- AI Applications in Chemical Risk Assessment and Safety Science: Toward Predictive, Scalable, and Reduced Animal Testing Approaches
-
Stars at Night Ballroom B4‚ Convention Center
Chair: Chao Ji, CDC/ATSDR
Co-Chair: Kan Shao, Indiana University Bloomington
Primary Endorser: Risk Assessment Specialty Section
Endorser(s): Computational Toxicology Specialty Section
Endorser(s): Biological Modeling Specialty Section
Chemical risk assessment faces a growing scalability challenge, with thousands of chemicals requiring evaluation while resources for traditional in vivo testing remain limited. At the same time, regulatory and scientific communities are increasingly prioritizing new approach methodologies (NAMs) that are predictive, efficient, and reduce reliance on animal testing. Recent advances in artificial intelligence (AI), including physics-informed approaches, agent-based systems, and data-driven models, offer new opportunities to improve predictive performance and integrate diverse data streams into more efficient risk assessment workflows. This session highlights how AI can support key stages of the risk assessment process: (1) mechanistic prediction of chemical-biological interactions, (2) scalable data generation using high-throughput in vitro methods, (3) system-level prediction of toxicity outcomes, and (4) integration of evidence streams for hazard identification and weight of evidence evaluation. Presentations span approaches from mechanistic modeling and simulation to predictive models built from large experimental datasets, demonstrating how complementary methods support a more efficient and predictive paradigm. Each presentation will emphasize practical applications in drug development, chemical risk assessment, and regulatory contexts, followed by a 10-minute Q&A. Topics include interpretation of model outputs, integration into workflows, and opportunities to reduce animal testing. The first presentation demonstrates how hybrid physics-AI models improve prediction of secondary pharmacology, achieving strong agreement with experimental binding affinity data and enabling interpretable, structure-based safety predictions. The second presentation showcases large-scale generation of in vitro dermal toxicity data (>3,000 compounds) to train machine learning models, demonstrating improved predictive performance relative to models trained on smaller public datasets. The third presentation presents ToxDT, a toxicological digital twin platform that generates multi-endpoint toxicity predictions from chemical structure and exposure inputs, validated against curated in vivo datasets and used to optimize experimental design. The fourth presentation introduces agentic AI systems that autonomously integrate chemical structure, bioactivity, regulatory databases, and literature to generate transparent hazard assessments, with outputs benchmarked against expert-derived evaluations in regulatory contexts (e.g., ICH M7.Challenges remain, including data quality, model bias, lack of standardized validation frameworks, and uncertainty quantification. These limitations, along with domain applicability, will be discussed to provide a balanced perspective. Attendees will gain a practical understanding of how different AI approaches can be applied across the risk assessment process, as well as considerations for evaluating their reliability and appropriate use. By presenting diverse but complementary perspectives, this session aims to support informed and realistic adoption of AI tools in modern toxicology and safety science. Together, these presentations illustrate the transition of AI from isolated predictive tools to integrated, validated, and decision-support systems that can support more efficient and increasing animal-free safety assessments. This session is particularly relevant for researchers, regulators, and practitioners seeking to leverage advanced computational tools to modernize chemical risk assessment for pharmaceuticals, consumer products, and environmental chemicals. By the end of this session, attendees will be able to (1) describe how physics-informed and machine learning models improve predictive performance in chemical safety assessment; (2) assess model validation approaches, including performance metrics, external validation, and uncertainty quantification; (3) understand the role of toxicological digital twins in simulating biological responses and optimizing study design; (4) evaluate how agentic AI systems integrate diverse data streams into transparent weight of evidence decisions; (5) identify limitations of AI approaches; and (6) apply practical strategies for incorporating AI tools into chemical risk assessment workflows.Session and Activity Type
Enter Note
Go to previous page in this tab
Event
