Falgun Shah, Merck Research Laboratories

Raymond Gonzalez, Merck Research Laboratories

There have been several recent debates on what artificial intelligence (AI) can or cannot do in the field of predictive toxicology. In general, skepticism exists in the utility of AI in drug discovery, development and environmental toxicology that can be attributed to (1) limited case studies in public domain showing impact of AI/machine learning (ML) in toxicology decision-making, and (2) a narrow applicability domain or limited prospective validation of existing machine learning (ML) models for toxicology endpoints. Also, missing considerations for dose/toxicokinetic information for end-organ toxicity prediction limits models' utility in decision-making. This session is intended to highlight successful applications of AI in the field of predictive toxicology in the form of case studies from government agencies, academia, software companies and pharmaceutical industries. Case studies will focus on three major areas: (1) how adverse outcome pathway (AOP) models built on diverse chemical space using structural properties and/or high-throughput assays are being used to prioritize pharmaceuticals or identify environmental chemicals' toxicity risks; (2) the impact of AI for hazard detection in toxicological pathology; (3) the knowledge graph search engine–based explainable AI providing insights into the mechanisms of toxicity. The first speaker from academia will share his view on AI in the field of predictive toxicology and will introduce the case study themes of the session. The second speaker will discuss a case study from NIEHS on how consensus ensemble ML models for endocrine disruption and acute systemic toxicity built on diverse chemical space are being used to prioritize environmental chemicals and/or to waive requirements for animal toxicology studies. The third case study from the academia will describe their adverse outcome pathways (AOP) models built using high-throughput screening toxicity assays in conjunction with toxicokinetic data to prioritize compounds with a lower likelihood of hepatotoxicity. The fourth case study from an AI software company will discuss their search engine–based AI approach leveraging knowledge graph generated using 700 million biomedical research data points to inform on potential off-target pharmacology or to identify mechanisms of toxicity. The next three case studies will be presented by speakers from the pharmaceutical industry focusing on successful applications of AI approaches in safety prediction at early discovery, in preclinical or a regulatory environment. In this section, the fifth speaker will share applications of the in silico AOP approach for the prediction of genotoxicity risk and its practical utility in early discovery to prioritize compounds with low mutagenicity risk. The sixth speaker will cover a developmental and reproductive toxicity case study for automated rabbit fetus skeleton assessment focusing on application of micro-CT imaging coupled with conventional neural network models for the segmentation and labeling of vertebrae. The final speaker will describe efforts on application of natural language processing to digitalized historical treatment–related in vivo data and leveraging this digitized data and a high-quality ML model to de-risk off-target toxicity.