tests/testthat/test_checkSeparation.R

In MikeDMorgan/miloR: Differential neighbourhood abundance testing on a graph

context("Testing checkSeparation function")

### Set up a mock data set using simulated data
suppressWarnings({
    library(miloR)
    library(SingleCellExperiment)
    library(scran)
    library(scater)
    library(irlba)
    library(MASS)
    library(mvtnorm)
})

set.seed(42)
r.n <- 1000
n.dim <- 50
block1.cells <- 500
# select a set of eigen values for the covariance matrix of each block, say 50 eigenvalues?
block1.eigens <- sapply(1:n.dim, FUN=function(X) rexp(n=1, rate=abs(runif(n=1, min=0, max=50))))
block1.eigens <- block1.eigens[order(block1.eigens)]
block1.p <- qr.Q(qr(matrix(rnorm(block1.cells^2, mean=4, sd=0.01), block1.cells)))
block1.sigma <- crossprod(block1.p, block1.p*block1.eigens)
block1.gex <- abs(rmvnorm(n=r.n, mean=rnorm(n=block1.cells, mean=2, sd=0.01), sigma=block1.sigma))


block2.cells <- 500
# select a set of eigen values for the covariance matrix of each block, say 50 eigenvalues?
block2.eigens <- sapply(1:n.dim, FUN=function(X) rexp(n=1, rate=abs(runif(n=1, min=0, max=50))))
block2.eigens <- block2.eigens[order(block2.eigens)]
block2.p <- qr.Q(qr(matrix(rnorm(block2.cells^2, mean=4, sd=0.01), block2.cells)))
block2.sigma <- crossprod(block2.p, block2.p*block2.eigens)
block2.gex <- abs(rmvnorm(n=r.n, mean=rnorm(n=block2.cells, mean=4, sd=0.01), sigma=block2.sigma))


block3.cells <- 650
# select a set of eigen values for the covariance matrix of each block, say 50 eigenvalues?
block3.eigens <- sapply(1:n.dim, FUN=function(X) rexp(n=1, rate=abs(runif(n=1, min=0, max=50))))
block3.eigens <- block3.eigens[order(block3.eigens)]
block3.p <- qr.Q(qr(matrix(rnorm(block3.cells^2, mean=4, sd=0.01), block3.cells)))
block3.sigma <- crossprod(block3.p, block3.p*block3.eigens)
block3.gex <- abs(rmvnorm(n=r.n, mean=rnorm(n=block3.cells, mean=5, sd=0.01), sigma=block3.sigma))

sim1.gex <- do.call(cbind, list("b1"=block1.gex, "b2"=block2.gex, "b3"=block3.gex))
colnames(sim1.gex) <- paste0("Cell", 1:ncol(sim1.gex))
rownames(sim1.gex) <- paste0("Gene", 1:nrow(sim1.gex))
sim1.pca <- prcomp_irlba(t(sim1.gex), n=50, scale.=TRUE, center=TRUE)

set.seed(42)
block1.cond <- rep("A", block1.cells)
block1.a <- sample(1:block1.cells, size=floor(block1.cells*0.9))
block1.b <- setdiff(1:block1.cells, block1.a)
block1.cond[block1.b] <- "B"

block2.cond <- rep("A", block2.cells)
block2.a <- sample(1:block2.cells, size=floor(block2.cells*0.05))
block2.b <- setdiff(1:block2.cells, block2.a)
block2.cond[block2.b] <- "B"

block3.cond <- rep("A", block3.cells)
block3.a <- sample(1:block3.cells, size=floor(block3.cells*0.5))
block3.b <- setdiff(1:block3.cells, block3.a)
block3.cond[block3.b] <- "B"

meta.df <- data.frame("Block"=c(rep("B1", block1.cells), rep("B2", block2.cells), rep("B3", block3.cells)),
                      "Condition"=c(block1.cond, block2.cond, block3.cond),
                      "Replicate"=c(rep("R1", floor(block1.cells*0.33)), rep("R2", floor(block1.cells*0.33)),
                                    rep("R3", block1.cells-(2*floor(block1.cells*0.33))),
                                    rep("R1", floor(block2.cells*0.33)), rep("R2", floor(block2.cells*0.33)),
                                    rep("R3", block2.cells-(2*floor(block2.cells*0.33))),
                                    rep("R1", floor(block3.cells*0.33)), rep("R2", floor(block3.cells*0.33)),
                                    rep("R3", block3.cells-(2*floor(block3.cells*0.33)))))
colnames(meta.df) <- c("Block", "Condition", "Replicate")
# define a "sample" as teh combination of condition and replicate
meta.df$Sample <- paste(meta.df$Condition, meta.df$Replicate, sep="_")
meta.df$Vertex <- c(1:nrow(meta.df))

sim1.sce <- SingleCellExperiment(assays=list(logcounts=sim1.gex),
                                 reducedDims=list("PCA"=sim1.pca$x))

sim1.mylo <- Milo(sim1.sce)

# build a graph - this can take a while for large graphs - will need to play
# around with the parallelisation options
sim1.mylo <- buildGraph(sim1.mylo, k=21, d=30)

# define neighbourhoods - this is slow for large data sets
# how can this be sped up? There are probably some parallelisable steps
sim1.mylo <- makeNhoods(sim1.mylo, k=21, prop=0.1, refined=TRUE,
                        d=30,
                        reduced_dims="PCA")
sim1.mylo <- calcNhoodDistance(sim1.mylo, d=30)

sim1.meta <- data.frame("Condition"=c(rep("A", 3), rep("B", 3)),
                        "Replicate"=rep(c("R1", "R2", "R3"), 2))
sim1.meta$Sample <- paste(sim1.meta$Condition, sim1.meta$Replicate, sep="_")
rownames(sim1.meta) <- sim1.meta$Sample

sim1.mylo <- countCells(sim1.mylo, samples="Sample", meta.data=meta.df)


test_that("Wrong input gives errors", {
    # no rownames - can't do subsetting
    rownames(sim1.meta) <- NULL
    expect_error(checkSeparation(sim1.mylo, sim1.meta, "Condition"),
                 "Please add rownames to design.df that are the same as the colnames of nhoodCounts",
                 fixed=TRUE)

    rownames(sim1.meta) <- paste0("NOTSAMPLE", sim1.meta$Sample)
    expect_error(checkSeparation(sim1.mylo, sim1.meta, "Condition"),
                 "rownames of design.df are not a subset of nhoodCounts colnames",
                 fixed=TRUE)

    # Asking for column not in coldata
    expect_error(checkSeparation(sim1.mylo, sim1.meta, "BLAH"),
                 "BLAH is not a variable in design",
                 fixed=TRUE)

    nhoodCounts(sim1.mylo) <- matrix(0L)
    expect_error(checkSeparation(sim1.mylo, sim1.meta, "Condition"),
                 "nhoodCounts not found - please run countCells() first",
                 fixed=TRUE)

})

test_that("Factor checking handles expected formats", {
    sim1.meta$Variable <- seq_len(nrow(sim1.meta))
    expect_error(checkSeparation(sim1.mylo, sim1.meta, "Variable"),
                 "Too many levels in Variable",
                 fixed=FALSE)
})

test_that("Output is the expected format", {
    # set the first 5 nhood counts to 0 for one condition level
    nhoodCounts(sim1.mylo)[c(1:5), sim1.meta$Sample[sim1.meta$Condition %in% c("A")]] <- 0
    expect_gte(sum(checkSeparation(sim1.mylo, sim1.meta, "Condition", min.val=1)), 5)

    expect_identical(length(checkSeparation(sim1.mylo, sim1.meta, "Condition", min.val=1)), nrow(nhoodCounts(sim1.mylo)))
})