One-shot Evaluation of Protein Mutability and Epistasis Score Using Structure-Based Model ESM3

Abstract

Background

Existing mutation scoring methods, often designed for predicting genetic diseases in the highly conserved human genome, lack generalizability across diverse protein fold types. Conventional alignment-based mutation scoring protocols lack causal explanations for conservation patterns, such as intra-chain and inter-chain interactions, due to insufficient structural awareness. Here, we leveraged the generative capabilities of the structure-based transformer model ESM3 to construct a UniProtKB-scale database of amino acid mutation and epistasis scores.

Methods

Mutation scores (M-scores) were calculated as average differences between mutant and wild-type embeddings. In contrast, epistasis scores (E-scores) were derived from variance among mutant embeddings, mathematically representing mutational variability due to epistatic interactions. To facilitate score calculations using Euclidean distance, the original 1536-dimensional ESM3 embeddings were compared to 312 dimensions.

Significant Findings

Our approach successfully captured protein mutability patterns beyond peptide flexibility and phylogenetic analysis. Benchmarking against reported in silico protein evolution datasets showed a significant correlation with mega-scale Bayesian optimization experimental results. M-score and E-score profiles exhibited intra-family variations compared to root mean square fluctuation (RMSF) profiles, revealing evolutionary constraints beyond kinetic energy and family-level phylogeny insights. This divergence proved functionally meaningful: in low-RMSF regions, gap formation between homologs could be explained by strong regional mutability and epistasis, indicated by high M- and E-scores. Overall, M-scores and E-scores serve as interpretable, target-agnostic metrics for structurally informed, epistasis-aware mutational analysis, advancing protein function and evolution understanding.