Computational design of novel protein binders based on structure mining and learning from data

我们的长期目标是将计算蛋白质设计变成一个颠覆性的技术平台， 能够常规和快速地产生用于检测蛋白质或干扰其功能的试剂。目前， 研究和治疗依赖于用于这些任务的小分子和/或抗体。这些都是强大的工具，但 它们的开发可能缓慢而昂贵，而且不能满足所有需求。设计肽或微型蛋白质 具有作为标记物结合细胞外或细胞内靶的高潜力（例如，用于成像）或作为功能 调制器（例如，相互作用抑制剂），用于基础和临床研究以及疾病诊断 和治疗。用于设计这种定制蛋白的现有工具依赖于实验文库筛选， 有时由结构的计算模型引导或支持。尽管价值巨大， 分子将带来基础生物医学研究和治疗发展，目前还没有快速， 获得具有所需性质的设计蛋白质的简便途径。计算方法可能 虽然这些技术可以满足这一需求，但现有技术对于日常使用来说还不够可靠、灵活或自动化。 与20世纪90年代中期相比，当现代计算蛋白质设计方法被开发出来时，我们 生活在一个数据和技术赋能的时代。这项建议的前提是，我们可以增加 这些问题可以通过计算设计来解决，并大大提高成功率， 充分利用编码在已知天然结构中的序列-结构相容性的已证实规则， 它们的同源序列蛋白质数据库（所有已知蛋白质结构的集合）已经增长 10-自2000年以来翻了一倍，使我们能够通过构建新蛋白质来设计新蛋白质， 自然界中使用的积木。我们已经实现了一个新的设计框架，该框架基于这一原则， 在基本方面不同于所有以前公布的替代品。对不同任务的测试表明， 非常成功为了进一步发展我们的方法，我们提出了方法上的进步，我们将 实施、测试并应用于蛋白质设计挑战，包括检测或抑制蛋白质识别 域.我们将开发和应用的方法：自动识别绑定到目标的设计策略 蛋白质，评分和排名具体的设计候选人，设计库，将被筛选，以提供丰富的 成功和失败的实验数据，并自动将实验数据反馈到模型中 以有原则的方式发展。成果将包括与社区分享的新方法， 高等级界面设计站点的计算预测，可以为结构分析提供信息， 途径，以及实验验证的设计分子，这些分子与对信号传导重要的蛋白质结构域结合， 疾病的传播途径。

Our long-term objective is to turn computational protein design into a disruptive technology platform that will enable the routine and rapid generation of reagents for detecting proteins or perturbing their functions. Currently, research and therapy rely on small molecules and/or antibodies for these tasks. These are powerful tools, but they can be slow and expensive to develop, and they do not meet all needs. Designer peptides or mini-proteins have high potential to bind extracellular or intracellular targets either as labels (e.g., for imaging) or as functional modulators (e.g., interaction inhibitors), for applications in basic and clinical research and in disease diagnosis and treatment. Existing tools for designing such custom proteins rely on experimental library screening, sometimes guided or supported by computational modeling of structure. Despite the immense value such molecules would bring to basic biomedical research and therapeutic development, there are not yet rapid and facile routes to obtaining designed proteins with desired properties. Computational methods can potentially address this need, but existing technology is not sufficiently reliable, flexible or automated for routine use. Compared to the mid-1990’s, when the modern approach to computational protein design was developed, we live in a data and technology-empowered age. The premise of this proposal is that we can increase the range of problems that can be solved using computational design, and also dramatically improve success rates, by making full use of the proven rules of sequence-structure compatibility encoded in known natural structures and their homologous sequences. The Protein Data Bank (the collection of all known protein structures) has grown 10-fold since 2000, placing us at a point where we can design novel proteins by constructing them from building blocks used in nature. We have implemented a new design framework that is based on this principle and that is different in fundamental aspects from all previously published alternatives. Tests on diverse tasks demonstrate outstanding success. To further develop our approach, we propose methodological advances that we will implement, test and then apply to protein design challenges involving detecting or inhibiting protein recognition domains. We will develop and apply methods to: automatically identify design strategies for binding to a target protein, score and rank specific design candidates, design libraries that will be screened to provide rich experimental data about successes and failures, and automatically feed experimental data back into model development in a principled way. Outcomes will include new methodology that will be shared with the community, computational predictions of high-ranked interface design sites that can inform analysis of structures and pathways, and experimentally validated designer molecules that bind to protein domains important for signaling in disease pathways.