Acquiring a GPU server to accelerate developing deep learning methods to reconstruct protein structures from cryo-EM data

项目摘要 本补充资料的目标是购买配备8个NVIDIA A100显卡的戴尔高性能计算服务器 处理单元(GPU)加速深度学习方法的开发以重建蛋白质 来自低温电子显微镜(CRYO-EM)的结构图像数据准确、自动。冷冻机 技术可以确定由以下组成的大型蛋白质复合体和组件的四级结构 许多链对于传统技术来说是困难的，甚至是不可能的，如X射线结晶学或 核磁共振(核磁共振)测定。因为低温电磁技术通常达到高分辨率 近年来，它给结构生物学领域带来了革命性的变化，并被广泛用于确定结构 大型蛋白质复合体和集合体。然而，蛋白质结构的计算重建 冷冻-EM图像数据仍然是一个耗时和劳动密集型的过程。先进的人工智能 (AI)深度学习等方法是自动化过程和改进重建的关键 精确度。本补充计划的家长R01拨款旨在开发尖端深度学习模式，如 2D和3D转换器可自动完成从冷冻-EM数据重建蛋白质结构的关键任务：(1) 在冷冻-EM图像(显微照片)中拾取蛋白质颗粒；(2)对由蛋白质构建的低温-EM密度图进行去噪 粒子图像；(3)从冷冻-EM密度图重建蛋白质结构；以及(4)集成这些方法 (1)、(2)和(3)作为从冷冻-EM图像自动重建高精度蛋白质结构的管道 无需人工干预的数据。 我们在这个项目的前八个月取得了实质性的进展，这表明拟议的 这些方法完全可行，前景看好。然而，大规模深度学习的训练和测试 在大型低温电磁数据集上高效地建立变压器模型需要大量的GPU计算 权力。使用我们目前可用的GPU资源，开发人员需要大约一年的时间才能完成 开发一种深度学习方法。尽管速度可以带来显著的进步，但还不够快。 最大限度地发挥母公司R01项目尖端深度学习方法的潜力和影响。这 补充将使我们能够获得由8个NVIDIA A100 80 GB组成的高性能计算服务器 GPU，大大加快了母公司R01项目的研究速度。这种GPU服务器可以减少计算时间 完成一个深度学习模式的开发，从大约一年减少到不到两个月；以及 从而极大地提高了开发人员的生产力，大大加快了发布和发布的速度 在这个项目中开发的方法和工具。此外，每个GPU的大内存(80 GB)将支持 美国将培训由数百万个参数组成的高质量深层变压器，以最大限度地提高 从低温电子显微镜图像数据重建蛋白质结构。

Project Summary The goal of this supplement is to acquire a Dell high-performance computing server with 8 Nvidia A100 Graphics Processing Units (GPUs) to accelerate the development of deep learning methods to reconstruct protein structures from cryogenic electron microscopy (cryo-EM) image data accurately and automatically. The cryo-EM technology can determine the quaternary structure of large protein complexes and assemblies consisting of many chains that are difficult or even impossible for traditional techniques such as X-ray crystallography or nuclear magnetic resonance (NMR) to determine. As the cryo-EM technology routinely reached high resolution in recent years, it has been revolutionizing the field of structural biology and widely used to determine structures of large protein complexes and assemblies. However, the computational reconstruction of protein structures from cryo-EM image data is still a time-consuming and labor-intensive process. The advanced artificial intelligence (AI) methods such as deep learning hold the key to automate the process and improve the reconstruction accuracy. The parent R01 grant of this supplement aims to develop cutting-edge deep learning models such as 2D and 3D transformers to automate the key tasks of reconstructing protein structures from cryo-EM data: (1) picking protein particles in cryo-EM images (micrographs); (2) denoising cryo-EM density maps built from protein particle images; (3) reconstructing protein structures from cryo-EM density maps; and (4) integrating the methods of (1), (2) and (3) as a pipeline to automatically reconstruct high-accuracy protein structures from cryo-EM image data without human intervention. Our substantial progress in the first eight months of this project has demonstrated that the proposed methods are fully feasible and highly promising. However, training and testing the large deep learning transformer models on big cryo-EM datasets efficiently and effectively need a large amount of GPU computing power. Using the current GPU resource available to us, it takes about one year for a developer to complete the development of one deep learning method. Although the speed can yield significant progress, it is not fast enough to maximize the potential and impact of the cutting-edge deep learning methods of the parent R01 project. This supplement will enable us to acquire a high-performance computing server consisting of 8 Nvidia A100 80GB GPUs to drastically speed up the research in the parent R01 project. This GPU servers can reduce the time of completing the development of one deep learning model from about one year to less than two months, and therefore drastically improve the productivity of the developers and greatly accelerate publishing and releasing the methods and tools developed in this project. Moreover, the large (80GB) memory of each GPU will enable us to train high-quality deep transformers consisting of millions of parameters to maximize the accuracy of reconstructing protein structures from cryo-EM image data.