Protein-Ligand Binding Affinity Prediction

The field predicts the strength of protein–ligand binding from molecular structures, protein and ligand sequences, or experimental activity data, helping researchers prioritize compounds for testing. Methods range from docking and physics-based scoring to graph neural networks, geometric interaction models, and pretrained protein or molecule encoders that combine sequence and structural features. Models are commonly assessed on benchmarks such as PDBbind, Davis, and KIBA, using affinity error and ranking metrics, with virtual screening used to test compound prioritization. A central challenge is that strong results on random splits may not transfer to new proteins or chemical scaffolds, and predicted scores do not always capture fine-grained binding energetics.