Four new AI tools automate mapping of cell membranes and proteins

Several research teams have separately built computational tools this year to automate a task that has long required slow, manual work: identifying and outlining cell membranes, proteins, and other structures inside three-dimensional images. At the University of Tsukuba, researchers built DeMemSeg, a deep-learning pipeline that segments the prospore membrane, a membrane structure that forms in budding yeast during sporulation, the process by which the yeast cell forms spores. The team reports that DeMemSeg's accuracy is statistically comparable to expert manual analysis, and that it also worked on mutant cells with abnormal membrane shapes that were not part of its training data.

A separate team built CellSeg3D, a toolbox for segmenting individual cells and nuclei directly in 3D, described in an eLife reviewed preprint. It was tested on whole mouse brain images taken with a light-sheet microscope, an imaging method that captures large 3D volumes of tissue with minimal damage to the sample. The researchers generated the first human-annotated 3D dataset of its kind for this type of neuron in the mouse neocortex, containing 2,632 neurons, since no such 3D-labeled dataset previously existed for benchmarking these methods.

Two more tools address different targets. MetaMP is a web application that unifies four separate membrane protein databases and applies machine learning to classify structures and detect data problems, according to researchers including those at the Robert Koch Institute and the University of Marburg. And a 3D image filter, developed for electron tomography, a technique that reconstructs 3D volumes of cells at near-molecular detail, automatically detects the edges of macromolecular structures inside noisy cellular images without requiring a person to tune its settings.

DeMemSeg is described as an AI-driven pipeline based on deep learning; the source material does not detail its internal architecture beyond that. Its first step is single-cell segmentation, identifying each individual cell in an image, which the researchers describe as the crucial foundation for both training the model and running large-scale analysis afterward.

CellSeg3D takes a different route, comparing several approaches side by side. Two established methods, Cellpose and StarDist, work by predicting either flow vectors that pull pixels toward each object's center, or distances from each pixel to the boundary of the object it belongs to. The CellSeg3D team also trained two other models, a transformer-based network called SwinUNetR and a convolutional network called SegResNet, and found both performed comparably to Cellpose and StarDist. They then built a self-supervised model, WNet3D, which needs no human-labeled examples at all; instead it works from the assumption that a cell's shape and boundary can be inferred just from patterns already present in the raw 3D image, such as local contrast and spatial coherence.

The electron tomography filter works by scoring every unit of volume in the 3D image, called a voxel, on how likely it is to sit on a true edge rather than in noise. It compares each voxel's brightness to its neighbors and produces a score between 0 and 1, where a score near 0 marks a likely edge and near 1 marks likely noise; scores are then averaged with more weight given to nearby voxels than distant ones. A voxel is finally classified as edge or non-edge using a cutoff of 0.85, a value the researchers arrived at through repeated testing rather than manual tuning by the user, which is what makes the filter parameter-free.

The electron tomography filter detected edges as narrow as 2 pixels and picked out signal only 0.72% above the average background noise, and its performance was found to be comparable to or better than two established filters, the 3D recursive filter and the Canny filter, while requiring no manual parameter tuning. CellSeg3D's WNet3D model was found to be as good as or better than the fully supervised models, particularly when only a small amount of training data was available, despite using no human-labeled data during training. DeMemSeg's segmentation accuracy was statistically comparable to expert manual analysis, and it also correctly detected and quantified irregular membrane shapes in mutant yeast cells that had never appeared in its training set.

MetaMP was tested through what its developers describe as a validation focused on statistics: it resolved 77% of the data discrepancies it encountered across the four databases it integrates, and correctly predicted the classification of newly identified membrane proteins 98% of the time, a rate the researchers say overtook expert curation in that test. The application also merges 3,569 membrane protein structure entries out of 3,795 total across the four source databases, MPstruc, the Protein Data Bank, OPM, and UniProt, giving researchers a single resource where the four previously had gaps and inconsistencies with each other.

Each result carries its own qualifications. The DeMemSeg findings come from a study of a single model system, the prospore membrane in budding yeast, and the researchers describe the work as a foundational technology whose relevance to human disease mechanisms involving similar membrane abnormalities is a possible future direction rather than a demonstrated one. CellSeg3D's self-supervised approach is presented as having trade-offs: images affected by artifacts from tissue clearing or light-sheet imaging may need an added filtering step to remove overly large or small detected particles, and the eLife posting is a reviewed preprint rather than a final peer-reviewed publication.

MetaMP's validation figures come from what the paper calls a validation focused on statistics, and the tool is described as an initial release still under version control, meaning it may change as it is developed further. The electron tomography filter's threshold of 0.85 was chosen because it repeatedly gave the best results in testing, but the paper does not claim it will hold for every possible imaging condition, and its output is intended as an input to further downstream steps such as sub-tomogram averaging rather than a finished analysis on its own.