Development of machine-learning methods for the optimization of binding selectivity for early stage drug discovery

该项目侧重于开发新的计算方法来支持药物开发过程，以减少优化临床前候选药物所需的大量化合物。具体来说，重点是开发机器学习模型来指导药物化学家，以便在优化药物结合选择性或乱交时更有效地选择化合物。近年来，“一药一靶点”模式受到了挑战，导致优化药物的重点转向所需的多药理学模式，这需要药物化合物结合多个不同的蛋白质靶点。细菌金属- β -内酰胺酶（MBL）抑制剂是高度期望药物结合杂乱性的一个例子。这类化合物是用来对付抗生素耐药性的。需要一种特定的选择性模式，候选药物应该对尽可能多的不同细菌MBLs有活性，同时不影响人类MBLs。目前，多靶点MBL抑制剂的开发正在进行中，目前还没有一种单一的抑制剂能够靶向4个最重要的细菌MBL靶点：VIM-1、VIM-2、IMP-1和NDM-1。该项目的第一个目标是开发计算方法，以设计具有针对MBL蛋白家族的所需选择性模式的新化合物，并在实验室中直接验证这些模型。最终，目标是创建一种方法，可用于预测和优化药物化合物对不同蛋白质家族的结合选择性，并可作为化学家创建选择性或混杂化合物的一般指导方针。因此，该项目旨在引入两个主要的新奇事物：首先，机器学习方法本身，将用于建议在实验室中制造新的化合物；第二，合成一种新的MBL抑制剂，对所有期望的细菌MBL靶标有效。在世界上，抗生素耐药性正在稳步增加，治疗方法的发展对制药工业来说是不经济的，迫切需要开发新的抗生素或抗生素耐药抑制剂。此外，对多药理学的关注和制药工业中药物发现成本的增加要求开发新的方法来设计选择性化合物，以使该过程更便宜和更快。因此，该项目属于EPSRC“人工智能技术”，“化学生物学和生物化学”和“计算与理论化学”研究领域，并与牛津大学有机化学系的Schofield教授以及葛兰素史克制药公司合作完成。

This project focuses on the development of novel computational methods to support the drug development process in order to reduce the large amount of compounds needed to optimize a pre-clinical drug candidate. Specifically, the focus is on developing machine learning models to guide medicinal chemists in order to choose compounds more efficiently when optimizing drugs for binding selectivity or promiscuity. In recent years, the "one drug one target" model has been challenged, resulting in a shift in emphasis of optimizing drugs towards desired polypharmacology patterns, which requires the drug compound to bind to multiple different protein targets. One example where binding promiscuity of a drug is highly desired are bacterial metallo-beta-lactamase (MBL) inhibitors. This class of compounds is designed to tackle antibiotic resistances. A specific selectivity pattern is required where drug candidates should be active against as many different bacterial MBLs as possible while not affecting human MBLs. Currently, the development of multi-target MBL inhibitors is ongoing and no single inhibitor that hits the four most important bacterial MBL targets[1]: VIM-1, VIM-2, IMP-1 and NDM-1 is known. The first goal of this project is the development of computational methods to design new compounds with the desired selectivity pattern against the MBL protein family as well as the direct validation of those models in the lab. Ultimately, the goal is to create a methodology which can be used to predict and optimize binding selectivity of drug compounds for different protein families and can act as a general guideline for chemists to create selective or promiscuous compounds. As such, the project aims to introduce two major novelties: first, the machine learning method itself, which will be used to suggest new compounds to make in the lab; and second, the synthesis of a new MBL inhibitor that is effective against all desired bacterial MBL targets. The development of new antibiotics or antibiotic resistance inhibitors is desperately needed in a world where antibacterial resistance is increasing steadily and the development of treatments uneconomical for the pharmaceutical industry. Furthermore, the focus on polypharmacology and the increasing cost of drug discovery in the pharmaceutical industry calls for the development of new methods for the design of selective compounds in order to make the process cheaper and faster. This project therefore falls within the EPSRC "Artificial Intelligence Technologies", "Chemical Biology and Biological Chemistry" and "Computational & Theoretical Chemistry" research areas and is done in collaboration with Prof. Schofield at the Department of Organic Chemistry at the University of Oxford as well as in collaboration with the pharmaceutical company Glaxo-Smith-Kline.