Directed Message Passing with Global Descriptors for Toxicity Prediction on Tox21

This work presents a reproducible molecular toxicity prediction model evaluated on the original Tox21 benchmark using the restored 2014–2015 challenge protocol. The approach is based on a Directed Message Passing Neural Network (D-MPNN) augmented with global physicochemical descriptors to capture both local substructure information and global molecular properties.Model development employs scaffold-based data splitting to promote generalization to unseen chemical scaffolds and to prevent test leakage. Class imbalance and missing labels are handled explicitly during training, and robustness is improved through ensemble averaging. Final evaluation is performed exclusively via a remote, held-out test server hosted on Hugging Face, ensuring strict separation between model development and test assessment.The accompanying report documents the full preprocessing pipeline, training procedure, evaluation protocol, and reproducibility artifacts required for leaderboard verification. This submission is intended as a transparent and verifiable contribution to the Tox21 leaderboard and as a reference implementation for rigorous toxicity prediction benchmarking.