Consistency Criterion for Particle Sorting in Single-Particle Cryo-EM

Abstract

Current cryo-EM datasets contain tens of thousands to millions of high quality single-particle images, representing snapshots of all states present in solution. They also contain many “bad” particles - denatured protein or other contaminants. High-resolution reconstructions hinge on identifying and eliminating “bad” particles, separating “good” particles into distinct classes, and accurate orientation estimation within each class. Particle sorting attempts to differentiate “good” and “bad” particles, but typically does not address persistent misalignment of “good” particles. To further high-resolution structure determination of macromolecules inherently difficult for image alignment because of lack of obvious structural features, such as purely transmembrane proteins embedded in lipid nanodiscs, we describe an approach differentiating “misaligned” and “well-aligned” particles based on consistency of alignment across multiple de novo 3D refinements with different random initializations. Orientation alignment differences of the same particles are calculated using geodesic distance in SO(3), and consistency is defined as having similar final orientations across all refinements. We applied this approach to the published cryo-EM dataset of the TRPV1 ion channel, and to a synthetic dataset that contains only the transmembrane domain of TRPV1. The entire channel has well-defined soluble domains, facilitating image alignment. On the contrary, particles of the synthetic dataset lack well-defined molecule shape. Compared to the whole TRPV1, the proportion of particles with inconsistent orientation assignment increases significantly (∼24% vs ∼70%) for the synthetic dataset. We additionally found that a 3D reconstruction calculated from the subset of particles with consistent orientation assignment can reach a higher resolution using the same resolution criterion. Interestingly, a 3D reconstruction calculated from the excluded particles with inconsistent angular assignment has generally worse features, but still reached a relatively high resolution. This suggests such particles are misaligned “good” particles, rather than residual “bad” particles.