Chemistry-informed non-covalent graphs and pharmacophore supervision for molecular property prediction

Graph neural networks describe a molecule by its covalent bonds, however, many molecular properties are shaped by contacts between atoms that are not bonded to each other. Existing models either ignore these non-covalent interactions or approximate them with a distance cutoff that cannot distinguish an interaction from mere proximity. Here we introduce a framework that injects chemical knowledge at two levels. First, intramolecular hydrogen bonds and van der Waals contacts become graph edges only when they satisfy interaction-specific topological and geometric criteria; these edges form a separate graph whose layer refines the atom states of a covalent encoder, and the two readouts are fused through a task-aware gate. Second, an auxiliary head supervises every atom with its pharmacophoric roles with expanded σ-hole interaction dictionary. We show that the rule-based graph is about twenty times sparser than a 5-Å cutoff graph, that the gated model reduces to a covalent model as its gate closes and adds a generalisation term bounded by a multiple of its maximal gate value, and that its predictions are Lipschitz in the atomic coordinates as long as the edge set is unchanged. On seven MoleculeNet classification benchmarks under scaffold splitting and without pretraining, ChemPrior-GAT achieves the best ROC-AUC on three benchmarks and the lowest average rank, and the same two modules also improve AttentiveFP, demonstrating that the proposed strategy is transferable to other GNN architecture. Pharmacophore-level attributions are consistent with established chemical principles. In prospective experimental validation involving 16 compounds, the model achieved a mean absolute error (MAE) of 0.50 log units between predicted and experimentally measured logD7.4 values.