Enabling Spectrally Resolved Single-Molecule Localization Microscopy at High Emitter Densities

Single-molecule localization microscopy (SMLM) is a powerful super-resolution technique for elucidating structure and dynamics in the life- and material sciences. Simultaneously acquiring spectral information (spectrally resolved SMLM, sSMLM) has been hampered by several challenges: an increased complexity of the optical detection pathway, lower accessible emitter densities, and compromised spatio-spectral resolution. Here we present a single-component, low-cost implementation of sSMLM that addresses these challenges. Using a low-dispersion transmission grating positioned close to the image plane, the +1stdiffraction order is minimally elongated and is analyzed using existing single-molecule localization algorithms. The distance between the 0th and 1st order provides accurate information on the spectral properties of individual emitters. This method enables a 5-fold higher emitter density while discriminating between fluorophores whose peak emissions are less than 15 nm apart. Our approach can find widespread use in single-molecule applications that rely on distinguishing spectrally different fluorophores under low photon conditions.

S uper-resolution microscopy, or nanoscopy, has revolutionized the life and material sciences as it allows surpassing the optical diffraction limit by more than an order of magnitude. 1−4 One frequently used implementation is single-molecule localization microscopy (SMLM), in which the stochastic activation of single fluorescent emitters leads to spatially separated point spread functions (PSFs) that are used to determine the position of each emitter with sub-50 nm accuracy. Localizations obtained via (direct) stochastic optical reconstruction microscopy ((d)STORM), 1,4 point accumulation for imaging in nanoscale topography (PAINT), 5−8 or photoactivated localization microscopy (PALM) 3,9,10 provide access to detailed structural images or can quantify dynamics and mobilities via single-particle tracking (spt). 10,11 In this capacity, SMLM has led to breakthroughs 12,13 in fields such as DNA−protein interactions, 14−17 cell biology, 18−20 and soft matter. 8,21−23 Improving throughput via multiplexing of different fluorophores in SMLM, 5 enabling microenvironmental characterization, 24,25 or studying fluorophore-to-fluorophore distance (via single-molecule Forster resonance energy transfer (smFRET)), 26,27 can be accomplished by combining SMLM with the additional spectral characterization of emitters. Spectral information on single emitters can be acquired and analyzed using various implementations that all rely on placing additional components into the optical detection pathway.
The first implementation of spectrally resolved SMLM uses the ratiometric distinction of spectral emission profiles and is based on placing one (or more) suitable dichroic mirror(s) in the emission pathway. 28−31 Photons emitted from the sample are separated based on their wavelength and directed toward two different detection channels. This entails either two separate detectors or using two areas on the same camera chip after using additional lenses or mirrors to guide the beams. Then the PSFs and their integrated intensities obtained in the two channels are matched, and the intensity ratio of photons is used to discriminate between the emission spectra of different fluorophores. Importantly, this method requires photons to be directed toward each channel, implicating that only a defined spectral range around the cutoff wavelength of the dichroic mirror can be accessed.
The second implementation uses point spread function engineering to obtain spectral information on emitters. Here, a spatial light modulator (SLM) or a phase mask (PM) is employed in the Fourier plane of the emission path. 32,33 The introduced phase offset by these elements depends on the incoming wavelength, which can be exploited to design a pattern so that different PSF shapes are realized when photons of different wavelength arrive at the detector. However, small spectral emission differences in the order of tens of nanometers in the peak emission cannot create sufficiently distinct PSF shapes, hindering discrimination of spectrally close fluorophores. Moreover, the voltage (for SLM) or phase (for PM) has to be specifically tuned for certain emission wavelengths, complicating this method when different fluorophores are used.
In the third implementation, spectral dispersion, a spectrally dispersive optical element is added, after which the spatial and spectral profiles are guided to different regions on a single camera chip or to completely separate detectors. 34−41 Generally, the spatial profile is then analyzed with regular single-molecule localization algorithms, 42−46 while the spectral profile is spread out over tens of pixels and is used to determine the corresponding emission profile. While this implementation allows a large spectral range to be used and allows the discrimination of fluorophores with similar emission spectra, it has various downsides. First, the entire emission pathway needs to be modified to separate the spatial from the spectral channel. 37, 47 Second, as the spectral information is widened over tens of pixels, the signal-to-noise ratio obtainable in this channel is compromised, leading to a loss in spectral accuracy. 48 The wide spreading of emission further directly limits the usable density of emitters in the sample, as overlapping spectral profiles cannot be resolved.
In comparing these implementations, it is obvious that there is a need to combine an easy implementation with a broad spectral range and good specificity. Here, we demonstrate spectrally resolved single-molecule localization microscopy (sSMLM) using an inexpensive (blazed) transmission grating that can be easily implemented in most microscope configurations allowing for widespread adoption while maximizing the achievable signal-to-noise ratio. By using a grating with a large line pitch and placing it close to the image plane to minimize the dispersion of the 1 st order, we created a lowdispersion sSMLM implementation in which the 0 th and 1 st diffraction order of every emitter is imaged and analyzed by existing single-molecule localization algorithms in a single field of view without the need for any additional optical elements. This concept can be straightforwardly combined with SMLM methods employing structured excitation profiles 49−54 and with 3-dimensional PSF engineering approaches, 43,44,55−59 further increasing the potential use of our implementation. Our novel implementation is capable of accurately determining spectral properties of single molecules at much higher emitter densities than other spectral-dispersing sSMLM implementations. With our sSMLM approach, we show the technical feasibility of spectral multiplexing and ability to distinguish between 0 and 15% FRET efficiency in single-molecule FRET experiments.
To maximize the signal-to-noise ratio and the achievable molecular density in spectrally resolved single-molecule localization microscopy (sSMLM), the available photon budget should be distributed over as few camera pixels as possible. 48 To fulfill this criterion, we placed a transmission grating with low dispersion (70 lines/mm) as close as possible to a camera chip representing the image plane (<1 cm, Figure  1a) by placing it inside the camera housing. This arrangement minimizes the separation of the 0 th and 1 st order diffraction patterns and thus results in the highest achievable fluorophore density and signal-to-noise ratio for the 1 st order diffraction pattern (Figure 1b,c). Notably, and distinctly different from earlier implementations, 34−37 our arrangement allows imaging of both spatial and spectral information in the same field of view, thereby maximizing the usable detection area on the sensor. Furthermore, our approach does not require any additional optical components such as mirrors, beam splitters, or secondary detectors. Figure 1. Implementation of low-dispersion spectrally resolved single-molecule localization microscopy. (a) A low-dispersion blazed transmission grating is placed in the emission path of a typical SMLM capable microscope such that the distance of the grating to the image plane is minimized. Around 50% of the light passing the grating will not have any dispersion, causing a 0 th order point spread function (PSF) to appear. The other 50% of the light is dispersed based on its wavelength and will create a second, slightly elongated 1 st order diffraction pattern. Image not to scale. (b) Simulation of low-dispersion sSMLM data of two spectrally different emitters with λ 1 (corresponding to d 1 ) < λ 2 (corresponding to d 2 ). Six emitters create in total five 0 th and 1 st order diffraction pairs on the detector, which can be linked together (two 1 st order diffraction patterns are not captured in the field of view). The obtained distances between the 0 th and 1 st order diffraction patterns (d 1 and d 2 ) are a measure for the average emission wavelength. (c) Single frame of raw data of a DNA-PAINT nanoruler sample showing 85 spatio-spectrally resolvable emitters in a 31 × 31 μm 2 field of view. The red outline is enlarged in the inset, in which the 0 th and 1 st order diffraction patterns are encircled in magenta and green, respectively. Scale bar represents 5 μm. (d) Comparison of achievable emitter density in standard (non-spectrally resolved) SMLM (black dotted line), our approach (red line), and sSMLM with 20−30 pixels spectral pattern elongation, taken from Bongiovanni et al. 37 The shaded background indicates the standard deviation as determined from repeating the simulations.

Nano Letters pubs.acs.org/NanoLett Letter
As the PSFs from the obtained 1 st order diffraction pattern show only minor excess width compared to the PSFs in the 0 th order, we were able to employ existing super-resolution algorithms to independently obtain subpixel localizations of the 0 th and 1 st order diffraction patterns. Next, we linked the localizations with each other in the dispersion direction, with the distance between the 0 th and 1 st order diffraction patterns (d 1 and d 2 in Figure 1b; further called "0 th -to-1 st -order distance") being a direct measure for the average emission wavelength of the emitter (λ 1 < λ 2 with d 1 < d 2 ). Moreover, the excess width of the 1 st order diffraction pattern compared to the width of the PSF in the 0 th order is a measure for the width of the emission spectrum. This directly results in a spectral accuracy being limited primarily by spatial single-molecule localization accuracy, an area of research that is progressing very rapidly via both software and hardware developments. 42,55,60 Using minimal dispersion, the achievable density in our implementation without the need for specialized highdensity fitting algorithms is around 5 times higher than earlier implementations of sSMLM, where the spectral information is spread out over 20−30 pixels (Figure 1d; Methods, Supporting Information). 37 We determined the distance between the dispersioninducing optics of the grating to the camera chip to be 6.9 ± 0.1 mm in our system (Methods, Supporting Information, Supplementary Figure 1a,b). The spectral dispersion (SD) was determined by calculating the 0 th -to-1 st -order distance of a sample labeled with ATTO542 and ATTO655 (Supplementary Figure 1c). From the median distance of these obtained distances and the mean emission profile of the fluorophores, a spectral dispersion of ∼0.21 nm/nm (spatial/spectral; equivalent here to ∼27 nm/px (spectral/spatial)) was determined. We did not observe a wavelength dependency on the angle between the spatial and spectral profiles (Supplementary Figure 1d).
The signal-to-noise ratio of the obtained 1 st order diffraction pattern, combined with high resolution of subpixel localization algorithms, indicates that small spectral differences can be elucidated. We imaged double-labeled fixated Cos7 cells, in which clathrin was labeled with the fluorophore CF660, and tubulin with CF680. Pseudocolor coding a super-resolved image based on 0 th -to-1 st -order distance reveals good separation of the labeled structures without further analysis (Figure 2a), even though these fluorophores only have a ∼10 nm intensity-weighted spectral separation in our microscope (CF660: 692 nm, CF680: 702 nm, Figure 2b; 12 nm difference in peak wavelength (CF660: 686 nm, CF680: 698 nm)).
Selecting image regions with mostly CF660-or CF680labeled structures (dotted outlines in Figure 2a) and fitting the corresponding 0 th -to-1 st -order distances with a Gaussian profile

Nano Letters
pubs.acs.org/NanoLett Letter reveals that CF660 has a 0 th -to-1 st -order distance of 3077 ± 2 nm (σ = 56 ± 2 nm; mean ±95% confidence interval (CI)) and 3128 ± 2 nm for CF680 (σ = 62 ± 2 nm; Figure 2c). This is a difference of 51 ± 2 nm in the raw data, which corresponds to a spectral distance of 10.9 ± 0.4 nm, in agreement with the weighted average and peak position difference of the microscope-corrected emission profiles. Next, we performed a DNA-PAINT experiment on polystyrene nanoparticles (NPs) that have DNA-PAINT docking strands for either ATTO647N or for ATTO655 (Figure 3a). These fluorophores have a weighted average emission wavelength separated by only ∼9 nm (685 nm and 693 nm, respectively, after correcting for optical components in our microscope, Figure 3b) and a peak emission wavelength separation of ∼14 nm (665 and 679 nm, respectively, after correcting for optical components in our microscope, Figure  3b). After isolating localizations belonging to individual beads and analyzing the 0 th -to-1 st -order distances of these emitters, two populations can be observed (Figure 3c).
The population with the lowest 0 th -to-1 st -order distance (green; Gaussian fit peak position: 3014 ± 1 nm, σ = 34 ± 1 nm, mean ±95% CI) was attributed to ATTO647N fluorophores. The population with the larger distances (magenta) was fitted with a combination of two Gaussian curves: one restricted to the fit of the first observed population; along with a unique Gaussian curve (Gaussian fit peak position: 3090 ± 2 nm, σ = 41 ± 3 nm). This population was attributed to ATTO655 fluorophores. The larger standard deviation of the ATTO655 population compared to the ATTO647N population can be attributed to a lower median localization accuracy (42 nm vs 50 nm), possibly caused by a difference in quantum yield (65% vs 30%). The spectral distance between these fitted peak positions (76 ± 2 nm distance; corresponding to 16.2 ± 0.4 nm spectral separation) is close to the difference between the emission peaks of both fluorophores but higher than the weighted average wavelength. This is possibly caused by deviations of the described wavelength-dependent efficiency of optical elements compared to our hardware implementation, which could lead to a shifted weighted mean emission wavelength of ATTO647N, as its emission maximum is close to the lower spectral cutoff (∼660 nm) of the filters and beam splitters used.
Next, all linked localizations were color-coded according to their distance (cutoff at the black dotted line in Figure 3c at 3054 nm). Visualization of the individual NPs (Figure 3d) then reveals their fluorophore distribution. This shows that the NPs are populated by either one DNA-PAINT docking strand or the other, with minimal cross-talk between the used fluorophores, which can be attributed to unspecific DNA− DNA interactions.
Taken together with the dSTORM data of the fixated cells, the obtained order of the mean emission wavelength of these four fluorophores (ATTO647N, CF660, ATTO655, CF680) coincided with that of the 0 th -to-1 st -order distance (calculated mean emission wavelength from spectra provided by the manufacturers and corrected for the optical properties of our setup): 685, 692, 693, 701 nm; mean 0 th -to-1 st -order distance: 3014, 3077, 3090, 3128 nm). The separation of these 0 th -to-1 storder distances suggests that simultaneous multiplexing of at least three fluorophores with single-wavelength excitation is technically possible under realistic experimental conditions (Supplementary Figure 2).
Next, we were interested whether we could expand the lowdispersion sSMLM to assess single-molecule Forster resonance energy transfer (smFRET). In a typical surface-based smFRET experiment, probes labeled with a donor and an acceptor fluorophore are immobilized and monitored over time. Depending on the experiment, changes in FRET and/or changes in acceptor/donor activity (i.e., blinking and bleaching) can be expected. Conventionally, a ratiometric spectral determination method is applied to separate the donor emission from the acceptor emission on different positions on a camera chip, and the intensity ratio between these is a measure for the relative FRET efficiency E*. In our implementation, we can use the full field of view of our camera and determine E* via the 0 th -to-1 st -order distance. We further can utilize the width of the 1 st order diffraction pattern as an additional way of discriminating FRET.
We performed smFRET measurements on well-characterized samples of immobilized double-stranded DNA that is dual-labeled with ATTO550 and ATTO647N. 27 Two samples respectively. First, we performed simulations using the known emission profiles corrected for the fluorophore's quantum yield and the optical elements in our microscope (Methods, Supporting Information; Figure 4a,b). These simulations of donor-only, 15% FRET, and 55% FRET samples show that the 0 th -to-1 st -order distance follows d donor < d 15% < d 55% , and the width of the 1 st order diffraction pattern follows σ donor < σ 15% < σ 55% . Experimentally, the immobilized DNA strands were imaged separately. Contrary to the multiplexing before, both the 0 th -to-1 st -order distance and the width of 1 st order diffraction pattern are measures for the FRET efficiency and were therefore visualized (Figure 4c,d). The experimental data agree with the simulations showing d donor < d 15% < d 55% (2620 nm, 2653 nm, 2781 nm, respectively) and σ donor < σ 15% < σ 55% (172 nm, 204 nm, 238 nm, respectively).
Next, we explored to which extent we can monitor dynamic behavior using spectrally resolved smFRET. While no direct state transitions are expected for this sample, there is occasional acceptor fluorophore blinking or bleaching, leading to a transition of FRET emission to donor-only emission. For this, we fitted the combined [d, σ] 2-dimensional histogram with four 2-dimensional Gaussian profiles (Figure 4c,d black crosses and ellipses). These profiles comprise donor-only, 15% FRET efficiency, 55% FRET efficiency, and "background" states (background state not shown). The "background" state is attributed to nonsense linkages occurring from sparse localizations unrelated to the FRET sample.
Time traces of individual emitters were further assessed (Figure 4e−h). The likelihood of an emitter belonging to the predetermined states was calculated (Methods, Supporting Information), and the most likely state dictated the background color of the graphs in Figure 4e and g. With this methodology, we were able to determine acceptor bleaching ( Figure 4e) and acceptor blinking (Figure 4g) in 15% and 55% (e, f) a bleaching acceptor fluorophore in a 15% FRET pair, and of (g, h) a blinking acceptor fluorophore in a 55% FRET pair. Horizontal gray lines with red and blue shading represent (e) 15% or (g) 55% FRET populations (red) and donor-only populations (blue), determined from the fits in panels c and d. The vertical green, yellow, and orange shading represent current FRET pair state, with green representing donor-only, yellow representing 15% FRET, and orange representing 55% FRET. The raw data corresponding to these FRET pairs throughout the observation time are shown in panels f and h. Scale bars in panels f and h represent 500 nm. Nano Letters pubs.acs.org/NanoLett Letter FRET experimental data. Accurate state determination of the 15% FRET measurement proved to be difficult due to the overlapping Gaussian profiles representing the FRET and donor-only states (Figure 4c), whereas this was better discriminable for 55% FRET.
Here we have demonstrated minimal-dispersion spectrally resolved SMLM (sSMLM), which fundamentally maximizes signal-to-noise and emitter density due to lowest possible photon spread on the detector. In our implementation, we used a single optical component add-on to the detection path leading to a spectral dispersion of just ∼0.2 nm/nm (spatial/ spectral), orders of magnitude lower than typical grating-based sSMLM implementations. With this implementation, we realized a five-times increased emitter density compared to similar approaches, achieved good separation of emitters with a ∼10 nm intensity-weighted spectral difference in STORM, and were able to observe changes between 0%, 15%, and 55% FRET efficiency in smFRET.
The low spectral dispersion allows us to use subpixel localization algorithms, a field that is advancing rapidly, 42 and all future developments are directly applicable to our sSMLM implementation, potentially benefiting from custom deeplearning routines addressing both spectral orders. 56,61 This could open up avenues for better spatial and spectral precision, more information obtained from the 1 st order pattern shape, and sSMLM at even higher emitter density. We additionally note that our implementation can readily be combined with SMLM techniques that modulate excitation patterns to increase localization precision. 49−54 Finally, low-dispersion sSMLM as presented here can be combined with threedimensional SMLM by engineering the PSF in the emission path. 57−59 Work by Song et al. employed subpixel localization algorithms for sSMLM, based on data obtained via the −1 st and +1 st orders of a nonblazed transmission grating. 38 While this solution is elegant and uses all photons that arrive on the camera chip for both spatial and spectral localization, blocking out the 0 th order leads to a significant loss of photons. Moreover, the implementation requires additional optical components (mirrors and lenses) to direct only the −1 st and +1 st orders on the camera chip. We additionally note that minimizing the spectral dispersion maximizes spectral precision by minimization of effects by shot-, background-, and read-out noise. 48 Further minimization of spectral dispersion in our implementation is possible by using a lower dispersion blazed grating by decreasing the effective distance of the grating to the camera chip, if the camera housing permits, or by placing the grating close to an intermediate image plane. While this could result in better spectral resolution and higher achievable sSMLM density, a trade-off of this minimization is decreased information content about the shape of the emission spectrum and the risk of imaging the grating itself on the camera. In this study, we used the shape of the emission spectrum only to discriminate FRET states from the donor-only state.
Taken together, we believe that our implementation of lowdispersion sSMLM will find widespread use due to its inherent simplicity and photon efficiency providing access to maximized spatiotemporal and spectral resolution. We further envision applications in which the photon efficient separation of spectrally different entities is desired such as in low-signal flow cytometry. Here, the ideas taken from super-resolution microscopy such as subpixel localization and spectral peak determination can be equally applied even for low-magnification configurations.

■ ASSOCIATED CONTENT Data Availability Statement
The code used in this manuscript can be found on Github, https://github.com/HohlbeinLab/sSMLMAnalyzer. Data underlying the manuscript are available at 10.5281/zenodo. 6778964.