iterativeWGCNA: a WGCNA extension

New Release Available

• iterativeWGCNA 1.1.6 now available

  • bug fix: saveTOMs disabled by default but can be enabled by with the option --wgcnaParameters "saveTOMs=TRUE"
  • bug fix: issue parsing boolean WGCNA parameters (e.g. saveTOMs=FALSE or cosineCorrelation=TRUE) resolved
  • new parameter added: --gzipTOMs which will gzip TOM .RData files as generated to save space
  • new dist available on PyPI

• iterativeWGCNA 1.1.3 now available

  • added script to adjust final module merge
    • see Add-ons and updated Output Files for more information
  • fixed Python 3.3+ bug with converting odict_values to ro.StrVector
  • added --debug option; currently only prints extensive debugging statements for module merge stage

Synopsis

iterativeWGCNA provides a Python-wrapped extension for the R program Weighted Gene Correlation Network Analysis (WGCNA) that improves the robustness of network-based classifications (modules) inferred from whole-transcriptome gene expression datasets.

How to Cite

When citing iterativeWGCNA, please use:

Greenfest-Allen et. al 2017. iterativeWGCNA: iterative refinement to improve module detection from WGCNA co-expression networks. bioRxiv doi:10.1101/234062

Contents

Setup and Installation

• Dependencies
• Installation

Usage

• Running iterativeWGCNA
• Add-ons

Troubleshooting

• libreadline.so.6: undefined symbol
• Cannot install rpy2 on OSX
• Segmentation Faults, missing C libs, etc

Setup and Installation

Dependencies

iterativeWGCNA has the following dependencies:

R language for statistical computing

R version 3.* must be available on the system and the binary executable in the system PATH.

NOTE: the most recent version of R that supports WGCNA is 3.3.x

iterativeWGCNA requires that the following R packages be installed:

• WGCNA: Weighted Gene Co-expression Network Analysis package and Bioconductor dependencies

Python

iterativeWGCNA requires Python version 2.7 or higher. It is designed to be future compatible with Python 3+. iterativeWGCNA requires the following Python packages:

• rpy2: a Python interface for R (v. 2.7.9+)
• matplotlib

NOTE: the most recent version of rpy2 requires python 3.x

If missing, rpy2 will be installed by the iterativeWGCNA installer. See below.

Installation

iterativeWGCNA can be run without installing the package as long as the requisite Python dependencies (rpy2) and R are already present on the system. Installing the package will install any missing Python dependencies.

iterativeWGCNA is reposited in the Python Package Index (PyPI) and can be installed via pip or easy_install.

pip install iterativeWGCNA

This package is tied to the tagged releases on GitHub.

To install the iterativeWGCNA package from the git master, clone and then run the python setup.py script as folows:

git clone https://github.com/cstoeckert/iterativeWGCNA.git
cd iterativeWGCNA
python setup.py install

NOTE: depending on your system this may require administrative (e.g., sudo) permissions.

As a work around, specify the --user switch to install iterativeWGCNA and its dependencies to a local (user) library (e.g., .local/bin on a Linux system) as follows:

git clone https://github.com/cstoeckert/iterativeWGCNA.git
cd iterativeWGCNA
python setup.py install --user

Alternatively, you can also use pip to install from the git master:

pip install git+git://github.com/cstoeckert/iterativeWGCNA.git

Usage

Running iterativeWGCNA

1. Quick Start
2. Command Line Options
3. WGCNA Parameters
4. Input File Format
5. Output Files

Quick Start

If installed via the pip or easy_install, iterativeWGCNA can be run using the iterativeWGCNA command. At minimum, the -i option (--inputFile) denoting the full path to the input file must be specified:

iterativeWGCNA -i <input_file_path>

iterativeWGCNA can also be run without installing the iterativeWGCNA package by executing the wrapper script run_iterative_wgcna.py in the iterativeWGCNA directory. At a minimum, the -i option (--inputFile) denoting the full path to the input file must be specified:

python run_iterative_wgcna.py -i <input_file_path> 

if the iterativeWGCNA package was installed, iterativeWGCNA can also be run at the package level using the -m switch:

python -m iterativeWGCNA -i <input_file_path> 

Command Line Options

Execute run_iterative_wgcna.py with the -h (--help) switch to see all command line options and additional usage information, including details on file formats.

python run_iterative_wgcna.py -h

-h, --help
   show help message and exit
    
-i <gene expression file>, --inputFile <gene expression file>
   full path to input gene expression file; if full path is not provided,
   assumes the file is in the working (output) directory
+ required
   
-o <output dir>, --workingDir <output dir>
   R working directory; where output will be saved
   default: current directory
   
-v, --verbose
   print status messages

-p <param list>, --wgcnaParameters <param list>
   comma separated list of parameters to be passed to WGCNA's blockwiseModules function
   e.g., power=6,randomSeed=1234875
   see 'blockwiseModules' section of the WGCNA manual for more information
   
--enableWGCNAThreads
    enable WGCNA to use threads
    
--skipSaveBlocks
    do not save WGCNA blockwise modules for each iteration
	also will not save TOMs
	NOTE: blocks are necessary to generate summary graphics
	
--gzipTOMs
    if the WGCNA parameter saveTOMs is set to TRUE, this will
	gzip the TOM .RData files
	NOTE: R is not able to read the .RData.gz files; uncompress
	first
	
-f, --finalMergeCutHeight <cut height>
	cut height (max dissimilarity) for final module merge
	(after algorithm convergence); [0, 1.0], default=0.05
    

WGCNA Parameters

iterativeWGCNA can accept any parameter valid for the WGCNA blockwiseModules function. See https://www.rdocumentation.org/packages/WGCNA/versions/1.41-1/topics/blockwiseModules for full details

To specify these parameters use the --wgcnaParameters flag followed by a comma separated list of parameter=value pairs.

For example:

--wgcnaParameters maxBlockSize=5000,corType=bicor,power=10

sets the maximum block size to 5000 genes, the correlation type to the biweight correlation, and the power-law scaling factor (beta) to 10

WGCNA's blockwiseModules function partitions the gene set into a set of blocks each containing at most maxBlockSize genes.

To run iterativeWGCNA in a single block, set maxBlockSize to a value > than the number of genes in your geneset.

NOTE: for large datasets (>10,000 genes or probes), adjacency and TOM matrix calculations done in a single block may fail due to memory allocation issues

see the WGCNA large dataset tutorial, section 2.c.2 for more details

If WGCNA parameters are not specified, iterativeWGCNA uses the default WGCNA settings for the blockwiseModules function, except for the following:

minModuleSize = 20 # minimum number of genes in a detected module
saveTOMs = FALSE # save the topological overlap matrices for each block in the block data structure
minKMEtoStay = 0.8 # minimum eigengene connectivity (kME) required for a gene to be retained in its assigned module
minCoreKME = 0.8 # if the module does not have minModuleSize genes with eigengene connectivity at least minCoreKME, the module is disbanded
reassignThreshold = 0.05 # per-set p-value ratio threshold for reassigning genes between modules
networkType = 'signed' # character string specifying network type. Allowed values are "unsigned", "signed", and "signed hybrid"
numericLabels = TRUE # label modules by numbers (e.g., 0,1,2) instead of colors

Input File Format

iterativeWGCNA expects a tab-delimited text file containing gene expression data arranged such that there is one row per gene and one column per sample. The first column should contain unique gene identifiers. For example:

Gene	Sample1	Sample2	Sample3
Gata1	500	715	1000
Phtf2	60	1000	1600

NOTE: We recommend using numeric gene identifiers to uniquely label genes in the input file as R will do some character substitutions (e.g., '.' for '-') and concatenate 'X' to gene symbols starting with a number, leading to erroneous mapping between data frames and potential loss of data.

iterativeWGCNA will accept gzipped input files.

Output Files

An iteration of iterativeWGCNA comprises one run of blockwiseWGCNA followed by an eigengene-connectivity (kME) goodness of fit assessment. A pass of iterativeWGCNA comprises multiple iterations applied to an expression dataset until no more residuals to the kME-fit are found. A new pass is initiated by creating a new expression dataset from all residuals to the kME-fit found during the previous pass.

Modules are uniquely identified by the numerical assignment and the iteration in which they were first detected: e.g., P1_I2_M1 is module 1, detected in the second iteration of the first pass. Unclassified genes are labeled UNCLASSIFIED with a kME of NA.

Results from each pass and iteration are saved in a series of directories, labeled as:

passM: results from the numbered (M) pass iN: results from the numbered (N) iteration

The directory structure and output files are as follows:

├── output_directory
│   ├── iterativeWGCNA.log: main log file for the iterativeWGCNA run
│   ├── iterativeWGCNA-R.log: log file for R; catches R errors and R warning messages
│   ├── gene-counts.txt: tally of number of genes fit and residual to the fit with each iteration
│   ├── final-eigengenes.txt: eigengenes for final modules after final network assembly (before merge)
│   ├── final-kme-histogram.pdf: histogram of eigengene connectivities (kME) in the final classification (before merge)
│   ├── final-membership.txt: gene-module assignments and kME after final iteration (before merge)
│   ├── merge-<finalMergeCutHeight>-eigengenes.txt: recalculated eigengenes for modules retained after merging close modules
│   ├── merge-<finalMergeCutHeight>-kme-histogram.pdf: histogram of eigengene connectivities (kME) after merging close modules
│   ├── merge-<finalMergeCutHeight>-membership.txt: gene-module assignments and kME after merging close modules
│   ├── passM
│   │   ├── initial-pass-expression-set.txt: pass input
│   │   ├── kme_histogram.pdf: histogram of eigengene connectivities for genes classified during pass
│   │   ├── membership.txt: gene-module assignments and kME for genes classfied during pass
│   │   ├── iN
│   │   │   ├── eigengenes.txt: eigengenes of modules detected during the iteration
│   │   │   ├── kme_histogram.pdf: kME histogram after pruning of WGCNA result based on kME
│   │   │   ├── membership.txt: gene membership after kME-based goodness of fit (Pruning)
|   │   │   ├── summary.txt: summaries pass (number genes input, classfied, residual, and number of detected modules)
│   │   |   ├── wgcna-blocks.RData: R data object containing input expression data (expression) and results from blockwise WGCNA (blocks)
│   │   │   ├── wgcna-kme_histogram.pdf: kME histogram based on WGCNA classification
│   │   │   ├── wgcna-membership.txt: gene membership from WGCNA classification
│   │   │   ├── passM_iN-TOM.block.X.RData(.gz): TOM for block X generated in passM, iN (if saveTOMs=TRUE; gzipped if --gzipTOMs option specified)

Note: as of release 1.1.3, iterativeWGCNA now outputs two sets of files containing the final classification. Those prefixed with final- report the penultimate module membership assignments and eigengenes; i.e. result at the algorithm convergence. Those prefixed with merge- report the final module assignements determined after merging close modules and reassessing module memberships after the merge.

Note: TOMs are only saved if the wgcnaParameter saveTOMs is set to TRUE. With large gene sets (>10,000 genes), these can be very large and take a while to write to file, dramatically slowing down the performace of the algorithm in the early iterations. To save disk space, specify the paratmer --gzipTOMs to gzip .RData files as generated. Again, this i/o operation may slow down the performance of the algorithm in the early iterations.

Add-ons

1. Merge Close Modules

Merge Close Modules

Script for running the final-module merge. Allows users to choose a different merge-threshold without having to rerun the entire iterativeWGCNA classification.

The merge script depends on the following options:

-i <gene expression file>, --inputFile <gene expression file>
   full path to input gene expression file; if full path is not provided,
   assumes the file is in the working (output) directory
+ required
   
-o <output dir>, --workingDir <output dir>
   R working directory; where output from the iterativeWGCNA run is stored
   default: current directory
+ at minimum files final-membership.txt and final-eigengenes.txt must be in the directory

   	
-f, --finalMergeCutHeight <cut height>
	cut height (max dissimilarity) for final module merge
	(after algorithm convergence); [0, 1.0], default=0.05

-p <param list>, --wgcnaParameters <param list>
   comma separated list of parameter=value pairs required to assess module similarity and gene reassignment
   The following parameters are required (defaults will be used if not specified):
+ minKMEtoStay --> should be the same as used when running iterativeWGCNA; default 0.8
+ reassignThreshold --> p-value cut-off for shifting a gene to module assignment; default 0.05

If the iterativeWGCNA package was installed, run as follows:

iterativeWGCNA_merge -i <input_file_path> -o <iterativeWGCNA_output_dir> --finalMergeCutHeight <float> -p minKMEtoStay=<float;same_as_iterativeWGCNA_run>

alternative, it can be run using the wrapper script in the iterativeWGCNA directory

python merge_close_modules.py -i <input_file_path> -o <iterativeWGCNA_output_dir> --finalMergeCutHeight <float> -p minKMEtoStay=<float;same_as_iterativeWGCNA_run>

Troubleshooting

libreadline.so.6 undefined symbol

Access to the readline library in the context of the rpy2 library and an Ananconda install can be problematic and has been reported elsewhere. In trying to run iterativeWGCNA, an error like the following would be observed:

Traceback (most recent call last):
  File "../iterativeWGCNA-master/run_iterative_wgcna.py", line 7, in <module>
    from iterativeWGCNA.iterativeWGCNA import IterativeWGCNA
  File "../iterativeWGCNA-master/iterativeWGCNA/iterativeWGCNA.py", line 17, in <module>
    import rpy2.robjects as ro
  File "../lib/python2.7/site-packages/rpy2/robjects/__init__.py", line 15, in <module>
    import rpy2.rinterface as rinterface
  File "../lib/python2.7/site-packages/rpy2/rinterface/__init__.py", line 100, in <module>
    from rpy2.rinterface._rinterface import *
ImportError: ../lib/python2.7/site-packages/rpy2/rinterface/../../../../libreadline.so.6: undefined symbol: PC

The workaround is to uncomment the readline import in the run_iterative_wgcna.py script:

# import readline

Cannot install rpy2 with latest R (version 3.4.x) on macOS

error: command 'gcc' failed with exit status 1

The build process for many R libraries explicitly uses gcc, instead of the system default, which on OSX is clang. To override the default compiler, set the CC environmental variable as folows:

export CC=clang

clang: error: unsupported option '-fopenmp'

This is a known issue with an open ticket in the rpy2 project (see issue #403). R 3.4.0 was built using the -fopenmp flag with Clang 4.0.0, which is not supplied by Apple. There are several suggested workarounds (e.g., installing the LLVM library via homebrew) that do not work for all system configurations. We recommend downloading the rpy2 source, unpacking, and editing the setup.py file after line 268 (the line above the comment # OS X's frameworks need special attention) as follows:

  if "-fopenmp" in unknown:  # remove linker argument
        unknown.remove("-fopenmp")

With this fix you should be able to build rpy2 from the downloaded source as follows:

python setup.py install

Segmentation Faults, missing C libs, etc

iterativeWGCNA is written in Python but has dependencies on R and the rpy2 Python-R interface that both rely on C libraries.

If iterativeWGCNA is crashing as soon as it starts due to a segmentation fault, or you get an error along the lines of

ImportError: <some C library>.so.0: cannot open shared object file: No such file or directory

then you are having C-related troubles.

Most likely, you are using the Anaconda package and environment system which has known issues with R and R-interfaces such as rpy2.

Many of these issues have already been addressed in user groups/issue trackers for Anaconda, conda-forge and ryp2.