Accurate protein–ligand binding poses are crucial for structure-based drug discovery, as downstream interpretation relies on plausible atomic contacts; while docking workflows can generate useful hypotheses, many poses remain geometrically imperfect and require refinement for confident use. Current post-docking refinement tools often rely on either computationally expensive physics-based relaxation or learned scoring models that fail to explicitly preserve three-dimensional molecular symmetry, limiting their reliability when input poses are rotated, translated, or structurally distinct from training examples. To address this, we propose an equivariant neural network that refines docked protein–ligand poses directly from atomic coordinates and interaction energies, ensuring that coordinate corrections respect rotational and translational symmetry while remaining informed by local protein–ligand physics. The protein–ligand complex is represented as a three-dimensional graph, with atoms as nodes and spatial contacts as edges; node features capture atomic identity and chemistry, while edge features encode distances, directions, and interaction energies such as electrostatics, van der Waals forces, and hydrogen bonds. Conceptually, this model is expected to guide docked ligands toward more physically plausible binding geometries without relying on fixed molecular orientations and can provide an interpretable quality score indicating whether the refined pose is suitable for downstream design. By combining symmetry-preserving architecture with energy-aware refinement, equivariant geometric deep learning offers a principled approach bridging rapid docking and computationally intensive physical methods, positioning this model as a valuable post-processing tool in modern drug discovery pipelines.