Discovering Novel Biological Mechanisms from Protein Language Models

A particularly promising direction in interpretability is the possibility that protein language models contain internal representations corresponding to previously unknown biological mechanisms. Rather than using interpretability solely to explain model behavior, this project treats the model itself as a source of scientific hypotheses. By identifying latent features or circuits that strongly influence protein function predictions but do not align with existing biological annotations, we may uncover new motifs, interaction patterns, or evolutionary constraints that have not yet been characterized experimentally. Interventions on these internal representations could then be used to generate novel protein variants and test whether the discovered features correspond to real biological phenomena. More broadly, this approach explores whether foundation models can reveal abstractions about biology that are useful, predictive, and potentially beyond current human scientific understanding.