The core challenge that MemBrain v2 addresses begins with segmentation. A cryo-electron tomogram is essentially a three-dimensional reconstruction assembled from a tilt series of two-dimensional projection images of a vitrified sample. Within that volumetric data, membranes appear as weak, noisy, curving sheets only a few nanometers thick, often distorted by the so-called missing wedge artifact that arises from physical limits on how far the sample can be tilted in the microscope. Early approaches to membrane segmentation relied on hand-crafted filters, edge detectors, and painstaking manual contouring, a process that could consume days of an expert’s time for a single tomogram and that varied substantially from one analyst to another. Such variability made large-scale, quantitative comparisons across cells, conditions, or organisms extremely difficult, effectively capping the throughput of the field.

MemBrain v2 builds on the deep-learning foundations laid by its predecessor, which introduced convolutional neural network models trained to recognize membrane density patterns that human eyes and classical algorithms routinely miss. The new version extends this philosophy across the entire analysis pipeline. Rather than providing a single segmentation model that researchers must then stitch into their own homemade workflows, the software packages the full sequence of steps: preprocessing of raw tomograms, automated segmentation of membranes, extraction of continuous membrane meshes from the segmented volumes, localization of membrane proteins relative to those meshes, and quantitative spatial statistics describing how proteins and curvature features are organized across the membrane surface. The result is a pipeline in which a user can move from a reconstructed tomogram to publication-ready membrane measurements with minimal manual intervention.

Technically, the segmentation engine in MemBrain v2 employs neural networks trained on curated, manually annotated datasets of tomograms spanning diverse biological contexts, from isolated organelles to intact cells prepared by focused ion beam milling. The models operate on three-dimensional patches of the tomogram, predicting per-voxel probabilities that a given location belongs to a membrane. Crucially, the developers have placed strong emphasis on generalization: the goal is that a network trained on one set of samples and imaging conditions will still perform reliably on tomograms acquired on a different microscope, with a different detector, or from a different biological system. This kind of cross-dataset robustness has been a persistent pain point in cryo-ET machine learning, where models frequently overfit to the specific noise characteristics and contrast regimes of their training data. Performance across heterogeneous datasets was therefore treated as a first-class evaluation criterion rather than an afterthought.

Once membranes are segmented, the software converts the voxel-level predictions into geometrically clean surface meshes, a step that is far less trivial than it sounds. Segmentation outputs are typically noisy at the pixel scale, containing holes, spurious protrusions, and disconnected fragments that would corrupt any downstream measurement of surface area, curvature, or protein spacing. MemBrain v2 incorporates methods to extract smooth, watertight representations of membrane surfaces, preserving fine topological features such as tight junctions, cristae junctions, organelle contact sites, and highly curved tubular structures. This geometric fidelity matters because many of the most interesting biological questions in modern cell biology hinge precisely on such features: how the folded inner membrane of a mitochondrion organizes its respiratory machinery, how the endoplasmic reticulum generates curved hubs that recruit certain proteins, or how viral replication compartments remodel host membranes into convoluted replication factories.

Perhaps the most consequential addition in version 2 is the integration of membrane protein localization into the same workflow. Subtomogram averaging has long allowed structural biologists to determine high-resolution structures of abundant, repetitive complexes, but mapping the broader landscape of membrane-associated proteins in situ, including those that are sparse, irregularly arranged, or conformationally heterogeneous, remains a formidable task. MemBrain v2 provides tools to detect and classify particles in the vicinity of segmented membranes, project their positions onto the extracted meshes, and compute spatial statistics that characterize their distribution: clustering tendencies, exclusion patterns, correlations with local membrane curvature, and nearest-neighbor relationships. By unifying segmentation and particle analysis in one software environment, the tool removes the error-prone handoffs between disconnected programs that have historically plagued multi-step cryo-ET analyses.

The significance of this integration extends well beyond convenience. Spatial organization of proteins on membranes is now understood to be a central regulatory principle in cell biology. The clustering of receptors can amplify signaling; the sorting of complexes into curvature-sensing domains can drive vesicle budding; the exclusion of certain proteins from contact sites can maintain organelle identity. Quantifying these patterns rigorously requires exactly the kind of joint membrane-and-protein analysis that MemBrain v2 enables. Instead of qualitative statements that a protein appears enriched on curved regions, researchers can now report statistically grounded measurements of enrichment, complete with the geometric context of the underlying membrane. This shift from descriptive to quantitative in situ analysis aligns cryo-electron tomography with the standards of rigor that fields such as light-sheet microscopy and single-cell genomics achieved years ago.

Accessibility has been a guiding concern throughout the design. Cryo-ET is a technique that sits at the intersection of biology, physics, and computer science, and many research groups possess deep biological questions but limited computational infrastructure. Earlier generations of analysis tools often assumed considerable programming expertise, requiring users to assemble pipelines from research code written by other laboratories, with little documentation and no guarantee of stability between versions. MemBrain v2 counters this with user-friendly interfaces, containerized installation options, thorough documentation, and standardized output formats designed to interoperate with the broader ecosystem of tomography software, including visualization platforms and subtomogram averaging packages. The developers have also made the trained models and processing tools openly available, consistent with a broader movement in structural biology toward open, reproducible computational methods that any laboratory can adopt, inspect, and extend.

The timing of such a tool is not accidental. Cryo-electron tomography is undergoing explosive growth, propelled by advances in direct electron detectors, automated data collection, focused ion beam milling of lamellae, and reconstruction algorithms that dramatically improve resolution and contrast. Large-scale initiatives now aim to image thousands of cellular volumes, generating data at a rate that makes manual analysis categorically impossible. In this environment, automated, reliable, generalizable analysis pipelines are not merely convenient; they are the precondition for the field’s ambition to build a quantitative atlas of cellular architecture. Tools like MemBrain v2 represent the computational counterpart to the hardware revolution, ensuring that the flood of tomographic data can be converted into biological insight rather than accumulating as unprocessed archives.

For the practicing researcher, the practical implications are immediate. A laboratory studying mitochondrial remodeling during stress, for example, could apply the pipeline across dozens of tomograms of control and treated cells, obtain consistent membrane segmentations, quantify changes in cristae morphology, map the redistribution of respiratory chain complexes, and test whether protein reorganization tracks with local membrane curvature, all within a unified framework. A virology group could chart how viral proteins reshape intracellular membranes as replication compartments mature. A neurobiologist could examine the nanoscale organization of synaptic vesicle clusters and presynaptic membrane specializations. In each case, the same underlying machinery of segmentation, mesh extraction, particle localization, and spatial statistics does the heavy lifting, freeing scientists to focus on experimental design and biological interpretation rather than on debugging bespoke image-analysis scripts.

As with any computational method, caveats remain. Deep-learning segmentations are probabilistic and can fail in unexpected regimes, particularly with extreme imaging conditions, unusual staining or labeling approaches, or membrane-like features such as dense protein coats that mimic lipid bilayers. Responsible use therefore still entails expert validation, and the developers emphasize that the tool is intended to accelerate and standardize expert workflows rather than eliminate judgment from the loop. Nevertheless, by demonstrating robust performance across diverse datasets and by wrapping the entire analysis chain into a coherent, accessible package, MemBrain v2 marks a meaningful step toward making the quantitative analysis of cellular membranes as routine as the imaging itself. If the trajectory of cryo-electron tomography continues on its current course, tools of this kind will become the standard infrastructure on which the next decade of in situ structural biology is built, transforming noisy three-dimensional images into a rigorous, comparative anatomy of the cell’s molecular geography.

Subject of Research: A deep-learning software pipeline for membrane segmentation and membrane protein spatial analysis in cryo-electron tomography

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

Article References: Lamm, L., Zufferey, S., Zhang, H., Righetto, R. D., Waltz, F., Wietrzynski, W., Yamauchi, K. A., Burt, A., Liu, Y., Martinez-Sanchez, A., Ziegler, S., Isensee, F., Schnabel, J. A., Engel, B. D., & Peng, T. (2026). MemBrain v2: an end-to-end tool for the analysis of membranes in cryo-electron tomography. Nature Methods. https://doi.org/10.1038/s41592-026-03178-8

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03178-8

Keywords: cryo-electron tomography, MemBrain v2, membrane segmentation, deep learning, membrane proteins, spatial analysis, structural biology, cell biology, organelle architecture, image analysis, subtomogram averaging, Nature Methods