Sequence-Based Protein Pocket Prediction via Prott5 Embeddings and Spatial Sampling

Protein pocket prediction is critical for elucidating protein-ligand interactions, with substantial implications in drug discovery. In this study, we used ProtT5 embeddings to represent protein pockets derived from protein-ligand complexes in the PDBBind database. Given that PDBBind provides exclusively positive (pocket) sequences, we developed a spatial sampling strategy to generate meaningful counter (non-pocket) examples. Classification performance was evaluated using Random Forest (RF) and Support Vector Machine (SVM), achieving accuracies up to 0.90. Overall, the SVM classifier consistently demonstrated superior performance, highlighting its greater effectiveness compared to RF. We further extended the classifier into a full pocket detection pipeline, enabling the identification of putative pockets directly from protein structures. A case study on the c-Src kinase (PDB entry 3G5D) demonstrated the applicability of this approach in mapping binding sites onto sequences, facilitating structure-based drug design. Our approach thus provides a valuable tool for drug discovery, particularly in the identification of novel and allosteric binding sites and guiding targeted therapeutic interventions.