Development of QSAR Model for Subchronic Inhalation Toxicity Using Random Forest Regression Method

Subchronic inhalation toxicity assessment is mainly used to evaluate both workers and occupational risk. Since the assessment requires repeated experiment of chemical exposure to model animals, this type of studies is laborious and expensive. Computational methods for estimating toxicity play an increasingly important role in assessing toxicity of chemicals. They not only predict toxicity of chemicals but can also analyze molecular features of a chemical causing toxic effect. Here we present a computational model to predict the subchronic inhalation toxicity. We retrieved repeated dose inhalation toxicity information and chemicals from OECD eChemPortal website and compile 143 chemical compounds with NOAEC (No Observed Adverse Effect Concentration) values. The random forest regression approach has been applied to learn inhalation toxicity data using a number of molecular descriptors, such as physicochemical properties and fingerprints. Recursive feature selection with nested cross-validation was applied to efficiently select essential molecular features. In results, we successfully obtained a model performing with r(ext)(2) = 0.70 and RMSEext = 0.74 for external set.