DynamicBind is a deep learning method designed to predict ligand-specific protein-ligand complex structures by employing equivariant geometric diffusion networks. It addresses the limitations of traditional docking methods, which often treat proteins as rigid, and molecular dynamics simulations, which are computationally expensive. DynamicBind constructs a smooth energy landscape, enabling efficient transitions between different equilibrium states. The method accurately recovers ligand-specific conformations from unbound protein structures without the need for holo-structures or extensive sampling. It demonstrates state-of-the-art performance in docking and virtual screening benchmarks, showing the ability to handle large conformational changes and identify cryptic pockets in unseen protein targets. DynamicBind's dynamic docking capability allows it to refine AlphaFold-predicted structures towards more native-like states, improving binding affinity predictions. The method also excels in virtual screening, achieving better performance than traditional docking methods and machine learning-based re-scoring methods in an antibiotics benchmark. Overall, DynamicBind offers significant potential for accelerating the development of small molecules for previously undruggable targets and expanding computational drug discovery.DynamicBind is a deep learning method designed to predict ligand-specific protein-ligand complex structures by employing equivariant geometric diffusion networks. It addresses the limitations of traditional docking methods, which often treat proteins as rigid, and molecular dynamics simulations, which are computationally expensive. DynamicBind constructs a smooth energy landscape, enabling efficient transitions between different equilibrium states. The method accurately recovers ligand-specific conformations from unbound protein structures without the need for holo-structures or extensive sampling. It demonstrates state-of-the-art performance in docking and virtual screening benchmarks, showing the ability to handle large conformational changes and identify cryptic pockets in unseen protein targets. DynamicBind's dynamic docking capability allows it to refine AlphaFold-predicted structures towards more native-like states, improving binding affinity predictions. The method also excels in virtual screening, achieving better performance than traditional docking methods and machine learning-based re-scoring methods in an antibiotics benchmark. Overall, DynamicBind offers significant potential for accelerating the development of small molecules for previously undruggable targets and expanding computational drug discovery.