MolScreen: A Graph Neural Network for Early-Stage Molecular Toxicity Screening with Substructure-Level Explainability and LLM-Generated Research Triage Notes

Pharmaceutical firms lose millions of dollars annually in lab experiments testing compounds later found to be toxic. MolScreen predicts toxicity from chemical structure alone to help prioritize which compounds move to lab testing. Given a molecular structure like those in Tox21, MolScreen runs a graph neural network, trained on 7,823 real compounds, to produce a toxicity risk prediction along 12 biological pathways. It also displays the atoms most important to each prediction and offers a plain-English summary of what the prediction means and what to do next. Across 10 independent training runs, MolScreen achieves a mean ROC- AUC of 0.7716 on 12 toxicity tasks under Bemis–Murcko scaffold splitting, the harder, more realistic protocol for assessing generalization to novel molecule types. The best previously published system on this evaluation reached 0.757, and its authors identified interpretability as an unsolved problem. MolScreen modestly outperforms that result while also addressing the interpretability gap, deployed as a live web application.