Cryogenic Single-Particle Electron Microscopy (SPEM) is a powerful method for investigating the structure and dynamics of biological macromolecules. Despite recent advances in both instrumentation and software, generating an accurate ab initio 3D reconstruction from large sets of SPEM projection images remains difficult. This problem is challenging due to the high level of noise in the images and the many parameters that need to be determined in an unsupervised and efficient manner. Small particles with low or no symmetry are the most difficult targets.
We present a new stochastic approach for dimensionality and noise reduction of large SPEM data sets. Once particles images have been identified and extracted, our new method is applied to reduce the dimensionality of the data ~100-fold and create representatives with improved SNR (class averages). Test on numerous experimental cryo-EM data sets reveal that our new method has dramatically improved noise robustness, permitting generation of ab initio 3D reconstructions with resolutions well below 10 Å from the analysis of class averages.