i-DILI: A Harmonized Multi-Source Database for Drug-Induced Liver Injury (v1.0.0)

i-DILI is an open, harmonized multi-source database for drug-induced liver injury (DILI) labeling and machine-learning benchmarking. Version 1.0.0 contains 17,778 unique compounds with binary DILI labels, a multi-source confidence score (Cf*), canonical SMILES where available, and computed molecular descriptors. Sources and harmonization The build pipeline integrates six public resources (LiverTox, DILIrank, DILIst, DrugBank, Open TG-GATEs, Tox21). SIDER is supported in code but contributes zero rows in this release because its public files use compound IDs rather than drug names. Conflicting labels are resolved by a fixed clinical-priority rule; each compound also receives a confidence score Cf* reflecting source agreement or conflict. Assay-heavy bias is mitigated by liver-relevant Tox21 filtering and DrugBank approved-drug negative controls. Dataset statistics (v1.0.0) 17,778 compounds (22.47% DILI-positive, 77.53% negative) Mean Cf* = 0.603 Valid SMILES: 15,164 compounds (85.3%) Multi-source drugs: 1,833; label conflicts among multi-source entries: 848 (46.3%) Files in this deposit i_dili_final.csv — primary harmonized table (one row per drug) SCHEMA.md — data dictionary and per-source mapping notes reports/ — source overlap, pairwise agreement (Cohen's κ), conflict cases, chemical-space summaries, benchmark metrics, and figure PNGs splits/ — frozen random 5-fold CV indices and Bemis–Murcko scaffold hold-out partition CHECKSUMS.sha256 — SHA-256 integrity manifest Python code — build_idili.py, prepare_datasets.py, analysis/, reproduce.ps1, requirements.txt to rebuild from locally obtained raw sources (not included; see fetch_data.py and README.md) docs/mdpi_template/idili.pdf — accompanying data descriptor manuscript PDF Technical validation Standard fingerprint classifiers (Random Forest, XGBoost, HistGradientBoosting, MLP, DNN) were evaluated under stratified random 5-fold CV and a scaffold hold-out. Results are in reports/benchmark_full.csv. This deposit is intended as a data resource, not a claim of state-of-the-art DILI prediction. How to cite Shende, S., & Tembhurne, J. (2026). i-DILI: A harmonized multi-source database for drug-induced liver injury (Version 1.0.0) [Data set]. Zenodo.