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

Abstract

Quantitative Structure–Activity Relationship (QSAR) models are increasingly discussed in the broader context of artificial intelligence (AI). Indeed, they formally meet certain regulatory definitions of AI as data-driven inference systems. However, in the context of chemical safety assessment, QSARs represent a distinct, domain-specific class of models shaped by decades of dialogue between computational toxicology and regulatory science. This work summarizes roundtable discussions from the 21st International workshop on QSAR in Environmental and Health Sciences (QSAR2025) held in Milan in June 2025, addressing structural similarity and local performance assessment, model selection, integration of multiple predictions, and challenges posed by black-box models that demand careful consideration of the balance between predictive performance and explainability. These discussions highlight the experience of integrating QSAR approaches into regulatory frameworks, supported by internationally harmonized OECD principles, standardized reporting formats (QMRF, QPRF, QRRF), and the OECD QSAR Assessment Framework (QAF). The QAF provides a structured basis for evaluating the reliability and regulatory relevance of QSAR predictions through transparent documentation, consideration of applicability domain, and expert-driven interpretation. The discussion is then broadened to examine how modern AI, particularly Large Language Models, may support toxicological risk assessment beyond QSAR modeling itself. Building on both the conference insights and this extended analysis, this work reflects on how principles established through decades of QSAR development and regulatory integration (including considerations on data quality, applicability domain, uncertainty, and expert judgment) may inform the governance of emerging AI applications in chemical safety assessment.