Overview

Enhancing the transparency and acceptance of artificial intelligence (AI) and machine learning (ML) techniques in chemical risk assessment is the focus of this initiative. With an increasing reliance on these advanced methods for predicting chemical behaviour and assessing risks, addressing the challenges posed by unclear documentation and limited access to underlying data and algorithms is essential. This lack of transparency often hampers understanding and complicates the evaluation of uncertainties in these models.

The initiative will involve a thorough review and analysis of existing AI and ML-based and data-driven quantitative structure-activity relationship (QSAR) models ↗, aiming to identify best practices for transparency and regulatory applicability. By establishing clear reporting standards and decision-making criteria, the project will enhance the reliability and communication of uncertainties associated with these predictive models. Additionally, exemplary models will be curated and made available as FAIR (Findable, Accessible, Interoperable, and Reusable) data through the QsarDB.org repository ↗.

Ultimately, this effort seeks to foster better-informed decision-making for the protection of human health and the environment while reducing the need for animal testing, thereby contributing to a more effective regulatory framework in chemical risk assessment.

Achievements & Results

• Project group established with 10 participating partners, ensuring collaborative efforts.
• Definition of research priorities and requirements, aimed at focusing on the project objectives.
• A survey has been conducted on the use of ML and AI approaches.
• Review to gather information on ML New Approach Methodologies (NAMs ↗) for has been completed for chemical hazard identification, providing also framework for assessing ML NAMs.
• Ongoing review to gather information on AI new approaches for chemical risk assessment, aiming to enhance understanding and application in regulatory contexts.
• Examples on interpretation of computational NAMs have been provided together with respective data publications.

Policy relevance

This project supports hazard identification broadly, as the computational models developed are not limited to a single regulatory framework. This flexibility is valuable, given that ML and AI-based QSAR approaches, also known as in silico or computational NAMs, are emerging as powerful tools across multiple regulatory domains. While conventional QSAR models are already applied in frameworks such as REACH ↗, BPR, and CLP (ECHA ↗), as well as in EFSA ↗-regulated areas like pesticides, cosmetics, and pharmaceutical additives, newer ML/AI models require further validation to ensure regulatory applicability.

Aligned with the OECD QSAR Assessment Framework ↗ (QAF), which offers general guidance regardless of the modelling technique, this project addresses a key gap: the lack of clear guidance on how to evaluate and apply complex ML and AI models in regulatory contexts. By doing so, it contributes to making these next-generation tools more accessible and acceptable for chemical risk assessment and regulatory decision-making. While traditional QSAR models are better understood and already used in regulatory context, the application of complex ML and AI models is still an open challenge. This project aims to provide additional guidance on handling complex ML and AI QSAR approaches, that is today largely missing in regulatory settings.