Data-driven regularisation lowers the size barrier of cryo-EM structure determination

Dari Kimanius, Kiarash Jamali, Max E. Wilkinson, Sofia Lövestam, Vaithish Velazhahan, Takanori Nakane, Sjors H. W. ScheresbioRxiv 2023 · DOI

TL;DR

This paper introduces "Blush regularisation," a novel deep learning-based approach that leverages denoising convolutional neural networks to improve the resolution and applicability of cryo-electron microscopy (cryo-EM) for determining the structures of smaller or more challenging biological macromolecules. By training a denoiser on existing cryo-EM data, Blush regularisation effectively acts as a data-driven prior, overcoming limitations of traditional methods and enabling the reconstruction of previously intractable samples.

What Problem Does This Paper Address?

Cryo-EM is a powerful technique for visualizing the 3D structures of biological molecules. However, a significant challenge is the inherent noise in the low-dose images required to protect the sample from electron beam damage. This noise makes it difficult to accurately align the images of individual particles, especially for smaller molecules where the signal is weaker. This alignment problem leads to an "ill-posed optimization problem," limiting the minimum size of macromolecules that can be reliably studied. Traditional methods often rely on a "smoothness prior" in Fourier space, which is a simplification that doesn't fully capture the complex nature of biological structures. The paper aims to address this by developing a more sophisticated, data-driven way to regularize the reconstruction process, thereby improving alignment and enabling the study of smaller or more complex biological assemblies.

Methods

The core of this work is the development of "Blush regularisation," which uses a denoising convolutional neural network (CNN). This CNN is trained using a "noise2noise" approach, meaning it learns to denoise by being trained on pairs of noisy images (in this case, cryo-EM half-maps) rather than requiring perfectly clean "ground truth" images.

Key aspects of their method include:

  • Denoising CNN Architecture: A U-net architecture with approximately 13 million parameters is employed.
  • Data Augmentation: To make the denoiser robust, various augmentation techniques are used, including anisotropic Gaussian filtering, affine transformations (rotation, translation), cropping, and voxel-value standardization. This helps the network learn to handle variations in particle orientation and image quality.
  • Iterative Application: The trained denoiser is integrated into the standard expectation-maximization (EM) algorithm used in cryo-EM reconstruction. At each iteration, the denoiser is applied to the intermediate reconstructions.
  • Spectral Trailing: To prevent overfitting and the hallucination of non-existent features, a technique called "spectral trailing" is introduced. This involves using the Fourier Shell Correlation (FSC) between independently refined half-maps to determine a resolution cutoff. Beyond this cutoff, information from the denoiser is not used, and the map is low-pass filtered. This ensures that the highest resolution features are still validated by the data itself.
  • RELION Integration: The Blush regularisation method has been implemented within the open-source cryo-EM software RELION-5, making it accessible to the broader community.

Datasets

The authors utilized several datasets for training and evaluation:

  • Training Data: Pairs of noisy half-maps were downloaded from the Electron Microscopy Data Bank (EMDB). Only entries with resolutions higher than 4 Å and deposited unfiltered half-maps were selected, totaling 422 pairs.
  • Evaluation Datasets:
    • Plasmodium falciparum Chloroquine Resistance Transporter (PfCRT) dataset (EMPIAR-10330).
  • Ste2 membrane protein dataset.
  • Saccharomyces cerevisiae pre-catalytic spliceosomal B complex dataset (EMPIAR-10180).
  • First intermediate amyloid (FIA) dataset.
  • A 40 kDa protein-nucleic acid complex dataset.

Key Insights and Ideas

  • Data-Driven Priors are Powerful: The success of Blush regularisation demonstrates that deep learning models can effectively learn complex, data-driven priors about biological structures, going beyond simple smoothness assumptions.
  • Overcoming Size Limitations: The most significant impact is the ability to reconstruct smaller or more challenging macromolecules (like the 40 kDa protein-nucleic acid complex) that were previously intractable with standard methods.
  • Balancing Prior Knowledge and Data Fidelity: The spectral trailing mechanism is crucial. It allows the denoiser to improve reconstructions by leveraging learned priors while ensuring that the final results are still strongly supported by the experimental data, thus mitigating overfitting.
  • Generalizability: The method shows broad applicability across diverse sample types, including membrane proteins, protein complexes, amyloid filaments, and protein-nucleic acid complexes, suggesting its utility for a wide range of cryo-EM studies.

Key Results

  • Improved Resolution: Application of Blush regularisation consistently led to improved resolutions across various datasets. For example, the PfCRT dataset improved from 3.8 Å to 3.4 Å, and the Ste2 dataset improved from 3.8 Å to 3.4 Å.
  • Enhanced Model Completeness: Automated model building (using ModelAngelo) showed significant improvements in completeness, indicating better quality and more interpretable density maps. For the Ste2 dataset, completeness increased from 19% to 43%.
  • Successful Reconstruction of Challenging Samples: The 40 kDa protein-nucleic acid complex, which was previously unresolvable, was successfully reconstructed to 2.5 Å resolution using Blush regularisation.
  • Mitigation of Overfitting: Even without spectral trailing, the denoiser showed no signs of overfitting on the PfCRT dataset, suggesting its inherent ability to regularize.

Limitations and Future Work

  • Heuristic Nature of Spectral Trailing: The authors acknowledge that the choice of the resolution cutoff 'p' in spectral trailing is heuristic and that the dampening effect of the network is not well-defined in Fourier space. Future work could explore alternative approaches that don't rely on such heuristics.
  • Potential for Overfitting in Other Datasets: While no overfitting was observed in their tested cases, they caution that it might be more prominent for other datasets, hence recommending the use of spectral trailing.
  • Future Directions: The authors suggest that future adaptations of algorithms like VDAM could incorporate Blush regularisation for initial model generation. They also anticipate that as machine learning advances, the community will continue to explore the optimal balance between incorporating prior knowledge and relying on experimental data for validation.

Summary updated

Abstract

Abstract Macromolecular structure determination by electron cryo-microscopy (cryo-EM) is limited by the alignment of noisy images of individual particles. Because smaller particles have weaker signals, alignment errors impose size limitations on its applicability. Here, we explore how image alignment is improved by the application of deep-learning to exploit prior knowledge about biological macromolecular structures that would otherwise be difficult to express mathematically. We train a denoising convolutional neural network on pairs of half-set reconstructions from the electron microscopy data bank (EMDB) and use this denoiser as an alternative to a commonly used smoothness prior. We demonstrate that this approach, which we call Blush regularisation, yields better reconstructions than existing algorithms, in particular for data with low signal-to-noise ratios. The reconstruction of a protein-nucleic acid complex with a molecular weight of 40 kDa, which was previously intractable, illustrates that regularisation through denoising will expand the applicability of cryo-EM structure determination for a wide range of biological macromolecules.
BibTeX
@misc{kimanius2023data,
  author = {Dari Kimanius and Kiarash Jamali and Max E. Wilkinson and Sofia Lövestam and Vaithish Velazhahan and Takanori Nakane and Sjors H. W. Scheres},
  title = {{Data-driven regularisation lowers the size barrier of cryo-EM structure determination}},
  year = {2023},
  doi = {10.1101/2023.10.23.563586},
  howpublished = {bioRxiv},
}
Figure