In Silico Toxicology Prediction Using Toxicogenomics Data

Toxicogenomics holds the promise of unprecedented advances in two broad, overlapping fields: mechanistic or investigative toxicology, and predictive toxicology. The progress of toxicogenomics has been supported by DNA microarray technology, a powerful tool for directly monitoring patterns of cellular perturbations through the identification and quantification of global shifts in gene expression resulting from pathological alterations within cells and tissues. Microarrays provide a large amount of transcriptional expression data for thousands of individual genes under various experimental conditions. Bioinformatics technologies can determine which genes are meaningful, facilitating the analysis of huge pools of toxicogenomics data in mechanistic and predictive toxicology. This chapter is devoted to computational approaches for the data mining of biomarker genes from toxicogenomics data, leading to toxicity prediction. Many algorithms have been developed for feature gene selection. Most studies on feature selection have found that wrapper methods are superior to filter methods, but many of these studies have over-emphasized prediction accuracy and over-looked the robustness of the selected genes. In fact, this study illustrates that intensity-based moderated t-statistics–support vector machine (SVM) produces more stable gene lists than recursive feature elimination–SVM. Therefore, we have to carefully gauge not only prediction performance but also the robustness of gene sets in feature gene selection. Keywords: biomarker; feature selection; gene selection; machine learning; microarray; support vector machine; toxicogenomics