Spatial transcriptomics skips the region map: 45.8% more DE genes in chronic stress, with attention maps showing where
Synopsis
Comparative spatial transcriptomics usually partitions tissue into anatomical regions or clusters first and then tests differential expression inside those fixed units, separating localization from testing. SpaceDX instead learns gene-specific spatial attention weights while testing, aggregating spots into an attention-weighted pseudobulk that enters a hierarchical Bayesian logistic regression for subject-level association. In 19 stressed and 19 control mice it found 175 DE genes at 5% empirical FDR, 45.8% more than regional DESeq2's 120; in an Alzheimer's model the two found similar numbers alone, and combining them raised discovery 19.1% over DESeq2 alone.
Interpretation
Differential expression testing no longer needs tissue regions to be specified beforehand, and each gene's association comes with a map of where in the tissue it occurs. Regional pseudobulk approaches first aggregate spots into anatomical regions or data-driven clusters and then run subject-level tests inside them, so where the test happens is fixed before analysis. In semisynthetic simulations built from real mouse brain expression, the two performed comparably when the signal aligned with the predefined region; as the signal deviated from that boundary, ROC-AUC declined for the region-based methods (DESeq2, edgeR, MAST) while SpaceDX held, and the advantage grew with effect size and number of affected spots.
On real tissue, the adaptive weighting recovers association signals that regional aggregation dilutes. Regional aggregation implicitly assumes differential expression aligns with fixed spatial boundaries, so signals confined to subregions, spanning regions, or following continuous gradients are diluted. In the chronic social defeat stress data (19 stressed, 19 control animals) at 5% empirical FDR, SpaceDX identified 175 DE genes versus 120 for DESeq2, 116 for edgeR and 69 for MAST; combining SpaceDX with DESeq2 via ACAT gave 180. In the Alzheimer's data (22 diseased, 18 control) SpaceDX found 364 and DESeq2 366, and the combination reached 436, 19.1% more than DESeq2 alone.
Each gene's attention weights form a spatial map that places gene-level significance onto specific brain structures. A regional test only says whether a gene is significant inside a predefined region, not where in the tissue the association concentrates. Among the top 20 genes in the stress data ranked by evidence that attention weighting beats uniform weighting, all localized strongly to at least one annotated region (AUC > 0.8) and 6 (30%) were not significant under regional pseudobulk testing; 24 of 40 highly localized gene-region pairs (60%) were missed by predefined regional testing. In the Alzheimer's data, 6 of 32 such pairs (18.8%) were missed.
The learned maps group differential genes into sets with cellular and anatomical meaning. A gene-level significance list alone gives little indication of which cell types or brain structures the associations correspond to. In the stress data, SpaceDX-derived genes yielded 98 significantly enriched cell-type terms in CellMarker analysis (FDR < 0.05) versus 2 for DESeq2; one group of genes with correlated attention maps contained six of nine oligodendrocyte markers and localized to deep subcortical structures, where Opalin and Gjb1 showed concordant attention across thalamus, cerebral nuclei and ventral diencephalic/pallidal regions.
Perspective
The demonstrated scope is semisynthetic simulations plus two multi-subject mouse brain 10x Visium datasets, a chronic social defeat stress model and a late-stage Alzheimer's disease model, with all methods compared after permutation-based calibration to 5% empirical FDR. The authors note that regional approaches remain valuable when biologically meaningful boundaries are known and align with the signals of interest.
The method relies on informative spot-level context embeddings, whose quality remains dataset-dependent; the authors suggest improved spatial representation learning and larger multimodal reference atlases may make this component more robust. How consistently learned localization patterns reproduce across independent datasets and biological settings is left to future work.