AI-Powered chemical-pathway-patient-directed polypharmacology for OUD therapy

Abstract Opioid Use Disorder (OUD) presents substantial health and societal challenges, particularly in the US. The proposal aims to discover novel small molecule therapeutics for safe and effective treatment of OUD. We will identify multi-cell type and multi-target polypharmacological compounds, focusing on opioid oxycodone in the Phase I. Recent studies highlight that not only neurons but also glial cells are involved in opioid addiction, affecting communication and neuronal plasticity. Furthermore, OUD impacts both the central nervous system and peripheral immune systems, with different opioids affecting various cell types distinctly. Thus, effective OUD treatment requires modulating multiple cell types and drug targets to restore brain function. Traditional drug discovery approaches are insufficient for OUD due to the complexity of the disease and the lack of specific drug targets and high-throughput phenotype readouts. Moreover, there are limited methods to understand OUD's complexity for polypharmacological lead discovery. This proposal seeks to overcome these challenges by leveraging Artificial Intelligence (AI) to develop a novel drug discovery strategy that integrates target-based and phenotype-based methodologies, aiming to identify new drug candidates for OUD efficiently. Specifically, we will develop the AI platform µ3.ai and apply it to OUD drug discovery. This platform will generate multi-cell type, multi- objective, multi-target lead compounds to restore OUD patient omics profiles to a non-addictive state. µ3.ai combines four AI models that 1) predicts dose-dependent chemical-induced cell viability and transcriptomics for phenotype-based drug discovery, 2) provides accurate genome-scale binding affinity predictions, 3) searches for molecules with desired efficacy and safety profiles, and 4) prioritizes cell type-specific OUD targets. Additionally, we will develop a human-in-the-loop active learning strategy. This approach involves close collaboration between medicinal chemists and AI scientists to iteratively refine µ3.ai and discover new polypharmacological therapeutics. This strategy utilizes Reinforcement Learning from Human Feedback (RLHF) to enhance the AI platform. It seamlessly integrates AI-powered molecular generation, property predictions, organic synthesis, in vitro assays, and compound testing in cell line models. In Phase I, established on the solid evidences from patient-derived single cell RNA-seq studies, we will target oxycodone dysregulates type I interferon signaling in neurons and astrocytes. Our goal is to restore the expression of multiple dysregulated genes (STAT1, ISG15, BST2, IFI44L, IFITM3, IFI27, PARP14, PLSCR1, B2M, P2RY14, etc.) in the type I interferon signaling to their normal state. The successful conclusion of this study will establish a robust groundwork for the exploration of OUD drug development endeavors in Phase II, thereby making substantial progress towards the objectives of the HEAL initiative of the National Institutes of Health (NIH).