Thomas Hartung
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Thomas Hartung has a total of 86 co-authors. Here are the top co-authors by number of shared publications:
Albrecht Wendel
Biochemical Pharmacology, Department of Biology, University of Konstanz, D-78457 Konstanz,1 and Untere Hofstatt 3, D-76703 Kraichtal,2 Germany
11 shared publicationsIsabel Diterich
University of Konstanz, Biochemical Pharmacology, Konstanz, Germany,1 Swiss Institute of Allergy and Asthma Research, Daros, Switzerland,2 Untere Hofstaat 3, Kraichtal, Germany,3 ECVAM, EU Joint Research Centre, IHCP, Ispra, Italy4
6 shared publicationsCarolin Rauter
University of Konstanz, Biochemical Pharmacology, Konstanz, Germany,1 Swiss Institute of Allergy and Asthma Research, Daros, Switzerland,2 Untere Hofstaat 3, Kraichtal, Germany,3 ECVAM, EU Joint Research Centre, IHCP, Ispra, Italy4
5 shared publicationsCorinna Hermann
University of Konstanz, Biochemical Pharmacology, Konstanz, Germany,1 Swiss Institute of Allergy and Asthma Research, Daros, Switzerland,2 Untere Hofstaat 3, Kraichtal, Germany,3 ECVAM, EU Joint Research Centre, IHCP, Ispra, Italy4
5 shared publicationsSiegfried Morath
Biochemical Pharmacology, University of Konstanz, Germany
5 shared publicationsArmin Geyer
Institute for Organic Chemistry, University of Regensburg, 93040 Regensburg, Germany
3 shared publications
Latest Publications
- Open Access
- 6 October 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1921273
Towards the One Exposome: harnessing artificial intelligence and digital health technologies across One Health toxicology, new approach methodologies, and environmental public health
The One Exposome framework integrates exposomics, One Health toxicology, New Approach Methodologies (NAMs) and artificial intelligence to advance environmental health from reactive risk assessment towards predictive, preventive, and participatory science. The paper positions One Health as the ethical and conceptual rationale for recognizing the interdependence of human, animal, and environmental health, the Exposome as the framework for measuring cumulative and life-course exposures, and NAMs together with digital and AI-enabled tools as the means to generate, integrate, and interpret mechanistic and adverse outcome pathway (AOP)-based evidence across biological scales. In this review, environmental health challenges such as antimicrobial resistance, endocrine disruption, and other complex multi-stressor exposures cannot be adequately addressed through compartmentalized, single-substance, or purely retrospective approaches. The One Exposome paradigm instead proposes a systems-based model that links exposure data, mechanistic toxicology, and public health intelligence across humans, animals, and ecosystems. It draws on human-relevant NAMs, including microphysiological systems, together with AI and multi-omics approaches, to support cross-species interpretation, earlier hazard identification, and more preventive intervention strategies. Within this framework, sentinel surveillance, integrated monitoring, and interoperable data systems are central for detecting shared risks across environmental, veterinary, food, and human health domains. The approach is also presented as requiring governance reform so that agricultural practice, environmental management, and clinical stewardship are aligned within a common preventive and planetary health perspective. Overall, the paper frames the One Exposome as a transdisciplinary architecture for environmental public health that combines measurable environments, mechanistic understanding, and computational integration. By connecting cumulative exposures to biological responses and population-level consequences, it offers a foundation for more anticipatory, equitable, and scientifically integrated environmental health decision-making.
- 6 October 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1921273
Open AccessTowards the One Exposome: harnessing artificial intelligence and digital health technologies across One Health toxicology, new approach methodologies, and environmental public health
The One Exposome framework integrates exposomics, One Health toxicology, New Approach Methodologies (NAMs) and artificial intelligence to advance environmental health from reactive risk assessment towards predictive, preventive, and participatory science. The paper positions One Health as the ethical and conceptual rationale for recognizing the interdependence of human, animal, and environmental health, the Exposome as the framework for measuring cumulative and life-course exposures, and NAMs together with digital and AI-enabled tools as the means to generate, integrate, and interpret mechanistic and adverse outcome pathway (AOP)-based evidence across biological scales. In this review, environmental health challenges such as antimicrobial resistance, endocrine disruption, and other complex multi-stressor exposures cannot be adequately addressed through compartmentalized, single-substance, or purely retrospective approaches. The One Exposome paradigm instead proposes a systems-based model that links exposure data, mechanistic toxicology, and public health intelligence across humans, animals, and ecosystems. It draws on human-relevant NAMs, including microphysiological systems, together with AI and multi-omics approaches, to support cross-species interpretation, earlier hazard identification, and more preventive intervention strategies. Within this framework, sentinel surveillance, integrated monitoring, and interoperable data systems are central for detecting shared risks across environmental, veterinary, food, and human health domains. The approach is also presented as requiring governance reform so that agricultural practice, environmental management, and clinical stewardship are aligned within a common preventive and planetary health perspective. Overall, the paper frames the One Exposome as a transdisciplinary architecture for environmental public health that combines measurable environments, mechanistic understanding, and computational integration. By connecting cumulative exposures to biological responses and population-level consequences, it offers a foundation for more anticipatory, equitable, and scientifically integrated environmental health decision-making.
- Open Access
- 30 July 2026
- Published by Springer Nature in npj Science of Food
- Vol. 10 (1)
Leveraging artificial intelligence for mycotoxin management in food systems
AI-driven methods have been applied across mycotoxin research in detection, prediction, and control, offering high laboratory accuracy ( > 90%), robust forecasting (75–99%), and emerging mitigation strategies (80–86%). Integrating AI with multi-omics data, mixture toxicity modeling, and standardized protocols promises to enhance food safety, streamline regulatory decision-making, and reduce animal testing.
- 30 July 2026
- Published by Springer Nature in npj Science of Food
- Vol. 10 (1)
Open AccessLeveraging artificial intelligence for mycotoxin management in food systems
AI-driven methods have been applied across mycotoxin research in detection, prediction, and control, offering high laboratory accuracy ( > 90%), robust forecasting (75–99%), and emerging mitigation strategies (80–86%). Integrating AI with multi-omics data, mixture toxicity modeling, and standardized protocols promises to enhance food safety, streamline regulatory decision-making, and reduce animal testing.
- Open Access
- 22 May 2026
- Published by Frontiers Media SA in Frontiers in Chemistry
- Vol. 14, 1801623
Green toxicology only becomes beautiful through AI
Green Toxicology extends the principles of Green Chemistry by embedding toxicological foresight into chemical design, with the aim of preventing hazards before substances reach markets or the environment. Its conceptual pillars—prevention, precaution, life-cycle thinking, and avoidance of regrettable substitutions—align closely with sustainability agendas such as the European Green Deal and the United Nations Sustainable Development Goals. Despite its promise, Green Toxicology has remained largely aspirational, limited by fragmented data, slow regulatory uptake, and reliance on new approach methodologies (NAMs) that still face validation and acceptance hurdles. Artificial intelligence (AI) offers a transformative solution by integrating heterogeneous datasets, enhancing predictive accuracy, and enabling probabilistic risk assessment. Deep learning, natural language processing, and explainable AI can mine legacy studies, link adverse outcome pathways, and design safer chemistries proactively. Coupled with microphysiological systems and omics, AI makes Green Toxicology predictive, human-relevant, and scalable. Together, they form a practical framework for guiding chemical innovation toward sustainability, reconciling industrial productivity with ecological integrity and public health protection.
- 22 May 2026
- Published by Frontiers Media SA in Frontiers in Chemistry
- Vol. 14, 1801623
Open AccessGreen toxicology only becomes beautiful through AI
Green Toxicology extends the principles of Green Chemistry by embedding toxicological foresight into chemical design, with the aim of preventing hazards before substances reach markets or the environment. Its conceptual pillars—prevention, precaution, life-cycle thinking, and avoidance of regrettable substitutions—align closely with sustainability agendas such as the European Green Deal and the United Nations Sustainable Development Goals. Despite its promise, Green Toxicology has remained largely aspirational, limited by fragmented data, slow regulatory uptake, and reliance on new approach methodologies (NAMs) that still face validation and acceptance hurdles. Artificial intelligence (AI) offers a transformative solution by integrating heterogeneous datasets, enhancing predictive accuracy, and enabling probabilistic risk assessment. Deep learning, natural language processing, and explainable AI can mine legacy studies, link adverse outcome pathways, and design safer chemistries proactively. Coupled with microphysiological systems and omics, AI makes Green Toxicology predictive, human-relevant, and scalable. Together, they form a practical framework for guiding chemical innovation toward sustainability, reconciling industrial productivity with ecological integrity and public health protection.
- Open Access
- 1 May 2026
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 8, 1790669
Next generation validation for next generation risk assessment
Next generation risk assessment (NGRA) demands a fundamental transformation in how toxicological test methods are validated. Traditional validation approaches, designed for animal tests and mainly for simple in vitro methods, are increasingly inadequate for evaluating complex New Approach Methodologies (NAMs), artificial intelligence (AI)-based approaches, and integrated testing strategies (ITS). This paper presents a comprehensive framework for “ next-generation validation ” that leverages artificial intelligence and modern computational capabilities to create more efficient, thorough, and dynamic validation processes. The proposed framework emphasizes human relevance over simple concordance with animal data and emphasizes key innovations including e-validation, mechanistic validation, and post-validation companion AI agents. Because AI can inherit biases, obscure failure modes, and drift over time, the framework treats AI as both a tool for, and a subject of, validation, requiring transparent performance criteria, uncertainty quantification, and explicit governance for model updates and lifecycle monitoring. To make the framework actionable, we define a method as “ NGV-validated for a stated context of use ” when it meets pre-specified acceptance criteria across five domains, i.e., technical reliability, biological relevance, predictive performance, uncertainty quantification, and data integrity, supported by defined governance roles, version control, and lifecycle re-review triggers. e-validation employs sophisticated algorithms for reference chemical selection, study simulation, and continuous performance monitoring, while mechanistic validation evaluates whether methods accurately capture relevant biological pathways and mechanisms of toxicity. The paper addresses critical implementation challenges including data quality standardization, regulatory acceptance, and international harmonization, providing specific recommendations for various stakeholders. Looking forward, validation will increasingly embrace dynamic, adaptive approaches that evolve alongside scientific understanding and technological capabilities. The integration of artificial intelligence will enhance analysis of complex data, enable real-time monitoring of method performance, and support more sophisticated uncertainty quantification. Success in this transformation requires coordinated effort across regulatory agencies, industry partners, and academic institutions. In summary, this paper emphasizes a five-pillar framework integrating mechanistic, probabilistic, and AI-driven elements to reform toxicological validation. The proposed framework, exemplified here for tests for developmental neurotoxicants and virtual control groups, represents a crucial step toward more efficient and accurate chemical safety assessment while maintaining necessary standards for public health protection.
- 1 May 2026
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 8, 1790669
Open AccessNext generation validation for next generation risk assessment
Next generation risk assessment (NGRA) demands a fundamental transformation in how toxicological test methods are validated. Traditional validation approaches, designed for animal tests and mainly for simple in vitro methods, are increasingly inadequate for evaluating complex New Approach Methodologies (NAMs), artificial intelligence (AI)-based approaches, and integrated testing strategies (ITS). This paper presents a comprehensive framework for “ next-generation validation ” that leverages artificial intelligence and modern computational capabilities to create more efficient, thorough, and dynamic validation processes. The proposed framework emphasizes human relevance over simple concordance with animal data and emphasizes key innovations including e-validation, mechanistic validation, and post-validation companion AI agents. Because AI can inherit biases, obscure failure modes, and drift over time, the framework treats AI as both a tool for, and a subject of, validation, requiring transparent performance criteria, uncertainty quantification, and explicit governance for model updates and lifecycle monitoring. To make the framework actionable, we define a method as “ NGV-validated for a stated context of use ” when it meets pre-specified acceptance criteria across five domains, i.e., technical reliability, biological relevance, predictive performance, uncertainty quantification, and data integrity, supported by defined governance roles, version control, and lifecycle re-review triggers. e-validation employs sophisticated algorithms for reference chemical selection, study simulation, and continuous performance monitoring, while mechanistic validation evaluates whether methods accurately capture relevant biological pathways and mechanisms of toxicity. The paper addresses critical implementation challenges including data quality standardization, regulatory acceptance, and international harmonization, providing specific recommendations for various stakeholders. Looking forward, validation will increasingly embrace dynamic, adaptive approaches that evolve alongside scientific understanding and technological capabilities. The integration of artificial intelligence will enhance analysis of complex data, enable real-time monitoring of method performance, and support more sophisticated uncertainty quantification. Success in this transformation requires coordinated effort across regulatory agencies, industry partners, and academic institutions. In summary, this paper emphasizes a five-pillar framework integrating mechanistic, probabilistic, and AI-driven elements to reform toxicological validation. The proposed framework, exemplified here for tests for developmental neurotoxicants and virtual control groups, represents a crucial step toward more efficient and accurate chemical safety assessment while maintaining necessary standards for public health protection.
- Open Access
- 10 April 2026
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 8
Towards learning and memory risk assessment with human brain organoids: barriers and opportunities
Neurodevelopmental conditions, including autism spectrum disorder, intellectual disability, and learning disabilities, as well as neurodegenerative disorders, affect millions of people in the United States alone. Both genetic and environmental factors contribute to their onset, yet traditional neurotoxicity testing often fails to identify specific risks or mechanisms underlying cognitive impairment. Human brain organoids (hBOs), also called neural organoids or brain microphysiological systems, are three-dimensional (3D) stem cell-derived models that recapitulate key features of human brain development and offer greater physiological relevance than traditional 2D in vitro or animal models. The emerging field of “organoid intelligence” integrates these systems with advanced bioengineering and artificial intelligence to model higher-order neural functions and assess learning and memory-relevant endpoints that were previously less explored in vitro . Despite this promise, we have identified four key barriers that hinder the application of hBOs for the hazard identification phase of functional neurotoxic risk assessments: [1] limited maturity and regional complexity, [2] lack of high-throughput defined procedures for assessing cognitive development and function in vitro , [3] limited standardization for reproducibility, and [4] challenges in translating in vitro results to human health outcomes. Here, we outline current efforts to overcome these challenges, i.e., scientific, technical, and regulatory advances. We also illustrate how hBO-based assays can be applied to advance both mechanistic understanding and the regulatory evaluation of environmental (developmental) neurotoxicants, using heavy metals as a model.
- 10 April 2026
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 8
Open AccessTowards learning and memory risk assessment with human brain organoids: barriers and opportunities
Neurodevelopmental conditions, including autism spectrum disorder, intellectual disability, and learning disabilities, as well as neurodegenerative disorders, affect millions of people in the United States alone. Both genetic and environmental factors contribute to their onset, yet traditional neurotoxicity testing often fails to identify specific risks or mechanisms underlying cognitive impairment. Human brain organoids (hBOs), also called neural organoids or brain microphysiological systems, are three-dimensional (3D) stem cell-derived models that recapitulate key features of human brain development and offer greater physiological relevance than traditional 2D in vitro or animal models. The emerging field of “organoid intelligence” integrates these systems with advanced bioengineering and artificial intelligence to model higher-order neural functions and assess learning and memory-relevant endpoints that were previously less explored in vitro . Despite this promise, we have identified four key barriers that hinder the application of hBOs for the hazard identification phase of functional neurotoxic risk assessments: [1] limited maturity and regional complexity, [2] lack of high-throughput defined procedures for assessing cognitive development and function in vitro , [3] limited standardization for reproducibility, and [4] challenges in translating in vitro results to human health outcomes. Here, we outline current efforts to overcome these challenges, i.e., scientific, technical, and regulatory advances. We also illustrate how hBO-based assays can be applied to advance both mechanistic understanding and the regulatory evaluation of environmental (developmental) neurotoxicants, using heavy metals as a model.
- Open Access
- 5 March 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1775284
AI redefine untargeted metabolomics: estimating chemical amounts for a Human Exposome Project
The Human Exposome Project aims to map the totality of environmental exposures, but its success relies on transforming qualitative detections into quantitative data. Following our review on AI-driven metabolite identification, this second installment addresses the next critical bottleneck: estimating chemical concentrations in untargeted metabolomics without authentic standards. Translating LC-HRMS signal intensities into absolute concentrations is hindered by the vast variability in ionization efficiency and matrix effects, particularly for xenobiotics where reference standards are unavailable. We review emerging strategies that leverage artificial intelligence—ranging from descriptor-based regression to deep learning on molecular point clouds—to predict ionization response factors. We further evaluate a “matrix-embedded” calibration approach that utilizes ubiquitous endogenous metabolites (e.g., amino acids, lipids) as internal anchors to normalize response scales across studies. These innovations enable “tiered semi-quantification,” allowing the classification of exposures into biologically relevant ranges (e.g., nanomolar vs. micromolar). This stratification facilitates direct integration with toxicological frameworks, such as the Threshold of Toxicological Concern (TTC) and high-throughput bioactivity data (e.g., ToxCast), for rapid risk prioritization. By integrating quantitative AI prediction models with robust quality assurance, untargeted metabolomics can evolve from a qualitative discovery tool into a quantitative engine for exposure science, providing the necessary evidence to link complex chemical exposures to human health outcomes.
- 5 March 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1775284
Open AccessAI redefine untargeted metabolomics: estimating chemical amounts for a Human Exposome Project
The Human Exposome Project aims to map the totality of environmental exposures, but its success relies on transforming qualitative detections into quantitative data. Following our review on AI-driven metabolite identification, this second installment addresses the next critical bottleneck: estimating chemical concentrations in untargeted metabolomics without authentic standards. Translating LC-HRMS signal intensities into absolute concentrations is hindered by the vast variability in ionization efficiency and matrix effects, particularly for xenobiotics where reference standards are unavailable. We review emerging strategies that leverage artificial intelligence—ranging from descriptor-based regression to deep learning on molecular point clouds—to predict ionization response factors. We further evaluate a “matrix-embedded” calibration approach that utilizes ubiquitous endogenous metabolites (e.g., amino acids, lipids) as internal anchors to normalize response scales across studies. These innovations enable “tiered semi-quantification,” allowing the classification of exposures into biologically relevant ranges (e.g., nanomolar vs. micromolar). This stratification facilitates direct integration with toxicological frameworks, such as the Threshold of Toxicological Concern (TTC) and high-throughput bioactivity data (e.g., ToxCast), for rapid risk prioritization. By integrating quantitative AI prediction models with robust quality assurance, untargeted metabolomics can evolve from a qualitative discovery tool into a quantitative engine for exposure science, providing the necessary evidence to link complex chemical exposures to human health outcomes.
- Open Access
- 17 December 2025
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 7, 1690491
Mechanistic read-across comes of age: a comparative appraisal of EFSA 2025 guidance, ECHA’s RAAF, and good read-across practice
Read-across has matured from an expert-driven extrapolation based largely on structural analogy into a rigorously documented, mechanistically informed cornerstone of next-generation risk assessment. Three pivotal frameworks are compared that now shape its regulatory use: the European Food Safety Authority’s (EFSA) 2025 guidance for food and feed safety, the European Chemicals Agency’s (ECHA) Read-Across Assessment Framework (RAAF) for industrial chemicals under REACH, and the community-driven Good Read-Across Practice (GRAP) principles. Using five analytical lenses—conceptual structure, scientific rigor, implementation tools, regulatory acceptance, and practical impact—we identified areas of complementarity and divergence. EFSA provides a seven-step, uncertainty-anchored workflow that actively embeds new approach methodologies (NAMs) and adverse outcome pathway reasoning, offering applicants a transparent “how-to” template. RAAF, in contrast, operates as an evaluator’s rubric: six scenario types and associated assessment elements delineate what evidence must be delivered, thereby standardizing regulatory scrutiny but leaving dossier construction to the registrant. GRAP supplies the conceptual glue, emphasizing mechanistic plausibility, exhaustive analogue selection, explicit uncertainty characterization, and the strategic use of NAMs; its influence is evident in both EFSA’s and ECHA’s evolving expectations. (Terminology note: the acronym “NAM” was popularized at an ECHA workshop in 2016; earlier documents such as RAAF and initial GRAP papers therefore may not use the term explicitly). Regulatory experience under REACH demonstrates that dossier quality and acceptance rates rise markedly when RAAF criteria are met, while EFSA’s new guidance is poised to catalyze similar gains in food and feed assessments. Globally, the convergence of these frameworks—reinforced by OECD initiatives and NAM-enhanced case studies—signals an emerging international consensus on what constitutes defensible read-across. In conclusion, harmonizing EFSA’s procedural roadmap with RAAF’s evaluative rigor and GRAP’s best-practice ethos can mainstream reliable, animal-saving read-across across regulatory domains, paving the way for fully mechanistic, AI-enabled chemical safety assessment.
- 17 December 2025
- Published by Frontiers Media SA in Frontiers in Toxicology
- Vol. 7, 1690491
Open AccessMechanistic read-across comes of age: a comparative appraisal of EFSA 2025 guidance, ECHA’s RAAF, and good read-across practice
Read-across has matured from an expert-driven extrapolation based largely on structural analogy into a rigorously documented, mechanistically informed cornerstone of next-generation risk assessment. Three pivotal frameworks are compared that now shape its regulatory use: the European Food Safety Authority’s (EFSA) 2025 guidance for food and feed safety, the European Chemicals Agency’s (ECHA) Read-Across Assessment Framework (RAAF) for industrial chemicals under REACH, and the community-driven Good Read-Across Practice (GRAP) principles. Using five analytical lenses—conceptual structure, scientific rigor, implementation tools, regulatory acceptance, and practical impact—we identified areas of complementarity and divergence. EFSA provides a seven-step, uncertainty-anchored workflow that actively embeds new approach methodologies (NAMs) and adverse outcome pathway reasoning, offering applicants a transparent “how-to” template. RAAF, in contrast, operates as an evaluator’s rubric: six scenario types and associated assessment elements delineate what evidence must be delivered, thereby standardizing regulatory scrutiny but leaving dossier construction to the registrant. GRAP supplies the conceptual glue, emphasizing mechanistic plausibility, exhaustive analogue selection, explicit uncertainty characterization, and the strategic use of NAMs; its influence is evident in both EFSA’s and ECHA’s evolving expectations. (Terminology note: the acronym “NAM” was popularized at an ECHA workshop in 2016; earlier documents such as RAAF and initial GRAP papers therefore may not use the term explicitly). Regulatory experience under REACH demonstrates that dossier quality and acceptance rates rise markedly when RAAF criteria are met, while EFSA’s new guidance is poised to catalyze similar gains in food and feed assessments. Globally, the convergence of these frameworks—reinforced by OECD initiatives and NAM-enhanced case studies—signals an emerging international consensus on what constitutes defensible read-across. In conclusion, harmonizing EFSA’s procedural roadmap with RAAF’s evaluative rigor and GRAP’s best-practice ethos can mainstream reliable, animal-saving read-across across regulatory domains, paving the way for fully mechanistic, AI-enabled chemical safety assessment.
- Top Cited PapersOpen Access
- 5 December 2025
- Published by Springer Nature in Current Environmental Health Reports
- Vol. 12 (1) , 51
Navigating the AI Frontier in Toxicology: Trends, Trust, and Transformation
Purpose of Review: The integration of artificial intelligence (AI) into toxicology marks a profound paradigm shift in chemical safety science. No longer limited to automating traditional workflows, AI is redefining how we assess risk, interpret complex biological data, and inform regulatory decision-making. This article explores the convergence of AI and other new approach methodologies (NAMs), emphasizing key trends such as multimodal learning, causal inference, explainable AI (xAI), generative modeling, and federated learning. Recent Findings: These technologies enable more human-relevant, mechanistically grounded, and ethically aligned toxicological predictions—surpassing the reproducibility and scalability of animal-based methods. However, the dynamic nature of AI models challenges traditional validation paradigms. To address this, we introduced the e-validation framework, which operationalizes the TREAT principles (Trustworthiness, Reproducibility, Explainability, Applicability, Transparency) and incorporates AI-powered modules for reference chemical selection, virtual study simulation, mechanistic cross-validation, and post-validation surveillance through companion agents. Ethical considerations—including bias audits, equity audits, and participatory governance—are also foregrounded as critical elements for responsible AI adoption. The emergence of a co-pilot model, where AI augments but does not replace human judgment, offers a pragmatic path forward. Supported by evidence from the 2025 Stanford AI Index and recent regulatory advances, we argue that the infrastructure, economics, and policy momentum are now aligned for global-scale deployment of AI-based toxicology. Summary: The future of the field lies not in replicating legacy practices, but in reinventing toxicology as an adaptive, transparent, and ethically grounded science that delivers more accurate, inclusive, and human-centric safety assessments. Lay Summary: Artificial intelligence (AI) is changing how we test chemicals for safety. Instead of using animals, new computer-based tools can predict how substances affect human health more quickly, accurately, and ethically. This article looks at how these technologies—like smart data systems, models that explain their reasoning, and even AI "agents" that run simulations—can improve toxicology. We also introduce a new idea called "e-validation", which uses AI to help validate these methods in real-time, not just once. This ensures the models stay up to date and reliable. But using AI safely means tackling big questions: Can we trust results we don't fully understand? How do we prevent unfairness or bias in the data? We suggest a "co-pilot" model, where AI supports, but doesn't replace, human experts. With better data sharing, strong ethics, and smarter oversight, AI can help make chemical safety testing more human-focused, fair, and effective.
- 5 December 2025
- Published by Springer Nature in Current Environmental Health Reports
- Vol. 12 (1) , 51
Top Cited PapersOpen AccessNavigating the AI Frontier in Toxicology: Trends, Trust, and Transformation
Purpose of Review: The integration of artificial intelligence (AI) into toxicology marks a profound paradigm shift in chemical safety science. No longer limited to automating traditional workflows, AI is redefining how we assess risk, interpret complex biological data, and inform regulatory decision-making. This article explores the convergence of AI and other new approach methodologies (NAMs), emphasizing key trends such as multimodal learning, causal inference, explainable AI (xAI), generative modeling, and federated learning. Recent Findings: These technologies enable more human-relevant, mechanistically grounded, and ethically aligned toxicological predictions—surpassing the reproducibility and scalability of animal-based methods. However, the dynamic nature of AI models challenges traditional validation paradigms. To address this, we introduced the e-validation framework, which operationalizes the TREAT principles (Trustworthiness, Reproducibility, Explainability, Applicability, Transparency) and incorporates AI-powered modules for reference chemical selection, virtual study simulation, mechanistic cross-validation, and post-validation surveillance through companion agents. Ethical considerations—including bias audits, equity audits, and participatory governance—are also foregrounded as critical elements for responsible AI adoption. The emergence of a co-pilot model, where AI augments but does not replace human judgment, offers a pragmatic path forward. Supported by evidence from the 2025 Stanford AI Index and recent regulatory advances, we argue that the infrastructure, economics, and policy momentum are now aligned for global-scale deployment of AI-based toxicology. Summary: The future of the field lies not in replicating legacy practices, but in reinventing toxicology as an adaptive, transparent, and ethically grounded science that delivers more accurate, inclusive, and human-centric safety assessments. Lay Summary: Artificial intelligence (AI) is changing how we test chemicals for safety. Instead of using animals, new computer-based tools can predict how substances affect human health more quickly, accurately, and ethically. This article looks at how these technologies—like smart data systems, models that explain their reasoning, and even AI "agents" that run simulations—can improve toxicology. We also introduce a new idea called "e-validation", which uses AI to help validate these methods in real-time, not just once. This ensures the models stay up to date and reliable. But using AI safely means tackling big questions: Can we trust results we don't fully understand? How do we prevent unfairness or bias in the data? We suggest a "co-pilot" model, where AI supports, but doesn't replace, human experts. With better data sharing, strong ethics, and smarter oversight, AI can help make chemical safety testing more human-focused, fair, and effective.
- Open Access
- 12 November 2025
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 13, 1687056
AI redefines mass spectrometry chemicals identification: retention time prediction in metabolomics and for a Human Exposome Project
The comprehensive identification of environmental and endogenous chemicals in human biospecimens is a critical bottleneck for realizing the Human Exposome Project. Untargeted metabolomics, particularly liquid chromatography–high-resolution mass spectrometry (LC–HRMS), offers unparalleled coverage of small molecules, but most detected features remain unidentified due to limited spectral libraries and structural ambiguity. Retention time (RT) prediction—based on quantitative structure–retention relationships (QSRR) and enhanced by artificial intelligence (AI)—is an underutilized orthogonal parameter that can substantially improve metabolite annotation confidence. This review synthesizes advances in machine learning–based RT prediction, probabilistic calibration, and cross-platform harmonization for liquid chromatography and gas chromatography, including deep learning, graph neural networks, and transfer learning approaches. We evaluate workflows integrating RT prediction with mass-based searches and network-based annotation tools, highlighting their potential to refine candidate ranking and reduce false positives in environmental exposure assessment. The use of endogenous compounds as internal calibrants is discussed as a practical strategy for improving RT transferability across laboratories. We further outline how RT-aware annotation supports non-targeted screening of emerging contaminants, transformation products, and exposure biomarkers, thereby enhancing the interpretability and reproducibility of exposomics data. By integrating RT prediction, QSRR modeling, and AI into untargeted metabolomics pipelines, researchers can move from qualitative detection toward quantitative, inference-driven mapping of environmental influences on human health, strengthening the scientific foundation for environmental health policy and preventive public health strategies.
- 12 November 2025
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 13, 1687056
Open AccessAI redefines mass spectrometry chemicals identification: retention time prediction in metabolomics and for a Human Exposome Project
The comprehensive identification of environmental and endogenous chemicals in human biospecimens is a critical bottleneck for realizing the Human Exposome Project. Untargeted metabolomics, particularly liquid chromatography–high-resolution mass spectrometry (LC–HRMS), offers unparalleled coverage of small molecules, but most detected features remain unidentified due to limited spectral libraries and structural ambiguity. Retention time (RT) prediction—based on quantitative structure–retention relationships (QSRR) and enhanced by artificial intelligence (AI)—is an underutilized orthogonal parameter that can substantially improve metabolite annotation confidence. This review synthesizes advances in machine learning–based RT prediction, probabilistic calibration, and cross-platform harmonization for liquid chromatography and gas chromatography, including deep learning, graph neural networks, and transfer learning approaches. We evaluate workflows integrating RT prediction with mass-based searches and network-based annotation tools, highlighting their potential to refine candidate ranking and reduce false positives in environmental exposure assessment. The use of endogenous compounds as internal calibrants is discussed as a practical strategy for improving RT transferability across laboratories. We further outline how RT-aware annotation supports non-targeted screening of emerging contaminants, transformation products, and exposure biomarkers, thereby enhancing the interpretability and reproducibility of exposomics data. By integrating RT prediction, QSRR modeling, and AI into untargeted metabolomics pipelines, researchers can move from qualitative detection toward quantitative, inference-driven mapping of environmental influences on human health, strengthening the scientific foundation for environmental health policy and preventive public health strategies.
- Open Access
- 10 September 2025
- Published by Frontiers Media SA in Frontiers in Artificial Intelligence
- Vol. 8, 1632520
AI: the Apollo guidance computer of the Exposome moonshot
The Exposome—the totality of environmental exposures across a lifetime—remains one of the most significant challenges in understanding and preventing human disease. Translating its vast, heterogeneous data streams into actionable knowledge requires artificial intelligence (AI) integrated with human-relevant experimental systems. We propose a unifying vision in which Microphysiological Systems (MPS) and multi-omics platforms generate high-quality, context-specific data that iteratively calibrate AI models, enabling the creation of digital twins of organs, individuals, and ultimately populations. This “Exposome Moonshot” parallels the Apollo program in ambition, with MPS as the rocket, multi-omics as the lunar module, and AI as the guidance computer. Early applications demonstrate that deep learning can already outperform canonical animal tests for several toxicological endpoints, while reducing cost and time to decision. Realizing the full potential of Exposome intelligence will require expanding the applicability domain of models, implementing robust data security, and prioritizing transparent, interpretable algorithms. By linking predictive AI with experimental feedback, we can move toward a prevention-driven, personalized paradigm for human health and regulatory science.
- 10 September 2025
- Published by Frontiers Media SA in Frontiers in Artificial Intelligence
- Vol. 8, 1632520
Open AccessAI: the Apollo guidance computer of the Exposome moonshot
The Exposome—the totality of environmental exposures across a lifetime—remains one of the most significant challenges in understanding and preventing human disease. Translating its vast, heterogeneous data streams into actionable knowledge requires artificial intelligence (AI) integrated with human-relevant experimental systems. We propose a unifying vision in which Microphysiological Systems (MPS) and multi-omics platforms generate high-quality, context-specific data that iteratively calibrate AI models, enabling the creation of digital twins of organs, individuals, and ultimately populations. This “Exposome Moonshot” parallels the Apollo program in ambition, with MPS as the rocket, multi-omics as the lunar module, and AI as the guidance computer. Early applications demonstrate that deep learning can already outperform canonical animal tests for several toxicological endpoints, while reducing cost and time to decision. Realizing the full potential of Exposome intelligence will require expanding the applicability domain of models, implementing robust data security, and prioritizing transparent, interpretable algorithms. By linking predictive AI with experimental feedback, we can move toward a prevention-driven, personalized paradigm for human health and regulatory science.