You are here:
Variability in in vivo toxicity data
Citation:
Pham, L. Variability in in vivo toxicity data. Presented at Society of Toxicology Annual Meeting, Baltimore, MD, March 10 - 14, 2019.
Impact/Purpose:
Predictive models for toxicity are often evaluated using results from animal toxicology studies. However, variability in these in vivo reference data limits the upper bounds of alternative model predictivity. This talk will discuss the variance within in vivo toxicity studies and its impact on the predictive performance of models that reference these data.
Description:
Integrating new approach methods (NAMs) with traditional toxicological data to fill information gaps has the potential to inform hazard identification for thousands of chemicals. However, regulatory acceptance often requires comparison between NAM and animal toxicity study results using a reference chemical set. This expectation often does not consider the presence of potential confounders and variability in the observed in vivo toxicity studies. Acceptance of NAMs should include some evaluation of the variability and/or reproducibility of the reference chemical in vivo data. Our work characterizes variance in systemic effects from studies summarized in the US EPA’s Toxicity Reference Database (ToxRefDBv1). We defined the variance explained by documented study parameters, as well as the unexplained variance, and in doing so set an upper limit on the anticipated R-squared for predictive modeling of systemic lowest effect levels (LELs). The total variance in systemic LEL values (with log¬10-mg/kg/day units) was ~1; unexplained variance, as approximated by the residual mean square error (MSE), was ~0.3, or approximately one-third of the total variance. We used the root mean square error in these data (0.58) to define a 95% confidence interval for prediction of systemic LEL values in this data set as 2.3 log10¬-mg/kg/day. Based on the relationship between MSE and R-squared for goodness-of-fit, an upper bound on the R-squared for a predictive model would approach ~70%. Our analysis suggests that a predictive model can only account for ~70% of the variance in the systemic LELs in ToxRefDB, and that a given observed systemic LEL of 1 mg/kg/day might be predicted to be within 0.07 to 14 mg/kg/day with 95% confidence. These results indicate that variability of in vivo studies used to train a given model or NAM will limit the prediction accuracy of the model. This abstract does not necessarily reflect U.S. EPA policy.