viruses
Communication
Computer-Aided Directed Evolution Generates Novel AAV
Variants with High Transduction Efficiency
Zengpeng Han 1,2,3,† , Nengsong Luo 4,† , Fei Wang 5 , Yuxiang Cai 4 , Xin Yang 6 , Weiwei Feng 6 ,
Zhenxiang Zhu 1,2,3 , Jie Wang 1,2 , Yang Wu 1,2 , Chaohui Ye 1,2 , Kunzhang Lin 3, * and Fuqiang Xu 1,2,3,4,7, *
Citation: Han, Z.; Luo, N.; Wang, F.;
Cai, Y.; Yang, X.; Feng, W.; Zhu, Z.;
Wang, J.; Wu, Y.; Ye, C.; et al.
Computer-Aided Directed Evolution
Generates Novel AAV Variants with
High Transduction Efficiency. Viruses
2023, 15, 848. https://doi.org/10.3390/
v15040848
Academic Editor: Hernan
Garcia-Ruiz
Received: 31 January 2023
Revised: 20 March 2023
Accepted: 24 March 2023
Published: 26 March 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Viruses 2023, 15, 848. https://doi.org/10.3390/v15040848
1 State Key Laboratory of Magnetic Resonance and Atomic and Molecular Physics, Key Laboratory of Magnetic
Resonance in Biological Systems, Wuhan Center for Magnetic Resonance, Innovation Academy for Precision
Measurement Science and Technology, Chinese Academy of Sciences, Wuhan 430071, China
2 University of Chinese Academy of Sciences, Beijing 100049, China
3 Shenzhen Key Laboratory of Viral Vectors for Biomedicine, Key Laboratory of Quality Control Technology for
Virus-Based Therapeutics, Guangdong Provincial Medical Products Administration, NMPA Key Laboratory
for Research and Evaluation of Viral Vector Technology in Cell and Gene Therapy Medicinal Products, the
Brain Cognition and Brain Disease Institute (BCBDI), Shenzhen Institute of Advanced Technology, Chinese
Academy of Sciences, Shenzhen-Hong Kong Institute of Brain Science-Shenzhen Fundamental Research
Institutions, Shenzhen 518055, China
4 Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology,
Wuhan 430074, China
5 iDrugDance Technology Co., Ltd., Shenzhen 518055, China
6 ZhunAo Biotechnology Co., Ltd., Wuhan 430056, China
7 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences,
Shanghai 200031, China
* Correspondence: kz.lin@siat.ac.cn (K.L.); fq.xu@siat.ac.cn (F.X.); Tel.: +86-150-7244-8709 (K.L.);
+86-139-9550-5336 (F.X.)
† These authors contributed equally to this work.
Abstract: Adeno-associated viruses (AAVs) have become safe and effective tools for therapeutic
in vivo gene drug delivery. Among many AAV serotypes, AAV2 is the most well-characterized.
Although many studies have been carried out on the engineering of the capsid VR-VIII region, few
attempts have been made in the VR-IV region. Here, we targeted amino acid positions 442–469 of
the VR-IV region and established an engineering paradigm of computer-aided directed evolution,
based on training samples from previous datasets, to obtain a viral vector library with high diversity
(~95,089). We further examined two variants selected from the library. The transduction efficiency of
these two novel AAV variants, AAV2.A1 and AAV2.A2, in the central nervous system was 10–15 times
higher than that of AAV2. This finding provides new vehicles for delivering gene drugs to the brain.
Keywords: adeno-associated virus; directed evolution; computer-aided design; Cap gene library;
transduction efficiency
1. Introduction
Adeno-associated viruses (AAVs) have become safe and effective tools for the in vivo
delivery of therapeutic gene drugs due to their relatively low immune responses and
broad tissue tropism [1,2]. The authorization of five AAV-based gene therapy drugs,
comprising Glybera (AAV1), Luxturna (AAV2), Zolgensma (AAV9), Upstaza (AAV2),
and Roctavian (AAV5), clearly demonstrated the enormous potential of AAV vectors for
human gene therapy [3]. However, with a few exceptions, the transduction efficiency and
specificity of natural capsids remain issues that impede certain critical applications and
limit therapeutic purposes [4,5]. As synthetic biology progresses, the AAV capsid can be
engineered to overcome these challenges through directed evolution, which is rooted in the
simulation of natural evolution, thus generating genetic variants with enhanced specificity
and advantageous characteristics through the process of mutation and selection [4,6].
https://www.mdpi.com/journal/viruses
Viruses 2023, 15, 848 2 of 14

Among many AAV serotypes, AAV2 is the most well-characterized; its rep gene and
inverted terminal repeat (ITR) sequences are commonly used in the preparation of other
serotypes of recombinant AAV [7]. The AAV capsid protein determines most of the proper-
ties of the vector. Several methods have been developed for the construction of a diversified
Cap gene mutation library, including global Cap gene libraries implemented by DNA shuf-
fling [6,8] or random point mutations [9] and site-specific libraries implemented by peptide
display [10–12] or computer-aided design. In site-specific library design, the capsid VR-VIII
region, especially the position between N587 and R588 at the spike around the 3-fold
axis, has been most commonly used as a mutation introduction site in AAV2 [10,12–14].
Since this region is highly variable and on the outermost side of the capsid, it is suitable
for the functional display of inserted peptides to retarget AAV2 to poorly susceptible
cells [12,15–17]. Apart from this site, attempts to construct libraries in other regions of the
Cap gene are rarely reported.
To better inform AAV capsid engineering, Ogden et al. [7] systematically examined the
effects of all single-codon substitutions, insertions, and deletions across all 735 positions
in the AAV2 Cap gene. On the one hand, some low-conserved regions suitable for library
construction were revealed (e.g., positions 442–469 and 561–588) by the AAV2 capsid fitness
landscape [7], and on the other hand, this information provides a training set for machine
learning-aided library design. Thus, they focused on positions 561–588 and used these data
to design mutant libraries based on machine learning [7,18].
Here, we targeted another position, 442–469, and established an engineering paradigm
of computer-aided directed evolution based on training samples from the previous two
public datasets [7,18]; then, we obtained a viral vector library with high diversity. As
proof of principle, we further screened mutants from the library with higher transduction
efficiency than AAV2, providing new vehicles for gene drug delivery in the brain.

2. Materials and Methods

2.1. Computer-Aided Design of the AAV2 Mutant Library
The most commonly modified position within the AAV2 capsid protein is the surface-
exposed loop (Figure 1A), which contains amino acids (AAs) 561–588, because it is the
site of heparan sulfate binding in AAV2 [19] and is suitable for peptide display [10,12–14].
Similarly, AA442–469 generated another highly variable spike near the 3-fold axis in
addition to the loop generated by AA561–588 (Figure 1B).
The training samples for our deep models are generated based on two public datasets
described in previous work [7,18]. Ogden et al. [7] generated all single-codon mutants of the
AAV2 cap gene and examined the effects of mutations systematically across all 735 positions.
Their comprehensive codon-scanning approach provided sufficient information for the
training of our models. Bryant et al. [18] generated multi-mutations at positions 561–588,
which provided a multi-mutation training set for our model.
To verify our deep learning models, we generated test sequences following the scheme
described in [7,18], which employs random mutations and rational design mutations on the
wild-type AAV2 Cap sequence, including amino acid substitutions, insertions, or deletions
involving at least 1 site and up to 5 sites at positions AA442–469. Since it is impractical to
generate all possible variants exhaustively, we sampled about 50,000,000 variants from the
entire search space, which was sufficient for the verification and potential valuable sequence
discovery. These sequences were then embedded into a high-dimension space using an
embedding matrix. At this stage, our goal was to predict whether the variants could be
assembled to the level of the wild-type AAV2 capsid. We designed several transformer-
based and convolutional neural network (CNN)-based models to predict the possibility of
the successful assembly of the variants, and the results are ensembled (Figure 1C). Here, we
have to emphasize that, in our multi-mutation sequence generation and experiments, our
target positions were AA442–469—which were completely different from the AA561–588
described in previous works and have no overlap with the previous libraries. The highly
dimensional data of the generated sequences were then fed into the models for prediction.
 several transformer-based and convolutional neural network (CNN)-based models to pre-
dict the possibility of the successful assembly of the variants, and the results are ensem-
bled (Figure 1C). Here, we have to emphasize that, in our multi-mutation sequence gen-
eration and experiments, our target positions were AA442–469—which were completely
different from the AA561–588 described in previous works and have no overlap with the
Viruses 2023, 15, 848 3 of 14
previous libraries. The highly dimensional data of the generated sequences were then fed
into the models for prediction.

Figure 1. Capsid engineering locations and diagram of computer-aided design. (A) AAV2 capsid
Figure 1. Capsid engineering locations and diagram of computer-aided design. (A) AAV2 capsid
surface model demonstrates the location of the protruding loop structures in milky. The icosahedral
surface model demonstrates the location of the protruding loop structures in milky. The icosahedral
3- and 5-fold axes are indicated. (B) Left, an exterior capsid surface representation for AAV2 VP3
trimer. Right, an3-enlarged
and 5-fold axes
side viewareofindicated.
the capsid(B) Left, the
shows an exterior capsidby
spike created surface representation for AAV2 VP3
the AA442–AA469
trimer. Right, an
(green) and the AA561–AA588 enlarged
(red) variableside viewatofthe
regions theicosahedral
capsid shows theaxis.
3-fold spike created
AA, aminoby the AA442–AA469
acid.
The figure (A,B) images were generated using UCSF ChimeraX v1.4 [20,21] (UCSF, San Francisco, AA, amino acid.
(green) and the AA561–AA588 (red) variable regions at the icosahedral 3-fold axis.
The figure
CA, USA). (C) Schematic (A,B) images
diagram were generated
of computer-aided using UCSF ChimeraX v1.4 [20,21] (UCSF, San Francisco,
design.
CA, USA). (C) Schematic diagram of computer-aided design.
2.2. Construction of Plasmid Library using Two-Step Cloning
2.2. Construction of Plasmid Library Using Two-Step Cloning
Cap library fragments were synthesized on a customed oligo array (GENEWIZ, Su-
Cap library fragments were synthesized on a customed oligo array (GENEWIZ, Suzhou,
zhou, China), and each specific design sequence corresponded to a unique barcode (BC)
China), and each specific design sequence corresponded to a unique barcode (BC) for
for subsequent sequencing, with two BsaI restriction sites incorporated between the Cap
subsequent sequencing, with two BsaI restriction sites incorporated between the Cap library
library and the BC for Golden Gate assembly (Figure 2A). We pre-built the “recipient plas-
and the BC for Golden Gate assembly (Figure 2A). We pre-built the “recipient plasmid” as a
mid” as a backbone, which contained the 1–441 AA sequence of the AAV2 Cap gene. To
backbone, which contained the 1–441 AA sequence of the AAV2 Cap gene. To complete the
complete the plasmid library construction, the first cloning step was to insert the synthe-
plasmid library construction, the first cloning step was to insert the synthesized Cap library
sized Cap library fragment onto the recipient plasmid through Gibson assembly using the
fragment onto the recipient plasmid through Gibson assembly using the Gibson Assembly
Gibson Assembly Master Mix (New England Biolabs, Ipswich, MA, USA, E2611). The sec-
Master Mix (New England Biolabs, Ipswich, MA, USA, E2611). The second cloning step
ond cloning step was to amplify the 470–735 AA fragments of the AAV2 Cap gene from
was to amplify the 470–735 AA fragments of the AAV2 Cap gene from pAAV2/2 (Addgene,
pAAV2/2 (Addgene, Watertown,
Watertown, MA, USA,MA,104,963)
USA, 104,963) and,tothen,
and, then, inserttothem
insertinto
them
theinto the
backbone via Golden
backbone via Golden Gate assembly
Gate assembly using Golden
using Golden Gate Assembly
Gate Assembly Mix (New MixEngland
(New England
Biolabs, Ipswich, MA,
USA, E1601) to generate a complete Cap reading frame (Figure 2A). The primers NGS-F/R
(Supplementary Table S1) were used to amplify the barcodes for the sequencing assay.
To provide the AAV2 Rep and assembly-activating protein (AAP), the construction
of Rep-AAP helper was implemented on pAAV2/2 based on the design of Deverman
et al. [11]. Five stop codons were introduced in the AAV2 Cap gene (at VP1 AAs: 6, 10, 142,
152, and 216) to eliminate VP1, VP2, and VP3 capsid protein expression; the stop codon at
AA216 was designed to not disturb the codon in the alternative reading frame of the AAP
protein (Figure 2B).
 the sequencing assay.
To provide the AAV2 Rep and assembly-activating protein (AAP), the construction
of Rep-AAP helper was implemented on pAAV2/2 based on the design of Deverman et
al. [11]. Five stop codons were introduced in the AAV2 Cap gene (at VP1 AAs: 6, 10, 142,
152, and 216) to eliminate VP1, VP2, and VP3 capsid protein expression; the stop codon at
Viruses 2023, 15, 848 4 of 14
AA216 was designed to not disturb the codon in the alternative reading frame of the AAP
protein (Figure 2B).

Figure 2. Construction of plasmid systems for AAV library preparation. (A) The plasmid library
Figure 2. Construction of plasmid systems for AAV library preparation. (A) The plasmid library was
was constructed using two-step cloning. Step 1, Cap gene library fragments, generated by custom-
constructed
ized oligonucleotide using
synthesis, two-step
were insertedcloning.
into theStep 1, Capbackbone
recipient gene library fragments,
plasmid through generated
Gibson by customized
oligonucleotide synthesis, were inserted into the recipient backbone
assembly. Step 2, AA470–735 fragments of the AAV2 Cap gene were inserted into the backbone plasmid through
via Gibson assem-
bly. Step 2, AA470–735 fragments of the AAV2 Cap gene were inserted
Golden Gate assembly to generate a complete Cap reading frame. AA, amino acid. (B) The schematicinto the backbone via Golden
shows the genomeGatestructure
assemblyoftothe
generate
Rep-AAP a complete Cap reading
helper plasmid. Stopframe.
codonsAA, amino
were acid. (B)
inserted intoThe
theschematic shows
Cap gene to terminate the expression
the genome structure of
of the
the VP1, VP2, helper
Rep-AAP and VP3 proteins,
plasmid. while
Stop the expression
codons were inserted of the
into the Cap gene to
AAP protein was not affected.
terminate (C) An AAV
the expression mutant
of the VP1, library was
VP2, and produced
VP3 proteins,using
whilethe
thetriple-transfec-
expression of the AAP protein
tion method in HEK293T cells. (C) An AAV mutant library was produced using the triple-transfection method in
was not affected.
HEK293T cells.
2.3. Cell Culture, AAV Preparation, and In Vitro Transduction Assay
HEK293T2.3.cellsCell
wereCulture, AAV Preparation,
purchased and In Vitro
from the American Type Transduction Assay (Manas-
Culture Collection
sas, VA, USA). DailyHEK293T cells were
maintenance, purchased from
cell preparation priorthe American
to viral Type Culture
packaging, and the Collection
triple- (Manassas,
VA, USA).
plasmid transfection Daily maintenance,
of conventional cellcompleted
viruses were preparation as prior to viral
previously packaging,
described [22]. and the triple-
For AAV plasmid
mutant transfection of conventional
library production, viruses
the plasmid were completed
library, together withas previously
Rep-AAP described [22].
helper and pAdDeltaF6For AAV mutant
(Addgene, library production,
Watertown, MA, USA,the plasmid
112,867), library, together
constituted a specialwith Rep-AAP
helper and pAdDeltaF6 (Addgene, Watertown, MA, USA,
triple-plasmid system (Figure 2C). For individual AAV packaging, the AAV2 plasmid 112,867), constituted a special
triple-plasmid system (Figure 2C). For individual AAV
(pAAV2/2) was obtained from Addgene (Watertown, MA, USA, 104,963), while mutant packaging, the AAV2 plasmid
AAV plasmids were derived from the AAV2 plasmid using the Fast Mutagenesis System while mutant
(pAAV2/2) was obtained from Addgene (Watertown, MA, USA, 104,963),
AAV plasmids
(TransGen Biotech, were derived
Beijing, China, FM111).from the AAV2
The primer plasmid(Mu-A1-F/R
sequences using the FastandMutagenesis
Mu- System
(TransGen Biotech,
A2-F/R) for site-directed Beijing,are
mutagenesis, China,
shownFM111). The primer sequences
in Supplementary Table S1.(Mu-A1-F/R
All viral and Mu-A2-
F/R) for site-directed mutagenesis,
vectors were aliquoted and stored at −80 °C until use. are shown in Supplementary Table S1. All viral vectors
were aliquoted and stored at − 80 ◦ C until use.
For in vitro transduction assay, the cells were seeded for 16 h in 24-well plates at 70%
For in vitro transduction
confluence before transduction. Then, the mediumassay,was
the replaced
cells werebyseeded for 16
Dulbecco’s h in 24-well plates at
minimum
70% confluence before transduction. Then, the medium was replaced by Dulbecco’s
minimum essential media (DMEM) with 2% fetal bovine serum (FBS), and the AAVs were
added into plates to infect these cells at a multiplicity of infection (MOI) of 35,000. At
24 h and 48 h post-infection, the EGFP signals were observed using an Olympus IX71
fluorescence microscope (Olympus, Tokyo, Japan).

2.4. Next-Generation Sequencing (NGS) Assay

The DNA specimens amplified with NGS-F/R primers (Supplementary Table S1)
were utilized as input material for DNA sample preparation. Sequencing libraries were
Viruses 2023, 15, 848 5 of 14

constructed using the NEB Next Ultra DNA Library Prep Kit for Illumina (New England
Biolabs, Ipswich, MA, USA). In brief, the DNA specimens underwent end repair, A-tailing,
and ligation with a full-length adapter for Illumina sequencing. Following the purification
of the DNA products using the AMPure XP system (Beckman Coulter, Beverly, MA, USA),
DNA concentration was determined using a Qubit 3.0 Fluorometer (Invitrogen, Carlsbad,
CA, USA). The library size distribution was analyzed using an Agilent 2100 Bioanalyzer
and quantified via real-time qPCR.
The index-coded specimens were clustered using a cBot Cluster Generation System,
and then the library preparations were sequenced on an Illumina Novaseq 6000 platform
(Novogene, Beijing, China), generating paired-end reads.

2.5. Research Animals and Stereotactic Intracerebral Injection

The 8–10-week-old C57BL/6 mice (Hunan SJA Laboratory Animal Company, Chang-
sha, China; n = 3 in each group; n = 21 in total) were anesthetized using 1% pentobarbital
intraperitoneally (i.p., 50 mg/kg of their body weight) and placed in a stereotaxic apparatus
(RWD, Shenzhen, China). The injection coordinates referred to Paxinos and Franklin’s The
Mouse Brain in Stereotaxic Coordinates, 4th edition [23]. The coordinate of the caudate
putamen (CPu) was 0.75 mm anterior relative to bregma, 2.00 mm lateral, and 3.30 mm
below the dura. The virus was injected at a rate of 0.03 µL/min using a stereotaxic injector
equipped with a pulled glass capillary (Stoelting, Wood Dale, IL, USA, 53,311). After
the injection was completed, the micropipette was held for an additional 10 min before
being withdrawn. Animals were allowed to recover from anesthesia on a heating pad.
For the intravenous administration, C57BL/6 mice were injected with AAV vectors at
a dose of 4 × 1011 VG/mouse. Three weeks after injection, we sacrificed the mice to
collect brain tissue via transcardiac perfusion with phosphate-buffered saline (PBS) and
4% paraformaldehyde.

2.6. Slice Preparation and Imaging

Brain slice preparation and imaging were completed according to the previously
reported methods [22,24]. The antibodies used were rabbit anti-NeuN (neuronal nuclei
antibody; 1:800; Abcam, Cambridge, MA, USA), goat anti-GFAP (glial fibrillary acidic
protein antibody; 1:800; Abcam, Cambridge, MA, USA), Cy3-conjugated goat anti-rabbit
immunoglobulin G (IgG) (1:400; The Jackson Laboratory, Bar Harbor, ME, USA) and
rabbit anti-goat IgG conjugated with Cy3 (1:400; The Jackson Laboratory, Bar Harbor, ME,
USA). The sections were incubated with the primary antibody overnight at 4 ◦ C. After
washing 3 times with PBS, the slices were incubated with the secondary antibody for 1 h at
37 ◦ C. After washing with PBS, all the brain slices attached to the microscope slides were
counterstained with 40 ,6-diamidino-2-phenylindole (DAPI) (1:4000; Beyotime, Shanghai,
China) and sealed with 70% glycerol. Imaging was performed using an Olympus VS120
virtual microscopy slide scanning system (Olympus, Tokyo, Japan) or a Leica TCS SP8
confocal microscope (Leica, Wetzlar, Germany).

2.7. Data Analysis and Protein Structure Prediction

GraphPad Prism 7.0 (GraphPad Software, La Jolla, CA, USA) was used for data
analysis. One-sixth of the brain slices from each animal were selected for cell counting. The
positive cells were quantified using ImageJ v1.8.0 (National Institutes of Health, Bethesda,
MD, USA). All statistical data were presented as the mean ± the standard error of the
mean (SEM). Statistical significance was set as *, p < 0.05; **, p < 0.01; ***, p < 0.001; and
****, p < 0.0001.
The amino acid sequence alignment was presented using Geneious Prime v2021.2
(Biomatters, Auckland, New Zealand). The VP3 protein structure predictions for AAV2.A1 and
AAV2.A2 were achieved using ColabFold v1.5 (https://github.com/sokrypton/ColabFold,
accessed on 31 January 2023) [25]. The three-dimensional structure alignment of the
 Viruses 2023, 15, 848 6 of 14

VP3 protein of AAV2, AAV2.A1, and AAV2.A2 was demonstrated using UCSF ChimeraX
v1.4 [20,21] (UCSF, San Francisco, CA, USA).

3. Results
3.1. Design and Evaluation of Computer-Aided Directed Evolution
We developed a systematic approach that combined computational design with in vivo
selection to obtain novel AAV variants (Figure 3A). Our deep learning model showed
promising results. The models were trained on both single-mutation (442–470) and multi-
mutation (561–588) datasets, as mentioned in the public datasets. We then used the en-
sembled models to predict whether the mutated sequences could assemble, meaning that
2023, 15, x FOR PEER REVIEW 7 of 15
the virion assembly efficiency would reach the capsid assembly efficiency of the wild-type
AAV2 capsid.

Figure 3. Paradigm design and evaluation of machine learning models. (A) Experimental workflow.
Figure 3. Paradigm design and evaluation of machine learning models. (A) Experimental workflow.
(1) A Cap gene library of variants within AA442–469 was generated using computer-aided design.
(1) A Cap gene library of variants within AA442–469 was generated using computer-aided design.
(2) The designed sequences (95,995) and the corresponding barcodes were synthesized in parallel
(2) The designed
on a custom oligonucleotide sequences
array. (95,995)
(3) A plasmid and the
library wascorresponding
generated using barcodes were
two-step synthesized
cloning. (4) in parallel on
The AAV library was manufactured through triple-transfection in HEK293T cells and characterized cloning. (4) The
a custom oligonucleotide array. (3) A plasmid library was generated using two-step
AAVsequencing.
by high-throughput library was manufactured through
(5) The AAV library triple-transfection
was in HEK293T
injected into C57BL/6 mice (ncells
= 3)and
and,characterized by
after 3 weeks of high-throughput
expression, tissues were harvested
sequencing. (5) Theand
AAVthelibrary
DNA waswas isolated, amplified,
injected into C57BL/6 andmice
pre-(n = 3) and, after
pared for sequencing.
3 weeks(6)of
After NGS, variants
expression, tissueswere
wereranked based
harvested andonthe
enrichment
DNA wasinisolated,
select tissues. (7) and prepared
amplified,
Top enriched variants were repackaged and characterized systematically. (B) The AUC curve of the
for sequencing. (6) After NGS, variants were ranked based on enrichment in select tissues. (7) Top
training of a single deep variants
enriched learning were
model. The AUC and
repackaged increased as the training
characterized proceeded.
systematically. (B) (C)
TheTheAUC curve of the
recall curve of the training. We empirically set the training batch size as 4096 and employed a step
training of a single deep learning model. The AUC increased as the training proceeded. (C) The recall
learning rate (LR) scheduler. The LR scheduler had an initial learning rate of 0.0001, the step size
curve of the training. We empirically set the training batch size as 4096 and employed a step learning
was set as 800, and gamma was set to 0.9.
rate (LR) scheduler. The LR scheduler had an initial learning rate of 0.0001, the step size was set as
Diversity 800, and gamma
at each was detected
step was set to 0.9. using NGS. The Cap gene library sequences
(95,995) and the corresponding barcode sequences (17 bp) were designed using a com-
puter and synthesized in a custom oligonucleotide array. After two-step cloning, the plas-
mid library covered approximately 99.49% (95,506) of the design sequences. The viral li-
brary covered 99.06% (95,089) of the sequences and had 22.57% mutants (21,457) whose
packaging efficiencies were higher than AAV2 (Figure 4A).
Viruses 2023, 15, 848 7 of 14

We generated 95,995 sequences that the models rated as could be assembled; among
them, about 1/3 were generated with random multi-mutation, 1/3 were generated through
rational design, and 1/3 were considered viable by our deep learning models. Based on
the experiment, the average assembly rate was 22.57%, while random multi-mutation
had an assembly rate of 7.20%, the rational design had an assembly rate of 11.70%, and
AI-guided sequences had an assembly rate of 54.15%. This meant that nearly every 1 out
of 2 sequences generated by models could actually be assembled, while the rate was less
than 8% among the randomly mutated sequences. Compared to conventional methods, our
approach greatly helped increase the efficiency of AAV mutant design, which conserves
a large number of experiment resources. These results confirm that the computer-aided
design strategy greatly improves the efficiency of library packaging, which is in line with
previous reports [7].
For the training of deep learning models, we used the area under the curve (AUC) and
recall rate analyses to evaluate the performance of the models (Figure 3B,C). We assume that
recall rate is a well-known concept. AUC stands for the area under the receiver operating
characteristic (ROC) curve and can be seen as a scale-invariant proxy of ROC. The ROC
curve is a common metric used to evaluate the performance of a classification model. A
ROC curve plots the true positive rate (TPR) vs. the false positive rate (FPR) at different
classification thresholds. The AUC ranges in value from 0 to 1. As can be easily deduced,
the higher the AUC is, the more accurate the model is.
Diversity at each step was detected using NGS. The Cap gene library sequences
(95,995) and the corresponding barcode sequences (17 bp) were designed using a computer
and synthesized in a custom oligonucleotide array. After two-step cloning, the plasmid
library covered approximately 99.49% (95,506) of the design sequences. The viral library
covered 99.06% (95,089) of the sequences and had 22.57% mutants (21,457) whose packaging
efficiencies were higher than AAV2 (Figure 4A).

3.2. In Vivo Library Selections Converged on Dominant AAV Variants

The AAV library was injected into the caudate putamen (CPu) brain region of C57BL/6
mice (Figure 4B). After 3 weeks of expression, the tissues were harvested, and the DNA
was isolated, amplified, and prepared for sequencing. Variants were ranked based on
enrichment in the selected tissues. The top 20 variants identified via NGS sequencing were
designated AAV2.A1 to AAV2.A20 based on their ranking—among which, 15 variants had
in vivo transduction levels exceeding that of wide-type AAV2 transduction (Figure 4C).
We further specifically characterized the top two mutants as AAV2.A1 and AAV2.A2.
Compared to AAV2, AAV2.A1 had one amino acid substitution (T456I) and one amino
acid deletion (R459∆), and AAV2.A2 had two amino acid substitutions (R447M and T456S)
(Figure 4D).
To visualize the 3D structural changes of the protein due to amino acid mutations, we
used ColabFold v1.5 (https://github.com/sokrypton/ColabFold, accessed on 31 January
2023) [25] to predict the VP3 protein structure of AAV2.A1 and AAV2.A2. This was possible
because the atomic structure of the AAV2 capsid was experimentally determined and the
number of mutations in AAV2.A1 and AAV2.A2 was small. ColabFold, which incorporates
homology modeling methods, was well-suited for this task. It should be noted that this
software-based structural prediction method was intended to be a reference and not an
in-depth investigation. The results showed that the two mutants had only minor structural
changes compared to AAV2 (Supplementary Figure S1).

3.3. Novel AAV Variants Exhibited Higher Transduction Efficiency Than AAV2 In Vitro
To systematically characterize AAV2.A1 and AAV2.A2, these two variants were indi-
vidually repackaged using the traditional triple-plasmid transfection method, and their
packaging efficiencies were evaluated. We found that the mutations of the amino acid
sequences did not significantly affect the packaging efficiency of AAV2.A1 or AAV2.A2
compared to AAV2 (Figure 5A,B).
FOR PEER REVIEW
Viruses 2023, 15, 848 8 of 15 8 of 14

Figure 4. Diversity characterization during library preparation and mutant enrichment during in
Figure 4. Diversity characterization during library preparation and mutant enrichment during in vivo
vivo screening. (A) Polar plot showing the absolute quantities of the unique mutant collected at each
screening. (A) Polar plot showing the absolute quantities of the unique mutant collected at each stage
stage of the process. of(B) In vivo screening for AAVs with high transduction efficiency; 7.74 × 109
the process. (B) In vivo screening for AAVs with high transduction efficiency; 7.74 × 109 vector
vector genomes (VG)genomes
of the AAV library were injected into the caudate putamen (CPu) of C57BL/6
(VG) of the AAV library were injected into the caudate putamen (CPu) of C57BL/6 mice
mice (n = 3). (C) The top
(n =20
3).enriched variants,
(C) The top detected
20 enriched using
variants, NGS during
detected in vivo
using NGS screening.
during (D) The (D) The
in vivo screening.
amino acid sequenceaminoalignments for AAV2, AAV2.A1, and AAV2.A2.
acid sequence alignments for AAV2, AAV2.A1, and AAV2.A2.

To evaluate
3.2. In Vivo Library Selections the transduction
Converged efficiency
on Dominant of novel AAV variants in vitro, we measured
AAV Variants
the mean fluorescence intensity, expressed by AAV2-, AAV2.A1-, and AAV2.A2-infected
The AAV library was injected into the caudate putamen (CPu) brain region of
HEK293T cells carrying the EGFP reporter driven by the CMV promoter. At 24 h post-
C57BL/6 mice (Figure 4B). Aftervariants
infection, both 3 weeks of expression,
showed the tissues
higher fluorescence werethan
expression harvested, anddifference
AAV2. This
the DNA was isolated,
became amplified,
even more and prepared
pronounced at 48for sequencing. with
h post-infection, Variants were
AAV2.A1 ranked
being 2-fold higher
based on enrichment in the selected tissues. The top 20 variants identified via NGS se-
than AAV2, in particular. (Figure 5C,D).
quencing were designated AAV2.A1 to AAV2.A20 based on their ranking—among which,
15 variants had in vivo transduction levels exceeding that of wide-type AAV2 transduc-
tion (Figure 4C). We further specifically characterized the top two mutants as AAV2.A1
and AAV2.A2. Compared to AAV2, AAV2.A1 had one amino acid substitution (T456I)
 3.3. Novel AAV Variants Exhibited Higher Transduction Efficiency Than AAV2 in Vitro
To systematically characterize AAV2.A1 and AAV2.A2, these two variants were in-
dividually repackaged using the traditional triple-plasmid transfection method, and their 9 of 14
Viruses 2023, 15, 848
packaging efficiencies were evaluated. We found that the mutations of the amino acid
sequences did not significantly affect the packaging efficiency of AAV2.A1 or AAV2.A2
compared to AAV2.3.4. (Figure 5A,B).
Novel AAV Variants Exhibited Higher Transduction Efficiency Than AAV2 In Vivo
To evaluate the transduction
To assess theefficiency of novelefficiency
in vivo transduction AAV variants
of the AAV invariants,
vitro, wewe measured
injected AAV2.A1
the mean fluorescence andintensity,
AAV2.A2 carrying
expressed an EGFP reporter AAV2.A1-,
by AAV2-, controlled byandthe CMV promoter into the CPu
AAV2.A2-infected
for comparison with AAV2. Both AAV2.A1 and AAV2.A2 showed larger transduction
HEK293T cells carrying the EGFP reporter driven by the CMV promoter. At 24 h post-
ranges than AAV2, with 15 and 10 times more EGFP-positive neurons in the CPu region,
infection, both variants showed
respectively higher
(Figure 6A,B). fluorescence
We also testedexpression than AAV2.
the ability of AAV2.A1 This differ-
and AAV2.A2 to cross the
ence became even more pronounced
blood–brain barrier viaattail
48vein
h post-infection,
injections in adultwith
C57BL/6AAV2.A1
mice (n =being 2-fold
3 in each group); no
higher than AAV2, in particular. (Figure 5C,D).
fluorescent signals were observed in the mouse brain slices (Supplementary Figure S2).

Figure 5. Characterization of the production and transduction properties of novel AAV variants.
Figure 5. Characterization of the production and transduction properties of novel AAV variants.
(A,B) Crude lysate PCR (A,B) Crudewere
assays lysate PCRperformed on performed
assays were media (A) on and
mediacellular lysateslysates
(A) and cellular (B) of(B)HEK239T
of HEK239T cells
cells subjected to triple
subjected to triple transfection for AAV2, AAV2.A1, and AAV2.A2. Values representmean
transfection for AAV2, AAV2.A1, and AAV2.A2. Values represent mean ±± SEM,
SEM, n = 3/group. Statistical analysis was completed using one-way ANOVA analysis
n = 3/group. Statistical analysis was completed using one-way ANOVA analysis followed by followed byTukey’s
Tukey’s multiple comparisons test with an
multiple comparisons testalpha
with anvalue
alpha of 0.05.
value ns, no
of 0.05. ns,significant
no significantdifference. (C)Fluorescence
difference. (C) Flu-
orescence expressionsexpressions
of AAV2,ofAAV2.A1, AAV2, AAV2.A1, and AAV2.A2
and AAV2.A2after infection
after infection into HEK293T
into HEK293T cells
cells forh 24
for 24 and 48 h,
h and 48 h, respectively. Scale bar = 200 μm. (D) Quantification of mean fluorescence intensities of AAV
respectively. Scale bar = 200 µm. (D) Quantification of mean fluorescence intensities of different
transduction
different AAV transduction in vitro.
in vitro. Statistical
Statistical analysis was
analysis wascompleted
completed using two-way
using ANOVA
two-way followed by
ANOVA fol-Tukey’s
multiple
lowed by Tukey’s multiple comparisons test
comparisons testwith an alpha
with value value
an alpha of 0.05.of
AU, arbitrary
0.05. AU, units; hpi, hours
arbitrary units;post-infection;
hpi,
hours post-infection; *,*, pp << 0.05;
0.05;**,**,p <p 0.01; p < 0.0001.
****,****,
< 0.01; p < 0.0001.
Then, the GFAP and NeuN antibodies were used to identify cell types via immunoflu-
orescence. Our results demonstrated that the vast majority of the EGFP-positive cells were
neurons, and astrocytes were almost absent for AAV2, AAV2.A1, and AAV2.A2 (Figure 6C).
Since the vast majority of the cells that AAV2 infects in the CNS are neurons [26], limited
amino acid mutations in the variants did not alter the cell-targeting specificity of AAV2.
 Then, the GFAP and NeuN antibodies were used to identify cell types via immuno-
fluorescence. Our results demonstrated that the vast majority of the EGFP-positive cells
were neurons, and astrocytes were almost absent for AAV2, AAV2.A1, and AAV2.A2
(Figure 6C). Since the vast majority of the cells that AAV2 infects in the CNS are neurons
[26],15,limited
Viruses 2023, 848 amino acid mutations in the variants did not alter the cell-targeting specificity10 of 14
of AAV2.

Figure 6. Characterization of transduction properties of novel AAV variants. (A) The fluorescence
Figure 6. Characterization of transduction properties of novel AAV variants. (A) The fluorescence
distribution of EGFP at the caudate putamen (CPu). AAV2-CMV-EGFP, AAV2.A1-CMV-EGFP, or
distribution of EGFP at the caudate putamen (CPu). AAV2-CMV-EGFP, AAV2.A1-CMV-EGFP, or
AAV2.A2-CMV-EGFPAAV2.A2-CMV-EGFP
viruses (300 nL, 3viruses × 109 VG, in total)
(300 nL, 3 × 109were
VG, ininjected into
total) were the CPu
injected of C57BL/6
into the CPu of C57BL/6
mice (n = 3 in each group, n = 9 in total). Scale bar = 1 mm. (B) Quantitative analysis of the
mice (n = 3 in each group, n = 9 in total). Scale bar = 1 mm. (B) Quantitative analysis number
of the number
of EGFP+ cells transduced with
+ different AAV vectors in CPu. Statistical analyses were completed
of EGFP cells transduced with different AAV vectors in CPu. Statistical analyses were completed
using one-way ANOVAs usingfollowed by Tukey’s
one-way ANOVAs multiple
followed comparisons
by Tukey’s tests with tests
multiple comparisons an alpha
with anvalue
alphaofvalue of
0.05. ****, p < 0.0001. (C) Immunofluorescence
0.05. of cell types byofneuronal
****, p < 0.0001. (C) Immunofluorescence cell types nuclei antibody
by neuronal (anti-NeuN)
nuclei antibody (anti-NeuN)
or glial fibrillary acidic protein antibody (anti-GFAP). DAPI, 40 ,6-diamidino-2-phenylindole. Scale
bar = 200 µm.

4. Discussion
We established a computer-aided directed evolution system paradigm and designed
a gene mutation library through machine learning. Then, we used two-step cloning to
complete the construction of a plasmid library and packaged a high-diversity AAV mutation
library. As proof of principle, two AAV2 variants—named AAV2.A1 and AAV2.A2—with
higher transduction efficiencies than AAV2, were obtained after in vivo screening.
Viruses 2023, 15, 848 11 of 14

The targeted mutagenesis of the variable regions (VRs) exposed to the capsid sur-
face is a common sense approach, as these regions are the least evolutionarily conserved
sequences and are known to mediate virus–receptor interactions [27]. The prominence
of capsid protrusions is also utilized in the most rational way, taking advantage of the
flexibility of its surface-exposed loops, especially VR-IV and VR-VIII [27]. A large number
of studies have been conducted to construct mutant libraries, develop rational designs
by inserting short peptides into the AAV capsid VR-VIII, and obtain AAV variants with
excellent characteristics, such as AAV2-7m8 [10], AAV2-retro [13], AAV-MNM004 [14], AAV-
PHP.B/AAV-PHP.eB [11,28], MyoAAV [29], AAV9-Retro [30], and AAV-ie [31]; however,
only a few studies of AAV capsid engineering have involved the VR-IV [32,33]. Based on the
data previously collected by the George Church group [7,18], our study constructed a mu-
tation library in the vicinity of VR-IV through computer-aided design and obtained novel
variants with high transduction efficiency through in vivo selection, thereby confirming
the research potential of AAV vector engineering in this region.
During the process of library construction, the paradigm of computer-aided directed
evolution designed in this study has shown significant advantages. Our deep learning
models showed prominent performance in predicting the viability of the random multi-
mutated sequences, which greatly conserves experimental resources, and thus, removes
almost all restrictions in sequence generation since we can confidently focus on the high-
scoring sequences without wasting significant effort on non-viable ones. That is, the loss
of library diversity at each step was low, which greatly ensured the richness of the library
(Figure 4A). Finally, an AAV library containing about 95,089 mutants was obtained, which
laid the foundation for subsequently more refined library screening.
As proof of principle, we first screened and then characterized two variants from
the library, AAV2.A1 and AAV2.A2, whose transduction efficiencies in the brain were
10~15 times higher than that of AAV2. This confirms the transformation potential of the
AA442–469 region. Through further iterative evolution, we may obtain better variants.
On the other hand, this also makes AAV2.A1 and AAV2.A2 new parents in engineering
modification, and many variants obtained through peptide display on AAV2 could be
transplanted into AAV2.A1 or AAV2.A2 to get better transduction effects.
Structural characterization showed that two amino acid changes in AAV2.A1 or AAV2.A2
did not significantly affect the three-dimensional structure (Supplementary Figure S2), and
these sites have not been reported to directly bind to cell surface receptors in previous
studies. The Ubiquitin–proteasome pathway plays a significant role in the intracellular
trafficking of AAV, and the phosphorylation of certain surface-exposed amino acid residues
on the capsid provides the primary signal for ubiquitination. Previously, Aslanidi et al. [34]
improved the transduction efficiency of AAV2 through site-directed mutagenesis replacing
AAV2 surface-exposed threonine (T) residues, including T455. Zhang et al. [35] reported
that the AAV2 capsid undergoes a conformational shift after AAVR binding, and this
movement forms close contact between AAVR F156 and AAV2 T456, which helps stabilize
the virus–receptor interaction. Given that the substitution of T456 occurs in both AAV2.A1
and AAV2.A2, perhaps the modification of T456 affects the above process and changes
the transduction effect of AAV2. For AAV2.A1, another mutation site is the deletion of
R459. The mutation of AAV2 residue R459 was previously reported to increase cellular
transduction three-fold [36]. In the combination of AAV2 and AAVR, R459 of AAV2 will
lead to unfavorable steric and/or electrostatic interactions with K399 in the AAVR PKD1
domain [37], so the deletion of R459 may facilitate a better interaction of AAV2 and AAVR.
For AAV2.A2, it is probably worth mentioning that R447 forms a salt bridge with E499 of
VR-V of a symmetry-equivalent subunit [38], and residues 499–503 contact the receptor
AAVR [37]. This salt bridge is not conserved among all AAVR-binding serotypes; thus, it is
fairer to note that R447 interacts with a region of AAV that is in direct contact. Based on
the above information, a more detailed mechanism for the wider spread of AAV2.A1 and
AAV2.A2 still depends on further exploration.
Viruses 2023, 15, 848 12 of 14

Supplementary Materials: The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/v15040848/s1. Table S1: Primer sequences used in the experi-
ment; Figure S1: VP3 protein structural comparison of AAV2, AAV2.A1 and AAV2.A2. Structural
superposition of AAV2 (light grey), AAV2.A1 (light purple) and AAV2.A2 (light blue) shown as
ribbon diagrams. The VP3 protein structures of AAV2.A1 and AAV2.A2 were predicted via Colab-
Fold v1.5 (https://github.com/sokrypton/ColabFold, accessed on 31 January 2023) [25], enlarged
view shows the structural changes by the mutants T456I and R459∆ in the AAV2.A1 capsid and the
mutants R447M and T456S in the AAV2.A2 capsid. This figure was generated using UCSF ChimeraX
v1.4 [20,21] (UCSF, San Francisco, CA, USA); Figure S2: Variants obtained from library screening
did not exhibit cross-blood-brain barrier effects. EGFP-expressing AAV2, AAV2.A1 or AAV2.A2
(5 × 1011 VG/mouse) was intravenously injected into adult C57BL/6 mice (n = 3 in each group, n = 9
in total). Scale bar = 1 mm.
Author Contributions: Conceptualization: Z.H., K.L. and F.X.; methodology: Z.H., F.W. and K.L.;
resources and funding acquisition: J.W., Y.W., C.Y. and F.X.; software: F.W.; validation: Z.H., W.F.
and K.L.; data curation and formal analysis: Z.H. and F.W.; investigation: Z.H., N.L., Y.C., X.Y. and
Z.Z.; visualization and original draft of the manuscript: Z.H., K.L. and F.W.; review and editing: K.L.
and F.X.; supervision: K.L. and F.X. All authors have read and agreed to the published version of the
manuscript.
Funding: This work was supported by the STI2030-Major Projects (2021ZD0201003), the National
Natural Science Foundation of China (31830035, 31771156, 21921004, 82151309), the Strategic Pri-
ority Research Program of the Chinese Academy of Sciences (XDB32030200), the Shenzhen Key
Laboratory of Viral Vectors for Biomedicine (ZDSYS20200811142401005), and the Key Laboratory of
Quality Control Technology for Virus-Based Therapeutics, Guangdong Provincial Medical Products
Administration (2022ZDZ13).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: All data are available upon reasonable request.
Acknowledgments: Schematic diagrams in the pictures in this article were created using BioRender.com.
Conflicts of Interest: A patent based on this manuscript has been submitted. The patent applicant
is ZhunAo Biotechnology Co., Ltd. and iDrugDance Technology Co., Ltd., Fei Wang and Weiwei
Feng are among the inventors of the patent. The application number is Chinese Patent Application
No.202211741124.6. The status of the application is pending. The specific aspect of the manuscript
covered in the patent application is “A recombinant adeno-associated viral vector and its application
in gene delivery.” Xin Yang is currently employed by ZhunAo Biotechnology Co., Ltd. The remaining
authors declare no conflicts of interest.

References
1. Russell, S.; Bennett, J.; Wellman, J.A.; Chung, D.C.; Yu, Z.F.; Tillman, A.; Wittes, J.; Pappas, J.; Elci, O.; McCague, S.; et al.
Efficacy and safety of voretigene neparvovec (AAV2-hRPE65v2) in patients with RPE65-mediated inherited retinal dystrophy:
A randomised, controlled, open-label, phase 3 trial. Lancet 2017, 390, 849–860. [CrossRef] [PubMed]
2. Mendell, J.R.; Al-Zaidy, S.; Shell, R.; Arnold, W.D.; Rodino-Klapac, L.R.; Prior, T.W.; Lowes, L.; Alfano, L.; Berry, K.; Church, K.;
et al. Single-Dose Gene-Replacement Therapy for Spinal Muscular Atrophy. N. Engl. J. Med. 2017, 377, 1713–1722. [CrossRef]
[PubMed]
3. Shen, W.; Liu, S.; Ou, L. rAAV immunogenicity, toxicity, and durability in 255 clinical trials: A meta-analysis. Front. Immunol.
2022, 13, 1001263. [CrossRef] [PubMed]
4. Wang, D.; Tai, P.W.L.; Gao, G. Adeno-associated virus vector as a platform for gene therapy delivery. Nat. Rev. Drug Discov. 2019,
18, 358–378. [CrossRef] [PubMed]
5. Colella, P.; Ronzitti, G.; Mingozzi, F. Emerging Issues in AAV-Mediated In Vivo Gene Therapy. Mol. Ther. Methods Clin. Dev. 2018,
8, 87–104. [CrossRef]
6. Ojala, D.S.; Sun, S.; Santiago-Ortiz, J.L.; Shapiro, M.G.; Romero, P.A.; Schaffer, D.V. In Vivo Selection of a Computationally
Designed SCHEMA AAV Library Yields a Novel Variant for Infection of Adult Neural Stem Cells in the SVZ. Mol. Ther. 2018, 26,
304–319. [CrossRef]
7. Ogden, P.J.; Kelsic, E.D.; Sinai, S.; Church, G.M. Comprehensive AAV capsid fitness landscape reveals a viral gene and enables
machine-guided design. Science 2019, 366, 1139–1143. [CrossRef]
Viruses 2023, 15, 848 13 of 14

8. Grimm, D.; Lee, J.S.; Wang, L.; Desai, T.; Akache, B.; Storm, T.A.; Kay, M.A. In vitro and in vivo gene therapy vector evolution via
multispecies interbreeding and retargeting of adeno-associated viruses. J. Virol. 2008, 82, 5887–5911. [CrossRef]
9. Maersch, S.; Huber, A.; Buning, H.; Hallek, M.; Perabo, L. Optimization of stealth adeno-associated virus vectors by randomization
of immunogenic epitopes. Virology 2010, 397, 167–175. [CrossRef]
10. Dalkara, D.; Byrne, L.C.; Klimczak, R.R.; Visel, M.; Yin, L.; Merigan, W.H.; Flannery, J.G.; Schaffer, D.V. In vivo-directed evolution
of a new adeno-associated virus for therapeutic outer retinal gene delivery from the vitreous. Sci. Transl. Med. 2013, 5, 189ra176.
[CrossRef]
11. Deverman, B.E.; Pravdo, P.L.; Simpson, B.P.; Kumar, S.R.; Chan, K.Y.; Banerjee, A.; Wu, W.L.; Yang, B.; Huber, N.; Pasca, S.P.; et al.
Cre-dependent selection yields AAV variants for widespread gene transfer to the adult brain. Nat. Biotechnol. 2016, 34, 204–209.
[CrossRef]
12. Muller, O.J.; Kaul, F.; Weitzman, M.D.; Pasqualini, R.; Arap, W.; Kleinschmidt, J.A.; Trepel, M. Random peptide libraries displayed
on adeno-associated virus to select for targeted gene therapy vectors. Nat. Biotechnol. 2003, 21, 1040–1046. [CrossRef]
13. Tervo, D.G.; Hwang, B.Y.; Viswanathan, S.; Gaj, T.; Lavzin, M.; Ritola, K.D.; Lindo, S.; Michael, S.; Kuleshova, E.; Ojala, D.; et al.
A Designer AAV Variant Permits Efficient Retrograde Access to Projection Neurons. Neuron 2016, 92, 372–382. [CrossRef]
14. Davidsson, M.; Wang, G.; Aldrin-Kirk, P.; Cardoso, T.; Nolbrant, S.; Hartnor, M.; Mudannayake, J.; Parmar, M.; Bjorklund, T.
A systematic capsid evolution approach performed in vivo for the design of AAV vectors with tailored properties and tropism.
Proc. Natl. Acad. Sci. USA 2019, 116, 27053–27062. [CrossRef]
15. Nicklin, S.A.; Buening, H.; Dishart, K.L.; de Alwis, M.; Girod, A.; Hacker, U.; Thrasher, A.J.; Ali, R.R.; Hallek, M.; Baker, A.H.
Efficient and selective AAV2-mediated gene transfer directed to human vascular endothelial cells. Mol. Ther. 2001, 4, 174–181.
[CrossRef]
16. Grifman, M.; Trepel, M.; Speece, P.; Gilbert, L.B.; Arap, W.; Pasqualini, R.; Weitzman, M.D. Incorporation of tumor-targeting
peptides into recombinant adeno-associated virus capsids. Mol. Ther. 2001, 3, 964–975. [CrossRef]
17. Girod, A.; Ried, M.; Wobus, C.; Lahm, H.; Leike, K.; Kleinschmidt, J.; Deleage, G.; Hallek, M. Genetic capsid modifications allow
efficient re-targeting of adeno-associated virus type 2. Nat. Med. 1999, 5, 1052–1056. [CrossRef]
18. Bryant, D.H.; Bashir, A.; Sinai, S.; Jain, N.K.; Ogden, P.J.; Riley, P.F.; Church, G.M.; Colwell, L.J.; Kelsic, E.D. Deep diversification
of an AAV capsid protein by machine learning. Nat. Biotechnol. 2021, 39, 691–696. [CrossRef]
19. Kern, A.; Schmidt, K.; Leder, C.; Muller, O.J.; Wobus, C.E.; Bettinger, K.; Von der Lieth, C.W.; King, J.A.; Kleinschmidt, J.A.
Identification of a heparin-binding motif on adeno-associated virus type 2 capsids. J. Virol. 2003, 77, 11072–11081. [CrossRef]
20. Pettersen, E.F.; Goddard, T.D.; Huang, C.C.; Meng, E.C.; Couch, G.S.; Croll, T.I.; Morris, J.H.; Ferrin, T.E. UCSF ChimeraX:
Structure visualization for researchers, educators, and developers. Protein Sci. 2021, 30, 70–82. [CrossRef]
21. Goddard, T.D.; Huang, C.C.; Meng, E.C.; Pettersen, E.F.; Couch, G.S.; Morris, J.H.; Ferrin, T.E. UCSF ChimeraX: Meeting modern
challenges in visualization and analysis. Protein Sci. 2018, 27, 14–25. [CrossRef] [PubMed]
22. Han, Z.; Luo, N.; Wu, Y.; Kou, J.; Ma, W.; Yang, X.; Cai, Y.; Ma, L.; Han, L.; Wang, X.; et al. AAV13 Enables Precise Targeting of
Local Neural Populations. Int. J. Mol. Sci. 2022, 23, 12806. [CrossRef] [PubMed]
23. Paxinos, G.; Franklin, K.B.J. Paxinos and Franklin’s the Mouse Brain in Stereotaxic Coordinates, 4th ed.; Elsevier Academic Press:
Amsterdam, The Netherlands; Boston, MA, USA, 2013.
24. Han, Z.; Luo, N.; Kou, J.; Li, L.; Ma, W.; Peng, S.; Xu, Z.; Zhang, W.; Qiu, Y.; Wu, Y.; et al. AAV11 permits efficient retrograde
targeting of projection neurons. bioRxiv 2022. [CrossRef]
25. Mirdita, M.; Schutze, K.; Moriwaki, Y.; Heo, L.; Ovchinnikov, S.; Steinegger, M. ColabFold: Making protein folding accessible to
all. Nat. Methods 2022, 19, 679–682. [CrossRef]
26. Haery, L.; Deverman, B.E.; Matho, K.S.; Cetin, A.; Woodard, K.; Cepko, C.; Guerin, K.I.; Rego, M.A.; Ersing, I.; Bachle, S.M.;
et al. Adeno-Associated Virus Technologies and Methods for Targeted Neuronal Manipulation. Front. Neuroanat. 2019, 13, 93.
[CrossRef]
27. Zolotukhin, S.; Vandenberghe, L.H. AAV capsid design: A Goldilocks challenge. Trends Mol. Med. 2022, 28, 183–193. [CrossRef]
28. Chan, K.Y.; Jang, M.J.; Yoo, B.B.; Greenbaum, A.; Ravi, N.; Wu, W.L.; Sanchez-Guardado, L.; Lois, C.; Mazmanian, S.K.; Deverman,
B.E.; et al. Engineered AAVs for efficient noninvasive gene delivery to the central and peripheral nervous systems. Nat. Neurosci.
2017, 20, 1172–1179. [CrossRef]
29. Tabebordbar, M.; Lagerborg, K.A.; Stanton, A.; King, E.M.; Ye, S.; Tellez, L.; Krunnfusz, A.; Tavakoli, S.; Widrick, J.J.; Messemer,
K.A.; et al. Directed evolution of a family of AAV capsid variants enabling potent muscle-directed gene delivery across species.
Cell 2021, 184, 4919–4938.e22. [CrossRef]
30. Lin, K.; Zhong, X.; Li, L.; Ying, M.; Yang, T.; Zhang, Z.; He, X.; Xu, F. AAV9-Retro mediates efficient transduction with axon
terminal absorption and blood-brain barrier transportation. Mol. Brain 2020, 13, 138. [CrossRef]
31. Tan, F.; Chu, C.; Qi, J.; Li, W.; You, D.; Li, K.; Chen, X.; Zhao, W.; Cheng, C.; Liu, X.; et al. AAV-ie enables safe and efficient gene
transfer to inner ear cells. Nat. Commun. 2019, 10, 3733. [CrossRef]
32. Tse, L.V.; Klinc, K.A.; Madigan, V.J.; Castellanos Rivera, R.M.; Wells, L.F.; Havlik, L.P.; Smith, J.K.; Agbandje-McKenna, M.;
Asokan, A. Structure-guided evolution of antigenically distinct adeno-associated virus variants for immune evasion. Proc. Natl.
Acad. Sci. USA 2017, 114, E4812–E4821. [CrossRef]
Viruses 2023, 15, 848 14 of 14

33. Havlik, L.P.; Simon, K.E.; Smith, J.K.; Klinc, K.A.; Tse, L.V.; Oh, D.K.; Fanous, M.M.; Meganck, R.M.; Mietzsch, M.; Kleinschmidt,
J.; et al. Coevolution of Adeno-associated Virus Capsid Antigenicity and Tropism through a Structure-Guided Approach. J. Virol.
2020, 94, e00976-20. [CrossRef]
34. Aslanidi, G.V.; Rivers, A.E.; Ortiz, L.; Song, L.; Ling, C.; Govindasamy, L.; Van Vliet, K.; Tan, M.; Agbandje-McKenna, M.;
Srivastava, A. Optimization of the capsid of recombinant adeno-associated virus 2 (AAV2) vectors: The final threshold? PLoS
ONE 2013, 8, e59142. [CrossRef]
35. Zhang, R.; Cao, L.; Cui, M.; Sun, Z.; Hu, M.; Zhang, R.; Stuart, W.; Zhao, X.; Yang, Z.; Li, X.; et al. Adeno-associated virus 2 bound
to its cellular receptor AAVR. Nat. Microbiol. 2019, 4, 675–682. [CrossRef]
36. Lochrie, M.A.; Tatsuno, G.P.; Christie, B.; McDonnell, J.W.; Zhou, S.; Surosky, R.; Pierce, G.F.; Colosi, P. Mutations on the external
surfaces of adeno-associated virus type 2 capsids that affect transduction and neutralization. J. Virol. 2006, 80, 821–834. [CrossRef]
37. Zane, G.; Silveria, M.; Meyer, N.; White, T.; Duan, R.; Zou, X.; Chapman, M. Cryo-EM structure of adeno-associated virus 4 at 2.2
A resolution. Acta Crystallogr. D Struct. Biol. 2023, 79, 140–153. [CrossRef]
38. Lerch, T.F.; Chapman, M.S. Identification of the heparin binding site on adeno-associated virus serotype 3B (AAV-3B). Virology
2012, 423, 6–13. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.