Training Generalized Segmentation Networks with Real and Synthetic Cryo-ET data.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Deep learning for cryo-electron tomograms (cryo-ET) is powerful but requires extensive ground truth data.
• Generating realistic synthetic cryo-ET data is crucial for training robust segmentation models.

Purpose of the Study:

• To develop an open-source software package, CryoTomoSim (CTS), for simulating cryo-ET data.
• To investigate the impact of microscope parameters on deep learning segmentation using simulated data.
• To train a generalized cellular segmentation network for cryo-ET.

Main Methods:

• CryoTomoSim was used to generate coarse-grained models of macromolecular complexes in vitreous ice and simulate tilt series.
• Deep learning segmentation networks were trained using synthetic datasets with varying molecular crowding and diversity.
• A co-training approach was employed, segmenting over 100 neuronal growth cone tomograms to develop the NeuralSeg network.

Main Results:

• Simulated data revealed the influence of dose, defocus, and pixel size on deep learning segmentation.
• Molecular crowding and diversity in synthetic datasets are essential for training effective cellular segmentation networks.
• The NeuralSeg network demonstrated the ability to segment cellular features across diverse life domains.

Conclusions:

• CryoTomoSim provides a valuable tool for generating synthetic cryo-ET data to train deep learning models.
• While synthetic data is effective for initial model training, real cellular data is necessary for optimal accuracy and generalizability.
• NeuralSeg represents a significant step towards automated cellular segmentation in cryo-ET across various biological systems.