We first create a toy example using an available large-scale scRNA-seq
library(TENxPBMCData)
sce <- TENxPBMCData("pbmc4k")
mat <- counts(sce)
rownames(mat) <- rowData(sce)$ENSEMBL_ID
colnames(mat) <- colData(sce)$Sequence
writeHDF5Array(mat, filepath = "counts.hdf5", name = "count")
saveRDS(list(row.names = rowData(sce)$ENSEMBL_ID , col.names = colData(sce)$Sequence), file = "dimnames.RDS")In the latest version of HDF5Array, dimension names can be written in to the hdf5 file directly writeHDF5Array(mat, filepath = "counts.hdf5", name = "count", with.dimnames=T)
Load all functions in funcs.R
source("funcs.R") #glmnet and HDF5Array will be loaded
setAutoRealizationBackend("HDF5Array") #supportedRealizationBackends(), getRealizationBackend()Read in the file we just created. The name can be verified via h5ls("counts.hdf5")
sce <- DelayedArray(seed = HDF5ArraySeed(filepath = "counts.hdf5", name = "count"))Read the dimension names
dimnamaes <- readRDS("dimnames.RDS")
rownames(sce) <- dimnamaes$row.names
colnames(sce) <- dimnamaes$col.namesRemove all NA values and genes that are not differentially expressed
sce[is.na(sce)] <- 0
data <- sce[which(DelayedMatrixStats::rowSds(sce) >0),]Read in prior information matrix
priorMat <- readRDS("canonicalPathways_ENSG.RDS")Run the DelayedPLIER
PLIER.res$residual, PLIER.res$Z and PLIER.res$B are stored as three on-disk files instead of variables in the memory. PLIER() function will realize and write these three to hard disk based on the output_path. Other values are still in the memory can be directly accessed via PLIER.res$'name of the values'
ptm <- proc.time()
PLIER.res <- PLIER(data, priorMat, output_path = "output/")
print(proc.time()-ptm)