GPU Optimization of Biological Macromolecule Multi-tilt Electron Tomography Reconstruction Algorithm

Three-dimensional (3D) reconstruction in cryo-electron tomography (cryo-ET) plays an important role in studying in situ biological macromolecular structures at the nanometer level. Owing to limited tilt angle, 3D reconstruction of cryo-ET always suffers from a “missing wedge” problem which causes severe accuracy degradation. Multi-tilt reconstruction is an effective method to reduce artifacts and suppress the effect of the missing wedge. As the number of tilt series increases, large size data causes high computation and huge memory overhead. Limited by the memory, multi-tilt reconstruction cannot be performed in parallel on GPUs, especially when the image size reaches 1 K, 2 K, or even larger. To optimize large-scale multi-tilt reconstruction of cryo-ET, we propose a new GPU-based large-scale multi-tilt tomographic reconstruction algorithm (GM-SIRT). Furthermore, we design a two-level data partition strategy in GM-SIRT to greatly reduce the memory required in the whole reconstructing process. Experimental results show that the performance of the GM-SIRT algorithm has been significantly improved compared with DM-SIRT, the distributed multi-tilt reconstruction algorithm on the CPU cluster. The acceleration ratio is over 300%, and the memory requirement only decreases to one-third of DM-SIRT when the image size reaches 2 K.