featcounts_deseq.R

#!/usr/bin/env Rscript
#+ include=FALSE


suppressMessages(require(docopt))

doc <- '
Convert a featureCounts results matrix into a dge-report using deseq2
Usage: featcounts_deseq.R [options] <count_matrix>

Options:
--contrasts=<tab_delim_table> Table with sample pairs for which dge analysis should be performed
--qcutoff <qcutoff>    Use a q-value cutoff of 0.01 instead of a q-value cutoff [default: 0.01]
--pcutoff <pcutoff>    Override q-value filter and filter by p-value instead
--min_count <min_count>   Minimal expression in any of the sample to be included in the final result list [default: 10]
--project <project_prefix>   Name to prefix all generated result files [default: ]
'

opts <- docopt(doc, commandArgs(TRUE))

#opts <- docopt(doc, "--contrasts ~/MPI-CBG_work/P5_DESeq/dba_contrasts_human.txt ~/MPI-CBG_work/P5_DESeq/countMatrix_human.txt")
## opts <- docopt(doc, "countMatrix.txt")
## opts <- docopt(doc, "../mapped/star_counts_matrix.txt")

require(knitr)
require(DESeq2)


devtools::source_url("https://raw.githubusercontent.com/holgerbrandl/datautils/v1.14/R/core_commons.R")
devtools::source_url("https://raw.githubusercontent.com/holgerbrandl/datautils/v1.14/R/ggplot_commons.R")
devtools::source_url("https://raw.githubusercontent.com/holgerbrandl/datautils/v1.14/R/bio/diffex_commons.R")


require.auto(DT)

require.auto(gplots)

count_matrix_file <- opts$count_matrix
contrasts_file <- opts$contrasts

#resultsBase <- count_matrix_file %>% basename() %>% trim_ext(".txt") %>% trim_ext(".count_matrix")
resultsBase <- if(str_length(opts$project)>0) paste0(opts$project, ".") else ""

pcutoff <- if(is.null(opts$pcutoff)) NULL else as.numeric(opts$pcutoff)
qcutoff <- if(is.numeric(pcutoff)) NULL else as.numeric(opts$qcutoff)
if(is.numeric(pcutoff)) opts$qcutoff <- NULL


########################################################################################################################
#' ## Data preparation

#' Working Directory: `r getwd()`

countData <- read.delim(count_matrix_file)

## save a backup into the current working directory as a reference
countData %>% write.delim(file=paste0(resultsBase, "input_counts.txt"))

#' strucutre of input matrix was
glimpse(countData, width=60)

countMatrix <- countData %>% column2rownames("ensembl_gene_id") %>% as.matrix()

# Expression Filtering
genesBefore <- nrow(countMatrix)
countMatrix %<>% filterByExpression(opts$min_count)
genesAfter <- nrow(countMatrix)
#' Counts were filtered to only retain genes which had more that `r opts$min_count` alignments in at least one replicates. This reduced the number of genes from `r genesBefore` to `r genesAfter`.


########################################################################################################################
#' ## Perform Differential Expression Analysis


#' DESeq is an R package to analyse count data from high-throughput sequencing assays such as RNA-Seq and test for differential expression. The latest version, DESeq2, is used.
#' DESeq uses a model based on the negative binomial distribution and offers the following features:
#' Count data is discrete and skewed and is hence not well approximated by a normal distribution. Thus, a test based on the negative binomial distribution, which can reflect these properties, has much higher power to detect differential expression.
#' Tests for differential expression between experimental conditions should take into account both technical and biological variability. Recently, several authors have claimed that the Poisson distribution can be used for this purpose. However, tests based on the Poisson assumption (this includes the binomial test and the chi-squared test) ignore the biological sampling variance, leading to incorrectly optimistic p values. The negative binomial distribution is a generalisation of the Poisson model that allows to model biological variance correctly. 
#' DESeq estimate the variance in a local fashion, using different coefficients of variation for different expression strengths. This removes potential selection biases in the hit list of differentially expressed genes, and gives a more balanced and accurate result.

#' See deseq reference [docs](http://master.bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.pdf) for details

#' To understand fold-change shrinking and estimation check out
#' [Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2](http://genomebiology.com/2014/15/12/550/abstract)


get_sample_from_replicate <- function(repName) str_match(repName, "(.*)_[R]?[0-9]{1}$")[,2]

# Define or load a contrasts matrix
if(!is.null(contrasts_file)){
    contrasts <- read.delim(contrasts_file, header=T) %>% fac2char()
}else{
    contrasts <- data.frame(sample=get_sample_from_replicate(colnames(countMatrix))) %>% distinct() %>%
        merge(.,., suffixes=c("_1", "_2"), by=NULL) %>%
        filter(ac(sample_1)>ac(sample_2)) %>%
#        filter(ac(sample_1)!=ac(sample_2)) %>%
        fac2char()
    write.delim(contrasts, paste(resultsBase, "contrasts.txt"))
}

#' The used contrasts model is
contrasts %>% kable()

## gene specific factors
#normalizationFactors(dds)

## sizeFactor R help:sigEnrResults
#dds <- makeExampleDESeqDataSet()
#dds <- estimateSizeFactors( dds )
#sizeFactors(dds)

## try again but now use lambda normalization
## see "3.11 Sample-/gene-dependent normalization factors" in the DEseq2 manual for details
colData <- data.frame(condition=colnames(countMatrix) %>% get_sample_from_replicate)

# valiadate that contrasts model is compatible with data
stopifnot(all((contrasts %>% gather %$% value %>% unique) %in% colData$condition))

dds <- DESeqDataSetFromMatrix(countMatrix, colData, formula(~ condition))
dds <- estimateSizeFactors(dds)


#' For details about size factor normalziation and calculation see https://www.biostars.org/p/79978/

##' Size Factors estimated by DESeq
sizeFactors(dds) %>%
    set_names(colnames(countMatrix)) %>% melt() %>%
    rownames2column("sample") %>%
    ggplot(aes(sample, value)) + geom_bar(stat="identity") + ggtitle("deseq size factors") + coord_flip()

##' From the DESeq docs about how size factors are used: The sizeFactors vector assigns to each column of the count matrix a value, the size factor, such that count values in the columns can be brought to a common scale by dividing by the corresponding size factor.

##' This means that counts are divied by size factors. So let's now load the lambda libraies and replace the predefined size factors with our custom ones

#' From DESeq manual: The regularized log transformation is preferable to the variance stabilizing transformation if the size factors vary widely.

#' Run Deseq Test
#dds <- DESeq(dds, fitType='local', betaPrior=FALSE)
#dds <- DESeq(dds, fitType='local')
dds <- DESeq(dds)

#res <- results(dds)
#resultsNames(dds)

#' Model Overview:
summary(results(dds))

########################################################################################################################
#' ## Quality Control

#' ### Data Dispersion

# TODO explain how this plot is useful
#' Plot of the per-gene dispersion estimates together with the fitted mean-dispersion relationship.
#' The dispersion plot shows how the estimates are shrunk from the gene wise values (black dots) toward the fitted estimates, with the final values used in testing being the blue dots.
#' The dispersion can be understood as the square of the coefficient of biological variation.  So, if a gene's expression typically differs from replicate to replicate sample by 20%, this gene's dispersion is: .20^2 = .04.
## The function estimateDispersions performs three steps. First, it estimates, for each gene and each (replicated) condition, a dispersion value, then, it fits, for each condition, a curve through the estimates. Finally, it assigns to each gene a dispersion value, using either the estimated or the fitted value.
#dds <- estimateDispersions(dds)
plotDispEsts(dds, main="Dispersion plot")

########################################################################################################################
#' ### PCA and Clustering

#' In order to assess overall similarity between samples two common statistical methods are used - Principal component analysis (PCA) and clustering.
#' This should provide answers to the questions: Which samples are similar to each other, which are different? Does this fit to the expectation from the experiment’s design?
#' Using a regularized log transformation of the raw counts provides the advantage that it stabilizes the variance across the mean.


#' Principal component analysis (PCA) is a technique used to emphasize variation and bring out strong patterns in a dataset.
#' Principal components are the underlying structure in the data. They are the directions where there is the most variance, the directions where the data is most spread out. The data points/samples are projected onto the 2D plane such that they spread out in the two directions that explain most of the differences. The x-axis is the direction that separates the data points the most and the y-axis is a direction that separates the data the second most.

# Regularized log transformation for clustering/heatmaps, etc
rld <- rlogTransformation(dds)

plotPCA(rld, intgroup = "condition")


#' The Euclidean distance between samples are calculated after performing the regularized log transformation.
#' Using the calculated distance matrix, the samples are projected on a two-dimensional graph such that the distance between samples approximately corresponds to the biological coefficient of variation between those samples.

#distsRL <- dist(t(assay(rld)))
#hc <- hclust(distsRL)
#distMatrix <- as.matrix(distsRL)
distMatrix <- as.matrix(dist(t(assay(rld))))
#rownames(distMatrix) <- colnames(distMatrix) <- with(colData(dds), paste(condition, sep=" : "))

labelcntData <- countData %>%
  distinct(ensembl_gene_id) %>% column2rownames("ensembl_gene_id") %>%
  fac2char %>% data.matrix() %>% round() %>% data.frame()

heatmap.2(distMatrix, labRow=colnames(labelcntData), labCol=colnames(labelcntData),
          symm=TRUE, trace="none",
          #key=F, col=colorpanel(100, "black", "white"),
          margin=c(8, 8), main="Sample Distance Matrix")

########################################################################################################################


## extract all de-sets according to our contrasts model
deResults <- adply(contrasts, 1,  splat(function(sample_1, sample_2){
        results(dds, contrast=c("condition", sample_1, sample_2)) %>%
        rownames2column("ensembl_gene_id") %>%
        as.data.frame() %>%
        ## see http://rpackages.ianhowson.com/bioc/DESeq2/man/results.html when using contrasts argument
        rename(s1_over_s2_logfc=log2FoldChange) %>%
        mutate(sample_1=sample_1, sample_2=sample_2)
}))

# deResults <- adply(contrasts, 1,  splat(function(sample_1, sample_2){
#  echo('samples', sample_1, sample_2)
#   return(data.frame())
# }))

deResults %>% ggplot(aes(s1_over_s2_logfc)) +
    geom_histogram(binwidth=0.1) +
    facet_grid(sample_1 ~ sample_2) + geom_vline(yintercept=0, color="blue") +
    xlim(-2,2) + 
    ggtitle("sample1 over sampl2 logFC ")


########################################################################################################################
#' ## Significnce of differential binding

#deResults %>% ggplot(aes(pvalue)) +
#    geom_histogram() +
#    xlim(0,1) +
#    facet_grid(sample_1 ~ sample_2) + geom_vline(yintercept=0.01, slope=1, color="blue") +
#    ggtitle("p-value distributions") #+ scale_x_log10()
#
#deResults %>% ggplot(aes(padj)) +
#    geom_histogram() +
##    xlim(0,1) +
#    facet_grid(sample_1 ~ sample_2) + geom_vline(yintercept=0.01, slope=1, color="blue") +
#    ggtitle("Adjusted p-value distributions") #+ scale_x_log10()


# report hit criterion
#+ results='asis'
if(!is.null(qcutoff)){
    echo("Using q-value cutoff of", qcutoff)
    deResults %<>% transform(is_hit=padj<=qcutoff)
}else{
    echo("Using p-value cutoff of", pcutoff)
    deResults %<>% transform(is_hit=pvalue<=pcutoff)
}
#deResults %<>% mutate(is_hit=pvalue<0.05)

deResults %<>% mutate(s1_overex=s1_over_s2_logfc>1)


normCounts <- counts(dds,normalized=TRUE) %>%
#    set_names(colData(dds)$condition) %>% rownames2column("ensembl_gene_id")
    set_names(colnames(countMatrix)) %>% rownames2column("ensembl_gene_id")


## .. should be same as input
#filter(countData, ensembl_gene_id=="ENSDARG00000000001")

########################################################################################################################
#' # MA and Volcano plots

# deseq approach
# plotMA(deResults, main="DESeq2", ylim=c(-2,2))

# Calculation of the base means per condition
baseMeanPerLvl <- sapply( levels(dds$condition), function(lvl) rowMeans( counts(dds,normalized=TRUE)[,dds$condition == lvl] ) ) %>% 
  rownames2column("ensembl_gene_id")

## add base means to diffßex summary
#numResultsBeforeMerge <- nrow(deResults)
deResults <- baseMeanPerLvl %>%
    gather(sample, norm_count, -ensembl_gene_id) %>%
    merge(.,., by="ensembl_gene_id", suffixes=c("_1", "_2")) %>%
#    filter(ac(sample_1)<ac(sample_2)) %>%
    # add diffex status
    merge(deResults)

#+ fig.width=16, fig.height=14
deResults %>% ggplot(aes(0.5*log2(norm_count_1*norm_count_2), log2(norm_count_2/norm_count_1), color=pvalue<0.05)) +
    geom_point(alpha=0.3) +
    geom_hline(yintercept=0, color="red") +
    facet_grid(sample_1 ~ sample_2)


#deResults %$% pvalue %>%  log10() %>% quantile(0.05, na.rm=T)

#' For the volcano plot we have fold change on the X-axis and -log10 of the p-value on the y-axis.
##Volcano plots
hitCounts <- filter(deResults, is_hit) %>%
        count(sample_1, sample_2, s1_overex) %>%
        rename(hits=n) %>%
        merge(data.frame(s1_overex=c(T,F), x_pos=c(-2.5,2.5)))

#+ fig.width=16, fig.height=14
deResults %>% ggplot(aes(s1_over_s2_logfc, -log10(pvalue), color=is_hit)) +
    geom_jitter(alpha=0.3, position = position_jitter(height = 0.2)) +
#    theme_bw() +
    xlim(-3,3) +
    scale_color_manual(values = c("TRUE"="red", "FALSE"="black")) +
#    ggtitle("Insm1/Ctrl") +
    ## http://stackoverflow.com/questions/19764968/remove-point-transparency-in-ggplot2-legend
    guides(colour = guide_legend(override.aes = list(alpha=1))) +

    ## tweak axis labels
    xlab(expression(log[2]("x/y"))) +
    ylab(expression(-log[10]("p value"))) +
#  ylim(0,50) +

    ## add hit couts
    geom_text(aes(label=hits, x=x_pos), y=2, color="red", size=10, data=hitCounts) +
    facet_grid(sample_1 ~ sample_2)



# Define absolute binding categories
#rawCounts <- counts(dds,normalized=F) %>%
#    set_names(colData(dds)$condition) %>% rownames2column("ensembl_gene_id")
#ggplot(rawCounts, aes(H3HA_Sphere)) + geom_histogram() + scale_x_log10() + ggtitle("raw H3HA_Sphere counts distribution")
#require.auto(Hmisc)
#
#bindCats <- rawCounts %>% transmute(ensembl_gene_id, bind_category=cut2(H3HA_Sphere, cuts=c(10, 100)))
#with(bindCats, as.data.frame(table(bind_category))) %>% kable()



# Load transcriptome annotations needed for results annotation
geneInfo <- quote({
  ## mart <- biomaRt::useDataset("drerio_gene_ensembl", mart = biomaRt::useMart("ensembl"))§
  mart <- biomaRt::useDataset(guess_mart(countData$ensembl_gene_id), mart = biomaRt::useMart("ensembl"))
  c("ensembl_gene_id", "external_gene_name", "description", "chromosome_name", "start_position", "end_position") %>%
        biomaRt::getBM(mart=mart)
}) %>% cache_it("geneInfo")


# Export the complete dataset for later analysis
deAnnot <- deResults %>%
  #inner_join(normCounts) %>% 
  #merge(., contrasts) %>%
  #merge(.,., suffixes=c("_1", "_2"), by=NULL) %>%
  #inner_join(baseMeanPerLvl, id='ensembl_gene_id') %>%
  left_join(geneInfo)

write.delim(deAnnot, file=paste0(resultsBase, "de_results.txt"))
#'  [deAnnot](`r paste0(resultsBase, ".dba_results.txt")`)

## also export results which we'll always need for pathways overlays
deAnnot %>% transmute(
        ensembl_gene_id,
        contrast=paste(sample_1, "vs", sample_2),
        plot_score=-log10(pvalue)*ifelse(s1_overex, 1, -1)
    ) %>% spread(contrast, plot_score) %>%
    write.delim(file=paste0(resultsBase, "plot_score_matrix.txt"))

#deAnnot %>% filter(ensembl_gene_id=="FBgn0000015")


## todo understand purpose and effeect of indpendentFiltering (see https://support.bioconductor.org/p/57128/)
#deResults %>% count(is_hit)
#deAnnot %>% count(is_hit)
#deAnnot %>% count(is.na(padj))

########################################################################################################################
#' ## Hits Summary

## Extract hits from deseq results
degs <- deAnnot %>% filter(is_hit)

# ggplot(degs, aes(paste(sample_1, "vs",  sample_2))) + geom_bar() +xlab(NULL) + ylab("# DBGs") + ggtitle("DBG count summary") + coord_flip()

# degs %>%
#     ggplot(aes(paste(sample_1, "vs",  sample_2), fill=s1_overex)) +
#     geom_bar(position="dodge") +
#     xlab(NULL) + ylab("# DBGs") +
#     ggtitle("DBG count summary by direction of expression") +
#     coord_flip()

# Export DBA genes
## disabled because we just subset the annotated data now to define degs
#degsAnnot <- degs %>%
#    inner_join(normCounts) %>%
#    left_join(geneInfo) %>%
#    left_join(bindCats)
degs %>% write.delim(file=paste0(resultsBase, "diffex_genes.txt"))
#'  [degs](`r paste0(resultsBase, ".diffbind_genes.txt")`)


#unloadNamespace("dplyr"); require(dplyr)

## export slim hit list for downstream enrichment analysis
degs %>% transmute(ensembl_gene_id, contrast=paste(sample_1, "vs", sample_2)) %>% write.delim(paste0(resultsBase, "degs_by_contrast.txt"))


# ## render interactive hit browser
# #+results='asis'
# degs %>%
#     left_join(geneInfo) %>%
#     mutate(
#         igv=paste0("<a href='http://localhost:60151/goto?locus=", chromosome_name,":", start_position, "-", end_position, "'>IGV</a>")
#     ) %>%
#     select(s1_over_s2_logfc, pvalue, ensembl_gene_id, sample_1, sample_2, external_gene_name, description, igv) %>%
# #    kable("html", table.attr = "class='dtable'", escape=F)
#     datatable(escape=F)


#' ### DEG Counts

#+ fig.height=nrow(contrasts)+2
#ggplot(degs, aes(paste(sample_1, "vs",  sample_2), fill=status)) + geom_bar() +xlab(NULL) + ylab("# DGEs") +ggtitle("DEGs by contrast") + coord_flip()
ggplot(degs, aes(paste(sample_1, "vs",  sample_2), fill=(s1_overex))) + geom_bar() + xlab(NULL) + ylab("# DGEs") + ggtitle("DEGs by contrast") + coord_flip()

#with(degs, as.data.frame(table(sample_1, sample_2, s1_overex))) %>% filter(Freq >0) %>% kable()
degs %>% count( sample_1, sample_2, s1_overex) %>% kable()


#'  DEGs (differentially expressed genes) are contained in the following table
write.delim(degs, file="degs.txt")
# degs <- read.delim("degs.txt")
#' [Differentially Expressed Genes](degs.txt)


#' DEGs can be browsed in Excel using the exported table or via the embedded table browser. To enable the IGV links, you need to set the port in the IGV preferences to 3334.
kableDegs <- degs
if(nrow(degs>2000)){
  kableDegs <- degs %>% arrange(padj) %>% head(1000)
  #  Error: object 'q_value' not found. So: padj used instead of '-q_value'
  print("just showing first 1000 most significant hits (highest p-value) in interactive hit table!!!! Use Excel to browser to browse the full set")
}

#+ results='asis'
kableDegs %>%
  #inner_join(geneLoci) %>%
  mutate(
    ensembl=paste0("<a href='http://www.ensembl.org/Multi/Search/Results?y=0;site=ensembl_all;x=0;page=1;facet_feature_type=Gene;q=",ensembl_gene_id, "'>",ensembl_gene_id, "</a>"),
    igv=paste0("<a href='http://localhost:60151/goto?locus=", chromosome_name,":", start_position, "-", end_position, "'>IGV</a>")
  ) %>%
  select(sample_1, sample_2, s1_over_s2_logfc, pvalue, padj, ensembl_gene_id, external_gene_name, description) %>%
#  kable("html", table.attr = "class='dtable'", escape=F)
   datatable(escape=F)

## just needed to restore environment easily
#save(degs, file=".degs.RData")
# degs <- local(get(load(".degs.RData")))


########################################################################################################################
#' ## MA Plots

#' Redo MA plots but now including hit overlays

maPlots <- deResults %>% group_by(sample_1, sample_2) %>% do(gg={ maData <-.
## todo why not s2_over_s1_log2fc
maData %>% ggplot(aes(0.5*log2(norm_count_1*norm_count_2), log2(norm_count_1/norm_count_2), color=is_hit)) +
  geom_point(alpha=0.3) +
  geom_hline(yintercept=0, color="red") +
  ggtitle(with(maData[1,], paste(sample_1, "vs", sample_2)))
}) %$% gg

#+ fig.height=6*ceiling(length(maPlots)/2)
multiplot(plotlist=maPlots, cols=min(2, length(maPlots)))


##--------------------------------------------------------

## Useful link: https://gist.github.com/stephenturner/f60c1934405c127f09a6

## TODO later, reenable child-inclusion of enrichment analysis
########################################################################################################################
## ## Term enrichment Data Preparation
##+ echo=FALSE
#
## This analysis was performed using [David](http://david.abcc.ncifcrf.gov/). The following ontologies were tested: `r paste(ontologies, collapse=', ')`
#
#geneLists <- degs %>%
#    transmute(ensembl_gene_id, list_id=paste(sample_1, "vs", sample_2))
#
#if(nrow(contrasts)<4){
#    geneLists <- rbind_list(
#        geneLists,
#        degs %>% filter(s1_overex) %>% transmute(ensembl_gene_id, list_id=paste(sample_1, ">", sample_2)),
#        degs %>% filter(!s1_overex) %>% transmute(ensembl_gene_id, list_id=paste(sample_1, "<", sample_2))
#    )
#}
#
### additional overlaps before doing the intersection analysis
#geneLists %>% count(list_id) %>% kable()
#
#intersectLists <- function(geneLists, listIdA, listIdB, intersectListId) {
#    commonGenes <- setdiff(filter(geneLists, list_id==listIdA)$ensembl_gene_id, filter(geneLists, list_id==listIdB)$ensembl_gene_id)
#    data.frame(list_id=intersectListId, ensembl_gene_id=commonGenes)
#}
#
#geneLists %<>% group_by(list_id)
#save(geneLists, file=".enrGeneLists.RData")
## geneLists <- local(get(load("enrGeneLists.RData")))
#
#
### redefine opts arguments and tweak knitr options
#opts_knit$set(root.dir = getwd())
#commandArgs <- function(x) c(paste("--overlay_expr_data ", count_matrix_file, " enrGeneLists.RData"))
##source("/home/brandl/mnt/mack/bioinfo/scripts/ngs_tools/dev/common/david_enrichment.R")
## child='/home/brandl/mnt/mack/bioinfo/scripts/ngs_tools/dev/common/david_enrichment.Rmd'