AI-powered chemical proteomics for drug discovery targeting orphan proteins

摘要 全基因组关联研究、全基因组测序和高通量技术 产生了大量不同的组学数据。然而，这些数据集还没有被充分挖掘出来 提高药物发现的有效性和效率。只有5%-10%的可用药蛋白质是 批准的药物。未下药的孤儿蛋白是尚未治愈的疾病的潜在靶点，但其 内源和外源配体是未知的。此外，在将药物与靶标联系起来方面存在知识差距。 将亲和力与临床结果相结合。我们几乎不知道目标是被粘合剂激活还是抑制(即， 功能活性：激动剂与拮抗剂)。到目前为止，很少有实验和计算工具可以确定 孤儿蛋白的全基因组蛋白质-配体相互作用及其配体诱导的功能活性 (LIFA)既适用于孤儿蛋白，也适用于大多数研究良好的蛋白质。现有的机器学习技术有 由于分布外(OOD)问题，在预测孤儿蛋白的配体方面大多不成功，即， 如果一种看不见的蛋白质与蛋白质中的蛋白质有显著不同，他们就不能可靠地预测它的功能。 在序列和结构方面训练数据。基于结构药物的常用计算工具 设计，如蛋白质-配体对接/评分和分子动力学模拟，既不能扩展，也不能 特别可靠。因此，我们对孤儿蛋白的化合物筛选能力有限。 这项提案寻求开发和实验验证预测全基因组pli的创新方法。 和LIFA，以应对上述挑战。基于我们成功的概念验证研究和我们的 在实验实验室和计算实验室之间密切的多学科合作，我们将开发一种 一种新的计算框架，通过集成来自化学品和药物的大数据，在多尺度上模拟药物作用 结构基因组学和开发创新的深度学习算法。具体地说，我们将制定一种结构- 增强的深度学习框架可靠准确地预测孤儿的蛋白质-配体相互作用 基因组尺度上的蛋白质。我们将结合功能基因组学和化学基因组学来预测配体- 诱导的功能活动。我们将应用开发的方法来设计和实验测试抑制剂 孤儿抗癌靶标AVIL和治疗阿片使用障碍的多巴胺受体双重拮抗剂(OUD)。这个 拟议的研究提供了一种创新的概念、方法和翻译应用。完成这项工作 研究将填补在理解生物系统中药物作用方面的关键知识空白，并显著 影响复杂疾病的药物发现，其中许多疾病缺乏有效和安全的治疗方法。已开发的 方法论和平台不仅会立即影响美国国立卫生研究院的“照亮可用药基因组” 该计划不仅在生物医学研究的其他领域具有潜在的广泛应用。

Abstract Genome-Wide Association Studies, whole-genome sequencing, and high-throughput techniques have generated vast amounts of diverse omics data. However, these sets of data have not yet been fully explored to improve the effectiveness and efficiency of drug discovery. Only 5-10% of druggable proteins are targeted by approved drugs. The undrugged orphan proteins are potential targets of yet-incurable diseases but whose endogenous and exogeneous ligands are unknown. Furthermore, there is a knowledge gap to link drug-target binding affinities to clinical outcomes. We know little if the target is activated or inhibited by the binder (i.e., function activity: agonist vs. antagonist). To date, few experimental and computational tools can determine genome-wide protein-ligand interactions (PLIs) for orphan proteins and ligand-induced functional activities (LIFAs) for both orphan proteins and majority of well-studied proteins. Existing machine learning techniques are mostly unsuccessful in predicting the ligand of orphan proteins due to an out-of-distribution (OOD) problem, i.e., they cannot reliably predict the function of an unseen protein if it is significantly different from the proteins in the training data in terms of sequence and structure. Commonly used computational tools for structure-based drug design, such as protein-ligand docking/scoring and Molecular Dynamics simulations, are neither scalable nor particularly reliable. As a result, we only have a limited capability of compound screening for orphan proteins. This proposal seeks to develop and experimentally validate innovative methods for predicting genome-wide PLIs and LIFAs to address aforementioned challenges. Building on our successful proof-of-concept studies and our close multidisciplinary collaborations between experimental and computational laboratories, we will develop a novel computational framework to model drug actions on a multi-scale by integrating big data from chemical and structural genomics and developing innovative deep learning algorithms. Specifically, we will develop a structure- enhanced deep learning framework to reliably and accurately predict protein-ligand interactions for orphan proteins on a genome-scale. We will integrate functional genomics with chemical genomics to predict ligand- induced functional activity. We will apply the methods developed to design and experimentally test inhibitors of orphan anti-cancer target AVIL and dual antagonists of dopamine receptors for opioid use disorder (OUD). The proposed research offers an innovative concept, methodology, and translational applications. Completing this research will fill a critical knowledge gap in understanding drug actions in a biological system and significantly impact drug discovery for complex diseases, many of which lack effective and safe treatments. The developed methodology and platform will not only immediately impact the NIH’s “Illuminating the Druggable Genome” Program but also has potentially broad applications in other areas of biomedical research.