The abstract is read at the publisher; the summary is JClub's.
PocketVec-based pocket similarity clustering selected 200 ligand-free representative pockets spanning 100 clusters per strategy, covering over 10,000 pockets in ~5,600 proteins (~36% of annotated druggable proteins).
- Why it matters: Large, unbiased protein–ligand datasets are scarce, limiting machine learning in drug discovery, especially for understudied human proteins. Maximizing diversity of selected targets requires a criterion that captures shared binding-site chemistry rather than sequence or fold.
- What they did: The authors mapped the human globular pocketome using PocketVec, predicting 51,535 druggable pockets across 15,559 proteins, then clustered them into 1,865 chemically similar binding-site groups. After removing 533 clusters with known ligands, they selected the 100 most populated clusters per prioritization strategy and chose two representative pockets per cluster.
- The result: The two complementary selections overlapped in 89 clusters and together cover over 10,000 pockets in ~5,600 proteins, about 36% of human proteins with annotated druggable pockets. Compared with random baselines, this yields a 2.8- to 4.4-fold increase in covered pockets, domains, and proteins, supporting Target 2035 target selection.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.