MemBrain v2

Cryo-electron tomography images cells frozen in a near-native state, and membranes are everywhere in the resulting volumes — but they are also the field's chief obstacle. Low signal-to-noise ratios, missing-wedge artifacts, and the sheer geometric complexity of a crowded cytoplasm mean that tracing membranes has traditionally consumed days of manual annotation per tomogram, and that the proteins embedded in those membranes are hard to detect against the surrounding density.

MemBrain v2, from the CellArch Lab at Helmholtz Munich with the Biozentrum at the University of Basel and collaborators across TUM, ETH Zürich, DKFZ, and the MRC Laboratory of Molecular Biology, bundles three tools into one pipeline. MemBrain-seg segments membranes from raw tomograms with a pretrained U-Net. MemBrain-pick localizes membrane-associated particles by learning directly on the resulting membrane meshes. MemBrain-stats turns the picked coordinates into spatial statistics — concentrations, geodesic nearest-neighbor distances, Ripley's functions. The preprint was first posted in January 2024 and substantially revised in April 2025; it has not been peer reviewed.

Within the cryo-ET catalog, MemBrain-seg occupies the "generalist segmenter" slot alongside SABER and CryoViT, while MemBrain-pick addresses the same particle-localization problem as ProPicker from the opposite direction: rather than a promptable general picker, it constrains the search to the membrane surface and learns a target-specific model from a handful of annotated membranes.

#Key Features

• Out-of-the-box membrane segmentation: A pretrained checkpoint segments new tomograms directly, and has been applied across the full CryoET Data Portal without dataset-specific retraining.
• Surface-Dice metric and loss: A skeleton-based extension of Centerline-Dice that rewards membrane continuity instead of voxel overlap, used both to evaluate models and, in differentiable form, to train them.
• Fourier-domain augmentation: Randomized rescaling of Fourier amplitude bands and synthetic missing wedges simulate acquisition variability, raising Surface-Dice on the held-out DeePiCt test set from 55% to 59%.
• Mesh-based particle picking: MemBrain-pick projects tomographic density onto a triangular membrane mesh and runs a DiffusionNet over that surface, shrinking the search space to the membrane neighborhood.
• Interactive annotation loop: A Napari plugin with 3D lasso selection isolates single membrane instances, which Surforama then displays for manual particle annotation — the workflow that generates MemBrain-pick's training data.

#Technical Details

MemBrain-seg is a U-Net whose architecture and training schedule follow nnU-Net conventions, trained with a composite of binary cross-entropy, Dice, and Surface-Dice losses plus an ignore label for ambiguous voxels. Its training set was assembled over five active-learning rounds: Spinacia oleracea and Chlamydomonas reinhardtii tomograms first, then collaborator contributions, synthetic data, and DeePiCt patches. Generalization improved with each round, with Dice on the DeePiCt test set rising from 39% to 66%. Against TARDIS on fully synthetic data with absolute ground truth, MemBrain-seg scored 66% Dice to TARDIS's 59%. Where domain shift is severe, fine-tuning on a few dataset-specific patches lifted Surface-Dice from 58% to 67%.

MemBrain-pick predicts, for each mesh vertex, the distance to the nearest particle center, then resolves positions by score-guided mean-shift clustering. On stacked spinach thylakoid membranes it reached 91% F1 for Photosystem II localization from a single annotated membrane and 93% from 29, against 90% for MemBrain v1, 60% for DeepFinder, 52% for template matching in PyTOM, 42% for MPicker with EPicker, and 26% for crYOLO. A spinach-trained model transferred to Chlamydomonas thylakoids at 84% F1. Applied end to end, MemBrain-stats recovered a phycobilisome spacing of roughly 35 nm in red algal chloroplasts, matching the 34.5 nm previously determined by manual annotation, and PSII concentrations within 2% of published manual counts for spinach.

#Applications

The pipeline serves structural and cell biologists analyzing membrane organization in situ — photosynthetic membranes, nuclear envelopes, mitochondrial cristae — where the question is how many copies of a complex sit on a membrane and how they are arranged. Because MemBrain-seg runs without retraining, it also functions as a general preprocessing step, supplying membrane masks that constrain template matching or particle picking for other tools.

#Impact

MemBrain-seg has become a widely used cryo-ET utility, with the preprint accumulating over a hundred citations while still unpublished, and its segmentations are browsable across the CryoET Data Portal. Three caveats bear on reuse. The pretrained segmentation weights are deposited on Zenodo under CC BY 4.0, and the training dataset alongside them, so the Google Drive link that appears in the documentation is a mirror rather than the only route to the checkpoint. The v2 monorepo carries no license file — the MemBrain-seg component repository is BSD-3-Clause, and the archived v2 codebase deposit is CC BY 4.0, but the umbrella repository's terms are unstated. And the pretrained-checkpoint convenience applies to segmentation alone: MemBrain-pick ships no general particle model, so each target complex requires annotating a small number of membranes in Surforama and training a model, which is what the paper's "data-efficient" framing describes.