Domain Elastic Transform Aligns Gene Activity Maps, Preserving Cell Detail

Researchers from Kanazawa University and Sapienza University of Rome have developed Domain Elastic Transform (DET), a computational algorithm that improves the alignment of spatial transcriptomics maps while preserving single-cell resolution. Led by Osamu Hirose and Emanuele Rodolà, the team published their findings in the 2026 issue of IEEE Transactions on Pattern Analysis and Machine Intelligence. The method addresses a persistent challenge in biological imaging: comparing gene activity profiles across tissue samples that often exhibit natural shape variations, rotation, or physical distortion due to slicing and preparation processes. Current registration techniques face significant limitations. Image-based methods typically convert sparse molecular measurements into regular pixel grids, which blurs fine cellular structures and reduces spatial detail. Approaches relying exclusively on geometric features may misalign regions with similar shapes but distinct biological functions. DET overcomes these issues by simultaneously leveraging spatial coordinates and gene-expression values. The algorithm functions in an unsupervised, training-free mode, iteratively estimating likely matches between cells based on positional proximity and transcriptomic similarity. It then smoothly deforms one digital map to coincide with a reference, adjusting cell positions to encourage neighboring cells to move together while retaining the original measurement locations and values. The researchers validated DET using MERFISH data from mouse brains. In 90 test cases involving random rotations and large translational shifts, DET outperformed competing methods on metrics assessing map overlap, neighbor preservation, and the concordance of gene-activity patterns. The method was also tested on Stereo-seq maps from the MOSTA atlas, aligning mouse embryo data from two distinct developmental stages. This application involved datasets containing over 100,000 measurement locations and demonstrated DET's ability to reconcile maps with evolving tissue morphologies without requiring verified reference correspondences. Computational scalability tests further confirmed the algorithm's utility for large-scale data. By synthetically expanding MERFISH datasets to one million points per slice, the researchers observed peak memory usage averaging less than one gigabyte. DET employs a landmark budget to estimate transformations, allowing computation time to depend primarily on the number of landmarks rather than the total data volume. This design supports efficient processing of massive tissue maps. The framework is intended to complement existing alignment tools, offering a solution for tasks requiring both fine spatial fidelity and smooth transformation, with potential applications extending to other scientific domains where datasets combine location data with quantitative measurements.