COMPUTATIONAL PREDICTION OF MOLECULAR TOXICITY USING MULTI-ENDPOINT RANDOM FORESTS AND GRAPH NEURAL NETWORKS

Accurate prediction of molecular toxicity is critical for early-stage drug discovery and chemical risk assessment.In this study, three artificial intelligence-based models were developed using the Tox21 dataset: (i) a descriptor-based Random Forest for global toxicity screening, (ii) a multiendpoint Random Forest for pathway-specific prediction, and (iii) a graph neural network that learns toxicity directly from molecular structure.The baseline model achieved a ROC-AUC of 0.75, while endpoint-specific learning improved performance to a mean ROC-AUC of 0.819.The graph-based model achieved a competitive ROC-AUC of 0.789 without relying on handcrafted descriptors.These results demonstrate that endpoint-aware ensemble learning and structurebased deep learning provide robust and scalable solutions for predictive toxicology.