Application of Artificial Intelligence and Machine Learning for Advancing Environmental Health Sciences (R41 Clinical Trial Not Allowed)

Background With the tremendous growth in data collection, coupled with advances in computing power and data accessibility, Artificial Intelligence (AI) and Machine Learning (ML) approaches are now broadly applied in many scientific disciplines. Applying AI and ML to the extensive toxicology data and environmental health sciences-related exposure data available in public resources and in published scientific literature has potential for better identifying hazardous substances that adversely affect human health and for preventing or reducing those exposures. A workshop in June 2019, held by the National Academies of Sciences, Engineering, and Medicine’s Standing Committee on the Use of Emerging Science for Environmental Health Decisions, explored leveraging AI and ML approaches to advance environmental health research and decisions, such as characterizing the sources of pollution, predicting chemical toxicity, estimating human exposures to contaminants, and identifying exposure-related health outcomes. The workshop addressed several fundamental questions including data availability, quality, bias, and uncertainty in the data used to develop machine learning algorithms. The workshop also discussed other barriers such as the lack of transparency and interpretability of AI systems that could impact reproducibility, replicability, accuracy of the results which ultimately would affect the trust in the quality of the methodologies and models. Despite these challenges, the workshop participants concluded that AI and ML have the potential to revolutionize environmental health and provide new avenues to address existing and emerging challenges. For example, for predicting toxicities of chemicals and for assessing the impact on human health often an understanding of the complex biological and environmental factors and how they work together is required. The ML approaches are ideally suited to model such complex interactions and can assess many such variables simultaneously to determine the risks and hazardous outcomes and help making better informed decisions. The brief proceedings of the workshop can be found at https://www.nap.edu/catalog/25520/leveraging-artificial-intelligence-and-machine-learning-to-advance-environmental-health-research-and-decisions. Objectives Through this Funding Opportunity Announcement (FOA), NIEHS is interested in supporting small business concerns (SBCs) to develop promising methodologies applying AI and ML approaches to advance environmental health research and decisions. The overall goal is to advance and adapt current AI and ML approaches by leveraging existing toxicity and environmental health data from published reports and public health records, including enhancing the accuracy of toxicity prediction or safety assessment, prioritizing chemicals for more comprehensive testing, identifying data or knowledge gaps in the field, and promoting novel approaches for exposure science such as estimating human exposures and health outcomes. The proposed approaches can focus on extracting and integrating information from environmental datasets or resources, developing algorithms and predictive models and applying those for predicting toxicity, and characterizing the biological responses or health consequences of chemical exposures. Some of the specific efforts encouraged through this FOA include: Data curation and integration Development of methods and resources for curating and integrating environmental health datasets. The source of the environmental health data could include one or more sources, such as the biomedical literature, established public datasets and other biochemical databases, high-throughput and high-content datasets, multi-omics datasets, legacy study reports or datasets, single-cell phenotyping and genetic diversity studies. Developing and automating tools for systematic reviews of environmental health scientific literature such as creating methods to map free text to standardized ontologies and facilitating data extraction from those sources and dataset creation. Incorporating details about specific chemical features and toxicity or physiological data points and improve data accessibility for computational, analysis and interpretation. Developing models and algorithms Develop and apply appropriate algorithms and predictive toxicology models on high-quality, curated environmental health datasets and apply machine learning approaches to enable data mining. Develop models to relate chemical structure or activity to biological and toxicity data to fill gaps, predict the potential toxicity of new compounds, inform mechanism based chemical risk-assessment, predictive exposure, predict chemical-physical properties or support untargeted metabolomics/ exposomics/ metabolism. Apply or develop in silico approaches to contextualize the exposure and toxic effects and provide mechanistic insights, such as Adverse Outcome Pathway (AOP) generation or predicting molecular targets, pathways or mechanisms. Predictive toxicology and interaction Use curated data, and algorithms or models for improved predictions of chemical toxicity and to validate the models for accuracy and reproducibility. Apply advanced in silico approaches for exposure science, such as improving the understanding of interactions between genes and environmental exposures, identifying differential susceptibility across human population, identifying biomarkers of exposure, and helping to characterize the exposome. Apply AI and ML approaches to develop better prediction models and systems for estimating exposures and thresholds, risk-assessment and for predicting the toxicity of chemical mixtures. Responsiveness Any datasets under consideration must be relevant to environmental health and can be related to chemical or drug toxicity. Approaches using datasets solely related to non-environmental health sciences such as general health sciences, ecotoxicology, or focused on pathogens rather than environmental chemicals or toxins (e.g. LPS, mycotoxins) will be nonresponsive. Applications focused on performing new studies involving laboratory or other toxicity testing either to generate new datasets and/or develop or validate models will also be considered non-responsive to this FOA. Additional Recommendations It is envisioned that whenever applicable, the FAIR guiding principles (findable, accessible, interoperable and reusable) are applied for data management. The proposed approaches can address more than one of the major areas described above. It is advised that appropriate steps/Indicators are provided to ensure high quality data are utilized and that this is also supported by an examination/evidence of uncertainties, data errors, sufficient sample sizes, and application of normalization and statistical approaches are in place. Expectation is that a multi-disciplinary team would be employed in the development of the AI/ML to further modify, adapt or extend existing methods and create new and more robust approaches. The team may include environmental toxicologists, epidemiologists, biostatisticians, data scientists and computational scientists. It is recommended that appropriate steps are taken to identify, mitigate, and account for biases of the datasets used and machine learning models developed and state any limitations and alternatives.