Unsupervised Deep Learning for Cryo-EM/ET Particle Classification

Cryogenic electron microscopy (cryo-EM) is widely used to determine protein structures from purified samples. With cryogenic electron tomography (cryo-ET), it is possible to image thin sections of cells and directly visualize diverse proteins in their native environment. However, separating the massive mixture of different types of proteins inside cells remains challenging, especially without human intervention. In this work, we present an unsupervised deep learning framework for classifying cryo-EM/ET particle images that jointly optimizes for pose and structure through an autoencoder-based multi-stage pipeline. On single-particle cryo-EM datasets, our framework can capture the continuous variation in particle views and separate images of real particles from contamination. Using in silico mixtures of multiple types of proteins, we demonstrated our method's capability of separating the protein classes without supervision. Finally in a real cryo-ET dataset of algal chloroplast, our deep learning model successfully detected the four main protein complexes of photosynthesis on native thylakoid membranes. Taken together, these results demonstrate that our framework offers a scalable, annotation-free solution for robust structural discovery across diverse cryo-EM and cryo-ET applications.