SpaCoDA: fast compositional deconvolution maps recurrent cellular niches across hundreds of spatial transcriptomics sections

Background Spot-based spatial transcriptomics such as 10x Visium measures a mixture of cells in each spot. Estimating the cell-type composition of each spot against a single-cell reference is the first step towards cell-level maps of tissue microenvironments, but the most accurate deconvolution methods are slow and memory-intensive at the scale of current atlases, and there still need an established way to make tissue niches comparable across hundreds of sections. Results We present SpaCoDA, which fits each spot by variance-weighted non-negative least squares against a reference signature, re-estimates a gene-level platform factor between iterations, and converts RNA shares into cell fractions using cell-size and marker-capture factors. Niches are identified within each section from fused compositional (Aitchison distances) and spatial distances, and their prototypes are integrated across sections into meta-niches. Across 123 simulated configurations under three reference-mismatch scenarios, SpaCoDA was the only one of nine methods ranked within the top two on all four evaluation metrics in every scenario, second to RCTD in mean rank. On 12 imaging-derived datasets with known composition (six Xenium datasets, each evaluated at two annotation levels, and six CosMx/Xenium datasets from three patients), it achieved the best mean rank in every summary, while running 4 to about 1,800-fold faster than the other eight methods (233-fold faster than RCTD by median runtime) with the lowest peak memory. Applied to 254 breast cancer and 181 glioblastoma sections (about 1.18 million spots), SpaCoDA identified 29 and 31 meta-niches, respectively. In breast cancer, the expression signature of a B-cell and plasma-cell meta-niche was associated with longer overall survival in TCGA-BRCA and METABRIC and was higher in tumors with a pathological complete response (pCR) to neoadjuvant therapy in I-SPY2. In glioblastoma, signatures of a perivascular antigen-presenting myeloid meta-niche and of a proliferative, hypoxic meta-niche were associated with shorter survival in TCGA-GBM and three CGGA glioma cohorts. Conclusions SpaCoDA combines competitive accuracy with orders-of-magnitude lower computational cost, making composition and niche analysis practical at cohort scale. It provides a practical framework for such analyses, enabling disease subtyping and the identification of novel therapeutic targets, and shows promise for clinical translation.