You are here:
Methodology for Uncertainty Analysis of Dynamic Computational Toxicology Models
Citation:
DAVIS, J., J. F. WAMBAUGH, R. GARCIA, AND R. W. SETZER. Methodology for Uncertainty Analysis of Dynamic Computational Toxicology Models. Presented at Computational Toxicology Board of Scientific Counselors Review, RTP, NC, September 29 - 30, 2009.
Impact/Purpose:
We discuss approaches for developing prior distributions quantifying the uncertainty in chemical-specific parameters based on comparisons of measured values with predicted values from computational or in vitro methods.
Description:
The task of quantifying the uncertainty in both parameter estimates and model predictions has become more important with the increased use of dynamic computational toxicology models by the EPA. Dynamic toxicological models include physiologically-based pharmacokinetic (PBPK) models, closely-related (and often coupled) pharmacodynamic models or biologically-based dose-response (BBDR) models, and models such as those for virtual tissues. Given a set of values for biological parameters describing the subject and chemical parameters describing the compound, biological models can make predictions that both allow assessment of the understanding of how data came to be (interpolation) as well as what might occur under different conditions (extrapolation). Careful consideration must be given to determining the value of model parameters and uncertainty about those values as well as the selection of one model over another given the overall uncertainty about different models. Quantitative uncertainty analyses are necessary for fully vetting models for applications such as risk assessments. Along with uncertainty, in these types of systems variability must also be accounted for on various levels, such as variation in model parameters across individuals in a population as well as variation in experimental data. Thus, the analysis of computational toxicology models requires valid statistical methodologies that are capable of handling both uncertainty and variability accurately. Several methodological issues are generic to these dynamic toxicological models. Often, values must be determined for parameters in the absence of hard chemicalspecific data, contributing parameter uncertainty. We discuss approaches for developing prior distributions quantifying the uncertainty in chemical-specific parameters based on comparisons of measured values with predicted values from computational or in vitro methods. Such informative priors are beneficial both in the context of Bayesian estimation, and for assessing uncertainty of predictions from dynamic models for which in vivo data are entirely lacking. One application wherein informative priors are particularly useful that is of interest to the Agency focuses on the development of better quantitative approaches for cumulative risk assessments of linked exposure-doseeffects models. A major part of this effort involves formulating PBPK models, which include well known physiological parameters as well as unknown psysicochemical parameters that describe the uptake and disposition of chemicals or toxins through the body. Using PBPK models as a motivating example, we discuss some of the advantages and drawbacks associated with the use of hierarchical Bayesian analysis in model calibration, uncertainty analysis, and model evaluation. Conventional computational methods for estimating parameters and evaluating their uncertainty, which were developed for substantially simpler non-linear models, require lengthy computations (e.g., weeks or even months) when applied to dynamic models. We discuss some attempts to standardize this analysis, address the issue of efficient computational time for deterministic (e.g., PBPK) models, and deal with uncertainty in stochastic models (e.g., agent-based virtual tissue models). Finally, we discuss evaluating how well models describe data and approaches to evaluating model uncertainty. This work was reviewed by EPA and approved for publication but does not necessarily reflect official Agency policy.
URLs/Downloads:
Methodology for Uncertainty Analysis of Dynamic Computational Toxicology Models (abs) (PDF, NA pp, 8 KB, about PDF)Methodology for Uncertainty Analysis of Dynamic Computational Toxicology Models (poster) (PDF, NA pp, 1014 KB, about PDF)