CRII: III: Deep Learning Methods for Protein Inter-residue Distance Prediction

蛋白质是生命的齿轮。我们牢房里的大部分工作都是由它们完成的。胰岛素和血红蛋白就是两个例子。与机械齿轮相似，蛋白质的功能是因为它们独特而精确的三维结构。例如，由于其精确的结构，血红蛋白在我们的体内携带氧气。了解蛋白质的精确结构有助于我们更好地了解生命过程和疾病机制。这种知识也使新药的设计成为可能。在实验室中确定蛋白质的结构可能需要几个月到几年的时间。这些实验有时会失败，没有结果。作为另一种选择，基于人工智能(AI)的计算机算法经常被用于预测结构。在过去的50多年里，计算预测蛋白质结构一直是最重要和最基本的挑战之一。最近，被称为卷积神经网络的特定类型的人工智能方法被发现是解决这个问题的最有效的方法。然而，即使是最好的人工智能方法也不足以处理许多蛋白质。目前，哪些类型的人工智能算法最适合解决这一问题是一个悬而未决的问题。该项目将推动目前基于人工智能的蛋白质结构预测方法的进展。该项目的主要新颖性将是开发特别适用于蛋白质结构相关任务的基于人工智能的方法，并调查哪些输入变量驱动可预测性。该项目将增加对哪种人工智能算法在解决基本生物学问题方面有效的理解，从蛋白质结构预测开始。最终，这将导致对许多不治之症的更好理解，并使治疗学得到更快的发展，这将对个人、社区和国家的健康和福利做出积极贡献。蛋白质结构预测领域的当前进展来自深度学习方法和从高通量测序获得的大量蛋白质序列的可用性。深度学习方法可以更准确地预测残基之间的接触和距离，这是结构预测的关键信息。尽管最近取得了进展，但总体问题仍远未得到解决。虽然卷积神经网络和残差网络的许多变体已经被用来解决接触和距离预测问题，但许多其他类型的深度学习方法的潜力还没有被开发出来，例如胶囊网络。工作范围设定为使用蛋白质残基间距离预测的新方法、改进的特征工程和改进的高质量多序列比对的选择来实现高精度的蛋白质结构预测。这项工作将使用代表蛋白质结构数据库的各种尺度的数据集--小的、中的和大的。这项研究将开发利用现有的和新的胶囊网络变体来预测蛋白质距离的新方法，用于预测多序列比对质量的深度学习方法，并调查哪些变量驱动深度学习模型的准确性。这些研究工作将增加我们对各种深度学习算法在解决蛋白质折叠问题中的适用性和局限性的理解。这些发现也将有助于解决计算生物学中的类似问题。这一奖项反映了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。

Proteins are the gears of life. They do most of the work in our cells. Insulin and hemoglobin are two examples. Similar to mechanical gears, proteins function because of their unique and precise three-dimensional structure. For example, hemoglobin carries oxygen in our body because of its precise structure. Learning the precise structures of proteins helps us to better understand life processes and disease mechanisms. This knowledge also enables the design of novel drugs. Determining the structures of a protein in a lab can take months to years. These experiments sometimes fail with no results. As an alternative, artificial intelligence (AI) based computer algorithms are often used to predict structures. Computationally predicting protein structures has been one of the most significant and fundamental challenges during the past 50+ years. Recently, specific types of AI methods, known as convolutional neural networks, have been found to be most effective for solving this problem. However, even the best AI methods are inadequate for many proteins. It is currently an open question which types of AI algorithms are best suited for solving this problem. This project will push forward the current advances in AI-based methods for predicting structures of proteins. The main novelty of this project will be in developing AI-based methods particularly suitable for protein structure-related tasks and investigating what input variables drive the predictability. The project will increase the understanding of which kinds of AI algorithms are effective for solving fundamental biological problems, beginning with protein structure prediction. Ultimately, this will lead to a better understanding of many incurable diseases and enable a much faster development of therapeutics that will positively contribute to health and welfare of individuals, communities, and the nation.Current progress in the field of protein structure prediction comes from deep learning methods and the availability of a large number of protein sequences obtained from high throughput sequencing. Deep learning methods can predict inter-residue contacts and distances, the key information for structure prediction, much more accurately. Despite recent advances, the overall problem remains far from being solved. While many variants of convolutional neural networks and residual networks have been investigated for solving the problem of contact and distance prediction, the potential of many other types of deep learning methods such as capsule networks have not been explored. The work scope is set to achieve high accuracy protein structure prediction using novel methods for protein inter-residue distance prediction, improved feature engineering, and improved selection of high-quality multiple sequence alignments. This work will use datasets at various scales - small, medium, and large - that are representative of the protein structure database. The research will develop novel methods for protein distance prediction using existing and novel variants of capsule networks, deep learning methods for predicting the quality of multiple sequence alignments, and investigate what variables drive the accuracy of deep learning models. These research efforts will increase our understanding of the applicability and limitations of diverse deep learning algorithms in solving the protein folding problem. The findings will also be helpful in solving similar problems in computational biology.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.