graphSpatialFDR: Control the spatial FDR
In MarioniLab/miloR: Differential neighbourhood abundance testing on a graph

graphSpatialFDR

R Documentation

Control the spatial FDR

Description

Borrowing heavily from cydar which corrects for multiple-testing using a weighting scheme based on the volumetric overlap over hyperspheres. In the instance of graph neighbourhoods this weighting scheme can use graph connectivity or incorpate different within-neighbourhood distances for the weighted FDR calculation.

Arguments

`x.nhoods`	A list of vertices and the constituent vertices of their neighbourhood
`graph`	The kNN graph used to define the neighbourhoods
`pvalues`	A vector of p-values calculated from a GLM or other appropriate statistical test for differential neighbourhood abundance
`k`	A numeric integer that determines the kth nearest neighbour distance to use for the weighted FDR. Only applicaple when using `weighting="k-distance"`.
`weighting`	A string scalar defining which weighting scheme to use. Choices are: max, k-distance, neighbour-distance or graph-overlap.
`reduced.dimensions`	(optional) A `matrix` of cells X reduced dimensions used to calculate the kNN graph. Only necessary if this function is being used outside of `testNhoods` where the `Milo` object is not available
`distances`	(optional) A `matrix` of cell-to-cell distances or a list of distance matrices, 1 per neighbourhood. Only necessary if this function is being used outside of `testNhoods` where the `Milo` object is not available.
`indices`	(optional) A list of neighbourhood index vertices in the same order as the input neighbourhoods. Only used for the k-distance weighting.

Details

Each neighbourhood is weighted according to the weighting scheme defined. k-distance uses the distance to the kth nearest neighbour of the index vertex, neighbour-distance uses the average within-neighbourhood Euclidean distance in reduced dimensional space, max uses the largest within-neighbourhood distance from the index vertex, and graph-overlap uses the total number of cells overlapping between neighborhoods (distance-independent measure). The frequency-weighted version of the BH method is then applied to the p-values, as in cydar.