Enhancing cell type deconvolution in spatial transcriptomics through incorporating multiple reference panels

Spatial transcriptomics (ST) captures gene expression with spatial context but often at resolutions where each spatial location contains multiple cells. Deconvolution of cell-type proportions is therefore essential, yet existing methods typically rely on a single scRNA-seq reference, limiting accuracy when references are incomplete or mismatched especially in complex tissues. These limitations highlight the need for a framework that can integrate multiple references while leveraging spatial information. We introduce SWIMR (Spatially Weighted Integration for Multi-Reference deconvolution), a statistical framework that integrates multiple scRNA-seq references while accounting for spatial correlations. SWIMR models each spatial location as a weighted combination of references, with both cell-type proportions and reference weights varying across tissue locations. Across comprehensive simulations and diverse ST datasets, SWIMR consistently outperformed single-reference approaches. In developing human heart tissue, it recovered ground-truth cell-type compositions with the highest accuracy. In human lung and breast cancer, SWIMR delineated tumor, stromal, and immune compartments, uncovered tumor-adjacent populations, and detected cell type-specific spatially variable genes missed by other methods. In mouse embryo and liver cancer, it leveraged context-specific references to resolve organ-specific cell types and increase sensitivity for biologically relevant pathways. Together, these results establish SWIMR as a robust framework for multi-reference deconvolution in spatial transcriptomics.