Systematic evaluation of robustness to cell type mismatch of deconvolution methods for spatial transcriptomics data

Sequencing-based spatial transcriptomics (ST) approaches preserve spatial information but with limited cellular resolution. Single-cell RNA-sequencing (scRNA-seq) techniques, on the other hand, provide single-cell resolution but lose spatial resolution because of the tissue dissociation step. With these complementary strengths in mind, computational tools have been developed to combine scRNA-seq and ST data. These approaches use deconvolution to identify cell types and their reoctive proportions present at each location in ST data, with the aid of a scRNA-seq reference dataset. It has been suggested that deconvolution methods are sensitive to the absence of cell types in the scRNA-seq reference, a problem referred to as cell type mismatch. Here, we used extensive simulations to systematically evaluate the robustness to cell type mismatch of six state-of-the-art deconvolution methods tailored for spatially resolved transcriptomics data, along with two deconvolution methods designed for bulk RNA-seq data. At baseline, that is, with no cell types missing from the reference data, cell2location, RCTD, and CARD were the best performing methods, while SPOTlight performed worst. By simulating various cell type mismatch scenarios, we found that the performance of deconvolution methods decreases proportionally to the number of cell types missing from the reference data. Moreover, for most deconvolution methods the decrease in performance is similar relative to their baseline performance. We also observed that those methods that perform well at baseline tend to assign the proportions of a missing cell type to the transcriptionally most similar cell types present in the reference data. This study highlights the adverse effects of cell type mismatch on the performance of deconvolution methods for ST data and stresses the need for methods that are more robust to this type of mismatch.