MemBrain

MemBrain detects membrane protein complexes in cryo-electron tomograms (cryo-ET) and estimates their positions and orientations to enable quantitative analysis of membrane-associated molecular organization.

Key Features:

• Deep learning evaluation: Uses a convolutional neural network (CNN) to evaluate sampled subvolumes along segmented membranes.
• Subvolume sampling: Samples subvolumes along segmented membranes for targeted analysis of membrane-bound proteins.
• Rotational subvolume normalization: Applies rotational subvolume normalization to simplify the detection task for the CNN.
• Tiny receptive field: Employs a tiny receptive field in the CNN architecture to reduce detection complexity and improve training.
• Scoring and clustering: Assigns a score to each subvolume and determines protein center positions using an efficient clustering algorithm.
• Protein orientation estimation: Determines protein orientations in addition to center positions from cryo-ET data.
• Annotation efficiency: Achieves high performance with minimal training labels, reporting an F1 score of 0.88 when trained on a single annotated membrane and 0.92 overall.
• Comparative performance: Outperforms classical computer vision and other CNN-based approaches (maximum reported F1 of 0.63 for existing methods).
• Generalizability: Pre-trained models can be applied across different cryo-ET acquisition methods and diverse cell types.

Scientific Applications:

• Membrane protein detection: Localization of membrane-bound protein complexes in cryo-ET tomograms.
• Spatial and interaction analysis: Analysis of protein spatial arrangements and interactions within cellular membranes using position and orientation estimates.
• Cross-method and cross-cell-type studies: Application of pretrained models to tomograms acquired with different cryo-ET methods or representing diverse cell types.
• Method benchmarking: Comparative evaluation and benchmarking of detection performance against classical computer-vision and other CNN-based approaches.

Methodology:

Subvolumes are sampled along segmented membranes, normalized by rotation, evaluated by a convolutional neural network (CNN) with a tiny receptive field that assigns scores to subvolumes, and protein center positions are extracted via an efficient clustering algorithm; protein orientations are also estimated.