Joint Model for Image Denoising and Detection of Proteins Imaged by Cryo-EM
Particle picking is a critical step in the single-particle cryo-electron microscopy (EM) structure determination pipeline. In order to successfully obtain the 3D reconstruction of a macro-molecular complex, hundreds of thousands of particles need to be accurately identified and extracted from micrographs so they can be used for 3D refinement. While the particle detection process has historically been tedious and time-consuming, recent deep-learning based algorithms have achieved promising results on single-particle cryo-EM datasets provided that a sufficiently large number annotations are available. However, in datasets from low molecular weight targets or very low defocus where images have poor signal-to-noise ratios (SNR), these algorithms yield less satisfying results. To overcome the low SNR and need for large image annotation sets, we propose a joint semi-supervised learning framework that performs image denoising while simultaneously detecting the particles of interest. We do this by leveraging prior data distribution knowledge and consistency regularization. Our framework denoises images without the need of clean images and is able to detect particles of interest under very low SNR conditions, even when less than 5% of the data is labeled. We validate our approach on a single-particle cryo-EM dataset of ribosomes under decreasing SNR conditions and show that our strategy outperforms existing state-of-the-art methods by a significant margin, both in the tasks of micrograph denoising and particle detection.
