The resolution and accuracy of various single-molecule localization microscopes (SMLMs) are routinely benchmarked using simulated data, calibration "rulers," or comparison to secondary imaging modalities. However, these methods cannot quantify the nanoscale accuracy of an arbitrary SMLM dataset. Here, we show that by computing measurement stability under a well-chosen perturbation with accurate knowledge of the imaging system, we can robustly quantify the confidence of individual localizations without ground-truth knowledge of the sample. We demonstrate our broadly-applicable method, termed Wasserstein-induced flux, for measuring the accuracy of various reconstruction algorithms directly on experimental 2D and 3D data of microtubules and amyloid fibrils. We further show that the estimated confidences can be used to evaluate the experimental mismatch of computational models, enhance the accuracy and resolution of reconstructed structures, and discover sample heterogeneity due to hidden molecular parameters. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 localization accuracy, statistical confidence, localization software, model mismatch, Wasserstein distance Single-molecule localization microscopy (SMLM) has be-1 come an important tool for resolving nanoscale structures 2 and answering fundamental questions in biology (1-4) and 3 materials science (5). SMLM uses repeated localizations of 4 blinking fluorescent molecules to reconstruct high-resolution 5 images of a target structure. In this way, quasi-static features 6 of the sample are estimated from noisy individual images cap-7 tured from a fluorescence microscope. These quantities, such 8 as fluorophore positions (i.e., a map of fluorophore density), 9blinking "on" times, emission wavelengths, and orientations, 10 influence the random blinking events that are captured within 11 an SMLM dataset. By using a mathematical model of the 12 microscope, SMLM reconstruction algorithms seek to estimate 13 the most likely set of fluorophore positions and brightnesses 14 (i.e., a super-resolution image) that is consistent with the 15 observed noisy images.
16A key question left unresolved by existing SMLM method-17 ologies is: How well do the SMLM data, i.e., the images of 18 blinking single molecules (SMs), support the super-resolved im-19 age produced by an algorithm? That is, what is our statistical 20 confidence in each localization? Intuitively, one's interpreta-21 tion of an SMLM reconstruction could dramatically change 22 by knowing how trustworthy each localization is.23Existing metrics for assessing SMLM image quality can 24 be categorized broadly into two classes: those that require 25 knowledge of the ground-truth positions of fluorophores (e.g.,
26Jaccard index and imaging DNA calibration "rulers") (6-9), 27 and those that operate directly on SMLM reconstructions 28 alone, possibly incorporating information from other measure-29 ments (e.g., diffraction-limited imaging) (10-12). While these 30 methods are able to provide summary or aggregate measures 31 of performan...