Human pluripotent stem cells such as embryonic stem (ES) and induced pluripotent stem (iPS) cells, combined with sophisticated bioinformatics methods, are powerful tools to predict developmental chemical toxicity. Because cell differentiation is not necessary, these cells can facilitate cost-effective assays, thus providing a practical system for the toxicity assessment of various types of chemicals. Here we describe how to apply machine learning techniques to different types of data, such as qRT-PCRs, gene networks, and molecular descriptors, for toxic chemicals, as well as how to integrate these data to predict toxicity categories. Interestingly, our results using 20 chemical data for neurotoxins (NTs), genotoxic carcinogens (GCs), and nongenotoxic carcinogens (NGCs) demonstrated that the highest and most robust prediction performance was obtained by using gene networks as the input. We also observed that qRT-PCR and molecular descriptors tend to contribute to specific toxicity categories.
Keywords: Chemical toxicity prediction; Developmental effect; Embryonic stem cell; Gene network; Molecular descriptor; Multi-kernel support vector machine.