Fig. 5. VESM achieves state-of-the-art VEP performance in both clinical and DMS benchmarks.

a, A comparison of model performance on clinical VEP (ClinVar AUC, same dataset as in Fig. 4a) versus experimental VEP from human protein DMS assays (Spearman correlation, ProteinGym fitness and activity assays). Points denote individual models; the dashed gray line shows the fitted linear regression. Shaded bands (95% CIs) around the regression indicate uncertainty in the fit. The models are color-coded based on the additional sources of information they use. b, Performance on the ProteinGym DMS benchmark (rank score), with assays categorized by selection type: fitness/activity (x axis) versus binding, stability and expression (y axis). VESM3 and VESM++ achieve state-of-the-art performance across both assay types, surpassing existing structure- and MSA-based methods. c, Pairwise win rates between models across 120 DMS fitness and activity assays, quantifying the proportion of assays in which a given model outperforms another. d, A comparison of VESM models and ESM baselines across nonhuman DMS assays, grouped by taxonomic category (eukaryote, prokaryote and virus). Base models for individual VESM models: ESM2 (3B), ESM2 (650M), ESM2 (150M), ESM2 (35M) and ESM3_struct. For VESM++ the baseline is given by the ensemble of the corresponding base models ESM2 (3B) and ESM3_struct.