You are here:
Comparative Analysis of Predictive Models for Liver Toxicity Using ToxCast Assays and Quantitative Structure-Activity Relationships (MCBIOS)
Citation:
Liu, J., R. Judson, M. Martin, H. Hong, AND I. Shah. Comparative Analysis of Predictive Models for Liver Toxicity Using ToxCast Assays and Quantitative Structure-Activity Relationships (MCBIOS). Presented at MCBIOS Annual Conference, Stillwater, OK, March 06 - 08, 2014. https://doi.org/10.23645/epacomptox.5196901
Impact/Purpose:
The results suggest the potential utility of machine learning techniques for predicting chronic liver injury and the relative advantage of HTS assays compared to QSAR-based descriptors alone.
Description:
Comparative Analysis of Predictive Models for Liver Toxicity Using ToxCast Assays and Quantitative Structure-Activity Relationships Jie Liu1,2, Richard Judson1, Matthew T. Martin1, Huixiao Hong3, Imran Shah1 1National Center for Computational Toxicology (NCCT), US EPA, RTP, NC, USA. 2University of Arkansas at Little Rock/University of Arkansas for Medical Sciences, Little Rock, AR, USA. 3Division of Bioinformatics and Biostatistics, National Center for Toxicological Research, U.S. Food and Drug Administration, Jefferson, AR, USA Keywords: ToxCastDB; ToxRefDB; high-throughput screening, QSAR. The U.S. EPA’s ToxCastTM program uses hundreds of high-throughput, in vitro assays to screen chemicals for bioactivity. Here our objective is to predict rodent liver toxicity from guideline testing studies using either ToxCast data or chemical structure information. We used in vitro assays and molecular structures for 880 ToxCast chemicals to classify in vivo rodent hepatotoxicity outcomes. The classification models were built using four machine learning approaches (LDA, Naïve Bayes, SVM and KNN). Chemicals were represented by either in vitro bioactivity assays, or quantitative structure-activity (QSAR) descriptors. Liver toxicity was defined by 3 broad categories of histopathologic effects including hypertrophy, injury and proliferative lesions. After 10-way cross-validation testing, we found the best predictive balanced accuracy of classifiers using ToxCast assays was 0.715, 0.690, and 0.656 for hypertrophy, injury and proliferative lesions, respectively, and this was significantly greater than the balanced accuracy of QSAR-based models (0.649, 0.639, and 0.606 for hypertrophy, injury and proliferative lesions, respectively). The results suggest the potential utility of machine learning techniques for predicting chronic liver injury and the relative advantage of HTS assays compared to QSAR-based descriptors alone. This abstract does not represent EPA policy.
URLs/Downloads:
DOI: Comparative Analysis of Predictive Models for Liver Toxicity Using ToxCast Assays and Quantitative Structure-Activity Relationships (MCBIOS)