Accurate Identification of Functional Residues Across the Human Proteome with TAMALE

The central challenge of the post-AlphaFold era is the ‘functional gap’. Despite having structure predictions for nearly every human protein, we remain unable to systematically distinguish residues that drive activity from those that merely maintain structural integrity. Here we introduce TAMALE, a machine learning model that calculates graded residue-level functional scores across the human proteome without prior annotation. TAMALE transforms structure models and variant effect predictions to identify residues involved in catalysis, ligand binding, nucleic acid interactions, and regulation. The model also distinguishes pseudo-enzymes from catalytically active homologs. The model was validated across 20 case studies along with experimental characterization of the FASTKD5 ribonuclease, demonstrating its utility for functional discovery. Applied proteome-wide to 19,528 human proteins, TAMALE generates testable hypotheses enabling mechanistic discovery at scale.