SMILESGNN Achieves 0.987 AUC-ROC for Toxicity Prediction
September 25, 2026
SMILESGNN is a multimodal architecture that fuses SMILES Transformer encoders with GATv2 graph encoders via cross-attention for clinical toxicity prediction. With only 0.4M parameters, it achieves an AUC-ROC of 0.987 on ClinTox while maintaining interpretability through GNNExplainer-based substructure analysis.
HOW THIS AFFECTS YOU
●
researcherYou can combine sequence and graph modalities with a very low parameter count for competitive performance.
●
healthThis offers a highly efficient, interpretable tool for reducing late-stage attrition in drug discovery.