TomoTwin

TomoTwin applies deep metric learning to locate and identify macromolecular particles within cryogenic electron tomography (cryo-ET) volumes for downstream subtomogram averaging and structural analysis.

Key Features:

• Deep metric learning: Uses deep metric learning to create embeddings that capture structural similarity among tomographic subvolumes.
• High-dimensional structural embeddings: Embeds tomograms into a high-dimensional space enriched with structural information to separate macromolecular classes.
• Automated particle picking: Automates particle localization and picking within tomogram volumes without requiring manually curated training data.
• De novo identification: Enables de novo identification of proteins and other macromolecules in tomograms without network retraining when locating new protein types.
• Robustness to low SNR and crowding: Distinguishes macromolecules in densely packed volumes with low signal-to-noise ratios and spatial crowding.
• Output for subtomogram averaging: Produces particle coordinates suitable for downstream subtomogram averaging and structural analysis.

Scientific Applications:

• Subtomogram averaging preparation: Provides particle coordinates for subtomogram averaging workflows to improve macromolecular structure determination from cryo-ET.
• In situ macromolecular localization: Locates proteins and macromolecular assemblies within native cellular contexts for structural and spatial studies.
• Cellular architecture analysis: Maps distributions of macromolecules in tomograms to support investigations of cellular architecture.

Methodology:

Applies deep metric learning to embed tomographic subvolumes into a high-dimensional structural embedding space and performs automated particle localization based on those embeddings without requiring manually curated training data or network retraining.