HiKG-DDI: Preprocessed feature matrices and benchmark splits (v1.0)

Abstract

Preprocessed feature matrices and train/val/test splits used to train and evaluate HiKG-DDI, a multi-modal drug-drug interaction (DDI) prediction model. Includes Morgan/ECFP4 fingerprints, RDKit 2D descriptors, four Jaccard+PCA biological similarity features (protein targets, metabolic enzymes, pathways, substructure indices), and 2-hop GraphSAGE-style HetioNet knowledge-graph embeddings for 1,731 drugs. Ready-to-use splits for DDI-Bench DrugBank (S0 transductive, S1/S2 cold-start), TWOSIDES (209 side effects), and DDIMDL (5-fold CV). Derived from public data sources under compatible licenses: DrugBank 5.x (CC BY-NC 4.0), TWOSIDES (Tatonetti et al. 2012, freely redistributable), HetioNet v1.0 (CC0), DDIMDL (Deng et al. 2020), and DDI-Bench (Zhang et al. 2023, MIT). See README inside the bundle for full attribution and citation details. Associated code: https://github.com/Omarmahmoud711/Khemical