# Molecular resolution imaging by repetitive optical selective exposure

An Author Correction to this article was published on 22 October 2019

## Abstract

We introduce an interferometric single-molecule localization method for super-resolution fluorescence microscopy. Fluorescence molecules are located by the intensities of multiple excitation patterns of an interference fringe, providing around a twofold improvement in the localization precision compared with the conventional imaging with the same photon budget. We demonstrate this technique by resolving nanostructures down to 5 nm in size over a large 25 × 25 μm2 field of view.

## Main

Various image-based central position estimation (termed centroid fitting) methods such as two-dimensional (2D) Gaussian fitting have been commonly used in single-molecule localization microscopy (SMLM) to precisely determine the location of each fluorophore1,2,3. The localization precision is mainly dependent on the number of detected photons4; therefore, tremendous efforts have been invested to increase the photon budget by developing improved fluorophores5, antibleaching agents6,7 and cryogenic imaging techniques8,9,10. Recently, a method called minimal photon fluxes (MINFLUX) was demonstrated to achieve higher localization precision with reduced photon budget11. However, the triangular scanning of doughnut spot and the sequential localization scheme for each molecule limit its field of view, making it impractical for super-resolution imaging of large areas. Interferometry, on the other hand, is another strategy for single-molecule localization that can be parallelized but has only been demonstrated to enhance the axial localization precision12,13,14. Improving the single-molecule lateral localization precision to molecular scale (<2 nm) for high-throughput nanostructure imaging remains a challenge.

In this work, we report repetitive optical selective exposure (ROSE) as a camera-based single-molecule localization technique in which emitters are localized by estimating their relative positions with respect to shifting interference fringes. The basic concept is shown in Fig. 1a. In the excitation optical path, a laser beam was modulated and split to generate a phase-controllable interference fringe in the sample plane. Three fringes with phases shifted by 120 degrees were used for localizing emitters in one direction; thus, a total of six excitation patterns were sufficient for lateral (x and y directions) localization. Along the imaging path, the fluorescence signal was reflected and switched by a resonant scanning mirror, through which single-molecule signals were sequentially focused onto six subregions of the two electron-multiplying charge-coupled devices (each was divided into three subregions) with a field of view of about 25 × 25 μm2 (Supplementary Figs. 1 and 2), corresponding to six excitation patterns. The excitation patterns were synchronized with the scanning mirror so that the fluorescence signal under each excitation pattern was projected to the corresponding subregion (Supplementary Figs. 2 and 3). During one camera exposure, several hundreds to thousands of rounds of scanning were performed, and each of the acquired images contained six subimages corresponding to the three phases and two directions of excitation patterns. The resonant scanner works at a frequency of 4 kHz, taking only 125 μs for one scanning cycle among the six subregions (Supplementary Fig. 3). In this way, the localization error caused by millisecond-scale on–off state transition and bleaching could be minimized (Supplementary Fig. 4).

The localization procedure integrates normal Gaussian fitting with ratiometric calculation of photon numbers in each of the six subimages. Gaussian fitting provides rough position estimates, while the ratiometric calculation provides relative location with respect to interference fringes. A higher localization precision can be achieved by the combination of these two strategies (details in Methods, Supplementary Figs. 5 and 6 and Supplementary Note 1). The calibration of the system (details in Methods and Supplementary Fig. 7) showed a 220-nm interference fringe period length, indicating a localization precision as below:

$${\mathrm{\sigma }}\sim \frac{{220}}{{\sqrt 2 \times {\mathrm{\pi}} \times \sqrt {{N}} }} \approx 50/\sqrt {{N}}$$
(1)

where N denotes the photon number of the single molecule. The influences of phase step error and modulation depth were also evaluated (Supplementary Figs. 8 and 9), showing that 2.3% and 8.7% of total localization error were caused by each of them, respectively.

To evaluate the performance of ROSE, conventional 2D Gaussian curve fitting was used as a centroid-fitting method for comparison. We first validated ROSE with simulated data that exclude the dynamic behaviors of single molecules. The result shows that ROSE can provide a twofold improvement in the localization precision compared with Gaussian fitting, with detected photon number ranging from 250 to 20,000 (Fig. 1b), which is consistent with the theoretical conclusions (Supplementary Note 2). The validation of the scanning speed was also tested by the simulation with various on and off times, showing that 125 μs scan cycle time is sufficient to eliminate the effect due to the fast on–off switching of single molecules (Supplementary Fig. 4). Details of the simulation can be found in the Methods. We further employed ROSE to image 40 nm fluorescent nanospheres. The centered relative position between two fluorescent nanospheres illustrated the superior performance of ROSE, which also shows a twofold improvement in localization precision (Fig. 1c,d). Both the simulated data and the fluorescent nanosphere imaging results suggest that ROSE can extract more precise position information than the Gaussian fitting method with the same number of photons.

To verify the performance of ROSE for single-molecule localization, DNA points accumulation for imaging in nanoscale topography (DNA-PAINT) imaging was conducted on a sample with mixed 20 nm and 10 nm lattice grids of DNA origami structures (Supplementary Fig. 10a,b)7,15. Imager strands labeled with Atto-655 dyes were imaged under 639 nm laser excitation. The images were reconstructed with centroid fitting and ROSE, as illustrated in Fig. 2a–d (larger images in Supplementary Fig. 11). Both methods could resolve the 20 nm structures, but ROSE provided a much narrower peak in the intensity profile with mean photon number of 5,677 (Fig. 2c–g). Multiple Gaussian fitting was performed on the profile, demonstrating an improved resolution (full width at half maxima, FWHM) of ~6 nm by ROSE. In contrast, the centroid-fitting method delivered a lateral resolution (FWHM) ~12 nm. Furthermore, as a demonstration of improved resolution, ROSE could clearly resolve a 4 × 6 lattice origami pattern with a 10 nm distance, which could not be resolved by centroid fitting (Fig. 2c,d). The Fourier ring correlation (FRC) analysis performed in Fig. 2a,b indicates that ROSE improved final resolution by 1.76-fold compared with the centroid-fitting method (Fig. 2i).

We further employed DNA origami with densely packed 5 nm lattice patterns to test whether ROSE has the ability to push the resolving power of SMLM to the molecular scale (<5 nm) with limited photon budgets. As shown in Fig. 3a–c (larger images in Supplementary Fig. 12), the 5 nm structure could be adequately resolved using ROSE at a resolution of 2.6–3.5 nm (FWHM) with a median photon budget of 12,600. The comparison between reconstructed structures and the ground truth pattern shows that reconstructed results match with the designed pattern (Supplementary Fig. 10c–i).

Compared with DNA-PAINT, fluorophores with faster on–off times are more widely applied in biological SMLM imaging techniques, such as (fluorescence) photoactivated localization microscopy ((f)PALM) and stochastic optical reconstruction microscopy (STORM)1,2,3. It is challenging to apply similar techniques in (f)PALM/STORM because the six illumination patterns need to share the same on state of each emitter. For an average on-time of 10 ms, the interference fringe shift and camera exposure should be deployed in 2 ms. ROSE circumvents this problem by ultrafast shifting of six interference fringes required for the lateral position estimate with 125 μs per cycle (Supplementary Figs. 3 and 4), which is far below the on time and the intensity fluctuations of fluorophores. To demonstrate ROSE imaging in-cell with (f)PALM/STORM-compatible probes, microtubule filaments of COS-7 cells were immunostained with primary antibodies and then with Alexa-647-labeled secondary antibodies. As illustrated in Fig. 3d–g, ROSE clearly resolved these cellular nanostructures, and even the hollow structure of single microtubule filaments was clearly visible with only 2,247 average collected photons, which could be barely resolved by the centroid-fitting method. We also verified ROSE imaging using cells stained with Alexa-647-labeled phalloidin. As illustrated in Fig. 3h–j, ROSE clearly resolves cellular nanostructures of the actin filament network. The FRC analysis performed in Fig. 3i,j indicates that ROSE improved final resolution by 2.19-fold compared with the centroid-fitting method (Fig. 3k). Two close filaments with a distance of 24 nm could be resolved. In contrast, centroid fitting could not clearly resolve this structure, as illustrated in Fig. 3l.

ROSE achieves the same localization precision with only one-fourth of the photon counts as compared with the conventional centroid-fitting method, making ROSE suitable for SMLM imaging with luorophores that are less bright. To demonstrate the advantage of ROSE with dimmer fluorophores, we further performed (f)PALM/STORM imaging of Alexa-647-immunolabeled clathrin-coated pits (CCPs), cellular structures used for receptor-mediated endocytosis, in COS-7 cells. A neutral density filter was placed in the imaging path to reduce the photon counts to simulate a low-photon-count condition. Compared with conventional wide-field fluorescence images (Supplementary Fig. 13a), ROSE and STORM revealed predominantly spherical structures, representing CCPs and vesicles (Supplementary Fig. 13b–d). ROSE clearly revealed the hollow structure even for small CCPs of less than 100 nm diameter with ~400 photons, which is similar to the photon budget of the fluorescence proteins used in SMLM imaging (Supplementary Fig. 13d–g). In contrast to ROSE, the centroid-fitting method only resolved the hollow structure of much bigger CCPs (Supplementary Fig. 13c).

In summary, we demonstrate that ROSE enables a twofold improvement in lateral resolution relative to the widely used Gaussian fitting method. We also demonstrate that ROSE provides molecular resolution by resolving structures separated by 5 nm. Our method is compatible with imaging a large field of view, which is only limited by the detection area of the electron-multiplying charge-coupled device and the illumination area. This large field of view was not demonstrated by MINFLUX due to its scanning scheme for every single fluorophore. Currently, ROSE uses a fixed scanning range to realize parallelization and hence high throughput. However, the fixed scanning range also limits the resolution of ROSE as compared with MINFLUX, which uses adjustable scanning range. We envisage further improvement of ROSE in resolution by using a smaller scanning range, higher laser power and numerical aperture objectives, controlled on-switching of fluorophores, more images, and so on. ROSE could be extended to three-dimensional nanometer-scale imaging by introducing additional excitation patterns along the axial direction. We envisage that this method could extend the application of SMLM in biomacromolecule dynamic analysis and structural studies at molecular scale.

## Methods

### Optical setup

The detailed optical setup can be found in Supplementary Fig. 2. The entire setup was constructed based on an Olympus IX71 inverted microscope. A 639 nm 400 mW single longitudinal mode laser was modulated by an acousto-optic tunable filter (AOTF) and coupled to a polarization-maintaining fiber. After being emitted from the fiber, the laser was collimated into a parallel beam with a fiber collimator. A half wave plate and a polarizer were used to adjust the polarization direction and ensure the polarization ratio of the beam. An electro-optic modulator (EOM) (EOM1 in Supplementary Fig. 2) was used to switch the beam between two optical paths with the polarizing beam splitter. After the polarizing beam splitter, the polarization of the beam was adjusted by a half wave plate, and then was modulated by another EOM (EOM2 or EOM3) before being expanded and split into two beams. The two beams were focused on the back focal plane of the objective to create an interference fringe on the sample; to increase the contrast of the fringe, the polarization of the laser beams was adjusted by a half wave plate. In our work, the modulation speed needs to be much faster than that in structured illumination microscopy16, so we adopt EOMs, which provide more than 1 MHz bandwidth of modulation. EOM2 and EOM3 are electron-optic phase modulators which are applied to modulate the relative phase between the parallel (p) and senkrecht (s) parts of the beams passed through them. After the modulators, the p and s parts were split by a polarizing beam splitter and then the polarization direction was adjusted by the half wave plates. With this method, these modulators could control the phase of the corresponding interference fringes. Please note that this approach seems less stable than the grating-based method as used in structured illumination microscopy16, and will need additional improvement to correct the drift of interference fringes.

For the imaging path, the fluorescence image was focused on a 4 kHz resonant scanning mirror resulting in 8 kHz scanning rate. The fluorescence signal was then reflected by the scanner and collimated by six pairs of conjugate lenses to project six subimages onto the CCD (charge-coupled device) chips (CCD1 and CCD2; each CCD contained three subimages). Here, the fluorescence image was focused at the surface of the scanning mirror, and the mirror was placed at the focal point of the six pairs of conjugate lenses; in this way, the image will not be blurred by the scanner when the fluorescent beam swipes within the optical aperture of one of six relay lenses (Supplementary Fig. 14). It should be noted that the laser needs to be turned off when the fluorescent beam locates at the junction of two lenses to avoid cross-talk and signal loss. The scanning rate of 8 kHz was similar to MINFLUX11, which has been shown to be sufficient to eliminate the influence of fast on–off events and other effects that could cause intensity fluctuations. The dipole rotation could cause intensity fluctuations in the nanoseconds scale17, which could be eliminated by 125 μs scanning time.

The resonant scanner, AOTF and EOMs were synchronized by a National Instruments data acquisition (DAQ) device with a program written in LabVIEW. During image acquisition, the resonant scanner continuously vibrated the mirror at its resonant frequency, and at the same time a synchronization signal indicating the scanning direction was generated by the scanner. The frequency of the synchronization signal from the resonant scanner was firstly multiplied by a complex programmable logic device (CPLD), and then used as an external reference clock for the DAQ device to generate the output signals. AOTF and EOMs were controlled by the output signals from the DAQ device based on the external reference clock. In this way, the AOTF and the EOMs could be synchronized to the resonant scanner during image acquisition, as showed in Supplementary Fig. 3. The details of image path switching and the resonant scan timing are shown in Supplementary Note 3.

The drift correction system consisted of a light-emitting diode (LED) light source, a condenser, a piezo stage, a dichroic mirror and a CCD. The light from the LED was focused by the condenser before passing through the sample. Then the LED light was reflected by the dichroic mirror, and focused on the CCD (CCD3) to form a bright-field image of the sample. A program written in Python was used to acquire the images, calculate the drift of the sample and control the piezo to correct the drift in real time.

### Single-molecule localization procedure of ROSE

The single-molecule localization procedure is different from the centroid estimation method, and the phase was estimated similarly to phase step interferometry. First, 2D Gaussian fitting was performed on the integrated image of the six subimages, by which the position, size of the spot and total photon number were estimated. The position estimated with 2D Gaussian fitting was used as rough localization information. Then, another 2D Gaussian fitting was performed on each subimage, with fixed position and size of the spot, to estimate the photon numbers in each subimage. For each x or y direction, three corresponding photon numbers were used to calculate the phase of the fringe. We further aligned the rough position and the phase information with linear regression, by which the molecules could be localized more precisely. Please refer to Supplementary Note 1 for more details of phase estimation and the phase–position alignment, and Supplementary Note 4 for more details of reconstruction.

### Drift correction

Sample drift is one of the major technical problems for SMLM, especially when pushing the resolution below 10 nm, as discussed in previous work7. Because of the interferometry involved in our setup, the system will also be affected by drift of the interference fringe, which can be caused by air flow or temperature changes. In this work, we tried to push the resolution to sub 5 nm, and so drift correction needed to be considered seriously.

To correct system drift over hours, real-time drift correction was adopted as described above in the optical setup section. A cover glass was fused with 5 μm polystyrene microspheres. During data acquisition, bright-field images of the microspheres were also acquired simultaneously, and the three-dimensional drift of the microspheres was estimated and corrected in real time.

The real-time drift correction provided high stability of the sample during the image acquisition, but the drift of the interference fringe could not be corrected by this system (Supplementary Fig. 15). To achieve a higher precision, a cross-correlation-based method was used to correct the residual drift. The localization information of single molecules was divided into several groups according to the time they were detected. Then, cross-correlation of each group to the first one was performed to estimate and correct the drift further.

For the DNA-PAINT imaging of the 10 and 5 nm structures, a geometry-based drift correction was applied additionally, as described in previous work7. With this kind of correction, the effect of drift could be minimized.

After the drift correction procedure, as shown in Fig. 3b,c, the minimum measured standard deviation (std.) of the profile was about 1.1 nm, and the theoretical localization precision at photon number 12,600 was 0.9 nm, so the maximum residual drift was below 0.63 nm in standard deviation, showing a high precision of the drift correction method.

### System calibration

To measure the period length of the interference fringe, fluorescence nanospheres were imaged with position scanning by a piezo stage. Then, the location and corresponding phase were fitted with linear regression to estimate the period length of the fringes in the x and y directions, as shown in Supplementary Fig. 7. The results show a period length of approximately 222.6 nm in the x direction and 215.7 nm in the y direction, which is close to the theoretical value of 214.4 nm.

### Simulation

The simulation program was written in MATLAB with pixel size of 150 nm, point spread function (PSF) standard derivation of 150 nm, fringe periodic length of 220 nm, electron multiplying (EM) gain of 100 and electron to analog-to-digital units (ADU) conversion factor of 10.5, which are similar to experimental conditions. For every simulated image, shot noise and read-out noise were added together with additional noise introduced by the EM gain. Next, position estimation was performed on both conventional images and ROSE images. The estimation results were compared with the ground truth and then converted to the localization precision. For each photon number condition in Fig. 1b, 1,000 images were generated for either ROSE or centroid fitting, which is sufficient for the localization precision analysis. For the simulation of on and off times, a series of on and off times was randomly generated according to the exponential distribution with the mean on and off times. Then, they were compared with the scanning time of each subimage to determine the intensity of each subimage.

Demo software of the simulation can be found in the Supplementary Software.

### Sample preparation

#### Materials and buffers

Unmodified DNA oligonucleotides and biotin-labeled DNA oligonucleotides were purchased from GENEWIZ. Fluorophore-modified DNA oligonucleotides were purchased from Ningbo Kangbei Biochemical. M13mp18 scaffold was purchased from New England BioLabs (catalog no. N4040S). BSA-biotin was obtained from GATTAquant (catalog no. 0082). NeutrAvidin protein was purchased from ThermoFisher (catalog no. 31000). Ultrafiltration concentrators were ordered from Millipore (catalog no. UFC510096). DMEM (catalog no. c11995500BT) and fetal bovine serum (catalog no. 16000-044) were purchased from Gibco. Penicillin-streptomycin solution was purchased from HyClone (catalog no. 30010). Anti-Clathrin heavy chain primary antibody was purchased from Abcam (catalog no. ab21679). Anti-α-Tubulin primary antibody was purchased from Sigma-Aldrich (catalog no. T9026). Goat anti-Rabbit IgG (H + L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor 647 (catalog no. A21245) and Donkey anti-Mouse IgG (H + L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor 647 (catalog no. A31571) were purchased from ThermoFisher. Paraformaldehyde (catalog no. 15710) and glutaraldehyde (catalog no. 16200) were ordered from Electron Microscopy Sciences. Normal goat serum (NGS) was purchased from Beyotime (catalog no. C0265). Glass-bottomed dishes with 10 mm (catalog no. D35-10-1-N) and 20 mm microwells (catalog no. D35-20-1-N) were purchased from Cellvis. Glucose (catalog no. 101141632), glucose oxidase (catalog no. G2133), catalase (catalog no. C1345) and 2-mercaptoethanol (catalog no. M3148) were ordered from Sigma-Aldrich. Polystyrene microspheres were purchased from Wuhan Huake Weike Technology (catalog no. PS-M-10075). FluoSpheres Carboxylate-Modified Microspheres, 0.04 µm, dark red fluorescent (660/680), were purchased from ThermoFisher (catalog no. F8789). Alexa-647-labeled phalloidin was purchased from ThermoFisher (catalog no. A22287).

The following buffers were used for the sample preparation and imaging:

DNA origami folding buffer: 1× TAE buffer, 5 mM NaCl and 15 mM MgCl2

DNA origami purification buffer: 1× TAE buffer and 10 mM MgCl2

DNA origami imaging buffer: 1× PBS buffer and 10 mM MgCl2

STORM imaging buffer: 1× PBS buffer, 10% glucose, oxygen removed, GLOX (glucose oxidase (0.6 mg ml−1), catalase (0.06 mg ml−1), dissolved in Tris-HCl buffer) and 143 mM 2-mercaptoethanol.

#### DNA origami synthesis

The DNA origami structures were assembled in a one-pot reaction with 50 μl total volume, containing 10 nM m13mp18, 100 nM unmodified staple strands, 100 nM biotin-modified strands and 400 nM strands with DNA-PAINT extensions in DNA origami folding buffer. The solution was annealed with a PCR device, heated to 90 °C for 15 min, cooled to 70 °C at a speed of 1 °C min−1, cooled to 30 °C at a speed of 3 °C min−1 and finally stopped at 4 °C.

The self-assembled DNA origami structures were characterized by agarose gel electrophoresis (2% agarose, 1× TAE, 10 mM MgCl2, and 1× DuRed prestain) at 9 V cm−1 for 1 h.

For purification, a 100 kDa ultrafiltration concentrator was used. First, 500 μl of DNA origami purification buffer was added, and the concentrator was centrifuged at 5,000g at 4 °C for 7 min. Then, 50 μl of anneal solution and 450 μl of DNA purification buffer were added, and the concentrator was centrifuged at 2,000g at 4 °C for 17 min. The last step was repeated until all annealing solution was added into the concentrator. Then, 450 μl of DNA purification buffer was added, and the concentrator was centrifuged at 2,000g at 4 °C for 17 min, and the step was repeated twice. The filter was inverted into a new tube and centrifuged at 1,000g at 4 °C for 2 min, and the solution was collected and stored at −20 °C.

#### DNA-PAINT imaging

For the sample preparation, a glass-bottomed dish with 10 mm microwell was used. First, 150 μl of biotin-labeled BSA (1 mg ml−1, dissolved in PBS) was pipetted into the microwell, incubated for 5 min and then washed three times with 150 μl of PBS. Then, 150 μl of NeutrAvidin (1 mg ml−1, dissolved in PBS) was pipetted into the microwell, incubated for 5 min and washed three times with 150 μl of PBS. The DNA origami structures were mixed with 5 μl of the 20 nm structure and 2.5 μl of the 10 nm or 5 nm structures, and then diluted with 50 μl of DNA origami imaging buffer. Next, 50 μl of the mixture was pipetted into the microwell, incubated for 5 min and then washed with 150 μl of DNA origami imaging buffer three times. After adding the imager strands, the sample was ready for imaging. The final concentration of the imager strands was approximately 3 nM.

For the 10 nm structure, exposure of 200 ms, EM gain of 50 and illumination power intensity of approximately 2–4 kW cm−2 were applied, and 40,000 frames were acquired for the reconstruction.

For the 5 nm structure, exposure of 300 ms, illumination power intensity of approximately 2–4 kW cm−2 and no EM gain were applied. The CCD readout bandwidth was set to 3 MHz, and 80,000 frames were acquired for the reconstruction.

#### Cell culture and immunostaining

COS-7 cells (3111C0001CCC000033, National Infrastructure of Cell Line Resources, China) were maintained in DMEM supplemented with 10% fetal bovine serum and 100 U ml−1 penicillin and streptomycin at 37 °C and 5% CO2. Cells were plated in 35 mm glass-bottomed dishes with 20 mm microwells at a density of 50,000 cells per well. After 48 h, cells were rinsed with PBS, then fixed with 3% paraformaldehyde and 0.1% glutaraldehyde in PBS for 10 min at room temperature. Reduction was done using 0.1% sodium borohydride for 7 min. After washing with PBS, the cells were permeabilized with 0.2% Triton X-100 for 15 min. Blocking was done for 90 min with 10% NGS and 0.05% Triton X-100 in PBS at room temperature, following by staining with the primary antibodies against tubulin or clathrin (50 μg ml−1 mouse anti-α tubulin or 1 μg ml−1 rabbit anti-clathrin heavy chain) for 60 min in antibody dilution buffer (5% NGS, 0.05% Triton X-100 in PBS). The cells were then washed with washing buffer (1% NGS, 0.05% Triton X-100 in PBS) five times and stained with the secondary antibodies labeled with Alexa Fluor 647 for 60 min. The cells were then washed with washing buffer five times and postfixed with 3% paraformaldehyde and 0.1% glutaraldehyde in PBS for 10 min. Next, the cells were washed three times with PBS, 5 min per wash, and then washed twice with ddH2O, 3 min per wash. Finally, the cells were stored with ddH2O at 4 °C.

#### Actin staining

COS-7 cells were cultured on 35 mm glass-bottomed dishes with 20 mm microwells at a density of 50,000 cells per well. After 48 h, cells were briefly washed with prewarmed (37 °C) PBS. Then, cells were fixed and permeabilized with 0.3% glutaraldehyde and 0.25% Triton X-100 in cytoskeleton buffer (10 mM MES pH 6.1, 150 mM NaCl, 5 mM EGTA, 5 mM glucose and 5 mM MgCl2) for 2 min. After that, cells were fixed with 2% glutaraldehyde in cytoskeleton buffer for 10 min. Reduction was done using 0.1% sodium borohydride for 7 min. Next, cells were washed with PBS three times, each time allowing for 10 min incubation. Cells were stained with Alexa-647-phalloidin (1.1 μM in PBS) at 4 °C overnight. Then, cells were briefly washed with PBS once and mounted for imaging.

#### STORM imaging

For the samples in Fig. 3d–j, exposure time of 50 ms, EM gain of 100 and illumination power intensity of approximately 4 kW cm−2 were used. In total, 20,000 frames were acquired for the reconstruction. During the imaging, a 405 nm laser was used to control molecule density.

For the sample in Supplementary Fig. 13, it was not feasible to perform imaging with photon-sensitive fluorescent proteins because of the 639 nm laser used in our setup, so we used an optical attenuator to simulate the low photon count condition in SMLM. A neutral density filter with an optical density of 1 was inserted into the imaging path to reduce the photon number. Exposure time of 50 ms, EM gain of 100 and illumination power intensity of approximately 4 kW cm−2 were used. In total, 20,000 frames were acquired for the reconstruction. A 405 nm laser was used to control molecular density.

### Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

## Data availability

The data that support the findings of this study are available from the corresponding author upon request.

## Code availability

MATLAB programs for simulation data generation and data analysis are freely available for academic use and are provided online with this paper as Supplementary Software. The LabVIEW program for device controlling and the Python program for drift correction during imaging, which are hardware-dependent, are available from the corresponding author upon request.

## Change history

• ### 22 October 2019

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

## References

1. 1.

Betzig, E. et al. Science 313, 1642–1645 (2006).

2. 2.

Hess, S. T., Girirajan, T. P. K. & Mason, M. D. Biophys. J. 91, 4258–4272 (2006).

3. 3.

Rust, M. J., Bates, M. & Zhuang, X. Nat. Methods 3, 793 (2006).

4. 4.

Thompson, R. E., Larson, D. R. & Webb, W. W. Biophys. J. 82, 2775–2783 (2002).

5. 5.

Zheng, Q. et al. Chem. Soc. Rev. 43, 1044–1056 (2014).

6. 6.

Vogelsang, J. et al. Angew. Chem. Int. Ed. 47, 5465–5469 (2008).

7. 7.

Dai, M., Jungmann, R. & Yin, P. Nat. Nanotechnol. 11, 798 (2016).

8. 8.

Weisenburger, S. et al. ChemPhysChem 15, 763–770 (2014).

9. 9.

Li, W., Stein, S. C., Gregor, I. & Enderlein, J. Opt. Express 23, 3770–3783 (2015).

10. 10.

Liu, B. et al. Sci. Rep. 5, 13017 (2015).

11. 11.

Balzarotti, F. et al.Science 355, 606–612 (2017).

12. 12.

Shtengel, G. et al. Proc. Natl Acad. Sci. USA 106, 3125–3130 (2009).

13. 13.

Aquino, D. et al. Nat. Methods 8, 353 (2011).

14. 14.

Huang, F. et al.Cell 166, 1028–1040 (2016).

15. 15.

Jungmann, R. et al. Nat. Methods 11, 313–318 (2014).

16. 16.

Gustafsson, M. G. L. J. Microsc. 198, 82–87 (2000).

17. 17.

Lew, M. D., Backlund, M. P. & Moerner, W. E. Nano Lett. 13, 3967–3972 (2013).

## Acknowledgements

We thank L. Pan for helping with the preparation of COS-7 cells stained with Alexa-647-labeled phalloidin, and Y. Zhang and B. Liu for discussions. This work was supported by the National Key Research and Development Program of China (grant no. 2017YFA0505300 to W.J., grant no. 2016YFA0500200 to T.X., grant no. 2016YFA0502400 to W.J.), the National Foundation of Natural Science of China (grant no. 31730054, 31661143041, 31127901 to T.X., grant no. 31700743 to L.G.), Beijing Municipal Science & Technology Commission project no. Z181100004218002 (to T.X.), the Instrument Development Project of CAS (grant no. GJJSTD20180001 to T.X., grant no. YJKYYQ20180069 to W.J.) and the Youth Innovation Promotion Association of CAS (grant no. 2013066 to W.J. and grant no. 2017135 to L.G.).

## Author information

Authors

### Contributions

W.J., T.X. and L.G. designed the experiments. W.J. and L.G. developed the ROSE system. L.G. performed data analysis. Y.L. and S.Z. conducted sample preparation and imaging. T.X., W.J., L.G., D.L., Y.X. and W.L. interpreted the results. W.J., T.X. and L.G. wrote the manuscript, which was modified by all of the other authors.

### Corresponding authors

Correspondence to Tao Xu or Wei Ji.

## Ethics declarations

### Competing interests

The authors declare no competing interests.

Peer review information Rita Strack was the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

## Integrated supplementary information

### Supplementary Figure 1 One pair of raw images from 2 EMCCDs.

Each image of 512*512 pixels contains 3 subimages indicated by subimage 1-6. The size of each subimage is about 170*170 pixels, with a pixel size of 150 nm, yielding a FOV of about 25*25 μm2. The yellow arrows indicate the same single molecule in six subimages. 5 experiments were repeated independently with similar results.

### Supplementary Figure 2 Optical setup of ROSE.

Laser1: MSL-FN-639, CNI; Laser2: MLL-III-405, CNI; AOTF: AOTFnC-400.650, AA; Fiber Coupler: HPUC-23AF-400/700-P-4.5AC-15, OZ; Fiber Collimator: HPUCO-23-400/700-P-6AC, OZ; PM fiber: QPMJ-3AF3S-400-3/125-3AS-3-1, OZ; Half waveplate: WPH05M-633, Thorlabs; Polarizer: GT10-A, Thorlabs; EOM1: LM0202, QiOptiq; EOM2 and 3: Model 350-50, ConOptics; Beam Expander: GCO-2501, CDH; PBS: PBS251, Thorlabs; Mirror: PFSQ10-03-P01, Thorlabs; DM1: Di02-R514-25x36, Semrock; DM2: Di03-R405/488/561/635-t1-25x36, Semrock; DM3: FF520-Di02-25x36, Semrock; CCD1 and 2: iXon 897 EMCCD, Andor; CCD3: Guppy F-033B, Allied Vision Technologies; Resonant Mirror: CRS 4 KHZ, CTI; DAQ: USB-6343, NI; CPLD: epm570t100c5n, Altera.

### Supplementary Figure 3 Time sequence diagram of ROSE.

This figure shows the signal for ATOF and EOMs to switch the illumination patterns, which was generated by DAQ and synchronized with the resonant scanner. This time sequence diagram was illustrated with 100 ms exposure. One exposure includes approximately 800 switch cycles, each with switch of 6 illumination patterns. The scanning cycle was 125 μs, with each illumination pattern of approximately 20 μs, with the AOTF on-time of 11.25 μs, resulting in a duty cycle of about 54% for the illumination.

### Supplementary Figure 4 The relation between on-off switching times and the localization precision under various scan cycle times.

The simulated off-time subjects to exponential distribution with average values of 1 ms, 5 ms, 15 ms, 50 ms, 150 ms, 500 ms, 1,500 ms and 5,000 ms, and the simulated average on-time subjects to exponential distribution with average values from 1 ms to 100 ms, which covers most experimental conditions. The scan cycle time of 0.125 ms, 2.5 ms and 25 ms were tested, which represents the typical scan cycle time for resonant scanner, standard galvo scanner and sequential exposure to get 6 subimages without scanning configuration, respectively. For each condition, 1000 frames were generated for the statistical analysis, each point represents the mean value with error bars representing a 95% confidential interval.

### Supplementary Figure 5 Localization procedure of ROSE along one of the lateral directions.

Three subimages were acquired corresponding to the three illumination patterns. Then Gaussian fitting was performed to estimate the rough position and three photon numbers. The three photon numbers were then used to estimate the phase. An alignment of the phase and the rough position will provide precise position.

### Supplementary Figure 6 Demonstration of periodical linear fitting in 1D.

Each localization contains both localization information and phase information, shown as the scatter plot of blue dots. After the periodical linear fitting, the relation between localization and the phase could be identified as shown by the red points.

### Supplementary Figure 7 Calibration of the fringes in x and y directions.

The position scanning was performed by a piezo stage, with steps of 20 nm. The linear regression results showed that the periodic length of fringe in the x direction (a) and y direction (b) are 222.6 nm and 215.7 nm, respectively. The excitation laser wavelength of 639 nm and the N.A. of 1.49, resulting in a theoretical minimal achievable periodic length of 214.4 nm. For the data in a to b, n = 20 images were acquired at each position for the analysis.

### Supplementary Figure 8 The modulation depth and its influence on the localization precision.

(a) Simulated data shows the relation between modulation depth and localization precision with 2000 photon number. (b and c) Experimental modulation depth estimated by measurement of 40 nm fluorescent nanospheres, which showed a high contrast of the fringe with modulation depth of 0.91 in x direction (b) and 0.92 in y direction (c). (d and e) Experimental modulation depth estimated by measurement of 40 nm fluorescent nanospheres, which showed a nearly uniform high contrast of the fringe within the range of FOV. For each condition in (a), 1,000 frames were generated for the statistical analysis, and each point represents the mean value with the error bar represents for a 95% confidential interval. 550 measurements were taken in x- and y- directions in b and c, respectively. For each fluorescent spheres in d and e, 20 measurements were performed and the number indicates the mean modulation depth. Scale bar: 5 μm in d and e.

### Supplementary Figure 9 The phase step error and its influence on localization precision and bias.

(a) Simulated data showed the relation between phase step error and the localization precision with an assumed pi*2/3 fixed step size and 2000 photon number. (b) Theoretical phase estimation bias with different phase step error. (c) The system phase step calibration results showed phase step error below 2 degree. 1,000 frames were generated for each condition in (a), and each point represents the mean value with the error bar represents for a 95% confidential interval. 500 measurements were taken in each condition in c.

### Supplementary Figure 10 Design and measurement results of DNA origami structures.

(a) AFM image of the synthesized DNA origami. (b) Design of the DNA origami structures for DNA-PAINT imaging. (c and e) Averaged image of the 20 nm DNA origami structures reconstructed by centroid fitting (c) and ROSE (e), respectively, with the center of each cluster (red cross) compared with the designed lattice pattern (green cross). (d and f) Measured 20 nm structure lattice constants of c and e, shown as design schematics. (g) Averaged image of the 5 nm DNA origami structures reconstructed by ROSE, with the center of each cluster (red cross) compared with the designed lattice pattern (green pattern). (h) Multi 2D Gaussian fitting result of g, with the designed lattice pattern (green cross). (i) Measured 5 nm structure lattice constants of g, shown as design schematics. Scale bars: 100 nm in a, 20 nm in c, e, g and h. For the data in a, 5 experiments were repeated independently with similar results. For the data in c to f, the total number of averaged structures was 92, and the center indicates the mean value. 5 experiments were repeated independently with similar results. For the data in g to i, the total number of averaged structures was 169 and the center indicates the mean value. 5 experiments were repeated independently with similar results.

### Supplementary Figure 11 The 20 nm and 10 nm DNA origami structures over a large field of view.

The images were reconstructed with centroid fitting (a) and ROSE (b), respectively, each image with a gallery of zoomed 20 nm (gallery bottom) and 10 nm (gallery top) structures. Scale bar: 200 nm (large image) and 50 nm (gallery). For data in a and b, 5 experiments were repeated independently with similar results.

### Supplementary Figure 12 The 5 nm DNA origami structures over a large field of view with ROSE.

With a gallery of zoomed 20 nm (gallery bottom) and 5 nm (gallery top) structures respectively. Scale bar: 200 nm (up) and 50 nm (down). 5 experiments were repeated independently with similar results.

### Supplementary Figure 13 Comparison of centroid fitting and ROSE imaging of clathrin-coated pits (CCP) in COS-7 cells.

(a) Wide field of the fluorescent image. (b) Reconstructed image with ROSE, showing the hollow structures of the CCPs. (c and d) Zoomed in images of the reconstruction with conventional centroid fitting (c) and ROSE (d), respectively. (e) Distribution of the photon number of single molecules showing a mean number of 451 and median number of 391. (f) Intensity profile of the cross-section indicated in c and d. The diameter of the hollow structure was 90.3 nm, which could not be clearly resolved in (c) by the centroid fitting method. (g) Intensity profile of the cross section indicated in c and d. Scale bars: 1 μm in a and b; 500 nm in d. The width of ROI in c and d is 30 nm. For data in a to d, 5 experiments were repeated independently with similar results.

### Supplementary Figure 14 Comparison of PSF width between ROSE and conventional images.

(a) 6 subimages of ROSE and conventional image of 40 nm fluorescent spheres, with zoomed in on the right indicated by the yellow rectangle. (b-e) Distribution of PSF width in standard derivation of ROSE (b and d) and conventional image (c and e) in x and y directions, respectively. Scale bars: 5 μm (left of each image pair in a) and 1 μm (right of each image pair in a). For data in a, 5 experiments were repeated independently with similar results.

### Supplementary Figure 15 Phase drift and direction stability of fringes.

(a) Phase drift of two fringes within 600 seconds, showing a 0.1 rad drift of fringe X and about 0.4 rad drift of fringe Y, respectively. (b and c) Direction stability of fringe X (b) and fringe Y (c), the direction was calculated as the angle to the x-axis, so the angle of fringe X is near 0 and the angle of fringe Y is near pi/2.

## Supplementary information

### Supplementary Information

Supplementary Figs. 1–15 and Notes 1–4.

## Rights and permissions

Reprints and Permissions

Gu, L., Li, Y., Zhang, S. et al. Molecular resolution imaging by repetitive optical selective exposure. Nat Methods 16, 1114–1118 (2019). https://doi.org/10.1038/s41592-019-0544-2

• Accepted:

• Published:

• Issue Date:

• ### A Picture Worth a Thousand Molecules—Integrative Technologies for Mapping Subcellular Molecular Organization and Plasticity in Developing Circuits

• Jacqueline A. Minehart
•  & Colenso M. Speer

Frontiers in Synaptic Neuroscience (2021)

• ### Photonic Technologies for Liquid Biopsies: Recent Advances and Open Research Challenges

• Francesco Dell'Olio
• , Judith Su
• , Thomas Huser
• , Virginie Sottile
• , Luis Enrique Cortés‐Hernández
•  & Catherine Alix‐Panabières

Laser & Photonics Reviews (2021)

• ### Development of small molecule inhibitor-based fluorescent probes for highly specific super-resolution imaging

• Qiang Wu
• , Yingying Jing
• , Tan Zhao
• , Jing Gao
• , Mingjun Cai
• , Haijiao Xu
• , Yi Liu
• , Feng Liang
• , Junling Chen
•  & Hongda Wang

Nanoscale (2020)

• ### Molecular resolution imaging by post-labeling expansion single-molecule localization microscopy (Ex-SMLM)

• Fabian U. Zwettler
• , Sebastian Reinhard
• , Davide Gambarotto
• , Toby D. M. Bell
• , Virginie Hamel
• , Paul Guichard
•  & Markus Sauer

Nature Communications (2020)

• ### Single-molecule biosensors: Recent advances and applications

• Namik Akkilic
• , Stefan Geschwindner
•  & Fredrik Höök

Biosensors and Bioelectronics (2020)