Adversarial domain adaptation for cross data source macromolecule in situ structural classification in cellular...

Area of Science:

• Cryo-electron tomography
• Structural biology
• Machine learning

Background:

• Supervised deep learning for macromolecule classification in cellular electron cryo-tomography (CECT) is scalable but requires extensive labeled training data.
• Creating CECT training data from the same source is laborious, necessitating data from separate sources.
• Cross-dataset predictions in CECT are hindered by domain shift, caused by differing image intensity distributions.

Purpose of the Study:

• To address the domain shift problem in CECT data analysis.
• To improve the accuracy of macromolecule classification when using training data from a different source than the prediction data.
• To enhance the recovery of novel macromolecular structures from CECT data.

Main Methods:

• Adaptation of a deep learning-based adversarial domain adaptation (3D-ADA) method.
• Utilizing a source domain feature extractor for training data.
• Adversarially training a target domain feature extractor to minimize feature distribution differences between datasets.

Main Results:

• 3D-ADA effectively addressed the domain shift problem in CECT data.
• The method demonstrated stable improvements in cross-dataset prediction accuracy.
• 3D-ADA outperformed two existing domain adaptation techniques on experimental and simulated datasets.
• The approach enhanced the cross-dataset recovery of novel macromolecular structures.

Conclusions:

• 3D-ADA is a robust method for overcoming domain shift in CECT.
• The technique enables reliable macromolecule classification using data from disparate sources.
• This approach facilitates more efficient and accurate structural analysis in CECT.