Dataset and Pre-Trained Weights for MoDAN: An Interpretable Modality-Disentangled Attention Network for Zero-Shot DDI Prediction

This archive contains the reproducibility materials for the manuscript: "MoDAN: An Interpretable Modality-Disentangled Attention Network for Zero-Shot Drug-Drug Interaction Prediction" (Under Review, RSC Digital Discovery). Contents of this archive: Model Weights: The pre-trained PyTorch weights (best_biomodal_model.pt) for MoDAN, trained on 2.47 million canonical interactions. Embeddings: The extracted multimodal feature dictionaries, including ChemBERTa (1D chemical structures), ESM-2 (protein sequences), and BioBERT (multi-hot biological pathways). Evaluation Splits: The strict inductive (S2) cold-start evaluation CSV files for our custom DrugBank v5.1.13 dataset, as well as the strictly isolated 100% blind subsets for the Stanford BIOSNAP and ZhangDDI public benchmarks. Note: Due to DrugBank licensing restrictions, the raw DrugBank XML database is not included in this repository. Researchers must obtain an academic licence directly from DrugBank to run the initial dataset extraction scripts available on our GitHub.