Accelerated Cryo-EM Structure Determination with Parallelisation using GPUs in Relion-2

Abstract

Single-particle electron cryo-microscopy (cryo-EM) structure determination is transforming structural biology by enabling near-atomic resolution for wide classes of structures. However, the most powerful Bayesian approaches for image reconstruction are very costly from a compuational point-of-view. This has turned the computational step into a new bottleneck that is limiting both throughput and new method development, in particular as an increasing number of facilities have access to direct-electron detectors. Here, we present a new implementation of the Bayesian regular likelihood optimization used in the RELION program that has been reformulated to use parallelization on graphics processors (GPUs) to address the most computationally intensive steps in the cryo-EM structure determination workflow. Both image classification and high-resolution refinement have been accelerated up to 50-fold, and template-based particule selection running on a single workstation now provides performance equivalent to 1000 CPU cores at a tiny fraction of the cost. Memory requirements of the reconstruction algorithms have been reduced to make it possible to use widely available low-cost consumer hardware, and we show that the use of single precision arithmetic does not adversely affect results. This routinely enables high-resolution cryo-EM structure determination in a matter of days on a single workstation even for the largest datasets, and in many cases the processing is completed in mere hours. This new processing approach removes the need for advanced computational infrastructure, and it will make it possible to use many more classes in reconstruction - which in turn will improve detection of alternative conformations and handle much larger datasets than previously possible.