TomoScore: A Neural Network Approach for Quality Assessment of Cellular cryo-ET

Abstract

Electron cryo-tomography (cryoET) is a powerful imaging tool that allows three-dimensional visualization of subcellular and molecular architecture without chemical fixation. Tomogram quality varies widely, particularly during large high-throughput data collections, and the most common strategy for initial quality assessment is empirical judgment by an expert. Tomograms are collected for two distinct purposes: 1) Annotation of subcellular features and cellular morphology, typically performed at lower magnifications and higher defocus. 2) Subtomogram averaging, generally at higher magnifications and closer to focus. For the first purpose, contrast and the ability to distinguish cellular features of interest are key, whereas for subtomogram averaging, recoverable signal at high resolution is the key factor. We have developed "TomoScore" a deep-learning based tomogram quality metric to assess the suitability of individual tomograms for cellular annotation based on the scale of features which can be distinguished. We further explore the relationship between accumulated electron dose and resulting quality, suggesting an optimum dose range for cryoET data collection. Overall, our study streamlines data processing and reduces the need for human involvement during pre-selection for tomogram segmentation.