FakET: Simulating cryo-electron tomograms with neural style transfer.

Area of Science:

• Structural Biology
• Microscopy Techniques
• Computational Biology

Background:

• Accurate particle localization and classification are crucial in cryo-electron microscopy (cryo-EM).
• Deep learning models excel but require large training datasets.
• Current physics-based models for synthetic data generation are time-consuming, limiting their use.

Purpose of the Study:

• To introduce FakET, a novel method for simulating cryo-transmission electron microscope (cryo-TEM) data.
• To enable rapid adaptation of synthetic datasets to reference data for improved cryo-EM analysis.

Main Methods:

• Developed FakET using neural style transfer to simulate the cryo-TEM forward operator.
• Applied FakET to adapt synthetic datasets for training deep learning models.
• Compared performance of models trained on FakET-generated data versus benchmark data.

Main Results:

• FakET generates high-quality simulated micrographs and tilt-series.
• Models trained with FakET data achieved performance comparable to those trained on benchmark data.
• FakET demonstrated a 750× increase in data generation speed and a 33× reduction in memory usage.

Conclusions:

• FakET significantly accelerates the generation of training data for cryo-EM deep learning.
• The method offers a computationally efficient and scalable solution for cryo-EM data simulation.
• FakET's performance and resource efficiency make it a valuable tool for advancing cryo-EM analysis.