spAlignDE unifies cross-sample and cross-modal spatial alignment with mismatch-aware differential expression

Abstract

Comparative analysis of spatial omics requires aligning data across samples and modalities to a common coordinate system. Existing methods can be computationally intensive for large datasets, and cross-modal alignment is difficult when datasets lack comparable molecular features. In addition, residual alignment errors can cause locations assigned to the same coordinates to represent different biological regions, producing false differential expression signals. Here we propose spAlignDE, a computational method that integrates structure-guided spatial alignment with mismatch-aware local differential expression analysis. spAlignDE represents structures from spatial transcriptomics, spatial ATAC-seq, histology, and anatomical atlases as continuous fields and aligns them by shooting-based diffeomorphic registration without requiring shared molecular features. In cross-sample benchmarks against 12 methods, spAlignDE achieved the highest agreement in gene expression patterns and anatomical annotations. It also scaled to 20 MERFISH brain sections containing 1.45 million cells. For cross-modal tasks, spAlignDE accurately aligned spatial transcriptomics with histology, the Allen Mouse Brain Common Coordinate Framework, and spatial ATAC-seq. After alignment, spAlignDE estimates local expression contrasts on a shared grid and inflates their variances according to mismatch risk estimated from putatively stable genes and local observation density. The analysis can also adjust for cell-type composition. Simulations showed improved false discovery control without systematic loss of power. In real-data applications, spAlignDE localized age-associated changes in gene expression and T-cell distribution in the mouse brain and spatially restricted expression differences between normal and injured kidney sections.