Multi-label GHS hazard classification dataset and trained models for 243,323 chemical compounds

This record contains the curated dataset, computed molecular descriptors andtrained models supporting the study "Interpretable Machine Learning forPredicting GHS Chemical Hazard Classifications". GHS hazard classifications were harvested from PubChem for 243,323 uniquechemical compounds and reconciled across five independent regulatory sources:the European Chemicals Agency, Regulation (EC) No 1272/2008, the HazardousSubstances Data Bank, NITE-CMC, and the Hazardous Chemical Information Systemof Safe Work Australia. Where sources disagreed, labels were resolved bymajority vote. Each compound is described by 1,218 moleculardescriptors combining physicochemical properties, Morgan (ECFP4) and MACCSfingerprints and topological indices, reduced to816 after variance filtering. Models were evaluated on a Bemis-Murcko scaffold split(194,619 training / 24,352 validation /24,352 test) in which no chemical skeleton is shared betweenpartitions. XGBoost performed best, with a mean AUC-ROC of 0.915 across thenine GHS hazard classes. CONTENTS- Raw and cleaned datasets with SMILES, InChIKey, CAS and the nine binary hazard labels- The molecular descriptor matrix and aligned label matrix- Trained Random Forest, XGBoost and support vector machine models- Full evaluation results with bootstrap confidence intervals- SHAP interpretability tables- Validation results for Malaysian industrial chemicals and the compounds implicated in the 2019 Sungai Kim Kim incident at Pasir Gudang, Johor A README inside the archive describes every file. Approximately 8 GB of intermediate NumPy arrays are deliberately excluded, asthey are reproducible exactly by re-running the analysis pipeline. ANALYSIS CODEhttps://github.com/sareer555/ghs-hazard-classification DISCLAIMERThese models are computational screening tools. They do not replace laboratorytesting or regulatory assessment under Malaysia's Occupational Safety andHealth (Classification, Labelling and Safety Data Sheet of Hazardous Chemicals)Regulations 2013, or equivalent legislation elsewhere. Predictions must not beused as the sole basis for any decision affecting human safety.