MFB: Targeting the Dark Proteome by Machine-learning-guided Protein Design

来自罗格斯大学的萨加尔·哈雷、让·鲍姆、亚当·葛姆雷、纪尧姆·拉穆雷和王思健博士将开发由新型机器学习（ML）方法指导的靶向蛋白质编辑器。精确编辑基因组的能力已经改变了现代生物技术和医学;然而，原位精确编辑蛋白质的技术-生物学的主力-一直缺乏。用这种方法靶向含有功能重要但结构无序的片段的蛋白质将是特别有用的，但它是具有挑战性的。跨学科研究团队将开发和应用新型机器学习算法来设计新的编辑蛋白，这些蛋白包含一个结合结构域和一个酶结构域，以确保在细胞中存在数千种其他蛋白质的情况下，分别选择性地识别和修饰所选的靶蛋白。迭代设计-构建-测试-学习循环将涉及生化实验和模型开发的密切相互作用，以获得强大的，可推广的和可解释的ML模型，这些模型可以预测选择性蛋白质编辑。该项目位于化学，生物物理学，机器人技术，计算机科学和统计学的接口，因此将为来自定量，生物和化学科学背景的学生提供独特的培训和研究机会。本科生团队（从罗格斯大学的各种本科生研究经验项目中招募）将使用计算蛋白质设计算法的游戏化版本来开发和改进蛋白质编辑器。蛋白质设计实践课程将在亲和组会议上举行，以接触STEM中代表性不足的群体的学生，并为他们提供学习和参与研究的机会。最近，用于蛋白质建模的新ML方法表明，可以利用序列和结构数据正确学习复杂的序列-结构关系并设计新型蛋白质。然而，丰富的功能和生物物理数据，例如，蛋白质-肽结合数据仍然在很大程度上未被利用。需要超越序列表示并基于分子结构实现语义更丰富的表示的新ML模型。这些表示可以捕获更具体的序列-功能和序列-能量学关系。研究团队将使用由ML模型驱动的设计-构建-测试-学习管道，这些模型在实验数据、分子结构和能量学方面都经过训练。该团队将这些方法应用于PDZ结构域的设计，这些结构域以高亲和力和特异性与目标内在无序蛋白质的C末端结合，并在目标蛋白质的内部序列位点选择性切割。ML模型将在PDZ和蛋白酶结构域的可用序列和实验生物物理数据上进行预训练，并用于生成新蛋白质。ML模型的设计验证和迭代改进将通过使用高通量测定、深度测序、生物分子NMR和机器人启用的结合和切割测定的实验表征来进行。成功设计的蛋白质将被融合以获得PDZ-蛋白酶蛋白编辑器。开发的方法将是通用的，广泛适用于允许原位蛋白质编辑的各种生物学相关的目标，特别是与内在无序的蛋白质或区域。新的试剂将有助于照亮所谓的“黑暗”蛋白质组。该项目由数学和物理科学（MPS）理事会的化学部（CHE）和数学科学部（DMS）以及计算机和信息科学与工程（CISE）的信息和智能系统部（IIS）支持。该奖项反映了NSF的法定使命，并被认为是值得通过使用基金会的学术价值和更广泛的影响审查标准。

Drs. Sagar Khare, Jean Baum, Adam Gormley, Guillaume Lamoureux, and Sijian Wang from Rutgers University will develop targeted protein editors guided by novel machine learning (ML) approaches. The ability to precisely edit genomes has transformed modern biotechnology and medicine; however, technology for the in-situ precision editing of proteins – the workhorses of biology – has been lacking. Targeting with such methods proteins contain functionally-important but structurally-disordered segments would be particularly useful but it is challenging. The interdisciplinary research team will develop and apply novel machine learning algorithms to the design of new editor proteins that contain a binding domain and an enzymatic domain to ensure the selective recognition and modification, respectively, of a chosen target protein in the presence of thousands other proteins present in the cell. Iterative Design-Build-Test-Learn cycles will involve a close interplay of biochemical experiments and model development to obtain robust, generalizable and interpretable ML models that predictively enable selective protein editing. The project lies at the interface of chemistry, biophysics, robotics, computer science, and statistics, and will therefore offer unique training and research opportunities for students from both quantitative and biological and chemical science backgrounds. Teams of undergraduates (recruited from various Research Experiences for Undergraduates programs at Rutgers) will use gamified versions of computational protein design algorithms to develop and refine protein editors. Hands-on protein design sessions will be held at affinity group conferences to reach students from groups underrepresented in STEM and to present them with opportunities to learn about, and participate in, research.New ML approaches for protein modeling have recently demonstrated that sequence and structure data can be leveraged to correctly learn complex sequence-structure relationships and design novel proteins. However, abundant functional and biophysical data, e.g., protein-peptide binding data, remain largely unexploited. New ML models that go beyond sequence representations and towards semantically richer representations based on molecular structure are needed. These representations can capture more specific sequence-function and sequence-energetics relationships. The research team will use a Design-Build-Test-Learn pipeline driven by ML models that are trained on both experimental data and molecular structure and energetics. The team will apply these methods to the design of PDZ domains that bind with high affinity and specificity to the C-termini of target intrinsically disordered proteins and selectively cleave at an internal sequence site in the targeted protein. ML models will be pre-trained on available sequence and experimental biophysical data on PDZ and protease domains and used for generating new proteins. Design validation and iterative improvements to the ML models will be carried out by experimental characterization using high-throughput assays, deep sequencing, biomolecular NMR and robotics-enabled assays of binding and cleavage. Successful designed proteins will be fused to obtain PDZ-protease protein editors. The developed methodology will be general and broadly applicable to allow in situ protein editing of a variety of biologically relevant targets, especially with intrinsically disordered proteins or regions. The novel reagents will help illuminate the so-called “dark” proteome. This project is supported by the Division of Chemistry (CHE) and the Division of Mathematical Sciences (DMS) in the Mathematics and Physical Sciences (MPS) Directorate and by the the Division of Information and Intelligent Systems (IIS) in the Computer and Information Science and Engineering (CISE) Directorate at the National Science Foundation.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.