Does spatial context transfer? Neighborhood-aware cell typing in seminiferous tubules under tissue-organization and species shift

Methods for annotating spatial transcriptomics data increasingly use a spot's tissue neighborhood as input, but they are benchmarked within one tissue and condition, so it is unknown whether the spatial signal they learn holds when tissue organization changes. In this study, we build an eight-puck benchmark from the public Slide-seqV2 testis pucks of Chen et al. (2021): three wild-type mouse, three diabetic ob/ob mouse, and two human pucks (252,898 beads), with cell-type labels from NMFreg (non-negative matrix factorization regression). Using germ-cell stage annotation as the task, we compare an expression-only classifier with neighborhood-augmented features, a GraphSAGE model, and test-time smoothing of the expression-only posterior, under held-out-puck, wild-type-to-ob/ob, and mouse-to-human transfer. Spatial context adds about one macro-F1 point within a species and up to three on human pucks. The ob/ob condition, where tubule organization is disrupted, is not a distribution shift for this task. Across species, expression-only accuracy drops by ten points. Neighborhood-augmented features and test-time smoothing keep their gains, whereas GraphSAGE, which matches them within a species, reverses its gain across species, ending five points below neighborhood-augmented features and losing fifteen F1 points on round spermatids. A representation-shift analysis and a depth ablation trace this to the neighborhood representation, which moves further between species than the bead's own in all 24 stage-by-puck comparisons. The gain from context tracks how informative the neighborhood is relative to the bead itself rather than the target tissue's neighborhood homogeneity. Because the two are indistinguishable in distribution and diverge under species shift, a benchmark confined to one condition cannot rank them.