Researchers have introduced two deep-learning tools that improve the precision and quality of cryo-electron microscopy imaging.
The developments address long-standing hurdles in structural biology, specifically the low signal-to-noise ratio that limits the accuracy of three-dimensional reconstructions. Cryo-EM allows near-atomic reconstruction of biological macromolecules without crystallization, yet the resulting images often require intensive manual screening or suffer from blurred structural details.
In a paper published in the Journal of Structural Biology: X, a team introduced TomoScore, a neural network that automates quality assessment of cellular cryo-electron tomography. The tool provides a metric to determine how suitable an individual tomogram is for cellular annotation based on the scale of distinguishable features.
Authors include Xuqian Tan, Xueting Zhou, Ethan Boniuk, Anisha Abraham, Zhili Yu, Valerie Dalton, Steven J Ludtke, and Zhao Wang. The team is affiliated with the Verna and Marrs McLean Department of Biochemistry and Molecular Pharmacology at Baylor College of Medicine in Houston, Texas. Ethan Boniuk is also associated with the Bioengineering Program at Rice University.
The TomoScore method aims to reduce the need for human experts to perform empirical judgment during large, high-throughput data collections. Electron cryo-tomography visualizes molecular and subcellular architecture in three dimensions while maintaining a native hydrated environment. Tomograms are typically collected for two purposes: annotation of cellular morphology and subcellular features at higher defocus and lower magnifications, and higher-resolution subtomogram averaging.
The system uses a deep-learning approach to predict a quality score between 0 and 1, helping scientists decide which purpose a tomogram serves best. As predicted, tomograms with a high TomoScore demonstrate better quality and possess more resolvable subcellular structures.
Existing denoising algorithms frequently encounter issues such as removing fine details, blurring structural boundaries, or providing limited benefits for 3D reconstruction.
CryoEDNet utilizes a multi-scale feature extraction module, an edge enhancement module, and a detail preservation module to protect these features. The network was trained using a multi-constraint loss that accounts for pixel-level reconstruction, perceptual similarity, structural similarity, and frequency-domain consistency. Testing on experimental datasets — specifically EMPIAR-10028, EMPIAR-10219, and EMPIAR-10330 — showed that the network improved particle representation, 2D class averaging, and reconstruction-related outcomes.
In synthetic tests using Gaussian white noise at an SNR of 0.1, the tool demonstrated improvements in image quality, 3D reconstruction results, and projection orientation estimation. Results support the transferability of CryoEDNet across these evaluated experimental datasets. The developers note that broader validation across more diverse imaging conditions and cryo-EM specimens remains necessary.
The TomoScore model, which uses a modified ResNet101 architecture, accurately predicts quality across various cell types. Effectiveness was observed across different species and cell types. In testing, the model captured various specimens including S. solfataricus, human blood platelet, Vibrio cholerae, and Magnetospirillum magneticum.
The research notes that additional examples from other species and measurements for particle features are available in the study’s supplementary materials. As these automated tools move into broader use, the ability to streamline data processing and enhance the signal in low-noise environments remains a central focus for the field.