Model Interpretability Discovery Team

Protein language models (PLMs) can identify pathogenic missense variants, but it is unclear how they internally make the predictions. Inspired by Anthropic’s discovery of the Jacobian lens (J-lens) in large language models, I hypothesized that a similar mechanism exists in PLMs. To test this, I used ESM1b, a PLM that reveals disease variant effects (Brandes et al., Nature Genetics, 2023). Using Claude Science, I adapted the J-lens to ESM1b’s bidirectional masked encoder architecture. I fit a local masked-site J-space on 1,000 proteins that were family- and homology-disjoint from the evaluation set. For benchmarking, I used the same ClinVar dataset from Brandes et al., containing clinically annotated benign and pathogenic variants. Compared with geometric-matched random controls, deleting a small number of J-space components degraded the variant effect prediction performance further, and the J-space better approximated hidden state gradients specific to each variant. Its contribution to variant effect prediction was comparable to that of the model’s native output head. I replicated this result across three ESM1v model variants. Overall, my project revealed how PLMs harbor a sparse, experimentally controllable internal representation linked to variant effect prediction. Intervening in this internal representation could improve variant effect prediction and elucidate the mechanisms of PLMs.