Comprehensive detection and characterization of human druggable pockets through novel binding site descriptors

Druggable pockets are protein regions that have the ability to bind organic small molecules, and their characterization is essential in target-based drug discovery. However, strategies to derive pocket descriptors are scarce and usually exhibit limited applicability. Here, we present PocketVec, a novel approach to generate pocket descriptors for any protein binding site of interest through the inverse virtual screening of lead-like molecules. We assess the performance of our descriptors in a variety of scenarios, showing that it is on par with the best available methodologies, while overcoming some important limitations. In parallel, we systematically search for druggable pockets in the folded human proteome, using experimentally determined protein structures and AlphaFold2 models, identifying over 32,000 binding sites in more than 20,000 protein domains. Finally, we derive PocketVec descriptors for each small molecule binding site and run an all-against-all similarity search, exploring over 1.2 billion pairwise comparisons. We show how PocketVec descriptors facilitate the identification of druggable pocket similarities not revealed by structure- or sequence-based comparisons. Indeed, our analyses unveil dense clusters of similar pockets in distinct proteins for which no inhibitor has yet been crystalized, opening the door to strategies to prioritize the development of chemical probes to cover the druggable space.