Abstract
Three-dimensional reconstruction of the electron-scattering potential of biological macromolecules from electron cryo-microscopy (cryo-EM) projection images is an ill-posed problem. The most popular cryo-EM software solutions to date rely on a regularization approach that is based on the prior assumption that the scattering potential varies smoothly over three-dimensional space. Although this approach has been hugely successful in recent years, the amount of prior knowledge that it exploits compares unfavorably with the knowledge about biological structures that has been accumulated over decades of research in structural biology. Here, a regularization framework for cryo-EM structure determination is presented that exploits prior knowledge about biological structures through a convolutional neural network that is trained on known macromolecular structures. This neural network is inserted into the iterative cryo-EM structure-determination process through an approach that is inspired by regularization by denoising. It is shown that the new regularization approach yields better reconstructions than the current state of the art for simulated data, and options to extend this work for application to experimental cryo-EM data are discussed.
Introduction
Determining the 3D structure of a protein from cryo-electron microscopy (cryo-EM) is inherently an "ill-posed" problem. Researchers capture thousands of 2D projection images of molecules frozen in random orientations. Due to the radiation sensitivity of biological samples, these images are acquired with low electron doses, resulting in extreme signal-to-noise ratios.
To overcome this, standard software like RELION employs a mathematical "prior"—an assumption about the expected structure to guide the reconstruction. Currently, the industry standard is a "smoothness prior," which assumes the scattering potential varies smoothly in space. However, this ignores the vast structural knowledge accumulated in the Protein Data Bank (PDB), such as the existence of alpha-helices, beta-sheets, and solvent regions. This paper introduces a framework to bridge this gap by teaching the reconstruction algorithm what a protein actually looks like.
Methodology
The proposed framework, Confidence-Weighted Regularization by Denoising (CW-RED), integrates deep learning into the standard refinement workflow.
The Learned Prior
The authors trained a 3D Convolutional Neural Network (U-Net architecture) to act as a "learned prior." Unlike human-defined formulas, this network learns the statistical essence of biological macromolecules from 543 atomic structures in the PDB. These were converted into volumetric maps and augmented with simulated cryo-EM artifacts (noise, Contrast Transfer Function effects) to create a training set of over 25,000 maps.
Integration and Safeguards
The neural network is inserted into the iterative Expectation-Maximization (EM) algorithm. In each iteration, the network denoises the current 3D estimate, nudging it toward a biologically plausible structure. To prevent "hallucination"—where the network might invent features not present in the data—the authors introduced a confidence-weighting heuristic based on the Fourier Shell Correlation (FSC). At low resolutions, the system relies on the standard Gaussian prior; as resolution improves, the neural network's influence increases.
Results
In simulated tests using four distinct PDB structures, the RED-based approach consistently outperformed the standard RELION Gaussian prior.
Key findings include:
- Higher Accuracy: Reconstructed maps showed significantly higher correlation with ground truth structures.
- Secondary Structure Recovery: The learned prior was notably better at recovering recognizable secondary structures, such as alpha-helices, in high-noise scenarios where traditional methods produced blurred results.
- Improved Angular Assignment: Cleaner 3D maps at each iteration allowed for more accurate determination of the 2D-to-3D orientations of the original noisy particles.
Discussion and Future Work
While promising, the current model was trained and tested on simulated data. Real-world cryo-EM data introduces complexities like ice thickness variations and beam-induced motion that require further adaptation. Additionally, the computational cost of running a 3D U-Net iteratively is high, suggesting a need for more efficient "patch-based" processing in the future.
As the field moves toward learned priors, new validation tools will be essential to ensure that recovered features are supported by the data and not merely "imagined" by the AI based on its training. The authors are currently working on adapting this framework for experimental data.
Summary updated