Cross-state representation learning enables prospective discovery of dynamic protein allosteric sites

Abstract

Allosteric regulation offers multifunctional induction in therapeutic interventions, yet conformational plasticity hinders computational identification of allosteric sites. Machine learning models rely on static holo (ligand-bound) structures, generalizing poorly to apo or substrate-bound states where binding pockets are transient or cryptic. We present AsBench2, a cross-state site-recognition benchmark, and AllositeUltra, a deep learning framework integrating multimodal protein representations (combining structure-aware language models with physical dynamics) and residue-guided candidate supplementation, effectively decoupling site detection from static pocket geometry. AllositeUltra demonstrates superior generalization, achieving Top-3 site recovery rates of 65.2% and 70.6% on apo and substrate-bound structures, respectively, representing relative improvements of 36.4% and approximately 50% over the best-performing baselines. Prospective application to MAGI2 identifies a previously unrecognized interdomain allosteric pocket, validated at atomic resolution through small-molecule screening and X-ray crystallography. Proteome-wide application yields a high-confidence atlas of targetable allosteric sites, bridging the gap between structural representation learning and prospective AI-driven drug discovery.