Fig. 1. Enriching for evolutionary signals by searching across multiple PLMs.

a, An example of ESM2 and ESM1b complementarity in detecting evolutionary conserved domains as mutationally sensitive; the 20xL heat map of all LLR scores obtained from each model is shown for two proteins (ZFP57 and ITM2B) that are involved in human disease. ESM1b is able to correctly identify the majority of mutations within the annotated KRAB domain in ZFP57 as damaging (yellow color), whereas ESM2 predicts them as neutral (blue color). The opposite is true for the BRICHOS domain in ITM2B. b, An evaluation of the capacity of different ESM models in detecting these domains across multiple proteins (n = 519 and n = 48 proteins with annotated KRAB and BRICHOS domains, respectively). The average score of all possible mutations that fall within each domain is calculated for each protein as a proxy for each model’s ability to detect it. The violin plot shows these average scores across all proteins for each model (median, center line; interquartile range, box; whiskers, 1.5× interquartile range). c, The average LLR per position as predicted by ESM2 and ESM1b is visualized for two other proteins (ZNF93 and PSPC) with experimentally determined structures of the two domains (PDB: 7Z36 and 2YAD). d, An illustration of the maximum confidence approach; signal‑enriched data are generated for all possible mutations by combining predictions from N models. Individual mutations are scored by choosing the most confident prediction (minimum LLR; indicated by ‘min()’) across all models (Methods).