Localization of macromolecules in crowded cellular cryo-electron tomograms from extremely sparse labels

Area of Science:

• Structural biology
• Cell biology
• Biophysics

Background:

• Accurate localization of macromolecules in cellular cryo-electron tomography (cryo-ET) is essential for determining their in situ structures.
• Conventional template matching methods are limited by template bias and low throughput.
• The scarcity of annotated data presents a significant hurdle for developing effective learning-based localization methods.

Purpose of the Study:

• To develop an annotation-efficient method for macromolecule localization in tomograms.
• To address the challenges posed by limited annotated data in cryo-ET image analysis.
• To improve the throughput and accuracy of macromolecule structure determination.

Main Methods:

• TomoPicker treats macromolecule localization as a voxel classification problem.
• The method employs two distinct positive-unlabeled learning strategies.
• Evaluation was performed on cryo-ET datasets from eukaryotic and prokaryotic cells.

Main Results:

• TomoPicker achieved performance comparable to state-of-the-art supervised methods using only 10 annotations.
• The method demonstrated up to 98% reduction in data requirements compared to supervised learning.
• Significant improvements were observed over existing methods under sparse annotation conditions.

Conclusions:

• TomoPicker offers a highly efficient solution for macromolecule localization in cryo-ET.
• The annotation-efficient approach democratizes in situ structural biology by reducing data burden.
• This method facilitates more accessible and rapid structural determination of cellular components.