Rule-based non-covalent graphs and pharmacophore supervision improve GNN property prediction
- A graph neural network adds a second graph of hydrogen-bond and van der Waals contacts, kept only when chemical geometry rules are met, and trains with an auxiliary atom-level pharmacophore task.
- On several MoleculeNet classification sets with scaffold splits and no pretraining, it was competitive with or better than baselines, and the modules also helped another GNN.
- A small prospective logD check on newly measured compounds showed moderate error. This is a chemRxiv preprint, not peer reviewed.
💡 Points Based on the abstract
Scaffold splits are the right choice, but I'd want the gain checked against a plain GNN with matched capacity and equal tuning, since a second graph and an auxiliary head add parameters. The pharmacophore labels are also rule-derived, so the model may just be relearning the rules. Try it on your own series, and judge the logD check against a simple baseline.