Accurate prediction of protein-ligand complexes and binding affinity is critical for hit identification and optimization for structure-based drug design. Traditional docking simulates binding processes with searching algorithms guided by energy-scoring functions, which are quite computationally expensive and time-intensive. In contrast, deep learning approaches offer a cost-effective alternative, yet often generate conformations with limited physicochemical validity and fail to account for protein flexibility. To address these pitfalls, we propose FlowDock, a multitask framework enhanced by Bayesian Flow Networks. FlowDock simultaneously generates accurate protein-ligand complex structures and predicts binding affinity while incorporating protein conformational flexibility. By leveraging multimodal intramolecular representations with a deep equivariant generative model, our method iteratively refines complex in latent space, ensuring rapid and stable generation. Benchmark evaluations demonstrate that FlowDock achieves state-of-the-art performance in binding pose prediction, especially physical plausibility, and virtual screening capability, alongside reliable binding affinity predictions. By providing deeper molecular insights into dynamic protein-ligand interactions, FlowDock represents a robust tool for accelerating the rational development of therapeutics.

Structure-based drug design  /  Protein–ligand docking  /  Binding affinity prediction  /  Machine learning in drug discovery  /  Bayesian flow networks  /  Deep equivariant generative models  /  Latent space optimization