4
Harnessing Data for Decision Making
Chirag Patel of Harvard Medical School opened by presenting strategies for using data to protect public health. The workshop participants then split up into breakout sessions to discuss important topics: alternative animal models (led by Robyn Tanguay, Oregon State University), cumulative risk assessment (led by Linda Birnbaum, National Institute of Environmental Health Sciences, retired), and biomarkers (led by Marie Fortin, Jazz Pharmaceuticals). Darryl B. Hood, Ohio State University, moderated the entire session.
STRATEGIES FOR USING DATA DELIVERING PUBLIC HEALTH1
Patel discussed leveraging the biobank scale to prioritize exposure–phenotype associations in children. Biobanks are a collection of biological samples (such as blood) and other health data, and examples include the UK Biobank,2 which contains genetic and health information from half a million UK participants; Lifelines, a large multigenerational, prospective cohort study of 10 percent of the northern population of the Netherlands; and All of Us, a U.S. one-million-person cohort study.
Patel is interested in whether it is possible to identify new and established exposures associated with health in big biobanked data. Biobank data have been beneficial in informing the discovery and replication of
___________________
1 This section is based on a presentation by Chirag Patel, associate professor, Harvard Medical School.
2https://www.ukbiobank.ac.uk/ (accessed November 3, 2022).
new genes. The large biobank datasets may allow for the interrogation of new biological pathways and the triangulation of causal relationships with data on biological mechanisms. Biobank data could also be used in meta-analysis and synthesis to inform policy research.
Genomics studies have been leveraging biobanks for some time, and the publications and relationships explored are growing exponentially. In 2018, about 3,567 publications and 71,673 genotype–phenotype associations were discovered. In 2022, 5,690 publications had 372,752 genotype–phenotype associations discovered. The massive datasets have led to replicated findings and meta-analysis across cohorts. New biological pathways have also been discovered.
Patel argues that these biobanks could also be leveraged to understand children’s environmental health, although the environment is very complex. The exposome includes both shared (air pollution) and non-shared (dietary) exposures. Exposures also have different effects depending on when they occur, which might change their prevalence and distribution. Patel thinks that biobanks with frequent measures may be able to hold us accountable for measuring these exposures in time-relevant windows to help inform measurement frequency.
According to Patel, some requirements must be met to use biobanks to identify exposures and help build policy around health and disease. First, and very important, is the participants’ consent (Klima et al., 2014).
Another is that you need the correct data to answer policy-relevant questions. Measures must be taken in utero to understand developmental impacts. Biosamples are also needed to ascertain environmental attributes, such as biomarkers. Information from both parents is needed to understand heritable and nonheritable risks. Health outcomes should be measured across life stages (adolescence and adulthood) to explore phenotypic relationships. Geospatial exposome biomarkers (such as climate and air pollution) will allow the evaluation of shared exposures across communities. Finally, data approaches to harmonize across cohorts for meta-analyses and systematic reviews will be necessary for informing policy decisions.
An emerging biobank that will be used to explore children’s environmental health is the Adolescent Brain Cognitive Development Study,3 said Patel. It is the largest U.S. long-term study of brain development and child health, with 11,880 children ages 9–10 invited to join. The study has begun to integrate multiple modalities, such as MRIs of the brain, geospatial markers, and biomarkers in blood to assess connections as the children develop.
Another example of a biobank study focusing on child development is the Environmental Determinants of Diabetes in the Young Study. It is recruiting individuals who are at high genetic and familial risk for Type 1 diabetes to determine the environmental triggers of type 1 diabetes. It collects information very frequently throughout early life. Blood samples are collected at 6-month intervals, stool samples are almost monthly, and other matrices to measure environmental exposure assays are also collected.
As an example of the type of predictions that can be made with large data, Patel described a study that used an insurance database of 45 million individuals and was able to evaluate genetics and environment by looking at a subset of 60,000 twin pairs (Lakhani et al., 2019). In this cohort, they can now evaluate how much genetics and environment play a role in disease outcomes and look at scale because the sample size is so large. Patel and his colleagues were able to link all the twin data with what they call the “environmental exposome warehouse,” a database of air pollution, weather, and census social deprivation index data.
Patel shared discoveries that his team has made with biobank data. Using the twin study, they were able to observe the proportion of genetic and environmental contributions to various diseases. Across different categories of outcomes, developmental, endocrine, digestive, and metabolic disorders have a large environmental component (Lakhani et al., 2019). Using the UK Biobank data, Patel’s colleagues completed an exposome-wide
___________________
3https://abcdstudy.org (accessed November 8, 2022).
association study in Type 2 diabetes. They prioritized nonmodifiable and modifiable exposome factors and put them into a model, the Poly eXposure Risk Score. It was almost as predictive as a Clinical Risk Score, based on blood pressure, BMI, glucose, HDL, and triglycerides, and a much better predictor than a Polygenic Risk Score (He et al., 2021). Patel shared numerous other examples from his group, including predicting pancreas and liver age, using deep learning or liver MRI images, and using biobanked samples to perform functional exposome-wide association studies to discover biologically relevant exposures (Chung et al., 2021; Le Goallec et al., 2022).
Some cautions in using big data are the possibilities for confounding, reverse causation, and missing data. Some potential methods to separate a true relationship from a spurious correlation are (1) sensitivity analyses or (2) meta-analyses across different cohorts, (3) and replication of findings in other cohorts, and (4) triangulating evidence from different data sources is necessary (trying to see how associations emerge or are concordant across different streams of information). According to Patel, it is possible to identify new and established exposures with health and big biobank data.
BREAKOUT SESSIONS
Participants formed three breakout groups: alternative animal models, cumulative risk assessment, and biomarkers of exposure. Participants in each group answered the following:
- What are opportunities and barriers?
- What are the best practices and potential solutions?
- How can the EPA play a role in advancing the field?
The descriptions of these breakout sessions are factual summaries of the breakout discussions that occurred during the workshop. Statements, recommendations, and opinions expressed are those of the presenter. They are not necessarily endorsed or verified by the breakout groups, the Planning Committee, or the National Academies, and they should not be construed as reflecting any group consensus or opinions of the EPA.
Alternative Animal Models4
Tanguay reported on her discussion with Heather Patisaul of North Carolina State University and Aramandla Ramesh of Meharry Medical
___________________
4 This section is based on a presentation by Robyn Tanguay, professor, Oregon State University.
College on alternative animal models. Alternative animal models are methods that have been used to limit the use of mammals in toxicology. The session focused on general principles and not any specific toxicity testing systems.
These models have several advantages, Tanguay said. A primary advantage is the prevention focus. Toxicity testing using alternative platforms can be conducted more quickly than with mammalian models, allowing for testing of more compounds and mixtures. Another advantage is exploring the mechanism across various life stages, which is better than animal toxicity tests or epidmiology for children’s environmental health, because the disease outcomes might be much later in life. They can also allow for using fewer animals. Another advantage is that chemical replacements can be screened for various toxic activity profiles rapidly before going to market. Animal alternatives allow for an understanding of biological mechanisms, making it possible to group chemical exposures by not just chemical class but also a biological mechanism. Activity relationship data could better arm people who are conducting human exposome testing. Tanguay added that her breakout group members argued that the bioactivity signals or a consistency across tests that chemical exposure is bioactive could be sufficient to make decisions. Tanguay explained that her breakout group members discussed that they would like to see a stronger focus on hazard characterization (the recognition that a stressor can cause harm) and not risk (the chance that the stressor will cause harm), which would lead to more protective decisions. Another point made was that in the big data era, scientists are better positioned to take the data generated from various alternative models and human participants and look across the different systems to evaluate gene and environment interactions. Tanguay then said that her breakout group discussed that scientists may now be able to detect biologically active and biologically accumulative chemicals. If such advanced testing had been possible 100 years ago, exposomics might not even be needed.
Tanguay then described her breakout group’s discussion on barriers to alternative animal models. A primary barrier is that standard animal models are still examples of the best science and practices for toxicity testing. Another barrier is the need to establish human relevancy for different alternative animal models. Often toxicity testing is too focused on single chemical interactions, and the doses are too high to be relevant to human exposure.
Tanguay described her breakout group’s discussion of best practices and potential solutions to overcome the barriers to using alternative animal models. Case studies could be performed with known toxicants to evaluate the alternative assay’s ability to predict human toxicity. The assays also can include improved genetic diversity to establish human
relevancy. Additionally, to overcome the barrier of how to better use bioactivity or hazard signals in decision making, Tanguay said that her breakout group participants felt there is not a safe level of a biologically active chemical. Tanguay also described that her working group felt that hazard determinations must incorporate molecular modeling tools such as those used in pharmaceutical development. “Right now, most people do not know how EPA makes policy decisions,” says Tanguay. The EPA could engage with the public more so that people understand how policy decisions are made.
The EPA has a large role in advancing the field of alternative animal models and improving toxicity testing. Data science approaches, integration of datasets, mining of different datasets, and predictive toxicology could be more of a focus. Even when data are available, they do not always seem to be fully used, and more hazard data need to be applied in decision making. The EPA could advocate for more independence from industry and be more inclusive of the public in decision making. Communities that study hazards need to come together to advance the field of environmental health; epidemiologists, toxicologists, and data scientists are often separated, and multidirectional discussions are needed to advance understanding. The EPA could also consider whether recent decisions protect children and account for total costs. The risk decision paradigm needs to take the health and protection of children into account; with a more upstream policy focus on hazards, less research on the probability that the hazard might occur would be needed.
Cumulative Risk Assessment5
Linda Birnbaum provided an overview of the discussion she led on cumulative risk assessment or the analysis, characterization, and possible quantification of the combined risks to human health from multiple agents or stressors. Perera of the Columbia Center for Children’s Environmental Health, Jeanne Briskin of the EPA, Nsedu O. Witherspoon of the Children’s Environmental Health Network participated in the discussion. There are opportunities for the use of cumulative risk assessment. Many scientists and health agencies have recognized the importance of cumulative risk assessment, and the science is available to allow risk assessment across chemical classes, various classes of social determinants of health, and broad groups of nutrients, Birnbaum described. New options in TSCA also may allow for evaluating chemical categories.
___________________
5 This section is based on a presentation by Linda Birnbaum, former director, National Institute of Environmental Health Sciences (retired).
Birnbaum described barriers that are the trade-offs between precision and big-picture accuracy. Information cannot always be high quality and complete; that is a limitation that needs to be recognized and not a reason for not moving forward. Real-world data that citizen scientists collect can be better incorporated into decision making. Risk assessment has sometimes been tangled up in the regulatory decision and the risk management process, which is not ideal as determinations of risk are best when separate from technological and cost considerations. More data are needed on children’s exposures; National Health and Nutrition Examination Survey (NHANES) data on young children and pregnant women are quite limited. Some alternative animal models and new approaches may miss important stages of development, leading to null findings.
Birnbaum described themes from her breakout group’s conversation about best practices and potential solutions. One avenue to explore cumulative risks is that big data could be mined for potential interactions such as gene–environment interactions or those between different chemical exposures, social determinants, diet, and stress. Health system data, such as that from Kaiser or Geisinger, could be leveraged more often to evaluate environmental impacts on human health. Researchers in North Carolina are just beginning to look at drinking water PFAS levels and finding associations with hospitalization records. Pattern recognition techniques could be applied using multiple lines of evidence. Citizen science and community-based participatory research may broaden how the EPA measures risk and synthesizes data.
Birnbaum reported on her breakout group’s discussion on the EPA’s role in advancing cumulative risk assessment and children’s health. The EPA has begun looking at social stressors only recently. The Integrated Risk Information System (IRIS) program within the EPA’s Office of Research and Development that conducts hazard evaluations for different chemicals has been looking at susceptible exposure windows, such as any stage with rapid cell differentiation. The EPA has made progress using the 10-fold adjustment factor. Still, Birnbaum said that her working group discussed that the 10-fold adjustment factor might not be sensitive enough to account for all developmental stages and endpoints, as mentioned by Woodruff and Miller. The EPA can also lead by looking more broadly at classes of chemicals, social determinants, nutrients, etc. Revisions to TSCA may offer a chance to evaluate the uses of different classes of chemicals. “One thing we want to encourage is that we don’t want the perfect to be the enemy of the good in advancing the field,” said Birnbaum.
Biomarkers6
Marie Fortin provided an overview of the breakout discussion on biomarkers. Hood, Patel, Skakkebæk, and Michael Langston, University of Tennessee, all participated.
Fortin first outlined several challenges associated with biomarkers of exposure. Some are measured with metabolites, which can be biomarkers of more than one compound, making exposure misclassification a possible risk. Another barrier is that exposomics data are hard to merge with social determinants due to data quality issues and co-linearity. Further, it can be difficult to measure cumulative exposures to chemicals with short biological half-lives. Life stage influences biomarkers in the blood; for example, increased blood volume throughout pregnancy can affect measurements. Taking biological samples from people is invasive, which limits what can be collected, and some matrices are difficult to process to measure exposures. Minority populations are also underrepresented in many studies using exposure biomarkers. Finally, diseases and chronic diseases are rarely attributed to a single cause or exposure, and therefore to understand disease etiology, it is important to measure many different exposures.
Biomonitoring is not without opportunities, said Fortin. Exposomics data are very rich compared to targeted biomarker approaches, and by using models that predict exposures, it can be possible to identify and link these exposures to external sources. Epigenetics is an untapped resource with great potential; epigenetic modifications are linked to different disease phenotypes. Opportunities exist to reconstruct exposures using different matrices such as hair, bone, and nails which are well-established for environmental contaminants such as lead and mercury. Subclinical effects, such as stunted growth, early menarche, and sperm changes, which are not necessarily associated with a disease phenotype, may be useful biomarkers of effect and help identify associations with environmental exposures. Another opportunity is that agnostic (non-hypothesis-driven) analytical approaches offered by tools characterized as “omics” may be able to elucidate the effect of mixtures and the component causes of disease.
Fortin reported her group discussed several best practices and potential solutions to the challenges outlined above. First is that to leverage all the data to discover and confirm associations, exposures, outcomes, and symptoms must be coded consistently. Further, technological advances, including machine learning and other data processing techniques, can
___________________
6 This section is based on a presentation by Marie Fortin, associate director of toxicology, Jazz Pharmaceuticals.
help with the deconvolution of massive datasets. An important consideration was that investigators likely need to consider how their science might be translated into policy.
The EPA has many opportunities to advance the use of biomarkers in research and decisions, said Fortin. It can support studies that measure exposures in federally qualified health centers and improve the representation of minority populations by providing incentives to participants and investigators. Often, people enrolled in health studies receive their clinical and environmental exposure results. It would benefit participants more if the tools and methods to report these data were culturally appropriate. The working group also discussed that the EPA could continue advancing the field with broader agnostic biomarker approaches to investigate disease clusters. International collaborators may have large disease registries and cohorts, which could help the EPA advance exposure science, but although some cohorts and registries may be more homogenous than the United States, Fortin added that the EPA could educate decision makers on the value and power of the new data processing and analytic techniques (such as multivariate analysis, machine learning) to improve their understanding and willingness to use big data.
This page intentionally left blank.
