Impact of Segmentation Errors in Analysis of Spatial Transcriptomics Data

Spatial transcriptomics aims to elucidate cell coordination within biological tissues by linking the state of the cell with its local tissue microenvironment. Imaging-based assays are particularly promising for exploring such interdependencies, as they can resolve molecular and cellular features with subcellular resolution in three dimensions. Quantification and analysis of cellular state in such data, however, ultimately depends on the ability to recognize which molecules belong to each cell. Despite computational and experimental progress, this cell segmentation task remains challenging. Here we re-analyze data from multiple tissues and platforms and find that segmentation errors currently confound most downstream analysis of cellular state, including analysis of differential expression, inference of neighboring cell influence, and ligand-receptor interactions. The extent to which mis-segmented molecules impact the results can be striking, often dominating the set of top hits. We show that factorization of molecular neighborhoods can be effective at isolating such molecular admixtures and minimizing their impact on downstream analysis, analogous to doublet filtering of scRNA-seq data. As applications of spatial transcriptomics assays become more widespread, we expect corrections for the confounding effect of segmentation errors to become increasingly important for being able to resolve molecular mechanisms of tissue biology.