DT-Kinase: curated kinase bioactivity dataset with Bemis–Murcko scaffold splits (v1.0.0)

Curated kinase–ligand bioactivity dataset accompanying the PhD thesis "Desenvolvimento de Modelos Atencionais para Triagem Computacional de Quinases Humanas e Não Humanas" (LNCC, 2026) and the attention-screening framework (Zenodo DOI: 10.5281/zenodo.20349742). Derived from ChEMBL release 35. Three corpora are provided: Non-Human (NH), Human (H), and a unified All set. Each (protein, ligand) pair is annotated with UniProt accession, full amino-acid sequence, ChEMBL molecule ID, canonical SMILES (RDKit, isomeric), Bemis–Murcko scaffold SMILES, pChEMBL activity value, binary class label (active if pChEMBL ≥ 6.0; inactive if pChEMBL < 5.0; intermediate values discarded), and partition assignment (train/val/test). Bemis–Murcko scaffold splitting was applied with a monotonic allocation that balances class distribution while keeping each scaffold in exactly one partition. A universal "shared_scaffold" flag is propagated across the three corpora, guaranteeing zero cross-corpus scaffold leakage in cross-domain evaluations. The deposit includes:- kinase_splits/{kinase_all,kinase_human,kinase_non_human}/{train,val,test}.csv — model-ready CSVs in the ConPLex format, consumed directly by the DT-Kinase, DrugBAN, GraphBAN and ConPLex training pipelines.- scaffold_splits/ — raw Bemis–Murcko split tables (TSV) and the "Sc" scenario used in the thesis, plus split-class distribution reports and the universal scaffold manifest (manifest.json, test_scaffolds_universal.json).- README.md, LICENSE and SHA256SUMS.txt for documentation, licensing and integrity verification. Intended uses: training and equitable benchmarking of drug–target interaction (DTI) models on kinase targets; cross-domain generalization studies (NH → H, H → All, etc.); statistical comparison of paired predictions under scaffold-aware splits. License: CC-BY-SA-4.0, inherited from ChEMBL's CC-BY-SA-3.0. Any redistribution or derivative work must attribute ChEMBL (Mendez et al., 2019, Nucleic Acids Research, doi:10.1093/nar/gky1075) and share under the same or a compatible license. Bit-identical regeneration of the splits requires the companion code at https://github.com/gmmsb-lncc/attention-screening (Zenodo: 10.5281/zenodo.20349742), ChEMBL release 35 and RDKit 2024.03.5.