Abstract
Predicting small molecule binding sites on proteins remains a key challenge in structure-based drug discovery. While AlphaFold3 has transformed protein structure prediction, accurate identification of functional sites such as ligand binding pockets remains a distinct and unresolved problem. Graph neural networks have emerged as promising tools for this task, but most current approaches focus on local structural features and are trained on relatively small datasets, limiting their ability to model long-range protein-ligand interactions. Here, we develop YuelPocket, a graph neural network that addresses these limitations. YuelPocket operates in two complementary modes: residue-level prediction for identifying contact residues and coordinate-level prediction for pinpointing pocket centers. Trained on the large-scale PLINDER dataset, YuelPocket achieves higher success rates in both Distance to Closest Atom and Center-to-Center metrics compared to the state-of-the-art methods. Crucially, YuelPocket demonstrates high robustness on AlphaFold-predicted structures, maintaining high accuracy for targets with deviations from experimental structures. We hope that YuelPocket will serve as a robust framework for accurate binding site identification, enabling reliable functional annotation and structure-guided drug discovery.