Evaluating and Improving Spectral Methods in Cryo-EM 2D Class Averaging

Abstract

2D class averaging is a crucial step in overcoming the issue of incredibly low signal-to-noise ratios (SNR) in Cryo-EM data sets. This method requires grouping of images according to their similarity, specifically the similarity of the orientation of the proteins which generate the images with respect to the electron beam and detector - their viewing directions. This poses a problem because the low SNR itself, which is precisely what class averaging aims to resolve, makes this step difficult. In this paper, we investigate two methods which attempt to improve groupings by viewing direction beyond the information which a raw similarity measure encodes, that is, methods which take pre-computed similarities as impute and give new, better measures of similarity as output. The two methods of interest to us are spectral embeddings and diffusion maps. We will first test the performance of these methods on simulated data sets, and then suggest manners in which their performance may be improved.