Assessment of 3D MINFLUX data for quantitative structural biology in cells

MINFLUX is a promising new development in single-molecule localization microscopy, claiming a resolution of 1-3 nm in living and fixed biological specimens. While MINFLUX can achieve very high localisation precision, quantitative analysis of reported results leads us to dispute the resolution claim and question reliability for imaging sub-100-nm structural features, in its current state.

biological feature.
residing inside the Nup96 octamer both laterally and axially. However, while dSTORM has previously resolved an 36 inner ring of the NPC with WGA-CF680 and measured a diameter of 41 ± 7 nm with WGA-AF647 (Löschberger 37 et al., 2012;Thevathasan et al., 2019), such structure was neither apparent in the MINFLUX data ( Fig. S1) nor 38 discussed. Even after segmentation and closer inspection of WGA distributions, we could not visually discern a 39 ring-like structure (Fig. S2). Therefore, we conclude that for this particular sample, MINFLUX,unlike dSTORM,40 failed to resolve a ∼40 nm ring structure. 41 FOV ensemble analysis. We next analyzed the distribution of MINFLUX localizations in a field of view (FOV) using 42 PERPL, a structure-based modelling technique designed for incomplete data, as is often the case for single-molecule 43 localization microscopy (SMLM) (Curd et al., 2020) (Fig. 2). Specifically, we calculated the relative position 44 distribution (RPD) of Nup96 localizations and used its components in the lateral (xy) (Fig. 2a-f) and axial (z) 45 directions ( Fig. 2g-j). 46 For structures in xy, a model RPD that performed well for dSTORM localisations of Nup107 (Curd et al., 2020) was fit to the histogram of the experimental distances between MINFLUX localizations ( Fig. 2a-f). Localization 48 precision estimates (σ xy ) were 0.98 ± 0.02 nm, 3.20 ± 0.05 nm and 3.31 ± 0.08 nm (fitted value ± 1 s.d. uncertainty) 49 for the 2D, 3D 1-color and 3D 2-color datasets, respectively, of Gwosch et al. (2020), which broadly agree with the 50 published analysis. Secondary peaks may indicate consistent substructure within clusters in the 2D dataset down to a 51 distance of 7 nm (Fig. 2a, not modelled). However, this detail is lost in the 3D and 2-color datasets (Fig. 2c,e). If 52 σ xy is in the range 1-3 nm, we question whether resolution is possible at 1-3 nm, as is claimed. For instance, FWHM 53 ≈ 2.355σ, so σ xy of 1-3 nm implies FWHM of 2.4-7.1 nm for a single molecule, and we would not expect to clearly 54 resolve molecules closer than this.

55
Estimates of the Nup96 ring diameters varied from 104 to 109 nm and orders of symmetry from 7-to 9-fold, for 56 the different datasets ( Fig. 2a-f). The best fits do not follow the experimental distance distribution as closely as for  In the axial (z) direction, using PERPL ( Fig. 2g-j), we estimated localization precision (σ z ) at 2.28 ± 0.05 nm and  Live MINFLUX and filtering. The raw MINFLUX data comes in a processed tabular format including the parameters 75 used for event filtering (Sup Table 1). We reproduced the published images of Gwosch et al. (2020) and noted an 76 increased level of filtering from 20% to 83% as we moved from 2D to 3D to live MINFLUX data (Fig. 3). We assume the choice of event filtering parameters in the case of the live data ( Fig. 3d) was needed to deal with increased 78 background and movement during localization, and visually appeared to succeed in finding a collection of ≤38 true 79 positive Nup96 localizations. To apply MINFLUX to biological research questions, an explanation of how to select 80 the values of the event filtering parameters is needed. In particular, how to optimize the number and proportion of 81 true positives in an investigation of an unknown sample structure, where this may be challenging.

82
With N = 2 NPCs and 38 localizations, there was too little data available to assess the claim of resolution at 83 1-3 nm in living cells. Circular fits found diameters of 102 nm and 104 nm (Figure 1f, ideal result 107 nm). We are 84 unsure of σ xy and a related resolution limit in this case since we could not use PERPL analysis on only two instances 85 of the NPC with missing data. Gwosch et al. did not estimate σ xy and resolution for localizations in the live sample. 86 However, in a fixed sample, they measured σ xy ∼2 nm for the same label, Nup96-mMaple (Gwosch et al. (2020) 87 Supplementary Fig. 7), which gives a possible resolution limit (≈ FWHM) of ∼5 nm in that case. Therefore, we 88 cannot quantitatively assess resolution and reliability in the case of live samples, although precision and resolution 89 are generally degraded when moving from fixed to live specimens, so we do not expect nanometer (1-3 nm) resolution 90 in the case of live MINFLUX imaging, based on these datasets. it generated less precise and reliable results than established SMLM methods and appeared unable to resolve a 40-nm 94 ring structure.

95
3D Nup96-AF647 localisation precision in fixed samples, after event filtering, is impressive at σ = 1-3 nm. However, 96 we advise against interpreting localisation precision as resolution, which is intuitively understood as the distance at 97 which two nearby objects can be distinguished, is larger than 2σ at a lower limit, and is affected by other factors 98 such as localisation density. Furthermore, these localisation precisions were found after event filtering, which reduces 99 localisation density and may be difficult to perform effectively on an unknown sample. We suggest assessing resolution, 100 detection efficiency and exploration of event filtering parameters on blind samples, to demonstrate the potential of 101 this new technology. For an initial discussion of these issues, see Prakash (2021).  Fig. 1a (a), 3f (c) and 5c (e) and fits to them (b, d, f,) of nuclear porin models from 6-to 10-fold symmetry, including repeated single-molecule localisations (σxy), intra-and inter-cluster distances within a NPC, and background≈inter-pore distances (Curd et al., 2020). Symmetry, nuclear pore diameter (D) and σxy for the model selected by AICc (Curd et al., 2020) in each experiment (b, d, f). Indications of resolved intra-cluster substructure in a (*). ∆z distribution for the data in Gwosch et al. (2020) Fig. 3f (g) and 5c (i) and fit with a model including two layers of localisations and repeated single-molecule localisations (σz) (h, j).