Diagnosis of drug-induced renal tubular toxicity using global gene expression profiles

Abstract

Toxicogenomics can measure the expression of thousands of genes to identify changes associated with drug induced toxicities. It is expected that toxicogenomics can be an alternative or complementary approach in preclinical drug safety evaluation to identify or predict drug induced toxicities. One of the major concerns in applying toxicogenomics to diagnose or predict drug induced organ toxicity, is how generalizable the statistical classification model is when derived from small datasets? Here we presented that a diagnosis of kidney proximal tubule toxicity, measured by pathology, can successfully be achieved even with a study design of limited number of training studies or samples. We selected a total of ten kidney toxicants, designed the in life study with multiple dose and multiple time points to cover samples at doses and time points with or without concurrent toxicity. We employed SVM (Support Vector Machine) as the classification algorithm for the toxicogenomic diagnosis of kidney proximal tubule toxicity. Instead of applying cross validation methods, we used an independent testing set by dividing the studies or samples into independent training and testing sets to evaluate the diagnostic performance. We achieved a Sn (sensitivity) = 88% and a Sp (specificity) = 91%. The diagnosis performance underscores the potential application of toxicogenomics in a preclinical lead optimization process of drugs entering into development.