#22 · Structural bio & protein design

Learned scores prioritize protein variants, binding pockets and complex interfaces

The dominant thread is learned scoring that ranks candidate residues, variants, pockets and interfaces for downstream follow-up. ProbJanusDDG adds standard and adaptive conformal prediction intervals to a protein-stability predictor to flag mutations with resolved stabilizing or destabilizing effects, while ESMfluc predicts residue-level Protein Block backbone variability from ESM2 and ESM3 embeddings and decomposes attention to give per-position evidence. A structure-based proton-wire model scores 40,166 mtDNA-encoded complex I missense variants, and tumours enrich electrostatic loading-site mutations. For predicted structures, analysis of 237,922 AlphaFold pockets across 12,255 proteins finds mostly holo-like pockets, which bears on pocket detection and docking. The ipSAE extension applies directional PAE scoring to AF3 and Boltz-2 cofolding complexes spanning protein, nucleic acid, ligand and ion interactions. What is new here is calibrated uncertainty, evidence attribution and confidence scoring extended beyond protein-protein contacts. The RFA Hi-C accessibility index serves as a counterexample: it anticorrelates with eQTL through distance-decay but offers no predictive value for ΔRNA.