QSAR in the AI Era: Reflections for Advancing Chemical Safety Assessment

Area of Science:

• Computational toxicology and regulatory science
• Chemical safety assessment
• Artificial intelligence in risk evaluation

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are data-driven inference systems meeting AI definitions.
• QSARs are domain-specific models developed through decades of computational toxicology and regulatory science dialogue.
• The 21st International workshop on QSAR in Environmental and Health Sciences (QSAR2025) focused on QSAR integration and AI in chemical safety.

Purpose of the Study:

• To summarize roundtable discussions on QSAR integration in regulatory frameworks.
• To explore the role of modern AI, including Large Language Models, in toxicological risk assessment.
• To reflect on how established QSAR principles can inform the governance of emerging AI in chemical safety.

Main Methods:

• Summary of roundtable discussions from QSAR2025.
• Analysis of QSAR integration challenges: structural similarity, local performance, model selection, multiple predictions, and black-box explainability.
• Examination of modern AI applications beyond QSAR modeling.

Main Results:

• QSAR integration into regulatory frameworks is supported by OECD principles and standardized reporting (QMRF, QPRF, QRRF).
• The OECD QSAR Assessment Framework (QAF) provides a structure for evaluating QSAR reliability and regulatory relevance.
• Modern AI, like LLMs, shows potential to enhance toxicological risk assessment.

Conclusions:

• Decades of QSAR development offer valuable principles (data quality, applicability domain, uncertainty, expert judgment) for governing AI in chemical safety.
• Transparent documentation, applicability domain consideration, and expert interpretation are crucial for regulatory acceptance of QSAR predictions.
• Established QSAR governance principles are essential for the responsible development and deployment of AI in chemical safety assessment.