hlaParallelAttrBagging: Build a HIBAG model via parallel computation
In zhengxwen/HIBAG: HLA Genotype Imputation with Attribute Bagging

hlaParallelAttrBagging

R Documentation

Build a HIBAG model via parallel computation

Description

To build a HIBAG model for predicting HLA types via parallel computation.

Usage

hlaParallelAttrBagging(cl, hla, snp, auto.save="",
    nclassifier=100L, mtry=c("sqrt", "all", "one"), prune=TRUE, na.rm=TRUE,
    mono.rm=TRUE, maf=NaN, stop.cluster=FALSE, verbose=TRUE,
    verbose.detail=FALSE)

Arguments

`cl`	`NULL`, `FALSE`, `TRUE`, an integer, or a cluster object created by the parallel-package; if `NULL` or `FALSE`, use the serial implementation; if `TRUE`, use the number of threads returned from `RcppParallel::defaultNumThreads()` (by default using all threads); if an integer, specify the number of threads; When `cl` is `TRUE` or an integer, the multithreading implementation will be used; when `cl` is a cluster, the multi-processing implementation will be used where each individual classifier is built within a child process
`hla`	training HLA types, an object of `hlaAlleleClass`
`snp`	training SNP genotypes, an object of `hlaSNPGenoClass`
`auto.save`	specify a autosaved file name for an R object (.rda, .RData or .rds); "", no file saving; see details
`nclassifier`	the total number of individual classifiers
`mtry`	a character or a numeric value, the number of variables randomly sampled as candidates for each selection. See details
`prune`	if TRUE, to perform a parsimonious forward variable selection, otherwise, exhaustive forward variable selection. See details
`na.rm`	if TRUE, remove the samples with missing HLA types
`mono.rm`	if TRUE, remove monomorphic SNPs
`maf`	MAF threshold for SNP filter, excluding any SNP with MAF < `maf`
`stop.cluster`	`TRUE`: stop cluster nodes after completing the calculation
`verbose`	if `TRUE`, show information
`verbose.detail`	if `TRUE`, show more information

Details

mtry (the number of variables randomly sampled as candidates for each selection): "sqrt", using the square root of the total number of candidate SNPs; "all", using all candidate SNPs; "one", using one SNP; an integer, specifying the number of candidate SNPs; 0 < r < 1, the number of candidate SNPs is "r * the total number of SNPs".

prune: there is no significant difference on accuracy between parsimonious and exhaustive forward variable selections. If prune = TRUE, the searching algorithm performs a parsimonious forward variable selection: if a new SNP predictor reduces the current out-of-bag accuracy, then it is removed from the candidate SNP set for future searching. Parsimonious selection helps to improve the computational efficiency by reducing the searching times of non-informative SNP markers.

An autosave function is available in hlaParallelAttrBagging when an new individual classifier is built internally without completing the ensemble.

Value

Return an object of hlaAttrBagClass if auto.save="", and NULL otherwise.

Author(s)

Xiuwen Zheng

References

Zheng X, Shen J, Cox C, Wakefield J, Ehm M, Nelson M, Weir BS; HIBAG – HLA Genotype Imputation with Attribute Bagging. Pharmacogenomics Journal. doi: 10.1038/tpj.2013.18. https://www.nature.com/articles/tpj201318

Examples

# make a "hlaAlleleClass" object
hla.id <- "A"
hla <- hlaAllele(HLA_Type_Table$sample.id,
    H1 = HLA_Type_Table[, paste(hla.id, ".1", sep="")],
    H2 = HLA_Type_Table[, paste(hla.id, ".2", sep="")],
    locus=hla.id, assembly="hg19")

# divide HLA types randomly
set.seed(100)
hlatab <- hlaSplitAllele(hla, train.prop=0.5)
names(hlatab)
# "training"   "validation"
summary(hlatab$training)
summary(hlatab$validation)

# SNP predictors within the flanking region on each side
region <- 500   # kb
snpid <- hlaFlankingSNP(HapMap_CEU_Geno$snp.id, HapMap_CEU_Geno$snp.position,
    hla.id, region*1000, assembly="hg19")
length(snpid)  # 275

# training and validation genotypes
train.geno <- hlaGenoSubset(HapMap_CEU_Geno,
    snp.sel = match(snpid, HapMap_CEU_Geno$snp.id),
    samp.sel = match(hlatab$training$value$sample.id,
    HapMap_CEU_Geno$sample.id))
test.geno <- hlaGenoSubset(HapMap_CEU_Geno,
    samp.sel=match(hlatab$validation$value$sample.id,
    HapMap_CEU_Geno$sample.id))


#############################################################################
# Multithreading

set.seed(100)

# train a HIBAG model in parallel with 2 cores
# please use "nclassifier=100" when you use HIBAG for real data
model <- hlaParallelAttrBagging(2, hlatab$training, train.geno, nclassifier=4)


#############################################################################
# Multicore & autosave

library(parallel)

# choose an appropriate cluster size, e.g., 2
cl <- makeCluster(2)
set.seed(100)

# train a HIBAG model in parallel
# please use "nclassifier=100" when you use HIBAG for real data
hlaParallelAttrBagging(cl, hlatab$training, train.geno, nclassifier=4,
    auto.save="tmp_model.RData", stop.cluster=TRUE)

mobj <- get(load("tmp_model.RData"))
summary(mobj)
model <- hlaModelFromObj(mobj)

# validation
pred <- hlaPredict(model, test.geno)
summary(pred)

# compare
hlaCompareAllele(hlatab$validation, pred, allele.limit=model)$overall


# since 'stop.cluster=TRUE' used in 'hlaParallelAttrBagging'
# need a new cluster
cl <- makeCluster(2)

pred <- hlaPredict(model, test.geno, cl=cl)
summary(pred)

# stop parallel nodes
stopCluster(cl)


# delete the temporary file
unlink(c("tmp_model.RData"), force=TRUE)

zhengxwen/HIBAG documentation built on April 16, 2024, 8:41 a.m.