Coordinates Alone: A Geometry-Only Null Model for Spatial Transcriptomics Benchmarks

Spatial transcriptomics methods are evaluated in two dominant regimes: unsupervised spatial-domain identification, scored by adjusted Rand index (ARI) against manual annotations on the same spots the model was fit to, and supervised prediction on spots held out from a single slide. We ask what such a score would be for a model with no access to the transcriptome at all. On the twelve-section LIBD human dorsolateral prefrontal cortex (DLPFC) Visium benchmark, a classifier given only each spot's (x,y) coordinates and the training labels reaches 0.975 mean layer accuracy under a random within-slide split, against 0.539 for an expression-only model and 0.901 for neighbourhood-averaged expression, winning in 12 of 12 sections. The advantage is short-range: it decays monotonically with the train-test spatial gap and survives a control holding the training-set size fixed. It is also not laminar biology: transferred to a different donor under a matched label space, the coordinate-only model falls to 0.140, below the 0.296 majority-class baseline, while expression holds at 0.377 and wins in 32 of 32 section pairs. In the unsupervised regime, clustering coordinates alone gives mean ARI 0.213, exceeding 71 of 72 expression-only pipelines in a preprocessing sweep and statistically indistinguishable from the best of them, while a random Voronoi partition of the tissue reaches 0.187: the effective null here is not zero. A geometry-only reference costs one model fit and should accompany every spatial benchmark number.