Supramolecular structure predictions validated from sparse experimental data

从生物学到先进材料，解开在大分子组装中遇到的复杂结构对于功能理解至关重要。对于生物分子(如蛋白质和DNA)，高分辨率结构测定技术(如结晶学和冷冻电子显微镜)对于结构-功能研究是不可或缺的。然而，强大的基于深度学习的高精度结构预测工具的出现给结构生物界带来了冲击波，并预示着结构研究的新纪元，在这个时代，繁重的实验高分辨率结构的常规生成可以被计算预测所取代。这些预测可以在实验验证和改进的基础上形成设计结构-功能研究的基础。被称为电子顺磁共振(EPR)谱的(生物)物理工具非常适合补充预测的结构。EPR探测由“自旋”引起的磁性，“自旋”是未成对电子的一种量子力学性质。电子包含在所有物质中，通常是成对的，使它们的磁性消失。然而，自由基等不成对的电子支撑着许多重要的生物过程，如光合作用、衰老和呼吸作用。利用EPR，这种自转之间的距离可以在纳米(十亿分之一米)的尺度上确定。在过去的20年里，这些距离测量已经发展成为研究复杂(生物)分子纳米世界的重要和强大的方法。分子生物学和化学可以通过选择性地引入自旋来标记生物分子中的特定位置，然后这些自旋可以用作分子的“灯塔”。引入两个这样的信标可以测量它们之间的距离。在这个项目中，基于深度学习的结构预测和建模工具将与最新的EPR技术(包括用于低浓度RIDME的正交铜(II)-SLIM标记和基于无偏深度学习的数据处理)相结合，以验证和细化蛋白质的结构模型，从而规避实验高分辨率结构确定。基于纯计算、高精度的结构预测，有可能生成蛋白质的信息EPR结构，其中分子信标将报告对结构和功能过程中的结构转变至关重要的特征。不同信标之间的距离将被用来反馈到结构模型中进行验证和改进。在功能过程中与结合伙伴的相互作用会导致结构变化，从而改变信标的距离和相对方向。用EPR确定这些变化将显示这种方法的潜力，并展示其产生广泛影响的机会。人工智能正越来越多地影响我们日常生活的许多方面。同样，深度学习彻底改变了结构性研究的执行方式。该项目展示了基于深度学习的结构预测与使用EPR的结构改进和验证之间的结合的好处。这里建立的方法和工作流程是完全可移植的，扩大了EPR在结构-功能研究中的应用范围，特别是关于目前无法实现的挑战性系统(由于其规模、复杂性、灵活性、膜环境或可实现的数量或浓度)。在这里，该方法被应用于一种结构未知的细菌表面蛋白，该蛋白与风湿性心脏病有关，拟议的实验有可能揭示宿主-病原体相互作用的结构机制。只有对生物纳米世界有了更多的了解，才有可能查明疾病的分子原因，并有助于制定预防和治疗战略。

Unravelling the complex structures encountered in macromolecular assemblies from biology to advanced materials is paramount to functional understanding. For biomolecules (such as proteins and DNA) high-resolution structure determination techniques (such as crystallography and cryo-electron microscopy) have been indispensable for structure-function studies. However, the emergence of powerful deep learning based high-accuracy structure prediction tools has sent shock waves through the structural biology community and heralds a new era for structural studies where the routine generation of laborious experimental high-resolution structures could be replaced with computational predictions. These predictions can form the basis to design structure-function studies upon experimental validation and refinement. The (bio)physical tool called electron paramagnetic resonance (EPR) spectroscopy is ideally suited to complement predicted structures. EPR detects the magnetism arising from the "spin", a quantum mechanical property of unpaired electrons. Electrons are contained in all matter and are commonly paired, quenching their magnetism. However, unpaired electrons such as free radicals underpin many important biological processes like photosynthesis, ageing, and respiration. Using EPR, distances in-between such spins can be determined on the nanometre (one billionth of a metre) scale. Over the past 20 years, these distance measurements have developed into an important and powerful method for investigating the nanoworld of complex (bio)molecules. Molecular biology and chemistry allow labelling specific sites in biomolecules by selectively introducing spins that can then be used as molecular "beacons". Introducing two such beacons allows measurement of the distance between them. With this approach structures of proteins and other macromolecules are successfully mapped, validated and refined.In this project, deep learning-based structure prediction and modelling tools will be combined with state-of-the-art EPR techniques (including orthogonal copper(II)-SLIM labelling for low-concentration RIDME and unbiased deep learning-based data processing), to validate and refine the structural model of a protein evading experimental high-resolution structure determination. Based on purely computational, high-accuracy structure prediction it is possible to generate informative EPR constructs of the protein where the molecular beacons will report on features critical for structure and structural transitions during function. The distances between different beacons will be used to feed back into the structural model for validation and refinement. Interaction with binding partners during function leads to structural changes which alter distance and relative orientation of beacons. Determination of these alterations with EPR will show the potential of this approach and demonstrate its opportunities for wide-reaching impact.Artificial intelligence is increasingly affecting many aspects of our everyday lives. Similarly, deep learning revolutionises the way structural studies are performed. This project showcases the benefits of the marriage between deep learning-based structure prediction and structural refinement and validation using EPR. The approach and workflows established here are fully transferable, widening the application scope of EPR for structure-function studies, especially regarding challenging systems currently beyond reach (owing to their size, complexity, flexibility, membrane environment or achievable amount or concentration). Here, the approach is applied to a bacterial surface protein of unknown structure implicated in rheumatic heart disease, and proposed experiments have the potential to uncover the structural mechanism of the host-pathogen interaction. Only with a greater knowledge of the biological nanoworld will it be possible to pinpoint the molecular causes of diseases, and aid in developing prevention and treatment strategies.