Prediction of Respiratory Irritation and Respiratory Sensitization of Chemicals Using Structural Alerts and Machine Learning Modeling

Inhalation of toxic substances and contaminants can have adverse effects on the respiratory tract, leading to a range of health problems, such as irritation and inflammation, allergic reaction and asthma, lung damage, or even death. It is not possible to experimentally evaluate respiratory toxicity for all the thousands of chemicals in use. Here, we generated structural alerts and developed machine learning (ML) classification models to predict respiratory irritation and respiratory sensitization hazards of chemicals using experimental data from publicly available databases and the literature. We identified 13 structural alerts for respiratory irritants and 18 structural alerts for respiratory sensitizers. We also developed a set of models for each hazard using different types of molecular descriptors and ML techniques. Five of the best performing models were combined into a consensus classification model for respiratory irritation, and four individual models were used to develop a consensus classification model for respiratory sensitization. The prediction accuracy of the respiratory irritation consensus model was 84% on the training set and 88% on the test set, and the accuracy of the respiratory sensitization consensus model was 86% on both training and test data sets. A combination of generated structural alerts and ML models was used to screen occupational- and military-relevant chemicals. Out of 687 screened occupational chemicals, 62 compounds were identified as respiratory irritants and 121 chemicals as respiratory sensitizers, while 47 chemicals were predicted as irritants and 36 compounds as sensitizers in the list of 525 military-relevant chemicals.