Modified Project Summary/Abstract Section Adverse drug reactions (ADRs) are dangerous and expensive. ADRS driven by immune-mediated hypersensitivity (including rashes, hepatotoxicity, and Steven-Johnson syndrome) are the most difficult to predict and occasionally can be severe as well as fatal. Hypersensitivity-driven ADRs are the leading cause of drug withdrawal and termination of clinical development. Yet a large proportion of drugs are not associated with hypersensitivity-driven ADRs, offering hope that new medicines could avoid these ADRs entirely if reliable models of bioactivation existed. Accurate prediction and identification of molecules prone to ADRs would revolutionize drug development by screening out ADR-prone candidates early, before exposure to patients, and guiding drug modifications to reduce ADR risk. Small molecules are not intrinsically immunogenic and instead, involve bioactivation into reactive metabolite is that then covalently modify proteins to create immunogenic antigens. “Structural alerts” are molecular substructures prone to bioactivation, and they are often used to identify small molecules prone to bioactivation, and at risk of bioactivation-mediated ADRs. Currently, bioactivation relevant alerts are defined by experts, and they have important limitations that this study overcomes. It is now possible to predict metabolism and reactivity and toxicity using machine learning approaches. Building on this foundation, this proposal systematically discovers new structural alerts by explicitly modeling the impact of metabolism on reactivity and hence the potential to form ADR-relevant adducts. We hypothesize that (1) known bioactivation reactions, (2) molecule citation data, and (3) new substructure mining algorithms can be used to identify emerging structural alerts. Aim 1. We will test this hypothesis by using a computational approach to systematically mine structural alerts from databases of known metabolism and reactivity reactions. Aim 2. We will computationally and experimentally validate structural alerts and assess their structural contingencies. Structural alerts are only conditionally bioactivated, depending on the precise molecule they appear. Newly proposed structural alerts, moreover, are most useful when there is experimental evidence that they in fact can be bioactivated. PubHlthRel: Structural alerts discovered in this study will help scientists avoid toxic molecules in drug development, and better understand why medicines on the market become toxic. Overcoming a fundamental limitation with structural alerts, machine learning models of bioactivation will clarify in which molecules alerts are and are not bioactivated. This knowledge will help scientists make safer medicines in the future, modify existing medicines to make them safer, and reduce ADRs by using existing medicines more safely.