PBPK models are powerful tools in environmental health, and can integrate various types and sizes of PK data sets into a common analysis framework. This approach can be enhanced significantly by incorporating the anatomy and physiology of the underlying biological system into the framework through the use of PBPK models, which have a much wider range of applicability than traditional empirical compartmental PK models. This greater applicability domain is particularly important when trying to understand population PK, since it is generally infeasible to obtain empirical data over the full range of variability in environmentally exposed populations and communities.

Unlike traditional compartmental models, PBPK model parameters are not uniquely identifiable based on experimental PK data alone, as it is infeasible for human populations to obtain all the information (tissue volumes and flows, tissue-specific metabolism and transport, and tissue-specific compound concentrations) needed to identify all their parameter values on the basis of maximum likelihood alone. It has therefore been proposed that population PBPK models, like other mechanistic models, are best parameterized in a Bayesian framework, which seamlessly integrates the prior information available from the literature on PBPK parameter values and correlations. A population model further helps Bayesian integration of prior information, since priors rest on population parameters (e.g., average blood volume) rather than on individual values.

The PBPK model–population model–Bayesian framework results in a clear, elegant, versatile, flexible approach that does not depend on asymptotic or Gaussian approximations. This approach has been demonstrated in multiple case studies of environmental or occupational toxicants, including several by Dr. Chiu. These include use of population PBPK models for bi-directional translation of mouse population-based models to human health and risk assessment.    

While in vitro data offer a great potential to fill knowledge gaps related to toxicity information, several factors that can significantly influence in vivo toxicity such as bioavailability, metabolic and renal clearance, and protein binding, are not accounted for in in vitro test data. Therefore, the incorporation of toxicokinetic dosimetry modeling is a necessary element of the vision to move toxicity testing away from in vivo rodent studies to human-based in vitro assays and enable direct comparisons with human exposures for evaluating risk.

IVIVE forms a critical link from Stressors to Responses, one of the Center’s themes. IVIVE models also account for population variability in toxicokinetics, and therefore also touch on the theme of Individuals to Populations. Moreover, IVIVE approaches have been demonstrated to have a substantial impact on hazard prioritization and ranking, thereby providing critical information to inform the theme of Community, Regulation, and Policy. Because of the large number of chemicals the Center will investigate, both individually and in combination, an essential element of the DSFC is to maintain, and augment as necessary, a repository of models and their parameters in a readily computable form. Pearce et al. recently developed an open-source library “httk” (high-throughput toxicokinetics) in the “R” statistical software program that includes toxicokinetic models and pre-calculated RTK factors, as well as computer code for dynamic simulation and Monte Carlo sampling using several different toxicokinetic model formulations. Center investigators have also extended these approaches to mixtures.   

A number of commercial and freely available software packages can conduct dose-response modeling, but their use in evaluating toxicological datasets is still quite limited compared to the traditional application of pair-wise statistical tests to determine “no observed adverse effect levels” (NOAELs).

The limitations of the NOAEL as a starting point for toxicological risk assessment have been recognized for decades, and it is generally accepted that the BMD, introduced by Crump, is more scientifically appropriate. The BMD is the dose associated with a specific size of effect, the benchmark response (BMR; we use M, for magnitude of effect, to denote this value). The BMD_M is estimated, with associated confidence interval/statistical distribution, by statistical model fitting to dose-response data. In order to facilitate use of toxicological data generated by Center investigators in regulatory decision-making, BMD modeling services will be provided.

We previously developed a standardized workflow for BMD modeling that was applied to hundreds of datasets, and which can be adapted for supported Center investigator needs. Additionally, the recent advent of high-throughput and population-based in vitro and in vivo models presents a new challenge. We have recently demonstrated development of a hierarchical Bayesian workflow for dose-response modeling of population data, which has a particular application to population-based in vitro data. This approach enabled a statistically rigorous analysis of population variability toxic responses, providing key information on toxicodynamic susceptibility.   

Methods for population-based risk prediction for the purposes of informing environmental health decisions continue to evolve, with increasing emphasis on use of in vitro data, dose-response modeling, and probabilistic analyses of uncertainty and variability. Recent National Academy of Sciences reports have recommended development of predictive risk estimates that provide stakeholders and decision-makers with a fuller characterization of the potential population human health effects associated with different levels of exposure.

Recently, the World Health Organization (WHO) developed a harmonized probabilistic approach based on the concept of the “target human dose” (denoted HD_M^I) that responds directly to these National Academy of Sciences recommendations. While similar to the traditional “reference dose” (RfD) in its derivation, it differs in several important ways. First, being based on dose-response modeling, it is always tied to a specific health effect, unlike a NOAEL or RfD. Second, it is a harmonized approach applicable to both cancer and non-cancer effects. Furthermore, it extends existing decision benchmarks by quantitatively defining the protection goals in terms of population incidence (I) and magnitude (M) of the critical effect in the human population. This allows for incorporation into regulatory impact analyses that require economic benefit-cost estimates. Finally, it uses uncertainty distributions derived from historical and/or chemical-specific data and analyses in order to derive confidence intervals that quantify the level of uncertainty in the overall result. 

The diversity of data stemming from different biological systems pose specific challenges to life sciences researchers during the data analysis steps, particularly to those who are not experienced in computer programming. Many of the tools developed within life sciences research are from other researchers in various academic institutions and deployed in a UNIX/LINUX computing environment, which requires at least a basic knowledge of those operating systems. The DSFC can support investigators and their staff by providing algorithm design and computer programming services to develop easy-to-use command-line tools and scripts, graphical user interfaces, and/or online web apps/interfaces that are customized to their data analysis workflow. Recent examples include:

•  lncRNApipe (https://github.com/biocoder/Perl-for-Bioinformatics/tree/master/NGS-Utils) is a command-line‒based software pipeline for identifying putative long non-coding RNA from RNA-Seq reads prepared with Total RNA-Seq protocol from Illumina. It is written in PERL programming language.
• SBEToolbox (https://github.com/biocoder/SBEToolbox) is a menu-driven graphical user interface written in MatLab to compute various topological metrics and deduce various clusters within biological networks.

gQTL (https://genomics.tamu.edu/gqtl) is a web-based application for QTL analysis of Collaborative Cross Mice. 

In some cases, novel bioinformatics methods may need to be developed to address Center project needs. Past examples include the following:

•  Regulatory Pathway Inference and Classification. Many associations between gene expression-activation and clinical response are correlative as opposed to mechanistic. As a result, these associations are often poor predictors of how cellular signaling pathways will behave if perturbed by environmental exposure modifiers such as diet and exercise. Using systems biology approaches, we have developed mathematical models to integrate the dynamics of reactions in cells and interactions between cells in response to environmental exposure, leading to a dynamic, potentially personalized, “molecular signature." We have recently demonstrated that the integration of existing pathway knowledge with expression data in a novel systems-based Bayesian approach can be used to achieve optimal classification accuracy, relative to standard data-driven approaches.

• Prediction of Toxicology Outcomes using Data Integration and Bridging from Cell and Animal Models to Human Disease. One approach to understanding mechanisms from Stressors to Response, one of the key themes of this center, is through predictive models that integrate the different modalities of data, including prior information, for the different animal models and interventional and observational studies (environmental exposure and intervention treatments). Some approaches include the concept of the Coefficient of Determination (CoD), Bayesian network models, and Canonical Correlation Analysis (CCA). The CoD paradigm has been used to detect the so-called “master-slave” regulatory relationships between genes in a Gene Regulatory Network (GRN) and more recently to estimate the probabilities of detecting true feature sets. The Bayesian network model allows for a proper interpretation of plausible causality relationships between regulatory and regulated elements in a given pathway of interest. CCA can integrate existing pathway knowledge with expression data to detect diet-dependent interactions between host transcriptome and microbiome. More recently, we have developed a novel methodology for detecting correlative structures in NextGen sequencing data. This methodology can be used, for example, to examine exposure to phthalates during two critical developmental windows—prenatal and early childhood—on the gut metagenome and host gut gene expression.