This function performs automatic network construction and module detection on large expression datasets in a block-wise manner.
blockwiseModules(
# Input data datExpr,
# Data checking options
checkMissingData = TRUE,
# Options for splitting data into blocks
blocks = NULL,
maxBlockSize = 5000,
blockSizePenaltyPower = 5,
nPreclusteringCenters = as.integer(min(ncol(datExpr)/20,
100*ncol(datExpr)/maxBlockSize)),
randomSeed = 12345,
# load TOM from previously saved file?
loadTOM = FALSE,
# Network construction arguments: correlation options
corType = "pearson",
maxPOutliers = 1,
quickCor = 0,
pearsonFallback = "individual",
cosineCorrelation = FALSE,
# Adjacency function options
power = 6,
networkType = "unsigned",
replaceMissingAdjacencies = FALSE,
# Topological overlap options
TOMType = "signed",
TOMDenom = "min",
# Saving or returning TOM
getTOMs = NULL,
saveTOMs = FALSE,
saveTOMFileBase = "blockwiseTOM",
# Basic tree cut options
deepSplit = 2,
detectCutHeight = 0.995,
minModuleSize = min(20, ncol(datExpr)/2 ),
# Advanced tree cut options
maxCoreScatter = NULL, minGap = NULL,
maxAbsCoreScatter = NULL, minAbsGap = NULL,
minSplitHeight = NULL, minAbsSplitHeight = NULL,
useBranchEigennodeDissim = FALSE,
minBranchEigennodeDissim = mergeCutHeight,
stabilityLabels = NULL,
stabilityCriterion = c("Individual fraction", "Common fraction"),
minStabilityDissim = NULL,
pamStage = TRUE, pamRespectsDendro = TRUE,
# Gene reassignment, module trimming, and module "significance" criteria
reassignThreshold = 1e-6,
minCoreKME = 0.5,
minCoreKMESize = minModuleSize/3,
minKMEtoStay = 0.3,
# Module merging options
mergeCutHeight = 0.15,
impute = TRUE,
trapErrors = FALSE,
# Output options
numericLabels = FALSE,
# Options controlling behaviour
nThreads = 0,
verbose = 0, indent = 0,
...)
Expression data. A matrix (preferred) or
data frame in which columns are genes and rows ar samples. NAs are
allowed, but not too many. See checkMissingData below and details.
logical: should data be checked for excessive numbers of missing entries in genes and samples, and for genes with zero variance? See details.
optional specification of blocks in which hierarchical clustering and module detection
should be performed. If given, must be a numeric vector with one entry per column (gene)
of exprData giving the number of the block to which the corresponding gene belongs.
integer giving maximum block size for module detection. Ignored if blocks
above is non-NULL. Otherwise, if the number of genes in datExpr exceeds maxBlockSize, genes
will be pre-clustered into blocks whose size should not exceed maxBlockSize.
number specifying how strongly blocks should be penalized for exceeding the
maximum size. Set to a lrge number or Inf if not exceeding maximum block size is very important.
number of centers for pre-clustering. Larger numbers typically results in better but slower pre-clustering.
integer to be used as seed for the random number generator before the function
starts. If a current seed exists, it is saved and restored upon exit. If NULL is given, the
function will not save and restore the seed.
logical: should Topological Overlap Matrices be loaded from previously saved files (TRUE)
or calculated (FALSE)? It may be useful to load previously saved TOM matrices if these have been
calculated previously, since TOM calculation is often the most computationally expensive part of network
construction and module identification. See saveTOMs and saveTOMFileBase below for when and how TOM
files are saved, and what the file names are. If loadTOM is TRUE but the files cannot be
found, or do not contain the correct TOM data, TOM will be recalculated.
character string specifying the correlation to be used. Allowed values are (unique
abbreviations of) "pearson" and "bicor", corresponding to Pearson and bidweight
midcorrelation, respectively. Missing values are handled using the pairwise.complete.obs option.
only used for corType=="bicor". Specifies the maximum percentile of data
that can be considered outliers on either
side of the median separately. For each side of the median, if
higher percentile than maxPOutliers is considered an outlier by the weight function based on
9*mad(x), the width of the weight function is increased such that the percentile of outliers on
that side of the median equals maxPOutliers. Using maxPOutliers=1 will effectively disable
all weight function broadening; using maxPOutliers=0 will give results that are quite similar (but
not equal to) Pearson correlation.
real number between 0 and 1 that controls the handling of missing data in the calculation of correlations. See details.
Specifies whether the bicor calculation, if used, should revert to Pearson when
median absolute deviation (mad) is zero. Recongnized values are (abbreviations of)
"none", "individual", "all". If set to
"none", zero mad will result in NA for the corresponding correlation.
If set to "individual", Pearson calculation will be used only for columns that have zero mad.
If set to "all", the presence of a single zero mad will cause the whole variable to be treated in
Pearson correlation manner (as if the corresponding robust option was set to FALSE). Has no
effect for Pearson correlation. See bicor.
logical: should the cosine version of the correlation calculation be used? The cosine calculation differs from the standard one in that it does not subtract the mean.
soft-thresholding power for network construction.
network type. Allowed values are (unique abbreviations of) "unsigned",
"signed", "signed hybrid". See adjacency.
logical: should missing values in the calculation of adjacency be replaced by 0?
one of "none", "unsigned", "signed". If "none", adjacency
will be used for clustering. If "unsigned", the standard TOM will be used (more generally, TOM
function will receive the adjacency as input). If "signed", TOM will keep track of the sign of
correlations between neighbors.
a character string specifying the TOM variant to be used. Recognized values are
"min" giving the standard TOM described in Zhang and Horvath (2005), and "mean" in which
the min function in the denominator is replaced by mean. The "mean" may produce
better results but at this time should be considered experimental.
deprecated, please use saveTOMs below.
logical: should the consensus topological overlap matrices for each block be saved and returned?
character string containing the file name base for files containing the
consensus topological overlaps. The full file names have "block.1.RData", "block.2.RData"
etc. appended. These files are standard R data files and can be loaded using the load
function.
integer value between 0 and 4. Provides a simplified control over how sensitive
module detection should be to module splitting, with 0 least and 4 most sensitive. See
cutreeDynamic for more details.
dendrogram cut height for module detection. See
cutreeDynamic for more details.
minimum module size for module detection. See
cutreeDynamic for more details.
maximum scatter of the core for a branch to be a cluster, given as the fraction
of cutHeight relative to the 5th percentile of joining heights. See
cutreeDynamic for more details.
minimum cluster gap given as the fraction of the difference between cutHeight and
the 5th percentile of joining heights. See cutreeDynamic for more details.
maximum scatter of the core for a branch to be a cluster given as absolute
heights. If given, overrides maxCoreScatter. See cutreeDynamic for more details.
minimum cluster gap given as absolute height difference. If given, overrides
minGap. See cutreeDynamic for more details.
Minimum split height given as the fraction of the difference between
cutHeight and the 5th percentile of joining heights. Branches merging below this height will
automatically be merged. Defaults to zero but is used only if minAbsSplitHeight below is
NULL.
Minimum split height given as an absolute height.
Branches merging below this height will automatically be merged. If not given (default), will be determined
from minSplitHeight above.
Logical: should branch eigennode (eigengene) dissimilarity be considered when merging branches in Dynamic Tree Cut?
Minimum consensus branch eigennode (eigengene) dissimilarity for
branches to be considerd separate. The branch eigennode dissimilarity in individual sets
is simly 1-correlation of the
eigennodes; the consensus is defined as quantile with probability consensusQuantile.
Optional matrix of cluster labels that are to be used for calculating branch
dissimilarity based on split stability. The number of rows must equal the number of genes in
multiExpr; the number of columns (clusterings) is arbitrary. See
branchSplitFromStabilityLabels for details.
One of c("Individual fraction", "Common fraction"), indicating which method
for assessing stability similarity of two branches should be used. We recommend "Individual fraction"
which appears to perform better; the "Common fraction" method is provided for backward compatibility
since it was the (only) method available prior to WGCNA version 1.60.
Minimum stability dissimilarity criterion for two branches to be considered
separate. Should be a number between 0 (essentially no dissimilarity required) and 1 (perfect dissimilarity
or distinguishability based on stabilityLabels). See
branchSplitFromStabilityLabels for details.
logical. If TRUE, the second (PAM-like) stage of module detection will be performed.
See cutreeDynamic for more details.
Logical, only used when pamStage is TRUE.
If TRUE, the PAM stage will
respect the dendrogram in the sense an object can be PAM-assigned only to clusters that lie below it on
the branch that the object is merged into.
See cutreeDynamic for more details.
a number between 0 and 1. If a detected module does not have at least
minModuleKMESize genes with eigengene connectivity at least minCoreKME, the module is
disbanded (its genes are unlabeled and returned to the pool of genes waiting for mofule detection).
see minCoreKME above.
genes whose eigengene connectivity to their module eigengene is lower than
minKMEtoStay are removed from the module.
p-value ratio threshold for reassigning genes between modules. See Details.
dendrogram cut height for module merging.
logical: should imputation be used for module eigengene calculation? See
moduleEigengenes for more details.
logical: should errors in calculations be trapped?
logical: should the returned modules be labeled by colors (FALSE), or by
numbers (TRUE)?
non-negative integer specifying the number of parallel threads to be used by certain parts of correlation calculations. This option only has an effect on systems on which a POSIX thread library is available (which currently includes Linux and Mac OSX, but excludes Windows). If zero, the number of online processors will be used if it can be determined dynamically, otherwise correlation calculations will use 2 threads.
integer level of verbosity. Zero means silent, higher values make the output progressively more and more verbose.
indentation for diagnostic messages. Zero means no indentation, each unit adds two spaces.
Other arguments.
A list with the following components:
a vector of color or numeric module labels for all genes.
a vector of color or numeric module labels for all genes before module merging.
a data frame containing module eigengenes of the found modules (given by colors).
numeric vector giving indices of good samples, that is samples that do not have too many missing entries.
numeric vector giving indices of good genes, that is genes that do not have too many missing entries.
a list whose components conatain hierarchical clustering dendrograms of genes in each block.
if saveTOMs==TRUE,
a vector of character strings, one string per block, giving the file names of files
(relative to current directory) in which blockwise topological overlaps were saved.
a list whose components give the indices of genes in each block.
if input blocks was given, its copy; otherwise a vector of length equal number of
genes giving the block label for each gene. Note that block labels are not necessarilly sorted in the
order in which the blocks were processed (since we do not require this for the input blocks). See
blockOrder below.
a vector giving the order in which blocks were processed and in which
blockGenes above is returned. For example, blockOrder[1] contains the label of the
first-processed block.
logical indicating whether the module eigengenes were calculated without errors.
Before module detection starts, genes and samples are optionally checked for the presence of NAs.
Genes and/or samples that have too many NAs are flagged as bad and removed from the analysis; bad
genes will be automatically labeled as unassigned, while the returned eigengenes will have NA
entries for all bad samples.
If blocks is not given and
the number of genes exceeds maxBlockSize, genes are pre-clustered into blocks using the function
projectiveKMeans; otherwise all genes are treated in a single block.
For each block of genes, the network is constructed and (if requested) topological overlap is calculated.
If requested, the topological overlaps are returned as part of the return value list.
Genes are then clustered using average linkage hierarchical clustering and modules are identified in the
resulting dendrogram by the Dynamic Hybrid tree cut. Found modules are trimmed of genes whose
correlation with module eigengene (KME) is less than minKMEtoStay. Modules in which
fewer than minCoreKMESize genes have KME higher than minCoreKME
are disbanded, i.e., their constituent genes are pronounced
unassigned.
After all blocks have been processed, the function checks whether there are genes whose KME in the module
they assigned is lower than KME to another module. If p-values of the higher correlations are smaller
than those of the native module by the factor reassignThresholdPS,
the gene is re-assigned to the closer module.
In the last step, modules whose eigengenes are highly correlated are merged. This is achieved by
clustering module eigengenes using the dissimilarity given by one minus their correlation,
cutting the dendrogram at the height mergeCutHeight and merging all modules on each branch. The
process is iterated until no modules are merged. See mergeCloseModules for more details on
module merging.
The argument quick specifies the precision of handling of missing data in the correlation
calculations. Zero will cause all
calculations to be executed precisely, which may be significantly slower than calculations without
missing data. Progressively higher values will speed up the
calculations but introduce progressively larger errors. Without missing data, all column means and
variances can be pre-calculated before the covariances are calculated. When missing data are present,
exact calculations require the column means and variances to be calculated for each covariance. The
approximate calculation uses the pre-calculated mean and variance and simply ignores missing data in the
covariance calculation. If the number of missing data is high, the pre-calculated means and variances may
be very different from the actual ones, thus potentially introducing large errors.
The quick value times the
number of rows specifies the maximum difference in the
number of missing entries for mean and variance calculations on the one hand and covariance on the other
hand that will be tolerated before a recalculation is triggered. The hope is that if only a few missing
data are treated approximately, the error introduced will be small but the potential speedup can be
significant.
Bin Zhang and Steve Horvath (2005) "A General Framework for Weighted Gene Co-Expression Network Analysis", Statistical Applications in Genetics and Molecular Biology: Vol. 4: No. 1, Article 17
goodSamplesGenes for basic quality control and filtering;
adjacency, TOMsimilarity for network construction;
hclust for hierarchical clustering;
cutreeDynamic for adaptive branch cutting in hierarchical clustering
dendrograms;
mergeCloseModules for merging of close modules.