In the field of structure-based drug design, accurately predicting the binding conformation of ligands to proteins is a long-standing objective. Despite recent advances in deep learning yielding various methods for predicting protein-ligand complex structures, these AI-driven approaches frequently fall short of traditional docking methods in practice and often yield structures that lack physical and chemical plausibility. To overcome these limitations, we present SurfDock, an advanced geometric diffusion network, distinguished by its ability to integrate multiple protein representations including protein sequence, three-dimensional structural graphs, and surface-level details into its equivariant architecture. SurfDock employs a generative diffusion model on a non-Euclidean manifold, enabling precise optimization of molecular translations, rotations, and torsions for reliable binding poses generation. Complemented by a mixture density network for scoring using the same comprehensive representation, SurfDock achieves significantly improved docking success rates over all existing methods, excelling in both accuracy and adherence to physical constraints. Equipped with post-docking energy minimization as an optional feature, the plausibility of generated poses is further enhanced. Importantly, SurfDock demonstrates excellent generalizability to unseen proteins and extensibility to virtual screening tasks with state-of-the-art performance. We consider it a transformative contribution that could serve as an invaluable asset in structure-based drug design.