DisCell: Separating Spillover from Niche Response in Spatial Transcriptomics

Spatial transcriptomics at single-cell resolution maps the location of each cell within a tissue, but errors in diffusion and segmentation also introduce transcripts from neighboring cells. This spillover makes a cell's expression resemble its neighbors, and a response to its niche causes cells in the same niche to resemble each other.

Both spillover and niche response link a cell's expression to its environment. However, the raw count data cannot differentiate between these two phenomena. Single-cell transcriptomes have identified genes that are more abundant in one niche compared to another or signaling genes that change based on the cell's neighborhood, but these findings may be spillover artifacts.

Existing models describe a cell's intrinsic state in conjunction with either its niche response or spillover, but no model currently captures all three aspects. The authors introduce DisCell, the first method that jointly models spillover, niche response, and intrinsic state, and assesses the robustness of each finding to spillover.

DisCell employs a generative model with two latent variables: an intrinsic state and a response to the niche. Spillover from neighboring cells is treated as a separate term. By fitting the model at varying levels of spillover and recomputing each finding on the spillover-free expression, DisCell quantifies how much spillover a result can withstand.

The method also estimates the maximum spillover a section can accommodate, and a finding that persists even at this level is deemed not influenced by spillover. In simulated tissues, DisCell accurately recovers each cell's intrinsic state and response, distinguishing genuine responses from spillover-induced findings. Tested on five ovarian and lung tissue sections, DisCell preserves more of each cell's unique characteristics in its intrinsic state compared to other methods that retain minimal neighborhood information.

Within each cell type, DisCell separates the variation in gene expression into intrinsic state, niche response, and spillover. The authors utilize DisCell to prioritize niche-related findings based on their resilience to spillover and validate that a previously reported finding remains robust under spillover conditions. For instance, esophageal cancer cells in niches infiltrated by immune or stromal cells express a greater proportion of an epithelial-mesenchymal transition program when analyzed with DisCell.

Furthermore, DisCell accurately predicts a cell type's gene expression in another niche while separating the spillover component, achieving lower error rates across most tasks than a competing method specifically designed for this purpose. Training DisCell takes less than ten minutes on a single consumer GPU for sections containing up to 400,000 cells.