Validating and Correcting Graph Neural Network Attention for Drug-Induced Liver Injury Prediction

Attention-based graph neural networks (GNNs) are increasingly used for drug-induced liver injury (DILI) and other toxicity prediction tasks on the claimed strength of built-in interpretability, but this claim is almost always supported by a handful of hand-selected examples rather than being tested systematically. Using a curated, scaffold-split gold-standard DILI data set (1,079 compounds), we show that a graph attention network's attention weights do not, in general, concentrate on seven literature-derived hepatotoxicity structural alerts across the full data set (pooled enrichment ratio of 0.35) and that the network itself does not outperform simple descriptor-based baselines under rigorous repeated-split evaluation (mean AUROC 0.625-0.649 vs. 0.714 for random forest, p = 0.011). We then show both gaps can be addressed: a chemistry-informed multitask extension, trained with an auxiliary structural-alert-recognition loss, improves attention alignment for two alerts by more than 5-fold (both Bonferroni/FDR-corrected p < 10-14), confirmed across five independent scaffold splits, without a robust cost to classification performance. All curated data, trained models, and analysis code are released to support reproducibility and further benchmarking of both predictive and interpretability claims for molecular GNNs.