Test neighbourhood-dependent differential expression

Completes a spiDE analysis: for each bandwidth, per-gene Wald/Brown inference is computed (block-wise, in parallel) if not already present; p-values are combined across bandwidths by a log-likelihood-weighted Cauchy combination; and a hierarchical (gene -> index cell type -> niche cell type) Benjamini- Hochberg procedure identifies significant neighbourhood-dependent DE. Results are returned as a tidy table via results().

Usage

testSpiDE(object, ...)

# S4 method for class 'SpiDEResults'
testSpiDE(
  object,
  spe = NULL,
  assay = "counts",
  fdr = 0.05,
  weight.thresh = 0.1,
  combine = c("cauchy", "brown"),
  block.size = NULL,
  backend = c("auto", "cpu", "gpu"),
  gpu.mem.budget = NULL,
  BPPARAM = BiocParallel::SerialParam(),
  ...
)

Arguments

object: a SpiDEResults from fitSpiDE().
...: ignored.
spe: the SpatialExperiment used for fitting (required to compute the Wald/Brown inference if it is not already present on the fits).
assay: a character, the counts assay (used only if inference must be computed).
fdr: a numeric in (0, 1], the target false discovery rate (default 0.05). Up/Down association directions are tested separately and combined, which is mathematically equivalent to gating the reported, combined fdr.gene/fdr.index/fdr.niche values at fdr directly - so fdr = 1 returns every (gene, index, niche) result, unfiltered.
weight.thresh: a numeric, Cauchy weights below this are set to zero.
combine: one of "cauchy" or "brown", the within-gene combiner for the correlated niche p-values (gene-level and per-index-cell-type). "cauchy" (default) is the correlation-agnostic Cauchy combination test (ACAT), which controls type-I error under correlation without estimating a correlation matrix; "brown" is the correlation-aware Brown's method. Used only if the Wald inference is not already present on the fits.
block.size: a numeric, genes per inference block (NULL = a single block on the CPU backend, or a memory-bounded auto-selected size on the GPU backend).
backend: a character, the compute backend for the inference stage ("auto", "cpu", or "gpu"). The GPU backend batches the per-gene Wald covariance and NB math across each block via SpaNorm's tensor engine; it also forces a serial BPPARAM (with a warning) to avoid multiple processes contending for one GPU device. Used only if the Wald inference is not already present on the fits.
gpu.mem.budget: a numeric, the GPU memory budget in bytes used to size inference blocks (NULL auto-detects; only relevant for the GPU backend). The batched Wald covariance is bounded separately, by a gene sub-batch sized from this budget on the GPU path and from getOption("spiDE.cov.mem.budget", 2e9) on the CPU path – raise the latter to trade memory for speed on wide (random-slope) designs.
BPPARAM: a BiocParallelParam for the inference stage.

Value

the input object with combined p-values and the tidy results table populated.

Examples

data(toySpiDE)
spe <- buildNiches(toySpiDE, sigma = 20)
res <- fitSpiDE(spe, condition = "condition", sigma = 20, verbose = FALSE)
res <- testSpiDE(res, spe = spe)
head(results(res))
#>   gene ct_index ct_niche bandwidth.max     coef       t DirectionGene
#> 1   G1        A        B            20 3.173422 4.63951            Up
#>   DirectionIndex DirectionNiche     fdr.gene    fdr.index    fdr.niche
#> 1             Up             Up 0.0005520939 2.760426e-05 9.201414e-06