Title: A “grammar” for protein-protein interactions ?
Abstract:
Protein recognition is everywhere in biology, but we still lack simple rules for why one sequence binds another. I will describe an effort to extract such rules from experimental data, which measure how a large designed binder library survives selection against protein targets. Building a deliberately simple energy-based model on top of protein language model embeddings, we find that binding affinity decomposes cleanly into residue-level contributions, recovering known hotspots and showing that binders converge onto a small number of distinct binding modes. Pushing further, we borrow ideas from sparse coding in neuroscience and from mechanistic interpretability of large language models to re-express the same model in a sparse “concept” basis rather than a residue basis, reducing binding to a small set of interpretable features — several of which correspond to known recognition motifs — and searching for these features across millions of natural protein sequences turns up plausible new binders, along with some unexpected biology. I will end with what these rules do and do not seem to generalize to, and ongoing experimental tests.
Bio: https://chemistry.uchicago.edu/suri-vaikuntanathan
Host: Professor Arun Yethiraj and Xuhui Huang