Ultra-fast haplotype phasing and genotype imputation service using a hybrid FPGA-GPU system

启动了大型国际社会项目，如美国100万名志愿者健康研究项目，对100多万人的基因组进行了排序。这一关键DNA参考数据的一个主要用途是允许阶段性和归因性，即估计个人单倍型和预测全球不同临床生物库的个人基因组范围数据的缺失基因类型。然而，利用超过32,000个欧洲个体的最新单倍型参考种群，甚至可以保持强大的专用集群系统，如Sanger Inputation Service(SIS)，运行数天。需要更快、更节能的计算架构来减少几个数量级的计算需求，以便真正允许使用不同的全球参考种群集合和更大的参考面板带来的所有新的实际好处。我们建议提供基于开发超快单倍型阶段化和基因拼接混合FPGA-GPU系统的阶段化和推算网络服务，以实现从不同参考种群集合的超快阶段化和推算。针对现场可编程门阵列(FPGA)和图形处理单元(GPU)的混合组合，为算法设计引入了一个新的有前途的科学研究领域。此外，我们还将在算法上改进SIS的HapHedge数据结构，以开发出另一种完全适合并行使用和FPGA结构的快速查找算法。对于基于最新的Eagle v2和PBWT工具的单倍型阶段化和推算，我们的悲观预期是，与16核CPU机器和Sanger推算服务相比，我们的预期已经显示在一个单一的FPGA处理器上分别加速了至少164和142，从而将中等大小基因组范围的数据集的运行时间从几天减少到几分钟。因此，对于阶段化和推算，我们计划为我们的服务提供的只有四个FPGA处理器的单个标准计算系统，与用于并行计算的650多个16核CPU核心的大型HPC计算集群一样强大，同时节省了99%以上的能源，这甚至没有考虑我们计划的GPU集成。此外，我们建议实施并为研究界提供一个可自由访问的图形决策Web界面。我们已经设计了一个“对话界面”原型，以便能够快速、轻松地从不同的参考面板集合中进行阶段划分和分配，并在运行时对阶段划分和分配精度进行“即时”评估。基于FPGA-GPU的混合分期和分配服务将为应用遗传学研究开辟全新的可能性，例如，在来自世界各地不同参考人群的分配小组之间切换，最终在个体水平上最大限度地提高分期和分配的准确性--这是个性化医学的重要前提。

Large-scale international community projects such as the United States 1-million-volunteer health study were launched to sequence the genomes of more than one million individuals. A major use of this key DNA reference data is to allow phasing and imputation, i.e. estimating personal haplotypes and predicting missing genotypes of individual genome-wide data of diverse clinical biobanks worldwide. However, utilizing the most recent haplotype reference population of more than 32,000 European individuals keep even powerful dedicated cluster systems, such as the Sanger Imputation Service (SIS), running for days. Faster and more energy-efficient computational architectures are required to reduce the computational requirements by several orders of magnitude in order to actually allow for all new practical benefits from using diverse sets of worldwide reference populations and larger reference panels.We propose to provide a phasing and imputation webservice based on the development of an ultra-fast haplotype phasing and genotype imputation hybrid FPGA-GPU system to enable ultra-fast phasing and imputation from diverse sets of reference populations. Targeting a hybrid composition of Field-Programmable Gate Arrays (FPGAs) and Graphical Processing Units (GPUs) introduces a new promising field of scientific research for algorithmic design. Besides, we will algorithmically improve the HapHedge data structure of the SIS to develop another fast lookup algorithm perfectly suited to parallel use and FPGA structure. For haplotype phasing and imputation based on the recent Eagle v2 and PBWT tools our pessimistic, i.e. lower bound, expectations already show a speedup of at least 164 and 142 on one single FPGA processor as compared to a 16-core CPU machine and the Sanger Imputation Service, respectively, thereby reducing the runtime for a medium-size genome-wide data set from days to minutes. Thus, for phasing and imputation, a single standard computing system with only four FPGA processors, as we plan to provide for our service, is as powerful as a large HPC computing cluster of more than 650 16-core CPU cores for parallel computing, while saving more than 99% of energy, and that does not even take our planned GPU integration into account.Further, we propose to implement and provide the research community with a freely accessible graphical decision-making web-interface. We already designed a "conversational interface" prototype to enable quick and easy phasing and imputation from diverse sets of reference panels in conjunction with an "on-the-fly" estimation of phasing and imputation accuracy during runtime. The hybrid FPGA-GPU-based phasing and imputation service will open up entirely new possibilities for applied genetic research, e.g. toggling between imputation panels from diverse sets of worldwide reference populations, to ultimately maximize phasing and imputation accuracy on the individual level - an important prerequisite for personalized medicine.