DynamicBind is a deep learning method that uses an equivariant geometric diffusion network to predict ligand-specific protein-ligand complex structures. Unlike traditional docking methods that treat proteins as rigid, DynamicBind efficiently adjusts the protein conformation from its initial AlphaFold prediction to a holo-like state. It accurately recovers ligand-specific conformations from unbound protein structures without requiring halo-structures or extensive sampling. DynamicBind demonstrates state-of-the-art performance in docking and virtual screening benchmarks, capable of accommodating large conformational changes and identifying cryptic pockets in unseen protein targets. It shows potential in accelerating the development of small molecules for previously undruggable targets and expanding computational drug discovery. DynamicBind is a geometric deep generative model designed for "dynamic docking." It accepts apo-like structures (AlphaFold-predicted conformations) and small-molecule ligands in various formats. During inference, the model randomly places the ligand around the protein and gradually translates and rotates it while adjusting internal torsional angles. The model uses an SE(3)-equivariant interaction module to generate predicted translation, rotation, and dihedral updates. It employs a morph-like transformation for decoy generation during training, allowing it to learn a funneled energy landscape that minimizes frustration between biologically relevant states. DynamicBind outperforms other methods in predicting ligand poses for the PDBbind and MDT datasets. It achieves higher accuracy in ligand pose prediction and improves the initial AlphaFold-predicted protein conformations. DynamicBind can capture ligand-specific protein conformational changes, such as the DFG-in to DFG-out transition in kinase proteins. It reveals cryptic pockets significant to drug discovery, such as the cryptic pocket targeted by EZM0414 in the SETD2 protein. DynamicBind also achieves better screening performance in an antibiotics benchmark, surpassing traditional docking methods and machine learning-based re-scoring methods. DynamicBind unifies protein conformation generation and ligand pose prediction into a single framework, making it significantly faster than traditional MD simulations. It can perform global docking, essential when the binding pocket is unknown. DynamicBind's ability to predict ligand-specific conformations may offer insights into the influence of ligands on proteins, clarifying structure-function relationships and enhancing mechanistic understanding. DynamicBind's performance in benchmarks indicates its potential for proteome-level virtual screening applications. The model's ability to generalize to new proteins and ligands is promising, and it benefits from advancements in Cryo-EM methods. DynamicBind's capacity to predict binding affinities and its efficiency in sampling conformational changes make it a valuable tool in drug discovery.DynamicBind is a deep learning method that uses an equivariant geometric diffusion network to predict ligand-specific protein-ligand complex structures. Unlike traditional docking methods that treat proteins as rigid, DynamicBind efficiently adjusts the protein conformation from its initial AlphaFold prediction to a holo-like state. It accurately recovers ligand-specific conformations from unbound protein structures without requiring halo-structures or extensive sampling. DynamicBind demonstrates state-of-the-art performance in docking and virtual screening benchmarks, capable of accommodating large conformational changes and identifying cryptic pockets in unseen protein targets. It shows potential in accelerating the development of small molecules for previously undruggable targets and expanding computational drug discovery. DynamicBind is a geometric deep generative model designed for "dynamic docking." It accepts apo-like structures (AlphaFold-predicted conformations) and small-molecule ligands in various formats. During inference, the model randomly places the ligand around the protein and gradually translates and rotates it while adjusting internal torsional angles. The model uses an SE(3)-equivariant interaction module to generate predicted translation, rotation, and dihedral updates. It employs a morph-like transformation for decoy generation during training, allowing it to learn a funneled energy landscape that minimizes frustration between biologically relevant states. DynamicBind outperforms other methods in predicting ligand poses for the PDBbind and MDT datasets. It achieves higher accuracy in ligand pose prediction and improves the initial AlphaFold-predicted protein conformations. DynamicBind can capture ligand-specific protein conformational changes, such as the DFG-in to DFG-out transition in kinase proteins. It reveals cryptic pockets significant to drug discovery, such as the cryptic pocket targeted by EZM0414 in the SETD2 protein. DynamicBind also achieves better screening performance in an antibiotics benchmark, surpassing traditional docking methods and machine learning-based re-scoring methods. DynamicBind unifies protein conformation generation and ligand pose prediction into a single framework, making it significantly faster than traditional MD simulations. It can perform global docking, essential when the binding pocket is unknown. DynamicBind's ability to predict ligand-specific conformations may offer insights into the influence of ligands on proteins, clarifying structure-function relationships and enhancing mechanistic understanding. DynamicBind's performance in benchmarks indicates its potential for proteome-level virtual screening applications. The model's ability to generalize to new proteins and ligands is promising, and it benefits from advancements in Cryo-EM methods. DynamicBind's capacity to predict binding affinities and its efficiency in sampling conformational changes make it a valuable tool in drug discovery.