In last month’s editorial my colleague Michael Sailor raised the interesting question of whether we will ever be able to completely computationally design a sensor from the ground up (10.1021/acssensors.1c01189). His short answer was probably not, as the multiparameter design space is likely too large for this to become reality. I agree with him, although this might also be because as humans we like to think that we are unique and indispensable. Of course this does not mean that computational design/modeling could not play a more important role in sensor engineering, and Michael also already mentioned some interesting examples where collaboration of computational scientists with experimental scientists and engineers proved very successful. One area where I believe the application of deep-learning and other advanced computational modeling is going to revolutionize our field is in the development of biosensors. Traditional biosensors rely on either the use of native proteins with a natural affinity and specificity for the analyte of interest (such as the enzymes used in glucose sensing) or the development of antibodies and other affinity reagents such as aptamers. In particular, antibodies have become the reagents of choice for many biosensors, as high-affinity antibodies can be developed for almost any molecular target. However, despite the development of modern protein engineering approaches such as directed evolution and in vitro selection procedures, the development of high affinity binders remains a technically complex and time-consuming process of sophisticated trial-and-error. If computational methods would be good enough to (1) reliably predict the optimal binding site for a specific analyte and (2) be able to predict the corresponding primary sequence that would fold into the desired 3-dimensional structure, obtaining a ligand binding domain would be as simple as ordering a piece of synthetic DNA. Regarding the second challenge, a team from Google’s DeepMind recently reported that their AlphaFold-2 program based on deep-learning algorithms achieved unprecedented accuracy in predicting the 3-dimensional structure of proteins solely based on their amino acid sequence.(1) The success of AlphaFold-2 illustrates that deep-learning approaches are particularly suited for complex processes such as protein folding, where the optimal solution is determined by many variables (many weak noncovalent interactions), and reliable data are available to properly train these neural networks (3-dimensional protein structures and primary sequences). In fact, this impressive progress leads some to conclude that the so-called protein folding problem has been solved.(2,3,4) Does the success of AlphaFold mean that we can design high affinity binders from scratch? Not so fast. This also requires one to calculate what the structure of the optimal binding site should be. The challenge there is that even small structural deviations on the order of 0.1–0.3 Å can have a major effect on binding affinity, let alone binding kinetics. Also, here new deep-learning approaches are emerging, e.g., for the prediction of molecular interactions.(5). For now, the main value of such approaches is that they allow one to much more efficiently explore sequence space to arrive at sequences/structures that are likely to have at least a modest affinity and specificity. These approaches thus provide an efficient starting point for the subsequent development of high-affinity binders using directed evolution approaches. An impressive recent example is the rapid development of stable miniproteins that bind and block the receptor binding domain of the SARS-CoV-2 spike protein with pM affinity.(6) These protein domains were developed completely from scratch in only a few months by the Baker group, using a combination of de novo design and efficient directed evolution approaches. In addition to their potential therapeutic application as potent inhibitors of viral infection, the same binding domains were also used to develop bioluminescent sensor proteins that allow detection of spike protein directly in solution.(7) The latter illustrates an important advantage of computational design. In contrast to antibodies or enzymes, whose native functional and structural properties determine and to some extent limit their use in biosensing, the use of de novo designed binding proteins allows one to take important properties for sensor integration into account from the start. Ultimately, one would like to computationally design a completely protein-based sensor, which requires not only highly accurate prediction of protein structure and protein-analyte interactions, but also the accurate modeling of allostery and conformational switching. Until recently, I thought that this would not happen any time soon, but now I am not so sure anymore. This article references 7 other publications.

中文翻译：

---

传感器蛋白的计算设计；它可能真的有效

在上个月的社论中，我的同事 Michael Sailor 提出了一个有趣的问题，即我们是否能够从头开始完全通过计算设计传感器 (10.1021/acssensors.1c01189)。他的简短回答可能是否定的，因为多参数设计空间可能太大而无法实现。我同意他的观点，尽管这也可能是因为作为人类，我们喜欢认为我们是独一无二的，不可或缺的。当然，这并不意味着计算设计/建模不能在传感器工程中发挥更重要的作用，迈克尔还提到了一些有趣的例子，其中计算科学家与实验科学家和工程师的合作证明是非常成功的。我相信深度学习和其他高级计算模型的应用将彻底改变我们领域的一个领域是生物传感器的开发。传统的生物传感器依赖于使用对目标分析物（例如葡萄糖传感中使用的酶）具有天然亲和力和特异性的天然蛋白质，或者开发抗体和其他亲和试剂，例如适体。特别是，抗体已成为许多生物传感器的首选试剂，因为几乎可以为任何分子靶标开发高亲和力抗体。然而，尽管发展了现代蛋白质工程方法，例如定向进化和体外选择程序，但高亲和力结合剂的开发仍然是技术上复杂且耗时的复杂试错过程。如果计算方法足以 (1) 可靠地预测特定分析物的最佳结合位点，并且 (2) 能够预测将折叠成所需 3 维结构的相应一级序列，那么获得配体结合域将就像订购一段合成 DNA 一样简单。关于第二个挑战，谷歌 DeepMind 的一个团队最近报告说，他们基于深度学习算法的 AlphaFold-2 程序在仅根据蛋白质的氨基酸序列预测蛋白质的 3 维结构方面取得了前所未有的准确度。(1) AlphaFold-2 说明深度学习方法特别适用于复杂过程，例如蛋白质折叠，其中最佳解决方案由许多变量（许多弱非共价相互作用）决定，可靠的数据可用于正确训练这些神经网络（3 维蛋白质结构和初级序列）。事实上，这一令人印象深刻的进展让一些人得出结论，所谓的蛋白质折叠问题已经解决。(2,3,4) AlphaFold 的成功是否意味着我们可以从头开始设计高亲和力的结合剂？没那么快。这也需要计算最佳结合位点的结构应该是什么。面临的挑战是，即使是 0.1-0.3 Å 量级的微小结构偏差也会对结合亲和力产生重大影响，更不用说结合动力学了。此外，这里出现了新的深度学习方法，例如，用于预测分子相互作用。(5)。目前，这种方法的主要价值在于它们允许人们更有效地探索序列空间，以获得可能至少具有适度亲和力和特异性的序列/结构。因此，这些方法为后续使用定向进化方法开发高亲和力粘合剂提供了有效的起点。最近一个令人印象深刻的例子是稳定的微型蛋白的快速开发，这些微型蛋白以 pM 亲和力结合和阻断 SARS-CoV-2 刺突蛋白的受体结合域。 (6) 这些蛋白质域是在短短几个月内完全从头开发的。 Baker 小组，结合了从头设计和高效的定向进化方法。除了它们作为病毒感染的有效抑制剂的潜在治疗应用之外，相同的结合域也用于开发生物发光传感器蛋白，允许直接检测溶液中的刺突蛋白。(7) 后者说明了计算设计的一个重要优势。与抗体或酶相比，它们的天然功能和结构特性决定并在一定程度上限制了它们在生物传感中的应用，使用从头设计的结合蛋白允许人们从一开始就考虑到传感器集成的重要特性。最终，人们希望通过计算设计一种完全基于蛋白质的传感器，这不仅需要对蛋白质结构和蛋白质-分析物相互作用进行高度准确的预测，还需要对变构和构象转换进行准确建模。直到最近，我还以为这不会很快发生，但现在我不太确定了。本文引用了 7 篇其他出版物。