Computational modulator design and machine learning to target protein-protein interactions

摘要 我的研究计划的总体目标是开发和应用计算工具，以促进Rational 用于治疗的重要细胞通路调节器的设计。蛋白质-蛋白质相互作用(PPI)是 细胞信号和基因调控网络中的中心因素。他们的不当监管与 多种疾病，包括癌症、神经退行性疾病、自身免疫性疾病和糖尿病。 不可避免的是，许多PPI是药物发现的生物学上令人信服的目标。但尽管有几个值得注意的 尽管取得了成功，但大多数PPI没有成功地被瞄准，仍然无法下药。最基本的 挑战来自其固有的结构特征：许多PPI的结合表面通常都很大 在面积、平面和动态方面。PPI通常是暂时的，涉及多价接触。目前，最多的一个 有前景的PPI抑制剂发现策略是使用微型蛋白结构域模拟(PDMS)来复制 自然界使用的关键接口联系人。PDMS作为中等大小的分子具有很大的优势 表面互补性和比典型小分子更广泛的接触点，但仍然有限 因为-根据定义-在相互作用中只捕获总PPI结合能的一部分。这个 合成结构域的结合亲和力往往低于同源全长蛋白。另一方面， 靶向共价抑制是一种适用于克服基本束缚的正交治疗方法 PPI的局限性，但有一个众所周知的缺点：典型共价弹头的高反应性导致 非特异性抑制和毒性。在这里，我们旨在为一种新的设计策略开发计算方法 这将利用这两种方法的优势-PDMS和共价抑制-同时 缓解了它们各自的限制。这项工作的重点是理性地发现有效的抑制剂 非共价识别，然后以精致的特异性共价靶向蛋白质-蛋白质结合界面。 此外，我们通过集成多任务机器学习和 分子建模将大大加快合理的药物发现过程。计划中的工作 基于我们在三种最先进的计算方法方面的最新进展：AlphaSpace for Fragment- PPI界面的中心地形图.模拟共价的从头算QM/MM分子动力学 抑制；以及一种新的增量机器学习策略，以同时提高得分、对接和 具有筛选性能的蛋白质配基评分功能。我们的设计努力将导致高度具体和 多种治疗上重要但以前不能用药的PPI接口的有效调节剂，提供 药物开发的新线索。

Abstract The overall goal of my research program is to develop and apply computational tools to facilitate the rational design of modulators of important cellular pathways for therapeutic use. Protein-protein interactions (PPIs) are central factors in cellular signaling and gene regulation networks. Their misregulation is associated with a variety of diseases, including cancer, neurodegenerative disease, autoimmune disease, and diabetes. Inevitably, many PPIs are biologically compelling targets for drug discovery. But despite a few notable successes, most PPIs have not been successfully targeted and remain undruggable. The fundamental challenge derives from their intrinsic structural features: the binding surfaces of many PPIs are generally large in area, flat, and dynamic. PPIs are often transient and involve multivalent contacts. Currently, one most promising PPI inhibitor discovery strategy is to use miniature protein domain mimetics (PDMs) to reproduce the key interface contacts utilized by nature. PDMs are advantageous as medium-sized molecules with high surface complementarity and a broader set of contact points than typical small molecules, but are still limited because—by definition—only a portion of the total PPI binding energy is captured in the interaction. The binding affinity of the synthetic domains is often lower than the cognate full-length proteins. On the other hand, targeted covalent inhibition is an orthogonal therapeutic approach fit to overcome the fundamental binding limitations at PPIs, but has a well-known drawback: the high reactivity of typical covalent warheads leads to nonspecific inhibition, and toxicity. Here we aim to develop computational methods for a new design strategy that will leverage the strengths of these two methods—PDMs and covalent inhibition—while simultaneously mitigating their respective limitations. The focus of the effort is to rationally discover potent inhibitors that will non-covalently recognize and then covalently target protein-protein binding interfaces with exquisite specificity. Furthermore, our development of robust scoring functions by integrating multitask machine learning and molecular modeling would significantly accelerate the rational drug discovery process. The planned work builds on our recent advances in three state-of-the-art computational approaches: AlphaSpace for fragment- centric topographical mapping of PPI interfaces; ab initio QM/MM molecular dynamics for modeling covalent inhibition; and a novel delta-machine learning strategy to simultaneously improve scoring, docking and screening performance of a protein-ligand scoring function. Our design efforts will result in highly specific and potent modulators of a variety of therapeutically important but previously undruggable PPI interfaces, providing new leads for drug development.