REGISTER NOW

Featuring:

Ivo Djidrovski, PhD: “Augmenting Chemical Hazard and Risk Assessment through Large Language Models and AI Agents”

ASCCT 14th Annual Meeting Ray Tice Tox21 Student Award

Nyssa Tucker: “RADISH: ROBOKOP Assisted Discovery of in silico Hypotheses - A case study of mechanisms linking nuclear receptors to liver disease”

ASCCT 14th Annual Meeting Travel Award Recipient

A brief Q&A session will follow each presentation.

---

ABSTRACTS

Augmenting Chemical Hazard and Risk Assessment through Large Language Models and AI Agents
The integration of Large Language Models (LLMs) and AI agents into toxicology offers a transformative opportunity to accelerate and scale hazard assessment. Trained on biomedical and regulatory corpora, LLMs provide context-aware reasoning and act as dynamic interfaces for decision-making in complex data environments. This work presents a prototype multi-agent system that leverages LLMs to support in silico hazard identification, analog selection, and read-across with emphasis on reproducibility, transparency, and regulatory alignment.
The system interfaces with the OECD QSAR Toolbox via an API-enabled framework. Agents autonomously query chemical databases, simulate metabolism, retrieve experimental and predicted data, and apply category formation logic to identify structurally and mechanistically relevant analogs. Through prompt engineering and multi-step reasoning, agents generate narratives that strengthen weight-of-evidence hazard decisions and mechanistic interpretation.
Case studies from the VHP4Safety project, including known regulatory substances, were used to benchmark agent performance against expert assessments. Results show AI agents can execute QSAR workflows, fill data gaps, and propose mechanistic hypotheses consistent with existing evidence. The modular architecture further integrates with LLM-native tools (e.g., ChemCrow, ChemThinker) and open-source components, underscoring adaptability across toxicological contexts.

A central feature is interoperability: modular design enables scalable, customizable, and trustworthy workflows that bridge regulatory frameworks with AI-native tooling. These agentic systems operationalize New Approach Methodologies (NAMs) and support transparent, collaborative risk assessment. Hybrid intelligence—human oversight augmented by AI-driven insight—emerges as a catalyst for sustainable, data-driven, and probabilistic toxicology.

RADISH: ROBOKOP Assisted Discovery of in silico Hypotheses - A case study of mechanisms linking nuclear receptors to liver disease
Adverse Outcome Pathways (AOPs) describe the cascade of biological events leading to toxic outcomes. An AOP is built through manual efforts of subject matter experts identifying empirical molecular or biological evidence for each node in the chain. Recent developments in biomedical information retrieval led to the emergence of biomedical knowledge graphs (KG), which integrate multiple elementary semantic triples, i.e., special ‘subject-predicate-object’ relationships between chemical and biological entities (such as ‘chemical causes hepatotoxicity’) described in biomedical literature or databases. One KG, ROBOKOP, is a biomedical publicly available graph integrating 50+ databases to include over 9M nodes and 140M edges.
ROBOKOP Assisted Discovery of in silico Hypotheses (RADisH) is a methodology for automated discovery of AOPs based on graph mining algorithms. We evaluate RADISH using expert-defined AOPs for fatty liver disease. Of the approximately 50 unique events described across the thirteen AOPs in AOPwiki, KG queries considering just one nuclear receptor, AhR, resulted in 34 genes of which 26 were partial or complete matches to AOP events implying a recall of ~70%. The remaining 8 genes not matched to existing AOP events were evaluated for plausibility within the context of fatty liver disease and investigated for potential applicability as hypothetical mechanistic event candidates. The use of these novel hypothetical mechanisms can support future development of quantitative models predicting chemical toxicity. Furthermore, these hypothetical mechanisms can be fed back to biomedical researchers to evaluate the plausibility of these connections and broaden the scope of biomedical knowledge.

---

ABOUT THE PRESENTERS

Ivo Djidrovski, PhD, is a computational toxicologist and AI Research Lead within The Virtual Human Platform for Safety Assessment (VHP4Safety) project at Utrecht University. His work focuses on developing auditable, interoperable AI-assisted workflows to support non-animal chemical safety assessment and next-generation risk assessment (NGRA). He leads the development of O-QT, an AI-assisted orchestration layer for the OECD QSAR Toolbox that formalizes read-across into structured, policy-aware execution traces. His research further explores federated orchestration architectures (ToxMCP) that enable transparent integration of read-across, mechanistic pathway knowledge and PBK/PBPK modelling. In parallel, he founded in4r.ai to translate agentic AI infrastructure into real-world industrial toxicology applications.

Nyssa Tucker is a 5th year Computational Toxicology Ph.D. Candidate at UNC Chapel Hill. Along with the research interests embedded in the associated abstract, Nyssa is also passionate about building strong networks between people towards a resilient community-focused system of collaboration addressing pressing social concerns. Nyssa is a member of SOT, UE150, and UNC's Molecular Modeling Laboratory.

---

The recording and select materials from this webinar will be posted on the ASCCT webinar archive: https://ASCCTox.org/Webinar-Archive

Download as iCal file