EPicker

EPicker implements exemplar-based continual learning to improve particle picking in single-particle cryo-electron microscopy (cryo-EM) by accumulating knowledge from limited samples while avoiding catastrophic forgetting.

Key Features:

• Exemplar-based continual learning: Retains exemplar samples to accumulate knowledge and prevent catastrophic forgetting during incremental updates.
• Mitigates limitations of supervised deep learning: Reduces dependence on supervised deep learning approaches that require extensive labeled training datasets with predefined features.
• Generalization to unseen datasets: Maintains identification performance across datasets with different features.
• Integration from limited samples: Efficiently incorporates new knowledge from limited samples of novel datasets without losing prior information.
• Robust performance across evolving datasets: Preserves particle identification accuracy during continuous updates in automated processing pipelines.
• Continuous expansion of identification capabilities: Enables ongoing adaptation to recognize additional biological object types.
• Supported biological object types: Targets protein particles, vesicles, and fibers (bio-macromolecules).

Scientific Applications:

• Single-particle cryo-EM particle picking: Improves automated selection of particle images for downstream cryo-EM reconstruction workflows.
• Automated cryo-EM processing pipelines: Provides continual learning capacity for routine pipeline updates and dataset variation.
• Detection of diverse bio-macromolecules: Facilitates identification of proteins, vesicles, and fibers across heterogeneous datasets.

Methodology:

Implements exemplar-based continual learning by retaining exemplar samples to accumulate knowledge and avoid catastrophic forgetting, contrasted with conventional supervised deep learning approaches that require extensive labeled datasets with predefined features.