Data and Code Supporting Leakage-Controlled Multi-Assay Antioxidant QSAR with Class-Conditional Learning and Evaluation of Geometric–Arithmetic Connectivity Indices

This record provides the data, code, and documentation supporting the study “Leakage-Controlled Multi-Assay Antioxidant QSAR with Class-Conditional Learning and Evaluation of Geometric–Arithmetic Connectivity Indices.” The deposited package contains the source workbook, processed compound–assay tables, descriptor-provenance files, executable Python scripts, reproducibility instructions, integrity checksums, and the editable TikZ source and vector output of the model-architecture figure. The dataset comprises 1,059 compound–assay records representing 898 standardized compounds and 330 Bemis–Murcko scaffolds across five antioxidant assay classes: ABTS, DPPH, superoxide, lipid-peroxidation, and hydroxyl-radical assays. Molecular representation includes 102 RDKit descriptors, 14 classical connectivity indices, and 14 geometric–arithmetic connectivity indices, giving a total of 130 imported molecular variables. The code implements fold-contained preprocessing, repeated compound-grouped cross-validation, scaffold-grouped evaluation, class-conditional XGBoost, random-forest and multilayer-perceptron models, known-assay and probability-weighted soft-gating routes, global-regressor and Morgan ECFP4 controls, descriptor-block ablation, response permutation, endpoint-sensitivity analysis, residual and calibration diagnostics, permutation importance, and applicability-domain assessment. The archive is organized as a versioned reproducibility package and includes a data dictionary, software requirements, execution instructions, package manifest, SHA-256 checksums, machine-readable citation information, and file-specific licensing information.