Abstract
Accurate predictions and large-scale identification of protein-protein interactions (PPIs) are crucial for understanding their inherent biological mechanisms and protein functions in virtually all biological processes. Nowadays, graph-based deep learning models have made significant contributions in modeling proteins with physicochemical and geometric features. However, most of these models rely on conventional graph construction methods, such as radial cutoff or k-nearest neighbor (k-NN), which often produce sparse and weakly connected graphs, limiting the ability of neural networks to exploit the spatial relationships between nodes. To address this, we introduce PT-PPI, a geometric deep learning framework that combines protein surface point clouds with geometric graphs. Protein surfaces are encoded as oriented point clouds enriched with geometric features, then transformed into sparse, well-connected graphs using the hyperparameter-free Spherical Convex Hull (SCHull) method. These graphs are processed by a Point Transformer network, with representations coupled to ProstT5 sequence embeddings. Evaluations on the PINDER dataset show that PT-PPI surpasses LLM-based (D-SCRIPT), graph-based (GCN, GAT, Struct2Graph), and hybrid sequence-structural-based models (SpatialPPIv2). Ablation studies confirm the complementary value of surface geometry and sequence information, demonstrating that geometric deep learning on protein surfaces and point cloud representations offers a promising approach that opens the doors for further research on large-scale interactome mapping and the understanding of protein function.