Reliable prediction of protein–ligand binding free energy is central to hit-to-lead optimization because affinity guides prioritization, analogue design, and resource allocation. Classical docking is fast and scalable, but its simplified scoring functions often struggle across chemically diverse ligand series. Purely data-driven models can learn dataset-specific chemical patterns that may not reflect realistic molecular recognition. Conversely, purely physics-based approaches can miss high-dimensional statistical regularities present in curated structural and affinity datasets. This article proposes a physics-informed graph neural network for predicting absolute binding free energy from protein–ligand complex structures. The model integrates docking scores, molecular-dynamics-derived interaction descriptors, and force-field-inspired energy terms within a unified graph representation. The proposed architecture uses three-dimensional, SE(3)-aware message passing over protein–ligand complex graphs. Atom and residue features are augmented with docking-derived pose information, molecular dynamics descriptors, and pre-computed electrostatic and van der Waals interaction terms. Conceptually, such a model would be expected to provide more physically consistent affinity estimates than purely empirical scoring functions. It could also support interpretable atom-level and residue-level interaction analysis for medicinal chemistry decision-making. Physics-informed graph learning offers a principled bridge between rigorous molecular thermodynamics and flexible deep representation learning. This framework provides a scalable model-oriented route toward more reliable in silico affinity optimization.