Acceleration of image processing algorithms for single particle analysis by electron microscopy

Abstract

Cryogenic Electron Microscopy (Cryo-EM) is a vital field in current structural biology. Unlike X-ray crystallography and Nuclear Magnetic Resonance, it can be used to analyze membrane proteins and other samples with overlapping spectral peaks. However, one of the significant limitations of Cryo-EM is the computational complexity. Modern electron microscopes can produce terabytes of data per single session, from which hundreds of thousands of particles must be extracted and processed to obtain a near-atomic resolution of the original sample. Many existing software solutions use high-Performance Computing (HPC) techniques to bring these computations to the realm of practical usability. The common approach to acceleration is parallelization of the processing, but in praxis, we face many complications, such as problem decomposition, data distribution, load scheduling, balancing, and synchronization. Utilization of various accelerators further complicates the situation, as heterogeneous hardware brings additional caveats, for example, limited portability, under-utilization due to synchronization, and sub-optimal code performance due to missing specialization. This dissertation, structured as a compendium of articles, aims to improve the algorithms used in Cryo-EM, esp. the SPA (Single Particle Analysis). We focus on the single-node performance optimizations, using the techniques either available or developed in the HPC field, such as heterogeneous computing or autotuning, which potentially needs the formulation of novel algorithms. The secondary goal of the dissertation is to identify the limitations of state-of-the-art HPC techniques. Since the Cryo-EM pipeline consists of multiple distinct steps targetting different types of data, there is no single bottleneck to be solved. As such, the presented articles show a holistic approach to performance optimization. First, we give details on the GPU acceleration of the specific programs. The achieved speedup is due to the higher performance of the GPU, adjustments of the original algorithm to it, and application of the novel algorithms. More specifically, we provide implementation details of programs for movie alignment, 2D classification, and 3D reconstruction that have been sped up by order of magnitude compared to their original multi-CPU implementation or sufficiently the be used on-the-fly. In addition to these three programs, multiple other programs from an actively used, open-source software package XMIPP have been accelerated and improved. Second, we discuss our contribution to HPC in the form of autotuning. Autotuning is the ability of software to adapt to a changing environment, i.e., input or executing hardware. Towards that goal, we present cuFFTAdvisor, a tool that proposes and, through autotuning, finds the best configuration of the cuFFT library for given constraints of input size and plan settings. We also introduce a benchmark set of ten autotunable kernels for important computational problems implemented in OpenCL or CUDA, together with the introduction of complex dynamic autotuning to the KTT tool. Third, we propose an image processing framework Umpalumpa, which combines a task-based runtime system, data-centric architecture, and dynamic autotuning. The proposed framework allows for writing complex workflows which automatically use available HW resources and adjust to different HW and data but at the same time are easy to maintain