Nonparametric kernel-based detection of spatially variable genes (SVGs) with adaptive shrinkage and scalable multi-sample inference.
CytoKspace summarises each gene with a quadratic form in a sparse
exponential nearest-neighbor kernel and calibrates that statistic by
permuting residuals across spatial locations. The only assumption is
exchangeability of residuals under the null, so p-values are exact in
finite samples regardless of whether expression is Gaussian, Poisson,
negative binomial or zero-inflated.
Install the development version from GitHub:
if (!requireNamespace("devtools", quietly = TRUE))
install.packages("devtools")
devtools::install_github("Ghoshlab/CytoKspace")library(CytoKspace)
library(SpatialExperiment)
spe <- CytoKspace(spe, phi_mode = "per_gene", seed = 1)
res <- as.data.frame(rowData(spe))
head(res[order(res$pval), c("Q_stat", "padj", "theta_shrunk", "lfsr",
"svg_class")])Multiple biological replicates:
res <- CytoKspaceCombine(spe_list, combine_method = "fisher", seed = 1)Two independent lines of evidence per gene:
| Column | Meaning |
|---|---|
pval, padj |
Permutation p-value and BH adjustment |
Q_stat |
Quadratic-form statistic |
phi |
Kernel bandwidth used for that gene |
nperm |
Permutations actually spent |
theta_shrunk |
Empirical Bayes posterior spatial effect |
posterior_sd |
Posterior standard deviation |
lfsr |
Local false sign rate |
svg_class |
Joint label from both tracks |
Rank genes on theta_shrunk rather than Q_stat: the raw statistic
scales with the number of spots and with total kernel affinity, the
shrunk effect does not.
- Distribution-free inference. Exact finite-sample permutation p-values under exchangeability alone.
- Per-gene spatial scales. Genes varying over tens of micrometres are
detected alongside genes varying over hundreds, at
O(N)cost. - Adaptive permutation schedule. The permutation budget concentrates on genes near the significance boundary.
- Adaptive shrinkage. Stabilised effect sizes and local false sign rates, which implicitly correct for the mean-variance relationship in log-transformed spatial data.
- Multi-sample scalability. Per-sample inference plus Fisher or Cauchy combination; memory does not grow with the number of samples.
vignette("CytoKspace")Code for every figure, table and benchmark in the accompanying manuscript lives in the separate reproducibility capsule archived on Zenodo. See the Data availability statement of the paper.
Ghosh T, Ghosh D (2026). Nonparametric kernel-based detection of spatially variable genes with adaptive shrinkage and scalable multi-sample inference. bioRxiv.