Advancing the implementation of artificial intelligence in regulatory frameworks for chemical safety assessment by defining robust readiness criteria

The integration of artificial intelligence (AI) into chemical risk assessment (CRA) is emerging as a powerful approach to enhance the interpretation of complex toxicological data and accelerate safety evaluations. However, the regulatory uptake of AI remains limited due to concerns about transparency, explainability, and trustworthiness. The European Partnership for the Assessment of Risks from Chemicals (PARC) project ReadyAI was established to address these challenges by developing a readiness scoring system to evaluate the maturity and regulatory applicability of AI-based models in CRA. The project unites a multidisciplinary consortium of academic, regulatory, and legal experts to define transparent and reproducible criteria encompassing data curation, model development, validation, explainability, and uncertainty quantification. Current efforts focus on identifying key priorities, including harmonized terminology, rigorous data quality standards, case studies, and targeted training of regulatory scientists. ReadyAI aims to deliver a practical, evidence-based scoring system that enables regulators to assess whether AI tools are sufficiently reliable for decision-making and guides developers toward compliance with regulatory expectations. By bridging the gap between AI innovation and regulatory applicability, ReadyAI contributes to the responsible integration of AI into chemical safety assessment frameworks, ultimately supporting human and environmental health protection.

1 Introduction

Recent advances in artificial intelligence (AI), encompassing machine learning and deep learning, are leading to a new era of data-driven innovation. These computational approaches are increasingly recognized as powerful tools across numerous disciplines. In regulatory toxicology, a field dedicated to protecting human health from adverse effects of chemical exposure, AI might offer significant support. This is especially true given the need to evaluate large volumes of datasets in the chemical risk assessment (CRA) process. Several recent publications highlighted the potential application of AI in safety assessment (Wittwehr et al., 2020; Lin and Chou, 2022; Wassenaar et al., 2024; Hartung and Kleinstreuer, 2025), emphasizing the importance of building trust in these technologies to facilitate their regulatory implementation into CRA. As a cornerstone of public health protection, CRA assesses chemical risk by integrating toxicological hazard data and exposure scenarios, within a framework of regulatory requirements. Its goal is to prevent environmental and human health adverse outcomes (e.g., acute/chronic diseases or developmental and reproductive disorders) associated with chemicals present in the environment or food (e.g., contaminants) or used in consumer products, pharmaceuticals, agriculture, and industrial processes.

The CRA process is currently undergoing an international paradigm shift moving from relying on endpoint measurements in animal studies toward human-relevant, mechanism-based, animal-free, and data-driven, integrating testing approaches using New Approach Methodologies (NAMs) (Schmeisser et al., 2023). This evolution entails not only the adoption of novel data generation techniques but also the development of fit-for-purpose frameworks for data evaluation, integration, and analysis. Within these frameworks, all available data are considered, provided they meet defined quality and validity criteria. However, the manual identification, retrieval, evaluation, and curation of such data are highly time-consuming. As a result, AI-driven computational approaches in CRA are increasingly recognized as powerful tools for various applications in toxicology, including quantitative structure–activity relationships (QSARs), exposure modeling, read-across, systematic reviews, and evidence integration for adverse outcome pathways development. Supported by AI, researchers, safety assessment scientists, and regulatory authorities, could work toward increasing accuracy, efficiency, time- and cost-effectiveness of assessment practices while decreasing reliance on in vivo animal testing (Blümmel et al., 2024). The latter aligns with worldwide trends on animal experimentation phaseout, announced by the European Commision (2025), United States Food and Drug Administration (2025), and United States National Institutes of Health (2025).

Similarly to several NAMs, the regulatory uptake of AI in toxicological sciences remains limited; there is no established regulatory requirements for incorporating AI-based tools into CRA, despite its growing number of potential use cases. The responsible regulatory uptake of AI is a topic of international discussion, supported by initiatives such as the Global Summit on Regulatory Science and the Global Coalition for Regulatory Science Research (2025). The enthusiastic discussion around the implementation of AI in CRA is now sharing space with debates regarding the uncertainties and transparency of AI. The integration of AI into regulatory frameworks, supporting regulatory decision-making, requires that these systems demonstrate trustworthiness, explainability and reliability. Moreover, regulatory scientists need adequate guidance and training to understand the principles, benefits, and limitations of AI algorithms. This ensures that regulatory scientists can critically evaluate AI-generated outputs and make well-informed decisions based on them. Ultimately, achieving a balance of human-AI interaction, responsibility, and accountability will be crucial for the successful integration of AI into regulatory decision-making (Hartung et al., 2025).

In this context, and as part of the European Union (EU) Partnership for the Assessment of Risks from Chemicals (PARC) initiative (Work Package 6, project P6.4.2.e), we initiated the project ReadyAI: Establishment of Readiness Criteria for AI-based Tools for Risk Assessment (Project lead: University of Basel, Switzerland & University of Gdańsk, Poland). This project aims to tackle key challenges in implementing machine- and deep learning in regulatory frameworks by developing a comprehensive scoring system to evaluate the development, training, validation, and maintenance of AI-based tools for use in CRA. To define these criteria, ReadyAI aims to foster effective communication between regulatory scientists, regulatory authorities, and AI developers. The resulting scoring system is intended to inform users about the readiness and associated uncertainties of AI-based tools, thereby enabling an objective assessment of their regulatory applicability (Figure 1).

Figure 1

Figure 1. ReadyAI aims to bridge the gap between AI development and regulatory application by defining an AI scoring system that enables regulatory scientists to assess the readiness, reliability, and applicability of AI-based models for chemical risk assessment. The project will promote structured communication among key stakeholders, including AI developers, researchers, computational scientists, regulatory scientists, as well as regulatory authorities. AI, Artificial intelligence; CRA, Chemical risk assessment. Figure was created in BioRender. Swiss Centre for Applied Human Toxicology (2025) https://BioRender.com/96ine4s.

2 Launch of the ReadyAI project

The ReadyAI project was officially launched with the kick-off meeting (May 19th, 2025, in Vienna, Austria), bringing together 21 partners from 9 institutions, including academia and regulatory authorities. This multidisciplinary group convened to align on the overarching goal of the project: the development of readiness criteria and a scoring system to assess the regulatory applicability of AI tools intended for use in CRA. The project aims at providing a foundational blueprint to integrate AI into regulatory processes in a scientifically robust, transparent, and responsible manner.

Regulatory readiness refers here to the extent to which an AI tool can be reliably applied within a decision-making process, in accordance with the constraints and requirements of chemicals safety regulation. Drawing inspiration from the established readiness framework for in vitro developmental neurotoxicity methods (Bal-Price et al., 2018; Blum et al., 2025) and quantitative structure–activity relationship (QSAR) (Gissi et al., 2024), ReadyAI aims to adapt and expand these concepts, creating a tailored scoring system for AI algorithms applied in CRA. The practical applicability of this envisioned scoring system will be assessed through a series of case studies, to be defined in alignment with regulatory stakeholders, which will be conducted over the duration of the project.

The kick-off meeting began with a session on the importance of building trust in AI for regulatory use. Presentations covered foundational concepts and applications of AI in CRA, including data extraction, QSAR quality evaluation, and AI development lifecycle steps (problem identification; data collection and processing; model training, selection, validation, deployment, and maintenance). Critical AI challenges such as data bias, adversarial attacks, and hallucinations were highlighted. The ReadyAI project objectives were outlined, with a focus on supporting technical training for regulatory scientists and developing a transparent and reproducible scoring system to assess the readiness and reliability of AI models used in CRA. The session also included a discussion of global regulatory perspectives, noting the current lack of overarching frameworks in some regions including the United States, and the United Kingdom, which adopt a less rigid, pro-innovation approach, as contrasted with the AI Act of the EU (European Commision, 2024b), which classifies AI systems according to risk levels. Within the EU, AI for CRA would likely fall under the high-risk category owing to its potential to impact human health and the environment. Legal scrutiny is therefore vital to the aims of the ReadyAI project.

The second part of the meeting focused on European institutional initiatives applying AI in CRA. Ongoing AI initiatives where the Joint Research Centre (JRC) is involved were presented, including the Designathon initiative by the European Partnership for Alternatives to Animal Experimentation (EPAA) (European Commision, 2024a), the Generative AI for Read-Across (GARA) initiative, and AI4AOP (Wittwehr, 2024). Other contributions showcased the integration of AI into systematic reviews, dossier completeness checks, and model-based hazard assessment within the European Food Safety Authority (EFSA) and the European Chemicals Agency (ECHA).

The closing discussion addressed critical factors for successful AI integration, including data curation, data quality assessment, model validation, transparent documentation, and clear communication of model strengths and limitations. The need for targeted training of regulatory scientists was emphasized, along with the importance of using precise and harmonized language when describing AI tool functionality and constraints. Participants also highlighted the risk of overcomplicating models, noting that algorithmic complexity does not always yield better performance. A strong consensus emerged on the necessity of high-quality training data to reduce bias and prevent overfitting.

The meeting concluded with a commitment to establishing a multi-agency, cross-sectoral steering committee and the delivery of a detailed project roadmap in 2025, which will guide the activities of the initiative through to its completion in 2029.

3 Perspectives for the establishment of regulatory readiness criteria for AI models

The integration of AI, ranging from classical machine learning models to advanced generative AI technologies, into regulatory frameworks requires a structured and transparent approach to evaluate whether AI tools are sufficiently mature, trustworthy, and fit for purpose. Although AI is rooted in computational sciences, many aspects of algorithm development and application rely on expert judgment rather than defined gold-standard rules. The establishment of regulatory readiness criteria for AI tools applied in safety assessment is, therefore, a critical step in bridging the gap among AI innovation, governance, and regulatory acceptance and implementation.

While the CRA is the primary focus of ReadyAI, the AI underlying principles and computational models have broad applicability. Therefore, we envision an intersectoral approach for our scoring system, supporting regulatory AI assurance across other domains such as preclinical safety and healthcare, where transparency, robustness, and accountability are equally critical. In line with that, the steering committee of ReadyAI has members of Swissmedic, EFSA, ECHA, JRC, and the Organization for Economic Co-operation and Development (OECD), complemented by an academic AI expert.

Our readiness scoring system will be grounded in core principles adapted from the OECD QSAR Assessment Framework (QAF) (Gissi et al., 2024), and extended to address the broader range of challenges presented by AI-based tools. While the QAF provides a well-established structure for evaluating model validity, transparency, and applicability, ReadyAI aims to broaden this approach so that it can be applied not only to QSAR models but also to other types of AI models used in CRA. This broader applicability requires expanding QAF key elements to capture aspects that are not traditionally part of QSAR evaluation. We anticipate that the scoring system will incorporate weighted criteria, acknowledging that certain elements carry greater regulatory relevance than others. Weighting will support proportional scoring and allow the system to distinguish between essential prerequisites and more context-dependent quality indicators. By doing so, the readiness scoring system will offer a unified, transparent framework that could support regulatory assessment of a variety of AI models, while maintaining coherence with established regulatory principles.

We envision that the scoring system will be built on some foundational pillars: data curation, overfitting, explainability, and uncertainty quantification.

AI performance relies on high-quality, well-curated training data, which affects bias, overfitting, and accuracy. Thus, data curation should be assigned substantial weight within the scoring system, as it directly impacts the quality and reliability of AI-based models.

Overfitting occurs when an AI model captures noise in the training data rather than learning generalizable patterns, leading to poor performance on unseen data. To achieve a high readiness score, developers must demonstrate the application of robust strategies to prevent overfitting.

Moreover, regulatory decision-making, particularly in contexts involving public health and safety, requires a clear understanding of how and why a model generates its outputs. As such, explainability is essential for regulatory acceptance and should be a key component of the scoring system.

Finally, regulatory scientists need to assess not only the output generated by the model but also its confidence in those predictions. Thus, precise description and quantification of uncertainty will enhance model transparency, foster trust, and ultimately support a more reliable and robust readiness assessment.

Together, these elements outline the conceptual foundations of the ReadyAI scoring system and illustrate how it aims to balance scientific rigor with regulatory applicability. As the project progresses, these components will be refined in consultation with regulatory stakeholders to ensure that the final framework is both scientifically robust and readily applicable in real-world regulatory contexts.

4 Conclusion remarks

The ambition of ReadyAI is to bridge the gap between AI innovation and regulatory usability and acceptance. By defining readiness levels aligned with scientific maturity and regulatory requirements, we aim to provide regulators with a practical tool to evaluate the applicability, limitations, and trustworthiness of AI models applied in CRA. In addition, the outcomes of this project will help AI developers and users, providing clear criteria that must be met for potential regulatory applicability. This will facilitate collaboration between developers and regulators to ensure responsible deployment of AI-methods.

To transition from animal-based safety assessment to the integration of NAMs, including algorithms and AI in regulatory processes, a governance structure and an intersectoral approach are essential.

The integration of AI into regulatory frameworks demands human oversight, responsibility, accountability, technical expertise, and traceability. While regulatory science is still evolving in this area, the successful uptake of AI will ultimately depend on our ability to translate technical excellence into trusted, interpretable, and actionable approaches that align with the high standards of human and environmental health protection and policymaking.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.

Author contributions

JPS: Conceptualization, Writing – review & editing, Writing – original draft. JB: Writing – review & editing. UM: Writing – review & editing. SS: Writing – review & editing. LD: Writing – review & editing. AČ: Writing – review & editing. LH: Writing – review & editing. RL: Writing – review & editing. VK: Writing – review & editing. LM: Writing – review & editing. SF-F: Writing – review & editing. AW: Writing – review & editing. GRi: Writing – review & editing. GRa: Writing – review & editing. AR: Writing – review & editing. EDC: Writing – review & editing. OT: Writing – review & editing. CB: Writing – review & editing. LW-R: Writing – review & editing. TP: Writing – review & editing. EF: Conceptualization, Funding acquisition, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This work has received funding from the Swiss Federal Office of Public Health (contract 142005152/643-5/1 and 142007580/641.3-57/1/1). This work was carried out in the framework of the European Partnership for the Assessment of Risks from Chemicals (PARC) and has received funding from the European Union’s Horizon Europe research and innovation program under Grant Agreement No 101057014. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the Health and Digital Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. UM and SS are also grateful for support to Ministry of Social Affairs, Republic of Estonia (grants 3-4/1593-1, 3-4/2332-1) and to the Ministry of Education and Research, Republic of Estonia, through the Estonian Research Council (grant number PRG1509).

Acknowledgments

The ReadyAI project steering committee members João Barroso (JRC), Patience Browne (OECD), Andrea Gissi (EFSA), Panagiotis Karamertzanis (ECHA), Elisabeth Klenke (Swissmedic), and Axel Mosig (Ruhr-University Bochum) are acknowledged for their valuable support and technical expertise throughout the course of the project.

Conflict of interest

TP is employed by QSAR Lab Ltd.

The remaining author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI was used only to refine the original draft manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Author disclaimer

The views expressed in this publication are those of the authors and do not necessarily represent the official positions of their affiliated institutions. These institutions assume no responsibility for any errors or inaccuracies that may be contained herein.

References

Bal-Price, A., Hogberg, H. T., Crofton, K. M., Daneshian, M., Fitzgerald, R. E., Fritsche, E., et al. (2018). Recommendation on test readiness criteria for new approach methods in toxicology: exemplified for developmental neurotoxicity. ALTEX 35, 306–352. doi: 10.14573/altex.1712081,

PubMed Abstract | Crossref Full Text | Google Scholar

Blum, J., Bartmann, K., Souza, J. D., and Fritsche, E. (2025). Developmental neurotoxicity as a case example for a six-step framework for the sustainable regulatory implementation of Nams. Curr. Opin. Toxicol. 42:100528. doi: 10.1016/j.cotox.2025.100528

Crossref Full Text | Google Scholar

Blümmel, T., Rehn, J., Mereu, C., Graf, F., Bazing, F., Kneuer, C., et al. (2024). Exploring the use of artificial intelligence (Ai) for extracting and integrating data obtained through new approach methodologies (Nams) for chemical risk assessment. EFSA Support. Public. 21:8567E. doi: 10.2903/sp.efsa.2024.EN-8567,

PubMed Abstract | Crossref Full Text | Google Scholar

European Commision. (2024a). EPAA designathon for human systemic toxicity. Available online at: https://single-market-economy.ec.europa.eu/calls-expression-interest/epaa-designathon-human-systemic-toxicity_en (Accessed July 25, 2025)

Google Scholar

European Commision. (2024b). The EU artificial intelligence act. Available online at: https://artificialintelligenceact.eu/ (Accessed July 25, 2025)

Google Scholar

European Commision. (2025). Roadmap towards phasing out animal testing. Available online at: https://single-market-economy.ec.europa.eu/sectors/chemicals/reach/roadmap-towards-phasing-out-animal-testing_en (Accessed July 25, 2025)

Google Scholar

Gissi, A., Tcheremenskaia, O., Bossa, C., Battistelli, C. L., and Browne, P. (2024). The OECD (Q)Sar assessment framework: a tool for increasing regulatory uptake of computational approaches. Comput. Toxicol. 31:100326. doi: 10.1016/j.comtox.2024.100326

Crossref Full Text | Google Scholar

Global Coalition for Regulatory Science Research. (2025). GCRSR. Available online at: https://gcrsr.net/. (Accessed July 25, 2025)

Google Scholar

Hartung, T., and Kleinstreuer, N. (2025). Challenges and opportunities for validation of Ai-based new approach methods. ALTEX 42, 3–21. doi: 10.14573/altex.2412291,

PubMed Abstract | Crossref Full Text | Google Scholar

Hartung, T., Whelan, M., Tong, W., and Califf, R. M. (2025). Is regulatory science ready for artificial intelligence? NPJ Digit. Med. 8:200. doi: 10.1038/s41746-025-01596-0,

PubMed Abstract | Crossref Full Text | Google Scholar

Lin, Z., and Chou, W. C. (2022). Machine learning and artificial intelligence in toxicological sciences. Toxicol. Sci. 189, 7–19. doi: 10.1093/toxsci/kfac075,

PubMed Abstract | Crossref Full Text | Google Scholar

Schmeisser, S., Miccoli, A., Von Bergen, M., Berggren, E., Braeuning, A., Busch, W., et al. (2023). New approach methodologies in human regulatory toxicology - not if, but how and when! Environ. Int. 178:108082. doi: 10.1016/j.envint.2023.108082,

PubMed Abstract | Crossref Full Text | Google Scholar

United States Food and Drug Administration. (2025). FDA announces plan to phase out animal testing requirement for monoclonal antibodies and other drugs. Available online at: https://www.fda.gov/news-events/press-announcements/fda-announces-plan-phase-out-animal-testing-requirement-monoclonal-antibodies-and-other-drugs (Accessed July 25, 2025)

Google Scholar

United States National Institutes of Health. (2025). NIH funding announcements to align with NIH initiative to prioritize human-based research. Available online at: https://grants.nih.gov/news-events/nih-extramural-nexus-news/2025/07/nih-funding-announcements-to-align-with-nih-initiative-to-prioritize-human-based-research (Accessed July 25, 2025)

Google Scholar

Wassenaar, P. N. H., Minnema, J., Vriend, J., Peijnenburg, W., Pennings, J. L. A., and Kienhuis, A. (2024). The role of trust in the use of artificial intelligence for chemical risk assessment. Regul. Toxicol. Pharmacol. 148:105589. doi: 10.1016/j.yrtph.2024.105589,

PubMed Abstract | Crossref Full Text | Google Scholar

Wittwehr, C. (2024). Ai4aop - artificial intelligence for Aops - background and overview. Available online at: https://zenodo.org/records/14515001 (Accessed July 25, 2025)

Google Scholar

Wittwehr, C., Blomstedt, P., Gosling, J. P., Peltola, T., Raffael, B., Richarz, A. N., et al. (2020). Artificial intelligence for chemical risk assessment. Comput. Toxicol. 13:100114. doi: 10.1016/j.comtox.2019.100114,

PubMed Abstract | Crossref Full Text | Google Scholar