baySeq (version 2.6.0)

getPriors: Estimates prior parameters for the underlying distributions of 'count' data.

Description

These functions estimate, via maximum or quasi-likelihood methods, the parameters of the underlying distributions for negative binomial distributions on count data, or for beta-binomial distributions on paired count data.

Usage

getPriors.NB(cD, samplesize = 1e5, samplingSubset = NULL,
equalDispersions = TRUE, estimation = "QL", verbose = TRUE, zeroML =
FALSE, consensus = FALSE, cl, ...)
getPriors.BB(cD, samplesize = 1e5, samplingSubset = NULL, verbose =
TRUE, cl, ...)

Arguments

cD

A countData or pairedData object.

samplesize

How large a sample should be taken in estimating the priors?

samplingSubset

If given, the priors will be sampled only from the subset specified.

equalDispersions

Should we assume equal dispersions of data across all groups in the 'cD' object? Defaults to TRUE; see Details.

estimation

Defaults to "QL", indicating quasi-likelihood estimation of priors. Currently, the only other possibilities are "ML", a maximum-likelihood method, and "edgeR", the moderated dispersion estimates produced by the 'edgeR' package. See Details.

verbose

Should status messages be displayed? Defaults to TRUE.

zeroML

Should parameters from zero data (rows that within a group are all zeros) be estimated using maximum likelihood methods (which will result in zeros in the parameters? See Details.

consensus

If TRUE, creates a consensus distribution rather than a separate distribution for each member of the groups structure in the `cD' object. See Details.

cl

A SNOW cluster object.

...

Additional parameters to be passed to the estimateTagwiseDisp function if 'estimation = "edgeR"'.

Value

A countData object.

Details

These functions empirically estimate prior parameters for the distributions used in estimating posterior likelihoods of each count belonging to a particular group. For unpaired count data, the distributions on the data are assumed to be negative binomial. For paired count data, the distributions on the data are assumed to be beta-binomial.

For priors estimated for the negative binomial methods, three options are available. Differences in the options focus on the way in which the dispersion is estimated for the data. In simulation studies, quasi-likelihood methods ('estimation = "QL"') performed best and so these are used by default. Alternatives are maximum-likelihood methods ('estimation = "ML"'), and the 'edgeR' packages moderated dispersion estimates ('estimation = "edgeR"'). The priors estimated for the negative binomial methods ('getPriors.NB') may assume that the dispersion of data for a given row is identical for all group structures defined in 'cD@groups' ('equalDispersions = TRUE'). Alternatively, the dispersions may be estimated individually for each group structure ('equalDispersions = FALSE'). Unless there is a strong reason for believing that the data are differently dispersed between groups, 'equalDispersions = TRUE' is recommended. If 'estimation = "edgeR"' then this parameter is ignored and dispersion is assumed identical for all group structures. If all counts in a given row for a given group are zero, then maximum and quasi-likelihood estimation methods will result in a zero parameter for the mean. In analyses where segment length is a factor, this makes it hard to differentiate between (for example) a region which contains no reads but is only ten bases long and one which likewise contains no reads but is ten megabases long. If 'zeroML' is FALSE, therefore, the dispersion is set to 1 and the mean estimated as the value that leaves the likelihood of zero data at fifty percent.

If `consensus = TRUE', then a consensus distribution is created and used for each group in the 'cD' object. This allows faster computation of the priors and likelihoods, but with some degradation of accuracy. A 'cluster' object is recommended in order to estimate the priors for the negative binomial distribution. Passing NULL to this variable will cause the function to run in non-parallel mode.

References

Hardcastle T.J., and Kelly, K. baySeq: Empirical Bayesian Methods For Identifying Differential Expression In Sequence Count Data. BMC Bioinformatics (2010)

See Also

countData, getLikelihoods

Examples

# See vignette for more examples.


# If we do not wish to parallelise the functions we set the cluster
# object to NULL.

cl <- NULL

# Alternatively, if we have the 'snow' package installed we
# can parallelise the functions. This will usually (not always) offer
# significant performance gain.

## Not run: try(library(snow))
## Not run: try(cl <- makeCluster(4, "SOCK"))

# load test data
data(simData)

# Create a {countData} object from test data.

replicates <- c("simA", "simA", "simA", "simA", "simA", "simB", "simB", "simB", "simB", "simB")
groups <- list(NDE = c(1,1,1,1,1,1,1,1,1,1), DE = c(1,1,1,1,1,2,2,2,2,2))
CD <- new("countData", data = simData, replicates = replicates, groups = groups)

#estimate library sizes for countData object
libsizes(CD) <- getLibsizes(CD)

# Get priors for negative binomial method
CDPriors <- getPriors.NB(CD, samplesize = 10^5, estimation = "QL", cl = cl)