Sparse Feature Changes Reveal Mechanisms of Missense Variant Effects

Understanding how amino-acid substitutions affect protein function is important to variant interpretation and protein engineering, yet most variant-effect predictors output scalar pathogenicity scores with limited mechanistic explanation. We represent each missense variant by its sparse feature delta, $\Delta F = F(\text{mutant}) - F(\text{wild-type})$, in an interpretable feature basis obtained by applying pretrained sparse autoencoders to ESM-2 representations. Sparse linear probes on these deltas reach an AUROC of 0.892 for ClinVar pathogenicity and a mean Spearman correlation of 0.445 for deep-mutational-scanning fitness in ProteinGym. Predictive signal concentrates in a small feature set that recovers most of the full model's performance, and feature-level ablation and patching materially alter probe predictions. Highly weighted features align with recognizable biophysical patterns, including collagen-glycine disruption and buried-core disruption, and localize to buried residues and annotated functional sites. A dense ESM-2 embedding-difference baseline attains similar predictive accuracy but lacks biologically annotatable components. Sparse feature deltas therefore preserve the predictive information in ESM-2 while transforming mutation effects into biologically legible, mechanistic hypotheses about how sequence changes affect protein function.