EAGER: Covariational Deep Learning for Protein Structure Prediction

该项目将提高目前蛋白质结构预测的能力。只有大约三分之一的已知蛋白质家族没有实验上可用的结构，可以进行同源建模；也就是说，有其他具有足够相似序列的蛋白质家族的结构已经在实验实验室中解决了。对于大多数已知的蛋白质家族，即所谓的暗蛋白质组，情况并非如此。建筑物不见了。能够通过实验或计算获得它们是了解蛋白质在关键细胞蛋白质中的作用、详细了解分子机制、指导治疗开发工作、设计具有特定功能的蛋白质等的关键。这个项目将用能够利用隐藏在蛋白质序列中的有用信号的新的信息学技术来推动暗蛋白质组的这一努力。该项目将评估多序列比对中的协变信号可以被利用来推进自由建模的假设。深度神经网络结构将被用于此目的。研究活动分为两个阶段：(1)通过对使用2D深度完全残差网络(FRN)预测的残基间距离界限进行比对，开发远距离同源折叠识别方法；(2)开发利用1D深度残差神经网络(ResNets)预测的每残基距离误差驱动的蛋白质模型质量估计方法。该项目使从事计算和生物学研究的不同社区的研究人员受益。计划的活动包括免费传播新的生物信息学工具和研究成果，通过创造性的指导和推广扩大K-12学生对计算机的参与，以及通过塞缪尔·吉恩工程学院的GING播客系列增加公众对跨学科科学的理解。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This project will advance current capabilities in protein structure prediction. Only about a third of known protein families with no experimentally-available structures are amenable to homology modeling; that is, have other proteins with sufficiently similar sequence profiles for which structures have been resolved in experimental laboratories. For the majority of known protein families, the so-called dark proteome, this is not the case. Structures are missing. Being able to obtain them experimentally or computationally is key to understanding the roles of proteins in key cellular proteins, obtaining a detailed view of molecular mechanisms, guiding efforts on therapeutic development, engineering proteins with specific functions, and more. This project will advance such efforts for the dark proteome with novel informatics techniques that are capable of harnessing useful signals hidden in protein sequences. The project will evaluate the hypothesis that covariational signals in multiple sequence alignment can be harnessed to advance free modeling. Deep neural network architectures will be utilized for this purpose. Research activities are organized in two thrusts: (1) development of distant-homology fold recognition methods by alignment of inter-residue distance bounds predicted using 2D deep fully residual networks (FRNs); and (2) development of protein model quality estimation methods driven by per-residue distance errors predicted using 1D deep residual neural networks (ResNets). The project benefits researchers in diverse communities that are working at the interface of computing and biology. Planned activities include free dissemination of novel bioinformatics tools and research results, broadening of participation of K-12 students in computing through creative mentoring and outreach, and increasing public understanding of interdisciplinary science via Samuel Ginn College of Engineering’s GINNing podcast series.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.