High-throughput toxicological classification of candidate drug compounds using gene expression, evolved neural networks, and a cell-based platform

All new pharmaceutical agents must be screened for potential toxicity in humans. This process includes a series of genotoxic screens in the discovery phase, and in the event the drug is designed for chronic use, a 2-year non-genotoxicity rodent study. Such non-genotoxicity studies are very expensive because of their duration, the amount of compound required, and the number of rodents required. Models capable of predicting genotoxicity during discovery would reduce these costs and increase favorable outcomes for drugs in a pipeline of development by reducing the rate of attrition. To that end, we have used gene expression data and evolved neural networks to classify compounds by their carcinogenicity or genotoxicity. 60 compounds were used for the training and testing of classifiers relative to gene expression from rat liver cells. Genes related to xenobiotic metabolism, proliferation, apoptosis, and DNA damage were identified. Our study demonstrates that evolved neural networks can be used to classify compounds as carcinogenic or genotoxic with reasonable accuracy.