TIDEST: post-imputation differential expression testing for spatial transcriptomics data

Abstract

Spatial transcriptomics is increasingly extended by computational reconstruction of unmeasured genes from matched single-cell RNA-sequencing references, enabling transcriptome-wide analyses from targeted or sparse assays. Yet downstream analyses typically treat reconstructed expression as experimentally observed, overlooking prediction error and latent spatial variation that can misattribute tissue architecture to biological regulation. Here we introduce TIDEST, a framework for statistically valid inference on reconstructed spatial transcriptomes. TIDEST calibrates reconstructed expression using measured genes and adjusts for latent spatial variation before differential expression analysis. Across realistic spatial tissue simulations, TIDEST controls false discoveries where existing approaches fail, while preserving power. Across mouse and human brain, glioblastoma and breast cancer, TIDEST changes biological interpretation by correcting misleading differential-expression calls and revealing disease-associated transcriptional programs obscured by spatial confounding. Our results show that prediction accuracy alone is insufficient for reliable biological discovery and establish valid statistical inference as an essential component of reconstructed spatial transcriptomics.