Machine learning for the prediction of drug-induced toxicity

The knowledge of toxicological properties of compounds (e.g. drugs, chemicals, and contaminants) is crucial for drug development, definition of toxicological thresholds and exposure limits. However, toxicological testing, either in vitro or in vivo, is time-consuming, labour intensive and expensive. An alternative to the classic experiments is the use of computational (in silico) approaches, such as machine learning. For machine learning, it is assumed that substances with comparable structure or molecular features also exhibit the comparable pharmacological or toxicological action. Based on the comparison of substances with known pharmacological or toxicological action to substances with unknown properties, models, which were generated using machine learning methods, are able to predict the action of the latter substances.
The aim of this work was the development of predictive machine learning models for the estimation of risk of hepatotoxicity and genotoxicity. These models were then applied on two different substance groups and the outcome was compared to available literature data.
The acute hepatotoxic potential of over 600 different pyrrolizidine alkaloids (PAs) was evaluated using the methods Random Forest and artificial Neural Networks. The predicted qualitative hepatotoxicity of both models was highly correlated. Furthermore, specific structural motives showed different hepatotoxic potential. Overall, the obtained results fitted well with already published in vitro and in vivo data on the acute hepatotoxic properties of PAs.
The genotoxic/ mutagenic potential of PAs was addressed using six different machine learning methods (LAZAR (Lazy Structure-Activity Relationships), Support Vector Machines, Random Forest and two Deep Learning Networks). Even though the models achieved only low to moderate accuracy rates, the best model clearly showed structural specific differences in the predicted genotoxic potential.
Furthermore, the acute hepatotoxic potential of 165 protein kinase inhibitors (PKIs) was predicted using Random Forest and artificial Neural Networks. The models confirmed clinical observations that PKIs have in general a high probability for inducing hepatotoxicity. However, interestingly, there seemed to be a target specific difference, with inhibitors of Janus kinases having the lowest hepatotoxic probability of 60-67%.
The greatest challenge is the performance of the models. This has to be validated e.g. by cross-validation before the model can be used on the substances of interest. Although group statements could be easily obtained, due caution has to be taken while interpreting the results of predictive models for single compounds and if possible, comparison to already published data is advisable, as a form of external validation.