III: Small: Integrated prediction of intrinsic disorder and disorder functions with modular multi-label deep learning

蛋白质是非凡的生物机器。在过去的二十年里，数以亿计的蛋白质序列被解码，造成了一个重大的知识鸿沟，这与我们不知道它们中的大多数是做什么的事实有关。破译蛋白质功能的一种常用方法依赖于序列-结构-功能范式，其中蛋白质功能是从序列产生的蛋白质结构中学习的。然而，最近的研究已经发现了一个大的内在无序蛋白家族，在生理条件下缺乏稳定的结构，因此无法使用基于结构的方法来表征。这些蛋白质在真核生物中特别丰富，并参与许多人类疾病的发病机制。内在无序蛋白质的发现促进了新一代计算方法的发展，这些方法可以直接从蛋白质序列中预测内在无序的存在。最近完成的一项蛋白质内在失调预测关键评估（CAID）实验表明，这些方法快速且提供准确的结果。然而，虽然内在紊乱可以很容易和准确地识别蛋白质序列，但其功能仍然是一个谜。该提案将概念化，设计，实施，测试和部署一种创新的机器学习方法，该方法可以直接从蛋白质序列中提供高度准确和集成的无序和无序功能预测。该团队将利用这种方法在数千万蛋白质的前所未有的规模上产生功能失调的注释，解决该蛋白质家族的知识缺口问题。从长远来看，这个项目将促进对内在无序蛋白质背景下的基本生物过程和相关人类健康问题的理解。该项目还将通过高中外展和多学科教学以及本科生、研究生和博士后研究人员的指导，培养STEM学生和研究人员，培养受工业界和学术界追捧的高技能研究人员。该团队解决了生物信息学和机器学习领域交叉的内在无序蛋白质结构的跨学科和具有挑战性的问题。基于内在无序计算分析方面的专业知识，并专注于技术创新，该项目将提供一种新颖的深度顺序多标签变压器架构，可提供对无序和无序功能的准确预测。该解决方案将被设计为适应蛋白质数据的生物学基础，例如固有的多标签结果，不平衡标签和蛋白质数据的顺序性质。此外，该架构将采用模块化设计，以方便转移到其他领域的蛋白质和核酸生物信息学。所产生的方法将广泛确定基准并加以传播，以最大限度地发挥影响。为了满足包括生物学家、生物化学家、生物物理学家和生物信息学家在内的广泛用户的需求，代码将被存入相关的公共存储库，并利用现代在线资源（如数据存储库和网络服务器）提供预先计算的内在紊乱的功能注释。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Proteins are remarkable biological machines. Hundreds of millions of protein sequences were decoded over the last two decades creating a significant knowledge gap related to the fact that we do not know what most of them do. A common way to decipher protein functions relies on the sequence-to-structure-to-function paradigm where protein function is learned from the protein structure that is produced from the sequence. However, recent research has identified a large family of the intrinsically disordered proteins that lack a stable structure under physiological conditions and which therefore cannot be characterized using the structure-based approaches. These proteins are particularly abundant in the eukaryotes and are involved in the pathogenesis of numerous human diseases. The discovery of the intrinsically disordered proteins has prompted the development of a new generation of computational methods that predict presence of intrinsic disorder directly from protein sequences. A recently completed Critical Assessment of protein Intrinsic Disorder prediction (CAID) experiment has shown that these methods are fast and provide accurate results. However, while intrinsic disorder can be readily and accurately identified in protein sequences, its function remains a mystery. This proposal will conceptualize, design, implement, test and deploy an innovative machine learning method that provides highly accurate and integrated predictions of disorder and disorder functions directly from protein sequences. The team will utilize this method to produce functional annotations of disorder on an unprecedented scale of dozens of millions of proteins, addressing the knowledge gap problem for this protein family. In the long run this project will advance understanding of fundamental biological processes and related human health issues in the context of the intrinsically disordered proteins. This project will also train STEM students and researchers via high-school outreach and multidisciplinary teaching and mentoring of undergraduate and graduate students and postdoctoral researchers, producing highly skilled researchers who are sought after by industry and academia.An interdisciplinary and challenging problem of the structure of intrinsically disorder protein structure at the intersection of bioinformatics and machine learning fields is addressed by the team. Building on expertise in the computational analysis of intrinsic disorder and with focus on technical innovation, this project will deliver a novel deep sequential multi-label transformer architecture that provides accurate predictions of disorder and disorder functions. The solution will be designed to accommodate for the biological underpinnings of protein data, such as the inherently multi-label outcomes, imbalanced labels and sequential nature of protein data. Moreover, this architecture will feature modular design to facilitate transfer to other areas of protein and nucleic acids bioinformatics. The resulting method will be extensively benchmarked and disseminated to maximize impact. The code will be deposited into relevant public repositories and pre-computed functional annotations of intrinsic disorder will be made available using modern online resources, such as data repositories and webservers, in order to meet the needs of a broad spectrum of users including biologists, biochemist, biophysicists and bioinformaticians.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.