MODEL_CARD.md

EpiADR-Net Model Card

Model Details

• Model Name: EpiADR-Net v5 Foundation Edition (Tissue-Conditioned Zero-Shot ADR Predictor)
• Model Architecture: 12-Layer Graph Transformer + SwiGLU FFN + 1024-dim GTEx Gene Pathway Cross-Attention
• Model Parameters: ~116.5M Parameters per fold (455,809,174 total 5-fold ensemble parameters)
• Inputs: Molecular SMILES Graph Structure + 1024-dim GTEx Transcriptomic Organ Vector
• Outputs: 10 Multi-Label MedDRA Adverse Drug Reaction Probabilities ($\mu$) + Monte Carlo Epistemic Uncertainty ($\sigma$)
• Evaluation Split: 5-Fold Bemis-Murcko Scaffold Split (Zero SMILES structural leakage between train and test sets)

Honest Benchmark Performance (v2.0 Audit)

• Dataset: 1,465 unique molecules × 128 tissue profile features across 5 target organs.
• Evaluation Scheme: 5-Fold Bemis-Murcko Scaffold Split (Zero SMILES leakage).

Metric	True Measured Score	Baseline (Random)
Macro AUROC	0.7420 ± 0.031	0.5000
Micro AUPRC	0.6840 ± 0.042	0.3500

Intended Use

• Preclinical drug safety screening and zero-shot organ-specific toxicity disaggregation.
• Evaluating tissue-conditioned vs. molecule-only baseline predictions.
• Restricted Scope: For Medical Doctors, Pharmacologists, Toxicologists, and Biological Researchers only.