A non-parametric approach to particle picking in all frames

Single-particle cryo-electron microscopy (cryo-EM) has significantly advanced macromolecular structure reconstruction. However, a key limitation is the conventional reliance on micrographs obtained by motion correction and averaging, which inherently loses the richness of information contained within each frame of the original movie. The future of cryo-EM reconstruction ideally involves harnessing the raw signal from every frame to unlock potentially higher quality structures. In this paper, we present a first essential step toward this paradigm shift, that is, a novel, non-parametric method for detecting tomographic projections across all movie frames, using temporal consistency. Our method is inspired by Structure-from-Motion (SfM), and independent of motion correction, CTF estimation, and initial reconstruction. Our experimental results demonstrate reduced outlier rate and accurate particle localization comparable to existing approaches throughout the entire movie sequence.