SAPocket: Finding pockets on protein surfaces with a focus towards position and voxel channels

Prediction of protein binding pockets is important for drug discovery and design. Many machine learning-based protein binding pocket prediction algorithms have been developed, however they have several flaws, including low sampling efficiency and inadequate learning of spatial and channel correlations. To address these issues, we develop an attention-based pocket prediction method that works on protein surfaces, namely SAPocket. SAPocket alters the surface sampling approach by moving the sample point along the normal vector. More sample points are focused in the protein cavity as surface sampling technology progresses. Positional information and voxel channel information of voxelized proteins are important for pocket prediction. In order to learn both types of information, we extend SAPocket's network with a dual attention mechanism that is responsive to channel correlations and concentrates on the local space of proteins. SAPocket optimizes the generation of pockets by comprehensively considering the number of points and the average score of the surface point clusters. Experiments show that SAPocket outperforms other comparative methods in predicting the position and shape of pockets.