A new registration method aligns single-cell spatial data without shrinking it into pixels first.
Researchers have built Domain Elastic Transform (DET), a grid-free framework that aligns both the shape and the biological signal in datasets like spatial transcriptomics slices, where gene-expression readings sit on scattered, irregular points instead of a neat grid. Existing tools force a tradeoff: bin the data into voxels to align it like an image, which erases single-cell resolution, or align sparse point sets while ignoring the functional signal entirely. DET treats the data as a function on an irregular domain, using a Bayesian model where the deformation between samples is driven by both geometry and gene-expression values. It's fully unsupervised and scales to large datasets by registering a sample of points, then interpolating the deformation across the rest.
On a 90-case MERFISH mouse-brain benchmark with heavy perturbations and no initial alignment, DET produced the best spatial overlap and tissue topology among the pipelines tested. A modified version of a rival tool, PASTE2, still won on a separate accuracy measure called label-transfer ARI, so no single method swept every metric.
The team's own atlas-scale mouse-embryo test had no cross-stage ground truth to check against, so those gains are self-referential rather than proof the alignment matches actual biology, a caveat worth remembering before anyone calls this settled.