Self-supervised learning for generalizable particle picking in cryo-EM micrographs.

We present cryoelectron microscopy masked autoencoder (cryo-EMMAE), a self-supervised method designed to overcome the need for manually annotated cryo-EM data. cryo-EMMAE leverages the representation space of a masked autoencoder to pick particle pixels through clustering of the MAE latent representation. Evaluation across different EMPIAR datasets demonstrates that cryo-EMMAE outperforms state-of-the-art supervised methods in terms of generalization capabilities. Importantly, our method showcases consistent performance, independent of the dataset used for training. Additionally, cryo-EMMAE is data efficient, as we experimentally observe that it converges with as few as five micrographs. Further, 3D reconstruction results indicate that our method has superior performance in reconstructing the volumes in both single-particle datasets and multi-particle micrographs derived from cell extracts. Our results underscore the potential of self-supervised learning in advancing cryo-EM image analysis, offering an alternative for more efficient and cost-effective structural biology research. Code is available at https://github.com/azamanos/Cryo-EMMAE.