EC-Dock: A Fast Equivariant Consistency Model for Molecular Docking and Virtual Screening

Molecular docking plays a crucial role in structure-based drug design. Although diffusion-based docking models have achieved remarkable success in this field, their practical application in real-world drug discovery remains limited by challenges in computational efficiency and physical validity. To address these issues, we propose EC-Dock, a fast diffusion-based docking model built upon an equivariant consistency framework, with ligand-protein atomic distance matrices serving as structural constraints. Leveraging SE(3)-equivariant graph neural networks and a consistency model, EC-Dock enables one-step generation of accurate and physically plausible docking poses. EC-Dock achieves state-of-the-art performance in docking accuracy and geometric stability, while drastically reducing inference time and largely avoiding structural distortions. On the widely adopted PoseBusters benchmark, EC-Dock attains a Top-1 success rate (RMSD < 2 Å) of 83% and a Best-1 success rate of 93%. On the TrueDecoy virtual screening benchmark, EC-Dock demonstrates competitive state-of-the-art performance, achieving the highest enrichment factor within the top 0.5% of selected molecules compared to other docking models. These results highlight the potential of EC-Dock as a rapid, robust, and scalable solution for structure-based drug discovery.