A cryptic pocket is a druggable site that is absent from a protein's resting structure and appears only once the backbone moves. It is the reason a target dismissed as undruggable from its crystal structure can still turn out to be druggable, and it is exactly what any method that reads a single static conformation will miss. The problem this paper takes on is finding these sites from an apo structure alone, on any protein rather than only the handful that come with a solved open-state or ligand-bound companion. Every established way of detecting a site that is visible only in the open state builds that state first, whether by sampling conformations or by predicting a bound complex. We do not build it. Our world model of molecular organization has already learned it and carries it in its latent, so we read cryptic pockets straight from that latent, per residue, from a single static apo structure, with no conformational sampling and no protein language model. The representation is never chained onto the output of a separate pocket finder, and a candidate pocket is a grown, disjoint set of residues rather than a ball of fixed radius. On CryptoBench, a set of 231 held-out apo chains, the detector localizes the cryptic site at top-1 0.848, top-3 0.931, and top-5 0.952 under a pocket-level overlap rule, using only the apo structure; on a 286-chain subset of CryptoBank it reads 0.846, 0.933, and 0.990. Scored residue by residue on CryptoBench, it reaches an AUC of 0.8465. On WRN helicase, held out of training entirely, it recovers at rank 1 the allosteric pocket bound by the clinical inhibitor HRO761, across three independent apo crystal forms and from the apo backbone alone; on the best-resolved form the rank-1 pocket covers 18 of the 22 annotated residues, including the C727 the drug contacts. Set against OpenDDE, the detector is complementary: on 190 targets held out of both training sets it recovers 119 sites OpenDDE misses, against 2 in the other direction, and at top-5 it misses none of the sites OpenDDE finds. It runs on any structure from the apo coordinates alone, including the mmCIF-only entries all recent depositions carry, and its accuracy keeps climbing as the training corpus grows. The intended use is prospective cryptic-site nomination on the targets that sequence and static structure leave without a starting point.
Shihabi, R., Taraman, S., Vaughan, B.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 0
- Comments 0