Fig. 3: ZeroBind is able to detect binding pockets of proteins in a weakly supervised way.

a The distribution of Jaccard similarity coefficients of the predicted binding pockets with the true binding pockets and randomly selected binding residues as binding pockets with the true binding pockets of individual DTIs. b The distribution of Jaccard similarity coefficients of the predicted binding pockets with the first-order neighbors of true binding pockets and randomly selected binding residues as binding pockets with the first-order neighbors of true binding pockets of individual DTIs. c The Serine/threonine-protein kinase N1 protein. Experimentally validated DTI binding pocket queried by BioLiP. The blue part represents the experimentally validated binding pocket along with the residue name and number. d The potential binding pocket predicted by IB-subgraph in ZeroBind. The red part represents the helix structure of protein, the yellow part represents the loop structure of protein and the green part represents the sheet structure of protein. Source data are provided as a Source Data file.