view template_script_DESeq2_CL.r @ 0:581d217c7337 draft

Planemo upload
author lgueguen
date Fri, 22 Jul 2016 05:39:13 -0400
parents
children de6d0b7c17af
line wrap: on
line source

#!/local/gensoft2/exe/R/3.1.2/bin/Rscript

# to run this script, use one of these commands:
# Rscript --no-save --no-restore --verbose template_script_DESeq2_CL.r -r raw -v group -c T0 > log.txt 2>&1
# Rscript template_script_DESeq2_CL.r -r raw -v group -c T0

# to get help:
# Rscript template_script_DESeq2_CL.r --help

################################################################################
### R script to compare several conditions with the SARTools and DESeq2 packages
### Hugo Varet
### April 20th, 2015
### designed to be executed with SARTools 1.1.0
################################################################################

rm(list=ls())                                        # remove all the objects from the R session
library(optparse)                                    # to run the script in command lines

# options list with associated default value.
option_list <- list( 
make_option(c("-P", "--projectName"),
			default=basename(getwd()),
			dest="projectName",
			help="name of the project used for the report [default: name of the current directory]."),

make_option(c("-A", "--author"),
			default=Sys.info()[7],
			dest="author",
			help="name of the report author [default: %default]."),

make_option(c("-t", "--targetFile"),
			default="target.txt",
			dest="targetFile",
			help="path to the design/target file [default: %default]."),

make_option(c("-r", "--rawDir"),
			default="raw",
			dest="rawDir",
			help="path to the directory containing the HTSeq files [default: %default]."),

make_option(c("-F", "--featuresToRemove"),
			default="alignment_not_unique,ambiguous,no_feature,not_aligned,too_low_aQual",
			dest="FTR",
			help="names of the features to be removed, more than once can be specified [default: %default]"),
			
make_option(c("-v", "--varInt"),
			default="group",
			dest="varInt", 
			help="factor of interest [default: %default]"),

make_option(c("-c", "--condRef"),
			default="WT",
			dest="condRef",
			help="reference biological condition [default: %default]"),

make_option(c("-b", "--batch"),
			default=NULL,
			dest="batch",
			help="blocking factor [default: %default] or \"batch\" for example"),

make_option(c("-f", "--fitType"),
			default="parametric",
			dest="fitType", 
			help="mean-variance relationship: [default: %default] or local"),

make_option(c("-o", "--cooksCutoff"),
			default=TRUE,
			dest="cooksCutoff", 
			help="perform the outliers detection (default is TRUE)"),

make_option(c("-i", "--independentFiltering"),
			default=TRUE,
			dest="independentFiltering",
			help="perform independent filtering (default is TRUE)"),

make_option(c("-a", "--alpha"),
			default=0.05,
			dest="alpha", 
			help="threshold of statistical significance [default: %default]"),

make_option(c("-p", "--pAdjustMethod"),
			default="BH",
			dest="pAdjustMethod", 
			help="p-value adjustment method: \"BH\" or \"BY\" [default: %default]"),

make_option(c("-T", "--typeTrans"),
			default="VST",
			dest="typeTrans", 
			help="transformation for PCA/clustering: \"VST\" ou \"rlog\" [default: %default]"),

make_option(c("-l", "--locfunc"),
			default="median",
			dest="locfunc", 
			help="median or shorth to estimate the size factors [default: %default]"),

make_option(c("-C", "--colors"),
			default="dodgerblue,firebrick1,MediumVioletRed,SpringGreen,chartreuse,cyan,darkorchid,darkorange",
			dest="cols",
			help="colors of each biological condition on the plots\n\t\t\"col1,col2,col3,col4\"\n\t\t[default: %default]")
)

# now parse the command line to check which option is given and get associated values
parser <- OptionParser(usage="usage: %prog [options]",
					   option_list=option_list, 
					   description="Compare two or more biological conditions in a RNA-Seq framework with DESeq2.",
					   epilogue="For comments, bug reports etc... please contact Hugo Varet <hugo.varet@pasteur.fr>")
opt <- parse_args(parser, args=commandArgs(trailingOnly=TRUE), positional_arguments=0)$options

# get options and arguments
workDir <- getwd()
projectName <- opt$projectName                       # name of the project
author <- opt$author	                             # author of the statistical analysis/report
targetFile <- opt$targetFile                         # path to the design/target file
rawDir <- opt$rawDir        						 # path to the directory containing raw counts files
featuresToRemove <- unlist(strsplit(opt$FTR, ","))   # names of the features to be removed (specific HTSeq-count information and rRNA for example)
varInt <- opt$varInt                                 # factor of interest
condRef <- opt$condRef                               # reference biological condition
batch <- opt$batch                                   # blocking factor: NULL (default) or "batch" for example
fitType <- opt$fitType                               # mean-variance relationship: "parametric" (default) or "local"
cooksCutoff <- opt$cooksCutoff                       # outliers detection threshold (NULL to let DESeq2 choosing it)
independentFiltering <- opt$independentFiltering     # TRUE/FALSE to perform independent filtering (default is TRUE)
alpha <- as.numeric(opt$alpha)                       # threshold of statistical significance
pAdjustMethod <- opt$pAdjustMethod                   # p-value adjustment method: "BH" (default) or "BY"
typeTrans <- opt$typeTrans                           # transformation for PCA/clustering: "VST" ou "rlog"
locfunc <- opt$locfunc                               # "median" (default) or "shorth" to estimate the size factors
colors <- unlist(strsplit(opt$cols, ","))            # vector of colors of each biologicial condition on the plots
	
# print(paste("workDir", workDir))
# print(paste("projectName", projectName))
# print(paste("author", author))
# print(paste("targetFile", targetFile))
# print(paste("rawDir", rawDir))
# print(paste("varInt", varInt))
# print(paste("condRef", condRef))
# print(paste("batch", batch))
# print(paste("fitType", fitType))
# print(paste("cooksCutoff", cooksCutoff))
# print(paste("independentFiltering", independentFiltering))
# print(paste("alpha", alpha))
# print(paste("pAdjustMethod", pAdjustMethod))
# print(paste("typeTrans", typeTrans))
# print(paste("locfunc", locfunc))
# print(paste("featuresToRemove", featuresToRemove))
# print(paste("colors", colors))

################################################################################
###                             running script                               ###
################################################################################
# setwd(workDir)
library(SARTools)

# checking parameters
problem <- checkParameters.DESeq2(projectName=projectName,author=author,targetFile=targetFile,
                       rawDir=rawDir,featuresToRemove=featuresToRemove,varInt=varInt,
                       condRef=condRef,batch=batch,fitType=fitType,cooksCutoff=cooksCutoff,
                       independentFiltering=independentFiltering,alpha=alpha,pAdjustMethod=pAdjustMethod,
                       typeTrans=typeTrans,locfunc=locfunc,colors=colors)
if (problem) quit(save="yes")
					   
# loading target file
target <- loadTargetFile(targetFile=targetFile, varInt=varInt, condRef=condRef, batch=batch)

# loading counts
counts <- loadCountData(target=target, rawDir=rawDir, featuresToRemove=featuresToRemove)

# description plots
majSequences <- descriptionPlots(counts=counts, group=target[,varInt], col=colors)

# analysis with DESeq2
out.DESeq2 <- run.DESeq2(counts=counts, target=target, varInt=varInt, batch=batch,
                         locfunc=locfunc, fitType=fitType, pAdjustMethod=pAdjustMethod,
                         cooksCutoff=cooksCutoff, independentFiltering=independentFiltering, alpha=alpha)

# PCA + clustering
exploreCounts(object=out.DESeq2$dds, group=target[,varInt], typeTrans=typeTrans, col=colors)

# summary of the analysis (boxplots, dispersions, diag size factors, export table, nDiffTotal, histograms, MA plot)
summaryResults <- summarizeResults.DESeq2(out.DESeq2, group=target[,varInt], col=colors,
                                          independentFiltering=independentFiltering, 
                                          cooksCutoff=cooksCutoff, alpha=alpha)

# save image of the R session
save.image(file=paste0(projectName, ".RData"))

# generating HTML report
writeReport.DESeq2(target=target, counts=counts, out.DESeq2=out.DESeq2, summaryResults=summaryResults,
                   majSequences=majSequences, workDir=workDir, projectName=projectName, author=author,
                   targetFile=targetFile, rawDir=rawDir, featuresToRemove=featuresToRemove, varInt=varInt,
                   condRef=condRef, batch=batch, fitType=fitType, cooksCutoff=cooksCutoff,
                   independentFiltering=independentFiltering, alpha=alpha, pAdjustMethod=pAdjustMethod,
                   typeTrans=typeTrans, locfunc=locfunc, colors=colors)