Article

A Simple Marker-Assisted 3D Nanometer Drift

Correction Method for Superresolution Microscopy
Hongqiang Ma,1 Jianquan Xu,1 Jingyi Jin,1,2 Yi Huang,3 and Yang Liu1,3,*
1
Biomedical and Optical Imaging Laboratory, Departments of Medicine and Bioengineering, University of Pittsburgh, Pittsburgh, Pennsylvania;
2
School of Medicine, Tsinghua University, Haidian District, Beijing, China; and 3University of Pittsburgh Cancer Institute, Pittsburgh,
Pennsylvania

ABSTRACT High-precision fluorescence microscopy such as superresolution imaging or single-particle tracking often re-
quires an online drift correction method to maintain the stability of the three-dimensional (3D) position of the sample at a nano-
meter precision throughout the entire data acquisition process. Current online drift correction methods require modification of
the existing two-dimensional (2D) fluorescence microscope with additional optics and detectors, which can be cumbersome
and limit its use in many biological laboratories. Here we report a simple marker-assisted online drift correction method in
which all 3D positions can be derived from fiducial markers on the coverslip of the sample on a standard 2D fluorescence mi-
croscope without additional optical components. We validate this method by tracking the long-term 3D stability of single-mole-
cule localization microscopy at a precision of <2 and 5 nm in the lateral and axial dimension, respectively. We then provide
three examples to evaluate the performance of the marker-assisted drift correction method. Finally, we give an example of a
biological application of superresolution imaging of spatiotemporal alteration for a DNA replication structure with both low-
abundance newly synthesized DNAs at the early onset of DNA synthesis and gradually condensed DNA structures during
DNA replication. Using an isogenic breast cancer progression cell line model that recapitulates normal-like, precancerous,
and tumorigenic stages, we characterize a distinction in the DNA replication process in normal, precancerous, and tumorigenic
cells.

INTRODUCTION
Fluorescence microscopy is a simple but powerful tech-
nique to visualize biological structures or track the dy-
namic process of macromolecular interactions at a high
precision in all three dimensions (3D). In particular,
the recent development in superresolution imaging and
single-particle tracking systems, such as single-molecule
localization microscopy (SMLM) (also known as (fluores-
cence) photoactivated localization microscopy (1,2) and
(direct) stochastic optical reconstruction microscopy ((d)
STORM) (3,4)), demands an extremely stable optical
system to maintain the 3D position of the sample down
to a few nanometers. System drift is one major source
for compromised precision, coming from various sources
such as mechanical vibration or thermal expansion, espe-
cially when long acquisition time is required. Different
methods have been developed to correct for lateral and
axial drift. They are generally categorized into two ap-
proaches. One commonly used approach on a standard
Submitted January 6, 2017, and accepted for publication April 20, 2017.
*Correspondence: liuy@pitt.edu
Editor: Julie Biteen.
http://dx.doi.org/10.1016/j.bpj.2017.04.025
Ó 2017 Biophysical Society.
two-dimensional (2D) fluorescence microscope is posterior
image processing method for lateral drift correction (5–9),
combined with a focus compensation hardware system
for axial drift correction (e.g., a perfect focus system im-
plemented in most commercial superresolution imaging
systems) (10). Most focus compensation systems use a
separate infrared light source and detector, and monitor
the reflected infrared light at the interface between the
cover glass and the sample due to their different refractive
indices. Another approach is based on fiducial markers
added as part of the sample. To correct for both lateral
and axial drift, a 3D localization microscope setup has
to be used, which requires additional optics, such as a
cylindrical lens inserted into the detection path or multifo-
cus configuration to localize the 3D positions of the fidu-
cial markers (11–13). These drift correction methods
have routinely shown the precision in the lateral position
of <10 nm and the axial position of
20–30 nm. A recent
report demonstrated the state-of-the-art overall correction
precision of
1.3 nm in the lateral position and
6 nm
in the axial position using the phase response of the nano-
particles (14).

2196 Biophysical Journal 112, 2196–2208, May 23, 2017


Current 3D drift correction methods suffer from certain
limitations. First, they all require modification to a stan-
dard 2D fluorescence microscopy system (e.g., additional
illumination light, optical components, special detectors)
or introduction of other imaging modalities (e.g., phase or
bright-field microscopy), which complicates the optical sys-
tem and can be difficult to implement in laboratories without
substantial optics expertise. On the other hand, the posterior
cross-correlation image processing method is a simple alter-
native, but it can only be used to correct for lateral drift in a
2D system, and a separate focus compensation hardware
system is often required to correct for the axial drift.
Here, we report, to our knowledge, a new and simple on-
line marker-assisted (MA) drift correction method in which
the entire 3D position can be derived from the fiducial
markers on the coverslip of the sample for the 3D drift
correction on a standard 2D fluorescence microscopy sys-
tem without introducing additional light source, optics, or
detectors. This method can routinely limit the effect of mo-
tion blur to be <2 nm in the lateral direction and <5 nm in
the axial direction during a long data acquisition process of

20 min for various imaging depths. Then we provide ex-
amples of superresolution imaging of low-abundance mole-
cules of interest and cells that move or deform during
imaging to show that the resolution and reliability of the
MA drift correction method is comparable to the state of
the art. Furthermore, to demonstrate the application of the
MA-based high-precision superresolution imaging system,
we investigate an important biological problem. By map-
ping the temporal alteration of nanoscale organization of
replication structures at the early onset of new DNA synthe-
sis in cells at three different stages of neoplastic progression,
we identify the distinction among three cell lines from an
isogenic human breast cancer progression cell line model
that recapitulates various stages of breast neoplastic trans-
formation from normal-like (benign) breast epithelial cells,
to precancerous, to a malignant form of breast cancer under
the same genetic background (15–19).
MATERIALS AND METHODS
Optical imaging system
The experiments were performed on our home-built fluorescence micro-
scope based upon a model No. IX71 inverted microscope frame (Olympus,
Melville, NY). The fluorophore was excited by 642 nm laser (model No.
VFL-P-1000-642-OEM3; MPB Communications, Point-Claire, Quebec,
Canada), whose intensity was controlled by a neutral density filter (model
No. NDC-50C-4-A; Thorlabs, Newton, NJ). The laser beam was then
expanded by a 10 beam expander (model No. T81-10X; Newport, Strat-
ford, CT), adjusted for appropriate field of view and focused onto the rear
pupil of a 100, NA-1.4 oil immersion objective (model No. UPLSAPO
100XO; Olympus) by an achromatic lens. The emitted fluorescence was
collected by the same objective, passing through a dichroic mirror (model
No. FF660-Di02; Semrock, Rochester, NY) and a band-pass emission filter
(model No. ET700/75m; Chroma Technology, Bellows Falls, VT), and then
focused by a tube lens and a 0.5 C-mount adaptor. The final image was
Simple Nanometer Drift Correction Method
recorded by a sCMOS camera (model No. pco.edge 4.2; PCO-Tech,
Romulus, MI), corresponding to a pixel size of 130 nm on the sample plane.
A closed-loop piezo nanopositioner (model No. Nano-F100S; Mad City
Labs, Madison, WI) was used for drift correction by adjusting the axial po-
sition of the objective in real time. Data acquisition, laser intensity control,
and drift correction were all integrated in our custom-designed software
written in LabVIEW (National Instruments, Austin, TX) and MATLAB
2015 (The MathWorks, Natick, MA). The field of view was set to be
512  512 pixels, corresponding to a region of 66  66 mm2.
For imaging cells, high oblique angle illumination was used to suppress
the background signal. Stage clips (model No. IX-SCL for IX stage;
Olympus) were used to firmly hold the sample on the microscope stage dur-
ing the entire imaging process. We acquired 40,000 frames under laser po-
wer density of 3 kW/cm2 with an exposure time of 20 ms for each image to
ensure the collection of sufficient blinking events. The exposure times for
each image (20 ms exposure time) and the drift correction rates (every
200 frames or 4 s) were used for all the experiments presented here.
Method to estimate 3D position from the fiducial
markers
In conventional 2D fluorescence microscopy, the axial position cannot be
determined from a 2D image, because the image pattern of the point spread
function (PSF) appears identical above and below the focal plane, as shown
in Fig. S1. Our marker-assisted method is a simple approach to directly
determine the 3D position from 2D images based on the asymmetric pat-
terns of PSF from the multiple fiducial markers (e.g., gold nanoparticle)
on the surface of the coverslip. First, a set of 2D images of the fiducial
markers is recorded at different axial positions set by the objective nanopo-
sitioner. Next, the center location information of the PSF (xc, yc) and the
width of the PSF of each marker (wx, wy) at each axial position is retrieved,
by fitting with the 2D elliptical Gaussian function expressed as follows:
"
2
P ðxðnÞ  xc ðnÞÞ
Iðx; y; nÞ ¼ exp  2
2pwx ðnÞwy ðnÞ 2ðwx ðnÞÞ
!#
2
ðyðnÞ  yc ðnÞÞ
þ  2 þ B; (1)
2 wy ðnÞ
where (x,y) is the spatial coordinate of the nth marker; I is the intensity dis-
tribution; (xc(n), yc(n)) is the lateral center location of the nth marker;
(wx(n), wy(n)) is the width of the PSF in the lateral (x-y) dimension for
nth marker; P is the total photon number; and B is the background. Then,
for each fiducial marker, the width and the center location information of
the PSF (in the x-y dimension) is plotted as a function of different axial po-
sitions (z), presented as each calibration curve, as shown in Fig. S3. Please
note that Fxc ðn; zÞ ¼ xc ðn; zÞ  x0 ðnÞ and Fyc ðn; zÞ ¼ yc ðn; zÞ  y0 ðnÞ are
defined as the center positon of the nth marker relative to (subtracting) the
position of its peak pixel at the axial position of 0 (x0(n), y0(n)). Then, a
fourth-order polynomial is used to fit each of the four calibration curves,
defined as Fwx ðn; zÞ, Fwy ðn; zÞ, Fxc ðn; zÞ, and Fyc ðn; zÞ for the nth marker.
Note that, if the PSF of the imaging system and the emission properties of
the markers are perfectly symmetric (as shown in Fig. S1), only the width of
PSF is dependent on the axial position (Fwx ðn; zÞ and Fwy ðn; zÞ in Fig. S3, a
and b), whereas the center location of the PSF should be the same (not
necessarily zero) for all axial positions. However, in practice, especially
when the gold nanoparticle is used as fiducial marker whose emission prop-
erties often do not present a perfectly symmetric PSF (as shown in Fig. S2),
the PSF center position (xc(n,z), yc(n,z)) of each marker varies slowly with
different axial positions, which can also affect the precision to determine
the lateral location of the fiducial marker. Therefore, the calibration curves
for the PSF center location (Fxc ðn; zÞ, Fyc ðn; zÞ in Fig. S3, c and d) are also
needed to correct this error. After the calibration curves for a set of fiducial
Biophysical Journal 112, 2196–2208, May 23, 2017 2197
Ma et al.
markers are built (see Figs. S3–S5 for more details on the specifics of the
calibration curve), we estimate the position of system drift via an online
drift correction method that is described in detail below.
First, the objective nanopositioner is moved to each of a set of axial po-
sitions z. Then, the center location data (xc(n), yc(n)) of the nth marker and
the width data (wx(n), wy(n)) of the nth marker in the x and y dimensions for
each fiducial marker in the set of N fiducial markers are retrieved by 2D
Gaussian function fitting as described above.
Next, a joint or combined axial position (zjoint) of the whole set of N fidu-
cial markers is calculated by taking the positions of all markers into account
to define the joint axial position zjoint as the axial position where the differ-
ence on the PSF width of the calibration curves for all N fiducial markers is
at a minimum. That joint axial position can be found via an optimization
problem using the following equation:
zjoint ¼ arg min
z highlight the candidate emitters from the background. Then, the peak posi-
XN qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2  2
 ðwx ðnÞ  Fwx ðn; zÞÞ þ wy ðnÞ  Fwy ðn; zÞ :
n¼1
(2)
Next, the calculated joint axial position zjoint is used to calculate the cen-
ter bias for each fiducial marker via the calibration curves of Fxc ðn; zÞ and
Fyc ðn; zÞ. The calculated center bias is then subtracted from the retrieved
center location (xc(n), yc(n)) of each fiducial marker to obtain a bias-
adjusted estimate of the center location (xc(n, zjoint), yc(n, zjoint)) for each
marker.
Then, a joint or combined center lateral location (in the x and y direc-
tions) for the whole set of N fiducial markers can be found using Eq. 3 as
follows:
8   X
N
   
>
> x z ¼ xc ðnÞ  Fxc n; zjoint  x0 ðnÞ n
< joint joint
n¼1
: Protocol to coat gold nanoparticles on the
>
>   X
N
   
: yjoint zjoint ¼ yc ðnÞ  Fyc n; zjoint  y0 ðnÞ n
n¼1
(3)
The above-described processes thus result in a single joint 3D position
(zjoint, xjoint, yjoint) for all of the N fiducial markers on the coverslip. The
3D drift during data acquisition may then be calculated as the position dif-
ference between the joint 3D positions determined from two different im-
ages during the data acquisition process. The retrieved axial position drift
(in the z direction) may be used to adjust objective nanopositioner to correct
the axial drift, and the retrieved lateral drift (in the x and y directions) may
be used in data reconstruction steps to correct for lateral drift. This process
is summarized in the flowchart shown in Fig. S6.
To perform 3D drift correction in the case of superresolution localization
microscopy, two different scenarios are considered. When the fiducial
markers are close to the imaging target (as shown in Fig. 1 f), given that
the fiducial markers can be detected in every imaging frame, the axial drift
can be corrected during the data acquisition either every frame or at a set
time interval (every 200 frames or 4 s used here). When the imaging plane
is located at a focal plane distant from that of fiducial markers (e.g., an im-
aging plane located at several microns above the surface of the coverslip
where the markers are already out of focus, as shown in Figs. 2, 3, and 4),
a jump strategy can be used to perform the drift correction. Specifically,
at a set time interval (e.g., 200 frames or 4 s used here), the imaging plane
jumps to the focal plane of fiducial markers via objective nanopositioner to
record their PSF patterns. Please note that the 100-ms exposure time used
here to accumulate sufficient photon number of >20,000 from gold nanopar-
ticles for precise drift estimation, given that the estimated precision in theory
should be inversely proportional to the photon number from the marker (20).
2198 Biophysical Journal 112, 2196–2208, May 23, 2017
We then calculate the sample drift according to the above-described
method, compensate the sample drift, and return to the imaging plane for
subsequent imaging. The total time used for the drift correction using this
jump strategy in this work is
500 ms. The axial drift has to be corrected
online, and the lateral drift can be corrected either online (if a motorized
translational stage is used) or during postprocessing. We use postprocessing
for lateral drift correction during superresolution image reconstruction.
Reconstruction of superresolution image
Our superresolution image reconstruction algorithm was implemented by
least square fitting to a single-emitter Gaussian function written in
MATLAB 2015 (21) and a diagram to detail all preprocessing and image
reconstruction steps is shown in Fig. S7. In brief, raw image data was first
denoised with a difference of Gaussian filter (s1 ¼ 1 pixel; s2 ¼ 3 pixels) to
tions of candidate single emitters were extracted by finding the local max-
ima in each region of 5  5 pixels, followed by fitting with a single-emitter
Gaussian function model in a region of 7  7 pixels. Those candidate emit-
ters that meet the following criteria were rejected in the localization pro-
cess: 1) total photon number <100, 2) point spread function width of
50% larger or smaller than the system PSF, and 3) peak intensity versus
background intensity <0.5. After applying the drift correction (marker-as-
sisted or cross-correlation-based), the final superresolution image was
rendered by accumulating all the valid emitters with a pixel size of
10 nm followed by a Gaussian smoothing filter (s2 ¼ 10 nm), without
any combination of consecutive frames. Please note that we did not account
for the sCMOS-specific noise statistics in our image reconstruction algo-
rithm, as it was previously shown that the localization bias caused by the
fixed pattern noise is <2 nm for most state-of-the-art sCMOS cameras
(22). However, if higher precision is needed, a sCMOS-specific algorithm
needs to be applied (23).
coverslip
We first coated the glass-bottom dish (Cat. No. FD3510-100; World Preci-
sion Instruments, Sarasota, FL) with poly-D-lysine (Cat. No. P7280;
Sigma-Aldrich, St. Louis, MO) for 20 min, followed by a 200-mL diluted
100-nm gold nanoparticle solution (1:60 with ddH2O, EM.GC100; BBI So-
lutions, Cardiff, UK) for 3 h. Finally, dishes were coated with another layer
of poly-D-lysine for 20 min to improve the cell adherence.
Cancer progression cell line model
MCF10A progression cell lines were gifts from Dr. Saraswati Sukumar
(Johns Hopkins University, Baltimore, MD), and were cultured in growth
medium as described in Santner et al. (15) and Debnath et al. (24). These
cell lines were authenticated by Genetica DNA Laboratories (Cincinnati,
OH) and the genetic profiles of all three cell lines (MCF10A, MCF10AT1K,
and MCF10ACA1a) match with each other, confirming that they are
isogenic and also match the commercial cell line MCF10A at 90%.
All cells were grown at 37 C with 5% CO2 and the MCF10A series pro-
gression cell lines were maintained in DMEM/F12 medium supplemented
with 5% horse serum, 10 mg/mL insulin, 20 ng/mL EGF, 0.5 mg/mL hydro-
cortisone, and 100 ng/mL cholera toxin. The cells were plated onto a poly-
D-lysine-coated glass-bottom dish at an initial confluency of
50% and
cultured overnight to let the cells attach to the dish.
Fluorescent labeling of newly synthesized DNA
DNA staining was conducted by using a Click-iT Plus 5-ethynyl-
20 deoxyuridine (EdU) Alexa Fluor 647 Imaging Kit (C10640; Thermo
 Simple Nanometer Drift Correction Method

FIGURE 1 High-precision superresolution imaging system. (a) Shown here is the schematic of the optical setup. (b) Shown here is the measured 3D point
spread function distribution of the eight fiducial markers attached to the surface of the coverslip. (c) Shown here is the estimated joint 3D position and (d) the
corresponding precision of the eight fiducial markers at 17 axial positions ranging from 400 to 400 nm. (e) Shown here is the tracking 3D position of our
high-precision imaging system for 40,000 frames or
15 min, with the SD of 1.3 nm in the lateral direction and 3.0 nm in the axial direction. Exposure time
for each image: 20 ms. Drift correction rate: every 200 frames. (f) Shown here is the reconstructed superresolution image of microtubules (labeled with
Alexa647) obtained with our high-precision imaging system. (g) A magnified region in (f) clearly shows the hollow structure of the microtubules. (h) Shown
here is the intensity profile of the hollow structures of microtubules. The full width half-maximum of each microtubule shown in (g) is 21 and 19 nm, with a
separation of 37 nm. To see this figure in color, go online.

Fisher Scientific, Waltham, MA). Click-iT Plus reaction cocktails were pre-
pared per the manufacturer’s instruction as follows. For a total volume of
500 mL, cocktails contained 440 mL 1X Click-iT reaction buffer, 10 mL cop-
per protectant, 1.2 mL Alexa Fluor 647 picolyl azide, and 50 mL reaction
buffer additive. All the components were provided by the manufacturer’s
imaging kits.
The cells were plated onto a gold-coated glass-bottom dish at an initial
confluency of
50% and cultured overnight to let the cells attach to the
dish. Diluted EdU in culture medium was added into the dish at a final con-
centration of 10 mM, and incubated with the cells for 1, 2, 5, and 10 min,
respectively. After incubation, we removed the media and fixed the cells
with 4% formaldehyde for 15 min. The cells were then washed three times
with PBS, permeabilized with 0.2% Triton X-100 for 15 min, and washed
three times with 3% BSA in PBS. The Click-iT Plus reaction cocktail
(Thermo Fisher Scientific) was added to detect EdU, incubating for
30 min at room temperature, protected from light. After removing the reac-
tion cocktail and washing twice with 3% BSA in PBS, the cells were stored
in PBS until imaging.
Immediately before imaging, the buffer was switched to the imaging
buffer (50 mM Tris-HCl pH 8.0, 10 mM NaCl, 1% v/v 2-mercaptoethanol
from 14.3 M pure liquid (Sigma-Aldrich), 10% w/v glucose, 0.56 mg/mL
glucose oxidase, 0.17 mg/mL catalase), adapted from the protocol recom-
mended by Nikon (Melville, NY). Our superresolution imaging focused
on the cells at the early S-phase, identified as those with bright and uni-
formly distributed fluorescence intensity throughout the nucleus, a charac-
teristic DNA replication pattern for early S-phase based on the literature
(25,26).
Fluorescence staining of microtubules
The cells were preextracted for 30 s with 0.5% Triton X-100 in BRB80
buffer (80 mM PIPES, 1 mM MgCl2, 1 mM EGTA, adjusted to pH 6.8
with KOH) supplemented with 4 mM EGTA, washed once with PBS,
and then fixed with cold Methanol at 20 C for 10 min. After we
removed the fixative and washed the cells with PBS, cells were then
blocked for 2 h in 3% BSA. The sample was then incubated with
anti-rabbit a-tubulin primary antibody (Cat. No. ab18251; Abcam,
Cambridge, UK) at 1:300 concentration diluted in 3% BSA at 4 C over-
night. Then the cells were washed 3 times with 0.05% Triton in PBS,
and the secondary donkey anti-rabbit antibody (Cat. No. 121165; Jack-
son ImmunoResearch Laboratories, West Grove, PA) labeled with

Biophysical Journal 112, 2196–2208, May 23, 2017 2199

Ma et al.
rected the axial drift using our online MA-based drift correction for all images presented. Here, we only compare the performance of lateral drift correc-
tion between the MA-based method and posterior image cross-correlation method. To see this figure in color, go online.
Alexa Fluor 647 (Cat. No. A20106; Thermo Fisher Scientific) was
added to the sample in 3% BSA at 1:200 concentration and then incu-
bated for 2 h, protected from light. The cells were washed again three
times with 0.05% Triton in PBS and stored in PBS before imaging.
Immediately before imaging, the buffer was switched to the imaging
buffer as described above.
Cluster analysis to automatically identify
replication clusters and quantitative analysis
We first manually segmented each nucleus in superresolution image.
Then for the first-step coarse segmentation of large clustered areas in
the nucleus, the reconstructed STORM image was converted into a bi-
nary image mask, where any pixel with more than one localization
event was designated as 1; otherwise, the pixel was designed as 0.
The MATLAB functions ‘‘imfill’’ and ‘‘bwboundaries’’ were used to
fill in holes (pixels with 0 value) of each enclosed area with 1 in the
binary mask.
Next, a difference of Gaussian filter (s1 ¼ 1 pixel; s2 ¼ 3 pixels) was
used to highlight the clusters from the noisy background. We set a
threshold of 0.8 localizations per subpixel area (10  10 nm) to identify
only those clusters with sufficient relative intensity compared to back-
ground. This threshold value was chosen such that most clusters that
can be visually identified can also be automatically selected by this cluster
analysis algorithm. Within each coarsely segmented large clustered area
(identified in the first step), we estimated the number of clusters and the
initial position via the cluster peaks recognized by finding the local max-
ima in a region of 11  11 pixels. We then retrieved their localization po-
sition list for each coarsely segmented, subsequent cluster analysis using a
Gaussian mixed model (MATLAB function ‘‘fitgmdist’’ (27)) to estimate
the cluster size (defined as full width at half-maximum), cluster center po-
sition, and number of localizations per cluster). The region of nucleoli
was rejected to ensure a more accurate estimation of cluster density (num-
ber of clusters per unit area).
2200 Biophysical Journal 112, 2196–2208, May 23, 2017
FIGURE 2 (a) Wide-field and reconstructed
superresolution images of replication structure
of a normal cell nucleus immediately after the
initiation of new DNA synthesis, labeled with
EdU-Alexa Fluor 647 after exposing to EdU for
1 min. (b) The magnified wide-field image and
superresolution images of the region is indicated
by the box in (a). (c) Shown here is the corre-
sponding superresolution image with only MA-
based axial drift correction. (d) Shown here is
the superresolution image after MA-based axial
drift correction, RCC-based lateral drift correc-
tion, and MA-based axial drift correction and (e)
the superresolution image after MA-based 3D
drift correction. The image resolution defined by
FRC is 84.5 5 1.1 nm, 75.8 5 1.4 nm, and 67.3
5 1.9 nm for (c)–(e), respectively. The white ar-
rows indicate the structure differences before
and after drift correction. (f) The distribution of
localization events along the vertical direction in
the region is indicated by the dotted box in (c)–
(e). (g) Shown here is a comparison of the lateral
drift estimated by (f) our marker-assisted method
and RCC. (h) Shown here is the axial drift esti-
mated by our method. Please note that given that
RCC-based drift correction can only compensate
the drift in the lateral dimension, we have cor-
Statistical analysis
To perform statistical analysis, we first take the mean value for each of the
cluster properties (e.g., cluster size, spot number per cluster, and cluster den-
sity) by averaging the values for all clusters within each cell. This mean
value was then used to calculate all statistical parameters (e.g., mean, stan-
dard deviation, 25th and 75th percentile) using
10 cells (from the same
dish) from each cell type, as shown in Table 1. The comparison between
two groups is performed by a nonparametric Mann-Whitney test, assuming
a 95% confidence interval, using the software GraphPad Prism 6 (GraphPad;
https://www.graphpad.com/scientific-software/prism/). The two-sided exact
p values for the comparisons of normal versus precancerous cells, precan-
cerous versus tumorigenic cells, and normal versus tumorigenic cells, are
listed in Table S3. The p value < 0.05 is considered statistically significant.
RESULTS
Marker-assisted 3D online drift correction method
In a conventional 2D SMLM imaging system, as shown in
Fig. 1 a, only the lateral 2D position can be accurately deter-
mined by fitting the image pattern of a fiducial marker (e.g.,
gold nanoparticle) with a Gaussian function model. To
obtain the axial position, additional optical elements are
required (e.g., cylindrical lens, phase mask, multifocal op-
tics) to encode the axial position of the marker by the shape
of its point spread function. Otherwise, their axial position
cannot be determined, because the image pattern appears
identical when the fiducial markers are located at the upper
and the lower focal planes (as shown in Fig. S1).
However, this fact holds only if one fiducial marker is in
the entire field of view or all of the fiducial markers are on
 Simple Nanometer Drift Correction Method

FIGURE 3 (a) Shown here are the wide-field and

(b)–(g) reconstructed superresolution images of
cancer cell nuclei labeled with histone protein
H2B-Alexa Fluor 647, (b and e) after MA-based
axial drift correction but before lateral drift correc-
tion, and after MA-based axial drift correction
and the lateral drift correction (c and f) using
RCC and (d) and (g) our marker-assisted methods.
Boxes 1 and 2 indicate two regions of the cell
nuclei that exhibit the clusters with different shapes
from RCC-based and MA-based drift correction
methods. The image resolution defined by FRC is
60.6 5 2.1, 49.4 5 1.7, 46.1 5 1.2, 61.0 5 1.9,
50.1 5 1.6, and 56.3 5 1.7 nm for (b)–(g), respec-
tively. Please note that we have corrected the axial
drift using our online MA-based drift correction for
all images presented here. To see this figure in co-
lor, go online.

the exact same focal plane. In practice, the surface of the 3D position on the sample itself without adding any addi-
substrate (i.e., coverslip) is not entirely flat, and such imper- tional optics. We recognize that, as shown in Fig. 1 b (and
fection in fact creates the opportunity to precisely locate the Fig. S3), the axial positions of the fiducial markers attached

Biophysical Journal 112, 2196–2208, May 23, 2017 2201

Ma et al.
FIGURE 4 Superresolution images of replication foci at four different time points (1, 2, 5, and 10 min) of (a1)–(d1) normal-like, (e1)–(h1) precancerous,
and (i1)–(l1) tumorigenic cells. The figure inset shows the corresponding wide-field fluorescence image. (a2)–(l2) Here we have the corresponding higher
zoom of the regions inside the boxes. To see this figure in color, go online.
to the coverslip surface are located at different depths and
their joint PSF distribution is asymmetric, suggesting that
the joint 3D position of these fiducial markers can be pre-
cisely determined simply from the 2D fluorescence image
without encoding the axial information to the shape of the
single PSF via additional optics.
We first measure the precision of this method by tracking the
joint 3D positions of the fiducial markers via stepwise move-
ment of axial positions at 50-nm step size, followed by
repeated localization of the 3D positions at each axial (z) posi-
tion. As shown in Fig. 1, c and d, the total error is <1.5 nm in
the lateral direction and <5 nm in the axial direction for the
entire depth range of [400, 400 nm], with a precision smaller
than 2 nm in all 3D positions for off-center axial positions of
the focal plane (e.g., 250–400 nm and 400 to 200 nm)
and even subnanometer accuracy for the lateral position.
To evaluate the routine performance of our high-precision
superresolution imaging system, we track the long-term sta-
bility (
15 minutes) of the 3D position during an entire im-
age acquisition process of 40,000 frames at an exposure time
of 20 ms per image (drift correction frequency is performed
once every 200 frames). Fig. 1 e shows that our system main-
tains the stability of 1.3 nm in the lateral direction and 3 nm
2202 Biophysical Journal 112, 2196–2208, May 23, 2017
in the axial direction, quantified by the SD of the 3D posi-
tions. The performance of our MA-based drift correction us-
ing different numbers of markers, photon numbers, and
numbers of imaging frames is shown in Tables S1 and S2,
which demonstrates the reproducibility and robustness of
our method with an overall precision of <2 nm in the lateral
dimension and <5 nm in the axial dimension. This value may
also be affected by the movement error of the objective nano-
positioner and high-frequency vibration besides the precision
of MA-based drift correction method itself. The ultimate per-
formance in imaging resolution using this high-precision
SMLM system is evaluated by imaging microtubules in a
MCF10A cell line. As shown in Fig. 1, f–h, our system can
achieve a spatial resolution of 11.2 nm defined by Fourier
ring correlation (FRC) (28), and in some areas, the hollow
structure of the microtubules with a separation distance of
<40 nm can be detected.
Superresolution imaging of cells using MA-based
3D drift correction
In the previous section, we demonstrated the performance of
MA-based 3D drift correction on a standard 2D fluorescence
 Simple Nanometer Drift Correction Method

TABLE 1 Summary of Mean, SD, 25th and 75th Percentile for Normal, Precancerous, and Tumorigenic Cells at 1, 2, 5, and 10 min
after Initiating the New DNA Synthesis
Number of Cells Mean SD 25th Percentile 75th Percentile
1 min cluster size (nm) normal 6 29.1 1.7 27.5 30.9
precancerous 11 33.8 3.2 31.1 36.7
tumorigenic 8 36.9 3.7 33.8 40.6
spot number per cluster normal 6 49 5 47 53
precancerous 11 69 13 62 82
tumorigenic 8 83 23 63 105
cluster density (mm2) normal 6 1.8 0.5 1.3 2.3
precancerous 11 6.0 2.4 4.2 8.1
tumorigenic 8 4.8 1.8 3.1 6.9
2 min cluster size (nm) normal 10 39.3 5.0 35.7 42.5
precancerous 15 46.0 6.1 42.1 48.9
tumorigenic 9 56.4 7.2 49.0 61.5
spot number per cluster normal 10 87 29 61 112
precancerous 15 132 48 98 143
tumorigenic 9 171 53 121 204
cluster density (mm2) normal 10 6.9 2.7 4.6 9.5
precancerous 15 12.6 1.7 11.9 13.3
tumorigenic 9 13.1 2.2 11.1 14.5
5 min cluster size (nm) normal 10 60.8 7.2 54.2 65.3
precancerous 11 70.7 5.1 65.7 76.2
tumorigenic 10 78.7 7.3 74.3 84.6
spot number per cluster normal 10 193 71 127 227
precancerous 10 288 55 237 342
tumorigenic 10 415 96 336 486
cluster density (mm2) normal 10 15.3 2.5 12.9 17.7
precancerous 10 19.1 1.6 18.3 20.1
tumorigenic 10 15.0 3.2 12.9 17.3
10 min cluster size (nm) normal 16 76.7 16.1 66.2 90.0
precancerous 13 87.7 8.2 80.6 95.5
tumorigenic 12 100.2 9.6 97.8 105.1
spot number per cluster normal 16 363 168 232 521
precancerous 13 448 80 381 533
tumorigenic 12 627 129 585 705
cluster density (mm2) normal 16 14.2 1.7 13.0 15.6
precancerous 13 17.5 1.1 16.6 18.4
tumorigenic 12 14.8 2.8 13.6 16.2

microscope to be comparable to the state of the art but
without additional optics. Next, we evaluate the perfor-
mance of MA-based drift correction with the most
commonly used posterior image registration method based
on redundant cross correlation (RCC) in superresolution im-
aging of nanoscale structures. Given that the RCC-based
drift correction cannot perform axial drift correction in a
2D system, we can only evaluate its performance in the
lateral drift correction. We present three different examples
to show that, even for lateral drift correction, the MA-based
drift correction method can achieve either comparable or su-
perior performance to RCC-based posterior image registra-
tion. We also demonstrate that our method can achieve more
reliable representation of the actual structure in cases when
the RCC-based method does not work effectively, such as a
low number of molecules when the localization density is
not sufficiently high, or when the sample (or part of the sam-
ple) deforms during image acquisition.
The first case is imaging a small number of molecules,
such as the early onset of new DNA synthesis in early
S-phase with ultrashort labeling of EdU for only 1 min.
EdU is a nucleoside analog of thymidine that is incorporated
into DNA during active DNA synthesis, followed by Alexa
647-Azide reaction, so that the fluorophore tags the newly
synthesized DNA (29). A spontaneously immortalized
normal-like breast epithelial cell line, MCF10A, is used.
As shown in Fig. 2 a, at 1 min of EdU exposure when the
amount of newly synthesized DNA is still very low, the
left side shows a conventional wide-field image and the right
side shows the superresolution image. The conventional im-
age does not provide sufficient structural detail within the
nucleus, whereas the superresolution image reveals spatially
isolated small clusters, indicating the individual replication
origins shortly after the initiation of new DNA synthesis.
Fig. 2, c–e, compares the reconstructed superresolution im-
ages after only the MA-based axial drift correction
(Fig. 2 c), with the commonly used drift correction method
based on image registration (6) (Fig. 2 d) and with our high-
precision SMLM system using MA-based drift correction
(Fig. 2 e). Given the small number of labeled molecules at

Biophysical Journal 112, 2196–2208, May 23, 2017 2203

Ma et al.
1 min after the new DNA synthesis, all three images exhibit
distinct small clusters, as indicated by the white arrows. The
image with our SMLM system with MA-based drift correc-
tion (Fig. 2 e) gives the best FRC resolution (67.3 5
1.9 nm), when compared to that (84.5 5 1.1 nm)
before lateral drift correction (Fig. 2 c) and that (75.8 5
1.4 nm) corrected with RCC-based lateral drift correction
(Fig. 2 d) (6). In addition, as shown in the cross-sectional
distribution of the localization events (Fig. 2 f) along the
vertical direction in the region indicated by the dotted green
box in Fig. 2, c–e, the MA-based drift correction also pro-
duces the tightest clusters with most localization events at
each cluster. As indicated by different arrows between
Fig. 2, d and e, and the difference in the derived positions
shown in Fig. 2, g and h, the RCC-based drift correction
shows different drift trajectories compared to those of the
MA-based drift correction method. Although ground truth
is difficult to obtain here, the small number of labeled mol-
ecules in principle does not provide a sufficiently reliable
average time-dependent reference image for the RCC-based
method, whereas the MA-based method relies on the highly
stable fiducial markers on the sample that remain the same
during
20 min of the image acquisition process.
The second case is imaging nucleosome organization
labeled by H2B-AlexaFluor 647 in the cell nucleus from
the same cell line via immunofluorescence staining.
Fig. 3, a–d, shows the comparison of conventional wide-
field, the reconstructed superresolution image after MA-
based axial drift correction but before the lateral drift
correction, and that performed after the 3D drift correction
by the MA-based and the lateral drift correction by the
RCC-based method. In the case where the target of interest
(histone protein) is densely labeled, the MA-based drift
correction method provides a slightly better image resolu-
tion defined by FRC.
The third case is an example to demonstrate that MA-
based drift correction helps in identifying the situation
when the cells may experience a possible deformation dur-
ing the imaging process. Fig. 3, e–g, shows that, for some
cell nuclei imaged using the same conditions as those in
Fig. 3, a–d, the RCC-based drift correction (Fig. 3 f) results
in different elongated clusters from those with MA-based
lateral drift correction (Fig. 3 g), as seen in the two regions
inside boxes 1 and 2. The RCC-based method results in a
better FRC resolution, but does it also better reflect the
actual structure of the cell? We explored two possible causes
for such inconsistency—the distortion of the imaging field
and the possible deformation in some parts of the cells.
We first examine the drift trajectories in other cells growing
on the same petri dish either in the same or other fields of
view (FOVs) that exhibit a round (or roundlike) shape
similar to those in Fig. 4 d. As shown in Fig. S8, a–c, the
drift trajectories from both cells and fiducial markers present
a similar direction and the resulting reconstructed image ex-
hibits round (or roundlike-shaped clusters (Fig. S8 c). On
2204 Biophysical Journal 112, 2196–2208, May 23, 2017
the other hand, as shown in Fig. S8, d–f, different parts of
the cell present distinct directions of drift trajectories, as
indicated by the green arrows in Fig. S8 e. But all markers
at similar locations in the same FOV exhibit similar drift tra-
jectories toward the same direction (Fig. S8 f). This result
suggests that parts of the cell may undergo deformation or
damage during image acquisition. Given that the markers
located in the similar FOV do not present different drift ori-
entations, it is unlikely to be due to field distortion. Further,
it is well documented that the gold nanoparticles are highly
stable, but the cell can be damaged during the long exposure
of high laser power density used in STORM imaging (30).
Therefore, this example demonstrates that the cell deforma-
tion may occur under the STORM imaging condition, even
though it only accounts for
5% of the cells in our experi-
ments. The MA-based drift correction helps in identifying
the types of cells that may be excluded for further analysis,
but the RCC-based drift correction may obscure such a sit-
uation despite its better FRC resolution.
Therefore, the above three examples demonstrate that our
MA-based 3D drift correction method overall provides the
state-of-the-art superresolution image and is highly reliable
for various scenarios that other drift correction methods may
not perform well.
Quantitative spatiotemporal imaging of DNA
replication structures in neoplastic progression
To demonstrate the robustness of MA-based SMLM imag-
ing on an important biological problem in which imaging
targeted molecules vary from low abundance to dense
compactly structures, we focus on spatiotemporal superre-
solution imaging of DNA replication in cells undergoing
neoplastic progression. Abnormal proliferation and DNA
replication in neoplastic progression are among the most
universal characteristics shared by most cancers. To charac-
terize the spatiotemporal alteration of replication structure
in cells at different stages of neoplastic progression, we
use an isogenic MCF10 cancer progression cell line model
(16,19) that recapitulates clinically significant stages
(normal, precancerous, and tumorigenic) of neoplastic
transformation in human breast cancer (15,24). Besides
the normal cell line (MCF10A) used above, we also use
its isogenic premalignant cell line (MCF10AT1K) generated
by HRas transformation of MCF10A, and a fully malignant
breast cancer cell line (MCF10CA1a) derived from
MCF10AT xenografts to form poorly differentiated malig-
nant tumors in the xenograft model.
Various microscopy techniques have been used to charac-
terize the in situ DNA replication structures in mammalian
cell nuclei, from phase contrast, confocal fluorescence
microscopy, to electron microscopy, to more recently super-
resolution fluorescence microscopy (26,29,31–36). How-
ever, conventional microscope image of replication foci
often can only be obtained after at least 510 min from
 Simple Nanometer Drift Correction Method

the initiation of new DNA synthesis to accumulate sufficient
signals to visualize replication foci and achieve sufficient
resolution. Such a process inevitably results in aggregates
of replication structure and may lead to incorrect or incon-
sistent characterization of replication structures (29). These
limitations hamper our ability to image temporal alteration
of spatial arrangement of newly synthesized DNA at a
high precision. Therefore, this is an important biological
problem where both high temporal and spatial resolution
are required. To capture the replication structure at the
earliest onset of new DNA synthesis, we use a very short
pulse-labeling time of 1 min, and then 2, 5, and 10 min.
Fig. 4, a1–d1, shows the temporal alteration of replication
structures at 1, 2, 5, and 10 min after the initiation of new
DNA synthesis for a normal cell. At 1 min, multiple replica-
tion origins are activated (Fig. 4, a1 and a2), shown as well-
separated nanoscale clusters with a small number of newly
synthesized DNA molecules. Next, two simultaneous pro-
cesses can be observed—1) sequential activation of new
sets of replication origins, shown as spatially separated
nanosized clusters similar to those detected at 1 min
(Fig. 4, a2–b2); and 2) heterogeneous spatial expansion of
newly synthesized DNA at each activated origin, shown as
larger and brighter clusters (Fig. 4, c2–d2). As DNA replica-
tion continues, we observe further spatial expansion and
more accumulated molecules (higher intensity in Fig. 4,
c1d1 and c2–d2) for each cluster, and many of them
form aggregated clusters at
10 min, which are likely those
replication foci observed under conventional light micro-
scopy (6,35).
We then compare the superresolution images of replica-
tion structure in normal, precancerous, and tumorigenic
cell lines after 1, 2, 5, and 10 min of exposure to EdU.
Whereas all three cell lines exhibit a similar progressive
change of replication structures—gradually increased num-
ber of replication origins and spatially expanded clusters
during new DNA synthesis—their timing is different.
More replication foci can be seen in precancerous cells
compared to those from normal cells at the early onset of
DNA synthesis (1–2 min shown in Fig. 4, e2–f2). Although
the gradual formation of large and heterogeneous replica-
tion foci can be observed in all three cell lines, those from
tumorigenic cells are formed much faster with larger size
and denser compaction (Fig. 4, j2–i2) than those from pre-
cancerous (Fig. 4, f2–h2) and normal cells (Fig. 4, b2–d2).
We next quantitatively characterize the temporal change
of replication structures in these three cell lines, as shown
in Fig. 5. A cluster analysis program is developed (described

FIGURE 5 Given here is the average histogram distribution of (a)–(c) spot number per cluster and (d)–(f) cluster size of normal, precancerous, and tumor-
igenic cells at 1, 2, 5, and 10 min of exposure to EdU. To eliminate the effect of nucleus size, the y axis uses cluster density, calculated as the number of
clusters per bin divided by the area of the cell nucleus. The number of cells averaged per group is shown in Table 1. (g–i) Shown here is the statistical
mean of cluster size, spot number per cluster, and cluster density at 1, 2, 5, and 10 min of exposure to EdU in normal, precancerous, and tumorigenic cells,
respectively. The error bar is 95% confidence interval of the mean. The detailed statistical analysis is shown in Tables 1 and S3–S5. To see this figure in color,
go online.

Biophysical Journal 112, 2196–2208, May 23, 2017 2205

Ma et al.
in Materials and Methods; Fig. S7) to automatically identify
each cluster and estimate the cluster size, spot number per
cluster, and cluster density (i.e., number of clusters per
mm2) on each superresolution image. Due to the complex
photophysics of Alexa Fluor 647, the spot number per clus-
ter is the actual localization number per cluster (for 20 ms
exposure time) without overblinking correction. Therefore,
this value is not equivalent to the number of labeled mole-
cules per cluster. Given the same imaging condition (e.g.,
illumination power density, number of imaging frames,
exposure time, and no activation with shorter wavelength)
used for all three cell lines and time points, their statistical
difference that relates to the number of labeled molecules
per cluster for these three cell lines can be compared.
Fig. 5, a–f, shows the temporal alteration of the average dis-
tribution of cluster size and spot number per cluster in
normal, precancerous, and tumorigenic cells. Fig. 5, g–i,
shows the statistical mean of cluster size, spot number per
cluster, and cluster density over time in the three cell lines
and the details are summarized in Table 1. These results
collectively quantify the structural changes of DNA replica-
tion in cells at different stages of neoplastic progression.
Right after the initiation of new DNA synthesis (1 min),
multiple replication origins are activated with an average
size of
29 nm at a low spatial density of
1.8 clusters
per mm2 in normal cells (as shown in Table 1). All three
cell lines exhibit a narrow distribution of cluster size
(
28–41 nm) with a low spot number per cluster, as shown
in blue curves of Fig. 5, a–f. Such characteristic clusters of
replication origin are present throughout the observed time
frame of 10 min. In precancerous and tumorigenic cells,
the activation of new replication origins in the first 2 min
is much faster compared to that of normal cells, reflected
as the significantly higher cluster density (
2–3 times the
density of that of normal cells) shown in Fig. 5 i, a manifes-
tation of their faster proliferation rate. The cluster density
reaches a plateau in all three cell lines, where precancerous
cells exhibit the highest cluster density at
5–10 min
(
25% higher, p value <0.0001), suggesting a character-
istic structural alteration in DNA replication of precancer-
ous cells. The cluster density in tumorigenic cells is
similar to that of precancerous cells in the first 2 min, but
significant aggregation of replication clusters at 5–10 min
in tumorigenic cells may bury some newly activated replica-
tion origins, resulting in some undetected replication origin
clusters (or lower cluster density). This data suggests faster
activation of replication origins as a possible characteristic
of tumorigenic cells as well.
Another process that occurs simultaneously with the
sequential activation of new replication origins is new
DNA synthesis around each activated origin, manifested
as spatial expansion of each cluster with larger size and
significantly increased spot number per cluster (suggesting
a fast-growing number of newly synthesized DNA mole-
cules) at each replication site over time, as shown in
2206 Biophysical Journal 112, 2196–2208, May 23, 2017
Fig. 5, g and h. The distribution of cluster size and spot num-
ber per cluster is also highly heterogeneous (wider distribu-
tion), as shown in Fig. 5, a–f, especially for those at 5 and
10 min, suggesting a diverse timing in the activation of
replication origins and new DNA synthesis on each replica-
tion site. Precancerous cells present only slightly larger
cluster size and spot number per cluster over time compared
to those of normal cells, and reach a similar value at 10 min
with no statistical significance (as shown in Table S3). On
the other hand, tumorigenic cells exhibit the highest rate
of increase for spot number per cluster (73% increase versus
normal cells at 10 min, p value <0.0001; 40% increase
versus precancerous cells at 10 min, p value ¼ 0.0002, as
shown in Table S3) and cluster size (31% increase versus
normal cells at 10 min, p value <0.0001; 14% increase
versus precancerous cells at 10 min, p value ¼ 0.0015).
This result suggests that tumorigenic cells exhibit a signifi-
cantly faster rate of new DNA synthesis at each replication
site.
DISCUSSION
The essence of our marker-assisted 3D drift correction
method is the ability to derive 3D position entirely from
the 2D point spread function of fiducial markers (gold or
fluorescent nanoparticles) on the coverslip of the sample,
thus eliminating the need for any additional optics, light
source, or detector. It can achieve superior routine perfor-
mance comparable to other superresolution imaging sys-
tems that require additional optics to derive the axial
position (5,10,11) and a similar precision compared to the
state of the art that is based on a complicated and expensive
optical setup (14) without being limited to the coverslip
surface. Besides the advantage of simplicity, we also
demonstrate that, in routine superresolution imaging exper-
iments, the MA-based drift correction method is more reli-
able compared to posterior image processing in correcting
lateral drift, especially in the case of a low number of
labeled molecules or altered subcellular structures during
the long image acquisition process. Our present work fo-
cuses on the use of gold nanoparticles as fiducial markers
without the need for an additional light source. However,
the same principle can also be extended to fluorescent
nanoparticles as fiducial markers if other laser sources are
easily available. Besides superresolution imaging demon-
strated in this report, this system can also be applied to
other fluorescence imaging applications that require main-
taining nanometer precision in 3D, such as single-particle
tracking.
With this MA-based SMLM imaging system, we investi-
gate characteristic spatiotemporal structural alteration of
DNA replication in which the newly synthesized DNA
changes from low abundance to highly condensed structure,
in cells undergoing different stages of neoplastic progres-
sion. Neoplastic transformation from normal cells to its

malignant form generally undergoes premalignant and
tumorigenic stage. In the premalignant stage, cells are often
proliferating with an increased risk to progress into malig-
nancy, but not tumorigenic. In the tumorigenic stage, cells
undergo uncontrolled cell proliferation that leads to
malignant tumor. Uncontrolled cell proliferation is a well-
recognized hallmark of cancer, but proliferation is not
necessarily a cancer marker. Hence, understanding the
distinction of DNA replication at different stages of cancer
development can provide significant insights into designing
effective diagnostic, therapeutic, and preventive strategies.
Although numerous molecular markers (e.g., genetic, prote-
omic, gene expression) (16–19) have been identified to
delineate precancerous from tumorigenic cells, their charac-
teristic structural differences in the fundamental process of
DNA replication and high-order chromatin organization
remain unclear. The robust performance of our MA-based
SMLM system is crucial to accurately characterize the sub-
tle structural alteration of DNA replication at the early onset
of new DNA synthesis at a high temporal and spatial resolu-
tion. DNA replication process covers a range of situations
from a small number of molecules at the initiation of new
DNA synthesis to highly condensed structures. The DNA
structure for very short pulse-labeling (i.e., 1–2 min) was
often too weak to be detected (34). Further, the labeling
size is another important technical factor for precise charac-
terization of replication structure. Most previous studies
used anti-BrdU or anti-PCNA immunostaining in which
each IgG adds
7 nm and the commonly used primary/sec-
ondary antibody labeling adds additional
14 nm, a size
comparable to replication structure, which can significantly
lower the labeling density and affect the accuracy in nano-
scale structural assessment. Our use of EdU-Alexa 647 label
has a fluorescent dye directly attached to the newly synthe-
sized DNA with negligible label size (<1 nm).
Our quantitative superresolution imaging shows that the
early phase of DNA replication is characterized as two
distinct processes that occur simultaneously—1) sequential
activation of multiple replication origins and 2) new DNA
synthesis at each activated replication site where replicated
DNA is folded into nanoscale clusters. The timing of these
two processes at individual replication sites is asynchro-
nous, reflected as heterogeneous size and amount of accu-
mulated newly synthesized DNA at each replication site.
In cells at different stages of neoplastic progression, both
precancerous and tumorigenic cells present a faster prolifer-
ation rate than normal cells. The faster proliferation of pre-
cancerous cells appears to be reflected as a faster temporal
activation of replication origins, shown as slightly increased
spatial expansion and spot number per cluster at each repli-
cation site; whereas the faster proliferation in tumorigenic
cells is reflected as a significantly faster rate of new DNA
synthesis. Previous studies have proposed a similar model
of two simultaneous processes in replication (37,38), but
their difference in neoplastic progression was not studied.
Simple Nanometer Drift Correction Method
Our findings suggest a new potential method to distinguish
tumorigenic cells from nontumorigenic proliferative cells
using fluorophore-tagged EdU superresolution imaging.
Future work on superresolution imaging of DNA replication
after functional treatment and in live cells is needed to
further our understanding of functional significance of chro-
matin organization to DNA replication and gene expression
in cancer initiation and progression.
SUPPORTING MATERIAL
Eight figures and five tables are available at http://www.biophysj.org/
biophysj/supplemental/S0006-3495(17)30441-1.
AUTHOR CONTRIBUTIONS
H.M. and Y.L. designed the research. H.M. and Y.L. developed the technical
methods, hardware, and software. Y.H. contributed cell lines. J.X. and J.J.
prepared the sample. H.M. performed data acquisition and data analysis.
H.M. and Y.L. summarized the data and wrote the paper based on the input
from all co-authors. Y.L. supervised the research.
ACKNOWLEDGMENTS
This work is supported by National Institutes of Health (NIH) grant Nos.
R01EB016657 and R01CA185363.
REFERENCES
1. Hess, S. T., T. P. K. Girirajan, and M. D. Mason. 2006. Ultra-high res-
olution imaging by fluorescence photoactivation localization micro-
scopy. Biophys. J. 91:4258–4272.
2. Betzig, E., G. H. Patterson, ., H. F. Hess. 2006. Imaging intracel-
lular fluorescent proteins at nanometer resolution. Science.
313:1642–1645.
3. Heilemann, M., S. van de Linde, ., M. Sauer. 2008. Subdiffraction-
resolution fluorescence imaging with conventional fluorescent probes.
Angew. Chem. Int. Ed. Engl. 47:6172–6176.
4. Rust, M. J., M. Bates, and X. Zhuang. 2006. Sub-diffraction-limit im-
aging by stochastic optical reconstruction microscopy (STORM). Nat.
Methods. 3:793–795.
5. Mlodzianoski, M. J., J. M. Schreiner, ., J. Bewersdorf. 2011. Sample
drift correction in 3D fluorescence photoactivation localization micro-
scopy. Opt. Express. 19:15009–15019.
6. Wang, Y., J. Schnitzbauer, ., B. Huang. 2014. Localization
events-based sample drift correction for localization microscopy
with redundant cross-correlation algorithm. Opt. Express. 22:
15982–15991.
7. Elmokadem, A., and J. Yu. 2015. Optimal drift correction for superre-
solution localization microscopy with Bayesian inference. Biophys. J.
109:1772–1780.
8. Geisler, C., T. Hotz, ., A. Egner. 2012. Drift estimation for single
marker switching based imaging schemes. Opt. Express. 20:7274–
7289.
9. McGorty, R., D. Kamiyama, and B. Huang. 2013. Active microscope
stabilization in three dimensions using image correlation. Opt. Nano-
scopy. 2:3.
10. Bates, M., S. A. Jones, and X. Zhuang. 2013. Stochastic optical recon-
struction microscopy (STORM): a method for superresolution fluores-
cence imaging. Cold Spring Harb. Protoc. 2013:498–520.
Biophysical Journal 112, 2196–2208, May 23, 2017 2207
Ma et al.
11. York, A. G., A. Ghitani, ., H. Shroff. 2011. Confined activation and
subdiffractive localization enables whole-cell PALM with genetically
expressed probes. Nat. Methods. 8:327–333.
12. Grover, G., W. Mohrman, and R. Piestun. 2015. Real-time adaptive
drift correction for super-resolution localization microscopy. Opt. Ex-
press. 23:23887–23898.
13. Lee, S. H., M. Baday, ., P. R. Selvin. 2012. Using fixed fiduciary
markers for stage drift correction. Opt. Express. 20:12177–12183.
14. Bon, P., N. Bourg, ., S. L
evêque-Fort. 2015. Three-dimensional nano-
metre localization of nanoparticles to enhance super-resolution micro-
scopy. Nat. Commun. 6:7764.
15. Santner, S. J., P. J. Dawson, ., F. R. Miller. 2001. Malignant
MCF10CA1 cell lines derived from premalignant human breast epithe-
lial MCF10AT cells. Breast Cancer Res. Treat. 65:101–110.
16. Worsham, M. J., G. Pals, ., S. R. Wolman. 2006. High-resolution
mapping of molecular events associated with immortalization, transfor-
mation, and progression to breast cancer in the MCF10 model. Breast
Cancer Res. Treat. 96:177–186.
17. Marella, N. V., K. S. Malyavantham, ., R. Berezney. 2009. Cytoge-
netic and cDNA microarray expression analysis of MCF10 human
breast cancer progression cell lines. Cancer Res. 69:5946–5953.
18. Kadota, M., H. H. Yang, ., M. P. Lee. 2010. Delineating genetic alter-
ations for tumor progression in the MCF10A series of breast cancer cell
lines. PLoS One. 5:e9201.
19. Shaw, P. G., R. Chaerkady, ., N. E. Davidson. 2013. Integrated prote-
omic and metabolic analysis of breast cancer progression. PLoS One.
8:e76220.
20. Rieger, B., and S. Stallinga. 2014. The lateral and axial localization un-
certainty in super-resolution light microscopy. ChemPhysChem.
15:664–670.
21. Huang, Z.-L., H. Zhu, ., S. Zeng. 2011. Localization-based super-res-
olution microscopy with an sCMOS camera. Opt. Express. 19:19156–
19168.
22. Long, F., S. Q. Zeng, and Z. L. Huang. 2014. Effects of fixed pattern
noise on single molecule localization microscopy. Phys. Chem.
Chem. Phys. 16:21586–21594.
23. Huang, F., T. M. Hartwich, ., J. Bewersdorf. 2013. Video-rate nano-
scopy using sCMOS camera-specific single-molecule localization al-
gorithms. Nat. Methods. 10:653–658.
24. Debnath, J., S. K. Muthuswamy, and J. S. Brugge. 2003. Morphogen-
esis and oncogenesis of MCF-10A mammary epithelial acini grown
2208 Biophysical Journal 112, 2196–2208, May 23, 2017
in three-dimensional basement membrane cultures. Methods. 30:
256–268.
25. Li, B., H. Zhao, ., M. Kimmel. 2014. Different rates of DNA repli-
cation at early versus late S-phase sections: multiscale modeling
of stochastic events related to DNA content/EdU (5-ethynyl-
20 deoxyuridine) incorporation distributions. Cytometry A. 85:785–797.
26. Nakayasu, H., and R. Berezney. 1989. Mapping replicational sites in
the eucaryotic cell nucleus. J. Cell Biol. 108:1–11.
27. McLachlan, G., and D. Peel. 2000. Finite Mixture Models. Wiley,
New York.
28. Nieuwenhuizen, R. P., K. A. Lidke, ., B. Rieger. 2013. Measuring im-
age resolution in optical nanoscopy. Nat. Methods. 10:557–562.
29. Berezney, R., D. D. Dubey, and J. A. Huberman. 2000. Heterogeneity
of eukaryotic replicons, replicon clusters, and replication foci. Chromo-
soma. 108:471–484.
30. W€aldchen, S., J. Lehmann, ., M. Sauer. 2015. Light-induced cell
damage in live-cell super-resolution microscopy. Sci. Rep. 5:15348.
31. Nakamura, H., T. Morita, and C. Sato. 1986. Structural organizations of
replicon domains during DNA synthetic phase in the mammalian nu-
cleus. Exp. Cell Res. 165:291–297.
32. Sporbert, A., A. Gahl, ., M. C. Cardoso. 2002. DNA polymerase
clamp shows little turnover at established replication sites but sequen-
tial de novo assembly at adjacent origin clusters. Mol. Cell. 10:
1355–1365.
33. Koberna, K., A. Ligasová, ., R. Berezney. 2005. Electron microscopy
of DNA replication in 3-D: evidence for similar-sized replication foci
throughout S-phase. J. Cell. Biochem. 94:126–138.
34. Ligasová, A., I. Raska, and K. Koberna. 2009. Organization of human
replicon: singles or zipping couples? J. Struct. Biol. 165:204–213.
35. Baddeley, D., V. O. Chagin, ., M. C. Cardoso. 2010. Measurement of
replication structures at the nanometer scale using super-resolution
light microscopy. Nucleic Acids Res. 38:e8.
36. Zessin, P. J., K. Finan, and M. Heilemann. 2012. Super-resolution fluo-
rescence imaging of chromosomal DNA. J. Struct. Biol. 177:344–348.
37. Chagin, V. O., J. H. Stear, and M. C. Cardoso. 2010. Organization of
DNA replication. Cold Spring Harb. Perspect. Biol. 2:a000737.
38. Guilbaud, G., A. Rappailles, ., O. Hyrien. 2011. Evidence for sequen-
tial and increasing activation of replication origins along replication
timing gradients in the human genome. PLOS Comput. Biol. 7:
e1002322.
Biophysical Journal, Volume 112

Supplemental Information

A Simple Marker-Assisted 3D Nanometer Drift Correction Method for

Superresolution Microscopy
Hongqiang Ma, Jianquan Xu, Jingyi Jin, Yi Huang, and Yang Liu
Supplementary Figures

Figure S1. Simulated point spread function (PSF) with Born & Wolf PSF model (1). Left panel: The
intensity distribution of PSF on x-z (or y-z) plane. Right panels: The corresponding PSF at each axial (z)
position.

Figure S2. Measured PSF pattern of (a-b) a fluorescence bead (100 nm) and (c-d) a gold nanoparticle
(100 nm) at different axial positions. The fluorescent bead exhibits a more symmetric PSF than that of the
gold nanoparticle.

2
 
  
Figure S3. Calibration curves ((a) Fwx  z  , (b) Fwy  z  , (c) Fxc  z  and (d) Fyc  z  ,)) of 8 random
distributed gold nanoparticles as a function of 17 different axial positons (step size of 50 nm). The proper
number of axial positions and the axial range depend on how fast the axial drift occurs in the fluorescence
microscopy system. We recommend setting the axial range to be larger than twice the maximum axial
drift of the system in each correction cycle, defined as the time interval between two drift corrections (4
seconds in this work). In the experiments shown in Figs. 1-3, we used 17 axial positions with an axial
range from -400nm to 400nm at a step distance of 50 nm to generate the calibration curve. If the imaging
plane of the fiduciary markers is located at z = 0 nm, this range is capable of correcting drift up to 400 nm
per correction cycle (~6 µm/min). Supplementary Table S1 shows the performance of our marker-assisted
drift correction method based on the calibration curve using 17 axial positions at the range of ± 400 nm.
In practice, we noticed that the axial drift of our localization microscope is less than 20 nm per
correction cycle (<300 nm/min) (as shown in Fig. 2g), so a long axial range of ± 400 nm is actually not
necessary, a shorter axial range with fewer number of positions should be sufficient for online drift
correction in our system, which is shown in Fig. S4.

3
 
Figure S4. (a) Calibration curves ( Fwx  z  ) of 7 random distributed gold nanoparticles and (b) the
corresponding precision to estimate 3D positions. This is an example of a simplified version of our
marker-assisted drift correction method using only 9 axial positions with an axial range from 100nm to
500nm. The imaging plane of the markers was chosen to be at 300 nm which is the axial position with
best precision to estimate 3D positons based on Fig. 1d to build the calibration curve. Due to the shorter
axial range, less number of axial positions and markers compared to Fig. S3, this simplified version can
be more convenient and efficient in practice. The performance of this simplified strategy is shown in
Table S2, with a correction precision of approximately <2 nm in the lateral dimension and <5 nm in the
axial dimension.

Figure S5. The precision of estimated 3D positions with only two bright markers (35000 photons) at a
separation of 100 nm in the axial direction. This is an example with one of the worst performance in the
case of only two markers used for drift correction, and their precision is still < 2 nm and < 5 nm in the
lateral and axial dimension, respectively.

4
 

Build calibration curves based
on fiduciary markers
Acquire axial image stack
of fiduciary markers
Use 2D Gaussian fitting to retrieve
PSF width and lateral bias for each
fiduciary marker
Fit a polynomial function to build
the calibration curves for 3D
positions based on the PSF width
and lateral bias
Figure S6. Flowchart for building calibration curves and online drift correction for marker-assisted drift
correction.
Figure S7. Flowchart for both super-resolution image reconstruction and subsequent cluster analysis.
Online drift estimation and
correction
Move to the focal plane of fiduciary
markers and acquire a 2D image
Use 2D Gaussian fitting to retrieve
PSF width and lateral bias
Retrieve PSF width and lateral bias
based on the calibration curves
Calculate the current 3D position of the
fiduciary markers based on their
retrieved PSF width and lateral bias
Correct the drift and move back to
the focal plane of the sample
5
Figure S8. (a) The fiduciary marker and (b) the reconstructed super-resolution images of cancer cell
nucleus labeled with histone protein H2B-Alexa Fluor 647 in the same FOV before lateral drift correction
but after MA-based axial direction correction. The green arrow indicates that both markers and different
parts of the cell exhibit similar drift trajectories orientated toward the same direction. (c) The
reconstructed image after MA-based 3D drift correction shows round-like shaped clusters. (d) The
fiduciary marker and (e) reconstructed super-resolution image of another cancer cell nucleus in the same
FOV before lateral drift correction but after MA-based axial direction. (f) The reconstructed image after
MA-based 3D drift correction. Please note that the direction of the local drift trajectory within each green
box is estimated by RCC method and indicated by the green arrow. Green arrows show that different parts
of the cell exhibit different drift trajectories toward different orientations; while the markers at the similar
locations in (d) in the same FOV during the same image acquisition period exhibit the drift trajectories
toward the same direction. The yellow arrows within the yellow boxes in (f) indicate the vector difference
between the drift trajectories for Regions 1 and 2 of the cell in (e) and the drift trajectory from the marker
within Region 1 in (d). However, no apparent vector difference is seen in other regions of the cell in (f).
This difference seen in yellow boxes of (f) suggests a possible cell movement or deformation during
image acquisition. The cell in (a-c) is the same cell as that in Figs. 3(a-d) and the cells in (d-f) are the
same cell as that in Fig. 3(e-g) from the main manuscript.

6
 
Supplementary Table S1. Drift correction performance based on calibration curves with an axial range
from -400nm to 400nm, in which the imaging plane of the markers is located at Z = 0 nm (exposure time
for each frame: 20ms).

Number of markers 8a 13 5 11 18 7
Mean photon number per
29000 24000 33000 24000 27000 36000
marker
Total number of imaging
40000 40000 40000 40000 60000 60000
frames
σX (nm) 1.3 1.1 1.4 1.4 1.2 1.2
σY (nm) 1.3 1.1 1.5 1.4 1.2 1.3
σZ (nm) 3 3.2 4.3 3.7 3.3 4.1
a
The detailed performance of this experiment is shown in Fig. 1 of the main manuscript.

Supplementary Table S2. Drift correction performance with calibration curve from 100nm to 500nm,
in which the imaging plane of the markers is at z = 300nm (exposure time for each frame: 20ms).

Number of markers 7 11 10

Mean photon number per marker 28000 27000 23000

Total number of imaging frames 40000 40000 60000

σX (nm) 1.4 1.2 1.3
σY (nm) 1.3 1.1 1.5
σZ (nm) 3.7 3.2 3.8

7
 
Supplementary Table S3. The statistical analysisa (p-valueb) of temporal changes of cluster size
between normal, pre-cancerous and tumorigenic cells.
Cluster size 1 min 2 min 5 min 10 min
Normal vs. 0.0031 0.0048 0.0011 0.589
Pre-cancerous ** ** ** No significance (NS)
Pre-cancerous vs. 0.0908 0.0022 0.0232 0.0015
Tumorigenic NS ** * **
Normal vs. 0.0007 <0.0001 0.0002 <0.0001
Tumorigenic *** **** *** ****
Supplementary Table S4. The statistical analysis (p-value) of temporal changes of spot number (or
localization number) per cluster between normal, pre-cancerous and tumorigenic cells.
Spot number per cluster 1 min 2 min 5 min 10 min

Normal vs. 0.0031 0.0048 0.0039 0.1227

Pre-cancerous ** ** ** NS
Pre-cancerous vs. 0.206 0.0698 0.0089 0.0002
Tumorigenic NS NS ** ***
Normal vs. 0.0007 0.0005 0.0002 <0.0001
Tumorigenic
*** *** *** ****
Supplementary Table S5. The statistical analysis (p-value) of temporal changes of cluster density
between normal, pre-cancerous and tumorigenic cells.
Cluster density 1 min 2 min 5 min 10 min

Normal vs. 0.0003 <0.0001 0.0016 <0.0001

Pre-cancerous
*** **** ** ****
Pre-cancerous vs. 0.27 0.22 0.0006 0.0002
Tumorigenic NS NS *** ***
Normal vs. 0.0007 0.0009 0.99 0.37
Tumorigenic *** *** NS NS
a.
The statistical analysis was performed using the mean value of cluster size, spot number per cluster and
cluster density from each cell and the total number of cells used in each cell type is shown in Table 1.
b.
*: p-value < 0.05; **: p-value < 0.01; ***: p-value < 0.001; ****: p-value < 0.0001, two-sided exact p-
values used.

Supporting References:
1. Born, M., and E. Wolf. 2003. Principles of Optics. Cambridge University Press. pp. 436–
445.

8