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

AllositeUltra identifies dynamic allosteric sites across apo and substrate-bound protein states, reaching Top-3 recovery rates of 65.2% (apo) and 70.6% (substrate-bound) on AsBench2.

• Why it matters: Static holo structures can leave models poorly equipped for cryptic or transient pockets in apo and substrate-bound conformations, limiting computational allosteric site discovery.
• What they did: The work introduces AsBench2 and trains AllositeUltra using multimodal protein representations that combine structure-aware language models with physical dynamics, plus residue-guided candidate supplementation.
• The result: In a prospective test on MAGI2, it predicts an interdomain allosteric pocket later validated by small-molecule screening and X-ray crystallography, and proteome-wide runs produce a high-confidence atlas of targetable sites.

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