Perform network construction and consensus module detection across several datasets.
blockwiseConsensusModules(
multiExpr, # Data checking options
checkMissingData = TRUE,
# Blocking options
blocks = NULL,
maxBlockSize = 5000,
blockSizePenaltyPower = 5,
nPreclusteringCenters = NULL,
randomSeed = 54321,
# TOM precalculation arguments, if available
individualTOMInfo = NULL,
useIndivTOMSubset = NULL,
# Network construction arguments: correlation options
corType = "pearson",
maxPOutliers = 1,
quickCor = 0,
pearsonFallback = "individual",
cosineCorrelation = FALSE,
# Adjacency function options
power = 6,
networkType = "unsigned",
checkPower = TRUE,
replaceMissingAdjacencies = FALSE,
# Topological overlap options
TOMType = "unsigned",
TOMDenom = "min",
suppressNegativeTOM = FALSE,
# Save individual TOMs?
saveIndividualTOMs = TRUE,
individualTOMFileNames = "individualTOM-Set%s-Block%b.RData",
# Consensus calculation options: network calibration
networkCalibration = c("single quantile", "full quantile", "none"),
# Simple quantile calibration options
calibrationQuantile = 0.95,
sampleForCalibration = TRUE, sampleForCalibrationFactor = 1000,
getNetworkCalibrationSamples = FALSE,
# Consensus definition
consensusQuantile = 0,
useMean = FALSE,
setWeights = NULL,
# Saving the consensus TOM
saveConsensusTOMs = FALSE,
consensusTOMFilePattern = "consensusTOM-block.%b.RData",
# Internal handling of TOMs
useDiskCache = TRUE, chunkSize = NULL,
cacheBase = ".blockConsModsCache",
cacheDir = ".",
# Alternative consensus TOM input from a previous calculation
consensusTOMInfo = NULL,
# Basic tree cut options
# Basic tree cut options
deepSplit = 2,
detectCutHeight = 0.995, minModuleSize = 20,
checkMinModuleSize = TRUE,
# Advanced tree cut opyions
maxCoreScatter = NULL, minGap = NULL,
maxAbsCoreScatter = NULL, minAbsGap = NULL,
minSplitHeight = NULL, minAbsSplitHeight = NULL,
useBranchEigennodeDissim = FALSE,
minBranchEigennodeDissim = mergeCutHeight,
stabilityLabels = NULL,
minStabilityDissim = NULL,
pamStage = TRUE, pamRespectsDendro = TRUE,
# Gene reassignment and trimming from a module, and module "significance" criteria
reassignThresholdPS = 1e-4,
trimmingConsensusQuantile = consensusQuantile,
minCoreKME = 0.5, minCoreKMESize = minModuleSize/3,
minKMEtoStay = 0.2,
# Module eigengene calculation options
impute = TRUE,
trapErrors = FALSE,
#Module merging options
equalizeQuantilesForModuleMerging = FALSE,
quantileSummaryForModuleMerging = "mean",
mergeCutHeight = 0.15,
mergeConsensusQuantile = consensusQuantile,
# Output options
numericLabels = FALSE,
# General options
nThreads = 0,
verbose = 2, indent = 0, ...)
A list with the following components:
module assignment of all input genes. A vector containing either character strings with
module colors (if input numericLabels was unset) or numeric module labels (if numericLabels
was set to TRUE). The color "grey" and the numeric label 0 are reserved for unassigned genes.
module colors or numeric labels before the module merging step.
module eigengenes corresponding to the modules returned in colors, in multi-set
format. A vector of lists, one per set, containing eigengenes, proportion of variance explained and other
information. See multiSetMEs for a detailed description.
a list, with one component per input set. Each component is a logical vector with one entry per sample from the corresponding set. The entry indicates whether the sample in the set passed basic quality control criteria.
a logical vector with one entry per input gene indicating whether the gene passed basic quality control criteria in all sets.
a list with one component for each block of genes. Each component is the hierarchical clustering dendrogram obtained by clustering the consensus gene dissimilarity in the corresponding block.
if saveConsensusTOMs==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 with one component for each block of genes. Each component is a vector giving
the indices (relative to the input multiExpr) of genes in the corresponding 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.
A vector of length nSets that
contains, for each set, the number of (calibrated) elements that were less than or equal the consensus for that element.
if the input getNetworkCalibrationSamples is TRUE, this component is a
list with one component per block. Each component is again a list with two components:
sampleIndex contains indices of the distance structure in which TOM is stored that were sampled,
and TOMSamples is a matrix whose rows correspond to TOM samples and columns to individual set.
Hence, networkCalibrationSamples[[blockNo]]$TOMSamples[index, setNo] contains the TOM entry that
corresponds to element networkCalibrationSamples[[blockNo]]$sampleIndex[index] of the TOM distance
structure in block blockNo and set setNo. (For details on the distance structure, see
dist.)
expression data in the multi-set format (see checkSets). A vector of
lists, one per set. Each set must contain a component data that contains the expression data, with
rows corresponding to samples and columns to genes or probes.
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 gene
of multiExpr 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 to be used in the preclustering. Defaults to smaller of
nGenes/20 and 100*nGenes/maxBlockSize, where nGenes is the nunber of genes (variables) in
multiExpr.
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.
Optional data for TOM matrices in individual data sets. This object is returned by
the function blockwiseIndividualTOMs. If not given, appropriate topological overlaps will be
calculated using the network contruction options below.
If individualTOMInfo is given, this argument allows to only select a subset
of the individual set networks contained in individualTOMInfo. It should be a numeric vector giving the
indices of the individual sets to be used. Note that this argument is NOT applied to multiExpr.
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 pariwise.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. Either a single number or a vector of the same length as the number of sets, with one power for each set.
network type. Allowed values are (unique abbreviations of) "unsigned",
"signed", "signed hybrid". See adjacency.
logical: should basic sanity check be performed on the supplied power? If
you would like to experiment with unusual powers, set the argument to FALSE and proceed with
caution.
logical: should missing values in the calculation of adjacency be replaced by 0?
one of "none", "unsigned", "signed", "signed Nowick",
"unsigned 2", "signed 2" and "signed Nowick 2". If "none", adjacency
will be used for clustering. See TOMsimilarityFromExpr for details.
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.
Logical: should the result be set to zero when negative? Negative TOM values can occur when
TOMType is "signed Nowick".
logical: should individual TOMs be saved to disk for later use?
character string giving the file names to save individual TOMs into. The
following tags should be used to make the file names unique for each set and block: %s will be
replaced by the set number; %N will be replaced by the set name (taken from names(multiExpr))
if it exists, otherwise by set number; %b will be replaced by the block number. If the file names turn
out to be non-unique, an error will be generated.
network calibration method. One of "single quantile", "full quantile", "none" (or a unique abbreviation of one of them).
if networkCalibration is "single quantile",
topological overlaps (or adjacencies if
TOMs are not computed) will be scaled such that their calibrationQuantile quantiles will agree.
if TRUE, calibration quantiles will be determined from a sample of network
similarities. Note that using all data can double the memory footprint of the function and the function
may fail.
determines the number of samples for calibration: the number is
1/calibrationQuantile * sampleForCalibrationFactor. Should be set well above 1 to ensure accuracy of the
sampled quantile.
logical: should samples used for TOM calibration be saved for future analysis?
This option is only available when sampleForCalibration is TRUE.
quantile at which consensus is to be defined. See details.
logical: should the consensus be determined from a (possibly weighted) mean across the data sets rather than a quantile?
Optional vector (one component per input set) of weights to be used for weighted mean
consensus. Only used when useMean above is TRUE.
logical: should the consensus topological overlap matrices for each block be saved and returned?
character string containing the file namefiles containing the
consensus topological overlaps. The tag %b will be replaced by the block number. If the resulting file
names are non-unique (for example, because the user gives a file name without a %b tag), an error
will be generated.
These files are standard R data files and can be loaded using the load
function.
should calculated network similarities in individual sets be temporarilly saved to disk? Saving to disk is somewhat slower than keeping all data in memory, but for large blocks and/or many sets the memory footprint may be too big.
network similarities are saved in smaller chunks of size chunkSize.
character string containing the desired name for the cache files. The actual file
names will consists of cacheBase and a suffix to make the file names unique.
character string containing the desired path for the cache files.
optional list summarizing consensus TOM, output of consensusTOM. It
contains information about pre-calculated consensus TOM. Supplying this argument replaces TOM calculation,
so none of the individual or consensus TOM calculation arguments are taken into account.
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.
logical: should sanity checks be performed on minModuleSize?
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.
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.
per-set p-value ratio threshold for reassigning genes between modules. See Details.
a number between 0 and 1 specifying the consensus quantile used for kME calculation that determines module trimming according to the arguments below.
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.
logical: should imputation be used for module eigengene calculation? See
moduleEigengenes for more details.
logical: should errors in calculations be trapped?
Logical: equalize quantiles of the module eigengene networks
before module merging? If TRUE, the quantiles of the eigengene correlation matrices (interpreted as a
single vectors of non-redundant components) will be equalized across the input data sets. Note that
although this seems like a reasonable option, it should be considered experimental and not necessarily
recommended.
One of "mean" or "median".
If quantile equalization of the module eigengene networks is
performed, the resulting "normal" quantiles will be given by this function of the corresponding quantiles
across the input data sets.
dendrogram cut height for module merging.
consensus quantile for module merging. See mergeCloseModules for
details.
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. At present these can include reproduceBranchEigennodeQuantileError that
instructs the function to reproduce a bug in branch eigennode dissimilarity calculations for purposes if
reproducing old reults.
Peter Langfelder
The function starts by optionally filtering out samples that have too many missing entries and genes
that have either too many missing entries or zero variance in at least one set. Genes that are filtered
out are left unassigned by the module detection. Returned eigengenes will contain NA in entries
corresponding to filtered-out samples.
If blocks is not given and
the number of genes exceeds maxBlockSize, genes are pre-clustered into blocks using the function
consensusProjectiveKMeans; 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 in each set. To minimize memory usage, calculated topological overlaps are optionally saved to disk in chunks until they are needed again for the calculation of the consensus network topological overlap.
Before calculation of the consensus Topological Overlap, individual TOMs are optionally calibrated. Calibration methods include single quantile scaling and full quantile normalization.
Single quantile
scaling raises individual TOM in sets 2,3,... to a power such that the quantiles given by
calibrationQuantile agree with the quantile in set 1. Since the high TOMs are usually the most important
for module identification, the value of calibrationQuantile is close to (but not equal) 1. To speed up
quantile calculation, the quantiles can be determined on a randomly-chosen component subset of the TOM matrices.
Full quantile normalization, implemented in normalize.quantiles, adjusts the
TOM matrices such that all quantiles equal each other (and equal to the quantiles of the component-wise
average of the individual TOM matrices).
Note that network calibration is performed separately in each block, i.e., the normalizing transformation may differ between blocks. This is necessary to avoid manipulating a full TOM in memory.
The consensus TOM is calculated as the component-wise consensusQuantile quantile of the individual
(set) TOMs; that is, for each gene pair (TOM entry), the consensusQuantile quantile across all input
sets. Alternatively, one can also use (weighted) component-wise mean across all imput data sets.
If requested, the consensus topological overlaps are saved to disk for later use.
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
consensus module membership kME (that is, correlation with module eigengene)
is less than minKMEtoStay.
Modules in which
fewer than minCoreKMESize genes have consensus 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 (in every set),
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.
Langfelder P, Horvath S (2007) Eigengene networks for studying the relationships between co-expression modules. BMC Systems Biology 2007, 1:54
goodSamplesGenesMS 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.