Graph Neural Networks for Kinase–Inhibitor Affinity Prediction Using Docking Scores and Binding-Site Features Download PDF

Kinases are major therapeutic targets, yet accurately predicting true inhibitor binding affinity remains a significant challenge. While molecular docking is useful for generating binding poses, docking scores alone often fail to capture the full complexity of kinase–inhibitor recognition. Current predictive models frequently overlook the spatial graph of the protein–ligand interface and the physicochemical characteristics of the binding site, limiting their effectiveness in compound ranking and reasoning about kinase selectivity. To address these limitations, this article proposes a graph neural network model that integrates kinase–inhibitor complex structures, docking-derived scores, and binding-site descriptors, aiming to provide an interpretable framework for affinity prediction that conceptually surpasses docking-only and ligand-only approaches. The model represents ligand atoms and kinase binding-site residues as graph nodes, with spatial contacts and docking-derived interaction terms encoded as edges, while pocket descriptors are incorporated as residue-level or graph-level features to convey local chemical context. This design enables the model to learn interaction patterns beyond what a single docking score can capture and to identify residues and ligand substructures that contribute most strongly to predicted affinity. By combining structure-based scoring with interpretable deep learning, a docking-informed graph neural network has the potential to advance kinase drug discovery, provided it is rigorously evaluated using kinase-specific benchmarks and prospective validation strategies.