InterPLM: discovering interpretable features in protein language models via sparse autoencoders

Despite their success in protein modeling and design, the internal mechanisms of protein language models (PLMs) are poorly understood. Here we present a systematic framework to extract and analyze interpretable features from PLMs using sparse autoencoders. Training sparse autoencoders on ESM-2 embeddings, we identify thousands of interpretable features highlighting biological concepts including binding sites, structural motifs and functional domains. Individual neurons show considerably less conceptual alignment, suggesting PLMs store concepts in superposition. This superposition persists across model scales and larger PLMs capture more interpretable concepts. Beyond known annotations, ESM-2 learns coherent patterns across evolutionarily distinct protein families. To systematically analyze these numerous features, we developed an automated interpretation approach using large language models for feature description and validation. As practical applications, these features can accurately identify missing database annotations and enable targeted steering of sequence generation. Our results show PLM representations can be decomposed into interpretable components, demonstrating the feasibility and utility of mechanistically interpreting these models.