Autofocused 3D Classification of Cryoelectron Subtomograms

Abstract

Classification of subtomograms obtained by cryoelectron tomography (cryo-ET) is a powerful approach to study the conformational landscapes of macromolecular complexes insitu. Major challenges in subtomogram classification are the low signal-to-noise ratio (SNR) of cryo-tomograms, their incomplete angular sampling, the unknown number of classes and the typically unbalanced abundances of structurally distinct complexes. Here, we propose a clustering algorithm named AC3D that is based on a similarity measure, which automatically focuses on the areas of major structural discrepancy between respectivesubtomogram class averages. Furthermore, we incorporate a spherical-harmonics-based fast subtomogram alignment algorithm, which provides a significant speedup. Assessment of our approach on simulated data sets indicates substantially increased classification accuracy of the presented method compared to two state-of-the-art approaches. Application to experimental subtomograms depicting endoplasmic-reticulum-associated ribosomal particles shows that AC3D is well suited to deconvolute the compositional heterogeneity of macromolecular complexes insitu.