FABind: Fast and Accurate Protein-Ligand Binding

Modeling the interaction between proteins and ligands and accuratelypredicting their binding structures is a critical yet challenging task in drugdiscovery. Recent advancements in deep learning have shown promise inaddressing this challenge, with sampling-based and regression-based methodsemerging as two prominent approaches. However, these methods have notablelimitations. Sampling-based methods often suffer from low efficiency due to theneed for generating multiple candidate structures for selection. On the otherhand, regression-based methods offer fast predictions but may experiencedecreased accuracy. Additionally, the variation in protein sizes often requiresexternal modules for selecting suitable binding pockets, further impactingefficiency. In this work, we propose $\mathbf{FABind}$, an end-to-end modelthat combines pocket prediction and docking to achieve accurate and fastprotein-ligand binding. $\mathbf{FABind}$ incorporates a unique ligand-informedpocket prediction module, which is also leveraged for docking pose estimation.The model further enhances the docking process by incrementally integrating thepredicted pocket to optimize protein-ligand binding, reducing discrepanciesbetween training and inference. Through extensive experiments on benchmarkdatasets, our proposed $\mathbf{FABind}$ demonstrates strong advantages interms of effectiveness and efficiency compared to existing methods. Our code isavailable at https://github.com/QizhiPei/FABind