SBIR Phase I: Empowering Discovery of COVID-19 Vaccine through Large Scale Unsupervised Deeplearning for Electron Microscopy Data

这个小企业创新研究第一阶段项目的更广泛影响将是通过消除计算瓶颈来缩短药物发现周期和上市时间，例如新冠肺炎疫情期间所需的瓶颈，从而使研究人员能够进行生物结构发现。大多数现代药物发现项目都是从识别和验证蛋白质靶标开始的，以获得“已验证的药物靶标”。结构疫苗学是产生有效疫苗的合理方法。例如，了解病毒表面蛋白质的结构可以指导疫苗设计，了解病毒酶的结构可以直接导致抑制复制的药物治疗。低温电子显微镜(Cryo-EM)是一种对重要结构进行成像的方法。这项拟议的创新通过使用先进的深度学习技术，实现了图像的快速分类和三维结构的创建。这将使分析和发现速度提高50%以上。这个小型企业创新研究(SBIR)第一阶段项目以基准精度实时为Cryo-EM数据集提供了一条自主的图像处理管道。平均而言，每台低温电子显微镜每天产生高达2TB的数据。来自Cryo-EM的高分辨率图像大小通常在10k x 10k像素的数量级，需要专门的软件来处理、分类并将这些图像转换为3D结构。传统的完全有监督的深度学习方法由于生物分子的多样性和缺乏标记图像而失败。该方法利用大规模的深度无监督学习方法，可以训练人工智能系统，并通过使用高性能计算(HPC)簇来提供准确的结果，而不需要关于全分辨率图像大小的标签数据。该项目涉及开发云基础设施软件，以消除GPU内存限制。此外，建议的机器学习方法显著加速和改进了低温EM粒子拾取，无论是在单粒子重建工作流程的2-D图像中，还是在基准精度的电子断层扫描工作流程中完整细胞的3-D图像中。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will be to empower researchers to make bio-structure discoveries by eliminating computational bottlenecks to shorten drug discovery cycles and time-to-market, such as those needed during the COVID-19 pandemic. Most modern drug discovery projects start with protein target identification and verification to obtain a 'verified drug target'. Structural vaccinology is a rational approach to generate an effective vaccine. For instance, knowing the structures of the proteins on the surface of the virus can guide vaccine design, and knowing the structures of viral enzymes can lead directly to drug treatments that inhibit replication. Cryo-electron microscopy (cryo-EM) is a method to image important structures. The proposed innovation enables fast classification of images and creation of 3-dimensional structures by using advanced deep learning techniques. This will accelerate analysis and discovery by more than 50%. This Small Business Innovation Research (SBIR) Phase I project delivers an autonomous image processing pipeline for Cryo-EM data sets in real time at benchmark accuracies. On average, each cryo-EM microscope produces up to 2 TB of data daily. High-resolution image sizes from Cryo-EM are typically on the order of 10k x 10k pixels, requiring specialized software to process, classify and convert these images into 3D structures. Traditional fully supervised deep learning methods fail due to diversity of biomolecules and lack of labeled images. The proposed approach utilizes large-scale “deep unsupervised” learning methods that can train AI systems and deliver accurate results without labeled data on full resolution image sizes by use of High Performance Computing (HPC) clusters. The project involves development of cloud infrastructure software that eliminates GPU memory limitations. Further, the proposed machine learning approach dramatically accelerates and improves cryo-EM particle picking, both in 2-D images from single particle reconstruction workflows and in 3-D images of intact cells in electron tomography workflows at benchmark accuracies.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.