CryoFSL: an annotation-efficient, few-shot learning framework for robust protein particle picking in cryo-electron...

Area of Science:

• Structural Biology
• Computational Biology
• Microscopy

Background:

• Accurate protein particle identification in cryo-electron microscopy (cryo-EM) is vital for high-resolution structure determination.
• Current methods struggle with large annotation needs and poor generalization, especially in low signal-to-noise ratio (SNR) conditions.

Purpose of the Study:

• To develop a novel few-shot learning framework for robust and annotation-efficient protein particle picking in cryo-EM.
• To reduce the reliance on extensive annotated datasets and improve generalization to new protein targets.

Main Methods:

• Introduced CryoFSL, a few-shot learning framework utilizing Segment Anything Model 2 with lightweight adapters.
• Employed a hierarchical adapter design for dynamic feature modulation to handle low-SNR and heterogeneous cryo-EM data.
• Evaluated performance using minimal labeled micrographs (as few as five).

Main Results:

• CryoFSL demonstrated superior performance compared to traditional and state-of-the-art deep learning methods in the few-shot setting.
• Achieved high recall, precision, and improved 3D reconstruction resolution with minimal supervision.
• Showcased stability across heterogeneous micrographs and consistent detection of high-quality particles with fewer false-positives.

Conclusions:

• CryoFSL significantly reduces the annotation burden for cryo-EM particle picking, enabling scalable and generalizable pipelines.
• The framework redefines efficiency and quality in cryo-EM analysis by achieving competitive resolution with fewer particles.
• Paves the way for more accessible and efficient high-resolution structure determination using cryo-EM.