Low-Rank Covariance for Cryo-EM Clustering

Cryo-electron microscopy (cryo-EM) provides 2D projections of 3D molecules by measuring electron absorption. Due to very high noise, a large number of projections are needed for accurate reconstruction of the 3D volume. In biological samples, however, a molecule may exist in a number of different states, any one of which may be portrayed in a given projection. By estimating the low-rank covariance matrix of the 3D volumes, we can properly describe molecular variability and cluster the projections. Once clustered, the projection images can be used to reconstruct each of the molecular states. We evaluate the performance of this method for both synthetic and experimental datasets, demonstrating its effectiveness for clustering heterogeneous cryo-EM samples.