AtomSurf-PPI: Protein-Protein Docking with Geometric Deep Learning Representations

Deep learning approaches to protein docking are fast, but do not yet reach the performance of traditional models. However, recent joint modeling of the surface and the graph of a protein enhance protein representations, notably for interaction prediction. In this paper, we show that embeddings learned by such models can efficiently guide classic point cloud alignment procedures and pose scoring models, resulting in a state-of-the-art protein-protein docking system. Specifically, we propose $\textbf{AtomSurf-PPI}$, a multi-stage framework for protein-protein docking that integrates a dual-representation encoder, an enhanced Top-K RANSAC procedure for candidate pose generation, and a Graph Transformer-based scorer for final evaluation. AtomSurf-PPI consistently outperforms other deep learning methods and achieves large speedups over traditional search-based and co-folding methods.