New algorithm aligns gene-activity maps of tissue without losing cell detail

Every organ in the body is a landscape. Its cells occupy precise positions, and the genes active within them vary from one neighborhood to the next, creating molecular geography that determines how the tissue functions. Spatial transcriptomics, one of the most powerful technologies in modern biology, allows researchers to chart that geography in extraordinary detail, recording the activity of thousands of genes across a thin slice of tissue at the resolution of individual cells. The resulting maps have transformed how scientists study development, organ architecture, and disease. Yet the technology has long faced an awkward practical problem: two of these molecular maps, even of the same organ, rarely line up with one another.

The difficulty is geometric. Tissue slices are thin sections cut from three-dimensional structures, and the act of cutting, mounting, and preparing them introduces rotations, shifts, and stretches. One slice of a mouse brain may be tilted a few degrees relative to another; one may be slightly distorted so that its cortical layers sit at different distances from the edge. When researchers overlay two such maps, equivalent regions do not automatically coincide. Comparing gene activity in the wrong regions produces misleading conclusions, so meaningful comparison first requires alignment, a computational task known as registration, in which one digital map is shifted, rotated, and smoothly reshaped until its regions correspond to those of another.

Researchers at Kanazawa University in Japan and Sapienza University of Rome have now introduced a method designed to perform this registration directly on the raw spatial transcriptomics data, without sacrificing the cell-level resolution that makes the technology valuable. The approach, called Domain Elastic Transform, or DET, was developed by Osamu Hirose of Kanazawa University in collaboration with Emanuele Rodolà of Sapienza University of Rome and published online in IEEE Transactions on Pattern Analysis and Machine Intelligence on September 15, 2026. The work addresses a persistent trade-off in the field between preserving fine spatial detail and achieving robust alignment.

That trade-off arises because existing alignment strategies each make compromises. Image-based methods can incorporate gene-activity information, but they generally begin by converting the measurements into a regular grid of pixels, a step that blurs fine structures and reduces cell-level detail. Point-based methods, by contrast, keep the measurements as individual points and align them using their locations alone, but this can confuse regions whose shapes are similar while their gene-expression patterns differ. Other approaches can link cells or regions across samples without estimating how the tissue map should be smoothly reshaped, leaving the underlying geometric distortion unresolved. Each strategy uses only part of the information that spatial transcriptomics provides.

DET takes a different route by treating every measured cell or location as carrying two kinds of evidence at once: where it sits in the tissue and which genes are active there. The method represents each cell by its position and its measured gene-activity profile, then proceeds through an iterative process with two alternating steps. First, it estimates likely matches between cells in the two maps, drawing on both their spatial positions and the similarity of their gene-activity values. Second, it adjusts the positions of cells in one digital map, encouraging neighboring cells to move together so that the transformation remains smooth rather than chaotic. Repeating these steps gradually refines the alignment until it stabilizes.

Two design choices distinguish the method. DET modifies the positions of cells in the digital map but never alters the recorded gene-activity values themselves; the expression data serve only as clues about which regions should match. And the method requires no collection of pre-aligned training examples and no manually identified correspondences, making it training-free and unsupervised. In principle, this means it can be applied to any pair of tissue maps without the laborious preparation of reference alignments, a significant practical advantage in a field where verified ground truth is scarce.

The researchers evaluated DET against competing methods in 90 test cases using mouse brain maps generated with MERFISH, a technique that maps gene activity at single-cell resolution. To make the task deliberately difficult, they digitally rotated the maps by random angles spanning the full range from 0 to 360 degrees and shifted them by large distances. Among the methods tested, DET achieved the highest scores on three measures: how well the tissue maps overlapped after alignment, whether neighboring cells remained neighbors in the aligned result, and how closely gene-activity patterns agreed between corresponding regions. For the gene-activity comparison, the researchers combined expression information from nearby cells before comparing patterns, providing a robust assessment of functional correspondence.

The team then confronted a harder biological scenario: developing embryos, whose shapes change over time along with the types and proportions of cells they contain. Using maps from the MOSTA atlas, a collection of mouse development maps created with Stereo-seq, a technique that records gene activity across tissue, they aligned mouse-embryo maps from two developmental stages one day apart. Each map contained more than 100,000 measurement locations. DET first adjusted the overall position and orientation of the maps, then smoothly reshaped one to refine the alignment. Because no verified reference existed showing exactly which locations should match across these stages, the experiment served primarily to demonstrate that the method could be applied to maps of this size and complexity, a scale at which many alignment approaches struggle.

Scalability was tested further by artificially expanding MERFISH-based data to as many as one million points per tissue slice. Peak memory use averaged less than one gigabyte in every setting tested. The method achieves this efficiency through a landmark strategy: it estimates the alignment using a selected subset of points, called landmarks, and extends the estimated movement to the full map. The number of landmarks, which the researchers call the landmark budget, controls how much computation the reshaping step requires. With the number of landmarks held fixed, peak memory use grew approximately in proportion to the total number of points, while computation time depended mainly on the number of landmarks used in the main alignment step rather than on the full dataset size. The researchers note that exact runtimes depend on the computer and algorithm settings, but the results support applying DET to very large tissue maps.

The implications extend beyond tissue slices. Many scientific datasets combine locations with measurements, and similar alignment challenges arise with data attached to three-dimensional surfaces or anatomical structures. The researchers position DET as a complement to existing alignment methods rather than a universal replacement, designed for tasks that require both fine spatial detail and a smooth transformation of the map rather than only a list of matching locations. The broader aim, as Hirose describes it, is to use both kinds of information that such datasets carry, where something is and what was measured there, when bringing different datasets into alignment, rather than treating shape and measurements separately. For spatial transcriptomics, that could mean more reliable comparisons of corresponding regions across samples, even when their shapes and gene-activity patterns differ, strengthening studies of how organs develop and how disease reorganizes the body’s molecular geography. The research was supported in part by the Japan Science and Technology Agency FOREST Program under grant JPMJFR242V.

Subject of Research: A computational registration method for aligning spatial transcriptomics maps of gene activity across tissue samples

Article Title: Making maps of gene activity easier to compare

Article References: Making maps of gene activity easier to compare. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: spatial transcriptomics, image registration, Domain Elastic Transform, gene expression, MERFISH, Stereo-seq, mouse brain, mouse embryo, MOSTA atlas, computational biology, Kanazawa University, IEEE TPAMI