SOT 2025 Poster Abstract- Characterizing Variability and Uncertainty for Parameter Subset Selection in PBPK Models
On this page:
Physiologically based pharmacokinetic (PBPK) models are regularly used to inform human health risk assessments of chemicals. By performing simulations with a PBPK model, one can estimate human exposure levels that result in internal doses equal to those predicted for laboratory animals exposed to substances according to specific experimental dosing regimens. Using point estimate parameter values representing an “average” adult human, one can apply a PBPK model to estimate a single “human equivalent dose” (HED), which refers to the human equivalent concentration (for inhalation exposure) or dose (for oral exposure) of a substance that is expected to induce the same magnitude of toxic effect for a human as that observed for laboratory animals exposed to a specific concentration or dose. However, point estimates do not address variability among humans or uncertainty in parameter values. To address human variability in chemical risk, experts have recommended probabilistic risk assessment paradigms which use parameter distributions to generate probabilistic reference values which can then be used to identify 1st, 99th or other population percentiles. Inevitably, these model predictions come with uncertainty, especially when model parameter distributions are not well-known. Thus, it is important to determine how the uncertainty in the output of a PBPK model can be attributed to different sources of uncertainty in its inputs/parameters, in particular to identify the most influential parameters.
We analyzed published PBPK models for dichloromethane (DCM) and chloroform and evaluated the relative importance of parameters when computing internal dose metrics for a variety of external doses. First, we identified globally influential parameters through the Morris method (which calculates “elementary effects,” or differences in output following one-at-a-time changes in input parameters) and the Sobol’ method (an approach which uses decomposition of variance to rank parameter effects). Next, we used Monte Carlo sampling and reverse dosimetry to generate HED distributions. We then analyzed the numerical stability of these distributions by fixing non-influential parameters to scalar values and quantifying the relative importance that each remaining parameter had on the model output. Finally, we conducted identifiability analysis to obtain an even smaller subset of important parameters.
We used distributional data for a relatively small subset of influential model parameters (i.e., variability was not considered for most model parameters) and obtained estimates of the 1st and 99th percentiles of HED distributions to within one percent of the “true” values of these percentiles (i.e., those calculated by varying all the model parameters). The list of important (influential) parameters (approximately 6-10 parameters for the DCM model and 13-18 parameters for the chloroform model) were also identified by the Morris and Sobol’ methods as parameters that accounted for greater than 98% of the model output variation for the DCM model and more than 88% of model output variation for the chloroform cases. However, the degree of influence of each parameter depends on the chemical, dose, route of exposure, and internal dose metric. Global sensitivity analysis methods also identified parameters that exert minimal influence on model output. Collecting highly accurate distributional data for such parameters may be less important.
Thus, predictions of dose metrics and HEDs (central estimates or extreme percentiles) can be greatly improved by having precise knowledge about certain input parameter distribution details (e.g., shape and variance), but not all. In the future, these methods can be used to identify parameters or situations for which it is important to allocate time and resources to collect data for more accurate representations of parameter variability..