Graph Neural Networks in Molecular Interaction Prediction

Abstract

Graph neural networks encode molecular structures as graphs with atoms as nodes and bonds as edges, learning representations that capture topological, geometric, and chemical features for predicting molecular interactions including protein-ligand binding, drug-drug interactions, and molecular property estimation, yet their adoption into molecular discovery pipelines varies across interaction types and architectural designs. We evaluated 218 GNN-based molecular interaction prediction programmes across centres in Switzerland, Spain, and Italy between 2018 and 2022, spanning five task categories: protein-ligand binding affinity prediction, drug-drug interaction classification, molecular property and ADMET prediction, protein-protein interaction scoring, and reaction outcome prediction. A GNN Molecular Interaction Index (GNNMII) was constructed from five sub-scores -- prediction accuracy on benchmark datasets, out-of-distribution generalisation, model interpretability, computational throughput for screening, and prospective experimental validation - with weights from regression against sustained adoption. GNNMII correlated with adoption at r = +0.84 and discriminated adopted from non-adopted models with an AUC of 0.884. Molecular property prediction scored highest (mean GNNMII 0.826), while reaction outcome prediction trailed at 0.602. Only 35.3 percent exceeded the 0.75 threshold. Prediction accuracy carried the largest weight (beta = +0.280), followed by out-of-distribution generalisation (beta = +0.228)