experts who’ve walked in your shoes.

cellmentor.com
 ll

Article
The PD-1/PD-L1-Checkpoint Restrains T cell
Immunity in Tumor-Draining Lymph Nodes
Floris Dammeijer,1,2,11,* Mandy van Gulijk,1,2,11 Evalyn E. Mulder,3 Melanie Lukkes,1 Larissa Klaase,1
Thierry van den Bosch,4 Menno van Nimwegen,1 Sai Ping Lau,1,3 Kitty Latupeirissa,1 Sjoerd Schetters,5
Yvette van Kooyk,5 Louis Boon,6 Antien Moyaart,4 Yvonne M. Mueller,7 Peter D. Katsikis,7 Alexander M. Eggermont,8
Heleen Vroman,1,2 Ralph Stadhouders,1,9 Rudi W. Hendriks,1 Jan von der Thu € sen,4 Dirk J. Gru
€ nhagen,2,3
Cornelis Verhoef,2,3 Thorbald van Hall,10,12,* and Joachim G. Aerts1,2,12,13,*
1Department of Pulmonary Medicine, Erasmus Medical Center, Rotterdam, the Netherlands
2Erasmus MC Cancer Institute, Erasmus Medical Center, Rotterdam, the Netherlands
3Department of Surgical Oncology, Erasmus Medical Center, Rotterdam, the Netherlands
4Department of Pathology, Erasmus Medical Center, Rotterdam, the Netherlands
5Department of Molecular Cell Biology and Immunology, Amsterdam University Medical Center, Amsterdam, the Netherlands
6Polpharma Biologics, Utrecht, the Netherlands
7Department of Immunology, Erasmus Medical Center, Rotterdam, the Netherlands
8Princess Máxima Center, Utrecht, the Netherlands
9Department of Cell Biology, Erasmus Medical Center, Rotterdam, the Netherlands
10Department of Medical Oncology, Oncode Institute, Leiden University Medical Center, Leiden, the Netherlands
11These authors contributed equally
12These authors contributed equally
13Lead Contact

*Correspondence: f.dammeijer@erasmusmc.nl (F.D.), t.van_hall@lumc.nl (T.v.H.), j.aerts@erasmusmc.nl (J.G.A.)

https://doi.org/10.1016/j.ccell.2020.09.001

SUMMARY

PD-1/PD-L1-checkpoint blockade therapy is generally thought to relieve tumor cell-mediated suppression in

the tumor microenvironment but PD-L1 is also expressed on non-tumor macrophages and conventional
dendritic cells (cDCs). Here we show in mouse tumor models that tumor-draining lymph nodes (TDLNs)
are enriched for tumor-specific PD-1+ T cells which closely associate with PD-L1+ cDCs. TDLN-targeted
PD-L1-blockade induces enhanced anti-tumor T cell immunity by seeding the tumor site with progenitor-ex-
hausted T cells, resulting in improved tumor control. Moreover, we show that abundant PD-1/PD-L1-interac-
tions in TDLNs of nonmetastatic melanoma patients, but not those in corresponding tumors, associate with
early distant disease recurrence. These findings point at a critical role for PD-L1 expression in TDLNs in gov-
erning systemic anti-tumor immunity, identifying high-risk patient groups amendable to adjuvant PD-1/PD-
L1-blockade therapy.

INTRODUCTION
Drugs targeting the PD-1/PD-L1 pathway have revolutionized
the treatment of multiple cancer types including non-small cell
lung cancer, renal cancer, and melanoma with a subset of pa-
tients experiencing durable responses. However, still a majority
of patients and cancer types do not, or only temporarily respond
to these immune checkpoint blocking (ICB) drugs (Dammeijer
et al., 2017a; Ribas and Wolchok, 2018). PD-1/PD-L1 blocking
antibodies are believed to act primarily in the tumor microenvi-
ronment (TME), by re-invigorating exhausted T cells and thereby
reviving pre-existing anti-tumor immunity (Ribas and Wolchok,
2018). Based on this hypothesis, several theories have been pro-
posed to explain the lack of ICB efficacy in patients, such as a
lack of PD-L1 expression or T cell infiltration in the TME and up-
regulation of other co-inhibitory receptors or suppressive mole-
cules following ICB therapy (Koyama et al., 2016; Nishino et al.,
2017). However, the predictive value of these proposed bio-
markers remains poor in the majority of tumors, while the rele-
vance of PD-L1-expression at other sites remains unknown.
Furthermore, results from recent trials evaluating combination
therapy with anti-PD-(L)1 and other co-inhibitory pathways in
the TME have been disappointing (Havel et al., 2019; Hellmann
et al., 2019; Long et al., 2019; Ready et al., 2019). Therefore, a
more comprehensive interrogation of the molecular and spatial
mechanisms of anti-PD-1/PD-L1 therapy is needed to further
boost immunotherapy efficacy.
Several recent insights exploring the biology of the PD-1 re-
ceptor and its corresponding ligand PD-L1 offer clues into
what drives ICB efficacy. Besides tumor cells, myeloid cells ex-
pressing PD-L1 have been revealed to be essential for ICB effi-
cacy as anti-PD-L1 antibodies remained effective in

Cancer Cell 38, 685–700, November 9, 2020 ª 2020 Elsevier Inc. 685
 ll
transplanted tumor cells lacking PD-L1 in a variety of models
(Kleinovink et al., 2017; Lau et al., 2017; Lin et al., 2018; Tang
et al., 2018). Furthermore, these seminal discoveries offer an
explanation for the rather unexpected finding that PD-1 primarily
acts by inhibiting signaling downstream of the CD28 costimula-
tory receptor following B7-ligation, proposedly by myeloid cells
(Hui et al., 2017; Kamphorst et al., 2017). Where this interaction
and therapeutic disruption in case of anti-PD-1/PD-L1 anti-
bodies takes place, however, is still unknown, as current genetic
and pharmacologic interventions (e.g., with the S1PR-blocking
agent FTY720) limit proper spatiotemporal analysis of ICB-
induced anti-tumor immune responses (Chamoto et al., 2017;
Lau et al., 2017; Lin et al., 2018). Additional insights into these dy-
namics may improve ICB-response prediction and future immu-
notherapy development.
Recent investigations analyzing T cell receptor clonotypes in
mouse and patient tumors before and after anti-PD-1 therapy
suggest the appearance of novel, previously non-existing clono-
types in ICB-treated tumors, and a limited expansion capacity of
tumor-resident T cells following treatment (Sade-Feldman et al.,
2018; Yost et al., 2019). In contrast to terminally differentiated tu-
mor-resident T cells, T cell factor 1 (TCF-1)-expressing progen-
itor-exhausted T cells have recently been described to generate
effector T cell progeny, however their exact origins remain un-
known (Jansen et al., 2019). These findings, together with an
abundance of B7-expressing antigen-presenting cells being
exposed to draining tumor-antigens prompted us to investigate
the role of tumor-draining lymph nodes (TDLNs) in efficacy of
anti-PD-L1-therapy in multiple pre-clinical tumor mouse models.
We find that TDLNs harbor significant proportions of tumor-spe-
cific PD-1+ T cells co-localizing with PD-L1 expressing myeloid
cells including conventional dendritic cells (cDCs). Selective tar-
geting of PD-L1 only in the TDLN, reveals that TDLN-localized
T cells are capable of mounting effective anti-tumor immune re-
sponses thereby demonstrating that TDLN-resident T cells are
able to generate ICB-mediated immunity. Finally, we show the
role of this PD-1/PD-L1 interaction in the TDLN of stage II mela-
noma patients, independent of known prognostic parameters.
TDLNs in patients with early disease recurrence are enriched
for PD-1/PD-L1 interactions which seem to primarily occur be-
tween T cells and CD11c+ DCs. In patients without disease
recurrence there are fewer PD-1/PD-L1 interactions in the
TDLN. These results offer unexplored insights in the role of
TDLNs in generating effective anti-tumor immunity and chal-
lenge the current dogma that PD-1/PD-L1-blockade occurs pri-
marily at the tumor site.
RESULTS
+
TDLNs Are Enriched for PD-1 Tumor-Specific T Cells
To gain insight into the activity of the PD-1/PD-L1-axis in LNs, we
analyzed the frequencies and phenotype of tumor antigen-spe-
cific CD8+ T cells in LNs of ovalbumin (OVA)-expressing AE17
mesothelioma tumors. AE17-OVA tumor cells were injected
intraperitoneally and at late-stage disease mediastinal LNs that
drain tumors in the peritoneal cavity (TDLNs) and inguinal control
LNs (non-TDLNs) were analyzed. We found higher frequencies of
OVA257-264-specific CD8+ T cells in the TDLN, compared with a
near absence of these cells in a non-TDLN (Figures 1A and
686 Cancer Cell 38, 685–700, November 9, 2020
Article
S1A). Tumor-specific CD8+ T cells in the TDLN were highly pro-
liferative, expressed PD-1 with higher frequencies and to higher
levels than those in non-TDLNs (Figure 1B). In the TDLN, the pro-
portions of PD-1-expressing CD8+ T cells strongly correlated
with the frequency of tumor-specific CD8+ T cells and could
therefore serve as a marker for tumor-specificity (Figures 1C
and S1A). Therefore, we enumerated PD-1+ CD4+ and CD8+
T cells in several solid tumor models and consistently found
higher frequencies of PD-1+ T cells in TDLNs compared with
non-TDLNs, irrespective of cancer type, mouse genetic back-
ground, tumor localization or T cell subset, except for the
KPC3 pancreatic cancer model (Figure 1D). This difference in
PD-1 expression did not appear to result from tumor metastasis
to the TDLN, as the frequency of CD45 cells was equally low in
both TDLNs and non-TDLNs (Figure S1B). The poorly immuno-
genic KPC3 pancreatic cancer cell line did not induce PD-1+
T cells in the TDLN, suggesting that PD-1 expression is related
to immunogenicity of the tumor. In line with this hypothesis, intro-
duction of the immunogenic OVA-antigen in KPC3 recapitulated
the results of the other tumor models (Figure 1D). In addition to
high PD-1 display, tumor antigen-specific CD8+ T cells in TDLNs
expressed other co-inhibitory receptors like TIM-3 and TIGIT,
resembling T cells from the tumor site (Figure 1E). In contrast
to PD-1 expression, other parameters of lymphocyte activation
such as proliferation were not consistently increased across
different models (Figure S1C), but absolute T cell numbers
were enhanced (Figure S1D).
To more comprehensively characterize both CD4+ and CD8+
PD-1+ tumor-specific T cells in TDLNs and their dynamics, we
infused naive CD45.1+ OT-I and OT-II cells in AE17-OVA-bearing
CD45.2+ mice and assessed their frequency and phenotype
across tissues over time (Figure 2A). As expected, CD45.1+ cells
initially accumulated in secondary lymphoid organs including the
TDLN, followed by egress and migration in blood to finally infil-
trate tumor tissue 6 days post-injection (Figure 2B). Upon hom-
ing to TDLNs, CD45.1+ T cells upregulated PD-1 and prolifer-
ated, suggesting PD-1 to be associated with activation rather
than T cell exhaustion (Figure 2C). Globally, PD-1+ T cells further
preferentially accumulated in TDLNs over time, where a sub-
group of cells expressed effector molecules including interleukin
(IL)-2, interferon (IFN)-g, and granzyme-B (Figure 2D). In addition
to increased expression of effector molecules, PD-1+ CD4+ and
CD8+ T cells expressed more CD28, CD44, CD69, and SLAMF6
(Ly108) compared with their PD-1 counterparts (Figure S2)
(Kurtulus et al., 2019). These results show initial activation of
early-effector T cells in TDLNs, but not in non-TDLNs, prior to
PD-1 expression and migration to the tumor.
TDLNs Contain Abundant PD-L1high- Expressing Myeloid
Cells including Migratory cDC2s
Next, we set out to quantify and characterize the ligands of PD-1
on LN myeloid cells. LNs harbor a complex architecture of
myeloid cells including CD11b+ dendritic cells (DCs) and macro-
phages lining the subcapsular (CD169+ subcapsular sinus mac-
rophages [SSMs]) and medullary sinuses (F4/80+ medullary si-
nus macrophages [MSMs]), type 1 and 2 conventional DCs
(cDC1 and cDC2), and granulocytes interspersed between
T cells in the paracortex (Eisenbarth, 2019; Glodde et al., 2017;
Gray and Cyster, 2012). TDLN-residing myeloid cells including
 ll
Article

A 4
*
C
non-TDLN TDLN

% OVA(257-264)-reactive
3 60
0.13% 1.54%
**

% OVA257-264-reactive
2 51.2%
CD3

40

CD3
1

20
0

LN -
OVA257-264 Tetramer+

TD on

LN
N

TD
OVA257-264 Tetramer+ 0

CD3

+
-1
PD-1+ OVA-specific

-1
PD
p<0.0001

PD
B OVA257-264-specific CD8+ T cells CD8+ T cells 8 r2= 0.923
0.39%
* * 6
*** 1400

%PD1+ CD8+
CD3
100 100 PD-1

1200
4
% Proliferation (Ki67)

80 80
1000
PD-1 MFI

60 60 800 2
%PD-1+

600 OVA257-264 Tetramer+

40 40
400 0
0 1 2 3 4
20 20 200 % OVA257-264-reactive CD8+ T-cells

0 0 0
LN -

or
T D on

LN
LN -

or
TD on
LN -

LN

m
or
T D on

LN

N
m

TD

Tu
m

TD
N

Tu
TD

Tu

D
15 CD8+ T cells 20 CD4+ Foxp3- T cells
* **
***
*** non-TDLN
TDLN 15 **
10
%PD1+
%PD1+

10
* *
*** *
5
p=0.09 ** *
p=0.07
5
p=0.43

0
AC29 AE17-OVA B16F10 MC38 s.c. MC38 i.p. KPC3 KPC3-OVA AC29 AE17-OVA B16F10 MC38 s.c. MC38 i.p. KPC3 KPC3-OVA

E
OVA257-264-reactive
CD8+ T cells CD4+ Foxp3- cells
CD8+ T cells

100 100 100

% positive

50 50 50

0 0 0
LN -

LN

or
TD on
LN -
LN -

LN

or
LN

or

TD on
TD on

m
N
m

TD
m

N
N

TD
TD

Tu
Tu
Tu

TP SP
DP Negative

Figure 1. Tumor-Draining Lymph Node Harbor PD-1+ Tumor-Specific T Cells

(A) Dot plots and quantification of CD8+ T cells in non-TDLNs and TDLNs showing percentages of OVA(257–264)-tetramer+ T cells.
(B) Proliferation (Ki-67) and PD-1-positivity were determined on ovalbumin (OVA) (257–264)-tetramer+ CD8+ T cells in non-TDLNs, TDLNs, and tumor.
Furthermore, PD-1-expression (MFI) was assessed on (PD-1+) OVA(257–264)-tetramer+ CD8+ T cells.
(C) Tetramer-positivity of PD1+ and PD1 CD8+ T cells in the TDLN. In addition, proportions of PD-1+ CD8+ T cells were plotted against proportions of OVA(257–
264)-tetramer+ CD8+ T cells in the TDLN and a Pearson correlation coefficient was calculated (r2).
(D) Comparison of frequencies of PD-1+ CD8+ (left) and CD4+ Foxp3-T helper (Th) cells (right) between TDLNs (circles) and non-TDLNs (squares) in different solid
tumor models (different colors) transfected with/without OVA, or injected orthotopically (i.p.) or subcutaneously (s.c.) in CBA/J (AC29) or C57BL/6 mice (n = 6–7
mice/group, n = 14 in case of KPC3-OVA).
(E) TIM-3, TIGIT, and PD-1-single-postitivity and co-expression were assessed on T cells in different tissues (SP = single positive, DP = double-positive, TP = triple
positive).
Means and SEMs are shown, paired t tests were performed to determine statistical significance.

Cancer Cell 38, 685–700, November 9, 2020 687

 ll
Article

A B CD8+ CD45.1+
1.25

1.00
3.0x105 Tumor
AE17-OVA i.p.
CD45.1+ OT-I/II-transfer i.v.
0.75 TDLN

% of Alive
CD45.2 +
ns
Blood
0 11 12 14 17 0.50 * * non-TDLN

***
**
Analysis
0.25

0.00
+24h +72hr +144hr

PD-1+ %Proliferated cells

C
100 100
** 100 **
100

80 80 80 80
% of CD4+ CD45.1

% of CD4+ CD45.1
% of CD8+ CD45.1
% of CD8+ CD45.1

60 60 60 60

40 40 40 40

20 20 20 20

0 0 0 0

h
4h

2h
4h

h
4h

2h

4h

2h
h
4h

2h

44
44
44

+2

+7
4
+2

+7

+2

+7
+2

+7

+1
+1

+1
+1

D PD-1+ IL-2+
IFN-γ+ Granzyme-B+
3 4 10 10 5 25
** *
8 8 *
* *** 6 6 4 20
3 4 4
% of CD4+

2 4 4 *
+

% of CD8+

% of CD8+
% of CD4+

% of CD8

15
% of CD8+

*** 3
2
**
* 3 3
2 10 *
1 2 * * 2
1 5
1 1
1

0 0 0 0 0 0
t=12 t=14 t=17 t=12 t=14 t=17 t=12 t=14 t=17 t=12 t=14 t=17 t=12 t=14 t=17 t=12 t=14 t=17

non-TDLN
TDLN

Figure 2. T Cells Are Activated in TDLNs Prior to Migration and Activation in the Tumor
(A) Experimental design (n = 5–6 mice per group per time point).
(B) Frequencies of CD45.1+ cells were determined in tumor, TDLN, blood, and non-TDLN.
(C) PD-1 positivity as well as the percentage of cells undergoing proliferation of CD8+CD45.1+ cells in the TDLN.
(D) Percentages of PD-1+, IL-2+, IFN-g+, and granzyme-B+ were compared for CD8+ T cells and CD4+ T cells in TDLN (circles) and non-TDLN (squares).
Means and SEMs are shown, paired t tests were used to calculate statistical significance. *p < 0.05, **p < 0.01, ***p < 0.001. TDLN = tumor-draining lymph node,
i.v. = intravenous, IL-2 = interleukin-2, IFN-g = interferon-gamma.

SSMs and cDCs but not granulocytes expressed high levels of types of macrophages (SSM and MSM) in all tested tumor
PD-L1, approaching levels found on their intratumoral counter- models (Figure 3D). These findings were corroborated by multi-
parts (Figure 3A). Interestingly, TDLN myeloid cells lacked or dis- color confocal microscopy of TDLN tissue, where F4/80 + MSMs
played low surface expression of PD-L2, whereas PD-L2 was in the medulla and CD11c + DCs in the LN cortex expressed the
strongly present on all investigated cells in the TME. PD-L1 highest levels of PD-L1, whereas expression levels were negli-
expression and myeloid cell numbers were consistently higher gible in granulocytes (Figure 3E).
in TDLNs compared with non-TDLNs in the investigated solid tu- DCs can be subdivided into migratory and resident subsets
mor models, paralleling PD-1 positivity on T cells (Figures 3B and based on differential expression of CD11c and MHC-II (Fig-
3C). To evaluate which myeloid cells particularly expressed PD- ure S3A) (Ohl et al., 2004). Based on this distinction, we detected
L1 and therefore could be involved in suppressing PD-1-ex- a particularly strong increase in migratory cDC2s, especially in
pressing T cells, we quantified PD-L1 on the aforementioned TDLNs (Figure S3B). These cells expressed high levels of PD-
cell types and found especially high levels on cDC2s and both L1 in addition to CD80 compared with cDC1s (Figures S3C

688 Cancer Cell 38, 685–700, November 9, 2020

 ll
Article

A B
Ratio PD-L1 expression SSM
SSM MSM cDC2 cDC1 Granulocytes
p=0.06
300 *
Tumor
**
4849 (TAM) 4849 (TAM) 7578 2837 4164
**
PD-L1

TDLN
3689 5267 6361 3044 867 250

%increase (MFI)
non-
TDLN 2646

3 4
4171

3 4
5258

3
2248
3 4
1196
3 4
p=0.07
** **
4

200

Tumor
PD-L2

2046 (TAM) 2046 (TAM) 3583 1186 3655 150

TDLN
682 1170 345 191 1300
non-
TDLN 941 1419 422 276 1916 100
3 4 3 4 3 4 3 4 3 4

AC29 AE17-OVA B16F10 MC38 s.c. MC38 i.p. KPC3 KPC3-OVA

C non-TDLN
SSM TDLN
cDC2 cDC1 Granulocytes
8000 1500 600 1000
** * ***
* * * 800
6000 p=0.06
1000 400
600
Counts

4000
400
500 200
2000 200

0 0 0 0
B16F10 MC38 s.c. KPC3-OVA B16F10 MC38 s.c. KPC3-OVA B16F10 MC38 s.c. KPC3-OVA B16F10 MC38 s.c. KPC3-OVA

D E
9000 PD-L1+ CD11c+ CD169+
cDC2 SSM

PD-L1 MFI 8000 SSM cDC1

7000
Cortex

6000
PD-L1 MFI

cDC2
5000
SSM
MSM
4000
cDC1 3000
Granu- 2000
locyte PD-L1+ F4/80+
1000
Medullary Sinus

0 AC29 AE17- B16F10 MC38 KPC3-

OVA (s.c.) OVA

PD-L1+ Ly6G+
Medulla

Figure 3. TDLN Myeloid Cells Express High Levels of PD-L1 Compared With Non-TDLNs
(A) Expression of PD-L1 and PD-L2 (KPC3-OVA) on myeloid cells in the tumor, TDLN and non-TDLN at late-stage disease.
(B) PD-L1 expression on subcapsular sinus macrophages (SSMs) was compared between TDLNs and non-TDLNs, with PD-L1 expression (MFI) on TDLN
macrophages divided over non-TDLN MFI multiplied 3100 to indicate percentage increase of expression.
(legend continued on next page)
Cancer Cell 38, 685–700, November 9, 2020 689
 ll
and S3D), which is in line with recent findings identifying migra-
tory PD-L1+ DCs to predominantly present tumor-derived anti-
gens in the TDLN (Maier et al., 2020). These data indicate
increased frequencies of PD-L1+ cells in TDLNs compared to
non-TDLNs, especially for macrophages and cDC2s.
Low-Dose Intrapleural PD-L1 Antibody Administration
Selectively Targets TDLNs
In order to study the effect of selectively targeting the PD-1/PD-L1
axis in the TDLN on tumor progression we set up a system by
which we selectively target TDLNs with therapeutic PD-L1 anti-
bodies prohibited antibody availability in the tumor environment it-
self. We examined the option to administer ICB antibody via the
intrapleural route. This location drains directly to mediastinal
LNs, which are the TDLNs of intraperitoneal tumors from where
excess antibody continues to enter the blood via the thoracic
duct (Kool et al., 2008). Therapeutic anti-PD-L1 blocking anti-
bodies were administered via intravenous (i.v.) or intrapleural
(i.pl.) routes and dispersed tissues were counterstained with a flu-
orescently labeled anti-rat IgG2a antibody, thereby obviating the
need to alter the therapeutic antibody itself potentially influencing
drug pharmacokinetics (Figure 4A). Irrespective of the route of
administration, we could readily detect in vivo binding of the ther-
apeutic PD-L1 antibody on tumor cells and TAMs by flow cytom-
etry within 24 hr post-injection (Figure 4B). The mediastinal TDLN
SSMs, MSMs and cDC2s were similarly reached by i.v. as well as
i.pl. administration and staining patterns paralleled those of direct
ex vivo staining with PD-L1 antibodies (Figure 4C). As expected
from the previous findings showing decreased PD-L1 expression
in non-TDLNs, binding of the PD-L1-antibody to myeloid cells was
decreased in the absence of tumor (Figure 4D). Interestingly, i.pl.
injection appeared to especially reach TDLN subcapsular cells
(SSMs & CD11b + cDC2s) more efficiently than the i.v. route in tu-
mor-free animals in contrast to tumor-bearing mice in which anti-
body binding on these cells was largely similar between injection
routes (Figure 4D). This was not due to differences in PD-L1
expression on cells between treatment arms as PD-L1 expression
(assessed in isotype-treated animals, Figure S4A) mirrored PD-L1
antibody binding in all the leukocyte subsets interrogated irre-
spective of injection route (Figure 4E). These data suggest that
antibody deposition in TDLNs is enhanced in the presence of tu-
mor, possibly via secondary drainage of antibody from permeable
tumor vasculature to afferent lymphatics, in addition to direct
TDLN targeting (Figure 4F). As we could now successfully target
TDLNs via i.pl. injections we next examined the extent to which
the efficacy of ICB depends on TDLNs and therefore titrated the
i.pl.-administered anti-PD-L1-antibody dose to a level that al-
lowed selective blockade in the mediastinal TDLN, with no anti-
body drainage to the intraperitoneal tumor or the circulation (Fig-
ures S4B and S4C). At a near 100-fold lower anti-PD-L1-dosage
of 2.5 mg, macrophages and cDCs in the TDLN still bound the anti-
(C) Absolute myeloid cell numbers in subcutaneous tumor models. Means and SEMs are shown and paired t tests were performed.
(D) PD-L1 expression on CD169+ SSMs, F4/80+ medullary sinus macrophages (MSM), type 1 (cDC1) and 2 conventional dendritic cells (cDC2), and Ly6G+
granulocytes from TDLNs.
(E) Multicolor confocal microscopy of a representative TDLN section from an untreated mouse bearing AC29 tumor. Slides were stained for PD-L1, CD11c (DCs),
CD169 (SSM and MSM), Ly6G (granulocytes) and F4/80 (MSM and medullary cord macrophages; MCM). A 3-dimensional composite image was created after
spectral unmixing using single stains (lower left image).
*p < 0.05, **p < 0.01, ***p < 0.001. TDLN = tumor-draining lymph node. Paired t tests were performed to determine statistical significance
690 Cancer Cell 38, 685–700, November 9, 2020
Article
body, albeit not completely, whereas the antibody did not reach
non-TDLN nor tumor cells, TAMs or circulating monocytes and
DCs (Figures S4B and S4C).
TDLN-Specific PD-L1 Antibody Elicits Anti-tumor T Cell
Immunity and Tumor Control
Using these established doses for local (2.5 mg i.pl.) TDLN target-
ing and for systemic (200 mg i.pl.) overall targeting of PD-L1, we
examined the therapeutic effect of this ICB in 2 syngeneic tumor
models, AC29 mesothelioma and MC38 colon carcinoma (Figures
5A and S5A). Systemic targeting of PD-L1 as well as local TDLN
targeting resulted in decreased tumor burden and increased sur-
vival (Figures 5B and S5B). These therapeutic effects were quite
strong, considering that not all PD-L1 molecules were blocked
in TDLN at these low doses (Figure S4B). CD8+ tumor-infiltrating
T cells (TILs) simultaneously expressing multiple co-inhibitory re-
ceptors were increased in MC38 tumor tissue, indicating a more
exhausted phenotype, which is in line with recent findings in the
same tumor model (Oh et al., 2020). Systemically administered
anti-PD-L1 increased T cell proliferation in blood, whereas both
TDLN-targeted and systemic treatment caused elevated fre-
quencies of CD69+ cells (Figure 5C). Whereas increased tumor
infiltration of T cells was markedly induced by systemic anti-PD-
L1 treatment, TDLN-targeted anti-PD-L1 specifically induced
KLRG-1+ effector T cell infiltration in the AC29 tumor model (Fig-
ure 5C). In addition, while nearly all TILs were TOX-positive,
TOX-MFI was significantly decreased in both treatment groups.
Recent research has shown that tumors may harbor TCF1+
CD8+ T cells with a stem-like phenotype, giving rise to distinct
populations, including terminally differentiated exhausted T cells
(Jansen et al., 2019). Similarly, we could identify a subset of TILs
with a stem-like phenotype termed TEXprog, characterized by the
surface marker SLAMF6+ (Ly108), being a surrogate marker for
TCF-1, displaying low Ki-67, TIM-3, and CD39 (Figures 5D, 5E,
and S5D–S5F) (Beltra et al., 2020). Intratumoral CD8+ and CD4+
TEXprog cells were increased by TDLN-targeted and systemic
PD-L1 blockade with decreased levels of proliferation indicating
preserved stemness following treatment (Figure 5F).
FTY720 Abrogates Systemic and TDLN-Specific Anti-
PD-L1 Immunotherapy Efficacy
In order to confirm the role of the T cells activated in TDLN
following PD-L1 treatment, we administered the S1P receptor
agonist FTY720, which abrogates T cell egress from lymphoid or-
gans (Yanagawa et al., 1998) during anti-PD-L1 treatment. Reten-
tion of T cells in lymphoid organs was confirmed by decreased fre-
quencies of T cells but not natural killer cells in peripheral blood
(Figure S6A). FTY720 administration abrogated anti-PD-L1 treat-
ment efficacy and prevented influx of CD4+Th- and CD8+ TILs
but retained T cells exhibiting increased PD-1 expression (Figures
5G and 5H). In addition, FTY720 administration neutralized both
 ll
Article

A B
αPD-L1-binding αPD-L1-binding αPD-L1-bound
2.5-100 μg αPD-L1 TAM
in 200 μl PBS
Tumor Cells TAM
isotype 100 ns 100 ns
200 μg αPD-L1 i.pl. Isotype
200 μg αPD-L1 i.v. 80 80

%Binding
%Binding
Mediastinal
LN (TDLN) 7
i.p. 10 AC29 60 60
Inguinal LN Tumor cells αPD-L1 i.pl
(non-TDLN)
40 40
αPD-L1 i.v
0 5 1011 20
20

aRat FITC (Donkey) 0

Tumor aPD-L1-BV711
0
Spleen
(MIH5, rat) iso i.pl. i.v. iso i.pl. i.v.
aPD-L1 (MIH5, rat)
PD-L1

C + Tumor D - Tumor

αPD-L1-bound SSM αPD-L1-bound MSM αPD-L1-bound SSM αPD-L1-bound MSM

ns
100 100 100 100
ns

80 80 80
** 80
ns
%positive

60
%positive
60 60 60

40 40 40 40

20 20 20 20

0 0 0 0
iso i.pl. i.v. iso i.pl. i.v. iso i.pl. i.v. iso i.pl. i.v.

αPD-L1-bound cDC2 αPD-L1-bound cDC1 αPD-L1-bound cDC2 αPD-L1-bound cDC1

ns
100 100 100 100
**
80 80 80 80
** *
%positive

60 60 60 60
%positive

40 40 40 40

20 20 20 20

0 0 0 0
iso i.pl. i.v. iso i.pl. i.v. iso i.pl. i.v. iso i.pl. i.v.

E αPD-L1 i.v.
F
αPD-L1 i.pl.
+ Tumor - Tumor
cDC2 100 p=0.0067 SSM
100 p=0.0067 SSM PD-L1
Rho=0.93 cDC2 αPD-L1
Rho=0.93
afferent
lymphatics
Mean PD-L1

MSM
Expression

Subcapsular
Mean PD-L1

MSM
Expression

Sinus
Granulocyte
50 Granulocyte
50
cDC1
cDC1
B cell B cell
T cell T cell
0 HEV
0
0 20 40 60 80 100
0 20 40 60 80 100

Mean αPD-L1-binding Mean αPD-L1-binding

Figure 4. Systemically Administered Anti-PD-L1 Antibodies Bind to PD-L1 Expressing Cells in the TDLN
(A) Experimental setup (n = 6–7 mice/group) and mode of antibody administration. Tumor-bearing mice (filled symbols) were treated with 200 mg isotype (gray)
intrapleurally (i.pl.) or alternatively with 200 mg anti-PD-L1 antibody i.pl (blue) or i.v. (turquoise) to target the TDLN directly or indirectly, respectively.
(B and C) (B) In vivo antibody expression was quantified on intratumoral cell types, and (C) on cells in the TDLN. Histograms showing PD-L1 expression patterns
are displayed.
(D) Quantification of antibody binding on cells in mediastinal LNs in the absence of tumor (open squares and circles).
(E) Mean PD-L1 expression on TDLN cells subsets derived from isotype-treated animals (stained with the BV711-antibody) was correlated to anti-PD-L1 antibody
binding (detected using the secondary anti-rat FITC-labeled antibody) for both i.pl. (left) and i.v. (right) injection routes. A Pearson correlation coefficient (Rho) was
determined.
(F) A graphical depiction of the proposed model showing that anti-PD-L1 antibodies reach TDLNs via intravascular and afferent lymphatics in the presence
of tumor.
Means and SEMs are shown and Mann-Whitney tests were performed indicating statistical significance. ns = not significant (p R 0.05), *p < 0.05, **p < 0.01. i.p. =
intraperitoneal, i.v. = intravenous, TDLN = tumor-draining lymph node, Ab = antibody, SSM = subcapsular sinus macrophages, MSM = medullary sinus mac-
rophages, cDC = conventional dendritic cell.

Cancer Cell 38, 685–700, November 9, 2020 691

 ll
Article

A B
100
10 ***
Isotype i.pl.
*** ** ***
80 2.5 μg i.pl.
***
200 μg i.pl. *** 9

Centimeters
10⁷ AC29 i.p. 200 μg Isotype
2.5 μg/200 μg αPD-L1 60

% Alive
8
Survival 40
0 10 13 17
7
20

0 6
0 20 40 60 80

l.

l.
-L g
-L g
l.
yp 0μg

i.p

i.p
D 0μ
i.p

D 5μ
Days

1
aP 20
aP 2.
20

e
ot
is
C Blood Isotype i.pl.
2.5 μg aPD-L1 i.pl. Tumor
200 μg aPD-L1 i.pl.

Ki-67+ CD69+ CD8+ T cells PD-1+ KLRG-1+ TOX+

20 20 30 100 10 100
**** **** **** * MFI
**** *** * ***
80 8000 ****
80
15 15 ** **
20

% of CD8+
6000

% of CD8+
% of CD45+

% of CD8+
60 60
% of CD8+

% of CD8+

10 10 5 4000

40 40
10 2000

5 5 20
20 0

0 0 0 0 0 0

D E F CD8+ TEXprog Ki-67+

15 100
SLAMF6-
5.2%
p=0.05
SLAMF6+ ** 2500 *** *
** 80 *
400 ** **

% of CD8+ TEXprog
2000
+ Isotype ** 10
300 % of CD8+ 60
** 1500
MFI
MFI

200 1000 40
9.0% 5
+ aPD-L1
2.5 μg 100 500 20

0 0 0
0
TIM-3

TIM3 Ki67 PD-1 TOX CD39 CD69

SLAMF6

G H
CD4+ Th cells CD8+ T cells PD-1+
ns
1500 10 30 100 p=0.05
***
*** *** ** *
8
Tumor weight (mg)

1000 20
% of CD45+

% of CD8+
% of CD45+

6 Isotype i.pl.
50 Isotype i.pl. + FTY720 (>day 9)
200 μg aPD-L1 i.pl.
4 200 μg aPD-L1 i.pl. + FTY720 (>day 9)
500 10
2

0 0 0 0

CD4+ TEXprog CD8+ TEXprog

I 60 15
p=0.065
* * ***
0.7%
p=0.05
40 10
% of CD4+

% of CD8+

TIM-3

20 5
SLAMF6
200 μg aPD-L1
0 0 + FTY720

Figure 5. Specific Targeting of PD-L1 in the TDLN by Low-Dose Intrapleural Injection of Anti-PD-L1 Enhances Clinical Responses in the AC29
Tumor Model and Is Abrogated by FTY720 Treatment
(A) Experimental setup (n = 8–15/group) with mice being treated with either LN-local (2.5 mg) or systemic anti-PD-L1 antibodies (200 mg) i.pl. and survival was
monitored.
(B) Kaplan-Meier curve of the experiment in A showing tumor survival. Log rank tests were used to determine statistical significance. In addition, abdominal
circumferences (being a measure of ascites volume) were measured on t = 21 post tumor-cell injection and displayed as violin-plots.

(legend continued on next page)

692 Cancer Cell 38, 685–700, November 9, 2020
 ll
Article

spontaneous and additional anti-PD-L1-mediated induction of tu-
mor CD4+ and CD8+ TEXprog cells following local- and systemically
administered anti-PD-L1 treatment (Figures 5I and S6B). These
results indicate that PD-L1 blockade does not solely rely on re-
invigoration of TME-localized T cells but in fact amplifies priming
and activation of T cells, including TEXprog, from the TDLN.
PD-L1 Antibody Blockade Amplifies TEXprog Induction
in TDLNs
To investigate the functional consequences of blocking the PD-1/
PD-L1 axis in TDLNs or TILs, we repeated the aforementioned
experiment in the aggressive AE17 mesothelioma expressing the
OVA model antigen allowing interrogation of tumor-specific
T cells (Figure S7A). CD45.1+ T cell proliferation was enhanced in
TDLNs compared with non-TDLN following anti-PD-L1 treatment
(Figure S7B). CD45.1+ T cells were proliferating more and were
more frequently present in TDLN compared with tumor early
following systemic anti-PD-L1 while no significant differences be-
tween isotype-treated animals were found (Figure S7C). Anti-PD-
L1 treatment markedly increased endogenous CD45.1 TEXprog
cells, starting in the TDLN after 24 h, followed by blood and tumor
on 72 and 144 h, respectively (Figure S7D). In an efficacy experi-
ment, repeated dosing of anti-PD-L1 locally in the TDLN and sys-
temically resulted in decreased AE17-OVA-tumor burden (Figures
6A and 6B), without overt increase in (tumor-specific) TILs (Fig-
ure 6C). To assess the replicative potential and phenotype of TILs
following anti-PD-L1 therapy, we developed an ex vivo T cell stim-
ulation assay in which CD8+ TILs were stimulated with cognate an-
tigen in the presence of their original TME with or without blocking
PD-L1 antibodies (Figure 6D). Stimulation with SIINFEKL-peptide
led to increased CD8+ T cell activation in this assay, with profound
upregulation of PD-1 following the addition of anti-PD-L1 in vitro
(Figure 6D). This induced activation system selectively triggered
CD8+ TILs and not CD4+ TILs due to the lack of MHC-II binding
OVA peptide. PD-L1 blockade in tumor cell cultures of mice treated
with LN-targeted therapeutic ICB-induced TIL activity to much
higher extent than non-treated mice, as measured by increased
cell frequencies, proliferation, and decrease in PD-1 levels (Fig-
ure 6E). These findings demonstrate that PD-L1 blockade in the
TDLN alleviates the suppressive impact on intranodal tumor-spe-
cific T cells, resulting in trafficking to the tumor site where they
display a much improved responsiveness to OVA tumor antigen.
Response to PD-L1-Blockade Occurs Independent from
TDLN Macrophages
In order to investigate which PD-L1-expressing cells in the TDLN
were most likely responsible for inhibiting T cell responses, we
made use of low-dose i.pl.-administered clodronate-encapsu-
lated liposomes (CELs), which specifically deplete LN macro-
phages upon phagocytosis (van Rooijen and Hendrikx, 2010).
SSMs and MSMs, but not cDC2, were effectively depleted
from the TDLN within 48 h at a dose of 5% CEL, whereas mac-
rophages in non-TDLNs and in the intraperitoneal tumor re-
mained essentially unaltered (Figures 7A and 7B). Importantly,
repeated TDLN-localized CEL administration failed to abrogate
anti-PD-L1 efficacy (Figures 7C and 7D), demonstrating a negli-
gible contribution of macrophages to anti-PD-L1 therapy in this
model, which is in line with recently published data in genetically
modified models (Oh et al., 2020). We then examined TDLNs us-
ing multicolor confocal microscopy to visualize co-localization of
anti-PD-L1-expressed cDC2s with CD8+ T cells. CD8+ T cells did
not co-localize with SSMs and MSMs in TDLNs, but we
frequently found clusters of PD-L1 positive DCs and CD8+
T cells (Figure 7E). Together our data indicate an active role for
TDLN in re-invigoration of tumor-specific T cells by ICB, most
likely via PD-L1-expressing cDC(2)s, but not macrophages.
PD-1/PD-L1 Interactions in TDLN but Not in Tumor
Correlate with Prognosis in Melanoma Patients
As all the aforementioned data were derived from pre-clinical
solid tumor models, it remained unclear whether PD-1/PD-L1
interaction takes place in patient TDLNs. Melanoma is a disease
with a highly variable prognosis depending on the presence of
distant and LN metastasis (stage III-IV), and in case of absence
of distant metastasis (stage I-II); tumor characteristics including
thickness (Breslow’s-depth), tumor histology, and presence of
ulceration (Schadendorf et al., 2015; Verver et al., 2018). The
melanoma setting is particularly suited for TDLN characteriza-
tion, as these tissues are generally extracted for staging pur-
poses with the aim of identifying patients who may benefit
from adjuvant systemic (immuno-)therapy (Eggermont et al.,
2018). To identify whether melanoma TDLNs feature PD-1/PD-
L1-axis activity in the absence of LN metastasis, we stained for
PD-1/PD-L1-interactions in TDLNs of systemic treatment-naive
stage II melanoma using a proximity ligation assay (PLA, Fig-
ure 8A). We studied PD-1/PD-L1-positivity in TDLNs of stage II
patients remaining disease free following surgery (n = 19) and pa-
tients with early distant disease recurrence (n = 15) (Table S1). As
expected, known prognostic factors from the primary tumor
such as Breslow’s-depth, presence of ulceration, and nodular
histology were overrepresented in the short-recurrence-free sur-
vival (RFS) cohort compared with patients remaining free of dis-
ease long-term (Schadendorf et al., 2015; Verver et al., 2019). We
detected a significantly higher PD-1/PD-L1-interaction density in
patients with a short RFS (<48 months) compared with patients
remaining disease free for more than 96 months (Figure 8B),

(C–E) (C) Blood was isolated +4 days after first injection with anti-PD-L1 treatment and characterized by multicolor flow cytometry. Tumor-infiltrating lymphocytes
(TILs) were characterized by multicolor flow cytometry for frequencies as well as for positivity for PD-1, KLRG1 and TOX (D) CD8+ TEXprog cells in tumor tissue were
characterized by positivity for SLAMF6 and further identified in (E) for expression of TIM3, Ki67, PD-1, TOX, CD39 and CD69.
(F) Frequencies of tumor-infiltrating CD8+ TEXprog cells and their level of proliferation (Ki67+).
(G) S1P receptor agonist FTY720 administration via drinking water and oral gavage at day 9 to 20 and isotype or anti-PD-L1 treatment (200 mg) i.p. at day 10, 13,
and 17 (9–10 mice per group).
(H) Frequencies of CD4+ Th cells and CD8+ T cells of CD45+ T cells and PD-1 positivity on CD8+ T cells were determined of TILs.
(I) Frequency of CD4+ and CD8+ TEXprog cells was determined for total CD4+ and CD8+ TILs, respectively.
Means and SEMs are shown and Mann-Whitney tests were performed indicating statistical significance. ns = not significant (p R 0.05), *p < 0.05, **p < 0.01. i.p. =
intraperitoneal, i.pl. = intrapleural, TEXprog = progenitor-exhausted T cells.

Cancer Cell 38, 685–700, November 9, 2020 693

 ll
Article

A B Tumor Weight C CD8+ T cells OVA(257-264)-tetramer+

(AE17-OVA)
ns CD8+ T cells
6 p=0.08 60
30 ns
20
15
1.0⁶ AE17-OVA i.p. 40

% of CD45+
200μg Isotype 4

% of CD8+
Grams
2.5μg/200μg αPD-L1
10

0 10 13 17 20 2 20
5

0 0 0

l.

l.
-L g
-L g

l.

l.
yp μg

-L g
l.

l.

l.
-L g
-L g
yp μg
i.p

i.p

-L g
D 0μ

l.

yp μ g

l.
D 5μ

i.p

i.p
D 0μ
i.p

i.p

i.p
D 0μ
D 5μ

D 5μ
i.p

i.p
ot 0

ot 0
aP 20

ot 0
1

1
aP 2.
e

1
is 20

aP 20

aP 20
aP 2.
e
is 20

aP 2.
e
is 20
OVA(257-264)-specific
CD8+ T cells
D E
+ Peptide & αPD-L1
Enzymatic
disgestion 7000
+Celltracker dye
+3d
6000
Analyze + Peptide
Untreated 5000

PD-1 MFI
2.5u g aPDL1 4000 No adds
200 ug aPDL1
3000
Tumor +/- OVA(257-264)-peptide
+/- 10μg/ml αPD-L1 antibody 2000
1000
0

OVA(257-264)-tetramer+ % increase % decrease % increase

CD8+ T cells CD4+ T cells
****
Cells PD-1 (MFI) Cell division
100 **** 100
***
250 ** p=0.05 150 **
**
OVA(257-264)-specific

80 ns
p=0.05
100 **
80 140
* ns
CD8+ T cells
%CD25+

%CD25+

ns
ns
60 200 130
60
*
ns
* * 120
40 40 50 ** *
150
110
20 20
100
100
0 0 0

Peptide - + + Peptide - + + p=0.07

*
ns
200 150
αPD-L1 - - + αPD-L1 - - + p=0.07
ns
100
ns *
140 ns *
* ns
ns
OVA(257-264)-tetramer + 130
Total CD8+

CD4 T cells
+
150
T cells

CD8+ T cells 120

50
**** 110
7000 **** **** 7000 p=0.05

*
6000 6000 * * 100
100 * **
5000 5000 0
PD-1 MFI
PD-1 MFI

4000 4000 *
* * 150
*
3000 3000 200 * 100 *
* 140 * ns
ns *
2000 2000
Total CD4+

ns
130
T cells

1000 1000
150 120
50
0 0
110
Peptide - + + Peptide - + +
αPD-L1 - - + αPD-L1 - - + 100
100

Pre-culture + - - - + - - - Peptide - + +
Post-culture - + + + - + + + αPD-L1 - - +
Peptide - - + + - - + +
αPD-L1 - - - + - - - +

Figure 6. Specific Targeting of PD-L1 in the TDLN Elicits Durable Anti-tumor Immune Responses Capable of Re-invigoration In Vitro
(A) Experimental design (n = 7–9 mice per group).
(B) Tumor weights of the different treatment groups.
(C) Frequencies of total CD8+ T cells and OVA(257–264)-specific CD8+ T cells in tumors as percentages of total alive CD45+ leukocytes or CD8+ T cells,
respectively.
(D) Design and validation of the in vitro culture system mimicking the tumor microenvironment. Means and SEMs are depicted with paired t tests used for
statistical analysis.
(E) Tumor single cell suspensions from isotype (gray), anti-PD-L1 LN-specifically (pink), and anti-PD-L1 systemically (purple) treated mice were cultured as
described in D.
Means and SEMs are shown, Mann-Whitney tests were used to evaluate statistical significance For E, treatment arms were normalized to isotype-treated animals
and baseline and post-3 day culture values for the several conditions are shown.

694 Cancer Cell 38, 685–700, November 9, 2020

 ll
Article

A SSM mLN MSM mLN cDC2 mLN

B
3000 600 1500 * ns
8000 25
PBS

1000 6000 * 5% CEL (i.pl.)

20
2000 400
Count

4000 ns
15

%alive
Count
500 ns
1000 200
2500
2000 ns
10
ns ns
0 0 1500
0
1000 5
500
Red Pulp Macrophage Marginal Metalophillic Ly6C-low Non-classical 0 0
1.0 Macrophage 6 Monocyte (blood) SSM MSM cDC1 cDC2 SSM Peritoneal RP TAM
0.4
(iLN) Mac Mac
0.8
0.3
4
0.6
%CD45+

0.2
0.4
2
0.1 0.2

0
0.0 0.0
l l l l l .
D
l l l l
BS .p .p .p .p .p p
l . l l l l
BS .p .p .p .p .p p
l .
P % i % i % i % i % i i.
BS .p .p .p .p .p p
P % i % i % i % i % i i.
100 Isotype *
P % i % i % i % i % i i. 5 10 2 0 5 0 00 0 % 5 1 0 2 0 5 0 00 0 %
5 10 20 50 00 0%
1 10 1 10 1 10 5% CEL
80 local aPDL1
5% CEL + local aPDL1

Percent survival
60
C
10⁷ AC29 i.p. 2.5 μg αPD-L1 40
5% CEL + 5% CEL
Survival
20
0 8 10 13 17

0
0 20 25 30 35
Days

E
αPD-L1
200 μg αPD-L1

αPD-L1+

CD169+ CD8a+ CD11c- αPD-L1+ CD169+ CD8a CD11c-

B220+ αPD-L1+ CD11c+
CD11c+ CD4+ CD11c- CD8a+ CD11c- αPD-L1+ CD11c

Figure 7. TDLN-Local Anti-PD-L1 Treatment Efficacy Occurs Independent of Macrophages

(A and B) CBA/J (non-tumor bearing, A; and tumor-bearing, B) mice were treated with a range of clodronate-encapsulated liposomes (CEL) concentrations in PBS
and myeloid cell subsets were analyzed 48 h later. In red is the dose established for subsequent experiments (n = 3–4 mice per group).
(C) Experimental setup (n = 7–8 per group).
(D) KM-curve of the survival of the experiment described in C. As isotype/5%CEL and anti-PD-L1/combination treatment showed significant overlap, these
groups were pooled and Log rank tests were performed.
(E) TDLN tissue 24 h following systemic (200 mg) anti-PD-L1 antibody treatment allowing for visualizing of antibody binding to different myeloid cell subsets in the
TDLN was assessed using 6-color confocal microscopy.
I.pl. = intrapleural, anti-PD-L1 = anti-PD-L1 antibody. ns = not significant (p R 0.05), *p < 0.05,**p < 0.01. i.p. = intraperitoneal, i.pl. = intrapleural, i.v. = intravenous,
SSM = subcapsular sinus macrophages, MSM = medullary sinus macrophage, cDC = conventional dendritic cell, Mac = macrophage, TAM = tumor-associated
macrophage, TDLN = tumor-draining lymph node, CEL = clodronate-encapsulated liposome.

reflecting possible ineffective anti-tumor immune surveillance. and CD68. Using this method, we observed high PD-1/PD-L1-
Similar results were obtained when contacts were numerically signaling in germinal centers largely devoid of CD8+ T cells,
quantified or alternatively calculated as percentage of total which is in line with previous data (containing primarily B cells
cell-surface area using automated software analysis (Methods and follicular T helper cells, excluded from analysis in 8A-C (Shi
S1). Interestingly, short RFS was associated with increased et al., 2018)). Also numerous interactions elsewhere in the LN
PD-1/PD-L1-contact density irrespective of the aforementioned cortex were present (Figure 8D). Similar to the murine data, these
prognostic factors, indicating that this process of PD-L1-medi- contacts were particularly established by CD8+ (and to a lesser
ated immune suppression arises independently of these primary degree CD8 ) T cells and CD11c+ DCs whereas CD68+ macro-
tumor characteristics (Figure 8C). To gain insight into which cells phages barely associated with T cells (Figure 8E). To investigate
were involved in PD-1/PD-L1-interaction, we combined the PLA whether PD-1/PD-L1-interactions occur in melanoma tumors to
with multicolor immunofluorescence staining for CD8, CD11c a similar extend as in TDLNs, we performed PLA on matched

Cancer Cell 38, 685–700, November 9, 2020 695

 ll
Article

A B ** ***
40X p=0.0064 p=0.0007
200 5

4
150

%cell-surface+
#contacts/hpf
PD-1 3
100
PD-L1 2

50 1

0 0
RFS RFS RFS RFS
<48m >96m <48m >96m

C ns ns ns

5 5 5

%cell-surface+
4 4 4
%cell-surface+
%cell-surface+

3 3 3

2 2 2

1 1 1

0 0 0
Ulcer- Ulcer- unknown SupSM NM other/
Brewlow Breslow unknown
ation+ ation- unknown
<4mm ≥4mm

D E
50
1. *
* *
%of cell-cell contacts

1.25X 40

30
*
1
20
10X
2 40X 10
2.
0

PD-1 CD8+ + + + + - - - -
CD11c+ CD8+

CD68+
CD11c+ - + - + - + - +
PD-1+ PD-L1+
PD-L1
+ - + + + +
CD68 - - -

F G PD-L1 (SP263) 20X PLA

5
4
3
%cell-surface+

2
1
0.50
ns

0.25

0.00
RFS RFS
<48m >96m

Figure 8. Stage II Melanoma TDLNs Harbor Frequent PD-1/PD-L1 Interactions, Which Associates With Early Distant Recurrence Following
Surgery, and Not in Primary Tumor Tissue
(A) Exemplary image magnified 340 of a stage II melanoma TDLN (sentinel node) displaying several PD-1/PD-L1-contacts stained using PLA.

(legend continued on next page)

696 Cancer Cell 38, 685–700, November 9, 2020
 ll
Article

primary tumor tissue of a subset of patients from whom material
was available. Intriguingly, PD-1/PD-L1 interactions in tumor tis-
sue of melanoma patients (n = 9) were scarce compared with the
TDLN with the extent of the remainder of interactions not corre-
lating with RFS (Figure 8F). PD-L1 alone, however, was ex-
pressed in tumor tissue showing that the lack of PLA-positivity
did not result from absence of PD-L1-expression (Figure 8G).
These findings support our initial findings in mice and highlight
a previously unidentified role for PD-1/PD-L1 interactions in hu-
man TDLNs, possibly identifying high-risk patient groups for
adjuvant immunotherapy.
DISCUSSION
Our data formally establish a role for TDLNs in generating pri-
mary anti-tumor immune responses following anti-PD-L1 ICBs,
a checkpoint that is generally regarded to act primarily at the
TME (Ribas and Wolchok, 2018). These data could offer an
explanation for the apparent clinical incongruences in which tu-
mor PD-L1-negative tumors may still respond to anti-PD-1
blockade, which has resulted in (chemo-)immunotherapy being
first-line treatment in several metastatic cancers irrespective of
PD-L1 positivity (Eggermont et al., 2018; Gandhi et al., 2018;
Paz-Ares et al., 2018). Although multiple mechanisms will likely
define response to ICB therapy, alleviating immune suppression
in the TDLN could propel systemic anti-tumor T cell immunity to
effectively control distant tumor sites. Our data reveal a critical
role of PD-L1 on cDCs in the TDLN, without negating the involve-
ment of this inhibitory ligand in the TME.
Our data show that PD-1 expression on T cells in the TDLN
seems to correlate with antigenicity of the tumor, which might
explain the conflicting results from others showing a minor de-
pendency on T cells from the TDLN (Siddiqui et al., 2019). Tu-
mor-mutational burden has been identified as a predictive
marker for ICB efficacy in patients, and our data strengthen the
role of TDLNs as possible mediators of ICB-responses as these
potentially immunogenic tumor-antigens are likely to drain or be
transported to TDLNs for further T cell induction (Samstein et al.,
2019). Others have previously hinted toward a role for LNs in
generating anti-tumor immunity following PD-1/PD-L1 ICB in pa-
tients by describing several dysfunctional T cell subsets arising
following treatment (Li et al., 2019; Sade-Feldman et al., 2018).
Our data formally establish these notions and directly comple-
ment recent findings on TIL-clonality in patients treated with
anti-PD-1 therapy, in which an extratumoral source of immuno-
therapy-elicited T cells was suggested (Yost et al., 2019).
TDLNs could therefore be pivotal in generating effective anti-
tumor T cell responses following liberation by anti-PD-1/PD-L1
antibodies, as has been previously shown to be the case for
other cancer immunotherapies (Spitzer et al., 2017). Surgical
removal of the TDLN in mouse tumor models revealed a contri-
bution of TDLNs to ICB efficacy, but as LNs are known to amplify
direct anti-tumor effects of immunotherapy, formal assessment
of their role in ICB therapy is lacking (Chamoto et al., 2017; Fran-
sen et al., 2018; Spitzer et al., 2017). The role of TDLNs as main
hubs in providing anti-tumor T cell immunity following ICB
furthermore fits with recent insights into PD-1/PD-L1 biology
showing that PD-1 inhibits CD28-B7-mediated T cell co-stimula-
tion, a process that takes place in a B7-rich environment such as
LNs, and is less likely in the immune-suppressive tumor microen-
vironment (Hui et al., 2017; Kamphorst et al., 2017). Furthermore,
anti-tumor T cells in TDLNs are less exhausted compared with
TIL and thus may have a proliferative advantage with ICB
(Hope et al., 2019).
Our tumor models evaluate the role of TDLN in a relatively early
stage, at a time that T cell priming following tumor cell inoculation
might still be occurring. The aggressiveness of the model pro-
hibits treatment in later phases. TILs were not significantly
decreased following FTY720 treatment on day 9, which suggests
that initial priming has largely occurred in the preceding time win-
dow (Figure 5H). Importantly, we do not exclude a role for PD-L1
blockade in the TME, but have technically not been able to selec-
tively administer antibody to the i.pl. or subcutaneously (s.c.)-
located tumors in order to directly compare the importance of
TDLN and TME to PD-L1 treatment efficacy. The exact contribu-
tion of TDLN versus TME during PD-1/PD-L1 checkpoint
blockade therapy remains to be elucidated; however, our data
clearly unraveled the involvement of lymph nodes.
We identify PD-L1+ cDCs, most likely cDC2s, as main targets
of anti-PD-L1 antibodies in the TDLN, and not macrophages.
Recently, Oh et al. showed that DC-specific genetic PD-L1-abla-
tion phenocopied the effects of complete PD-L1-knockout mice
in bearing MC38 tumors, whereas macrophage-direct PD-L1-
elimination did not (Oh et al., 2020). However, as PD-L1 was ab-
lated systemically via the CD11c promotor, analyses into the site
and preferred mechanism of PD-L1-PD-1-blockade mediated
anti-tumor rejection remained elusive. Classically, a division of
labor has been proposed with cDC2s predominantly inducing

(B) Quantification of the average number of contacts, in patients with an early (<48 months) recurrence of disease following surgery (n = 15) and no recurrence
after 96 months (n = 19) either via manual (left) and automated (right) quantification analysis. Statistical significance was determined using Mann-Whitney tests
and means plus SEMs are depicted.
(C) Using the outcome measure project in right panel B, patient tumors were divided according to Breslow’s-depth (tumor-thickness), presence of tumor ul-
ceration and histological subtype.
(D) TDLN-overview and zoom-in of germinal center naturally rich in PD-1/PD-L1-interactions (1) and paracortical area containing PD-1+ CD8+ T cells and PD-L1+
myeloid cells (2).
(E) Multiplexed images from 6 patient TDLNs high in PD-1/PD-L1 contacts (5 short, 1 long RFS) were constructed combining the PLA with CD8 (green), CD11c (yellow),
and CD68 (red), followed by counterstaining with hematoxylin (blue). Images magnified 340 were acquired from cortical LN regions, excluding germinal centers
(F) PD-1/PD-L1 interactions in primary tissue of stage II melanoma patients were stained using PLA (n = 9). The average number of contacts was enumerated from
patients with an early recurrence (n = 5) and no recurrence after 96 months (n = 4).
(G) A 320 magnified exemplary image of primary tumor tissue of stage II melanoma patients displaying PD-L1 expression (clone SP263; left) and PD-1/PD-L1
interactions using PLA (right).
Means and SEMs are shown, Mann-Whitney tests were used to calculate statistical significance. TDLN = tumor-draining lymph node, RFS = recurrence-free
survival, PLA = Proximity Ligation Assay, SupSM = Superficial spreading melanoma, NM = nodular melanoma.

Cancer Cell 38, 685–700, November 9, 2020 697

 ll
Article

CD4+ T cell activation and cDC1s priming CD8+ T cells (Wculek following blockade of PD-L1 in the TDLN. These data challenge
et al., 2020). We show, however, that cDC2s are significantly the current dogma that PD-1/PD-L1-checkpoint act primarily at
more PD-L1-positive and frequent in TDLNs compared with the effector (tumor) site and offers additional avenues for
cDC1s and co-localize with CD8+ T cells in the TDLN cortex, biomarker and combination-immunotherapy discovery.
challenging the current dogma. In agreement with our findings
are recent publications by different groups showing cDC2s to STAR+METHODS
be efficient CD8+ T cell stimulators in the context of solid cancer
and viral infection, in addition to their well-described role in acti- Detailed methods are provided in the online version of this paper
vating CD4+ T cells (Bosteels et al., 2020; Ruhland et al., 2020). and include the following:
Moreover, comprehensive analysis of tumor- and TDLN-DCs
by Maier et al. shows that both cDC1s and cDC2s adopt a reg- d KEY RESOURCES TABLE
ulatory phenotype upon apoptotic tumor-cell ingestion and up- d RESOURCE AVAILABILITY
regulate PD-L1 thereby preventing proper T cell induction (Maier B Lead Contact
et al., 2020). These combined findings provide a rationale for B Material Availability
manipulating cDC2s for the benefit of cancer immunotherapy. B Data and Code Availability
Our analysis of PD-1/PD-L1-axis activity in nonmetastatic TDLNs d EXPERIMENTAL MODEL AND SUBJECT DETAILS
of melanoma patients complements our findings in murine solid tu- B Mouse Models
mor models by identifying a subset of patients with high PD-1/PD- B Mouse Tumor Cell Lines
L1 interaction density that was associated with early disease B Melanoma Patient Cohort
relapse following surgery. Although these early stage (II) patients d METHOD DETAILS
generally have a favorable prognosis following primary resection, B In Vivo Experiments in Mouse Tumor Models
a minor subset eventually presents with distant recurrences bearing B Preparation of Single Cell Suspensions
a poor prognosis (Verver et al., 2019; von Schuckmann et al., 2019). B Flow Cytometry
It is tentative to speculate that PD-L1-mediated suppression of anti- B Multicolor Confocal Microscopy
tumor T cell responses in the TDLNs of these patients is involved in B TIL Re-stimulation Culture
the development or progression of distant metastasis, and future B PLA and Multiplex Staining (cmIHC)
research will likely shed more light on the validity of this hypothesis. B Quantification of PLA and Multiplex Stainings
An ancillary observation supporting this hypothesis is the fact that d QUANTIFICATION AND STATISTICAL ANALYSIS
the 2 patients who received primary anti-PD-1 antibodies at disease B Statistical Analysis
recurrence resulting in durable complete responses were the sec-
ond- and third-highest expressers of PD-1/PD-L1 in the TDLN (Fig- SUPPLEMENTAL INFORMATION
ure S8). Although the TDLNs bearing these high-density contacts
were excised at primary disease presentation, it is likely that distant Supplemental Information can be found online at https://doi.org/10.1016/j.
micrometastases bearing a similar genetic and TME-makeup ccell.2020.09.001.

would be susceptible to later anti-PD1 ICB therapy.

Perhaps unexpectedly, we found PD-1/PD-L1-interactions to
occur sporadically in primary melanoma tumors compared with
corresponding TDLNs. Moreover, PD-L1 expression in melanoma
tumors has been variably linked to favorable prognosis, whereas
its predictive value is limited (Daud et al., 2016; Hodi et al.,
2018; Morrison et al., 2018). Our data suggest that PD-1/PD-L1-
axis activity in TDLNs rather than the tumor itself could be a pri-
mary target for PD-1/PD-L1 checkpoint blocking antibodies,
thereby amplifying anti-tumor T cell induction. As PD-1 ICB is
currently being administered adjuvant to surgery in stage IIIB-C
disease, it will be of interest to assess the validity of PD-1/PD-
L1 expression in the TDLN of these patients. A different approach
currently being investigated is the neo-adjuvant administration of
anti-PD-1, which shows early encouraging results (Huang et al.,
2019). This is in line with our hypothesis claiming that targeting
the PD-1/PD-L1 checkpoint when the TDLN is still in situ could
harness effective anti-tumor immunity. Further evidence support-
ing this notion comes from neo-adjuvant and metastatic PD-1-
blockade studies showing increased proliferation of activated
CD8+ PD-1+/HLA-DR+ T cells in peripheral blood early after start
of ICB treatment, followed by increased T cell infiltrates in the tu-
mor (Herbst et al., 2014; Huang et al., 2019).
In summary, our findings implicate TDLNs as key orchestra-
tors of anti-tumor T cell immune responses that can be induced
ACKNOWLEDGMENTS
F.D. was funded by the Van Herk foundation. We would like to acknowledge all
involved technicians from the immunology and pulmonary medicine depart-
ment, and the animal facility at the Erasmus Medical Center for their valuable
contributions to this project.
AUTHOR CONTRIBUTIONS
F.D., M.G., T.H., J.A., R.H. P.K., and Y.M. designed the experiments. F.D.,
M.G., M.L. L.K., M.N., S.L., and K.L. performed murine experiments and
F.D. & M.G. analyzed the data. Y.K., S.S., F.D., and M.G. designed and per-
formed confocal microscopy and analysis. P.K., Y.M., and S.S. provided tumor
cell lines and transgenic mice. A.M., D.G., C.V. J.T. T.B., and E.M. coordinated
inclusion and selection of clinical samples. T.B., J.T., E.M. K.L., F.D., and M.G.
designed, performed, and analyzed PLA and cmIHC assays. L.B. provided
therapeutic anti-PD-L1 antibody. F.D. M.G. T.H., and J.A. wrote the manu-
script. All authors were involved in the critical review of the manuscript. All au-
thors read and approved the manuscript.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: March 26, 2020
Revised: June 28, 2020
Accepted: August 31, 2020
Published: October 1, 2020

698 Cancer Cell 38, 685–700, November 9, 2020


Article
REFERENCES
Beltra, J.C., Manne, S., Abdel-Hakeem, M.S., Kurachi, M., Giles, J.R., Chen,
Z., Casella, V., Ngiow, S.F., Khan, O., Huang, Y.J., et al. (2020).
Developmental relationships of four exhausted CD8(+) T cell subsets reveals
underlying transcriptional and epigenetic landscape control mechanisms.
Immunity 52, 825–841.e8.
Bosteels, C., Neyt, K., Vanheerswynghels, M., van Helden, M.J., Sichien, D.,
Debeuf, N., De Prijck, S., Bosteels, V., Vandamme, N., Martens, L., et al.
(2020). Inflammatory type 2 cDCs acquire features of cDC1s and macro-
phages to orchestrate immunity to respiratory virus infection. Immunity 52,
1039–1056.e9.
Chamoto, K., Chowdhury, P.S., Kumar, A., Sonomura, K., Matsuda, F.,
Fagarasan, S., and Honjo, T. (2017). Mitochondrial activation chemicals syner-
gize with surface receptor PD-1 blockade for T cell-dependent antitumor ac-
tivity. Proc. Natl. Acad. Sci. U S A 114, E761–E770.
Dammeijer, F., Lau, S.P., van Eijck, C.H.J., van der Burg, S.H., and Aerts, J.
(2017a). Rationally combining immunotherapies to improve efficacy of immune
checkpoint blockade in solid tumors. Cytokine Growth Factor Rev. 36, 5–15.
Dammeijer, F., Lievense, L.A., Kaijen-Lambers, M.E., van Nimwegen, M.,
Bezemer, K., Hegmans, J.P., van Hall, T., Hendriks, R.W., and Aerts, J.G.
(2017b). Depletion of tumor-associated macrophages with a CSF-1R kinase
inhibitor enhances antitumor immunity and survival induced by DC immuno-
therapy. Cancer Immunol. Res. 5, 535–546.
Daud, A.I., Wolchok, J.D., Robert, C., Hwu, W.J., Weber, J.S., Ribas, A., Hodi,
F.S., Joshua, A.M., Kefford, R., Hersey, P., et al. (2016). Programmed death-
ligand 1 expression and response to the anti-programmed death 1 antibody
pembrolizumab in melanoma. J. Clin. Oncol. 34, 4102–4109.
Eggermont, A.M.M., Robert, C., and Ribas, A. (2018). The new era of adjuvant
therapies for melanoma. Nat. Rev. Clin. Oncol. 15, 535–536.
Eisenbarth, S.C. (2019). Dendritic cell subsets in T cell programming: location
dictates function. Nat. Rev. Immunol. 19, 89–103.
Fransen, M.F., Schoonderwoerd, M., Knopf, P., Camps, M.G., Hawinkels, L.J.,
Kneilling, M., van Hall, T., and Ossendorp, F. (2018). Tumor-draining lymph no-
des are pivotal in PD-1/PD-L1 checkpoint therapy. JCI Insight 3, e124507.
Gandhi, L., Rodriguez-Abreu, D., Gadgeel, S., Esteban, E., Felip, E., De
Angelis, F., Domine, M., Clingan, P., Hochmair, M.J., Powell, S.F., et al.
(2018). Pembrolizumab plus chemotherapy in metastatic non-small-cell lung
cancer. N. Engl. J. Med. 378, 2078–2092.
Glodde, N., Bald, T., van den Boorn-Konijnenberg, D., Nakamura, K.,
O’Donnell, J.S., Szczepanski, S., Brandes, M., Eickhoff, S., Das, I., Shridhar,
N., et al. (2017). Reactive neutrophil responses dependent on the receptor
tyrosine kinase c-MET limit cancer immunotherapy. Immunity 47, 789–802.e9.
Gray, E.E., and Cyster, J.G. (2012). Lymph node macrophages. J. Innate
Immun. 4, 424–436.
Havel, J.J., Chowell, D., and Chan, T.A. (2019). The evolving landscape of bio-
markers for checkpoint inhibitor immunotherapy. Nat. Rev. Cancer 19,
133–150.
Hellmann, M.D., Paz-Ares, L., Bernabe Caro, R., Zurawski, B., Kim, S.W.,
Carcereny Costa, E., Park, K., Alexandru, A., Lupinacci, L., de la Mora
Jimenez, E., et al. (2019). Nivolumab plus ipilimumab in advanced non-
small-cell lung cancer. N. Engl. J. Med. 381, 2020–2031.
Herbst, R.S., Soria, J.C., Kowanetz, M., Fine, G.D., Hamid, O., Gordon, M.S.,
Sosman, J.A., McDermott, D.F., Powderly, J.D., Gettinger, S.N., et al. (2014).
Predictive correlates of response to the anti-PD-L1 antibody MPDL3280A in
cancer patients. Nature 515, 563–567.
Hodi, F.S., Chiarion-Sileni, V., Gonzalez, R., Grob, J.J., Rutkowski, P., Cowey,
C.L., Lao, C.D., Schadendorf, D., Wagstaff, J., Dummer, R., et al. (2018).
Nivolumab plus ipilimumab or nivolumab alone versus ipilimumab alone in
advanced melanoma (CheckMate 067): 4-year outcomes of a multicentre,
randomised, phase 3 trial. Lancet Oncol. 19, 1480–1492.
Hope, J.L., Spantidea, P.I., Kiernan, C.H., Stairiker, C.J., Rijsbergen, L.C., van
Meurs, M., Brouwers-Haspels, I., Mueller, Y.M., Nelson, D.J., Bradley, L.M.,
ll
et al. (2019). Microenvironment-dependent gradient of CTL exhaustion in the
AE17sOVA murine mesothelioma tumor model. Front. Immunol. 10, 3074.
Huang, A.C., Orlowski, R.J., Xu, X., Mick, R., George, S.M., and Yan, P.K.
(2019). A single dose of neoadjuvant PD-1 blockade predicts clinical outcomes
in resectable melanoma. Nat. Med. 25, 454–461.
Hui, E., Cheung, J., Zhu, J., Su, X., Taylor, M.J., Wallweber, H.A., Sasmal, D.K.,
Huang, J., Kim, J.M., Mellman, I., and Vale, R.D. (2017). T cell costimulatory
receptor CD28 is a primary target for PD-1-mediated inhibition. Science 355,
1428–1433.
Jansen, C.S., Prokhnevska, N., Master, V.A., Sanda, M.G., Carlisle, J.W.,
Bilen, M.A., Cardenas, M., Wilkinson, S., Lake, R., Sowalsky, A.G., et al.
(2019). An intra-tumoral niche maintains and differentiates stem-like CD8
T cells. Nature 576, 465–470.
Kamphorst, A.O., Wieland, A., Nasti, T., Yang, S., Zhang, R., Barber, D.L.,
Konieczny, B.T., Daugherty, C.Z., Koenig, L., Yu, K., et al. (2017). Rescue of
exhausted CD8 T cells by PD-1-targeted therapies is CD28-dependent.
Science 355, 1423–1427.
Kleinovink, J.W., Marijt, K.A., Schoonderwoerd, M.J.A., van Hall, T.,
Ossendorp, F., and Fransen, M.F. (2017). PD-L1 expression on malignant cells
is no prerequisite for checkpoint therapy. Oncoimmunology 6, e1294299.
Kool, M., Soullie, T., van Nimwegen, M., Willart, M.A., Muskens, F., Jung, S.,
Hoogsteden, H.C., Hammad, H., and Lambrecht, B.N. (2008). Alum adjuvant
boosts adaptive immunity by inducing uric acid and activating inflammatory
dendritic cells. J. Exp. Med. 205, 869–882.
Koyama, S., Akbay, E.A., Li, Y.Y., Herter-Sprie, G.S., Buczkowski, K.A.,
Richards, W.G., Gandhi, L., Redig, A.J., Rodig, S.J., Asahina, H., et al.
(2016). Adaptive resistance to therapeutic PD-1 blockade is associated with
upregulation of alternative immune checkpoints. Nat. Commun. 7, 10501.
Kurtulus, S., Madi, A., Escobar, G., Klapholz, M., Nyman, J., Christian, E.,
Pawlak, M., Dionne, D., Xia, J., Rozenblatt-Rosen, O., et al. (2019).
Checkpoint blockade immunotherapy induces dynamic changes in PD-1(-)
CD8(+) tumor-infiltrating T cells. Immunity 50, 181–194.e6.
Lau, J., Cheung, J., Navarro, A., Lianoglou, S., Haley, B., Totpal, K., Sanders,
L., Koeppen, H., Caplazi, P., McBride, J., et al. (2017). Tumour and host cell
PD-L1 is required to mediate suppression of anti-tumour immunity in mice.
Nat. Commun. 8, 14572.
Li, H., van der Leun, A.M., Yofe, I., Lubling, Y., Gelbard-Solodkin, D., van
Akkooi, A.C.J., van den Braber, M., Rozeman, E.A., Haanen, J., Blank, C.U.,
et al. (2019). Dysfunctional CD8 T cells form a proliferative, dynamically regu-
lated compartment within human melanoma. Cell 176, 775–789.e18.
Lin, H., Wei, S., Hurt, E.M., Green, M.D., Zhao, L., Vatan, L., Szeliga, W.,
Herbst, R., Harms, P.W., Fecher, L.A., et al. (2018). Host expression of PD-
L1 determines efficacy of PD-L1 pathway blockade-mediated tumor regres-
sion. J. Clin. Invest. 128, 1708.
Long, G.V., Dummer, R., Hamid, O., Gajewski, T.F., Caglevic, C., Dalle, S.,
Arance, A., Carlino, M.S., Grob, J.J., Kim, T.M., et al. (2019). Epacadostat
plus pembrolizumab versus placebo plus pembrolizumab in patients with un-
resectable or metastatic melanoma (ECHO-301/KEYNOTE-252): a phase 3,
randomised, double-blind study. Lancet Oncol. 20, 1083–1097.
Maier, B., Leader, A.M., Chen, S.T., Tung, N., Chang, C., LeBerichel, J.,
Chudnovskiy, A., Maskey, S., Walker, L., Finnigan, J.P., et al. (2020). A
conserved dendritic-cell regulatory program limits antitumour immunity.
Nature 580, 257–262.
Morrison, C., Pabla, S., Conroy, J.M., Nesline, M.K., Glenn, S.T., Dressman,
D., Papanicolau-Sengos, A., Burgher, B., Andreas, J., Giamo, V., et al.
(2018). Predicting response to checkpoint inhibitors in melanoma beyond
PD-L1 and mutational burden. J. Immunother. Cancer 6, 32.
Nishino, M., Ramaiya, N.H., Hatabu, H., and Hodi, F.S. (2017). Monitoring im-
mune-checkpoint blockade: response evaluation and biomarker develop-
ment. Nat. Rev. Clin. Oncol. 14, 655–668.
Oh, S.A., Wu, D.C., Cheung, J., Navarro, A., Xiong, H., Cubas, R., Totpal, K.,
Chiu, H., Wu, Y., Comps-Agrar, L., et al. (2020). PD-L1 expression by dendritic
cells is a key regulator of T-cell immunity in cancer. Nat. Cancer 1, 681–691.
Cancer Cell 38, 685–700, November 9, 2020 699
 ll
Ohl, L., Mohaupt, M., Czeloth, N., Hintzen, G., Kiafard, Z., Zwirner, J.,
Blankenstein, T., Henning, G., and Förster, R. (2004). CCR7 governs skin den-
dritic cell migration under inflammatory and steady-state conditions. Immunity
21, 279–288.
Paz-Ares, L., Luft, A., Vicente, D., Tafreshi, A., Gumus, M., Mazieres, J.,
Hermes, B., Cay Senler, F., Csoszi, T., Fulop, A., et al. (2018).
Pembrolizumab plus chemotherapy for squamous non-small-cell lung cancer.
N. Engl. J. Med. 379, 2040–2051.
Ready, N., Hellmann, M.D., Awad, M.M., Otterson, G.A., Gutierrez, M., Gainor,
J.F., Borghaei, H., Jolivet, J., Horn, L., Mates, M., et al. (2019). First-line nivo-
lumab plus ipilimumab in advanced non-small-cell lung cancer (CheckMate
568): outcomes by programmed death ligand 1 and tumor mutational burden
as biomarkers. J. Clin. Oncol. 37, 992–1000.
Ribas, A., and Wolchok, J.D. (2018). Cancer immunotherapy using checkpoint
blockade. Science 359, 1350–1355.
Ruhland, M.K., Roberts, E.W., Cai, E., Mujal, A.M., Marchuk, K., Beppler, C.,
Nam, D., Serwas, N.K., Binnewies, M., and Krummel, M.F. (2020).
Visualizing synaptic transfer of tumor antigens among dendritic cells. Cancer
Cell 37, 786–799.e5.
Sade-Feldman, M., Yizhak, K., Bjorgaard, S.L., Ray, J.P., de Boer, C.G.,
Jenkins, R.W., Lieb, D.J., Chen, J.H., Frederick, D.T., Barzily-Rokni, M.,
et al. (2018). Defining T cell states associated with response to checkpoint
immunotherapy in melanoma. Cell 175, 998–1013.e20.
Samstein, R.M., Lee, C.-H., Shoushtari, A.N., Hellmann, M.D., Shen, R.,
Janjigian, Y.Y., Barron, D.A., Zehir, A., Jordan, E.J., Omuro, A., et al. (2019).
Tumor mutational load predicts survival after immunotherapy across multiple
cancer types. Nat. Genet. 51, 202–206.
Schadendorf, D., Fisher, D.E., Garbe, C., Gershenwald, J.E., Grob, J.J.,
Halpern, A., Herlyn, M., Marchetti, M.A., McArthur, G., Ribas, A., et al.
(2015). Melanoma. Nat. Rev. Dis. Primers 1, 15003.
Shi, J., Hou, S., Fang, Q., Liu, X., Liu, X., and Qi, H. (2018). PD-1 controls follic-
ular T helper cell positioning and function. Immunity 49, 264–274.e4.
Siddiqui, I., Schaeuble, K., Chennupati, V., Fuertes Marraco, S.A., Calderon-
Copete, S., Pais Ferreira, D., Carmona, S.J., Scarpellino, L., Gfeller, D.,
Pradervand, S., et al. (2019). Intratumoral Tcf1(+)PD-1(+)CD8(+) T cells with
stem-like properties promote tumor control in response to vaccination and
checkpoint blockade immunotherapy. Immunity 50, 195–211.e10.
Spitzer, M.H., Carmi, Y., Reticker-Flynn, N.E., Kwek, S.S., Madhireddy, D.,
Martins, M.M., Gherardini, P.F., Prestwood, T.R., Chabon, J., Bendall, S.C.,
et al. (2017). Systemic immunity is required for effective cancer immuno-
therapy. Cell 168, 487–502.e15.
700 Cancer Cell 38, 685–700, November 9, 2020
Article
Tang, H., Liang, Y., Anders, R.A., Taube, J.M., Qiu, X., Mulgaonkar, A., Liu, X.,
Harrington, S.M., Guo, J., Xin, Y., et al. (2018). PD-L1 on host cells is essential
for PD-L1 blockade-mediated tumor regression. J. Clin. Invest. 128, 580–588.
van der Ploeg, A.P., van Akkooi, A.C., Schmitz, P.I., Koljenovic, S., Verhoef, C.,
and Eggermont, A.M. (2010). EORTC Melanoma Group sentinel node protocol
identifies high rate of submicrometastases according to Rotterdam Criteria.
Eur. J. Cancer 46, 2414–2421.
van Rooijen, N., and Hendrikx, E. (2010). Liposomes for specific depletion of
macrophages from organs and tissues. Methods Mol. Biol. 605, 189–203.
Verver, D., van Klaveren, D., Franke, V., van Akkooi, A.C.J., Rutkowski, P.,
Keilholz, U., Eggermont, A.M.M., Nijsten, T., Grunhagen, D.J., and Verhoef,
C. (2019). Development and validation of a nomogram to predict recurrence
and melanoma-specific mortality in patients with negative sentinel lymph no-
des. Br. J. Surg. 106, 217–225.
Verver, D., van Klaveren, D., van Akkooi, A.C.J., Rutkowski, P., Powell, B.,
Robert, C., Testori, A., van Leeuwen, B.L., van der Veldt, A.A.M., Keilholz,
U., et al. (2018). Risk stratification of sentinel node-positive melanoma patients
defines surgical management and adjuvant therapy treatment considerations.
Eur. J. Cancer 96, 25–33.
von Schuckmann, L.A., Hughes, M.C.B., Ghiasvand, R., Malt, M., van der Pols,
J.C., Beesley, V.L., Khosrotehrani, K., Smithers, B.M., and Green, A.C. (2019).
Risk of melanoma recurrence after diagnosis of a high-risk primary tumor.
JAMA Dermatol. 155, 688–693.
Wculek, S.K., Cueto, F.J., Mujal, A.M., Melero, I., Krummel, M.F., and Sancho,
D. (2020). Dendritic cells in cancer immunology and immunotherapy. Nat. Rev.
Immunol. 20, 7–24.
Yanagawa, Y., Sugahara, K., Kataoka, H., Kawaguchi, T., Masubuchi, Y., and
Chiba, K. (1998). FTY720, a novel immunosuppressant, induces sequestration
of circulating mature lymphocytes by acceleration of lymphocyte homing in
rats. II. FTY720 prolongs skin allograft survival by decreasing T cell infiltration
into grafts but not cytokine production in vivo. J. Immunol. 160, 5493–5499.
Yost, K.E., Satpathy, A.T., Wells, D.K., Qi, Y., Wang, C., Kageyama, R.,
McNamara, K.L., Granja, J.M., Sarin, K.Y., Brown, R.A., et al. (2019). Clonal
replacement of tumor-specific T cells following PD-1 blockade. Nat. Med.
25, 1251–1259.
Zhao, M., Kiernan, C.H., Stairiker, C.J., Hope, J.L., Leon, L.G., van Meurs, M.,
Brouwers-Haspels, I., Boers, R., Boers, J., Gribnau, J., et al. (2020). Rapid
in vitro generation of bona fide exhausted CD8+ T cells is accompanied by
Tcf7 promotor methylation. PLoS Pathog. 16, e1008555.
 stands for interdisciplinary

iScience is an interdisciplinary
open access journal from Cell Press
iScienceW\ISPZOLZIHZPJHUKHWWSPLKYLZLHYJO[OH[HK]HUJLZHZWLJPÄJÄLSKHJYVZZSPMLWO`ZPJHS
HUKLHY[OZJPLUJLZ0[»ZHUVWLUHJJLZZQV\YUHS^P[OJVU[PU\V\ZW\ISPJH[PVUZVYLZLHYJOPZ
PTTLKPH[LS`HJJLZZPISL6\YUVUVUZLUZLHWWYVHJO[VZ\ITPZZPVUZPZZPTWSLMHZ[HUKMHPYHUK
V\YJVTTP[TLU[[VPU[LNYP[`TLHUZ^LW\ISPZO[YHUZWHYLU[TL[OVKZ^P[OOPNOLKP[VYPHSZ[HUKHYKZ

So if you’re looking for a place to be inspired by science, welcome home.

Submit your paper to iScience.

cell.com/iscience
What is your superpower?
Gene functional analysis is ours.

YOUR INPUT YOUR BRIDGE TO DISCOVERY YOUR RESULTS

Cell or Animal Models • CRISPR / RNAi Libraries & Genetic Screens Functionally Important Genes
Biological Samples • DriverMap™ Targeted RNA Expression Profiling & Biomarkers
• CloneTracker™ Barcode Libraries

Real expertise delivering results. Learn more at cellecta.com.

Who we are
Cellecta is a leading provider of genomic products and services. Our
functional genomics portfolio includes gene knockout and knockdown
screens, custom and genome-wide CRISPR and RNAi libraries, construct
services, cell engineering, NGS kits and targeted expression profiling
products and services.

We can advance your discovery efforts.

www.cellecta.com info@cellecta.com +1 877-938-3910 or +1 650-938-3910 © 2020 Cellecta, Inc. 320 Logue Ave. Mountain View, CA 94043 USA
THE DIFFERENCE OF
BREAKING THROUGH BARRIERS
WITH VIBRANCE

ENABLING GREATER EXPERIMENTAL INSIGHTS WITH

THE POWER OF BD HORIZON™ RED 718 REAGENTS
Innovative small-molecule fluorochrome designed to provide improved
sensitivity and resolving power to help analyze low-density markers
without compromising quality.
Excited by the red laser and with an emission maximum of 718 nm,
BD Horizon™ Red 718 Reagents offer:
• Greater brightness and resolving power than Alexa Fluor® 700 for CD4+ IL-17+ cells are better resolved with BD Horizon™
both surface and intracellular markers R718 Reagents than Alexa Fluor® 700.

• Reduced background staining compared to Alexa Fluor® 700 and

APC-R700
• Minimal spread into the APC channel
• A broad portfolio of conjugates through BD OptiBuild™ On-Demand
Reagents

BD Horizon™ Red 718 Reagents can help resolve difficult to analyze

BD Horizon™ Red 718 Reagents can resolve FoxP3+ cells.
markers and are the reagents of choice for the 700 nm detector.

Switch to BD Horizon™ Red 718 Reagents to help achieve your resolution goals.

Learn more at bdbiosciences.com/R718

Class I Laser Products. For Research Use Only. Not for use in diagnostic or therapeutic procedures.
23-22635-00

BD, the BD Logo, Horizon, OptiBuild and Pharmingen are trademarks of Becton, Dickinson and Company or its affiliates.
All other trademarks are the property of their respective owners. © 2020 BD. All rights reserved.
 I S R U P T IN
DE THERAP RUG DIGSC
EUTIC D
N

OV
U
IMM

ERY

OX I I E S
J A X NGL E

OT T U D
CI T Y
A SI
CY T

KIN
ON N

S
O

ST E R
UD AT I
Y TO ELEASE EVALU IMMU
TEST EFFICACY AND

USE A SINGLE HUMAN PBMC DONOR USE MULTIPLE HUMAN PBMC DONORS
to perform dose ranging studies and to recapitulate donor-to-donor variability
to allow lead selection. you’d expect to find in the clinic.

CYTOKINE RELEASE BUILD ON THE OPTIMAL IMMUNODEFICIENT

PLATFORM

EVALUATION STUDIES to investigate your specific targets of interest.

Utilize this platform to evaluate

immunomodulation of human immune cells
YOUR HUMANIZED PLATFORM
against human specific targets. This assay is ideal for evaluating relative cytokine levels from
provides a fast, accurate, and reliable method to immunomodulating therapies and for assessing
anti-inflammatory therapies.
screen novel immunotherapies for efficacy and
cytokine release syndrome.

SHAPE YOUR DISCOVERY AT

JAX.ORG/CRES