FSCC: Few-Shot Learning for Macromolecule Classification Based on Contrastive Learning and Distribution Calibration...

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron tomography (Cryo-ET) enables 3D visualization of macromolecules in near-native states.
• Accurate classification of millions of macromolecules in tomograms is crucial for structural recovery.
• Existing supervised deep learning methods struggle to classify novel macromolecular classes due to extensive labeling requirements.

Purpose of the Study:

• To develop a novel few-shot learning method (FSCC) for accurate classification of unseen macromolecular classes in Cryo-ET.
• To reduce the significant time and labor associated with data labeling for new macromolecular structures.

Main Methods:

• Proposed a two-stage training strategy for the FSCC method.
• Stage 1: Contrastive learning pre-trains the model on a large dataset of labeled macromolecules.
• Stage 2: Distribution calibration re-trains the classifier to adapt to novel classes using limited labeled data.

Main Results:

• FSCC achieves competitive performance on synthetic datasets compared to state-of-the-art supervised methods, requiring only 5 labeled examples per novel class versus 1100-1500.
• On real datasets, FSCC improves classification accuracy by 5%-16% over baseline models.
• Demonstrates strong generalization ability for classifying novel macromolecules with minimal labeled data.

Conclusions:

• The proposed FSCC method effectively classifies novel macromolecular structures using few-shot learning.
• Contrastive learning and distribution calibration significantly enhance model generalization for unseen classes.
• FSCC offers a practical solution to the data labeling bottleneck in Cryo-ET macromolecular classification.