Abstract
High-throughput data collection in single-particle cryo-electron microscopy (EM) necessitates fast, accurate, and generalizable methods to assess micrograph quality. Manual micrograph curation scales poorly to large datasets and often misclassifies images due to sample-specific variability. Fully supervised deep-learning methods show promise in scalability and feature learning. However, dependence on annotated data limits generalizability. We present prismPYP, a self-supervised, data-driven framework that uses domain-specific image augmentations to perform label-free feature learning on micrographs and power spectra. From the learned, low-dimensional image representations, we perform feature-based image clustering that reveals distinct and consistent indicators of image quality. For validation, we used the resulting high-quality images to determine high-resolution structures that matched the quality of maps determined using manual curation, but using fewer particles. prismPYP generalizes across experimental conditions, imaging hardware, and both conventional single-particle and time-resolved cryo-EM. It is both interpretable and computationally efficient, and enables rapid, scalable quality assessment for cryo-EM micrographs.