Positive-unlabeled learning predicts which protein pockets are druggable without relying on unreliable negative labels
- PocketBagger is a machine-learning model that predicts whether a protein pocket can be drugged.
- It trains on pockets known to bind drug-like ligands as positives and treats all other pockets as unlabeled, avoiding unreliable negative labels.
- The stated aim is druggability prediction that generalizes beyond the training pockets; details of benchmarking against existing tools were not available in the source summary.
💡 Points Based on the abstract
Is "unlabeled" really neutral? Many pockets without a known ligand are simply unexplored, so positive-unlabeled training is a sensible fix for false negatives. But generalization claims hinge on the split: if homologous proteins sit on both sides, the score flatters. A fairer test is to run it on your own target class and on pockets that were drugged later.
PocketBagger: generalizable pocket druggability prediction via positive-unlabeled learning ↗