Extended Data Fig. 1: Performance of META-SiM in various downstream tasks.

a, The area under the receiver-operating characteristic curve (ROC AUC) for trace classification by META-SiM and DeepFRET compared to manual analysis. Error bars are +/- one standard deviation (s.d.) from 10 times sub-sampling experiments. b, Representative FRET histograms based on traces curated and segmented by META-SiM versus manual analysis. c, A distribution of dwell time predicted by the model is fit with single exponential distributions to yield transition rate constants. d, A representative confusion matrix comparing the labels from manual counting (‘True Label’) to the labels predicted by META-SiM. e, Concordance between manual counting and predictions by META-SiM that either match exactly or differ by no more than one step. f, Standard curve for T790M generated by META-SiM and HMM analysis. g, h, Evaluation of performance in trace idealization on 131 time traces for META-SiM (Fine-Tuned), and benchmarking against 14 other common analysis tools¹³, on the basis of measured rate constants (g) and FRET efficiencies (h): (1) Pomegranate; (2) Tracy(HMM); (3) FRETboard; (4) Hidden-Markury; (5) SMACKS(SS); (6) SMACKS; (7) Correlation; (8) Edge finding(CK); (9) Edge finding(k-means); (10) Step finding; (11) STaSI; (12) MASH-FRET(bootstrap); (13) MASH-FRET(prob); (14) postFRET. Error bars in g and h are +/- one s.d. of fitted rate constants and fitted Gaussian distribution of FRET efficiency, respectively.