Neighbor transcripts leak into spatial cells: DisCell splits spillover from niche response and gives each finding a breakdown point
Synopsis
Single-cell spatial transcriptomics counts neighboring cells' transcripts inside a cell, and this spillover and a cell's real response to its niche both tie expression to surroundings, so counts alone cannot separate them. DisCell models intrinsic state, niche response and spillover jointly, refits over a range of spillover fractions, and assigns each finding a breakdown point. On ovarian and lung sections its intrinsic state keeps more cell-cycle information than any competing method holding as little neighborhood information, and a published esophageal EMT finding survives the modeled spillover range.
Interpretation
In spatial transcriptomics, a cell shifting toward its neighbors can be either a response to the niche or a measurement artifact in which neighbors' transcripts are counted in the cell, and existing models describe intrinsic state with either niche response or spillover, never all three. Earlier practice either corrected counts by an assumed spillover amount before analysis, or filtered spillover after fitting or tested it at fixed levels, so whether a niche-program finding is real depends on how much of the shift toward neighbors the modeling assumptions assign to spillover. In simulated tissue DisCell recovers each cell's intrinsic state and response and separates spillover-made findings from genuine responses; on five ovarian and lung sections its intrinsic state keeps more of each cell's own properties, measured by cell-cycle phase, than any competing method that holds as little neighborhood information.
The model writes each cell's expression as a mixture of the cell's own clean profile and its neighbors' profiles mixed in at one fraction shared across the section. The clean profile comes from the intrinsic state through a network, the niche response shifts expression along a set of gene programs, and neighbors' clean profiles are averaged with weights from shared boundary length and center distance, with the fraction fixed per fit. The model is fitted independently at spillover fractions of 0, 0.05, 0.1, 0.2, 0.3 and 0.4 with three random seeds each, giving every finding a breakdown point, the smallest tested fraction at which it no longer holds, read against the section's count-based upper bound on spillover.
Every finding computed from cell expression gets a breakdown point, which ranks findings by robustness to spillover and flags those that may be artifacts. Counts bound the spillover fraction only from above and never from below, so a single fit cannot say how much spillover a finding depends on. On the ovarian and lung sections most findings hold beyond the tested fraction of 0.4, above each section's estimated bound; but the fresh-frozen ovarian section's bound of 0.49 lies above the tested range, so findings there that hold up to 0.4 are only undecided; on that same section the signaling-gene lean and tumor axis break at 0.2, below the bound, so spillover is not ruled out there.
Perspective
The test covers only spillover modeled as one fraction per section, shared across genes and reaching one ring of neighbors; on simulated tissue and ovarian, lung and esophageal cancer sections, findings are reported as breakdown points read against each section's estimated bound.
That bound constrains only the modeled spillover and is not a measurement of the physical share of misassigned transcripts; the model also cannot separate intrinsic variation organized in space, such as a cancer subclone occupying one region, from response, and such variation is read as response.