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

Tristan Bepler, K. Kelley, Alex J. Noble, Bonnie BergerNature Communications 2020 · DOI

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.

The Challenge of Noise in CryoEM

In structural biology, cryo-electron microscopy (cryoEM) is a powerful tool for visualizing proteins at atomic resolution. However, it faces a fundamental trade-off: proteins are delicate and easily destroyed by high-energy electron beams (radiation damage). To preserve samples, researchers must use extremely low electron doses, resulting in images with a typical signal-to-noise ratio (SNR) as low as 0.1.

These noisy images make it difficult to identify small proteins or find specific orientations, such as "top-down" views, leading to gaps in the final 3D structure. While standard low-pass filters can help, they often blur out essential structural details. Furthermore, the time required to collect enough high-quality data remains a major bottleneck in the field.

Methodology: Noise2Noise and U-Net

The authors developed Topaz-Denoise, utilizing a U-Net architecture—a convolutional neural network common in medical imaging. The technical breakthrough lies in the training process. Typically, denoising requires "clean" ground-truth images, which do not exist in cryoEM. To bypass this, the authors employed the Noise2Noise framework.

The Topaz-Denoise training pipeline using the Noise2Noise framework on split movie frames.

Modern direct detector devices collect data as "movies" (sequences of frames). By splitting these frames into two independent sets (even and odd) and summing them into two separate noisy micrographs of the same object, the network can be trained to map one noisy version to the other. Because the noise is independent between the sets, the network learns to predict the underlying common signal to minimize loss.

Key Results and Efficiency Gains

Topaz-Denoise provides a pre-trained "general" model trained on thousands of micrographs across diverse imaging conditions. This allows researchers to denoise new datasets "out of the box" without additional training.

Comparison of raw micrographs against low-pass filtering and Topaz-Denoise (affine and U-net models).

Key performance metrics include:

  • SNR Improvement: The U-Net model improved SNR by >20 dB (100-fold) over raw micrographs and significantly outperformed conventional low-pass filtering.
  • Biological Discovery: The denoiser identified 2.15x more particles of clustered protocadherin than previous methods, resolving a 3D structure at 12 Å (improving on the previous 35 Å) and identifying a new "partially open" conformation.
  • Microscope Throughput: By enabling the use of lower electron doses (10–25% of usual levels), the model increased collection efficiency by 65% (178 exposures/hour vs. 108).
  • 3D Denoising: The framework successfully extends to 3D cryoET tomograms, markedly improving the visibility of ribosomes and membrane bilayers.

Limitations and Best Practices

The authors include a critical warning regarding the "Hallucination" Problem. Deep neural networks can occasionally generate features or imprints from their training data that are not present in the sample. Consequently, denoised images should be used for visualization and particle picking only. The actual 3D reconstruction and high-resolution refinement must still be performed using the raw, noisy data to avoid over-fitting and false confidence in the final structure.

Summary updated

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.
BibTeX
@inproceedings{bepler2020topaz,
  author = {Tristan Bepler and K. Kelley and Alex J. Noble and Bonnie Berger},
  title = {{Topaz-Denoise: general deep denoising models for cryoEM and cryoET}},
  booktitle = {Nature Communications},
  year = {2020},
  doi = {10.1038/s41467-020-18952-1},
}
Figure