You are here:
Developing Predictive Toxicity Signatures Using In Vitro Data from the EPA ToxCast Program
Citation:
JUDSON, R., D. J. DIX, K. A. HOUCK, M. T. MARTIN, H. M. MORTENSEN, AND R. J. KAVLOCK. Developing Predictive Toxicity Signatures Using In Vitro Data from the EPA ToxCast Program. Presented at Society of Toxicology Annual Meeting, Baltimore, MD, March 15 - 19, 2009.
Impact/Purpose:
We present a variety of machine learning approaches to mine this complex data set for toxicity signatures with both high sensitivity and specificity. These include linear discriminant analysis, support vector machines and neural networks. In addition to these automated approaches, we also present more hypothesis-driven, mechanism-based signatures including ones probing nuclear receptor-mediated liver toxicities. An important observation from these first analyses is that one often needs to include assays of multiple types, probing multiple mechanisms or pathways, to adequately predict in vivo toxicity across a wide range of chemicals.
Description:
A major focus in toxicology research is the development of in vitro methods to predict in vivo chemical toxicity. Numerous studies have evaluated the use of targeted biochemical, cell-based and genomic assay approaches. Each of these techniques is potentially helpful, but provides only a partial view of the complex biology that leads to tissue, organ or whole organism toxic effects. Here we present the first results from the ToxCast program that combines multiple types of assays into “toxicity signatures” that are optimally predictive of particular in vivo toxicity endpoints. A toxicity signature is in essence a function that takes as input the results of a set of assays run on a chemical, and produces a prediction of the toxicity of that chemical for a specific in vivo endpoint. We used the EPA ToxCast program Phase I data, which includes about 600 in vitro assays run on a total of 320 chemicals, the majority of which are pesticide active compounds. The corresponding in vivo toxicity data is taken from the Toxicology Reference database (ToxRefDB), which is tabulated data from guideline animal studies for chronic, cancer, developmental and reproductive toxicity endpoints.