Deciphering a hexameric protein complex with Angstrom optical resolution

  1. Hisham Mazal
  2. Franz-Ferdinand Wieser
  3. Vahid Sandoghdar  Is a corresponding author
  1. Max Planck Institute for the Science of Light, Germany
  2. Max-Planck-Zentrum für Physik und Medizin, Germany
  3. Friedrich-Alexander University of Erlangen-Nürnberg, Germany

Abstract

Cryogenic optical localization in three dimensions (COLD) was recently shown to resolve up to four binding sites on a single protein. However, because COLD relies on intensity fluctuations that result from the blinking behavior of fluorophores, it is limited to cases where individual emitters show different brightness. This significantly lowers the measurement yield. To extend the number of resolved sites as well as the measurement yield, we employ partial labeling and combine it with polarization encoding in order to identify single fluorophores during their stochastic blinking. We then use a particle classification scheme to identify and resolve heterogenous subsets and combine them to reconstruct the three-dimensional arrangement of large molecular complexes. We showcase this method (polarCOLD) by resolving the trimer arrangement of proliferating cell nuclear antigen (PCNA) and six different sites of the hexamer protein Caseinolytic Peptidase B (ClpB) of Thermus thermophilus in its quaternary structure, both with Angstrom resolution. The combination of polarCOLD and single-particle cryogenic electron microscopy (cryoEM) promises to provide crucial insight into intrinsic heterogeneities of biomolecular structures. Furthermore, our approach is fully compatible with fluorescent protein labeling and can, thus, be used in a wide range of studies in cell and membrane biology.

Editor's evaluation

This paper will be of interest to the structural biology community and people working on cryogenic fluorescence microscopy. This paper is a clear step forward in the use of single-molecule localization microscopy at Å resolution, thanks to low-temperature polarized super-resolution imaging and advanced data processing algorithms.

https://doi.org/10.7554/eLife.76308.sa0

Introduction

Proteins and their various assemblies are among the main constituents of all living systems and govern every aspect of cellular physiology in both healthy and disease states (Nelson, 2017; Mavroidis et al., 2004; Schliwa and Woehlke, 2003; Alberts, 1998). These biomolecular structures adopt sophisticated three-dimensional (3D) configurations during their multifaceted conformational changes. A full understanding of their spatial arrangements and associated heterogeneous configurations is crucial for elucidating their molecular mechanisms, helps guide the engineering of new proteins, and is a great asset for drug discovery (Renaud et al., 2018). Indeed, since the pioneering work of Perutz on protein crystals (Fersht, 2008), a variety of techniques such as X-ray crystallography (Shi, 2014) and nuclear magnetic resonance (NMR) spectroscopy (Kanelis et al., 2001) have been explored for gaining insight into protein structure and function. Advances in sample preparation, detector technology, and image processing based on single-particle analysis have also ushered in atomic resolution in cryogenic electron microscopy (cryoEM) studies of protein structure (Nakane et al., 2020; Kühlbrandt, 2014). The inherent resolution of this method is highly desirable, but lack of specific labeling makes it challenging to identify small variations such as structural inhomogeneities. Fluorescence microscopy, on the other hand, draws its success from an exquisite specificity in labeling, but has traditionally suffered from a limited resolution.

The recent advent of super-resolution (SR) fluorescence microscopy has opened new avenues for studying subcellular organization and is on the way to become a workhorse for biological studies (Lelek et al., 2021; Sahl et al., 2017; Weisenburger and Sandoghdar, 2015). However, conventional SR microscopy performed at room temperature is still not considered as a contestant in the arena of structural biology, where Angstrom-level information about the molecular architecture of proteins and protein complexes is sought after. To push the limit of fluorescence microscopy, one can perform measurements under cryogenic conditions (Böning et al., 2021, Hoffman et al., 2020; Dahlberg et al., 2020; Moser et al., 2019; Wang et al., 2019; Hulleman et al., 2018b, Xu et al., 2018; Weisenburger et al., 2017; Furubayashi et al., 2017; Li et al., 2015; Weisenburger et al., 2014; Weisenburger et al., 2013). In addition to slowing down photochemistry, which allows each fluorophore to emit several orders of magnitude more photons than at room temperature (Li et al., 2015; Weisenburger et al., 2014; Weisenburger et al., 2013), a key advantage of cryogenic temperatures is in offering superior sample preservation and high stability for Angstrom-scale structural studies. In one implementation, cryogenic optical localization in 3D (COLD) was introduced, where the stochastic intensity blinking of organic dyes gave access to the positions of up to four labeling sites on a single protein (Weisenburger et al., 2017). We recently employed a more robust protocol to identify individual fluorophores by exploiting the polarization of the fluorescence light dictated by the fluorophore orientation (Böning et al., 2021). This latter method was validated by measuring single distances on one-dimensional DNA nanorulers (Böning et al., 2021). Control of the fluorescence signal via polarization modulation has also been shown to offer an alternative to random blinking (Hulleman et al., 2018a, Hafi et al., 2014). In our current study, we introduce polarCOLD, which exploits polarization encoding for resolving several fluorophores in 3D. Importantly, we show that the distances and arrangements of protein complexes can be determined by combining images recorded from under-sampled structures. To demonstrate this, we first resolve three fluorophores on a trimer protein complex with Angstrom resolution. Next, we use partial labeling, a supervised particle classification procedure to solve a complete hexameric protein arrangement. We discuss the limits of our methodology for resolving structures with a certain degree of disorder as well as its promise for combination with cryoEM.

Results

polarCOLD on a protein trimer

Proliferating cell nuclear antigen (PCNA) is a central functional unit in genome repair and replication (Bruck and O’Donnell, 2001). The structure of this complex protein was solved by X-ray crystallography (Georgescu et al., 2008) and more recently by cryoEM (Madru et al., 2020). These studies have shown that PCNA forms a stable homo-trimer with a pseudo-hexameric shape. The stability and simple configuration of its structure make PCNA a good model system for benchmarking our imaging methodology. To study human PCNA with polarCOLD, we first fully labeled it via a His-tag linker on the N-terminal side of each subunit of the protein, forming an equilateral triangle (see Materials and methods section and Figure 1—figure supplement 1). PCNA complexes were then embedded in a hydrophilic poly-vinyl alcohol (PVA) matrix at sub-nanomolar concentration and spin-coated on a mirror-enhanced substrate (see Böning et al., 2021 for in-depth characterization of the mirror-enhanced substrate). The resulting density corresponds to fewer than one protein per μm2 on average so that individual proteins can be easily identified in diffraction-limited imaging. The samples were immediately imaged in our custom-built microscope (see Materials and methods section).

Figure 1a shows a schematic of the imaging setup. A liquid helium cryostat houses the cold stage as well as a scannable microscope objective (Weisenburger et al., 2017; Böning et al., 2021). A polarizing beam splitter in the detection path allows us to determine the orientation of dipole-like emitters projected onto the angular interval θ ϵ [0°, 90°] in the imaging plane. The high photostability of the fluorophores at T=4 K allows us to collect on average 260 photons per frame (per 14 ms) from a single fluorophore with a total number of registered photons exceeding 106 after 50,000 frames (Figure 1—figure supplement 2). Figure 1b–c (blue trace) displays two examples in which three different polarization states recur at various times. The long photo-blinking off-times (Figure 1—figure supplement 2d) allow one to identify each of the three fluorophores on a given protein complex separately (Weisenburger et al., 2017). More insight into polarization trace processing can be found in Figure 1—figure supplement 3. To resolve the polarization histograms in cases where they partially overlap (see, e.g., Figure 1c), we fit the data using an algorithm that combines unsupervised statistical learning tools with change-point detection in a model-independent manner (White et al., 2020). As illustrated by the red traces in Figure 1b and c, we can robustly identify the polarization states over time and hence assign the signal in each frame to a single fluorophore. We also verified the robustness of our assignment procedure by performing random assignment of frames, which resulted in single, unresolved spots (see Materials and methods section and Figure 1—figure supplement 4).

Figure 1 with 6 supplements see all
Photo-physics and co-localization of fluorophores at cryogenic temperatures.

(a) Schematics of the cryogenic optical microscope. Polarization-resolved detection allows for direct measurement of the in-plane dipole moment of fluorophores. Here we use circularly polarized light from a laser at λ=635 nm. A polarizing beam splitter in the detection path allows one to resolve the polarization state of each individual molecule. (b, c) Exemplary polarization time traces of two single proteins. (b) demonstrates a case of well-separated polarization states, whereas (c) displays a case with smaller separations between polarization states. Blue traces present the experimental polarization values for each frame, and the red lines show the polarization determined by the algorithm (White et al., 2020). Top panels shows the residuals of the fit. The blinking kinetics are exceptionally slow with on/off times in the range of seconds to minutes.

Figure 1—source data 1

Exemplary polarization time traces of two single proteins for Figure 1b-c.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig1-data1-v1.zip
Figure 1—source data 2

Overview of the recorded data and the experimental yield.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig1-data2-v1.zip

The number of fluorophores that can be simultaneously resolved depends on the blinking on-off dynamics (see Materials and methods section). Furthermore, the shot noise determines the angular resolution and, thus, the maximum number of resolvable polarization states per protein. This, in turn, directly affects the yield of resolved particles (see Figure 1—figure supplements 56). For example, in our current experiment, we used one polarization basis and projected all orientations to the limited space of θ ϵ [0°, 90°]. By taking the experimental angular resolution of ca. 5° (Figure 1—figure supplements 56), we can theoretically expect to resolve 70% of the particles which contain 3 fluorophores, and roughly 15% of the particles which contain 6 fluorophores. Adding a second polarization basis at a tilt of 45°, would allow one to double the angular space (Stallinga and Rieger, 2012). Moreover, using polarized illumination (Backer et al., 2016; Zhanghao et al., 2019) and a full 3D characterization of the dipole moments (Lieb et al., 2004; Mortensen et al., 2010; Hulleman et al., 2021) would result in less overlap and thus enhanced capacity for identifying different polarization states. We remark, however, that even in our current scheme, one can improve the number of resolved polarization states by excluding the ambivalent cases, albeit at the expense of the overall yield (see Figure 1—figure supplement 6 and Figure 1—source data 2 for the statistics of this analysis).

Having identified the individual fluorophores, we generate super-resolved images by clustering the respective coordinates and taking their averages (see Figure 2—figure supplements 1 and 2 for moving from traces to 2D resolved image). The top and bottom rows in Figure 2a display a selection of the measured and simulated 2D projection maps. To quantify their similarity, we computed a correlation score ranging from 0 to 1. We obtained 2D correlation scores of 0.92 or higher, representing nearly perfect agreement. To obtain a 3D model from our 2D localization maps, we use a single-particle reconstruction algorithm (Dvornek et al., 2015; Weisenburger et al., 2017; see Materials and methods section and Figure 2—figure supplement 3 for complete data set). Figure 2b shows that the reconstructed fluorophore volumes (red spheroids) agree well with the crystal structure of the PCNA protein (PDB: 1AXC) containing three identical subunits in an equilateral triangle. The slight asymmetry and deviation from the actual crystal structure can in part be attributed to the uncertainty introduced by the dye linker, 6-histidine linker and possibly the restricted rotational mobility of the dye itself, resulting in a minor localization bias. Indeed, by taking the dye linker into account and calculating the accessible volume (Kalinin et al., 2012), we found that our 3D reconstructed volumes correlate very well (0.96 correlation score) with the simulated accessible volumes (see Figure 2b, Animation 1).

Figure 2 with 3 supplements see all
3D reconstruction of the PCNA protein trimer.

(a) Experimentally obtained super-resolved 2D images (top row) of single proteins and simulated images based on the crystal structure (bottom row). The color code represents the occupation probability determined by the localization precision for each fluorophore. The localization precision in the simulated data was normalized. Scale bar is 3 nm. (b) Overlay of the crystal structure of human PCNA with the reconstructed fluorophore volumes shown as red spheroids (see online Animation 1). The transparent white clouds represent the accessible volume of the dye linker attached to the N-terminal side of the protein, calculated using the parameter of ATTO647N as provided in Kalinin et al., 2012. By fitting the reconstructed 3D volumes obtained from polarCOLD into the theoretical accessible volumes of the dyes, we find a correlation score of 0.96, indicating a correct 3D reconstruction. (c) The Fourier shell correlation (FSC, blue curve) of the two half data sets gives a resolution of 4.9 Å based on the half-bit criterion (red curve). (d) Distribution of the projected side lengths (blue) obtained from the localized positions shown in (b). The model fit (red) takes the finite localization uncertainty and the random particle orientation into account, resulting in 9.9±0.6 nm. The error of the model fit was estimated from 200 fits. The reconstructed 3D volume was calculated from 119 particles.

Animation 1
Human PCNA 3D reconstruction.

To evaluate the resolution of our 3D reconstructed volumes, we used the well-established method of Fourier shell correlation (FSC) (van Heel and Schatz, 2005). Here, we divide the 2D image data set into two randomly chosen groups and then determine their 3D reconstruction separately. Then we assess the cross-correlation (similarity) between the two 3D volumes in Fourier space as a function of spatial frequency. The overall resolution of a 3D reconstructed volume is thus obtained by finding the maximum spatial frequency corresponding to a correlation above a specific threshold value. Here, we used the half-bit criterion, which is a standard threshold curve used in single-particle cryoEM. As shown in Figure 2c, the intersection of this curve (red) with the FSC curve (blue) indicates at which spatial frequency we have collected a sufficient amount of information in order to interpret the 3D reconstructed volumes accurately (van Heel and Schatz, 2005). We find a remarkable resolution of 4.9 Å. We further quantified the size of the protein-dye conjugate via the pair-wise distances between the localized sites on each particle. The histogram in Figure 2d plots the distribution of the side lengths of the projected triangles and is well-described by a fit that considers the localization uncertainty as well as the random particle orientation (Böning et al., 2021, Weisenburger et al., 2017). We determine a side length of 9.9±0.6 nm in excellent agreement with the expected value. The uncertainty was determined via bootstrapping and is consistent with the resolution obtained from the FSC curve. We note that the high signal-to-noise ratio (SNR) of the method (see, e.g, Figure 2a), and the comparatively low information density per particle, deliver a good results from a total of 119 particles (see Figure 1—source data 2 for overall statistics), which is two to three orders of magnitude lower than the number required for typical cryoEM measurements (Cheng et al., 2015). Indeed, the low SNR in cryo-EM requires data from a large number of particles to be first averaged to establish 2D classes before using them for 3D reconstruction (Rosenthal and Henderson, 2003). In our case, each 2D projection directly contributes to the 3D reconstruction process.

Resolving a hexameric protein complex using partial labeling

The trimer structure discussed above involves only a single dye-dye distance. We now turn to resolving an example of more complex higher-order protein structures. Considering that a limited number of fluorophores can be resolved via stochastic blinking (Figure 1—figure supplement 6), we pursued a strategy of partial labeling of the sites of interest on a given individual protein complex. The piece-wise information, which involves various fluorophore arrangements and distances is then assembled to solve for the full architecture using prior knowledge of the symmetry. This concept has been successfully used to build structural models in NMR (Fiaux et al., 2002) and more recently in SR microscopy (Heydarian et al., 2018; Molle et al., 2018). To demonstrate this technique, we examine the homo-hexamer Caseinolytic Peptidase B (ClpB) of Thermus thermophilus in its quaternary structure (Lee et al., 2003; Diemand and Lupas, 2006; Figure 3—figure supplement 1a). ClpB is a molecular machine that rescues proteins from aggregation within cells (Doyle et al., 2013), and its structure was shown to be very stable in the presence of ATP at low concentrations (Mazal et al., 2019). Here, we labeled the M-domain of the protein at residue S428C such that the distance between two adjacent labeling sites is expected to be 9 nm after accounting for the dye linker (Figure 3—figure supplement 1a). We deliberately reduced the labeling efficiency to 50% in order to allow for 1/3 of the particles to carry three fluorophores as estimated by the binomial distribution (Figure 3—figure supplement 1b).

Labeled ClpB complexes were imaged in the same fashion as before in the presence of 2 mM ATP to stabilize the protein assembly. We obtained about 510 photons per frame on average and a significantly higher total number of photons after 100,000 frames since 87% of the molecules survived until the end of the acquisition interval (see Figure 3—figure supplement 1c-g for photo-physics charectrization). The photo-blinking behavior was similar to that observed in PCNA with a slight increase in the average off-on ratio, allowing robust fluorophore assignment. Hence, we fitted our polarization traces as described previously and selected only those cases that contained three fluorophores. Particles carrying three labels inherently fall into three distinct classes (I, II, and III) with multiple pair-wise distances of approximately 9, 15, and 18 nm (Figure 3—figure supplement 1a).

To analyze the resulting triangle images, we used a supervised classification scheme (Gao et al., 2004) to assign each 2D image to one of the three classes, taking the angle of its plane into account. Here, we simulated large data sets of 2D projections for each class and applied a template matching procedure based on the 2D cross-correlation between an experimental image and the simulated images. A control based on simulated ground truth images obtained from the crystal structure of ClpB showed an accuracy of 98% in template matching. Figure 3a displays examples of the 2D super-resolved images of each class (see Figure 3—figure supplement 2 for more examples of different projections, and Figure 3—source data 1 and Figure 3—source data 2 for full dataset). We assigned each image to the class that yielded the highest correlation score. Images that matched all classes with a difference below 10% for class 1 and 3, and below 3% for class 2 in the score (estimated from a simulation analysis, see Figure 3—figure supplement 3) were excluded from further analysis in order to avoid smearing (see Figure 1—source data 2 for all statistics). Next, we picked 2D images of each class with a localization precision better than 3 nm (Figure 3—figure supplement 1c) and correlation score better than 0.9 and calculated their respective 3D structures as described previously for PCNA. Overall, we obtained 232, 100, and 135 particles for classes 1, 2, and 3, respectively, representing a yield of ~7% for all the particles detected with three polarization states (see Figure 1—source data 2). Remarkably, as illustrated in Figure 3b, the reconstructed 3D volumes fit very well to the probable locations of the dye on the protein structure given by the accessible volumes (circles) on each protomer after taking the dye linker into account. The correlation scores between the reconstructed 3D volumes and the accessible volumes of the dyes were 0.98, 0.96, and 0.86 for classes I, II, and III, respectively. As a control, we also performed a reconstruction of unclassified 2D images, and confirmed that no structure was identified (Figure 3—figure supplement 4). Again, the FSC curves of the volumes (see Figure 3—figure supplement 5a-c) suggest an exquisite resolution of 4.0, 7.9, and 6.4 Å, for classes I, II, and III, respectively. Following the successful assignment of the three classes, we merged them as shown in Figure 3c to obtain the complete 3D shape of the hexamer structure (see Animation 2 and 3).

Figure 3 with 5 supplements see all
3D reconstruction of the ClpB hexamer protein.

(a) Top view of super-resolved 2D images for classes I, II, and III, as obtained from single-particle classification procedure. Scale bar is 3 nm. (b) Result of single-particle classification and averaging. The reconstructed 3D volume of each class nicely sits in the simulated accessible volume of the fluorophores (grey spheres). Red, blue and yellow spheres represent classes I, II, and III with correlation values of 0.98, 0.96, and 0.86, respectively. (c) 3D reconstruction of the complete hexamer obtained from merging the 3D volumes (red spheroid) of the three classes. Top figure shows the top view of the reconstructed 3D volume of the hexamer shape, and the bottom figure shows its 90° rotation (see online Animation 2 and 3Animation 2 and 3). Crystal structure of ClpB is shown as a cartoon in gold (PDB: 1QVR) (Lee et al., 2003; Diemand and Lupas, 2006). Reconstructed 3D volumes were calculated from 232, 100, and 135 particles for classes 1, 2, and 3, respectively.

Figure 3—source data 1

Full dataset coordinates of the 2D images used for 3D reconstruction of each class for Figure 3a.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig3-data1-v1.zip
Figure 3—source data 2

Full raw dataset coordinates of the unclassified particles for Figure 3a.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig3-data2-v1.zip
Figure 3—source data 3

3D reconstituted maps of each class for Figure 3b.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig3-data3-v1.zip
Figure 3—source data 4

3D reconstituted maps of the hexamer complex for Figure 3c.

https://cdn.elifesciences.org/articles/76308/elife-76308-fig3-data4-v1.zip
Animation 2
Thermus thermophilus ClpB 3D reconstruction (side view).
Animation 3
Thermus thermophilus ClpB 3D reconstruction (top view).

An efficient classification benefits from some prior knowledge of the structure. To examine the applicability of our method for samples with unknown symmetry or side lengths, we followed a similar procedure as demonstrated by Curd et al., 2021. First, by inspecting the raw distance histogram of our unclassified particles obtained from single distance measurements (two polarization states), we could identify a peak at ~9 nm as the most probable side length of our molecules (see Figure 4a). Next, we assumed three models with different symmetries for pentamers, hexamers and heptamers, but all sharing the same side length of 9 nm. We simulated an equivalent number of projections for each model and correlated them with our experimental data. Then, we used the Akaike information criterion (Portet, 2020; Curd et al., 2021) to find the best model that describes our experimental data (see Materials and methods section). As shown in Figure 4b, we found that the hexamer structure matches our data significantly better than the other models. In addition, we examined a different case of hexamer model with reduced symmetry and found that this model resulted in a considerably worse fit to our data (Figure 4b).

Figure 4 with 2 supplements see all
Quantitative model selection for classification.

(a) Histogram of pair-wise distances from particles with two polarization states, showing a clear peak at ~9 nm as the most probable side length of the ClpB molecules. (b) Based on the identified most probable side length we built models of the oligomer with different symmetries, but sharing the same side length of 9 nm, and performed single-particle classification of the experimental images. The AIC criterion shows that the hexamer is the best model. Reducing the symmetry of the hexamer results in a worse fit. (c) Schematic of a model deviating from perfect symmetry. Each corner is allowed to be shifted within a circle of radius R. The classification procedure remains accurate for R ≤ 1 nm.

Quantification of the robustness of our approach for samples without symmetry and order goes beyond the scope of our current study, but one simple strategy would be to consider a symmetric structure such as a hexamer and allow for each corner to deviate within a circle of radius R (see Figure 4c). Simulations show that classification of our current data becomes less robust for R>1 nm. We point out that solving a completely disordered structure without any prior knowledge would only be possible for complete labeling.

Classification and reconstruction have recently been applied to other single-molecule localization microscopy studies, and in some cases specific algorithms have been developed to handle some degree of partial labeling (Heydarian et al., 2018; Sieben et al., 2018; Salas et al., 2017). As shown in Figure 4—figure supplement 1, however, the existing algorithms do not provide a satisfactory solution for our data. We attribute this to the fact that the structures in our investigations are randomly oriented in 3D, are strongly under-labeled, and involve classes that are similar. A mathematical analysis of the performance of various algorithms for structures of different morphologies is beyond the scope of our current work.

Discussion

polarCOLD can be further improved and extended to other modalities, for example via combination with genetic labeling (Hoffman et al., 2020; Dahlberg et al., 2018; Liu et al., 2015; Creemers et al., 2000) or by exploiting the narrow absorption and emission spectra at low temperatures for co-localization of larger numbers of fluorescent markers. However, the currently achieved SNR and Angstrom resolution of cryogenic light microscopy is already capable of providing pivotal solutions for quantitative structural analysis of small proteins as well as large protein complexes and aggregates. A specially promising line of study concerns the conformation of membrane proteins in their native environment. Indeed, an estimated 20% of the human genome encodes membrane proteins and many of them are potential drug targets (Piccoli et al., 2013). The structure and active configuration of many proteins and protein complexes, however, remain out of reach because it is very challenging to crystallize or resolve them in their crowded surrounding (Cheng, 2018). In particular, heterogeneities caused by mobile domains and intrinsic distributions in assemblies cannot be preserved and classified in a robust fashion so that they lead to a smearing effect in the final reconstruction in EM studies (Nwanochie and Uversky, 2019; Serna, 2019; Scheres, 2016; Scheres et al., 2007, Orlova and Saibil, 2004; Papai et al., 2020). A future direction is, thus, to vitrify biological samples via rapid freezing (Dubochet and McDowall, 1981) or high-pressure freezing (Studer et al., 2008; Hoffman et al., 2020) in order to exploit the high spatial resolution, sensitivity and specificity of polarCOLD for identifying and determining the positions and orientations of individual particles of interest in low-contrast EM micrographs.

The combination of cryogenic light microscopy with cryoEM has attracted attention for cellular and tissue imaging, although the existing reports do not surpass a resolution of about 10 nm (Dahlberg and Moerner, 2021; Hoffman et al., 2020; Dahlberg et al., 2020; Wu et al., 2020; Tuijtel et al., 2019; Wolff et al., 2016; de Boer et al., 2015; Loussert Fonta and Humbel, 2015; Kaufmann et al., 2014b, Chang et al., 2014; Kaufmann et al., 2014a, Faas et al., 2013; Watanabe et al., 2011; Mironov and Beznoussenko, 2009; Agronskaia et al., 2008; Sartori et al., 2007). A very exciting prospect is now to combine polarCOLD and cryoEM on the single-particle level. As illustrated in Figure 4—figure supplement 2, by labeling a functional domain within a single molecule with two or three fluorophores, it should be possible to classify the particles based on conformational changes as indicated by their intermolecular distances. Similarly, by labeling two different protomers or functional domains within a protein complex, it will be possible to group the particles based on intramolecular distances that result from conformational changes between subunits. In addition, while we currently screen different symmetries to find the best model that fits to the structures under study, we plan to explore unsupervised classification schemes for identifying and classifying particles in an unknown sample. In addition, one can exploit the axial information directly and follow a similar approach by Heydarian et al., 2021. Altogether, these approaches will help to overcome various challenges connected to low purification yields and large backgrounds from cell membranes, paving the way to studying samples containing a heterogeneous distribution of similar structures and mapping their energy landscapes.

Materials and methods

Key resources table
Reagent type(species) or resourceDesignationSource or referenceIdentifiersAdditional information
Peptide, recombinant proteinPCNASigma AldrichCat. #: SRP5117
Peptide, recombinant proteinClpBhttps://doi.org/10.1038/s41467-019-09474-6
Chemical compound, drugATTO647N Ni-NTASigma AldrichCat. #: 02175–250 UG-F
Chemical compound, drugATTO647N MaleimideSigma AldrichCat. #: 05316–1 MG-F
Chemical compound, drugPoly-vinyl alcohol (PVA)Sigma AldrichCat. #: 360,627
Chemical compound, drug6-Hydroxy-2,5,7,8-tetramethylchroman-2-carboxylic AcidTCI Deutschland GmbHCat. #: H0726
Software, algorithmMATLAB 2020aMathWorks
Software, algorithmUCSF ChimeraXhttp://www.rbvi.ucsf.edu/chimerax
Software, algorithmDISCDOI: 10.7554/eLife.53357
Software, algorithm3D reconstructionhttps://doi.org/10.1016/j.jsb.2015.03.009
Software, algorithmFSChttps://www.ebi.ac.uk/emdb/validation/fsc/

Protein labeling

Request a detailed protocol

His-tagged human PCNA protein was purchased from Sigma Aldrich (catalogue number SRP5117), at a concentration of 6.6 µM. The protein was specifically labeled via the histidine linker on the N-terminal side of the protein with the dye ATTO647N containing a Ni-NTA functional group which was also purchased from Sigma Aldrich (catalogue number 02175–250 UG-F). First, the protein was desalted to a labeling buffer (25 mM HEPES, 25 mM KCl, pH 7.8) using a 7 K MWCO Zeba desalting column (ThermoFischer, cat. 89882) and then reacted with dyes at a ratio of 1:4 for 2 hr at RT. The protein was then desalted from the excess of dyes using the same desalting column. The labeling efficiency was estimated using an absorption spectrometer (Nanodrop 2000, ThermoFischer) confirming ~100% labeling efficiency. SDS-page and native gel indicated that the protein is indeed assembled and labeled. N-terminal His-tagged ClpB protein, mutated at residue 428 from alanine to cysteine, was purified as described in a previous publication (Mazal et al., 2019). The labeling of the single cysteine was done following the same procedure as describe above with ATTO647N maleimide as a specific labeling agent of the cysteine. Absorption spectroscopy, SDS-page and native gel indicate complete assembly of the protein, see previous publication (Mazal et al., 2019).

Sample preparation

Request a detailed protocol

Both protein samples were prepared in a similar way, and all the samples were prepared freshly on the same day. Proteins were diluted to a stock solution of 50 nM, in 25 mM HEPES, 25 mM KCl, 10 mM MgCl2, 0.5 mM TCEP at pH 7.8 (working buffer). In the case of ClpB 2 mM ATP was added to stabilize protein assembly. The stock solution of poly-vinyl alcohol (PVA) was prepared as follows: 15 µl of 8% PVA (0.3% final concentration) was diluted into 345 µl working buffer containing 1 mM Trolox and 2 mM ATP in the case of ClpB. Then 2 µl of the protein was mixed with the 360 µl PVA solution and filtered with a 100 nm spin filter (Whatman Anotop-10). 5 µl of this mixed solution was then spin coated (30 s at 1000 rpm followed by 3000 rpm for 60 s) onto a plasma-cleaned mirror-enhanced substrate, prepared in house (Böning et al., 2021), and immediately loaded into our custom-built cryogenic microscope.

Experimental setup

Request a detailed protocol

All experiments were performed in a cryogenic microscope that is built around a Janis ST-500 flow cryostat and operates at liquid helium temperature. Samples are loaded onto a cold finger and imaged by a 0.95 NA objective (Olympus MPLAPO 100 x), which is mounted in vacuum, onto two separate EMCCD cameras (Andor iXon) in a polarization-resolved configuration. The field of view spans 211 × 313 pixels with a pixel size of 190 nm. A more detailed description of the optical setup can be found in Böning et al., 2021. The laser intensity used in all experiments was set to ~1 kW/cm2 and images were recorded with 14 ms exposure time, more than five times faster than the typical off-time. Therefore, we are able to capture individual bursts and minimize the probability of overlapping emission from two or more fluorophores in a single frame. For each field of view, we collected a total of 50,000 or 100,000 frames for PCNA and ClpB, respectively.

Image analysis

Request a detailed protocol

We analyzed raw image stacks from two polarization channels with custom-written MATLAB software. Briefly, we perform dual-channel localization using maximum-likelihood estimation with a Gaussian PSF model. The minor asymmetry due to the 3D dipole orientation causes a small bias of less than 3 nm on average after filtering the data. Mechanical drift during the long acquisition times is first corrected based on image cross-correlation. Localized coordinates from both channels are then registered via an affine transformation and a non-linear correction. Registered PSFs are grouped and their polarization is calculated from the intensities. Single fluorophores are identified via this feature and used as fiducial markers to perform additional drift correction and establish a more precise image alignment via an interpolated map. The residual drift is well below 1 nm and the median registration error is less than 2 nm. Polarization time-traces of multi-fluorophore particles are then analyzed in more detail (see Figure 1—figure supplement 3) by a routine that is based on the DISC algorithm (White et al., 2020) to find the polarization states of multi-fluorophore conjugated particles. Once all polarization states are identified, the associated coordinates are averaged by taking their median to generate a 2D super-resolved image. We took particles which were fitted best to three polarization states and used them for further analysis (see Figure 1—source data 2 for statistics).

Single-molecule 2D image fitting

Request a detailed protocol

To fit the 2D maps of the human PCNA, we used the theoretical distances obtained from the structural model of human PCNA (PDB: 1AXC) and then we performed a least-squares fit over rotation angles and translation using a simulated annealing algorithm. Two-dimensional cross-correlation of the experimental and simulated maps yield a score better than 0.92. In the case of the hexamer protein, ClpB, we generated a large space of multiple 2D projections, ~25,000 images for each class, using a rotation matrix over x, y, and z axis with an angular resolution of 10°. The exact distances were evaluated from the real crystal structural model taking the dye linker into account (ClpB model, PDB: 1QVR, and the full assembly model taken from Diemand and Lupas, 2006). Then we performed 2D template matching between experimental data and all the images from the simulated data. Here, we used a 2D cross-correlation algorithm to calculate the similarity index between two images. The correlation score ranges from 0 to 1, where 1 indicates a 100% match. For each experimental image, we select the fit with the highest score as the best fit to our data, and to assign it to a class. Images that were fitted to all classes with a difference below 10% for classes 1 and 3, and 3% for class 2 in the score were excluded from further analysis to avoid smearing (see Figure 3—figure supplement 3).

Tomography and 3D reconstruction

Request a detailed protocol

We selected the particles which were fitted properly to the simulated 2D projections, that is particles with a cross-correlation score better than 0.9 and a localization precision better than 3 nm. The 2D maps of the particles were then normalized (maximum and width) to perform the 3D reconstruction. For a full 3D reconstruction, we used the subspaceEM algorithm (Dvornek et al., 2015) with default settings and 100 runs. The algorithm was initialized with an elliptical Gaussian placed in the center of the volume. The 2D maps of the trimer protein, PCNA, were constructed using a grid size of 120 × 120 with a pixel size of 1.5 Å. In the case of the hexamer protein, the grid size was 200 × 200 with a pixel size of 1.5 Å. The 3D volumes were then processed, fitted to the crystal structures/maps of the dye location and edited using ChimeraX (Goddard et al., 2018). The Fourier shell correlation curves (FSC), were computed using the freely available FSC server, provided by the Protein Data Bank in Europe website (https://www.ebi.ac.uk/pdbe/emdb/validation/fsc/). Here we divide the 2D images into two equally large data sets, and calculate their 3D reconstruction as described above. Then we align the two reconstructed volumes and calculate the FSC and determine the resolution based on half-bit criteria (van Heel and Schatz, 2005).

Quantitative model selection using the Akaike Information Criterion (AIC)

Request a detailed protocol

To show that our method can infer the proper symmetry of a protein complex, we followed a similar approach as demonstrated by Curd et al., 2021, but taking the random projections into account. First, we simulated multiple projections of a regular pentamer, hexamer and heptamer, all with the same side length of 9 nm. Given their symmetry, we obtain 2, 3, and 4 classes when considering three randomly placed fluorophores, respectively. We also simulated a deformed hexamer (as. 6-mer) composed of two isosceles triangles of 9 nm length on opposing sides of a square with a side length of 4 nm. Here, we need to consider a total of 5 classes. Then, we performed 2D cross correlation of the experimental data with these templates, and we picked the particles with a localization precision better than 3 nm and correlation score better than 0.5. The normalized sum of square residuals (nSSR) for each model is calculated as

nSSR=j=1Ni=1n1-xi2n

where j is the number of classes, n is the number of images per class, and xi is the score of the classified image. Then, we assessed how well each model fits the data by calculating its AIC value (Portet, 2020; Curd et al., 2021) via AIC=MlognSSR+2K, where M is the total number of particles in each model and K is the number of parameters (8, 10, 12, 13 for the pentamer, hexamer, heptamer and deformed hexamer, respectively). K was calculated based on the number of sides in each model, number of distances, and number of classes. For example, the pentamer has 5 sides, 1 distance, and 2 classes.

Distance fitting

Request a detailed protocol

For more quantitative analysis of the pair-wise distances we used a model described in Weisenburger et al., 2017; Böning et al., 2021. In short, we convolve the Rician distribution of the distance between two spots with finite localization uncertainty with the projection onto the image plane given by a cosine function. In the case of the PCNA trimer, we only need to consider a single distance due to the symmetry of the structure.

Determination of expected measurement yield depending on angular resolution

Request a detailed protocol

Our co-localization procedure relies on identifying the unique polarization signature belonging to one of a maximum of N fluorophores within a diffraction limited area. To this end, we can choose to consider the complete distribution of polarization values and identify well-separated peaks (histogram method), or consider polarization information in the time domain and identify recurring levels and transitions between them (DISC algorithm White et al., 2020). Both methods are limited in resolution, that is depending on the signal-to-noise ratio and switching rates there is a smallest difference between any two fluorophore dipole orientations that we are able to resolve. We consider a particle with N fluorophores to be completely resolved if all N-1 pairwise angular separations are above this angular resolution. Thus, the yield of completely resolved particles depends on how well two polarization states can be distinguished. We performed a Monte Carlo simulation to estimate this yield in our measurements. To this end, we generate large sets of N random fluorophore angles in the interval [0°,90°] and compute the pairwise differences. For a given angular resolution s between 0.1° and 90° we then calculate the fraction of particles for which all differences are larger than s. This simple estimation shows that for a case of just 2 fluorophores an angular resolution of 20° is already sufficient to achieve a yield of 50%. The same yield for the case of 6 fluorophores, however, requires an angular resolution of 2°.

Data availability

Raw data generated or analyses during this study are included in the manuscript file and supporting files. Source data files have been provided for Figure 1b, c, Figure 1 - figure supplement 1 and 3. Figure 2a, b, c, d, Figure 2 - figure supplement 3. Figure 3a, b, c, Figure 3 - figure supplement 2,4, 5. Figure 4a.

References

  1. Book
    1. Nelson DL
    (2017)
    Lehninger Principles of Biochemistry
    ‎W.H. Freeman.
  2. Conference
    1. Weisenburger S
    2. Jing B
    3. Renn A
    4. Sandoghdar V
    (2013)
    Proceeding of Spie
    SPIE.

Decision letter

  1. Felix Campelo
    Reviewing Editor; Institute of Photonic Sciences, Spain
  2. Volker Dötsch
    Senior Editor; Goethe University, Germany
  3. Sophie Brasselet
    Reviewer; Aix Marseille Université, CNRS, France

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Deciphering a hexameric protein complex with Å optical resolution" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Volker Dötsch as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Sophie Brasselet (Reviewer #1).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

All three reviewers of this manuscript have pointed out important questions that I'm convinced will help the authors to resubmit a revised paper that is much stronger and clearer. The detailed reviews are enclosed in this email, and I would strongly encourage the authors to address all the questions and comments raised in those reports. That being said, the most important points are listed here (see referee reports for the details):

1) About the non-Gaussian shape of a 3D-oriented dipole PSF: what is its effect on localization precision?

2) Discuss the limitations of the method: requirement for no dipole rotation, orientation discrimination, etc.

3) Discuss a more general/wider applicability of the method: how can this be applied to non-symmetric/unknown complexes. How necessary is it the use of particle symmetry to determine the structures?

4) Discuss the reliability of particle discrimination based on the fluorophore polarization state.

5) Provide a better/fairer comparison of this technique to EM (e.g. number of needed particles, low yield in determination, etc.)

6) Expand on the methods (especially because this is a methods paper), in particular in the image analysis bit.

Reviewer #1 (Recommendations for the authors):

Some points, detailed here, need to be addressed to clarify the findings, their conclusion and the claims.

1. Effect of the 3D orientation of the fluorophores.

1.1 Even though the used NA is low and the incident polarization is in the sample plane, the localization precision required here is extreme and necessarily comes in competition with the inacuracy of PSF-centroid determination when the dipole is oriented in 3D. In particular, PSFs from 3D dipoles are not symetric Gaussians anymore. It is quite surprising that the 3D-orientation bias does not surpass, in some situations, the other bias sources in the reported data. This bias needs to be estimated and introduced in the localization accuracy estimation, together with the other inaccuracy sources (noise, drift).

1.2 Is the presence of the mirror-enhanced substrate modifying the expected brightness from dipoles at different axial positions, as well as the 3D-orientation bias mentioned above?

2. Orientational and time emission dynamics of the fluorophores.

2.1 The orientation of the dipoles is considered fixed at low T. The authors should give the expected number for the rotational diffusion of the fluorophore. We would expect that if some degree of rotational motion is permitted by the linker between the dipole and protein, this could be an issue for the reconstruction of the information, even at low T since more orientations will be probable for the measured dipoles.

2.2 It is not clear how the time dynamics of emission of the fluorophores affects the reliability of the discrimination of the different oriented dipoles in the time traces recorded. Is there any need for fine-tuning this on-off dynamics as well as the recording time of the measurement such as only one dipole emits at a time ? Optimal settings rules should be mentioned.

3. Orientational discrimination of fluorophores.

3.1 The authors mention in line 126 “The maximum number of resolvable polarization states per protein is currently limited to fewer than 5, dictated by the width of the polarization histogram”

How is this number 5 extracted from the width of the histograms, and how does this number depend on the signal level, possible presence of orientational mobility (see above)?

3.2 Would this number increase if the measurement also includes the quantification of the 3D orientation of the dipoles as mentioned below? If the measurement also includes spectral discrimination in addition to polarization discrimination?

3.3 The previous work by the authors (Boning et al.,) involved important signal thresholding, which was determining in the final dipole-distances assessment. How necessary is the signal thresholding here and what how is the choice of this filtering performed?

3.4 The authors mention in line 128 the necessity to exclude cases where the expected number of polarization states is not well resolved. What is the origin of this effect and is this also a limit of the unsupervised statistical learning algorithm used to treat the time traces?

3.5 The authors mention an angular resolution required for resolving 3 vs 6 fluorophores in a structure. How is the experimental angular resolution measured and what are the experimental sources of errors that can affect this resolution?

4. Classification and reconstruction.

4.1 The fluorophores assignment relies on a robust supervised classification procedure and template matching, which decreases the number of unknowns. How would this classification perform if the structure would contain unknown parameters?

4.2 In Figure 3 —figure supplement 2, it is not visible by eye that some of these figures belong to one class rather than another: some of them look very similar (e.g. aligned 3 dots) and belong to different classes.

4.3 Table S1 shows that the yield to reconstruct 3D information is very low (for ClpB, 100-200 particles are used out of more than 23000 particles detected initially). A lot of particles are also excluded from the analysis because a low-confidence score (line 202). What would be needed to increase this yield and increase the confidence score?

5. Methods

5.1 Is the fact that there is an ambiguity on the projection angle retrieval a problem? Since the orientation measured is confined in the [0-90{degree sign}] range, can't the orientation discrimination be problematic if a whole angle sector is missed in the angles determination?

5.2 How is the 100 photon/frame threshold (Figure S3 legend) chosen, what is the rationale behind in terms of estimation precision and accuracy? Isn't intensity thresholding also excluding 3D dipoles orientations?

5.3 The authors mention in line 160 a model to extract the side lengths of the projected triangles, what are the detail and hypotheses of this model?

5.4 The 3D protein orientation is linked to the spatial repartition of the fluorophores, however there is no mention of the possibility to use the 3D z position of the fluorophores to infer such a 3D information. Can't this axial position be exploited?

5.5 There are several recent techniques developed to infer the 3D orientation from emitting dipoles (using PSF engineering and/or polarization splitting, possibly using radial-polarization filtering). The measurement of the 3D orientation of the fluorophores would be of great value to complete the information with an additional angular parameter, which could increase the number of measurable fluorophores. The authors should comment on this, and possibly test this extension of the approach.

5.6 The abstract (line 43) states that the method “promises to provide crucial insight into intrinsic, environmental and dynamic heterogeneities of biomolecular structures.” The present method requires a large amount of data to be collected in the context of a known, fixed protein structure which density is highly controlled. It seems therefore very robust in the limited framework of low-level labelling, immoblilized and low concentration proteins, however it is difficult to envision its application in a native, dynamic configuration where diffusion potentially comes into play. The authors should provide more detail about what would be the ingredients needed to enlarge the application range.

Reviewer #2 (Recommendations for the authors):

In Weissenburger 2017, the authors already presented the reconstruction of four fluorophore sites within streptavidin. Overall, I think the manuscript presents an advance of previous developments by the authors, but the generalizability and applicability to more complex samples remain partly open. This could be discussed in more detail.

1) Fluorophore discrimination depends on the correct assignment of the different polarization states. It would be good to include a representation of the uncertainties in fluorophore assignment and the resulting filtering procedure. For clarity, it would also help to color-code the different fluorophores as identified from the polarization state in the scatter plots. For example in Figure 2A, some spots appear larger or have an irregular shape, which could be associated with the uncertainty (as mentioned above) or the result of overlapping molecules for which a color code could help.

2) I am wondering about how generalizable the described reconstruction approach is or can be with respect to non-symmetric complexes or mixed samples of multiple protein species. To what extent does the reconstruction rely on prior knowledge of the investigated proteins? For example, the simulations used to classify the triangle images in the case of ClpB. What are the boundaries of these simulations (distance, angle, etc)? Is it possible to combine this with existing pattern extraction schemes (i.e. Curd, 2020)? Also, would more classes be helpful in increasing the precision or applicability of the approach? Particle reconstruction typically relies on many more classes unless some of them are redundant due to symmetries.

Reviewer #3 (Recommendations for the authors):

The paper by Mazal, Wieser and Sandoghdar presents a method to image molecular complexes with light microscopy, but still obtain resolutions/localization uncertainties of fluorescent labels that have been limited to cryo EM.

This submission build upon earlier work from the same group (Weisenburger 2017 and Boning 2021). The polarisation detection does not increase the sparsity compared to Weisenburger, but on the detection side they can identify different emitters based on their fixed dipole emission as shown in Boning. They further improve the sparsity, by explicit under labelling which later is compensated by particle registration and averaging. Compared to Boning the submission adds 3D reconstructions of two molecular complexes from 2D under labelled structures. The polarisation and localization method is improved technically, but the concept was already there.

The presented methodology and results are overall very nice.

In the following a few remarks and details that could help to improve the manuscript or that are unclear to me.

– Abstract: many it would good to directly state to how many sites you extend the method (l34).

In how far was the use of a supervised classification and use of particle symmetry really needed? (l93) Typically in cryoEM the use of symmetry is only needed to increase the resolution or in the absence of a large number of particles. What does the method deliver if the symmetry is not implied? That is a strong prior knowledge assumption and hinders discovery of new insights. The supervised classification is similar as model assumptions could be brought into the reconstruction.

In the introduction it would be nice to mention the idea of Hafi et al., Fluorescence nanoscopy by polarization modulation and polarization angle narrowing. Nature Methods, 11:579-584, 2014. and also Hulleman et al., Fluorescence polarization control for on-off switching of single molecules at cryogenic temperatures. Small Methods, page 1700323, 2018. Similar ideas to exploit polarisation of the fixed dipole emitters have been introduced here too.

l112 the polarisation state give the in-plane dipole angle. Would the method benefit from an estimation of the polar angle in addition? around l125 this could increase the number of resolvable polarisation states, but this might not be the limiting factor but the sparsity of the blinking?

l132 After localizations and identifying emitters by their polarisation state, how is the average position computed? Is that done weighted by the photon count?

l162 The remark the only 119 particles are needed for a reconstruction and not many more as in cryo EM is a bit misplaced. The information density in the LM is very low (only the few coordinates of the particles), while the EM map is a full image with millions of pixels that contain information. In addition early LM particle averaging results have used also hundredth of particles only typically. In addition the benefit of increasing the number of particles in LM is diminishing, because the additional particles cannot improve the localization precision but only fill in missing information thereby increasing the SNR. For cryo EM the resolution scale as sqrt(N) with number of particles, for LM this is much worse and eventually zero.

l181 Could you provide the t_on and t_off for the emitters. The 50% underlabelling is based on their ratio?

l203 "better 3 nm" -> better than 3 nm.

l205 yield of 7,46% point and comma not correct. 2 digits behind comma not needed. In addition why was the yield so low? Even if you consider that the polarisation state removes 3/4 of the particles the yield is still low. In cryoEM most particles needs to be removed due to damage resulting from stress close to the air water interface. What could be the reason here?

l 217 The claim to "all possible orientations" is a bit strong. I do not see why the 3 citations, in particular Huijben and Sieben, cannot deal with data in all possible orientations. I would argue that up to now, either the particles in a top-down configuration have been images because the structure was 2D or in the case of the much used NPC because the nuclear envelope is running nearly parallel to the cover slip. Therefore the NPC only have a limited tilt range. In conclusion, the claim to being the first is a big too much, as this is certainly not a methodical advance.

l277 the amount of Trolox 1mM seems quite a lot. Is this needed to suppress triplet states?

l288 What is the physical pixel size of the camera. The model is not stated, but the back projected size is stated as 190 nm with a 100x objective. This information is also not present in the supplement of Boning2021.

l290-298 Image analysis.

This part is vague on the details. Is the code available? I think as the analysis is not standard this needs to be open. The Poisson-weighted Gaussian MLE, what is that? Fitting a Gaussian function with Poisson likelihood? Will not polarisation effects of the restricted dipole play a role already at 0.95NA? (see Stallinga 2010, Optics Express 18:24461 and Engelhardt 2010, NanoLetters 11:209). The drift correction is non-linear, but for the rest it is totally unclear how it works. However, that is quite important as the acquisition time is a bit less than 1h.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your article "Deciphering a hexameric protein complex with Å optical resolution" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Volker Dötsch as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1) Please, consider answering the 3 specific comments from Reviewer #3, which need no new experiments, just clarification in the text and maybe some simple extra data analysis.

Reviewer #2 (Recommendations for the authors):

The authors have done well in responding to my comments, which at least partly came from misreading and/or misunderstanding on my side. The paper remains a difficult read but shows a clear step forward in using low temperature to increase SNR in single-molecule imaging.

Reviewer #3 (Recommendations for the authors):

Overall the authors did a very nice job with the revision. I would support the publication of the manuscript.

point (2) I strongly disagree that classification is needed and an assumption of symmetry. Many published averaging techniques for LM deal exactly with the problem described by the authors "In that sense, the incompletely labelled object inherently behaves as a heterogenous sample with multiple labelling configurations and distances". This is why most of the cryo-EM algorithms fail when applied to strongly under labelled data, they see the different labelling states as different classes. Please have a look at the work of Heydarian et al., (Nat Methods 2018, Nat Com 2021), Sieben et al., (Nat Methods 2018), Salas et al., (PNAS 2017), Shi et al., (PLOS one 2019) and maybe others.

The question at hand is, if the under labelling here is so severe that indeed the above methods cannot deliver a good reconstruction. From my own experience somewhere between 30-50% degree of labelling is needed to obtain a meaningful reconstruction. However, this depends also on the absolute number of sites. Some of the above papers have open access code, it might be worthwhile to try this at least.

The approach to validate their model via the AIC is good. The idea from Curd et al., have been well incorporated as this method typically only can learn something for the nearest neighbour distance, exactly what it has been used for her.

point (5) The average best position of a set of localisations from the same molecule is typically done by photon weighted averaging not median computation? If there are strong outliers which suggest the use of the median, maybe it would be better to remove the outliers first and then compute a weighted average?

point (13) I see that the author fit a 2D Gaussian, at the expense of a bit of blurring due to model mismatch. I would expect that the random dipole orientation do not average about but give rise to a large standard deviation of the localization than strictly needed with a correct model fitting. The localization will be unbiased due to the randomness, but it will result in a bit of extra blurring.

https://doi.org/10.7554/eLife.76308.sa1

Author response

Essential revisions:

All three reviewers of this manuscript have pointed out important questions that I’m convinced will help the authors to resubmit a revised paper that is much stronger and clearer. The detailed reviews are enclosed in this email, and I would strongly encourage the authors to address all the questions and comments raised in those reports. That being said, the most important points are listed here (see referee reports for the details):

1) About the non-Gaussian shape of a 3D-oriented dipole PSF: what is its effect on localization precision?

2) Discuss the limitations of the method: requirement for no dipole rotation, orientation discrimination, etc.

We address these points in section 1, 2 and 3 of Reviewer 1.

3) Discuss a more general/wider applicability of the method: how can this be applied to non-symmetric/unknown complexes. How necessary is it the use of particle symmetry to determine the structures?

We address this point in section 4 of Reviewer 1, section 1 and 3 of Reviewer 2, and section 2 of Reviewer 3.

4) Discuss the reliability of particle discrimination based on the fluorophore polarization state.

We discuss this point in sections 2, 3 and 5 of Reviewer 1, point 2 of Reviewer 2, and point 4 of Reviewer 3.

5) Provide a better/fairer comparison of this technique to EM (e.g. number of needed particles, low yield in determination, etc.)

We discuss this point in section 4 of Reviewer 1, section 3 of Reviewer 2, and sections 2, 6 and 9 of Reviewer 3.

6) Expand on the methods (especially because this is a methods paper), in particular in the image analysis bit.

We address this issue in section 13 of Reviewer 3

Reviewer #1 (Recommendations for the authors):

Some points, detailed here, need to be addressed to clarify the findings, their conclusion and the claims.

1. Effect of the 3D orientation of the fluorophores.

1.1 Even though the used NA is low and the incident polarization is in the sample plane, the localization precision required here is extreme and necessarily comes in competition with the ymmetric of PSF-centroid determination when the dipole is oriented in 3D. In particular, PSFs from 3D dipoles are not ymmetric Gaussians anymore. It is quite surprising that the 3D-orientation bias does not surpass, in some situations, the other bias sources in the reported data. This bias needs to be estimated and introduced in the localization accuracy estimation, together with the other inaccuracy sources (noise, drift).

The reviewer points to an important consideration. The localization bias due to the asymmetry of the dipole PSF is a major contribution to the overall localization uncertainty. However, the overall extent of this contribution is reduced by the fact that we preferentially excite the in-plane components and detect them more efficiently due to the lower NA. In addition, by discriminating against less bright fluorescence signals, we eliminate contributions from molecules that have dipole moments with substantial axial component. We estimate the contribution to the localization inaccuracy to be no more than 3 nm on average as was discussed in our recent publication (Böning et al., ACS Photonics 2021). Nevertheless, it is possible to reach subnanometer resolution with a more selective choice of the particles that are included in the averaging process, as validated in Figure 2b-d and Figure 3b.

In the revised manuscript, we extended the image analysis part of the methods section to include our estimates for the contributions to the localization uncertainty.

1.2 Is the presence of the mirror-enhanced substrate modifying the expected brightness from dipoles at different axial positions, as well as the 3D-orientation bias mentioned above?

Yes, the effect of the mirror is two-fold. It increases the power radiated into the upper half space of the objective depending on the axial position and, thus, affects the collection efficiency. Moreover, due to interference of the incoming plane wave with the reflected wave the intensity of illumination is modulated in the axial direction. This creates an optical sectioning effect such that the brightness of fluorophores also depends on their axial position with respect to this standing wave pattern. We choose the spacer thickness on top of the mirror to be approximately a quarter-wavelength, such that the first antinode coincides with the spacer-sample interface. As mentioned in the main text (line 95) we refer the readers to our recent publication (Böning et al., ACS Photonics 2021) for more details on the properties of the mirror substrate.

L95: “see (Böning et al., 2021 for in-depth characterization of the mirror-enhanced substrate).

2. Orientational and time emission dynamics of the fluorophores.

2.1 The orientation of the dipoles is considered fixed at low T. The authors should give the expected number for the rotational diffusion of the fluorophore. We would expect that if some degree of rotational motion is permitted by the linker between the dipole and protein, this could be an issue for the reconstruction of the information, even at low T since more orientations will be probable for the measured dipoles.

We assume the orientation of dipoles to be fixed since the environment of the fluorophore is in a fully frozen solid state. Polarization-resolved imaging supports this assumption since for single fluorophores we find shot-noise limited polarization angles that remain constant over time. The shot noise limit of our polarization signal is ca. 4-5 degrees (see Author response image 1). Moreover, for single dye molecules we do not observe any jumps or diffusion of their polarization angle during 25 min of measurement time. This is fully in line with the wisdom of more than three decades of low temperature single-molecule spectroscopy. Thus, we can safely conclude that the rotational diffusion is less than a few degrees based on the angular resolution. This information is mentioned in the following passage of the main text, including the figure as well.

L118: “The number of fluorophores that can be simultaneously resolved depends on the blinking on-off dynamics (see method section). Furthermore, the shot noise determines the angular resolution and, thus, the maximum number of resolvable polarization states per protein. This, in turn, directly affects the yield for resolved particles (see Figure 1 —figure supplement 5-6). For example, in our current experiment, we used one polarization basis and projected all orientations to the limited space of θ ϵ [0°, 90°]. By taking the experimental angular resolution of ca. 5° (Figure 1 —figure supplement 5-6), we can theoretically expect to resolve 50% of the particles which contain 4 fluorophores, and roughly 15% of the particles which contain 6 fluorophores.”

Author response image 1
Polarization histogram of single fluorophore ATTO 647N imaged at 4K.

The histogram shows single population with a width of 5 degrees, which matches the expected shotnoise value estimated from 100 photons.

2.2 It is not clear how the time dynamics of emission of the fluorophores affects the reliability of the discrimination of the different oriented dipoles in the time traces recorded. Is there any need for fine-tuning this on-off dynamics as well as the recording time of the measurement such as only one dipole emits at a time? Optimal settings rules should be mentioned.

The dynamics of emission dictates the admissible fluorophore density and the necessary imaging speed. As a rule of thumb, to resolve N labels unambiguously within a diffraction-limited spot, we require an on-off ratio of less than 1/N. In addition, the frame rate of the camera needs to be faster than the average off-time to minimize the probability of overlapping contributions from many molecules in a single image. Unfortunately, it is generally not straightforward to predict the on- and off-times of fluorophores at cryogenic temperatures from their values in solution and common strategies cannot be directly used to engineer them. While we have confirmed an intensity dependence of the on-times in isolated fluorophores, the overall statistics still depends critically on the local environment. Here, in accordance with previous work at cryogenic temperatures (Weisenburger et al., Nat. Methods 2017; Hoffman et al., Science 2020), we find exceptionally long dark states. Since N is less than a handful of labels in our case, both criteria stated above are fulfilled. For the protein-fluorophore conjugates investigated in this work we found average on- and off-times of 17 ms and 73 ms, respectively. Therefore, an exposure time of 14 ms is sufficient to resolve single fluorophore blinking. We now mention this in the experimental setup part of the methods section.

L320: “The laser intensity used in all experiments was set to ~1 kW/cm2 and images were recorded with 14 ms exposure time, more than 5 times faster than the typical off-time. Therefore, we are able to capture individual bursts and minimize the probability of overlapping emission from two or more fluorophores in a single frame.”

3. Orientational discrimination of fluorophores.

3.1 The authors mention in line 126 “The maximum number of resolvable polarization states per protein is currently limited to fewer than 5, dictated by the width of the polarization histogram”

How is this number 5 extracted from the width of the histograms, and how does this number depend on the signal level, possible presence of orientational mobility (see above)?

Since fluorophores are orientated randomly with respect to each other, their polarization angles can only be distinguished if they are separated by more than the angular resolution given by the associated shot-noise, i.e., the angular resolution depends on the average number of photons per localization. Since fluorophores show a variation in their brightness, the angular resolution in assessing their dipoles also varies. In figure supplement 2 of Figure 1, we show the obtained total photons per frame. Taking an average of 100 photons, we arrive at an upper bound for the width of ca. 3-4 degrees. As stated before, our measurements of single fluorophores show that we reach this shot-noise limit (Author response image 1), suggesting no influence of orientational mobility. In figure supplement 6 of Figure 1 we performed a simulation to determine the fraction of particles for which all fluorophore polarizations are separated by more than the angular resolution. At a resolution of 5 degrees, we would obtain less than 30% yield beyond 5 fluorophores before filtering. This is indeed not a strict theoretical limitation. We have now removed this strict statement and only explain the important conditions for 3D localization.

3.2 Would this number increase if the measurement also includes the quantification of the 3D orientation of the dipoles as mentioned below? If the measurement also includes spectral discrimination in addition to polarization discrimination?

The reviewer makes very interesting suggestions, which are on the to-do list of future experiments. Including the 3D orientation of dipoles and/or their spectral information would certainly increase the dimensions along which fluorophores can be distinguished and would therefore scale up our method. Indeed, we could capitalize on the advantage of fixed orientation as well as narrower linewidths at low temperatures. However, acquiring more information comes at the expense of more sophisticated experimental protocols and analysis as well as longer integration times.

3.3 The previous work by the authors (Boning et al.,) involved important signal thresholding, which was determining in the final dipole-distances assessment. How necessary is the signal thresholding here and what how is the choice of this filtering performed?

The distribution of single dye-dye distances assumes an asymmetric shape when the localization uncertainty becomes comparable to the distance itself. In our previous work we demonstrated the narrowing of this distribution for a dye-dye separation of 6.5 nm by filtering the data for the brightest emitters. Here, the filtering process is less strict since the smallest distance we resolve is around 9 nm. The exact choice of the filtering process, however, also depends on the desired or required localization precision. In our current work, we could resolve the distance histogram (Figure 2 d) with a precision of 0.6 nm by selecting the particles with a localization precision below 3 nm. This particle selection comes over the yield, while in (Boning et al., 2020) the yield was ~ 60% , here we obtain ~ 7 %.

3.4 The authors mention in line 128 the necessity to exclude cases where the expected number of polarization states is not well resolved. What is the origin of this effect and is this also a limit of the unsupervised statistical learning algorithm used to treat the time traces?

There is no single algorithm which can achieve 100 % accuracy on the estimation of the number of polarization states. Therefore, in some cases the algorithm might over or under-estimate the number of polarization states, thus affecting the number of detected molecules. In particular, if the projections of two polarization are very close, or if the SNR is very low it might not be possible to distinguish them in a reliable fashion. In our case, we prefer to use an unsupervised algorithm as it is very user friendly and adapted very well to our pipeline. We modified our statement in the main text accordingly:

L118:” The number of fluorophores that can be simultaneously resolved depends on the blinking on-off dynamics (see method section). Furthermore, the shot noise determines the angular resolution and, thus, the maximum number of resolvable polarization states per protein. This, in turn, directly affects the yield for resolved particles (see Figure 1 —figure supplement 5-6). For example, in our current experiment, we used one polarization basis and projected all orientations to the limited space of θ ϵ [0°, 90°]. By taking the experimental angular resolution of ca. 5° (Figure 1 —figure supplement 5-6), we can theoretically expect to resolve 50% of the particles which contain 4 fluorophores, and roughly 15% of the particles which contain 6 fluorophores. Adding a second polarization basis at a tilt of 45°, would allow one to double the angular space (Stallinga and Rieger, 2012). Moreover, using polarized illumination (Backer et al., 2016, Zhanghao et al., 2019) and a full 3D characterization of the dipole moments (Lieb et al., 2004, Mortensen et al., 2010, Hulleman et al., 2021) would result in less overlap and thus enhanced capacity for identifying different polarization states. We remark, however, that even in our current scheme, one can improve the number of resolved polarization states by excluding the ambivalent cases, albeit at the expense of the overall yield (see Figure 1 —figure supplement 6 and Table supplement 1 for the statistics of this analysis).”

3.5 The authors mention an angular resolution required for resolving 3 vs 6 fluorophores in a structure. How is the experimental angular resolution measured and what are the experimental sources of errors that can affect this resolution?

Unlike the brightness, which is subject to fluctuations due to time-dependent changes in the blinking behaviour, the lifetime or the quantum yield, the polarization angle is remains unchanged. We estimate the experimental angular resolution based on the number of photons per frame (see figure supplement 2 of Figure 1 and Author response image 1) to be ca. 5 degrees. However, we want to emphasize that this is an average value and some fluorophores will show considerably broader polarization features. In particular, those with a polarization close to either 0° or 90° are not visible in one channel, narrowing the actual range of detection. It should also be noted that even 6 or more fluorophores could be resolved with our method if one is willing to accept the lower fraction of such particles extracted from a measurement.

4. Classification and reconstruction.

4.1 The fluorophores assignment relies on a robust supervised classification procedure and template matching, which decreases the number of unknowns. How would this classification perform if the structure would contain unknown parameters?

We would like to clarify that the fluorophore assignment does not depend on any model it is done using DISC algorithm in a model-free manner. However, the arrangement of the fluorophores on the protein is determined using prior knowledge/initial models. In general, if we have a homogenous sample, such as the trimer protein PCNA which adopts a single conformation only, or a homogenously and fully labelled object (Weisenburger et al., Nat. Methods 2017), our method would work without any prior assumptions. It should, of course, be evident that more data might be necessary to collect.

However, the challenge we want to address in this work is particle classification of an incompletely labelled system and reconstruction of a high-order protein complex based on that. In that sense, the incompletely labelled object inherently behaves as a heterogenous sample with multiple labelling configurations and distances, as described for the hexamer protein (class 1, class 2 and class 3). In addition, multiple different 3D orientations of proteins make the problem even more challenging. Thus, to classify these particles and converge to the true structure one needs to assume an initial model. To the best of our knowledge, currently, there are no algorithms that solve such an issue in a model-free manner (unsupervised). Even in the cryo-EM community it is common practice to work with existing structures as initial models for classification and alignment. Thus, the fact that we can nevertheless distinguish and resolve the three classes at high resolution, considering the subtle changes between them, is still remarkable. In fact, this establishes an important future step for our method and toward combination with other structural methods such as cryo-EM.

To address the reviewer concern, we reduced the number of known parameters, following a similar approach to (Curd et al., 2020), as also mentioned by Reviewer 2 and 3. Here we show that the symmetry assumptions could indeed be relaxed. First, by inspecting the raw distance histogram of our unclassified particles (Author response image 2a) we could identify a peak of ~ 9 nm as the major side length of our protein complex. Second, we assumed three different models, with different symmetries (pentamer, hexamer, and heptamer), but sharing the same side length of 9 nm (author response image 2b). Then we cross-correlated our experimental images with these different models and calculated the best model which fitted our data using the Akaike information criteria (AIC), as discussed now in the method section. Interestingly, we found that the best model is a symmetric hexamer (Author response image 2b). In addition, we examined a different case of hexamers. We simulated a deformed shape of hexamer structure, which composed of two side lengths 9 and 4 nm (Author response image 2b), and we found that this model fitted our experimental data significantly worse (Author response image 2b). Next, we also tried to test how minor deviations from a perfect hexamer could affect the classification procedure. Here, we found that changes up to ± 1 nm could still successfully classify our experimental data (Author response image 2c). Given these results, we believe that our approach does not rely on a strict assumption/knowledge of the particle symmetry since one can blindly screen different initial models for classification, and find the best choice which describes the experimental data. Quantification of the robustness conditions for a general particle with limited prior knowledge is an interesting mathematical task, which is beyond the scope of our current work.

To address the reviewer’s concerns, we have modified the main text as follows:

L225: “Efficient classification benefits from some prior knowledge of the structure. To examine the applicability of our method for samples with unknown symmetry or side lengths, we followed a similar procedure as demonstrated by (Curd et al., 2021). First, by inspecting the raw distance histogram of our unclassified particles obtained from single distance measurements (two polarization states), we could identify a peak at ~9 nm as the most probable side length of our molecules (see Figure 4a). Next, we assumed three models with different symmetries for pentamers, hexamers and heptamers, but all sharing the same side length of 9 nm. We simulated an equivalent number of projections for each model and correlated them with our experimental data. Then, we used the Akaike information criterion (Portet, 2020, Curd et al., 2021) to find the best model that describes our experimental data (see method section). As shown in Figure 4b, we found that the hexamer structure matches our data significantly better than the other models. In addition, we examined a different case of hexamer model with reduced symmetry and found that this model resulted in a considerably worse fit to our data (Figure 4b).

Quantification of the robustness of our approach for samples without symmetry and order goes beyond the scope of our current study, but one simple strategy would be to consider a symmetric structure such as a hexamer and allow for each corner to deviate within a circle of radius R (see Figure 4c). Simulations show that classification of our current data becomes less robust for R>1 nm. We point out that solving a completely disordered structure without any prior knowledge would only be possible for complete labelling.”

Author response image 2
Quantitative model selection for classification.

(a) Histogram of pair-wise distances from particles with two polarization states, showing a clear peak at ~9 nm as the most probable side length of the ClpB molecules. (b) Based on the identified most probable side length we built models of the oligomer with different symmetries, but sharing the same side length of 9 nm, and performed single-particle classification of the experimental images. The AIC criterion shows that the hexamer is the best model. Reducing the symmetry of the hexamer results in a worse fit. (c) Schematic of a model deviating from perfect symmetry. Each corner is allowed to be shifted within a circle of radius R. The classification procedure remains accurate for R<=1 nm.

4.2 In Figure 3 —figure supplement 2, it is not visible by eye that some of these figures belong to one class rather than another : some of them look very similar (e.g. aligned 3 dots) and belong to different classes.

The reviewer is right in that all the classes presented here are similar to some extent, for example all of them share at least one similar side length. For example, Class 1 and Class 2 share similar distances of 15 nm. Class 1 and 3 share the distances 9 and 15 nm and so on. If our projection were top views only, this would be obvious for us and we would not need any type of supervised classification. However, due to multiple orientations of the particles, the side lengths are somewhat distorted (looks shorter in some cases) and thus, it hard to judge by eye which particle is related to which class. In addition, all classes contain triangles, so that at certain projections they would look similar.

As verified by simulations, the 2D cross correlation algorithm turns out to be powerful in determining the class of each particle. The classification based on simulation yield 98% accuracy. Among the 2% of the misclassified particles, results showed that 46% were wrong assignment of particles from class 3 into class 1, 23% of particles form class 3 into 2, 23 % of class 2 into 1, and 8 % are of class 1 into 3. By examining the correlation score of these misclassified particles, we found that they had very similar scores. As seen in Figure 3 —figure supplement 3, some have a negligible score of 0 and others are up to 5 %. However, as described in the main text, we include information from particles with a very high correlation score above 0.9, and at the same time, we exclude particles with score difference below 10 %. Thus, theoretically, we should be able to exclude them completely from our classification and 3D reconstruction. We would like to note that the 2% misclassification is due to a number of simulated projections. In fact, we found that by increasing the number of projections (finer steps of rotational degrees) we could reduce misclassifications to 0.5%, leading a difference between the scores that is essentially 0.

We now clarify the sentence regarding to this and add a supplement figure:

L207: “Images that were fitted to all classes with a difference below 10% for class 1 and 3, and below 3% for class 2 (estimated from a simulation analysis, see Figure 3 —figure supplement 3) in the score were excluded from further analysis in order to avoid smearing.”

4.3 Table S1 shows that the yield to reconstruct 3D information is very low (for ClpB, 100-200 particles are used out of more than 23000 particles detected initially). A lot of particles are also excluded from the analysis because a low-confidence score (line 202). What would be needed to increase this yield and increase the confidence score?

The harsh reduction from ~ 23000 particles into ~ 6000, in the case of ClpB, has different origins. First, as described in the main text, we started by 50% labelling of the protein complexes. As shown in Figure 3 —figure supplement 1 b, 70% of the particle are disqualified as they are labeled with 1, 2, or 4-6 fluorophores. Second, some particles are also missed during the time trace fitting using DISC algorithm, for example, particles assigned to two polarization states rather than three. Third, proteins might be damaged during preparation and handling (e.g., spin coating). Fourth, although the photophysics is generally improved at low temperatures, some fluorophores bleach spontaneously. In addition, as pointed out in point 5.1, overlap of polarization projections can also contribute to reducing our yield.

The reduction of particles from 6000 particles to ~500 particles happens as a result of filtering: (1) Based on the correlation score (here we consider particles with a correlation score above 0.9). (2) Based on localization precision (here we consider particles with a localization precision better than 3 nm). (3) We further filter particles with a correlation score difference of up to 10% for class 1 and 3, and 3% for class 2. In fact, it is hard to judge the true yield, as here we rely on the performance of the DISC algorithm, which dictates the number of particles with three polarizations (as mentioned in point 3.4). Furthermore, we exclude particles where the SNR of the trace fit falls below 2. In the trimer case we might also be confronted with another issue. The labelling in that case was done on the histidine tag linker, which is not covalently bound, and might thus disassemble once we dilute the protein.

To increase the yield, one can take several measures. First, one could improve the assignment of fluorophore identification with better algorithms, or by including more information such as spectral information or resolving the 3D orientation of each fluorophore. For example, in the case of ClpB, specific labelling of each monomer in the complex would allow one to achieve full labelling with three fluorophores, thus increasing our yield. However, this approach is still very challenging and out of reach. In fact, we can also turn into fully labeling of the complex and exploit the 4-5 labeled molecules as well in order to gain more information. Second, regarding the classification, at the moment we select the best particles which match our model structure. In practice, we can increase our library for template matching, to include 1 nm variation at each distance, which would increase our yield. Lastly, one could expect to improve the sample quality by switching to 100% aqueous medium, such as vitreous ice, as used in the cryo-EM community.

5. Methods

5.1 Is the fact that there is an ambiguity on the projection angle retrieval a problem? Since the orientation measured is confined in the [0-90{degree sign}] range, can't the orientation discrimination be problematic if a whole angle sector is missed in the angles determination?

The reviewer is right to point out a limitation of the polarization-resolved detection. As we only measure in one polarization basis, all angles are mapped into the first quadrant. The resulting ambiguity reduces the range over which fluorophores can be distinguished by their dipole orientation. In principle, one can do better by measuring in two different polarization bases to extend the range to [0°,180⁰], and by extracting the 3D orientation of each fluorophore, e.g. by engineering the PSF. However, these approaches generally require a higher numerical aperture and signal. We now point these out in the manuscript.

L118:” The number of fluorophores that can be simultaneously resolved depends on the blinking on-off dynamics (see method section). Furthermore, the shot noise determines the angular resolution and, thus, the maximum number of resolvable polarization states per protein. This, in turn, directly affects the yield for resolved particles (see Figure 1 —figure supplement 5-6). For example, in our current experiment, we used one polarization basis and projected all orientations to the limited space of θ ϵ [0°, 90°]. By taking the experimental angular resolution of ca. 5° (Figure 1 —figure supplement 5-6), we can theoretically expect to resolve 50% of the particles which contain 4 fluorophores, and roughly 15% of the particles which contain 6 fluorophores. Adding a second polarization basis at a tilt of 45°, would allow one to double the angular space (Stallinga and Rieger, 2012). Moreover, using polarized illumination (Backer et al., 2016, Zhanghao et al., 2019) and a full 3D characterization of the dipole moments (Lieb et al., 2004, Mortensen et al., 2010, Hulleman et al., 2021) would result in less overlap and thus enhanced capacity for identifying different polarization states. We remark, however, that even in our current scheme, one can improve the number of resolved polarization states by excluding the ambivalent cases, albeit at the expense of the overall yield (see Figure 1 —figure supplement 6 and Table supplement 1 for the statistics of this analysis).”

5.2 How is the 100 photon/frame threshold (Figure S3 legend) chosen, what is the rationale behind in terms of estimation precision and accuracy? Isn't intensity thresholding also excluding 3D dipoles orientations?

The threshold for photons/frame is a trade-off between localization precision and accuracy of photon assignment. While a higher value excludes frames from analysis, possibly lowering the localization precision, it also increases the accuracy of assignment as the remaining frames have lower shot noise and can therefore be assigned with higher confidence. Dipoles with larger outof-plane angles are more likely to be completely excluded, which is, however, desirable given their larger localization bias. The exact threshold of 100 photon/frame (total counts from both cameras) was manually set as a reasonable compromise. We now add a figure in the supplement which shows the relation between localization precision as a function of the segment length (number of frames for each fluorophore). It shows that the longer the segment, the higher the localization precision.

5.3 The authors mention in line 160 a model to extract the side lengths of the projected triangles, what are the detail and hypotheses of this model?

The fit shown in Figure 2d is based on the following model: The distance distribution between two points in a plane, each determined with a finite localization uncertainty, is mathematically described by a so-called Rician function. If the expected distance value is large compared to the localization uncertainty it resembles a regular Gaussian. In our measurement, however, particles are randomly oriented within the sample, such that the distance between two localized fluorophores is a projection of the real distance. We include this projection by convolving the Rician with a cosine function, which results in the characteristic shape shown in Figure 2d. In the case of the PCNA trimer, we only need to consider a single distance due to the symmetry of the structure. We add this sentence in the relevant method section.

5.4 The 3D protein orientation is linked to the spatial repartition of the fluorophores, however there is no mention of the possibility to use the 3D z position of the fluorophores to infer such a 3D information. Can't this axial position be exploited?

We currently infer the third dimension by averaging many projections. This comes with its own complications as pointed out by all reviewers. By directly localizing fluorophores in 3D we could circumvent this step and facilitate our analysis. Instead of a 3D reconstruction, we could then perform particle alignment directly on the localization. However, most options for 3D localization (e.g. astigmatism, double helix, Vortex etc.) will enlarge the PSF and therefore reduce the SNR. Furthermore, the axial position is usually determined with lower precision than the lateral coordinates. As more groups join the emerging efforts in cryogenic super-resolution microscopy, we expect these variations to be examined in different arrangements.

5.5 There are several recent techniques developed to infer the 3D orientation from emitting dipoles (using PSF engineering and/or polarization splitting, possibly using radial-polarization filtering). The measurement of the 3D orientation of the fluorophores would be of great value to complete the information with an additional angular parameter, which could increase the number of measurable fluorophores. The authors should comment on this, and possibly test this extension of the approach.

We thank the reviewer for bringing up this issue. Measuring full 3D dipole orientation indeed adds information that improves the applicability of our method to more complex situations, while also eliminating the known dipole bias issue in a rigorous way. However, these approaches generally reduce the SNR because the emitted photons are either split between more channels or more pixels. Radial polarization as recently demonstrated by the group of W.E. Moerner decrease the signal by 30% even for perfectly horizontal dipoles. Nevertheless, we consider full 3D orientation estimation to be a valuable addition to be explored further. We have added the following statement to the manuscript to reflect this:

L125: “Adding a second polarization basis at a tilt of 45°, would allow one to double the angular space (Stallinga and Rieger, 2012). Moreover, using polarized illumination (Backer et al., 2016, Zhanghao et al., 2019) and a full 3D characterization of the dipole moments (Lieb et al., 2004, Mortensen et al., 2010, Hulleman et al., 2021) would result in less overlap and thus enhanced capacity for identifying different polarization states. We remark, however, that even in our current scheme, one can improve the number of resolved polarization states by excluding the ambivalent cases, albeit at the expense of the overall yield (see Figure 1 —figure supplement 6 and Table supplement 1 for the statistics of this analysis).”

5.6 The abstract (line 43) states that the method “promises to provide crucial insight into intrinsic, environmental and dynamic heterogeneities of biomolecular structures.” The present method requires a large amount of data to be collected in the context of a known, fixed protein structure which density is highly controlled. It seems therefore very robust in the limited framework of low-level labelling, immoblilized and low concentration proteins, however it is difficult to envision its application in a native, dynamic configuration where diffusion potentially comes into play. The authors should provide more detail about what would be the ingredients needed to enlarge the application range.

First, we would like to clarify that our usage of the word dynamic was unfortunate in that we did not mean to refer to temporally changing events. Rather, we wanted to emphasize conformational heterogeneity that might have happened due to dynamic processes prior to our sample preparation. We have now modified the abstract to avoid misunderstandings.

Our method is complementary to single particle cryo-EM methodology in the sense that both methods work on homogenous and diluted samples of interest. While cryo-EM measurements suffer from the sensitivity of electron interactions with biological matter, in our case we enjoy a high sensitivity that is intrinsic to the large cross section of light-matter interaction. Thus, here, we can use the information of known structures to resolve heterogenous ligand binding pockets or localizing protein ligand bindings within a protein complex, a realm which is challenging for cryo-EM. Second, as we show in the revised manuscript (Figure 4), our method can also tackle protein complexes with lower degree of symmetry, and less a priori knowledge of the structure to be resolved. Third, we believe that our approach would indeed be capable of resolving membrane proteins in their native environment, such as cell membrane, which is challenging for other biophysical methods, such as X-ray and cryo-EM. By preserving the sample in vitreous ice, we would be able to combine our optical classification scheme with cryo-EM studies, thus combining the strengths of the two methods. As demonstrated in our sample, optical particle classification can be fed to single particle cryo-EM to resolve particle orientation and conformation at high sensitivity. By doing so multiple conformational states can be identified and completely resolved (on the level of amino acids) at high resolution.

Reviewer #2 (Recommendations for the authors):

In Weissenburger 2017, the authors already presented the reconstruction of four fluorophore sites within streptavidin. Overall, I think the manuscript presents an advance of previous developments by the authors, but the generalizability and applicability to more complex samples remain partly open. This could be discussed in more detail.

A similar point was also raised by Reviewer 1. We attach our answer to points 4.1 and 5.6 of Reviewer 1 below and we address the question in depth in point 3:

In general, if we have a homogenous sample, such as the trimer protein PCNA which adopts a single conformation only, or a homogenously and fully labelled object (Weisenburger et al., Nat. Methods 2017), our method would work without any prior assumptions. It should, of course, be evident that more data might be necessary to collect.

However, the challenge we want to address in this work is particle classification of an incompletely labelled system and reconstruction of a high-order protein complex based on that. In that sense, the incompletely labelled object inherently behaves as a heterogenous sample with multiple labelling configurations and distances, as described for the hexamer protein (class 1, class 2 and class 3). In addition, multiple different 3D orientations of proteins make the problem even more challenging. Thus, to classify these particles and converge to the true structure one needs to assume an initial model. To the best of our knowledge, currently, there are no algorithms that solve such an issue in a model-free manner (unsupervised). Even in the cryo-EM community it is common practice to work with existing structures as initial models for classification and alignment. Thus, the fact that we can nevertheless distinguish and resolve the three classes at high resolution, considering the subtle changes between them, is still remarkable. In fact, this establishes an important future step for our method and toward combination with other structural methods such as cryo-EM.

Our method is complementary to single particle cryo-EM methodology in the sense that both methods work on homogenous and diluted samples of interest. While cryo-EM measurements suffer from the sensitivity of electron interactions with biological matter, in our case we enjoy a high sensitivity that is intrinsic to the large cross section of light-matter interaction. Thus, here, we can use the information of known structures to resolve heterogenous ligand binding pockets or localizing protein ligand bindings within a protein complex, a realm which is challenging for cryo-EM. Second, as we show in the revised manuscript (Figure 4), our method can also tackle protein complexes with lower degree of symmetry, and less a priori knowledge of the structure to be resolved. Third, we believe that our approach would indeed be capable of resolving membrane proteins in their native environment, such as cell membrane, which is challenging for other biophysical methods, such as X-ray and cryo-EM. By preserving the sample in vitreous ice, we would be able to combine our optical classification scheme with cryo-EM studies, thus combining the strengths of the two methods. As demonstrated in our sample, optical particle classification can be fed to single particle cryo-EM to resolve particle orientation and conformation at high sensitivity. By doing so multiple conformational states can be identified and completely resolved (on the level of amino acids) at high resolution.

1) Fluorophore discrimination depends on the correct assignment of the different polarization states. It would be good to include a representation of the uncertainties in fluorophore assignment and the resulting filtering procedure. For clarity, it would also help to color-code the different fluorophores as identified from the polarization state in the scatter plots. For example in Figure 2A, some spots appear larger or have an irregular shape, which could be associated with the uncertainty (as mentioned above) or the result of overlapping molecules for which a color code could help.

We thank the reviewer for his/her suggestions. We now provide the residuals of the DISC fit to better represent the uncertainty associated with this step (Figure 1 main text). Also, we provide a colour code of two different trajectories, and mention the associated localization precision for each fluorophore. In addition, we attach a figure which shows the localization precision as a function of the length of each assigned polarization (number of frames). As expected, the longer the segment, the higher the achievable localization precision. We refer the readers to these two new figures as well (Figure 2 —figure supplement 1-2).

Regarding the super-resolved projection images in Figure 2a, we can unfortunately not follow the Reviewer’s concern. All but the last image in Figure 2a, as well as the majority of images in the corresponding figure supplement 1, show three clearly separated spots and we describe in the figure caption that the spot size is given by the localization precision of the respective fluorophore. In some cases, the shape looks elongated as a result of two overlapping spots.

2) I am wondering about how generalizable the described reconstruction approach is or can be with respect to non-symmetric complexes or mixed samples of multiple protein species. To what extent does the reconstruction rely on prior knowledge of the investigated proteins? For example, the simulations used to classify the triangle images in the case of ClpB. What are the boundaries of these simulations (distance, angle, etc)? Is it possible to combine this with existing pattern extraction schemes (i.e. Curd, 2020)? Also, would more classes be helpful in increasing the precision or applicability of the approach? Particle reconstruction typically relies on many more classes unless some of them are redundant due to symmetries.

We have answered similar question raised by Reviewer 1 (point 4.1) and have revised the manuscript:

To address the reviewer concern, we reduced the number of known parameters, following a similar approach to (Curd et al., 2020), as mentioned by the Reviewer. Here we show that the symmetry assumptions could indeed be relaxed. First, by inspecting the raw distance histogram of our unclassified particles (Author response image 2a) we could identify a peak of ~ 9 nm as the major side length of our protein complex. Second, we assumed three different models, with different symmetries (pentamer, hexamer, and heptamer), but sharing the same side length of 9 nm (Author response image 2b). Then we cross-correlated our experimental images with these different models and calculated the best model which fitted our data using the Akaike information criteria (AIC), as discussed now in the method section. Interestingly, we found that the best model is a symmetric hexamer (Author response image 2b). In addition, we examined a different case of hexamers. We simulated a deformed shape of hexamer structure, which composed of two side lengths 9 and 4 nm (Author response image 2b), and we found that this model fitted our experimental data significantly worse (Author response image 2b). Next, we also tried to test how minor deviations from a perfect hexamer could affect the classification procedure. Here, we found that changes up to ± 1 nm could still successfully classify our experimental data (Author response image 2c). Given these results, we believe that our approach does not rely on a strict assumption/knowledge of the particle symmetry since one can blindly screen different initial models for classification, and find the best choice which describes the experimental data. Quantification of the robustness conditions for a general particle with limited prior knowledge is an interesting mathematical task, which is beyond the scope of our current work.

L225: “Efficient classification benefits from some prior knowledge of the structure. To examine the applicability of our method for samples with unknown symmetry or side lengths, we followed a similar procedure as demonstrated by (Curd et al., 2021). First, by inspecting the raw distance histogram of our unclassified particles obtained from single distance measurements (two polarization states), we could identify a peak at ~9 nm as the most probable side length of our molecules (see Figure 4a). Next, we assumed three models with different symmetries for pentamers, hexamers and heptamers, but all sharing the same side length of 9 nm. We simulated an equivalent number of projections for each model and correlated them with our experimental data. Then, we used the Akaike information criterion (Portet, 2020, Curd et al., 2021) to find the best model that describes our experimental data (see method section). As shown in Figure 4b, we found that the hexamer structure matches our data significantly better than the other models. In addition, we examined a different case of hexamer model with reduced symmetry and found that this model resulted in a considerably worse fit to our data (Figure 4b).

Quantification of the robustness of our approach for samples without symmetry and order goes beyond the scope of our current study, but one simple strategy would be to consider a symmetric structure such as a hexamer and allow for each corner to deviate within a circle of radius R (see Figure 4c). Simulations show that classification of our current data becomes less robust for R>1 nm. We point out that solving a completely disordered structure without any prior knowledge would only be possible for complete labelling.”

Regarding the comment on the classes. There is some difference between the classes we define here and the definition of classes raised by the reviewer. In cryo-EM, due to a very low SNR, they perform class average of particles which share the same in-plane orientation in order to increase the SNR for 3D reconstruction. Thus, the number of classes is defined by the total number of inplane orientation/projections. In our case, we do not need any class average, as our SNR is already very high, and these particles can be fed immediately into 3D reconstruction. This avoids the introduction of wrong classifications and some noise into the system. The three classes we define here as in the case of the hexamer protein, are based on conformational state/heterogeneity within the same protein complex. Each class here (act as different sample) adopts different 3D orientation and has different distances. As we have shown, give some degree of prior knowledge, we could identify these classes/conformations and resolve them separately in 3D. Such an approach can be combined with cryo-EM to help identifying particle orientation and heterogeneity with the same protein complex to resolve each state at high resolution without smearing.

We now mention this in the text:

L169: “Indeed, the low SNR in cryo-EM requires data from a large number of particles to be first averaged to establish 2D classes before using them for 3D reconstruction (Rosenthal and Henderson, 2003). In our case, each 2D projection directly contributes to the 3D reconstruction process.”

Reviewer #3 (Recommendations for the authors):

The paper by Mazal, Wieser and Sandoghdar presents a method to image molecular complexes with light microscopy, but still obtain resolutions/localization uncertainties of fluorescent labels that have been limited to cryo EM.

This submission build upon earlier work from the same group (Weisenburger 2017 and Boning 2021). The polarisation detection does not increase the sparsity compared to Weisenburger, but on the detection side they can identify different emitters based on their fixed dipole emission as shown in Boning. They further improve the sparsity, by explicit under labelling which later is compensated by particle registration and averaging. Compared to Boning the submission adds 3D reconstructions of two molecular complexes from 2D under labelled structures. The polarisation and localization method is improved technically, but the concept was already there.

The presented methodology and results are overall very nice.

In the following a few remarks and details that could help to improve the manuscript or that are unclear to me.

– Abstract: many it would good to directly state to how many sites you extend the method (l34).

We now state clearly to how many sites we extend this method.

“We showcase this method (polarCOLD) by resolving the trimer arrangement of proliferating cell nuclear antigen (PCNA) and six different sites of the hexamer protein Caseinolytic Peptidase B (ClpB) of Thermus thermophilus in its quaternary structure, both with Å resolution.”

In how far was the use of a supervised classification and use of particle symmetry really needed? (l93) Typically in cryoEM the use of symmetry is only needed to increase the resolution or in the absence of a large number of particles. What does the method deliver if the symmetry is not implied? That is a strong prior knowledge assumption and hinders discovery of new insights. The supervised classification is similar as model assumptions could be brought into the reconstruction.

Similar points were raised by Reviewer 1 (point 4.1) and Reviewer 2 (point 1 and 3). We attach our answer again for convenience:

In general, if we have a homogenous sample, such as the trimer protein PCNA which adopts a single conformation only, or a homogenously and fully labelled object (Weisenburger et al., Nat. Methods 2017), our method would work without any prior assumptions. It should, of course, be evident that collecting more data might be necessary.

However, the challenge we want to address in this work is particle classification of an incompletely labelled system and reconstruction of a high-order protein complex based on that. In that sense, the incompletely labelled object inherently behaves as a heterogenous sample with multiple labelling configurations and distances, as described for the hexamer protein (class 1, class 2 and class 3). In addition, multiple different 3D orientations of proteins make the problem even more challenging. Thus, to classify these particles and converge to the true structure one needs to assume an initial model. To the best of our knowledge, currently, there are no algorithms that solve such an issue in a model-free manner (unsupervised). Even in the cryo-EM community it is common practice to work with existing structures as initial models for classification and alignment. Thus, the fact that we can nevertheless distinguish and resolve the three classes at high resolution, considering the subtle changes between them, is still remarkable. In fact, this establishes an important future step for our method and toward combination with other structural methods such as cryo-EM.

To address the reviewer concern, we reduced the number of known parameters, following a similar approach to (Curd et al., 2020), as also mentioned by Reviewer 2. Here we show that the symmetry assumptions could indeed be relaxed. First, by inspecting the raw distance histogram of our unclassified particles (Author response image 2a) we could identify a peak of ~ 9 nm as the major side length of our protein complex. Second, we assumed three different models, with different symmetries (pentamer, hexamer, and heptamer), but sharing the same side length of 9 nm (Author response image 2b). Then we cross-correlated our experimental images with these different models and calculated the best model which fitted our data using the Akaike information criteria (AIC), as discussed now in the method section. Interestingly, we found that the best model is a symmetric hexamer (Author response image 2b). In addition, we examined a different case of hexamers. We simulated a deformed shape of hexamer structure, which composed of two side lengths 9 and 4 nm (Author response image 2b), and we found that this model fitted our experimental data significantly worse (Author response image 2b). Next, we also tried to test how minor deviations from a perfect hexamer could affect the classification procedure. Here, we found that changes up to ± 1 nm could still successfully classify our experimental data (Author response image 2c). Given these results, we believe that our approach does not rely on a strict assumption/knowledge of the particle symmetry since one can blindly screen different initial models for classification, and find the best choice which describes the experimental data. Quantification of the robustness conditions for a general particle with limited prior knowledge is an interesting mathematical task, which is beyond the scope of our current work.

To address the reviewer’s concerns, we have modified the main text as follows:

L225: “Efficient classification benefits from some prior knowledge of the structure. To examine the applicability of our method for samples with unknown symmetry or side lengths, we followed a similar procedure as demonstrated by (Curd et al., 2021). First, by inspecting the raw distance histogram of our unclassified particles obtained from single distance measurements (two polarization states), we could identify a peak at ~9 nm as the most probable side length of our molecules (see Figure 4a). Next, we assumed three models with different symmetries for pentamers, hexamers and heptamers, but all sharing the same side length of 9 nm. We simulated an equivalent number of projections for each model and correlated them with our experimental data. Then, we used the Akaike information criterion (Portet, 2020, Curd et al., 2021) to find the best model that describes our experimental data (see method section). As shown in Figure 4b, we found that the hexamer structure matches our data significantly better than the other models. In addition, we examined a different case of hexamer model with reduced symmetry and found that this model resulted in a considerably worse fit to our data (Figure 4b).

Quantification of the robustness of our approach for samples without symmetry and order goes beyond the scope of our current study, but one simple strategy would be to consider a symmetric structure such as a hexamer and allow for each corner to deviate within a circle of radius R (see Figure 4c). Simulations show that classification of our current data becomes less robust for R>1 nm. We point out that solving a completely disordered structure without any prior knowledge would only be possible for complete labelling.”

In the introduction it would be nice to mention the idea of Hafi et al., Fluorescence nanoscopy by polarization modulation and polarization angle narrowing. Nature Methods, 11:579-584, 2014. and also Hulleman et al., Fluorescence polarization control for on-off switching of single molecules at cryogenic temperatures. Small Methods, page 1700323, 2018. Similar ideas to exploit polarisation of the fixed dipole emitters have been introduced here too.

We thank the reviewer for these suggestions. We have inserted these citations in our revised manuscript.

L75: “Control of the fluorescence signal via polarization modulation has also been shown to offer an alternative to random blinking (Hulleman et al., 2018, Hafi et al., 2014).”

l112 the polarisation state give the in-plane dipole angle. Would the method benefit from an estimation of the polar angle in addition? around l125 this could increase the number of resolvable polarisation states, but this might not be the limiting factor but the sparsity of the blinking?

The reviewer is correct. By also measuring the polar angle we could increase the space for discriminating fluorophores based on their dipole orientation (a 3-fold improvement considering the available angles in a quarter circle vs. a half-sphere). But the reviewer is also right in recognizing that the limiting factor is indeed the blinking kinetics. To significantly scale the density of fluorophores would require much lower on-off ratios than we currently observe.

l132 After localizations and identifying emitters by their polarisation state, how is the average position computed? Is that done weighted by the photon count?

The calculation of the localized position is done with a median.

l162 The remark the only 119 particles are needed for a reconstruction and not many more as in cryo EM is a bit misplaced. The information density in the LM is very low (only the few coordinates of the particles), while the EM map is a full image with millions of pixels that contain information. In addition early LM particle averaging results have used also hundredth of particles only typically. In addition the benefit of increasing the number of particles in LM is diminishing, because the additional particles cannot improve the localization precision but only fill in missing information thereby increasing the SNR. For cryo EM the resolution scale as sqrt(N) with number of particles, for LM this is much worse and eventually zero.

The reviewer raises an interesting and subtle point. Indeed, our method delivers far less information about the structure than would be obtained from a cryoEM density map, i.e. we obtain a similar resolution but only for the locations of a few labelling sites. Perhaps unsurprisingly, we therefore require fewer particles to obtain a high-resolution reconstruction. Nevertheless, we wanted to raise awareness for this conceptual difference between LM and cryoEM. Following the reviewer’s justified remark, we attempted a better formulation of our original statement.

We now address this issue in the manuscript, where we write:

L165:“ We note that the high signal-to-noise ratio (SNR) of the method (see, e.g, Figure 2a), and the comparatively low information density per particle, deliver a good results from a total of 119 particles (see Table supplement 1 for overall statistics), which is two to three orders of magnitude lower than the number required for typical cryoEM measurements (Cheng et al., 2015). Indeed, the low SNR in cryo-EM requires data from a large number of particles to be first averaged to establish 2D classes before using them for 3D reconstruction (Rosenthal and Henderson, 2003). In our case, each 2D projection directly contributes to the 3D reconstruction process.”

l181 Could you provide the t_on and t_off for the emitters. The 50% underlabelling is based on their ratio?

We provide now the on time (17 ms) and the off time (73 ms) of the emitters. The 50% labeling was used intentionally to achieve a high percentage of 3 fluorophores/particle as described in the main text and in Figure 3 —figure supplement 1 b.

l203 "better 3 nm" -> better than 3 nm.

Thanks, we corrected this typo.

l205 yield of 7,46% point and comma not correct. 2 digits behind comma not needed. In addition why was the yield so low? Even if you consider that the polarisation state removes 3/4 of the particles the yield is still low. In cryoEM most particles needs to be removed due to damage resulting from stress close to the air water interface. What could be the reason here?

We answered this question in Reviewer 1 comment, point 4.3:

The harsh reduction from ~ 23000 particles into ~ 6000, in the case of ClpB, has different origins. First, as described in the main text, we started by 50% labelling of the protein complexes. As shown in Figure 3 —figure supplement 1 b, 70% of the particle are disqualified as they are labeled with 1, 2, or 4-6 fluorophores. Second, some particles are also missed during the time trace fitting using DISC algorithm, for example, particles assigned to two polarization states rather than three. Third, proteins might be damaged during preparation and handling (e.g., spin coating). Fourth, although the photophysics is generally improved at low temperatures, some fluorophores bleach spontaneously. In addition, as pointed out in point 5.1, overlap of polarization projections can also contribute to reducing our yield.

The reduction of particles from 6000 particles to ~500 particles happen as a result of filtering: (1) Based on the correlation score (here we consider particles with a correlation score above 0.9). (2) Based on localization precision (here we consider particles with a localization precision better than 3 nm). (3) We further filter particles with a correlation score difference of up to 10% for class 1 and 3, and 3% for class 2. In fact, it is hard to judge the true yield, as here we rely on the performance of the DISC algorithm, which dictates the number of particles with three polarizations (as mentioned in point 3.4). Furthermore, we exclude particles where the SNR of the trace fit falls below 2. In the trimer case we might also be confronted with another issue. The labelling in that case was done on the histidine tag linker, which is not covalently bound, and might thus disassemble once we dilute the protein.

We note that cryo-EM studies have similarly low yields of 10-30% (Merk et al., 2016), as a result of iterative brute-force filtration at the 2D class average steps, and consequently at the 3D classification process. This not explicitly due to the damage of molecules at the air water interface, but also from other factors as well, mainly particle heterogeneity.

l 217 The claim to "all possible orientations" is a bit strong. I do not see why the 3 citations, in particular Huijben and Sieben, cannot deal with data in all possible orientations. I would argue that up to now, either the particles in a top-down configuration have been images because the structure was 2D or in the case of the much used NPC because the nuclear envelope is running nearly parallel to the cover slip. Therefore the NPC only have a limited tilt range. In conclusion, the claim to being the first is a big too much, as this is certainly not a methodical advance.

We agree with the reviewer comment, and we modified the text accordingly.

l277 the amount of Trolox 1mM seems quite a lot. Is this needed to suppress triplet states?

Correct, based on our own previous works in the lab, and of others in field, the concentration was found to be acceptable ((Cordes et al., 2009, Gust et al., 2014)). However, we have not investigated the effect of Trolox concentration on photoblinking.

l288 What is the physical pixel size of the camera. The model is not stated, but the back projected size is stated as 190 nm with a 100x objective. This information is also not present in the supplement of Boning2021.

We thank the reviewer for spotting this missing detail. We use two Andor iXon 897 cameras with a physical pixel size of 16µm. We work at a magnification of 80x to adjust the pixel size in the sample plane according to the optimum in Mortensen’s formula.

l290-298 Image analysis.

This part is vague on the details. Is the code available? I think as the analysis is not standard this needs to be open. The Poisson-weighted Gaussian MLE, what is that? Fitting a Gaussian function with Poisson likelihood? Will not polarisation effects of the restricted dipole play a role already at 0.95NA? (see Stallinga 2010, Optics Express 18:24461 and Engelhardt 2010, NanoLetters 11:209). The drift correction is non-linear, but for the rest it is totally unclear how it works. However, that is quite important as the acquisition time is a bit less than 1h.

The MATLAB code for our image analysis is currently not fully available. Our PSF model is a 2D Gaussian that we fit with a Poisson likelihood. As the reviewer suspects, the dipole orientation does introduce a minor localization bias even at a numerical aperture of 0.95. However, the extent of this error is considerably reduced by selective excitation in epi-illumination and by postselection (at the expense of reducing the yield of the measurement). Furthermore, since the dipole orientation was found to be random with respect to the protein structure, the residual shift is in a random direction and therefore averaged out in the 3D reconstruction.

As for the rest of the image processing pipeline: we correct for drift in a coarse fashion based on image cross-correlation. To this end, we render a super-resolution image every 500 frames (batch size depends on frame rate and SNR) from all localizations. Since each particle has a finite spatial extent with several fluorophores contributing to the localizations, this procedure is not yet sufficient for our purpose. To push the residual drift below 1 nm, we e apply another correction step using occasional single fluorophores as fiducial markers, after they have been distinguished by their polarization, and subtract this averaged drift trajectory. A similar issue arises in image registration: in order to achieve highest accuracy, one needs to identify individual fluorophores as fiducial markers for a non-linear mapping between the imaging channels. Therefore, after coarse registration based on an affine transformation to pair up PSFs, we establish the non-linear matrix for precise registration based on the same single fluorophores identified previously. We have attempted to clarify the workflow in the methods section.

Image analysis

L327” We analyzed raw image stacks from two polarization channels with custom-written MATLAB software. Briefly, we perform dual-channel localization using maximum-likelihood estimation with a Gaussian PSF model. The minor asymmetry due to the 3D dipole orientation causes a small bias of less than 3 nm on average after filtering the data. Mechanical drift during the long acquisition times is first corrected based on image cross-correlation. Localized coordinates from both channels are then registered via an affine transformation and a non-linear correction. Registered PSFs are grouped and their polarization is calculated from the intensities. Single fluorophores are identified via this feature and used as fiducial markers to perform additional drift correction and establish a more precise image alignment via an interpolated map. The residual drift is well below 1 nm and the median registration error is less than 2 nm. Polarization time-traces of multi-fluorophore particles are then analyzed in more detail (see Figure 1 —figure supplement 3) by a routine that is based on the DISC algorithm (White et al., 2020) to find the polarization states of multi-fluorophore conjugated particles. Once all polarization states are identified, the associated coordinates are averaged by taking their median to generate a 2D super-resolved image. We took particles which were fitted best to three polarization states and used them for further analysis (see Table supplement 1 for statistics).”

References

Backer, A. S., Lee, M. Y. and Moerner, W. E. 2016. Enhanced DNA imaging using super-resolution microscopy and simultaneous single-molecule orientation measurements. Optica, 3, 659.

Böning, D., Wieser, F.-F. and Sandoghdar, V. 2021. Polarization-Encoded Colocalization Microscopy at Cryogenic Temperatures. ACS Photonics, 8, 194-201.

Cheng, Y., Grigorieff, N., Pawel and Walz, T. 2015. A Primer to Single-Particle Cryo-Electron Microscopy. Cell, 161, 438-449.

Cordes, T., Vogelsang, J. and Tinnefeld, P. 2009. On the Mechanism of Trolox as Antiblinking and Antibleaching Reagent. Journal of the American Chemical Society, 131, 5018-5019.

Curd, A. P., Leng, J., Hughes, R. E., Cleasby, A. J., Rogers, B., Trinh, C. H., Baird, M. A., Takagi, Y., Tiede, C., Sieben, C., Manley, S., Schlichthaerle, T., Jungmann, R., Ries, J., Shroff, H. and Peckham, M. 2021. Nanoscale Pattern Extraction from Relative Positions of Sparse 3D Localizations. Nano Letters, 21, 1213-1220.

Gust, A., Zander, A., Gietl, A., Holzmeister, P., Schulz, S., Lalkens, B., Tinnefeld, P. and Grohmann, D. 2014. A Starting Point for Fluorescence-Based Single-Molecule Measurements in Biomolecular Research. Molecules, 19, 15824-15865.

Hafi, N., Grunwald, M., Van Den Heuvel, L. S., Aspelmeier, T., Chen, J.-H., Zagrebelsky, M., Schütte, O. M., Steinem, C., Korte, M., Munk, A. and Walla, P. J. 2014. Fluorescence nanoscopy by polarization modulation and polarization angle narrowing. Nature Methods, 11, 579-584.

Hulleman, C. N., Huisman, M., Moerland, R. J., Grünwald, D., Stallinga, S. and Rieger, B. 2018. Fluorescence Polarization Control for On-Off Switching of Single Molecules at Cryogenic Temperatures. Small Methods, 2, 1700323.

Hulleman, C. N., Thorsen, R. Ø., Kim, E., Dekker, C., Stallinga, S. And Rieger, B. 2021. Simultaneous orientation and 3D localization microscopy with a Vortex point spread function. Nature Communications, 12.

Li, J. K., Wang, N. And Wu, X. S. 1998. Poly(vinyl alcohol) nanoparticles prepared by freezing–thawing process for protein/peptide drug delivery. Journal of Controlled Release, 56, 117-126.

Lieb, M. A., Zavislan, J. M. and Novotny, L. 2004. Single-molecule orientations determined by direct emission pattern imaging. Journal of the Optical Society of America B, 21, 1210-1215.

Merk, A., Bartesaghi, A., Banerjee, S., Falconieri, V., Rao, P., Davis, M. I., Pragani, R., Boxer, M. B., Earl, L. A., Milne, J. L. S. and Subramaniam, S. 2016. Breaking Cryo-EM Resolution Barriers to Facilitate Drug Discovery. Cell, 165, 1698-1707.

Mortensen, K. I., Churchman, L. S., Spudich, J. A. and Flyvbjerg, H. 2010. Optimized localization analysis for single-molecule tracking and super-resolution microscopy. Nature Methods, 7, 377-381.

Portet, S. 2020. A primer on model selection using the Akaike Information Criterion. Infectious Disease Modelling, 5, 111-128.

Rosenthal, P. B. and Henderson, R. 2003. Optimal Determination of Particle Orientation, Absolute Hand, and Contrast Loss in Single-particle Electron Cryomicroscopy. Journal of Molecular Biology, 333, 721-745.

Stallinga, S. and Rieger, B. 2012. Position and orientation estimation of fixed dipole emitters using an effective Hermite point spread function model. Optics Express, 20, 5896.

White, D. S., Goldschen-Ohm, M. P., Goldsmith, R. H. And Chanda, B. 2020. Top-down machine learning approach for high-throughput single-molecule analysis. eLife, 9.

Zhanghao, K., Chen, X., Liu, W., Li, M., Liu, Y., Wang, Y., Luo, S., Wang, X., Shan, C., Xie, H., Gao, J., Chen, X., Jin, D., Li, X., Zhang, Y., Dai, Q. And Xi, P. 2019. Super-resolution imaging of fluorescent dipoles via polarized structured illumination microscopy. Nature Communications, 10.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Essential revisions:

1) Please, consider answering the 3 specific comments from Reviewer #3, which need no new experiments, just clarification in the text and maybe some simple extra data analysis.

We addressed all reviewer #3 comments below.

Reviewer #3 (Recommendations for the authors):

Overall the authors did a very nice job with the revision. I would support the publication of the manuscript.

We thank the reviewer for the insightful and instructive comments again, and we are happy that he/she is satisfied with our previous answers.

point (2) I strongly disagree that classification is needed and an assumption of symmetry. Many published averaging techniques for LM deal exactly with the problem described by the authors "In that sense, the incompletely labelled object inherently behaves as a heterogenous sample with multiple labelling configurations and distances". This is why most of the cryo-EM algorithms fail when applied to strongly under labelled data, they see the different labelling states as different classes. Please have a look at the work of Heydarian et al., (Nat Methods 2018, Nat Com 2021), Sieben et al., (Nat Methods 2018), Salas et al., (PNAS 2017), Shi et al., (PLOS one 2019) and maybe others.

The question at hand is, if the under labelling here is so severe that indeed the above methods cannot deliver a good reconstruction. From my own experience somewhere between 30-50% degree of labelling is needed to obtain a meaningful reconstruction. However, this depends also on the absolute number of sites. Some of the above papers have open access code, it might be worthwhile to try this at least.

The approach to validate their model via the AIC is good. The idea from Curd et al., have been well incorporated as this method typically only can learn something for the nearest neighbour distance, exactly what it has been used for her.

We thank the reviewer for his/her suggestion. In general, other approaches could be used to deal with the data presented in our study. However, to the best of our knowledge there is no algorithm which is specifically designed for our problem, i.e., a partially labeled, randomly oriented structure with only a few fluorophores and 2D localizations. Nevertheless, we followed the reviewer’s suggestion and tested the code published by Heydarian et al., Nat Methods 2018.

First, we fed the coordinates of our 2D super-resolved experimental images obtained from the set of 467 particles used to generate the hexamer structure into the algorithm. As shown in Author response image 3, no clear reconstruction was obtained. To explore this code further, we simulated 3D rotations to the structure “tud_ralf” use in Heydarian et al. Nat Methods 2018 with 100% labelling efficiency. However, we applied rotations around the x axis within ± 180 degrees (i.e., not only in-plane rotations as in the original code) and used the code all2all to solve for the structure. Again, Author response image 4 shows that no clear reconstruction was obtained. Last, we also used a simulated data set of our own hexamer structure to avoid all noise. Here, we simulated 500 particles with 900 localization each, at localization precision below 1 nm, labelling efficiency of 50%, and for different 3D orientations in order to mimic our experimental conditions. Again, the algorithm fails to deliver a clear reconstruction (Author response image 5). We also remark that Sieben et al., Nat Methods 2018 do account for the 3D particle orientation, and they are able to separate different species with multi-color imaging. However, their algorithm is based on a cryo-EM classification, which does not deal with strongly under-labelled data. Similarly, the work published by Salas et al., (PNAS 2017).

We believe, there are several factors that limit the use of the other published methods. The first issue concerns the fact that in our samples, structure are randomly oriented in 3D whereas the existing algorithms accounts for translations and rotations in the image plane, i.e., they require particles to be oriented in the top view. Furthermore, our structures are strongly under-labelled and contain classes that are quite similar in terms of sides lengths and angles. It would indeed be interesting to develop a rigorous mathematical formalism for analyzing the performance of each algorithm for structures of different morphological nature.

Recent reports on 3D measurements (Heydarian et al., Nat Com 2021) offer an interesting alternative to reconstruction based on 2D projections. These are currently not applicable to our data as we only have access to 2D measurements. We now mention the prospect of 3D measurements and refer to the general issues in the modified manuscript and Figure 4 —figure supplement 1:

L245: “Classification and reconstruction have recently been applied to other single-molecule localization microscopy studies, and in some cases specific algorithms have been developed to handle some degree of partial labeling (Heydarian et al., 2018, Sieben et al., 2018, Salas et al., 2017). As shown in Figure 4 —figure supplement 1, however, the existing algorithms do not provide a satisfactory solution for our data. We attribute this to the fact that the structures in our investigations are randomly oriented in 3D, are strongly under-labeled, and involve classes that are similar. A mathematical analysis of the performance of various algorithms for structures of different morphologies is beyond the scope of our current work.”

L284: “In addition, while we currently screen different symmetries to find the best model that fits to the structures under study, we plan to explore unsupervised classification schemes for identifying and classifying particles in an unknown sample. In addition, one can exploit the axial information directly and follow similar approach by (Heydarian et al., 2021).”

Author response image 3
Experimental data obtained from partially labelled hexamer protein, as described in the main text.

Our experimental data samples 3D different orientations. (a-b) Examples of different initial 2D projections. (c) We fed the coordinates and localization precision of each of the 467 particles to the all2all algorithm developed by Heydarian et al., (Heydarian et al., 2018). The final reconstruction results fail to deliver the ground truth structural model. Instead, we obtained a scatter of points.

Author response image 4
Simulation of the structural model “tud_ralf” from (Heydarian et al., 2018) at 100% labelling efficiency (see inset in c) under the application of 3D rotations (± 180 degrees) around the x-axis only.

(a-b) Examples of different initial arrangements for two different orientations. (c) 256 particles were fed to the all2all algorithm (Heydarian et al., 2018). The final reconstruction fails to deliver the ground truth structural model.

Author response image 5
Simulation data of our hexamer structural model at 50% labeling efficiency.

Here we simulated 500 particles with 900 localizations each and localization precision below 1 nm, but sampling 3D different orientations. (a-b) Example of different initial particles, adopting different 3D orientations. (c) We fed the data to the all2all algorithm (Heydarian et al., 2018), and the final reconstruction, fails to deliver the ground truth structural model.

point (5) The average best position of a set of localisations from the same molecule is typically done by photon weighted averaging not median computation? If there are strong outliers which suggest the use of the median, maybe it would be better to remove the outliers first and then compute a weighted average?

We followed up on the reviewer’s hint at the difference between median and photon-weighted mean in the calculation of the coordinates. Indeed, even in the absence of strong outliers (see scatter plot example, Author response image 6) there is a minor difference in the coordinates, but it is less than 0.3 nm on average. We argue that since this is less than the localization precision, and small compared to the intermolecular distances in our sample, both estimators for the fluorophore position are acceptable. However, we share the reviewer’s view that the photon-weighted mean should be preferred, as it accounts for the different variances of the data points and results in a smaller standard error, especially in cases with higher fluctuation of the obtained photons per frame. For future experiments, we will therefore change the calculation of coordinates to photon-weighted means.

Author response image 6
Scatter plot showing the difference between median and photon-weighted mean in the calculation of the coordinates.

Even in the absence of strong outliers (a) there is a minor difference (less than 0.3 nm on average) in the coordinates (b). Please note the different axis scales between (a) and (b).

point (13) I see that the author fit a 2D Gaussian, at the expense of a bit of blurring due to model mismatch. I would expect that the random dipole orientation do not average about but give rise to a large standard deviation of the localization than strictly needed with a correct model fitting. The localization will be unbiased due to the randomness, but it will result in a bit of extra blurring.

The reviewer is right and we would like to clarify this point again. The localizations contain a small error due to the model mismatch of a Gaussian with a dipole-like PSF. This However, in order to obtain the final reconstruction, we collect many particles. Since the dipole orientation is random, the small shift occurs in a random direction. The ‘averaged’ 3D structure will therefore be unbiased. We have answered a similar question by reviewer 1:

“The reviewer points to an important consideration. The localization bias due to the asymmetry of the dipole PSF is a major contribution to the overall localization uncertainty. However, the overall extent of this contribution is reduced by the fact that we preferentially excite the in-plane components and detect them more efficiently due to the lower NA. In addition, by discriminating against less bright fluorescence signals, we eliminate contributions from molecules that have dipole moments with substantial axial component. We estimate the contribution to the localization inaccuracy to be no more than 3 nm on average as was discussed in our recent publication (Böning et al., ACS Photonics 2021). Nevertheless, it is possible to reach sub-nanometer resolution with a more selective choice of the particles that are included in the averaging process, as validated in Figure 2b-d and Figure 3b”

References

Heydarian, H., Joosten, M., Przybylski, A., Schueder, F., Jungmann, R., Werkhoven, B. V., Keller-Findeisen, J., Ries, J., Stallinga, S., Bates, M. and Rieger, B. 2021. 3D particle averaging and detection of macromolecular symmetry in localization microscopy. Nature Communications, 12.

Heydarian, H., Schueder, F., Strauss, M. T., Van Werkhoven, B., Fazel, M., Lidke, K. A., Jungmann, R., Stallinga, S. and Rieger, B. 2018. Template-free 2D particle fusion in localization microscopy. Nature Methods, 15, 781-784.

Salas, D., Le Gall, A., Fiche, J.-B., Valeri, A., Ke, Y., Bron, P., Bellot, G. and Nollmann, M. 2017. Angular reconstitution-based 3D reconstructions of nanomolecular structures from superresolution light-microscopy images. Proceedings of the National Academy of Sciences, 114, 9273-9278.

Sieben, C., Banterle, N., Douglass, K. M., Gönczy, P. and Manley, S. 2018. Multicolor single-particle reconstruction of protein complexes. Nature Methods, 15, 777-780.

https://doi.org/10.7554/eLife.76308.sa2

Article and author information

Author details

  1. Hisham Mazal

    1. Max Planck Institute for the Science of Light, Erlangen, Germany
    2. Max-Planck-Zentrum für Physik und Medizin, Erlangen, Germany
    Contribution
    Conceptualization, Data curation, Designed the experiment, Designed the experiment and carried out the measurements, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review and editing
    Contributed equally with
    Franz-Ferdinand Wieser
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2071-9552
  2. Franz-Ferdinand Wieser

    1. Max Planck Institute for the Science of Light, Erlangen, Germany
    2. Max-Planck-Zentrum für Physik und Medizin, Erlangen, Germany
    3. Friedrich-Alexander University of Erlangen-Nürnberg, Erlangen, Germany
    Contribution
    Data curation, Designed and prepared the optical setup, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review and editing
    Contributed equally with
    Hisham Mazal
    Competing interests
    No competing interests declared
  3. Vahid Sandoghdar

    1. Max Planck Institute for the Science of Light, Erlangen, Germany
    2. Max-Planck-Zentrum für Physik und Medizin, Erlangen, Germany
    3. Friedrich-Alexander University of Erlangen-Nürnberg, Erlangen, Germany
    Contribution
    Conceptualization, Designed the experiment, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Validation, Visualization, Writing - review and editing
    For correspondence
    vahid.sandoghdar@mpl.mpg.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2594-4801

Funding

Planning and Budgeting Committee of the Council for Higher Education of Israel (2020 - VATAT Postdoctoral fellowships for outstanding candidates from the Arab sector)

  • Hisham Mazal

Max Planck Society

  • Hisham Mazal
  • Franz-Ferdinand Wieser
  • Vahid Sandoghdar

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Tobias Utikal for assistance with cryogenic experiments, reading and commenting on the manuscript and Michelle Küppers for reading and commenting on the manuscript. We also thank Houman Mirzaalian Dastjerdi and Harald Koestler for fruitful discussions. We thank Simone Ihloff for support with materials, Alexander Gumann for preparing the mirror-enhanced substrates and Jan Renger for help with measuring the PVA film thickness. We are further grateful to Prof. Gilad Haran for the generous gift of ClpB proteins and plasmids. HM was supported by the Planning & Budgeting Committee of the Council of Higher Education of Israel. We are grateful to the Max Planck Society for financial support.

Senior Editor

  1. Volker Dötsch, Goethe University, Germany

Reviewing Editor

  1. Felix Campelo, Institute of Photonic Sciences, Spain

Reviewer

  1. Sophie Brasselet, Aix Marseille Université, CNRS, France

Version history

  1. Preprint posted: November 19, 2021 (view preprint)
  2. Received: December 13, 2021
  3. Accepted: May 12, 2022
  4. Version of Record published: May 26, 2022 (version 1)

Copyright

© 2022, Mazal, Wieser et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,679
    Page views
  • 232
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Hisham Mazal
  2. Franz-Ferdinand Wieser
  3. Vahid Sandoghdar
(2022)
Deciphering a hexameric protein complex with Angstrom optical resolution
eLife 11:e76308.
https://doi.org/10.7554/eLife.76308

Further reading

    1. Microbiology and Infectious Disease
    2. Structural Biology and Molecular Biophysics
    Dasvit Shetty, Linda J Kenney
    Research Article Updated

    The transcriptional regulator SsrB acts as a switch between virulent and biofilm lifestyles of non-typhoidal Salmonella enterica serovar Typhimurium. During infection, phosphorylated SsrB activates genes on Salmonella Pathogenicity Island-2 (SPI-2) essential for survival and replication within the macrophage. Low pH inside the vacuole is a key inducer of expression and SsrB activation. Previous studies demonstrated an increase in SsrB protein levels and DNA-binding affinity at low pH; the molecular basis was unknown (Liew et al., 2019). This study elucidates its underlying mechanism and in vivo significance. Employing single-molecule and transcriptional assays, we report that the SsrB DNA-binding domain alone (SsrBc) is insufficient to induce acid pH-sensitivity. Instead, His12, a conserved residue in the receiver domain confers pH sensitivity to SsrB allosterically. Acid-dependent DNA binding was highly cooperative, suggesting a new configuration of SsrB oligomers at SPI-2-dependent promoters. His12 also plays a role in SsrB phosphorylation; substituting His12 reduced phosphorylation at neutral pH and abolished pH-dependent differences. Failure to flip the switch in SsrB renders Salmonella avirulent and represents a potential means of controlling virulence.

    1. Plant Biology
    2. Structural Biology and Molecular Biophysics
    Jiyu Xin, Yang Shi ... Xiaoling Xu
    Research Article

    Carotenoid (Car) pigments perform central roles in photosynthesis-related light harvesting (LH), photoprotection, and assembly of functional pigment-protein complexes. However, the relationships between Car depletion in the LH, assembly of the prokaryotic reaction center (RC)-LH complex, and quinone exchange are not fully understood. Here, we analyzed native RC-LH (nRC-LH) and Car-depleted RC-LH (dRC-LH) complexes in Roseiflexus castenholzii, a chlorosome-less filamentous anoxygenic phototroph that forms the deepest branch of photosynthetic bacteria. Newly identified exterior Cars functioned with the bacteriochlorophyll B800 to block the proposed quinone channel between LHαβ subunits in the nRC-LH, forming a sealed LH ring that was disrupted by transmembrane helices from cytochrome c and subunit X to allow quinone shuttling. dRC-LH lacked subunit X, leading to an exposed LH ring with a larger opening, which together accelerated the quinone exchange rate. We also assigned amino acid sequences of subunit X and two hypothetical proteins Y and Z that functioned in forming the quinone channel and stabilizing the RC-LH interactions. This study reveals the structural basis by which Cars assembly regulates the architecture and quinone exchange of bacterial RC-LH complexes. These findings mark an important step forward in understanding the evolution and diversity of prokaryotic photosynthetic apparatus.