henhaobio
Sign in

Nature MethodsBioinformaticsPeer-reviewed journal article

MemBrain v2: an end-to-end tool for the analysis of membranes in cryo-electron tomography

Lorenz LammiD, Simon ZuffereyiD, Hanyi ZhangiD, Ricardo Diogo RighettoiD, Florent WaltziD, Wojciech WietrzyñskiiD, Kevin A. YamauchiiD, Alister BurtiD, Ye LiuiD, Antonio Martínez-SáncheziD, Sebastian ZiegleriD, Fabian IsenseeiD, Julia A. SchnabeliD, Benjamin D. EngeliD, Tingying PengiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Cryo-electron tomography (cryo-ET) provides unique insights into macromolecular complexes in their native environments, yet membrane analysis remains a major bottleneck due to low signal-to-noise ratios, missing wedge artifacts, and the complexity of membrane-associated proteins. Existing tools often require extensive manual annotation, struggle with generalization across datasets, and lack integrated solutions for segmentation, protein localization, and quantitative analysis. We introduce MemBrain v2, a deep learning-enabled framework that unifies these tasks into a streamlined pipeline. MemBrain-seg leverages a diverse, collaboratively generated training dataset and specialized model training strategies to achieve generalizable membrane segmentation across variable tomographic conditions. MemBrain-pick enables data-efficient localization of membrane-bound proteins by integrating geometric constraints with deep learning, reducing the need for extensive manual annotation. MemBrain-stats provides quantitative insights into protein distributions, computing spatial metrics to analyze intra-membrane particle organization. MemBrain v2 integrates seamlessly into cryo-ET workflows, providing an accessible and structured approach to membrane analysis. The full package is available at https://github.com/CellArchLab/MemBrain-v2.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (22 Apr 2025) · CC BY 4.0

MemBrain v2 couples a generalist U-Net cryo-ET membrane segmenter, a mesh-based picker reaching 91% F1 for photosystem II from one annotated membrane, and geodesic particle statistics.

Membrane analysis in cryo-electron tomography is limited by low signal-to-noise, missing-wedge distortion, segmenters trained on single membrane types, and pickers needing dense volumetric labels, while existing statistics tools treat membrane geometry and protein distributions separately. MemBrain v2 unifies segmentation, sparse-label membrane-protein picking, and surface-based spatial statistics in one open workflow.

Key results

  1. ShownIteratively expanding a community-annotated training set made MemBrain-seg generalize to unseen tomogram sources, and it outscored TARDIS on both real and fully synthetic membrane patches, although the authors note that the real-data comparison is biased toward MemBrain-seg.

    Five-fold cross-validation across five cumulative annotation rounds: Dice on the DeePiCt test set rose from 39% to 66% (reaching 55% in rounds lacking DeePiCt training patches). Fully synthetic patches with absolute GT: MemBrain-seg 66% (Round 3) vs TARDIS 59% Dice; MemBrain-seg also had higher Dice than TARDIS on the real test sets. · MemBrain-seg: a generalized approach for…; MemBrain-seg Ablation Studies; Incremental Training; TARDIS Comparison

  2. ShownFourier-domain augmentations improve transfer to an unseen dataset, and brief fine-tuning that monitors the skeleton-based Surface-Dice metric adapts the model to a shifted domain better than Dice-only monitoring.

    Round 4 models (no DeePiCt training data), DeePiCt Surface-Dice: 55% plain, 59% with Fourier amplitude, 57% with missing wedge augmentation. Fine-tuning with DeePiCt patches only: Surface-Dice 58% to 67%; monitoring Surface-Dice rather than Dice alone gave 67% vs 60%. · MemBrain-seg: a generalized approach for…; Model Fine-tuning; Data Augmentations; Fine-tuning with Surface-Dice

  3. ShownWorking directly on membrane meshes, MemBrain-pick localizes photosystem II accurately from very sparse membrane-level labels. It outperforms volumetric, flattened-membrane and template-matching pickers, and a model trained on spinach transfers to Chlamydomonas.

    PSII F1 (match radius 63.36 Å; test set ten membranes, two tomograms): 91% from one training membrane; 93% with 29 membranes vs MemBrain v1 90%, MPicker+EPicker 42%, PyTOM template matching 52%, DeepFinder 60% and crYOLO 26% (25 training membranes each). Spinach-trained model on Chlamyold: F1 84%. · MemBrain-pick: an interactive tool to…; Evaluation Metric; Spinach data; Chlamy Generalization

  4. ShownStatistics from automated picks come close to earlier manual estimates of thylakoid particle density and phycobilisome periodicity, while all-particle neighbor distances run slightly shorter than the published PSII-only values.

    Concentration (particles/µm): Spinach 1688 vs 1714 published; Chlamyold 1492 and Chlamynew 1316 vs 1292. All-particle NN distances: Spinach 18.8 vs 21.2 nm; Chlamyold 20.7 and Chlamynew 21.2 vs 24.4 nm (published values PSII-only). Ripley's O on EMD-31244: about 35 nm spacing vs 34.5 nm manual. · MemBrain-stats: quantitative analysis of…; Test application: phycobilisome organization on…; Phycobilisomes

  5. ShownAn end-to-end run on the Chlamydomonas nuclear envelope produced ribosome picks that subtomogram averaging confirmed. The picks showed a bias toward similar neighbor orientations, consistent with poly-ribosome chains.

    EMPIAR-11830: 97 of 140 segmented tomograms retained; a model trained on 13 annotated membranes predicted 4515 positions on the remaining 79. STOPGAP average of all picks reached 22.7 Å after 14 iterations, showing TRAP density; 3-NN minimum in-plane angles skewed lower than a uniform random model. · Test application: poly-ribosome chain…; MemBrain-pick prediction; Subtomogram Averaging; Ribosome Chains

Methods, data and reproducibility

MemBrain-seg: 3D U-Net (nnU-Net design, MONAI) trained over five active-learning rounds with Fourier amplitude and missing-wedge augmentations plus a skeleton-based Surface-Dice loss. MemBrain-pick: DiffusionNet on membrane meshes with normal-projected densities regressing particle distance maps, decoded by mean shift; MemBrain-stats computes geodesic metrics.

  • MemBrain-seg training patches from spinach chloroplast tomograms (Round 1), first labeled with TomoSegMemTV and then manually refined · Spinacia oleracea · n = 69 patches · EMPIAR-12612
  • Chlamydomonas tomogram patches covering thylakoids, mitochondria and Golgi (Round 2) · Chlamydomonas reinhardtii · n = 33 patches · EMPIAR-11830
  • Collaborator-contributed re-annotated patches (Round 3) and synthetic patches from PolNet and CryoTomoSim (Round 4) · n = 27 community patches; 40 synthetic patches (20 per simulator)
  • Externally annotated DeePiCt segmentation patches (Round 5), also the held-out generalization test set · n = 15 patches · EMPIAR-10988
  • Spinach thylakoid membranes with refined PSII, b6f and unknown-density labels for MemBrain-pick training and testing · Spinacia oleracea · n = 45 membranes, nine tomograms (ten test membranes) · EMPIAR-12612; GT annotations in Zenodo record 15090084
  • Chlamyold (published GT) and Chlamynew (no GT; Cryo-CARE and IsoNet processed) thylakoid tomograms for cross-species transfer · Chlamydomonas reinhardtii · n = 28 membranes (four tomograms); 23 membranes (four tomograms)
  • Red algal chloroplast tomogram with phycobilisome-PSII supercomplexes for the end-to-end phycobilisome analysis · red algae (species not named in text) · n = 29 membrane segmentations from one tomogram · EMD-31244
  • Chlamydomonas nuclear envelope and ER tomograms (Cryo-CARE denoised) for ribosome picking and subtomogram averaging · Chlamydomonas reinhardtii · n = 140 tomograms segmented, 97 retained; 4515 ribosome picks · EMPIAR-11830
Methods & tools
MemBrain-seg (3D U-Net following nnU-Net design, implemented in MONAI), MemBrain-pick (DiffusionNet on membrane meshes, score-guided mean shift plus DBSCAN), MemBrain-stats (concentrations, geodesic nearest neighbors, Ripley's K/L/O), Napari, Surforama and the napari-lasso-3d plugin, TARDIS (membrane segmentation baseline), DeepFinder and crYOLO (volumetric picking baselines), MPicker with EPicker; MemBrain v1 (membrane-specialized picking baselines), PyTOM template matching (pytom-match-pick), PolNet and CryoTomoSim (synthetic tomogram simulators), STOPGAP and IMOD (subtomogram averaging, CTF-corrected reconstruction), Cryo-CARE and IsoNet (denoising and missing-wedge restoration), TomoSegMemTV and MITK Workbench (initial annotation and correction)
Software & environment
pip install membrain (PyPI); segmentations used the MemBrain_seg_v10_alpha model; TARDIS model downloaded 6th February 2024; MemBrain-seg implemented with MONAI, mean shift clustering in PyTorch on GPU.
Assay conditions & compute
MemBrain-seg trained on RTX 4090 GPUs for 1000 epochs (SGD, batch size 2, 160-voxel patches); fine-tuning capped at 100 epochs; inference rescaling and mean shift clustering run on GPU.
Reproducibility gaps
The nuclear-envelope ribosome average is listed as EMD-XXXXX (not yet deposited). No wall-clock runtimes or dependency version pins are reported.
Limitations (6)
  • The segmentation ground truth was built by correcting MemBrain-seg predictions, so Dice comparisons with TARDIS on real data likely favor MemBrain-seg. The less biased fully synthetic benchmark is also less realistic. · stated by the authors
  • Predicted segmentation thickness need not match true bilayer thickness. It can reflect annotator choices such as brush size in the training labels. · stated by the authors
  • Out-of-the-box segmentation may degrade under strong domain shift, such as different microscopy conditions; fine-tuning on typically 10 to 20 annotated patches is offered to improve quality in such cases. · stated by the authors
  • Heterogeneous membrane-protein targets make a single universal MemBrain-pick model hard to achieve. Faint or occluded proteins may be missed in low-contrast, noisy tomograms, and strong membrane density can mask small integral proteins. · stated by the authors
  • The picking benchmarks rest on one spinach PSII test split (ten membranes, two tomograms). Volumetric baselines were trained on 25 membranes, and cross-dataset F1 was measured only on Chlamyold because Chlamynew lacks GT. · noted in this summary
  • The phycobilisome annotation counts conflict (six annotated and 23 predicted in Results vs five and 24 in Methods). The ribosome orientation shift versus a uniform model is reported without a statistical test. · noted in this summary

Read this if you segment membranes or pick membrane-embedded proteins in cryo-ET tomograms and want pretrained, fine-tunable models with sparse-annotation picking and geodesic spatial statistics.

Citations
153
FWCI
116.55
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLamm '22Selvaraj '25Siggel '24Francisco '22Seghiri '26Schönnenbec '25
84% semantic matchBioinformatics
Lorenz Lamm, Ricardo Diogo Righetto, Wojciech Wietrzyñski+4 more·Computer Methods and Programs in Biomedicine·Jul 1, 2022· 122 citations