A new model predicts drug toxicity and can explain which part of a molecule triggered the warning.
Researchers built SMILESGNN, which fuses two ways of reading a molecule: a SMILES Transformer that treats chemical structures like text, and a GATv2 graph neural network that treats them like graphs. The two feed into each other through cross-attention rather than being mashed together at the end. On the ClinTox dataset, it hit an AUC-ROC of 0.987 and an F1 score of 0.906 using only 0.4 million parameters, matching or beating larger baselines. A second version, SMILESGNN-PT, swaps in a pretrained ChemBERTa-2 backbone and scored a mean AUC-ROC of 0.750 across the 12 toxicity tasks in Tox21, in line with ChemBERTa-2 alone.
Drug toxicity is one of the biggest reasons candidates die late in development, after a company has already sunk years and money into them. Catching it early with software is not new, but most accurate models are black boxes. This one keeps an explicit graph branch so it can run GNNExplainer analysis and point to the specific substructures associated with a toxic prediction, which is closer to what a chemist or regulator actually needs.
That interpretability claim is the real pitch here, not the accuracy numbers, since plenty of models already hit similar AUC scores. Worth remembering this is a preprint tested on standard benchmarks, not a system validated in an actual drug pipeline.