1. Introduction

Fluorescence super-resolution microscopy consists of several methods that have revolutionized biological research by elucidating nanometer scale details of biological phenomena with high target specificity. Among these methods, techniques such as (d)STORM [1,2], (f)PALM [3,4], and DNA-PAINT [5] permit nanometer resolutions on a standard widefield fluorescence microscope, making them easily accessible super-resolution techniques. These methods are all examples of single-molecule localization microscopy, a technique which provides super-resolution by localization of spatiotemporally independent fluorescent molecules (fluorophores) attached to the structure of interest.

A common strategy used to investigate the relative organization of different components in a single cell is to target these components with spectrally distinct fluorophores. Imaging spectrally distinct fluorophores requires either a specialized multicolor detector [6] or the use of multiple emission channels, which can lead to degradation in multicolor overlays due to optical path differences. Furthermore, different fluorophores exhibit different photophysical properties (e.g., on/off duty cycle, photobleaching rate, photon emission rate) in the same imaging buffer, leading to sub-optimal image quality for one or more of the cellular targets. Optimizing the buffer conditions to extract the best qualities for each fluorophore has proven to be hard to implement [7,8].

An alternative to simultaneously imaging multiple colors is to sequentially label and image components with the same label. The label can be chosen so that it has the most desirable photophysical properties for the given imaging modality. Several schemes have been developed around this idea, including Exchange-PAINT [9], sequential multi-target STORM [10], and sequential multi-target dSTORM [11]. These methods allow for multi-target super-resolution with a single emission path and optimal fluorophore selection on any (d)STORM, (f)PALM, or DNA-PAINT capable microscope. Although such techniques simplify the experimental design (by reducing/eliminating the need for careful selection of multiple fluorophores), they typically necessitate either imaging only a single region of interest (ROI) per sample or adding some fiducial markers to reliably return to the same ROI for each sequential imaging round. A complication of sequential imaging is that changes in sample morphology due to the stress of repeated buffer exchanges, incomplete/imperfect chemical fixation, and/or prolonged inter-target imaging durations may reduce the fidelity of multi-target super-resolution reconstructions in terms of their representation of the relative associations between cellular component being imaged. In some cases, shifts greater than 25 nm have been observed in the resulting overlays [12][present study].

Recently, an automated multi-target serial staining dSTORM microscope was demonstrated on a large number (15-16) of targets, with the resulting multi-target reconstruction overlays having roughly 25-30 nm shifts [12]. Automated localization microscopy systems with even higher throughput have also been demonstrated on 100s of cells [10] via automated cell detection for single targets, or on 10,000 cells for multiple targets in a split emission path scheme [13]. Such methods reduce the time burden on localization microscopy users by reducing imaging time and/or user interaction with the microscope.

In this work, we describe a custom-built automated sequential super-resolution microscope, which we term Sequential Super-resolution Microscope (SeqSRM), that is optimized for Alexa Fluor 647 (AF647) or spectrally similar fluorophores. The SeqSRM is capable of imaging multiple structures on 10s of cells. User interaction with the SeqSRM is limited to cell selection and buffer exchange during sequential relabeling steps. The SeqSRM allows for relatively high-throughput localization microscopy in an automated manner on multiple targets on multiple cells, without the need for split emission paths or sub-optimal fluorophore selection. We demonstrate the reliability of the SeqSRM by performing sequential dSTORM imaging of $\alpha$- and $\beta$-tubulin on 40 cells over approximately 32 hours, with user interaction with the microscope limited to approximately 1 hour of initial cell selection and two sample replacements taking 5 minutes each between the imaging and photobleaching rounds of imaging. $\alpha$- and $\beta$-tubulin are the repeated subunits of the same microtubule structures, and hence overlain images of these two targets should coincide. We further demonstrate the stability and flexibility to different localization microscopy modalities by imaging $\alpha$-tubulin on a single cell using DNA-PAINT for approximately 21 consecutive hours. Finally, we develop an algorithm to detect and quantify sample movements between rounds of sequential imaging, allowing us to define trust regions within our final super-resolved images using only brightfield images of the sample. In other words, we quantify local sample movements from brightfield images and mask regions of the super-resolved images where the sample appears to have changed.

2. Materials and methods

2.1 Optical design

The SeqSRM was designed for imaging of AF647. As such, all light sources and optical components were optimized for excitation and collection of the light emitted by AF647. The details of our design are provided below.

An sCMOS camera (C11440-22CU, Hamamatsu) was used to detect the AF647 emission light. A high power 647 nm fiber laser (2RU-VFL-P-500-647-B1R, MPB Communications) was used as the excitation laser. A 405 nm diode laser (DL5146-101S, Thorlabs) can optionally be used to accelerate the dark to fluorescent state transition if needed. A 100X silicone oil immersion objective (UPLSAPO100XS, Olympus) was used to collect the AF647 emission light. A 708/75 nm bandpass filter (FF01-708/75-25-D, Semrock) was placed in the AF647 emission light path. A simplified optical diagram is presented in Fig. 1.

 

Fig. 1. SeqSRM Optical Diagram. A 647 nm fiber laser ("647 Laser") passes through an electronically controllable shutter (Shutter) and a neutral density filter on an electronically controllable flip mount (ND Filter). A 405 nm diode laser (405 Laser) is combined with the 647 nm laser before both are passed through an active vibrating membrane diffuser (Diffuser) before being coupled into a multimode fiber (MM). The tip of the multimode fiber is then imaged to the sample plane. Fluorescence emission light is collected by a 100X objective (100X) and is passed through a dichroic mirror (DM) and an emission filter (EF) before being imaged onto the sensor (sCMOS). The mirror indicated by "M*" is placed in a plane conjugated to the back focal plane of the objective lens to allow for placement of point-spread function engineering components (e.g., a deformable mirror). All lens focal lengths are given in units of millimeters (e.g., "f=150" indicates a lens with focal length of 150 mm.))

Download Full Size | PDF

The excitation path is designed to image the tip of a multimode fiber onto the sample plane, yielding a highly uniform illumination profile within a well-defined field of view when combined with a laser speckle reducer (Optotune LSR-3005) [14]. The emission path is designed to relay the emission light onto the sCMOS camera. The mirror indicated as "M*" in Fig. 1 is conjugated to the back focal plane of the objective lens and can be swapped with a reflective type phase pattern generator (e.g., a deformable mirror or a spatial light modulator) to permit point spread function (PSF) engineering. Representative photographs of the microscope are shown in Fig. S7.

The microscope is additionally equipped with a 660 nm LED lamp (M660L3, Thorlabs) for brightfield transmission imaging of the sample. The light used for brightfield imaging is collected through the same path as the fluorescence emission path, with the brightfield images collected by the same sCMOS used for fluorescence imaging.

2.2 Sample holder and stage design

We constructed the SeqSRM around a Thorlabs NanoMax 3D stage (Thorlabs MAX381/M) which is used to move the sample up to 4 mm with stepper motors in X, Y, and Z and down to 5 nm with piezoelectric actuators. The sample is fixed to the stage and held in place above the objective lens with custom sample holders which place the sample aside the stage (see Fig. S7(c,d)). The sample holder was designed to reduce rotational remounting offsets, as the mount has a dovetail mated to a corresponding inlet attached to the SeqSRM sample stage. CAD drawings of the sample holder and the mating part attached to the Thorlabs NanoMax stage are provided in Fig.s S8-S10. A tip/tilt stage (Edmund Optics 125mm Metric Micrometer Tilt Stage, stock #66-555) was placed beneath the 3D stage to allow for accurate control of the sample tilt with respect to the objective lens.

2.3 Instrument control

The SeqSRM was controlled by custom-written MATLAB and MATLAB mex codes written in C. A graphical user interface (GUI) was implemented to simplify the use of the SeqSRM for end-users. A screenshot of the GUI is shown in Fig. S11. Additionally, a flowchart detailing the SeqSRM control workflow is shown in Fig. S12.

2.4 Sample preparation

All HeLa cell samples used in this manuscript were prepared identically before labeling. The samples were prepared as follows, with experiment specific methods provided in Sections 2.4.1 and 2.4.2 below. HeLa cells were plated on 25 mm round #1.5 coverslips (Warner instruments) and allowed to adhere overnight. The cells were washed with PBS and then chemically fixed in two steps. The cells were first treated with 0.6% paraformaldehyde (PFA), 0.1% glutaraldehyde (GA), and 0.25% Triton X-100 in PBS for 60 seconds. The cells were then fixed for 2.5 hours in 4% PFA and 0.2% GA in PBS. Autofluorescence was quenched using 0.1% sodium borohydride in PBS for 5 minutes. The cells were washed with PBS and then treated with 10 mM Tris in PBS twice for minutes each to quench the fixation chemicals. Potential non-specific binding sites were then blocked by treating the cells for 15 minutes in a solution of 5% BSA and 0.05% Triton X-100 in PBS.

2.4.1 $\alpha$- and $\beta$-tubulin dSTORM samples

HeLa cells prepared as described above were labeled with anti-$\beta$-tubulin-AF647 (abcam ab204686) at 2.5 $\mu$g/mL in PBS (total volume 0.5 mL), 2% BSA, and 0.05% Triton X-100 (PBS/BSA/TX) for 2 hours and washed three times for 5 minutes with PBS before placing in the dSTORM buffer. We note that our $\beta$-tubulin labeling efficiency tends to be less than that of $\alpha$-tubulin, hence we image $\beta$-tubulin first to reduce cross-talk. After imaging $\beta$-tubulin, remaining AF647 were photobleached and quenched as described in [11]. The sample was then relabeled with anti-$\alpha$-tubulin-AF647 (Sigma-Aldrich Cat. #05-829-AF647) at 2.5 $\mu$g/mL in PBS (total volume 0.5 mL), 2% BSA, and 0.05% Triton X-100 (PBS/BSA/TX) for 2 hours and washed three times for 5 minutes with PBS before placing in fresh dSTORM buffer for imaging.

For dSTORM imaging, the coverslip was mounted on an Attofluor cell chamber (A-7816, life technologies) with 1.5 ml of the imaging buffer. Imaging buffer consisted of an enzymatic oxygen scavenging system and primary thiol: 50 mM Tris, 10 mM NaCl, 10% w/v glucose, 168.8 U/ml glucose oxidase (Sigma #G2133), 1404 U/ml catalase (Sigma #C9332), and 30 mM 2-aminoethanethiol (MEA), pH 8. The chamber was sealed by placing an additional coverslip over the chamber. The oxygen scavenging reaction was allowed to proceed for 30 minutes at room temperature before the imaging started.

2.4.2 $\alpha$-tubulin DNA-PAINT sample

HeLa cells were fixed and treated as described above. The cells were then labeled with anti-$\alpha$-tubulin mouse monoclonal antibody (Sigma-Aldrich T6074) at 2.5 $\mu$g/mL for 1 hour and washed three times for 5 minutes with PBS. Secondary labeling was performed using a MASSIVE Photonics DNA-PAINT kit (MASSIVE-sdAB 2-PLEX with FluoTag-XM-QC Anti-Mouse IgG kappa light chain) following the manufacturer’s instructions.

For DNA-PAINT imaging, the coverslip was mounted on an Attofluor cell chamber (A-7816, life technologies) with 1.5 ml of the imaging solution. The imaging solution consisted of 50 pM Imager 1 (MASSIVE Photonics Imager 1 Cy3B, ATTO 655) in MASSIVE Photonics imaging buffer.

2.5 Data collection

Cells were selected manually by the user by first finding desirable ROIs in brightfield using the 660 nm lamp illumination ("660 Lamp" in Fig. 1). The desired focal plane was further refined using fluorescence imaging with the 647 nm laser. Once a cell was selected, a stack of brightfield images was collected at 21 focal planes 50 nm apart to define a reference stack, which was saved for later brightfield registration as in [15].

dSTORM imaging was performed by illuminating the sample using 1-3 kW/cm$^2$ 647 nm illumination and collecting data at 100 frames per second. Immediately before dSTORM data collection, each cell was illuminated by the 647 nm laser for approximately 10 seconds to convert the majority of the AF647 into a dark state. DNA-PAINT imaging was performed at roughly 1 kW/cm$^2$ and 20 frames per second. The optional 405 nm laser was not used for any of the data presented in this manuscript. Brightfield registration was performed before collection of each sequence of super-resolution data (approximately every 1-2 minutes), which involved moving the 3D sample stage to maintain X, Y, and Z alignment of the sample based on shifts estimated from brightfield images (see [15] for a detailed description of this process). Before and after collection of each complete sequence of super-resolution data, a sequence of brightfield images at the focal plane was collected and saved for later use in drift-correction and sample movement characterization.

2.6 Sample mounting and remounting

For all experiments presented in this work, samples were not removed from the SeqSRM after initial mounting. For example, when relabeling the dSTORM sample with anti-$\alpha$-tubulin-AF647 for the second round of imaging described above, we left the sample on the SeqSRM and simply removed (and subsequently replaced) the coverslip used to seal the Attofluor cell chamber. This was done to ensure maximum mechanical stability of sample throughout the experiment.

We note, however, that we designed the SeqSRM to permit removal and remounting of the sample if necessitated by the experimental protocols. The custom sample mount shown in Fig. S7(d) and Fig. S8 securely fixes the Attofluor cell chamber in place with spring loaded fasteners (indicated by green arrows in Fig. S7(d), and shown as a CAD drawing in Fig. S9). The entire mount can be removed and replaced with the Attofluor cell chamber remaining in place. Upon remounting and securing with the set screws (indicated by magenta arrows in Fig. S7(d)), the user can click the "Find Coverslip Offset" button shown in Fig. S11, which instructs the SeqSRM stage to return to the (X, Y, Z) position of a reference cell in the sample. The user is then prompted to click the center of the cell as seen in a brightfield image to guide additional coarse-tuning of the remounted position. Remaining linear (X, Y, Z) remounting offsets are corrected for by the brightfield registration procedure, which is performed before each sequence of data collection. Rotational remounting offsets are inherently reduced by the mechanical design of the sample mount as described in Section 2.2.

2.7 Super-resolution image analysis

Data was analyzed via a 2D localization algorithm based on maximum likelihood estimation [16]. To eliminate false and low-quality localizations, we performed thresholds on photons (>= 200), localization error (>= 15 nm), PSF width ($50 \, \mathrm {nm} <= \sigma _\mathrm {PSF} <= 200 \, \mathrm {nm}$), and goodness of fit as defined by a p-value (accept localizations with p-value >= 0.01) [17]. The accepted emitters were used to reconstruct the super-resolution image. Gaussian super-resolution images were generated by placing a 2D Gaussian at the position of the localized emitter, with the covariance determined by the computed Cramér-Rao Lower Bounds (CRLB) for the X and Y localizations. Circle super-resolution images were generated by placing a circle at the position of the localized emitter with the circle radius being the square root of the average CRLB for the X and Y position estimates. To improve the contrast and appearance of the super-resolution images in Fig. 3(a), Fig. 4, and Fig. S1, pixel intensities were thresholded such that pixels brighter or dimmer than 95% of other pixels were reset to the 95-th and 5-th percentile pixel intensities, respectively.

Repeated localizations of the same emitter for the same blinking event were combined using the algorithm presented in [18]. Residual X and Y drift was corrected for by subtracting a linear drift model defined piecewise between each round of brightfield registration. The linear drift model was estimated by computing the shift between brightfield images taken in the focal plane immediately before and after each round of brightfield registration, with shift estimation performed using a modified version of the brightfield registration algorithm presented in [15] (see Supplement 1 Section A).

2.8 Defining trust regions in super-resolution images using brightfield images

As described in Section 2.5, we collect brightfield images at the focal plane immediately before each sequence of super-resolution data is collected. We use these brightfield images to define the trust regions as follows. We divide each of the brightfield images into a uniform grid of 32x32 pixel sub-ROIs (corresponding to 3.1$\mu$m × 3.1$\mu$m regions in the sample plane). Within each sub-ROI, we compute the shift between that sub-ROI and the corresponding sub-ROI from the opposite label (e.g., we compute the shift between the pre-sequence 1 sub-ROI collected during $\alpha$-tubulin imaging and the pre-sequence 1 sub-ROI collected during $\beta$-tubulin imaging). A description of the algorithm we developed to compute the sub-pixel brightfield shifts is given in Supplement 1 Section A. If we collect $N$ sequences of super-resolution data, this yields $N$ inter-label shifts for each sub-ROI. We then compute the median of the $N$ shifts for each sub-ROI and define that to be the apparent inter-label shift for that sub-ROI. We then use the inter-label shifts computed for each brightfield image sub-ROI to mask out the corresponding sub-ROIs in the final super-resolution reconstruction images. A flowchart illustrating the process of trust region generation for two-target data is shown in Fig. 2. Additional details regarding the selection of an optimal sub-ROI size, as well as suggestions for application to single-label data and data with >2 labels, are provided in Supplement 1 Section B.

 

Fig. 2. Trust region generation flowchart. Step-by-step flowchart illustrating the procedure used to generate trust regions for two-target data. In Step 1, we collect and save a brightfield image immediately before each sequence of super-resolution data is collected for each label. In Step 2, we divide the $N$ brightfield images for each of the labels into 32x32 pixel sub-ROIs and compute the shifts between corresponding sub-ROIs in each sequence of each label. In Step 3, we compute the median of the $N$ resulting shifts within each sub-ROI and generate a mask by comparing the median shifts to a user-defined threshold. Finally, in Step 4 we use the mask to identify regions of the super-resolved multi-target overlay that we deem trustworthy.

Download Full Size | PDF

2.8.1 Validation of trust region selection

To verify the application of our trust region definitions to super-resolution overlay reconstructions of multiple targets, we performed a validation experiment by sequentially imaging $\alpha$-tubulin for two rounds of imaging using DNA-PAINT. We chose this experiment to ensure that our final multi-target super-resolution overlay images exactly correspond to imaging the same structure twice (in contrast to our sequential dSTORM results of $\alpha$- and $\beta$-tubulin, in which the correlation between the resulting super-resolution localizations of each label is reduced by the incomplete labeling efficiency). DNA-PAINT imaging was performed as described in Section 2.4.2. Since our trust region definitions are based on shifts computed from the brightfield images, we compare the computed brightfield shifts within sub-ROIs to the corresponding super-resolution shifts computed between the multi-target localizations (which are used to reconstruct the multi-target super-resolution overlay images) within the same sub-ROIs.

To determine the shifts in the final multi-target super-resolution reconstruction images, we compute a coordinate-based shift between localizations within each sub-ROI. To compute the coordinate based shift, we use the thresholded nearest-neighbor cost function described in [15]. We emphasize that these coordinate-based super-resolution shifts are not used during the definition of the trust regions, and are only used for comparison in this experiment. Briefly, we do this by minimizing the sum of the thresholded distances between nearest-neighbor localizations from the first round of imaging to the second with respect to a linear shift model, where we use the distance threshold of $l=2$ pixels (191 nm) suggested in [15]. The linear shift that minimizes this cost function is then taken to be the super-resolution overlay shift for that sub-ROI.

We plot the median brightfield shift within each sub-ROI across all sequences of data collection versus the super-resolution overlay shift (Fig. 3(b)). We then perform a weighted least-squares fit with the model $\text {BF shift} = \text {slope}\cdot \text {SR shift} + \text {offset}$ to quantify the correlation between the two shifts, with the weighting factor being the inverse variance of the brightfield shifts within each sub-ROI.

 

Fig. 3. Local brightfield shifts versus super-resolution shifts. (a) Representative super-resolution overlay result from imaging $\alpha$-tubulin twice sequentially (first round in magenta, second round in green, with overlapping localizations appearing white due to the pixel sum of the magenta and green colors in reconstruction images) on three separate cells as described in Section 2.8. Yellow arrows represent the brightfield shifts computed within each sub-ROI and cyan arrows represent the corresponding super-resolution shifts, with the shift magnitudes scaled by a factor of 25 to improve visibility. The super-resolution image was also rescaled as described in Section 2.7 to improve visual contrast. (b) Local brightfield shifts within 32x32 pixel sub-ROIs versus the corresponding super-resolution shifts for the 3 cells imaged as described in Section 2.8, where a representative example is shown in (a). The solid red line indicates a weighted least-squares fit for the shown data, with the weighting being the inverse variance of the brightfield shifts. The data shown includes approximately 81% of the total sub-ROI shifts (i.e., shifts within 50 nm shown in the plot), with the outliers excluded from the plot attributed to sub-ROIs falling outside of the cell. Data points indicated by hollow black circles represent points with brightfield shift variances greater than the 513 nm$^2$ scale of the colorbar. (c) Cumulative distribution of brightfield shift variances color-coding points in (b), with the black dotted line indicating the visual cutoff of the colorbar. The white scale bar in (a) is 3 $\mu$m wide.

Download Full Size | PDF

3. Results

3.1 Validation of trust region definitions

$\alpha$-tubulin was imaged twice in a row using DNA-PAINT as described in Section 2.4.2 on three user-selected cells. The brightfield shift within 32x32 pixel sub-ROIs as well as the corresponding super-resolution coordinate shifts were computed as described in Section 2.8. An example super-resolution overlay is shown in Fig. 3(a), with magnified brightfield and super-resolution shift vectors overlain onto each sub-ROI. A plot of the brightfield versus super-resolution overlay shift magnitudes within each sub-ROI for all three cells is shown in Fig. 3(b), where approximately 81% of the total sub-ROI shifts are shown within the scale of the plot (we attribute the remaining outliers to sub-ROIs with poor brightfield contrast and/or regions with very few localizations). A linear fit weighted by the inverse of the brightfield shift variance gives a slope of 0.997, and the Pearson correlation coefficient was computed to be 0.63. Notably, points that fall further from the fit line tend to have a higher brightfield variance.

3.2 Sequential dSTORM imaging of $\alpha$- and $\beta$-tubulin

$\alpha$- and $\beta$-tubulin were imaged sequentially on a total of 40 user-selected HeLa cells as described in Section 2.4.1. Initial cell selection required approximately 50 minutes (contiguous) of user interaction with the microscope. The microscope then automatically returned to the user-selected cell ROIs and imaged $\beta$-tubulin, with the total acquisition taking approximately 10 hours. The photobleaching round required an additional 5 minutes of user interaction with the microscope, followed by automated photobleaching of each cell ROI, which took approximately 5 hours. Finally, an additional 5 minutes of user interaction with the microscope was required after sequential labeling and sample preparation for imaging the second structure ($\alpha$-tubulin). A selection of super-resolved overlay results are shown in Fig. 4, with $\beta$-tubulin shown in green and $\alpha$-tubulin shown in magenta. The results in Fig. 4 were masked to ignore sub-ROIs with apparent shifts greater than 25 nm, with the masked out regions shown as empty black squares. The results for all 40 cells are shown in Fig. S1.

 

Fig. 4. $\beta$- and $\alpha$-tubulin dSTORM. (a) Selected results from sequential imaging of (green) $\beta$- and (magenta) $\alpha$-tubulin on 40 cells/ROIs on a single coverslip. Pixels corresponding to both $\beta$- and $\alpha$-tubulin localizations appear white due to the sum of the color channels in the reconstruction image. Super-resolution localizations are represented by Gaussians with standard deviations equal to the estimated localization precision for each localization. Sub-ROIs were masked out in areas with shifts greater than 25 nm apparent from collected brightfield images. The white scale bar is 3 $\mu$m wide. The full set of 40 cells is provided in the Supplement 1. (b) Zoom in of selected sub-ROIs (indicated by yellow boxes) of the images in (a). (c) Image sub-ROIs that were masked out in (a). (d) Zoom in of selected sub-ROIs (indicated by yellow boxes) of the images in (c). The white scale bar is 3 $\mu$m wide. The yellow sub-ROIs are 1.1$\mu$m × 1.1$\mu$m. All images were rescaled as described in Section 2.7 to improve visual contrast.

Download Full Size | PDF

3.3 Long term DNA-PAINT imaging of $\alpha$-tubulin

$\alpha$-tubulin was imaged using DNA-PAINT as described in Section 2.4.2 for approximately 21 contiguous hours on a single cell. User interaction with the microscope was limited to approximately 5 minutes for cell selection, with the remaining acquisition running automatically without user intervention. The results are shown in Fig. S2 as circle super-resolution images (circles with radii equal to localization precision placed at each localization coordinate) color coded from blue to yellow with increasing time. The super-resolution image was masked to hide sub-ROIs with shifts greater than 15 nm identified from brightfield images. The remaining (non-masked) regions demonstrate the stability of SeqSRM and its ability to maintain the same focus/ROI for at least tens of hours. Additionally, Fig. S3 shows the distribution of the magnitudes of residual drift corrected for by the post-processing drift algorithm described in Section 2.7, demonstrating that the microscope maintains the sample in the same XY field of view throughout the 21 hour experiment to within approximately 5-10 nm.

4. Discussion

SeqSRM is a robust platform for fluorescence super-resolution microscopy which greatly simplifies experimental workflows and reduces user interaction with the microscope for high-throughput studies. The sequential dSTORM $\beta$- and $\alpha$-tubulin overlay results shown in Fig. 4 and Fig. S1 demonstrate that SeqSRM can image multiple targets on at least tens of user-selected cells with nanometer-scale inter-target overlay accuracy. Comparing to other automated localization microscopy techniques, such as [12,13], SeqSRM allows for highly-customized user-selection of multiple ROIs without the need for fiducial markers. Furthermore, the use of brightfield registration [15] to automatically realign the sample after moving between ROIs enables selection of a comparatively inexpensive sample stage. For samples with inherently less contrast, such as those with index matched buffers, contrast could potentially be improved by averaging over multiple brightfield exposures of the sample, enabling SeqSRM to image such samples.

The SeqSRM platform can also be readily modified to allow for an increased field of view (permitting higher throughput in terms of imaged sample area) by simply shifting or swapping out the lenses after the multi-mode fiber ("MM" in Fig. 1) and, if needed, using a suitably higher-powered imaging laser. The current SeqSRM design readily permits future upgrades for 3D studies, as the mirror "M*" in Fig. 1, which is placed at a pupil plane of the objective, can be swapped with a deformable mirror or a reflective phase SLM to allow for PSF engineering. The current design is also capable of increased throughput via rapid dSTORM imaging at even higher laser power (we note, however, that recent work has suggested such rapid imaging might degrade image quality [19]). This flexibility will allow for future customization and expansion for new experimental demands.

Finally, our results demonstrate that local, nanometer-scale changes in cell morphology occur during the course of imaging and can severely degrade the overlay fidelity of multi-target sequential super-resolution data as well as the relative intra-label association fidelity of single-target super-resolution data collected over long durations. We demonstrate that we can detect these local sample changes from brightfield images collected periodically throughout the experiment, allowing us to identify subregions of the resulting super-resolution data that can be trusted (Fig. 3). For most applications of SeqRSM, we do not know a priori the relative association (or lack thereof) between the imaged cellular targets. As such, the ability to quantify local changes in cell morphology from brightfield images alone is a powerful tool that allows us to define trust regions of our data, allowing for more reliable quantitative analyses of the resulting data. We anticipate that this technique will be adopted into other super-resolution imaging workflows and improve the trustworthiness (in terms of the fidelity of the resulting super-resolved overlays with respect to the underlying biological structures) of biological conclusions made from such studies. Furthermore, we anticipate that the quantification of such local sample shifts can be adapted into more sophisticated processing techniques. For example, knowledge of the local changes to cell morphology might be used to define local transforms that can be applied to multi-target super-resolution overlays, effectively recovering data that is otherwise masked out in our application.

5. Conclusion

We have developed a dedicated sequential super-resolution microscope, SeqSRM, which is capable of semi-automatically imaging multiple targets on at least tens of cells with minimal user interaction, with the number of cells limited only by practical considerations such as degradation of the sample/imaging buffer. We have designed SeqSRM around AF647 to allow for optimal imaging of each target in a sequential manner, avoiding the need for multiple optical emission paths and dyes. Using sequential dSTORM imaging [11] of $\beta$- and $\alpha$-tubulin, we have demonstrated that SeqSRM can reliably image multiple structures on tens of cells with nanometer accuracy. SeqSRM greatly simplifies the high-throughput super-resolution studies needed by biological researchers by reducing the need for sophisticated experimental design (e.g., selecting spectrally distinct fluorophores suitable for dSTORM imaging of multiple targets) and the burden of excessive microscope interaction. Another major conclusion of our work is that local, nanometer-scale changes to cell morphology can occur both during sequential super-resolution imaging and during long-duration single-target super-resolution imaging, and that these morphological changes severely impact the fidelity of the resulting super-resolution data. We have shown that we can detect these morphological changes from brightfield images collected throughout the experiment, allowing us to restrict subsequent analyses of resulting super-resolution data to "trusted" regions of the sample whose relative changes fall below a tolerance threshold.