Nonlinear dimensionality reduction and enhanced sampling in molecular simulation using auto-associative neural networks

伊利诺伊大学香槟分校的安德鲁·弗格森获得了化学部化学理论、模型和计算方法项目的支持，以建立新的理论方法和计算工具，以加速蛋白质折叠的分子模拟。该项目是由材料研究部凝聚态物质和材料理论计划共同资助的。蛋白质是执行生命基本功能的分子主力。蛋白质已经进化成能够完成这些任务的形状。确定蛋白质的形状和运动有助于揭示它是如何工作的，并为如何设计新的蛋白质来帮助治疗疾病、生产生物燃料或制造新材料提供信息。蛋白质的计算机模拟非常有用，因为它们可以识别所有组成原子的精确位置和运动。然而，对于除了最小的蛋白质以外的所有蛋白质来说，即使使用强大的超级计算机，它也太过计算密集，无法准确预测它们的结构和运动。已经开发出了加速这些模拟的方法，但要想很好地发挥作用，它们需要对蛋白质将进行的结构重排做出良好的估计。这是一个问题，因为这通常是模拟试图回答的问题。在这项工作中，弗格森教授正在开发一种新的方法，使用一种被称为人工神经网络的机器学习来加速蛋白质折叠模拟。人工神经网络之所以被称为人工神经网络，是因为它们是基于大脑中神经元的结构。神经网络允许计算机确定这些重要的结构路径，并使用它们来使模拟运行得更快。这一新方法正被用来帮助理解与癌症和艾滋病毒感染有关的大蛋白。它还被整合到流行的模拟软件中，可供免费公开下载。弗格森教授正在为本科生提供研究机会，与他在这个项目上合作，他正在开发计算材料科学的动手研讨会，作为伊利诺伊大学女孩学习材料(GLAM)夏令营的一部分。这项工作的目的是建立一种非线性机器学习方法来发现蛋白质折叠的集体变量，并使用这些变量在分子动力学模拟中执行增强采样。增强的采样技术在加速构象采样方面的成功是基于与重要的分子运动相关的良好的集体变量(CV)的可用性。现有的非线性降维技术(如扩散图、ISOMAP、陆地局部线性嵌入)可以很好地发现好的简历，但没有提供显式的坐标映射，因此必须在代理变量中进行低效和间接的有偏采样。这项工作建立了一种新的基于自联想人工神经网络(“自动编码器”)的增强型采样方法，以发现作为原子坐标的显式可微函数的CV，并允许计算解析偏向力。这种方法被称为MESA(具有自动编码器的分子增强采样)。MESA在短肽丙氨酸二肽和色氨酸-笼子上得到验证，并被用于发现在许多癌症中过度表达的激酶和HIV表面存在的包膜蛋白的亚稳状态和结构转变。通过与OpenMM的开发者合作，MESA被分子模拟社区广泛使用，并致力于为这些软件包的未来发布做出贡献。积极的研究经验对本科生的成功和留住有很大的好处，这项工作支持在每年的奖项期间有10周的暑期研究机会。弗格森教授还在为伊利诺伊大学非常成功的女孩学习材料(GLAM)夏令营开发新的令人兴奋的内容，以在女高中生中演示和推广计算材料科学，并提高STEM学位课程的女性入学率。

Andrew Ferguson of the University of Illinois at Urbana-Champaign is supported by an award from the Chemical Theory, Models and Computational Methods Program in the Chemistry Division to establish new theoretical approaches and computational tools to accelerate molecular simulations of protein folding. This project is cofunded by the Condensed Matter and Materials Theory Program in the Division of Materials Research. Proteins are molecular workhorses that perform the essential functions of life. Proteins have evolved to adopt shapes that enable them to do these tasks. Determining the shape and motions of a protein can help reveal how it works and inform how to design new proteins to help treat disease, produce biofuels, or make new materials. Computer simulations of proteins are very useful in that they can identify the precise locations and motions of all the constituent atoms. For all but the smallest proteins, however, it is too computationally intensive to accurately predict their structure and motions even with powerful supercomputers. Ways to accelerate these simulations have been developed, but to work well they need good estimates of the structural rearrangements that the protein will make. This is a problem, since this is usually the question the simulations are trying to answer. In this work, Professor Ferguson is developing a new approach to accelerate protein folding simulations using a type of machine learning known as artificial neural networks so-called because they are based on the structure of neurons in the brain. Neural networks allow computers to both determine these important structural pathways and use them to make simulations run faster. This new approach is being used to help understand large proteins involved in cancer and HIV infection. It is also being incorporated into popular simulation software available for free public download. Professor Ferguson is providing research opportunities for undergraduates to work with him on this project and he is developing hands-on workshops in computational materials science as part of the Girls Learning About Materials (GLAM) summer camp at the University of Illinois. The aim of this work is to establish a nonlinear machine learning approach to discover collective variables for protein folding and to use these variables to perform enhanced sampling in molecular dynamics simulations. The success of enhanced sampling techniques in accelerating conformational sampling is predicated on the availability of good collective variables (CVs) correlated with important molecular motions. Existing nonlinear dimensionality reduction techniques (e.g., diffusion maps, Isomap, land ocally linear embedding) can ably discover good CVs, but do not furnish the explicit coordinate mapping so that biased sampling must be conducted inefficiently and indirectly in proxy variables. This work establishes a new enhanced sampling approach based on auto-associative artificial neural networks ("autoencoders") to discover CVs that are explicit differentiable functions of the atomic coordinates and to permit calculation of analytical biasing forces. This approach is termed MESA (Molecular Enhanced Sampling with Autoencoders). MESA is validated on the short peptides alanine dipeptide and tryptophan-cage, and deployed to discover metastable states and structural transitions in a kinase overexpressed in many cancers and an envelope protein presented on the surface of HIV. MESA is made broadly available to the molecular simulation community by collaborating with the developers of OpenMM and PLUMED to contribute the approach to future releases of these software packages. Positive research experiences have great benefits for undergraduate success and retention and this work supports 10-week summer research opportunities during each year of the award. Professor Ferguson is also developing new and exciting content for the highly successful Girls Learning About Materials (GLAM) summer camp at the University of Illinois to illustrate and promote computational materials science among female high school students and elevate female enrollment in STEM degree programs.