Development and experimental validation of a deep-learning based pipeline for user-centric protein design.

蛋白质是为所有生物提供大部分复杂功能的分子。它们由20种不同的氨基酸组成，这些氨基酸以不同的顺序组合成长链。氨基酸的不同形状和化学性质导致链折叠成独特的3D结构。正是这种结构使蛋白质在自然界中扮演着不同的角色，无论是消化食物，移动你的身体还是仅仅保持头顶的温暖。尽管生命在40多亿年前就出现了，但由于其固有的复杂性，只有少数可能的蛋白质结构被进化探索过。由于蛋白质结构与功能直接相关，这意味着有大量未开发的具有功能的蛋白质可用于解决医学、生物技术、能源和农业等领域的问题。如果我们可以从头开始设计新的蛋白质，我们就可以用我们创造的新蛋白质解决其中的一些问题。如前所述，蛋白质是复杂的，因此很难设计新的蛋白质，但为了使它更容易，我们可以编写程序，可以在计算机模拟中创建和测试大量的设计。这提高了设计氨基酸序列的机会，当我们在实验室中创建它时，它将采用我们想要的结构。然而，即使使用最先进的方法在计算机上设计蛋白质，也只有少数序列采用我们想要的结构，这使得蛋白质设计成本高昂且不可靠。我打算创造一种新的方法来设计蛋白质，这种方法使用一种称为深度神经网络的人工智能（见http://playground.tensorflow.org的交互式示例）。这项技术将用于学习生成稳定蛋白质的复杂规则，这些蛋白质隐藏在我们已经观察到的蛋白质结构的氨基酸序列中。一旦了解了规则，我们就可以用它们来创建新的氨基酸序列，这些序列是采用我们需要的结构的良好候选者。在为我们的预期应用推荐最佳设计之前，该方法将形成自动化管道的一部分，该管道将在计算机模拟中创建和测试蛋白质结构。这将使蛋白质设计过程更加可靠。为了了解这种方法的有效性，我将通过在实验室中创建数百种管道推荐的蛋白质设计来测试它，使用机器人技术来加速这一过程。一旦测试成功，我计划通过设计新的蛋白质来展示这种方法，这些蛋白质可以进行工业上有用的化学反应。这将使进行这些化学反应更便宜，更环保，为设计更多具有有用功能的蛋白质铺平道路，以解决人类目前面临的挑战。

Proteins are the molecules that provide most of the complex functionality in all living things. They are made of 20 different building-block types called amino acids, which are combined in different sequences to make long chains. The varying shapes and chemistries of the amino acids cause the chains to fold into a distinct 3D structure. It is this structure that enables proteins to perform the different roles they have in nature, whether it's digesting your food, moving you around or simply keeping the top of your head warm.Even though life emerged over 4 billion years ago, only a small number of possible protein structures have been explored by evolution due to their inherent complexity. As protein structure is directly related to function, this means that there is a huge pool of unexplored proteins with functions that could be applied to solve problems in medicine, biotechnology, energy and agriculture. If we can design new proteins from scratch, we can address some of these problems with the new proteins that we create.As mentioned previously, proteins are complex, and so it is difficult to design new proteins, but to make it easier we can write programs that can create and test huge numbers of designs in computer simulations. This improves the chance of designing a sequence of amino acids that will adopt our desired structure when we create it in the laboratory. However, even with state-of-the-art methods for designing proteins on a computer, only a small number of sequences adopt the structures we intend them to, making protein design costly and unreliable.I intend to create a new method for designing proteins that uses a type of artificial intelligence called a deep-neural network (see http://playground.tensorflow.org for an interactive example). This technique will be used to learn the complex rules for generating stable proteins that are hidden inside the amino-acid sequences of protein structures we have already observed. Once the rules have been learned, we can use them to create new sequences of amino acids that are good candidates for adopting the structure we require. This method will form part of an automated pipeline that will create and test protein structures in computer simulations, before recommending the best designs for our intended application. This will make the process of protein design much more reliable.To get an understanding of how effective this method is, I will test it by creating hundreds of the protein designs recommended by the pipeline in the laboratory, using robotics to accelerate this process. Once tested, I plan to showcase the method by designing new proteins that can perform chemical reactions that are useful industrially. This will make performing these chemical reactions much cheaper and more environmentally friendly, paving the way for the design of many more proteins with useful functions that address the challenges that the human race currently faces.