Fig. 5. The PocketMiner graph neural network accurately detects sites of cryptic pocket formation in experimental structures.

A PocketMiner predicts a high likelihood of cryptic pocket formation at the site of ligand binding. The ligand (cyan) from the aligned holo structure is shown on the apo structure to highlight the steric clash between the ligand and a loop that must move to create the holo binding site. Though PocketMiner only uses the apo structure to generate a prediction (blue indicates low probabilities of cryptic pocket formation while red indicates high probabilities), the predicted labels are also shown on the holo structure to highlight that those high predicted labels cluster near the ligand binding site. B PocketMiner correctly predicts that the probability of pocket formation is low for a highly rigid helical bundle (PDB: 4TQL⁷⁹) that did not form large cryptic pockets in simulation. C Receiver Operating Curve for residue-level cryptic site detection shows that PocketMiner achieves a better performance than CryptoSite despite running >1000x faster. D A precision-recall curve highlights that at high levels of recall (0.6–0.8) PocketMiner predicts fewer false positives. Source data are provided as a Source Data file.