Topaz-Denoise: general deep denoising models for cryoEM and cryoET

Abstract

Cryo-electron microscopy (cryoEM) is becoming the preferred method for resolving protein structure. Low signal-to-noise (SNR) in cryoEM images reduces the confidence and throughput of structure determination during several steps of data processing, resulting in impediments such as missing particle orientations. Denoising cryoEM images can not only improve downstream analysis but also accelerate the time-consuming data collection process by allowing lower electron dose micrographs to be used for analysis without compromising structural interpretability. Here, we present Topaz-Denoise, a deep learning method for reliably increasing the SNR of cryoEM images in seconds. By training on a dataset composed of thousands of micrographs collected across a wide range of imaging conditions, we are able to learn models capturing the complexity of the cryoEM image formation process. While this general idea has been deployed successfully in natural imaging, protein threading, and proteomics, it has yet to be applied systematically in cryoEM where the lack of ground truth signal has been a long-standing limitation. To address this, we make the key insight that forming paired, independent micrographs from even and odd camera movie frames enables us to train denoising models without observing ground truth signal. We demonstrate that our denoising model improves SNR by roughly 100x over raw micrographs and 1.8x over other methods. Notably, we show that denoising with our general model enables solving the first 3D single particle structure of clustered protocadherin, an elongated particle with previously-elusive views. Topaz-Denoise and pre-trained general models are now included in Topaz ( https://github.com/tbepler/topaz ), a free and open-source software package that focuses on particle picking, and has been integrated into the Appion cryoEM suite. We expect that Topaz-Denoise will be of broad utility to the cryoEM community for improving micrograph interpretability and accelerating analysis.

Enhancing SNR and generating contrast for cryo-EM images with convolutional neural networks

“…Visualization of particles is significantly enhanced, facilitating more efficient manual image evaluation and particle picking. While we preprocess our images differently, use a novel CNN architecture, and train a single CNN model per dataset, our results are broadly consistent with other efforts of using similar denoising approaches (Bepler et al, 2019a;Bepler et al, 2019b;Tegunov and Cramer, 2019) and illustrate the robustness of the noise2noise algorithm.…”

“…While we were preparing this work, several other groups have introduced noise2noise-based denoising CNNs for enhancing contrast of cryo-EM images (Bepler et al, 2019a;Bepler et al, 2019b;Tegunov and Cramer, 2019). The application so far is, however, limited to the level of visualization and particle picking.…”

CryoGAN: A New Reconstruction Paradigm for Single-Particle Cryo-EMviaDeep Adversarial Learning

“…Deep learning has already had a profound impact in a wide range of image-reconstruction applications [40][41][42]; however, their current utilization in cryo-EM is mostly restricted to preprocessing steps such as micrograph denoising [43] or particle picking [44][45][46][47][48]. A recent work uses neural networks to model continuous generative factors of structural heterogeneity [49].…”

A guided approach for subtomogram averaging of challenging macromolecular assemblies

“…2 A). For ease in identification of complexes of interest in subsequent steps in this procedure, it is recommended that particles are initially extracted from tomograms reconstructed using iterative reconstruction methods such as Simultaneous Iterative Reconstruction technique (SIRT) ( Gilbert, 1972 ) or Simultaneous Algebraic Reconstruction Technique (SART) ( Andersen and Kak, 1984 ), or from other deconvolution or denoising approaches that boost contrast ( Bepler et al, 2019 , Buchholz, 2018 , Tegunov and Cramer, 2019 ). A list of coordinates denoting the approximate centroid of the identified complexes should be generated and used to extract subvolumes from each tomogram, so that the complex of interest is roughly positioned in the center of the extracted volume.…”