Characteristics of genomic signatures derived using univariate methods and mechanistically anchored functional descriptors for predicting drug- and xenobiotic-induced nephrotoxicity.

ABSTRACT The ideal toxicity biomarker is composed of the properties of prediction (is detected prior to traditional pathological signs of injury), accuracy (high sensitivity and specificity), and mechanistic relationships to the endpoint measured (biological relevance). Gene expression-based toxicity biomarkers ("signatures") have shown good predictive power and accuracy, but are difficult to interpret biologically. We have compared different statistical methods of feature selection with knowledge-based approaches, using GeneGo's database of canonical pathway maps, to generate gene sets for the classification of renal tubule toxicity. The gene set selection algorithms include four univariate analyses: t-statistics, fold-change, B-statistics, and RankProd, and their combination and overlap for the identification of differentially expressed probes. Enrichment analysis following the results of the four univariate analyses, Hotelling T-square test, and, finally out-of-bag selection, a variant of cross-validation, were used to identify canonical pathway maps-sets of genes coordinately involved in key biological processes-with classification power. Differentially expressed genes identified by the different statistical univariate analyses all generated reasonably performing classifiers of tubule toxicity. Maps identified by enrichment analysis or Hotelling T-square had lower classification power, but highlighted perturbed lipid homeostasis as a common discriminator of nephrotoxic treatments. The out-of-bag method yielded the best functionally integrated classifier. The map "ephrins signaling" performed comparably to a classifier derived using sparse linear programming, a machine learning algorithm, and represents a signaling network specifically involved in renal tubule development and integrity. Such functional descriptors of toxicity promise to better integrate predictive toxicogenomics with mechanistic analysis, facilitating the interpretation and risk assessment of predictive genomic investigations.