Predicting Organ Toxicity Using <i>in Vitro</i> Bioactivity Data and Chemical Structure

Animal testing alone cannot practically evaluate the health hazard posed by tens of thousands of environmental chemicals. Computational approaches making use of high-throughput experimental data may provide more efficient means to predict chemical toxicity. Here, we use a supervised machine learning strategy to systematically investigate the relative importance of study type, machine learning algorithm, and type of descriptor on predicting <i>in vivo</i> repeat-dose toxicity at the organ-level. A total of 985 compounds were represented using chemical structural descriptors, ToxPrint chemotype descriptors, and bioactivity descriptors from ToxCast <i>in vitro</i> high-throughput screening assays. Using ToxRefDB, a total of 35 target organ outcomes were identified that contained at least 100 chemicals (50 positive and 50 negative). Supervised machine learning was performed using Naïve Bayes, k-nearest neighbor, random forest, classification and regression trees, and support vector classification approaches. Model performance was assessed based on F1 scores using 5-fold cross-validation with balanced bootstrap replicates. Fixed effects modeling showed the variance in F1 scores was explained mostly by target organ outcome, followed by descriptor type, machine learning algorithm, and interactions between these three factors. A combination of bioactivity and chemical structure or chemotype descriptors were the most predictive. Model performance improved with more chemicals (up to a maximum of 24%), and these gains were correlated (ρ = 0.92) with the number of chemicals. Overall, the results demonstrate that a combination of bioactivity and chemical descriptors can accurately predict a range of target organ toxicity outcomes in repeat-dose studies, but specific experimental and methodologic improvements may increase predictivity