Pocket similarity guided selection of protein targets for systematic protein ligand dataset generation

The scarcity of large, diverse, and unbiased protein ligand binding datasets limits the performance of machine learning approaches in drug discovery, particularly for understudied proteins. The Target 2035 initiative aims to address this gap by generating open-access binding data for more than 2,000 human proteins, a goal whose value depends on maximizing the diversity of the selected targets. Because proteins with different folds can share chemically analogous binding sites, we reasoned that pocket similarity, rather than sequence or fold similarity, is the appropriate criterion for selecting representative targets that maximize coverage of the human pocketome. Here, we use PocketVec, an alignment-free descriptor of binding sites, to map the human globular pocketome, identifying 51,535 predicted druggable pockets across 15,559 proteins. A two level hierarchical clustering, first within protein families and then across the whole proteome, grouped these pockets into 1,865 clusters of chemically similar binding sites. After discarding 533 clusters containing pockets with known ligands, we selected the 100 most populated of the remaining 1,332 ligand free clusters and chose two high quality representative pockets per cluster, ranked by their resemblance to known ligand-binding sites. Finally, we implemented two complementary prioritization strategies, one favoring structurally unexplored proteins and the other proteins amenable to experimental follow up, and selected 200 pockets from 100 different clusters from each strategy, with 89 clusters in common. Both selections cover over 10,000 pockets in ~5,600 proteins, approximately 36% of the human proteins with annotated druggable pockets, which represents a 2.8- to 4.4-fold increase in the number of covered pockets, domains, and proteins than random selection baselines. This pocket based framework offers a principled strategy to select diverse targets for Target 2035 and other large scale protein ligand data generation efforts in chemical biology.