Tensor Core Based Algorithms For Cryo Electron Microscopy

Cryo-electron microscopy is a powerful imagining technique that can generate highly accurate structural models for macromolecules larger than 200 kDa. However, for macromolecules having smaller masses, the amount of information used to generate structural models is frequently insufficient to result in successful structure determination. The barrier defined by macromolecular mass arises from cryo-EM reconstruction starting with only low-resolution information, with the higher resolution information added gradually during subsequent steps of reconstruction. For macromolecules having low molecular mass, current methods do not use enough information at the beginning of the process to achieve successful convergence for an accurate structural model at the end. Unfortunately, the use of high-resolution information from the beginning of the computations is prohibitively expensive in terms of computational complexity. We propose to use Tensor Cores-based computations enabled by recent progress in GPU technology to make such computations possible. In this Phase I SBIR proposal, the algorithms that start the process of cryo-EM reconstruction will be re-designed so that the computations can be performed on Tensor Cores (Phase I), tested and validated (Phase I), implemented into a user-friendly software module (Phases I & II), which will become part of a larger software package that will be commercially distributed (Phase II & III). The redesign of computations will permit the use of full resolution at the start of the 2D classification and would result in dramatic expansion of the cryo-EM field. Such an expansion would enable studies of macromolecules and their complexes that are currently inaccessible to structural studies due to their size in the case of cryo-EM and due to their inability to form a crystal lattice in the case of X-ray crystallography. The potential for successful commercialization is high.