Learning Constraints in Protein Evolution beyond Known Substitution Frequencies.

Deep learning using an ESM2 encoder predicts the position and nature of amino-acid substitutions in ancestral protein evolution better than statistics-based baselines, with over 37% accuracy for substitution types.

• Why it matters: Conventional substitution frequencies and conservation capture broad patterns but are too general to forecast how specific protein sequences evolve, motivating tests for additional context-dependent predictive information.
• What they did: The study trained deep learning models on an ESM2 encoder to predict substitution rate, position, and type from ancestral sequence context, benchmarking against simpler evolutionary heuristics and outgroup-based evolutionary information.
• The result: Substitution rate is predicted equally well by deep learning and outgroup evolutionary information, while substitution position is predicted more accurately by deep learning; substitution types reach over 37% accuracy overall, are harder in rapidly evolving sequences, favor chemically similar amino acids, and rely mainly on a local 16-residue context.

AI summary of the abstract · may contain errors · not medical advice · read the paper.