kCluster: Constructs co-expression dissimilarities from k-means...

In tjh48/clusterSeq: Clustering of high-throughput sequencing data by identifying co-expression patterns

Description

This function aims to find pairwise distances between genes. It does this by constructing k-means clusterings of the observed (log) expression for each gene, and for each pair of genes, finding the maximum value of k for which the centroids of the clusters are monotonic between the genes.

Usage

kCluster(cD, maxK = 100, matrixFile = NULL, replicates =
  NULL, algorithm = "Lloyd", B = 1000, sdm = 1)

Arguments

cD	A countData object containing the raw count data for each gene, or a matrix containing the logged and normalised values for each gene (rows) and sample (columns).
maxK	The maximum value of k for which k-means clustering will be performed. Defaults to 100.
matrixFile	If given, a file to write the complete (gzipped) matrix of pairwise distances between genes. Defaults to NULL.
replicates	If given, a factor or vector that can be cast to a factor that defines the replicate structure of the data. See Details.
algorithm	The algorithm to be used by the kmeans function.
B	Number of iterations of bootstrapping algorithm used to establish clustering validity
sdm	Thresholding parameter for validity; see Details.

Details

In comparing two genes, we find the maximum value of k for which separate k-means clusterings of the two genes lead to a monotonic relationship between the centroids of the clusters. For this value of k, the maximum difference between expression levels observed within a cluster of either gene is reported as a measure of the dissimilarity between the two genes.

There is a potential issue in that for genes non-differentially expressed across all samples (i.e., the appropriate value of k is 1), there will nevertheless exist clusterings for k > 1. For some arrangements of data, this leads to misattribution of non-differentially expressed genes. We identify these cases by adapting Tibshirani's gap statistic; bootstrapping uniformly distributed data on the same range as the observed data, calculating the dissimilarity score as above, and finding those cases for which the gap between the bootstrapped mean dissimilarity and the observed dissimilarity for k = 1 exceeds that for k = 2 by more than some multiple (sdm) of the standard error of the bootstrapped dissimilarities of k = 2. These cases are forced to be treated as non-differentially expressed by discarding all dissimilarity data for k > 1.

If the replicates vector is given, or if the replicates slot of a countData given as the 'cD' variable is complete, then the k-means clustering will be done on the median of the expression values of each replicate group. Dissimilarity calculations will still be made on the full data.

Value

A data.frame which for each gene defines its nearest neighbour of higher row index and the dissimilarity with that neighbour.

Author(s)

Thomas J. Hardcastle

See Also

makeClusters makeClustersFF associatePosteriors

Examples

#Load in the processed data of observed read counts at each gene for each sample. 
data(ratThymus, package = "clusterSeq")

# Library scaling factors are acquired here using the getLibsizes
# function from the baySeq package.
libsizes <- getLibsizes(data = ratThymus)

# Adjust the data to remove zeros and rescale by the library scaling
# factors. Convert to log scale.
ratThymus[ratThymus == 0] <- 1
normRT <- log2(t(t(ratThymus / libsizes)) * mean(libsizes))

# run kCluster on reduced set.
normRT <- normRT[1:1000,]
kClust <- kCluster(normRT)
head(kClust)

# Alternatively, run on a count data object:
# load in analysed countData (baySeq package) object
library(baySeq)
data(cD.ratThymus, package = "clusterSeq")

# estimate likelihoods of dissimilarity on reduced set
kClust2 <- kCluster(cD.ratThymus[1:1000,])
head(kClust2)

tjh48/clusterSeq documentation built on May 31, 2019, 3:40 p.m.