Harman (version 1.0.2)

harman: Harman batch correction method

Description

Harman is a PCA and constrained optimisation based technique that maximises the removal of batch effects from datasets, with the constraint that the probability of overcorrection (i.e. removing genuine biological signal along with batch noise) is kept to a fraction which is set by the end-user (Oytam et al, 2016).

Harman expects unbounded data, so for example, with HumanMethylation450 arrays do not use the Beta statistic (with values constrained between 0 and 1), instead use the logit transformed M-values.

Usage

harman(datamatrix, expt, batch, limit = 0.95, numrepeats = 100000L, randseed, forceRand = FALSE, printInfo = FALSE)

Arguments

datamatrix

matrix or data.frame, the data values to correct with samples in columns and data values in rows. Internally, a data.frame will be coerced to a matrix. Matrices need to be of type integer or double.

expt

vector or factor with the experimental variable of interest (variance to be kept).

batch

vector or factor with the batch variable (variance to be removed).

limit

numeric, confidence limit. Indicates the limit of confidence in which to stop removing a batch effect. Must be between 0 and 1.

numrepeats

integer, the number of repeats in which to run the simulated batch mean distribution estimator using the random selection algorithm. (N.B. 32 bit Windows versions may have an upper limit of 300000 before catastrophic failure)

randseed

integer, the seed for random number generation.

forceRand

logical, to enforce Harman to use a random selection algorithm to compute corrections. Force the simulated mean code to use random selection of scores to create the simulated batch mean (rather than full explicit calculation from all permutations).

printInfo

logical, whether to print information during computation or not.

Value

A harmanresults S3 object.

Details

The datamatrix needs to be of type integer or numeric, or alternatively a data.frame that can be coerced into one using as.matrix. The matrix is to be constructed with data values (typically microarray probes or sequencing counts) in rows and samples in columns, much like the `assayData` slot in the canonical Bioconductor eSet object, or any object which inherits from it. The data should have normalisation and any other global adjustment for noise reduction (such as background correction) applied prior to using Harman. For converge, the number of simulations, numrepeats parameter should probably should be at least 100,000. The underlying principle of Harman rests upon PCA, which is a parametric technique. This implies Harman should be optimal when the data is normally distributed. However, PCA is known to be rather robust to very non-normal data.

References

Oytam, et al. (2016).

See Also

harman, reconstructData, pcaPlot, arrowPlot

Examples

library(HarmanData)
data(OLF)
expt <- olf.info$Treatment
batch <- olf.info$Batch
olf.harman <- harman(olf.data, expt, batch)
plot(olf.harman)
olf.data.corrected <- reconstructData(olf.harman)

## Reading from a csv file
datafile <- system.file("extdata", "NPM_data_first_1000_rows.csv.gz",
package="Harman")
infofile <- system.file("extdata", "NPM_info.csv.gz", package="Harman")
datamatrix <- read.table(datafile, header=TRUE, sep=",", row.names="probeID")
batches <- read.table(infofile, header=TRUE, sep=",", row.names="Sample")
res <- harman(datamatrix, expt=batches$Treatment, batch=batches$Batch)
arrowPlot(res, 1, 3)