New Bayesian statistical mechanical approaches to integrative structural biology using unassigned NMR and mass spectrometry

蛋白质数据库中的125，000多个结构为我们提供了一些关于生物学如何在分子水平上发挥作用的最丰富的见解。蛋白质是生命的小机器。这些微小的分子负责复制我们的DNA，将食物转化为能量，收缩我们的肌肉，并在我们的神经系统中传导电信号-以及许多其他活动。每个蛋白质的功能取决于其结构：每个蛋白质折叠成的特定三维形状，由其氨基酸序列编码。这些结构有助于揭示生命如何在分子水平上发挥作用的非凡故事。然而，这个故事是不完整的，因为有许多蛋白质和蛋白质复合物（蛋白质组粘在一起，形成更大的结构），我们仍然不知道。这是因为并非所有的蛋白质或复合物都适合结构生物学家常规使用的实验技术。但机器学习应用的最新进展有望改变这一点。谷歌DeepMind的一个团队最近开发了一种名为AlphaFold 2的计算机程序，可以非常准确地预测蛋白质结构。在许多情况下，预测的结构非常准确，有时甚至与通过黄金标准实验方法确定的结构一样好。然而，在其他一些情况下，预测会在一些细节上出错，而这些细节对于某些任务来说可能至关重要，比如寻找新药。更糟糕的是，一些预测是完全错误的。因此，一个紧迫的挑战是开发能够快速验证预测的方法，即快速确定预测是否可信。另一个挑战是，虽然AlphaFold 2擅长预测单个蛋白质的结构，但它在生物化学和细胞生物学中普遍存在的多蛋白质复合物上并不总是准确的。该提案的主要目的是开发新的计算和实验方法，通过快速验证结构并将其组装成复合物，从而利用这一引人注目的新预测工具。我们的方法是集中在核磁共振和交联质谱实验结合先进的计算机建模。这些实验可以应用于各种生化系统，但数据可能很难解释。我们正在开发建模工具，使用贝叶斯统计和统计力学的工具来理解这些具有挑战性的数据。解决这些挑战将为蛋白质结构测定开辟新的途径。该项目的长期目标是为结构生物学提供一系列新方法，使我们能够揭示目前隐藏的蛋白质结构。通过提供新的工具，我们将有助于更好地了解生命在分子水平上的功能，并为药物发现和蛋白质工程提供新的途径。

The 125,000+ structures in the protein databank provide some of our richest insights into how biology functions at the molecular level. Proteins are life's little machines. These tiny molecules are responsible for replicating our DNA, turning food into energy, contracting our muscles, and conducting electrical signals in our nervous systems-among many other activities. The function of each protein depends on its structure: the specific three-dimensional shape that each protein folds into, which is encoded by its amino acid sequence. These structures have helped to reveal the extraordinary story of how life functions at the molecular level. This story, however, is incomplete because there are many proteins and protein complexes (groups of proteins that stick together and form larger structures) that we still do not know. This is because not all proteins or complexes are amenable to the experimental techniques that structural biologists routinely use. But recent advances in the application of machine learning are poised to change this. A team at Google's DeepMind recently developed a computer program called AlphaFold2 that can predict the protein structures with remarkable accuracy. In many cases predicted structures are very accurate, sometimes even as good as structures determined through gold-standard experimental approaches. However, in some other cases, the predictions get some details wrong, and these details can be critical important for some tasks, like searching for new drugs. Even worse, some fraction of predictions is simply wrong. Thus, one pressing challenge is to develop methods that can rapidly validate predictions, that is, to determine quickly if they can be trusted or not. Another challenge is that while AlphaFold2 excels at predicting the structures of individual proteins, it is not always accurate on multi-protein complexes, which are ubiquitous in biochemistry and cell biology. The main aim of this proposal is to develop new computational and experimental approaches that can leverage this remarkable new prediction tool by rapidly validating structures and assembling them into complexes. Our approach is focused on nuclear magnetic resonance and cross-linking mass spectrometry experiments in combination with sophisticated computer modeling. These experiments can be applied to a variety of biochemical systems, but the data can be difficult to interpret. We are developing modeling tools that use tools from Bayesian statistics and statistical mechanics to make sense of this challenging data. Solving these challenges will unlock new avenues for protein structure determination. The long-term goal of this project is to enable a range of new approaches to structural biology that will allow us to reveal protein structures that are currently hidden. By providing new tools, we will contribute to a better understanding of how life functions at the molecular level and provide new avenues for drug discovery and protein engineering.