Abstract

AbstractIn vitro and in silico methods continue to be evaluated for their potential to inform chemical toxicology evaluation. The research arm of the Environmental Protection Agency has been one of many research bodies evaluating the potential of such methods as part of their ToxCast initiative. We set out to advance the ongoing discussion of improving toxicity testing by exploring whether or not ToxCast physiochemical properties and high-throughput assay data could be used as covariates in predictive models to accurately classify chemicals that either do not or do induce hepatocarcinogenesis in vivo. ToxCast physiochemical and high-throughput assay data were evaluated against known chemicals and in vivo endpoints from the ToxRef curated data set. Hepatocarcinogen-causing chemicals were found to be larger and more lipophilic and complex in shape than control group chemicals. Adjusted logistic regression models using physiochemical properties as covariates accurately classified 71% of the chemicals into the case or control groups, with overall higher specificity than sensitivity. ToxCast in vitro high-throughput assays revealed that the activity of two transcription factors exhibited differences across the case and control groups: Nrf2 and e2f. Logistic regression using high-throughput assay data as covariates resulted in an adjusted model that correctly classified 71% of the chemicals into the case or control groups, also with overall higher specificity than sensitivity. A combined logistic model using physiochemical properties and high-throughput assay data as covariates exhibited similar performance compared to the two adjusted models previously discussed. We found that logistic regression models using physiochemical properties and high-throughput assay data as covariates perform similarly well, accurately classifying chemicals at similar sensitivity and specificity. This analysis suggests that either form of data can be used in the accurate classification of hepatocarcinogenesis, and possibly other apical endpoints. This finding represents a valuable, incremental step forward in the use of such data in the evaluation of chemicals against apical endpoints of health concern. Further study is needed particularly with regard to sensitivity across models, irrespective of the use of physiochemical properties or high-throughput assay data.