dexus (version 1.12.0)

dexss: Detection of Differential Expression in a semi-supervised Setting

Description

Performs the DEXSS algorithm for detection of differentially expressed genes in RNA-seq data for a semi-supervised setting, i.e. that the condition of some samples is known, and for some samples the condition is unkown.

Usage

dexss(X, nclasses = 2, G = 1, alphaInit, cyc = 20,
    labels, normalization = "RLE", kmeansIter = 10,
    ignoreIfAllCountsSmaller = 1, theta = 2.5, minMu = 0.5,
    rmax = 13, initialization = "kmeans",
    multiclassPhiPoolingFunction = NULL, quiet = FALSE,
    resultObject = "S4")

Arguments

X

either a vector of counts or a raw data matrix, where columns are interpreted as samples and rows as genomic regions. An instance of "countDataSet" is also accepted.

nclasses

The number of conditions, i.e. mixture components. (Default = 2)

G

The weight of the prior distribution of the mixture weights. Not used in the supervised case. (Default = 1).

cyc

Positive integer that sets the number of cycles of the EM algorithm. (Default = 20).

alphaInit

The initial estimates of the condition sizes, i.e., mixture weights. Not used in the supervised case. (Default = c(0.5,0.5)) .

labels

The labels for the classes, will be coerced into an integer. For this semi-supervised version the known labels/conditions must be coded as integers starting with 1. The samples with the label 1 will be considered as being in the "major condition". For the samples with unknown labels/conditions an "NA" must be set.

normalization

method used for normalizing the reads. "RLE" is the method used by (Anders and Huber, 2010), "upperquartile" is the Upper-Quartile method by (Bullard et al., 2010), and none deactivates normalization. (Default = "RLE").

kmeansIter

number of times the K-Means algorithm is run. (Default = 10).

ignoreIfAllCountsSmaller

Ignores transcript for which all read counts are smaller than this value. These transcripts are considered as "not expressed" (Default = 1).

theta

The weight of the prior on the size parameter or inverse dispersion parameter. Theta is adjusted to each transcript by dividing by the mean read count of the transcript. The higher theta, the lower r and the higher the overdispersion will be. (Default = 2.5).

minMu

Minimal mean for all negative binomial distributions. (Default = 0.5).

rmax

Maximal value for the size parameter. The inverse of this parameter is the lower bound on the dispersion. In analogy to (Anders and Huber, 2010) we use 13 as default. (Default = 13).

initialization

Method used to find the initial clusters. Dexus can either use the quantiles of the readcounts of each gene or run k-means on the counts. (Default = "kmeans").

multiclassPhiPoolingFunction

In "multiClass" mode the dispersion is either estimated across all classes at once (NULL), or separately for each condition, i.e., class. The size parameters or dispersion per class are then joined to one estimate by the mean ("mean"), minimum ("min") or maximum ("max"). In our investigations estimation across all classes at once performed best. (Default = NULL).

quiet

Logical that indicates whether dexus should report the steps of the algorithm. Supresses messages from the program if set to TRUE. (Default = FALSE).

resultObject

Type of the result object; can either be a list ("list") or an instance of "DEXUSResult" ("S4"). (Default="S4").

Value

• "list" or "DEXUSResult". A list containing the results and the parameters of the algorithm or an instance of "DEXUSResult".

Details

The read count $x$ is explained by a finite mixture of negative binomials:

$$p(x) = \sum_{i=1} ^n \alpha_i\ \mathrm{NB}(x;\ \mu_i, r_i),$$

where $\alpha_i$ is the weight of the mixture component, $\mathrm{NB}$ is the negative binomial with mean parameter $\mu_i$ and size parameter $r_i$. The parameters are selected by an EM algorithm in a Baysian framework.

Each component in the mixture model corresponds to one condition.

• If the groups, conditions, replicate status, or labels are unknown, DEXUS tries to estimate these conditions. For each transcript DEXUS tries to explain the read counts by one negative binomial distribution. If this is possible, the transcript is explained by one condition and therefore it is not differentially expressed. If more than one negative binomial distribution is needed to explain the read counts of a transcript, this transcript indicates that it is differentially expressed. Evidence for differential expression is strong if a large amount of samples participate in each condition and the mean expression values are well separated. Both of these criteria are measured by the informative/non-informative (I/NI) call.
• If there are more than two groups given by the vectorlabels, DEXUS uses a generalized linear model to explain the data in analogy to (McCarthy, 2012).
• If there are two groups given by the vectorlabels, DEXUS uses the exact test for count data to test between the sample groups, as implemented by (Anders and Huber, 2010) in the package "DESeq".

References

Anders, S. and Huber, W. (2010). Differential expression analysis for sequence count data. Genome Biol, 11(10), R106.

Bullard, J. H., Purdom, E., Hansen, K. D., and Dudoit, S. (2010). Evaluation of statistical methods for normalization and differential expression in mRNA-seq experiments. BMC Bioinformatics, 11, 94.

McCarthy, D. J., Chen, Y., and Smyth, G. K. (2012). Differential expression analysis of multifactor RNA-Seq experiments with respect to biological variation. Nucleic Acids Res, 40(10), 4288-4297.

Examples

data(dexus)
labels1 <- substr(colnames(countsBottomly),1,2)
labels2 <- c()
labels2[which(labels1=="D2")] <- 1
labels2[which(labels1=="B6")] <- 2
labels2[c(3,7,8,10,12,15)] <- NA
res <- dexss(countsBottomly[1:100, ],labels=labels2,nclasses=2,G=0)