CBERTaXGB: A Domain-Specific Transformer-XGBoost Hybrid for Chemical Toxicity and Flammability Prediction

Abstract

Accurate prediction of chemical toxicity and flammability is essential for the effective implementation of the NFPA 704 Hazard Rating System, underpinning safety protocols in chemical lifecycle management. Traditional experimental methods, such as OECD 423 toxicity tests and ASTM E681 flammability assessments, face practical limitations due to extensive chemical inventories and continuous emergence of novel substances. This study introduces the CBERTaXGB framework, a hybrid model integrating Transformer-based molecular feature extraction (ChemBERTa) with XGBoost's optimized decision architecture, demonstrating exceptional predictive performance for organic hazardous chemicals. The model outperformed existing baseline models, including PR-AUC values of 0.994 and 0.996, AU-ROC of 0.971 and 0.923, and F1-scores of 0.972 and 0.996 for toxicity and flammability classification, respectively, while maintaining high precision and recall. Interpretability analysis confirmed CBERTaXGB's ability to identify critical molecular features such as aromatic stability patterns, electrophilic functional groups, and ester bond configurations, which align with established chemical mechanisms. t-SNE visualizations further validated distinct structural clusters in molecular embeddings. Application to environmental hazard screening of organic substances predicted hazardous properties in 1,319 chemicals, highlighting its practical utility in prioritizing chemicals for enhanced safety measures. CBERTaXGB thus offers a robust, interpretable, and regulatory-compliant approach, significantly advancing intelligent chemical hazard assessment and contributing to environmental protection and public health safety.