MolGraphormer: An Interpretable GNN-Transformer For Uncertainty-Aware Molecular Toxicity Prediction

Authors: Akshay Balaji

Abstract: Accurate and Interpretable toxicity prediction re- mains fundamental in computational chemistry and drug discov- ery. We propose MolGraphormer, a Transformer-GNN hybrid architecture integrating Graph Neural Network message passing with self-attention mechanisms for molecular property prediction. Our model incorporates substructure-aware embeddings via multi-head attention, edge-conditioned message passing, and hierarchical graph aggregation, enabling both local and global molecular reasoning. Evaluated on the Tox21 benchmark dataset, MolGraphormer achieves competitive performance with F1-Score of 0.6697 and AUC-ROC of 0.7806, while maintaining strong recall (0.7787) for identifying toxic compounds. We employ Monte Carlo Dropout and Temperature Scaling for uncertainty quantification, Combined with uncertainty quantification and attention-based interpretability, MolGraphormer offers a practical framework for drug safety assessment and regulatory toxicology.