FakET: AI-Powered Breakthrough in Electron Microscopy

Cryogenic electron microscopy (cryo-EM) has revolutionized structural biology by providing high-resolution images of biological processes at the near-atomic level. However, challenges like limited viewing angles, radiation damage and shot noise make accurate particle identification difficult and labor-intensive. To address these limitations, a study lead by David Haselbach, technology platform head for cryo-EM at the Research Institute of Molecular Pathology (IMP) and Pavol Harar, machine learning research engineer at the University of Vienna, developed “FakET”. The new method, published in the journal Structure, creates “fake” electron microscopy images to train artificial intelligence (AI), reducing manual work in particle identification.

Overcoming data gaps in AI-driven cryo-ET image analysis

Cryo-EM has made major strides in recent years, enabling high-resolution imaging of biological structures. Cryo-electron tomography (cryo-ET) builds on the core principles of cryo-EM, combining them with tomographic imaging to produce three-dimensional (3D) reconstructions of cellular environments. Despite this progress, several limitations continue to hinder the streamlined application of cryo-ET. Equipment constraints restrict imaging to a 140-degree range, causing a “missing wedge” of data in 3D reconstructions. High-energy electron beams damage biological samples, requiring low doses that increase image noise. Identifying and analyzing molecules within noisy reconstructions is challenging, requiring multiple images to be averaged for clarity.

To support this process, researchers are turning to machine learning algorithms, a subfield of AI, to identify particles in cryo-ET images. To achieve the desired accuracy, this requires extensive training data with precise annotations, also known as the “ground truth” in the machine learning field. However, the structural biology field is currently lacking readily available and well-annotated data to develop and test such software.

AI simulations boost molecule identification

The development of “FakET” required a multidisciplinary effort, bringing together Haselbach’s expertise in cryo-EM and Harar’s skills in data science and machine learning. The researchers built a machine learning tool that creates realistic cryo-ET images without relying on slow, physics-based simulations. Instead, they used neural style transfer (NST) – a deep learning technique – to copy the unique noise patterns from real electron microscope images and apply them to synthetic data. This new data-driven method can create realistic, computer-generated images that mimic those produced by real electron microscopes, helping scientists train their analysis software in a much less labor-intensive workflow.

The key findings of the paper were:

• Machine learning and neural style transfer could be used to generate synthetic cryo-ET images.
• Realistic, synthetic training data could be produced with minimal manual effort.
• Cryo-ET workflows could be streamlined and efficiency improved in structural biology research by utilizing this method.

Streamlining cryo-ET image analysis with deep learning

The tool not only simplifies image analysis but also enhances cryo-ET imaging by making the process more efficient and accessible. By leveraging a pre-trained model, researchers can generate high-quality synthetic images from any cryo-electron microscope, as long as they have a reference image.

This capability eliminates the need for extensive manual data annotation, which is typically a time-consuming and labor-intensive process, and accelerates the development of AI-driven analysis tools. It also has the potential to drive new deep-learning applications that further streamline cryo-ET image analysis.

While FakET offers a powerful way to generate realistic synthetic cryo-ET images, it does, however, have some limitations. It relies on high-quality reference images, meaning variations in microscope settings could affect its accuracy. The method also focuses on mimicking noise patterns but may not fully capture all physical distortions seen in real tomograms. Additionally, while FakET-generated images help train AI models, their effectiveness in real-world biological studies still needs more validation. Although it’s faster than traditional simulations, it still requires significant computational power, which may limit accessibility for some labs. Despite these challenges, FakET is a major step forward in improving AI-driven cryo-ET analysis and reducing the need for manual data annotation.

Toward smarter and faster cryo-ET image analysis

“This project has been a multidisciplinary collaboration,” says Haselbach. “I brought the challenge from structural biology, while Pavol brought his expertise in machine learning. Together, we developed a tool that not only simplifies image analysis but has the potential to advance cryo-ET imaging capabilities by streamlining it further.” With a pre-trained model ready to use, researchers can quickly and efficiently produce synthetic images from any cryo-electron microscope they have a reference image from. “Our approach is innovative in replacing the time-consuming and laborious task of manual data annotation,” says Haselbach. “We think it will pave the way for new applications based on deep learning that can further streamline cryo-ET image analysis.” The next steps will aim to develop an algorithm that can handle the entire analysis process in one go – from raw microscope images to final structural insights – without requiring multiple complex steps. While this is still a long-term objective, FakET provides the groundwork for achieving more automated and efficient cryo-ET analysis in the future.

Reference: Harar P, Herrmann L, Grohs P, Haselbach D. FakET: Simulating cryo-electron tomograms with neural style transfer. Structure. 2025. doi:10.1016/j.str.2025.01.020

Have you recently published research?

You can raise awareness of your work with an article like this too.

Find out more