LieOTDock: A Protein Docking Framework via Lie Group Optimization and Optimal Transport

Protein-protein interactions (PPIs) are fundamental to cellular function, and structurally characterizing them is a key goal in molecular biology. Computational docking, a crucial tool for predicting the structure of protein complexes, faces significant challenges in search efficiency and scoring accuracy. Here, we introduce LieOTDock, a novel framework that leverages the mathematics of Lie groups and Optimal Transport (OT) for highly efficient protein docking pose generation. The core of our method is a data-driven surface sampling technique that identifies salient geometric features for initial pairing. These initial poses are then refined via a gradient-based optimization directly on the Lie group SE(3), the natural space of rigid-body motions. This optimization is guided by a differentiable score based on Sinkhorn-regularized Optimal Transport, which measures the geometric complementarity of surface patches. We tested LieOTDock on the 1AVW benchmark, where it successfully generated a near-native pose of 3.43 [A] RMSD from a 5000-candidate run in under three minutes on a consumer-grade Apple M4 Mini. Our results demonstrate that LieOTDock, by its principled use of Lie theory and OT, provides a powerful and efficient framework for generating high-quality candidate structures. This work decouples the challenge of geometric sampling from final scoring, offering a robust and modular approach to the protein docking problem, though future work is needed to integrate more physically realistic scoring functions for improved candidate ranking.