GPU workstation for deep learning-based protein design and cryo-EM data processing

摘要 我们正在申请资金购买GPU工作站，用于（1）深度学习 在蛋白质设计的上下文中，以及（2）用于使用冷冻EM数据求解蛋白质结构。 正如深度学习方法所表明的那样，目前蛋白质建模正在发生一场革命 在蛋白质结构预测和蛋白质设计方面都取得了令人瞩目的成功。我们正在利用 这些新方法以几种方式需要访问GPU处理器。首先我们得 开发了一种新的蛋白质设计管道，在蛋白质结构预测与 AlphaFold和序列优化与图形神经网络，以发展序列， 具体功能。该管道的初步实验验证非常令人鼓舞，我们正在 现在渴望在各种蛋白质设计问题上测试它，包括竞争性蛋白质的设计。 抑制剂、蛋白质开关和生物传感器。流水线需要大块的GPU时间 （多个处理器的几天），以确定有前途的设计进行实验验证。 其次，我们正在训练新的神经网络，以提高蛋白质设计的性能。 快速测试替代网络架构需要访问GPU， 超参数许多蛋白质设计项目的最后一步是验证设计模型 高分辨率的结构。我们的一些项目涉及到足够大的系统， 研究与冷冻EM，我们现在需要的计算资源的结构， cryo-EM数据几乎所有现代低温EM包都依赖GPU来处理数据。

Abstract We are requesting funds to purchase a GPU workstation that will be used for (1) deep learning in the context of protein design and (2) for solving protein structures using cryo-EM data. Currently there is a revolution going on protein modeling as deep learning methods have shown impressive success in both protein structure prediction and protein design. We are leveraging these new approaches in several ways that require access to GPU processers. First, we have developed a new protein design pipeline that iterates between protein structure prediction with AlphaFold and sequence optimization with a graph neural network to evolve sequences for a specific function. Initial experimental validation of the pipeline is very encouraging, and we are now eager to test it on a variety of protein design problems including the design of competitive inhibitors, protein switches and biosensors. The pipeline requires large blocks of GPU time (several days with multiple processors) to identify promising designs for experimental validation. Second, we are training new neural networks for improved performance in protein design. Access to GPUs is required for rapidly testing alternative network architectures and hyperparameters. The last step of many protein design projects is validation of the design model with a high-resolution structure. Some of our projects involve systems large enough to be studied with cryo-EM and we are now in need of computational resources to structures from cryo-EM data. Almost all modern cryo EM packages rely on GPUs for processing data.