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

Bepler T, Kelley K, Noble AJ, Berger B. · 2020

Read Topaz-Denoise: general deep denoising models for cryoEM and cryoET. by Bepler T, Kelley K, Noble AJ, Berger B. (2020) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc7567117.txt.

Introduction

Bepler, Kelley, Noble and Berger present Topaz-Denoise, a deep learning tool for cryo-EM images. Cryo-EM is now a favored route to protein structures. Its images suffer from very low signal compared with noise.

The models were trained on thousands of micrographs. Conditions varied widely across the training set. The general model denoises new data without retraining. Cleaner images are easier to read and process. The tool recovered views of an elongated particle. Low-dose imaging also became more usable with denoising. That can cut data collection time for labs. A 3D model for cryo-ET tomograms is included. The software ships as part of the Topaz package.

Catalog record

CREATORBepler T, Kelley K, Noble AJ, Berger B.
YEAR2020
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC7567117

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Topaz-Denoise: general deep denoising models for cryoEM and cryoET By Tristan Bepler, Kotaro Kelley, Alex J Noble, Bonnie Berger Published in: Nature Communications DOI: 10.1038/s41467-020-18952-1 ABSTRACT Abstract Cryo-electron microscopy (cryoEM) is becoming the preferred method for resolving protein structures. Low signal-to-noise ratio (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. Here, we present Topaz-Denoise, a deep learning method for reliably and rapidly increasing the SNR of cryoEM images and cryoET tomograms. 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. The general model we present is able to denoise new datasets without additional training. Denoising with this model improves micrograph interpretability and allows us to solve 3D single particle structures of clustered protocadherin, an elongated particle with previously elusive views. We then show that low dose collection, enabled by Topaz-Denoise, improves downstream analysis in addition to reducing data collection time. We also present a general 3D denoising model for cryoET. Topaz-Denoise and pre-trained general models are now included in Topaz. We expect that Topaz-Denoise will be of broad utility to the cryoEM community for improving micrograph and tomogram interpretability and accelerating analysis. Subject terms: Machine learning, Cryoelectron microscopy INTRODUCTION Visualization of micrographs from cryo-electron microscopy (cryoEM) of biological specimens is primarily limited by the phase contrast of proteins, the low electron dose conditions required due to radiation damage accrued by the proteins, and the thickness of the ice. As researchers push towards smaller and smaller proteins, these issues hinder downstream analyses because these proteins become increasingly difficult to distinguish from noise. Certain orientations of large, non-globular proteins can also have low signal, leading to missing views. The typical signal-to-noise ratio (SNR) of a cryoEM micrograph is estimated to be only as high as 0.11, amongst the lowest in any imaging field, and no ground truth exists. Nonetheless, several steps during collection and processing of micrographs in single particle cryoEM rely on properly human-inspecting micrographs, identifying particles, and examining processed data. Conventional cryoEM methods for improving contrast in micrographs include downsampling, bandpass filtering, and Wiener filtering2,3. However, these methods do not address the specific noise properties of micrographs and often do not provide interpretable results, which increasingly hinders attempts to resolve small and non-globular proteins4,5. At the same time, there is a push in the field to fund large research facilities for high-throughput cryoEM. These and smaller facilities are moving towards the synchrotron model of data collection and need to increase their throughput to meet rising demand. One approach to speed up collection would be to collect shorter micrograph exposures. However, reducing total dose would exacerbate SNR-related analysis problems. Better micrograph denoising provides the opportunity to reduce total dose and increase collection throughput without compromising interpretability or downstream results. Image denoising has long been a topic of significant interest in the computer vision and signal processing community6, but has recently seen a surge in interest from the machine learning community. Advances in deep neural networks have enabled substantial improvements in image restoration and inpainting (i.e. filling in missing pixels) by learning complex, non-linear priors over the applied image domain. However, these methods require ground truth images to provide supervision for learning the denoising model7,8, and are hence limited to domains where ground truth is available. To overcome this barrier, Lehtinen et al.9 presented a general machine learning (ML) framework, called Noise2Noise, for learning denoising models from paired noisy images rather than paired noisy and ground truth images. This method has been followed by several others for learning denoising models without ground truth10–12. These methods offer new approaches for training deep neural network models for denoising in challenging domains. In cryoEM, neural network denoising software has only just started to emerge for dataset-by-dataset cryo-electron tomogram (cryoET) denoising13,14 and single particle micrograph denoising15. However, there have not been any systematic evaluations of these methods to date nor have pre-trained general denoising models been developed. Here, we develop Topaz-Denoise, large-scale, publicly available denoising models for cryoEM and cryoET. Conventional cryoEM and cryoET denoising methods are ad-hoc filters that do not model the complex image generative process. To address this, our goal is to learn the denoising process directly from data. However, deep denoising models typically require ground truth signal which is not available in cryoEM. We make the key insight that the individual movie frames collected by modern direct detector devices (DDD) are many independent observations of the same underlying signal and, hence, can be used to learn denoising models directly via the Noise2Noise framework. Trained on thousands of micrographs from DDD - K2, Falcon II, and Falcon III - across a variety of imaging conditions, these general models (also called pre-trained models) provide robust denoising without the need to train on a dataset-by-dataset basis. We test and compare these denoising models on several micrographs of typical particles and of small particles, study improvements in SNR, and use denoising combined with Topaz particle picking16 to obtain 3D single particle cryoEM structures of clustered protocadherin, an elongated particle with previously-elusive views and a putative new conformation. We also show that denoising enables more rapid data collection by allowing micrographs to be collected with a lower electron total dose (10–25% typical exposure times) without sacrificing interpretability or downstream processing. Shorter exposure times allow for higher throughput microscope usage, which reduces research cost and increases research efficiency. In addition, we develop a general 3D denoising model for cryoET tomograms, trained on dozens of cryoET tomograms, and show that our general denoising model performs comparably to models trained on a dataset-by-dataset basis. These models are integrated into Topaz allowing easy access to the community, along with the denoising framework that allows users to train their own cryoEM and cryoET denoising models. Topaz-Denoise source code is freely available as part of Topaz (http://topaz.csail.mit.edu) and can be installed through Anaconda, Pip, Docker, Singularity, and SBGrid17, and is now integrated into CryoSPARC18, Relion19, Appion20 and Scipion21. As with Topaz, Topaz-Denoise is designed to be modular and can easily be integrated into other cryoEM software suites. Topaz-Denoise includes several pre-trained models and the ability for the user to train their own models. Topaz-Denoise 2D training and inference runs efficiently on a single GPU computer, while 3D training and inference runs efficiently on multi-GPU systems. Both 2D and 3D denoising are integrated into the standalone Topaz GUI to assist with command generation. RESULTS Denoising with Topaz improves micrograph interpretability and SNR We develop a general cryoEM micrograph denoising model by training a neural network using the Noise2Noise framework on dozens of representative datasets of commonly used imaging conditions (Fig. 1, “Methods”). By learning the denoising model directly from data, we avoid making specific assumptions about the noise-generating process leading to superior denoising performance. [Figure: Illustration of the training framework and comparison of denoising methods on two example micrographs.a The Noise2Noise method requires paired

Read the complete text