1 Introduction

Ribo-Seq and polyRibo-Seq are a specific form of RNA-Seq gene expression experiments utilizing mRNA subpopulations directly bound to ribosomes. Compared to standard RNA-Seq, their readout of gene expression provides a better approximation of downstream protein abundance profiles due to their close association with translational processes. The most important difference among the two is that polyRibo-Seq utilizes polyribosomal RNA for sequencing, whereas Ribo-Seq is a footprinting approach restricted to sequencing RNA fragments protected by ribosomes (Ingolia et al. 2009; Aspden et al. 2014; Juntawong et al. 2015).

The workflow presented in this vignette contains most of the data analysis steps described by (Juntawong et al. 2014) including functionalities useful for processing both polyRibo-Seq and Ribo-Seq experiments. To improve re-usability and adapt to recent changes of software versions (e.g. R, Bioconductor and short read aligners), the code has been optimized accordingly. Thus, the results obtained with the updated workflow are expected to be similar but not necessarily identical with the published results described in the original paper.

Relevant analysis steps of this workflow include read preprocessing, read alignments against a reference genome, counting of reads overlapping with a wide range of genomic features (e.g. CDSs, UTRs, uORFs, rRNAs, etc.), differential gene expression and differential ribosome binding analyses, as well as a variety of genome-wide summary plots for visualizing RNA expression trends. Functions are provided for evaluating the quality of Ribo-seq data, for identifying novel expressed regions in the genomes, and for gaining insights into gene regulation at the post-transcriptional and translational levels. For example, the functions genFeatures and featuretypeCounts can be used to quantify the expression output for all feature types included in a genome annotation (e.g. genes, introns, exons, miRNAs, intergenic regions, etc.). To determine the approximate read length of ribosome footprints in Ribo-Seq experiments, these feature type counts can be obtained and plotted for specific read lengths separately. Typically, the most abundant read length obtained for translated features corresponds to the approximate footprint length occupied by the ribosomes of a given organism group. Based on the results from several Ribo-Seq studies, these ribosome footprints are typically ~30 nucleotides long (Ingolia, Lareau, and Weissman 2011; Ingolia et al. 2009; Juntawong et al. 2014). However, their length can vary by several nucleotides depending upon the optimization of the RNA digestion step and various factors associated with translational regulation. For quality control purposes of Ribo-Seq experiments it is also useful to monitor the abundance of reads mapping to rRNA genes due to the high rRNA content of ribosomes. This information can be generated with the featuretypeCounts function described above.

Coverage trends along transcripts summarized for any number of transcripts can be obtained and plotted with the functions featureCoverage and plotfeatureCoverage, respectively. Their results allow monitoring of the phasing of ribosome movements along triplets of coding sequences. Commonly, high quality data will display here for the first nucleotide of each codon the highest depth of coverage computed for the 5’ ends of the aligned reads.

Ribo-seq data can also be used to evaluate various aspects of translational control due to ribosome occupancy in upstream open reading frames (uORFs). The latter are frequently present in (or near) 5’ UTRs of transcripts. For this, the function predORFs can be used to identify ORFs in the nucleotide sequences of transcripts or their subcomponents such as UTR regions. After scaling the resulting ORF coordinates back to the corresponding genome locations using scaleRanges, one can use these novel features (e.g. uORFs) for expression analysis routines similar to those employed for pre-existing annotations, such as the exonic regions of genes. For instance, in Ribo-Seq experiments one can use this approach to systematically identify all transcripts occupied by ribosomes in their uORF regions. The binding of ribosomes to uORF regions may indicate a regulatory role in the translation of the downstream main ORFs and/or translation of the uORFs into functionally relevant peptides.

1.1 Experimental design

Typically, users want to specify here all information relevant for the analysis of their NGS study. This includes detailed descriptions of FASTQ files, experimental design, reference genome, gene annotations, etc.

2 Workflow environment

2.1 Load packages and sample data

The systemPipeR package needs to be loaded to perform the analysis steps shown in this report (H Backman and Girke 2016). The package allows users to run the entire analysis workflow interactively or with a single command while also generating the corresponding analysis report. For details see systemPipeR's main vignette.

library(systemPipeR)

2.2 Generate workflow environment

systemPipeRdata package is a helper package to generate a fully populated systemPipeR workflow environment in the current working directory with a single command. All the instruction for generating the workflow template are provide in the systemPipeRdata vignette here.

systemPipeRdata::genWorkenvir(workflow = "riboseq", mydirname = "riboseq")
setwd("riboseq")

After building and loading the workflow environment generated by genWorkenvir from systemPipeRdata all data inputs are stored in a data/ directory and all analysis results will be written to a separate results/ directory, while the systemPipeRIBOseq.Rmd script and the targets file are expected to be located in the parent directory. The R session is expected to run from this parent directory. Additional parameter files are stored under param/.

To work with real data, users want to organize their own data similarly and substitute all test data for their own data. To rerun an established workflow on new data, the initial targets file along with the corresponding FASTQ files are usually the only inputs the user needs to provide. Please check the initial targets file details below.

For more details, please consult the documentation here. More information about the targets files from systemPipeR can be found here.

2.3 Experiment definition provided by targets file

The targets file defines all FASTQ files and sample comparisons of the analysis workflow.

targetspath <- system.file("extdata", "targetsPE.txt", package = "systemPipeR")
targets <- read.delim(targetspath, comment.char = "#")[, 1:4]
targets
##                      FileName1                   FileName2
## 1  ./data/SRR446027_1.fastq.gz ./data/SRR446027_2.fastq.gz
## 2  ./data/SRR446028_1.fastq.gz ./data/SRR446028_2.fastq.gz
## 3  ./data/SRR446029_1.fastq.gz ./data/SRR446029_2.fastq.gz
## 4  ./data/SRR446030_1.fastq.gz ./data/SRR446030_2.fastq.gz
## 5  ./data/SRR446031_1.fastq.gz ./data/SRR446031_2.fastq.gz
## 6  ./data/SRR446032_1.fastq.gz ./data/SRR446032_2.fastq.gz
## 7  ./data/SRR446033_1.fastq.gz ./data/SRR446033_2.fastq.gz
## 8  ./data/SRR446034_1.fastq.gz ./data/SRR446034_2.fastq.gz
## 9  ./data/SRR446035_1.fastq.gz ./data/SRR446035_2.fastq.gz
## 10 ./data/SRR446036_1.fastq.gz ./data/SRR446036_2.fastq.gz
## 11 ./data/SRR446037_1.fastq.gz ./data/SRR446037_2.fastq.gz
## 12 ./data/SRR446038_1.fastq.gz ./data/SRR446038_2.fastq.gz
## 13 ./data/SRR446039_1.fastq.gz ./data/SRR446039_2.fastq.gz
## 14 ./data/SRR446040_1.fastq.gz ./data/SRR446040_2.fastq.gz
## 15 ./data/SRR446041_1.fastq.gz ./data/SRR446041_2.fastq.gz
## 16 ./data/SRR446042_1.fastq.gz ./data/SRR446042_2.fastq.gz
## 17 ./data/SRR446043_1.fastq.gz ./data/SRR446043_2.fastq.gz
## 18 ./data/SRR446044_1.fastq.gz ./data/SRR446044_2.fastq.gz
##    SampleName Factor
## 1         M1A     M1
## 2         M1B     M1
## 3         A1A     A1
## 4         A1B     A1
## 5         V1A     V1
## 6         V1B     V1
## 7         M6A     M6
## 8         M6B     M6
## 9         A6A     A6
## 10        A6B     A6
## 11        V6A     V6
## 12        V6B     V6
## 13       M12A    M12
## 14       M12B    M12
## 15       A12A    A12
## 16       A12B    A12
## 17       V12A    V12
## 18       V12B    V12

3 Workflow environment

systemPipeR workflows can be designed and built from start to finish with a single command, importing from an R Markdown file or stepwise in interactive mode from the R console.

This tutorial will demonstrate how to build the workflow in an interactive mode, appending each step. The workflow is constructed by connecting each step via appendStep method. Each SYSargsList instance contains instructions needed for processing a set of input files with a specific command-line or R software and the paths to the corresponding outfiles generated by a particular tool/step.

To create a Workflow within systemPipeR, we can start by defining an empty container and checking the directory structure:

library(systemPipeR)
sal <- SPRproject()
sal

3.1 Required packages and resources

The systemPipeR package needs to be loaded (H Backman and Girke 2016).

cat(crayon::blue$bold("To use this workflow, following R packages are expected:\n"))
cat(c("'rtracklayer", "GenomicFeatures", "grid", "BiocParallel",
    "DESeq2", "ape", "edgeR", "biomaRt", "BBmisc", "pheatmap",
    "ggplot2'\n"), sep = "', '")
### pre-end
appendStep(sal) <- LineWise(code = {
    library(systemPipeR)
    library(rtracklayer)
    library(GenomicFeatures)
    library(ggplot2)
    library(grid)
    library(DESeq2, quietly = TRUE)
    library(ape, warn.conflicts = FALSE)
    library(edgeR)
    library(biomaRt)
    library(BBmisc)  # Defines suppressAll()
    library(pheatmap)
    library(BiocParallel)
}, step_name = "load_SPR")

3.2 Read preprocessing

3.2.1 Preprocessing with preprocessReads function

The function preprocessReads allows to apply predefined or custom read preprocessing functions to all FASTQ files referenced in a SYSargsList container, such as quality filtering or adapter trimming routines. Internally, preprocessReads uses the FastqStreamer function from the ShortRead package to stream through large FASTQ files in a memory-efficient manner.

Here, we are appending a new step to the SYSargsList object created previously. All the parameters are defined on the preprocessReads/preprocessReads-pe_riboseq.yml file.

appendStep(sal) <- SYSargsList(step_name = "preprocessing", targets = "targetsPE.txt",
    dir = TRUE, wf_file = "preprocessReads/preprocessReads-pe.cwl",
    input_file = "preprocessReads/preprocessReads-pe_riboseq.yml",
    dir_path = system.file("extdata/cwl", package = "systemPipeR"),
    inputvars = c(FileName1 = "_FASTQ_PATH1_", FileName2 = "_FASTQ_PATH2_",
        SampleName = "_SampleName_"), dependency = c("load_SPR"))

The function trimbatch used trims adapters hierarchically from the longest to the shortest match of the right end of the reads. If internalmatch=TRUE then internal matches will trigger the same behavior. The argument minpatternlength defines the shortest adapter match to consider in this iterative process. In addition, the function removes reads containing Ns or homopolymer regions. More detailed information on read preprocessing is provided in systemPipeR's main vignette.

yamlinput(sal, step = "preprocessing")$Fct
# [1] ''trimbatch(fq, pattern=\'ACACGTCT\',
# internalmatch=FALSE, minpatternlength=6, Nnumber=1,
# polyhomo=50, minreadlength=16, maxreadlength=101)''
cmdlist(sal, step = "preprocessing", targets = 1)

3.2.2 FASTQ quality report

The following seeFastq and seeFastqPlot functions generate and plot a series of useful quality statistics for a set of FASTQ files, including per cycle quality box plots, base proportions, base-level quality trends, relative k-mer diversity, length, and occurrence distribution of reads, number of reads above quality cutoffs and mean quality distribution. The results are written to a png file named fastqReport.png.

This is the pre-trimming fastq report. Another post-trimming fastq report step is not included in the default. It is recommended to run this step first to decide whether the trimming is needed.

Please note that initial targets files are being used here. In this case, we used the getColumn function to extract a named vector.

appendStep(sal) <- LineWise(code = {
    fq_files <- getColumn(sal, "preprocessing", "targetsWF",
        column = 1)
    fqlist <- seeFastq(fastq = fq_files, batchsize = 10000, klength = 8)
    png("./results/fastqReport.png", height = 162, width = 288 *
        length(fqlist))
    seeFastqPlot(fqlist)
    dev.off()
}, step_name = "fastq_report", dependency = "preprocessing")
Figure 1: FASTQ quality report. To zoom in, right click image and open it in a separate browser tab.


3.3 Alignments

3.3.1 Read mapping with HISAT2

The following steps will demonstrate how to use the short read aligner Hisat2 (Kim, Langmead, and Salzberg 2015). First, the Hisat2 index needs to be created.

appendStep(sal) <- SYSargsList(step_name = "hisat2_index", dir = FALSE,
    targets = NULL, wf_file = "hisat2/hisat2-index.cwl", input_file = "hisat2/hisat2-index.yml",
    dir_path = "param/cwl", dependency = "load_SPR")

3.3.2 HISAT2 mapping

The parameter settings of the aligner are defined in the workflow_hisat2-pe.cwl and workflow_hisat2-pe.yml files. The following shows how to construct the corresponding SYSargsList object.

appendStep(sal) <- SYSargsList(step_name = "hisat2_mapping",
    dir = TRUE, targets = "targetsPE.txt", wf_file = "workflow-hisat2/workflow_hisat2-pe.cwl",
    input_file = "workflow-hisat2/workflow_hisat2-pe.yml", dir_path = "param/cwl",
    inputvars = c(FileName1 = "_FASTQ_PATH1_", FileName2 = "_FASTQ_PATH2_",
        SampleName = "_SampleName_"), dependency = c("hisat2_index"))

To double-check the command line for each sample, please use the following:

cmdlist(sal, step = "hisat2_mapping", targets = 1)

3.3.3 Read and alignment stats

The following provides an overview of the number of reads in each sample and how many of them aligned to the reference.

appendStep(sal) <- LineWise(code = {
    fqpaths <- getColumn(sal, step = "hisat2_mapping", "targetsWF",
        column = "FileName1")
    bampaths <- getColumn(sal, step = "hisat2_mapping", "outfiles",
        column = "samtools_sort_bam")
    read_statsDF <- alignStats(args = bampaths, fqpaths = fqpaths,
        pairEnd = TRUE)
    write.table(read_statsDF, "results/alignStats.xls", row.names = FALSE,
        quote = FALSE, sep = "\t")
}, step_name = "align_stats", dependency = "hisat2_mapping")

3.5 Read distribution across genomic features

The genFeatures function generates a variety of feature types from TxDb objects using utilities provided by the GenomicFeatures package.

3.5.1 Obtain feature types

The first step is the generation of the feature type ranges based on annotations provided by a GFF file that can be transformed into a TxDb object. This includes ranges for mRNAs, exons, introns, UTRs, CDSs, miRNAs, rRNAs, tRNAs, promoter and intergenic regions. In addition, any number of custom annotations can be included in this routine.

appendStep(sal) <- LineWise(code = {
    txdb <- suppressWarnings(makeTxDbFromGFF(file = "data/tair10.gff",
        format = "gff3", dataSource = "TAIR", organism = "Arabidopsis thaliana"))
    feat <- genFeatures(txdb, featuretype = "all", reduce_ranges = TRUE,
        upstream = 1000, downstream = 0, verbose = TRUE)
}, step_name = "genFeatures", dependency = "hisat2_mapping",
    run_step = "mandatory")

3.5.2 Count and plot reads of any length

The featuretypeCounts function counts how many reads in short read alignment files (BAM format) overlap with entire annotation categories. This utility is useful for analyzing the distribution of the read mappings across feature types, e.g. coding versus non-coding genes. By default the read counts are reported for the sense and antisense strand of each feature type separately. To minimize memory consumption, the BAM files are processed in a stream using utilities from the Rsamtools and GenomicAlignment packages. The counts can be reported for each read length separately or as a single value for reads of any length. Subsequently, the counting results can be plotted with the associated plotfeaturetypeCounts function.

The following generates and plots feature counts for any read length.

appendStep(sal) <- LineWise(code = {
    outpaths <- getColumn(sal, step = "hisat2_mapping", "outfiles",
        column = "samtools_sort_bam")
    fc <- featuretypeCounts(bfl = BamFileList(outpaths, yieldSize = 50000),
        grl = feat, singleEnd = FALSE, readlength = NULL, type = "data.frame")
    p <- plotfeaturetypeCounts(x = fc, graphicsfile = "results/featureCounts.png",
        graphicsformat = "png", scales = "fixed", anyreadlength = TRUE,
        scale_length_val = NULL)
}, step_name = "featuretypeCounts", dependency = "genFeatures",
    run_step = "mandatory")
Figure 2: Read distribution plot across annotation features for any read length.


3.5.3 Count and plot reads of specific lengths

To determine the approximate read length of ribosome footprints in Ribo-Seq experiments, one can generate and plot the feature counts for specific read lengths separately. Typically, the most abundant read length obtained for translated features corresponds to the approximate footprint length occupied by the ribosomes.

appendStep(sal) <- LineWise(code = {
    fc2 <- featuretypeCounts(bfl = BamFileList(outpaths, yieldSize = 50000),
        grl = feat, singleEnd = TRUE, readlength = c(74:76, 99:102),
        type = "data.frame")
    p2 <- plotfeaturetypeCounts(x = fc2, graphicsfile = "results/featureCounts2.png",
        graphicsformat = "png", scales = "fixed", anyreadlength = FALSE,
        scale_length_val = NULL)
}, step_name = "featuretypeCounts_length", dependency = "featuretypeCounts",
    run_step = "mandatory")
Figure 3: Read distribution plot across annotation features for specific read lengths.


3.6 Adding custom features to workflow

3.6.1 Predicting uORFs in 5’ UTR regions

The function predORF can be used to identify open reading frames (ORFs) and coding sequences (CDSs) in DNA sequences provided as DNAString or DNAStringSet objects. The setting mode='ORF' returns continuous reading frames that begin with a start codon and end with a stop codon, while mode='CDS' returns continuous reading frames that do not need to begin or end with start or stop codons, respectively. Non-canonical start and stop condons are supported by allowing the user to provide any custom set of triplets under the startcodon and stopcodon arguments (i.e. non-ATG start codons). The argument n defines the maximum number of ORFs to return for each input sequence (e.g. n=1 returns only the longest ORF). It also supports the identification of overlapping and nested ORFs. Alternatively, one can return all non-overlapping ORFs including the longest ORF for each input sequence with n="all" and longest_disjoint=TRUE.

appendStep(sal) <- LineWise(code = {
    txdb <- suppressWarnings(makeTxDbFromGFF(file = "data/tair10.gff",
        format = "gff3", organism = "Arabidopsis"))
    futr <- fiveUTRsByTranscript(txdb, use.names = TRUE)
    dna <- extractTranscriptSeqs(FaFile("data/tair10.fasta"),
        futr)
    uorf <- predORF(dna, n = "all", mode = "orf", longest_disjoint = TRUE,
        strand = "sense")
}, step_name = "pred_ORF", dependency = "featuretypeCounts_length")

To use the predicted ORF ranges for expression analysis given genome alignments as input, it is necessary to scale them to the corresponding genome coordinates. The function scaleRanges does this by transforming the mappings of spliced features (query ranges) to their corresponding genome coordinates (subject ranges). The method accounts for introns in the subject ranges that are absent in the query ranges. The above uORFs predicted in the provided 5’ UTRs sequences using predORF are a typical use case for this application. These query ranges are given relative to the 5’ UTR sequences and scaleRanges will convert them to the corresponding genome coordinates. The resulting GRangesList object (here grl_scaled) can be directly used for read counting.

appendStep(sal) <- LineWise(code = {
    grl_scaled <- scaleRanges(subject = futr, query = uorf, type = "uORF",
        verbose = TRUE)
    export.gff3(unlist(grl_scaled), "results/uorf.gff")
}, step_name = "scale_ranges", dependency = "pred_ORF")

To confirm the correctness of the obtained uORF ranges, one can parse their corresponding DNA sequences from the reference genome with the getSeq function and then translate them with the translate function into proteins. Typically, the returned protein sequences should start with a M (corresponding to start codon) and end with * (corresponding to stop codon). The following example does this for a single uORF containing three exons.

appendStep(sal) <- LineWise(code = {
    translate(unlist(getSeq(FaFile("data/tair10.fasta"), grl_scaled[[7]])))
}, step_name = "translate", dependency = "scale_ranges")

3.6.2 Adding custom features to other feature types

If required custom feature ranges can be added to the standard features generated with the genFeatures function above. The following does this for the uORF ranges predicted with predORF.

appendStep(sal) <- LineWise(code = {
    feat <- genFeatures(txdb, featuretype = "all", reduce_ranges = FALSE)
    feat <- c(feat, GRangesList(uORF = unlist(grl_scaled)))
}, step_name = "add_features", dependency = c("genFeatures",
    "scale_ranges"))

3.6.3 Predicting sORFs in intergenic regions

The following identifies continuous ORFs in intergenic regions. Note, predORF can only identify continuous ORFs in query sequences. The function does not identify and remove introns prior to the ORF prediction.

appendStep(sal) <- LineWise(code = {
    feat <- genFeatures(txdb, featuretype = "intergenic", reduce_ranges = TRUE)
    intergenic <- feat$intergenic
    strand(intergenic) <- "+"
    dna <- getSeq(FaFile("data/tair10.fasta"), intergenic)
    names(dna) <- mcols(intergenic)$feature_by
    sorf <- suppressWarnings(predORF(dna, n = "all", mode = "orf",
        longest_disjoint = TRUE, strand = "both"))
    sorf <- sorf[width(sorf) > 60]  # Remove sORFs below length cutoff, here 60bp
    intergenic <- split(intergenic, mcols(intergenic)$feature_by)
    grl_scaled_intergenic <- scaleRanges(subject = intergenic,
        query = sorf, type = "sORF", verbose = TRUE)
    export.gff3(unlist(grl_scaled_intergenic), "sorf.gff")
    translate(getSeq(FaFile("data/tair10.fasta"), unlist(grl_scaled_intergenic)))
}, step_name = "pred_sORFs", dependency = c("add_features"))

3.7 Genomic read coverage along transripts or CDSs

The featureCoverage function computes the read coverage along single and multi component features based on genomic alignments. The coverage segments of component features are spliced to continuous ranges, such as exons to transcripts or CDSs to ORFs. The results can be obtained with single nucleotide resolution (e.g. around start and stop codons) or as mean coverage of relative bin sizes, such as 100 bins for each feature. The latter allows comparisons of coverage trends among transcripts of variable length. Additionally, the results can be obtained for single or many features (e.g. any number of transcripts) at once. Visualization of the coverage results is facilitated by the downstream plotfeatureCoverage function.

3.7.1 Binned CDS coverage to compare many transcripts

appendStep(sal) <- LineWise(code = {
    grl <- cdsBy(txdb, "tx", use.names = TRUE)
    fcov <- featureCoverage(bfl = BamFileList(outpaths[1:2]),
        grl = grl[1:4], resizereads = NULL, readlengthrange = NULL,
        Nbins = 20, method = mean, fixedmatrix = FALSE, resizefeatures = TRUE,
        upstream = 20, downstream = 20, outfile = "results/featureCoverage.xls",
        overwrite = TRUE)
}, step_name = "binned_CDS_coverage", dependency = c("add_features"))

3.7.2 Coverage upstream and downstream of start and stop codons

appendStep(sal) <- LineWise(code = {
    fcov <- featureCoverage(bfl = BamFileList(outpaths[1:4]),
        grl = grl[1:12], resizereads = NULL, readlengthrange = NULL,
        Nbins = NULL, method = mean, fixedmatrix = TRUE, resizefeatures = TRUE,
        upstream = 20, downstream = 20, outfile = "results/featureCoverage.xls",
        overwrite = TRUE)
    png("./results/coverage_upstream_downstream.png", height = 12,
        width = 24, units = "in", res = 72)
    plotfeatureCoverage(covMA = fcov, method = mean, scales = "fixed",
        extendylim = 2, scale_count_val = 10^6)
    dev.off()
}, step_name = "coverage_upstream_downstream", dependency = c("binned_CDS_coverage"))

3.7.3 Combined coverage for both binned CDS and start/stop codons

appendStep(sal) <- LineWise(code = {
    fcov <- featureCoverage(bfl = BamFileList(outpaths[1:4]),
        grl = grl[1:4], resizereads = NULL, readlengthrange = NULL,
        Nbins = 20, method = mean, fixedmatrix = TRUE, resizefeatures = TRUE,
        upstream = 20, downstream = 20, outfile = "results/featureCoverage.xls",
        overwrite = TRUE)
    png("./results/featurePlot.png", height = 12, width = 24,
        units = "in", res = 72)
    plotfeatureCoverage(covMA = fcov, method = mean, scales = "fixed",
        extendylim = 2, scale_count_val = 10^6)
    dev.off()
}, step_name = "coverage_combined", dependency = c("binned_CDS_coverage",
    "coverage_upstream_downstream"))
Figure 4: Feature coverage plot with single nucleotide resolution around start and stop codons and binned coverage between them.


3.8 Nucleotide level coverage along entire transcripts/CDSs

appendStep(sal) <- LineWise(code = {
    fcov <- featureCoverage(bfl = BamFileList(outpaths[1:2]),
        grl = grl[1], resizereads = NULL, readlengthrange = NULL,
        Nbins = NULL, method = mean, fixedmatrix = FALSE, resizefeatures = TRUE,
        upstream = 20, downstream = 20, outfile = NULL)
}, step_name = "coverage_nuc_level", dependency = c("coverage_combined"))

3.9 Read quantification per annotation range

3.9.1 Read counting with summarizeOverlaps in parallel mode using multiple cores

Reads overlapping with annotation ranges of interest are counted for each sample using the summarizeOverlaps function (Lawrence et al. 2013). The read counting is preformed for exonic gene regions in a non-strand-specific manner while ignoring overlaps among different genes. Subsequently, the expression count values are normalized by (RPKM). The raw read count table (countDFeByg.xls) and the corresponding RPKM table (rpkmDFeByg.xls) are written to separate files in the results directory of this project. Parallelization is achieved with the BiocParallel package, here using 8 CPU cores.

appendStep(sal) <- LineWise(code = {
    txdb <- loadDb("./data/tair10.sqlite")
    eByg <- exonsBy(txdb, by = c("gene"))
    bfl <- BamFileList(outpaths, yieldSize = 50000, index = character())
    multicoreParam <- MulticoreParam(workers = 8)
    register(multicoreParam)
    registered()
    counteByg <- bplapply(bfl, function(x) summarizeOverlaps(eByg,
        x, mode = "Union", ignore.strand = TRUE, inter.feature = FALSE,
        singleEnd = FALSE, BPPARAM = multicoreParam))
    countDFeByg <- sapply(seq(along = counteByg), function(x) assays(counteByg[[x]])$counts)
    rownames(countDFeByg) <- names(rowRanges(counteByg[[1]]))
    colnames(countDFeByg) <- names(bfl)
    rpkmDFeByg <- apply(countDFeByg, 2, function(x) returnRPKM(counts = x,
        ranges = eByg))
    write.table(countDFeByg, "results/countDFeByg.xls", col.names = NA,
        quote = FALSE, sep = "\t")
    write.table(rpkmDFeByg, "results/rpkmDFeByg.xls", col.names = NA,
        quote = FALSE, sep = "\t")
    ## Creating a SummarizedExperiment object
    colData <- data.frame(row.names = SampleName(sal, "hisat2_mapping"),
        condition = getColumn(sal, "hisat2_mapping", position = "targetsWF",
            column = "Factor"))
    colData$condition <- factor(colData$condition)
    countDF_se <- SummarizedExperiment::SummarizedExperiment(assays = countDFeByg,
        colData = colData)
    ## Add results as SummarizedExperiment to the workflow
    ## object
    SE(sal, "read_counting") <- countDF_se
}, step_name = "read_counting", dependency = c("featuretypeCounts"))

When providing a BamFileList as in the example above, summarizeOverlaps methods use by default bplapply and use the register interface from BiocParallel package. If the number of workers is not set, MulticoreParam will use the number of cores returned by parallel::detectCores(). For more information, please check help("summarizeOverlaps") documentation.

Sample of data slice of count table:

read.delim(system.file("extdata/countDFeByg.xls", package = "systemPipeR"),
    row.names = 1, check.names = FALSE)[1:4, 1:5]
##           M1A M1B A1A A1B V1A
## AT1G01010  57 244 201 169 365
## AT1G01020  23  93  69 126 107
## AT1G01030  41  98  73  58  94
## AT1G01040 180 684 522 664 585

Sample of data slice of RPKM table

read.delim(system.file("extdata/rpkmDFeByg.xls", package = "systemPipeR"),
    row.names = 1, check.names = FALSE)[1:4, 1:5]

Note, for most statistical differential expression or abundance analysis methods, such as edgeR or DESeq2, the raw count values should be used as input. The usage of RPKM values should be restricted to specialty applications required by some users, e.g. manually comparing the expression levels among different genes or features.

3.9.2 Sample-wise correlation analysis

The following computes the sample-wise Spearman correlation coefficients from the rlog transformed expression values generated with the DESeq2 package. After transformation to a distance matrix, hierarchical clustering is performed with the hclust function and the result is plotted as a dendrogram and written to a file named sample_tree.png in the results directory.

appendStep(sal) <- LineWise(code = {
    ## Extracting SummarizedExperiment object
    se <- SE(sal, "read_counting")
    dds <- DESeqDataSet(se, design = ~condition)
    d <- cor(assay(rlog(dds)), method = "spearman")
    hc <- hclust(dist(1 - d))
    png("results/sample_tree.png")
    plot.phylo(as.phylo(hc), type = "p", edge.col = "blue", edge.width = 2,
        show.node.label = TRUE, no.margin = TRUE)
    dev.off()
}, step_name = "sample_tree", dependency = "read_counting")
Figure 5: Correlation dendrogram of samples.


3.10 Analysis of differentially expressed genes with edgeR

The analysis of differentially expressed genes (DEGs) is performed with the glm method from the edgeR package (Robinson, McCarthy, and Smyth 2010). The sample comparisons used by this analysis are defined in the header lines of the targetsPE.txt file starting with <CMP>.

appendStep(sal) <- LineWise(code = {
    countDF <- read.delim("results/countDFeByg.xls", row.names = 1,
        check.names = FALSE)
    cmp <- readComp(stepsWF(sal)[["hisat2_mapping"]], format = "matrix",
        delim = "-")
    edgeDF <- run_edgeR(countDF = countDF, targets = targetsWF(sal)[["hisat2_mapping"]],
        cmp = cmp[[1]], independent = FALSE, mdsplot = "")
}, step_name = "run_edgeR", dependency = "read_counting")

Add functional gene descriptions, here from biomaRt.

appendStep(sal) <- LineWise(code = {
    m <- useMart("plants_mart", dataset = "athaliana_eg_gene",
        host = "https://plants.ensembl.org")
    desc <- getBM(attributes = c("tair_locus", "description"),
        mart = m)
    desc <- desc[!duplicated(desc[, 1]), ]
    descv <- as.character(desc[, 2])
    names(descv) <- as.character(desc[, 1])
    edgeDF <- data.frame(edgeDF, Desc = descv[rownames(edgeDF)],
        check.names = FALSE)
    write.table(edgeDF, "./results/edgeRglm_allcomp.xls", quote = FALSE,
        sep = "\t", col.names = NA)
}, step_name = "custom_annot", dependency = "run_edgeR")

3.10.1 Plot DEG results

Filter and plot DEG results for up and down regulated genes. The definition of up and down is given in the corresponding help file. To open it, type ?filterDEGs in the R console.

appendStep(sal) <- LineWise(code = {
    edgeDF <- read.delim("results/edgeRglm_allcomp.xls", row.names = 1,
        check.names = FALSE)
    png("results/DEGcounts.png")
    DEG_list <- filterDEGs(degDF = edgeDF, filter = c(Fold = 2,
        FDR = 20))
    dev.off()
    write.table(DEG_list$Summary, "./results/DEGcounts.xls",
        quote = FALSE, sep = "\t", row.names = FALSE)
}, step_name = "filter_degs", dependency = "custom_annot")
Figure 6: Up and down regulated DEGs.


3.10.2 Venn diagrams of DEG sets

The overLapper function can compute Venn intersects for large numbers of sample sets (up to 20 or more) and plots 2-5 way Venn diagrams. A useful feature is the possibility to combine the counts from several Venn comparisons with the same number of sample sets in a single Venn diagram (here for 4 up and down DEG sets).

appendStep(sal) <- LineWise(code = {
    vennsetup <- overLapper(DEG_list$Up[6:9], type = "vennsets")
    vennsetdown <- overLapper(DEG_list$Down[6:9], type = "vennsets")
    png("results/vennplot.png")
    vennPlot(list(vennsetup, vennsetdown), mymain = "", mysub = "",
        colmode = 2, ccol = c("blue", "red"))
    dev.off()
}, step_name = "venn_diagram", dependency = "filter_degs")
Figure 7: Venn Diagram for 4 Up and Down DEG Sets


3.11 GO term enrichment analysis of DEGs

3.11.1 Obtain gene-to-GO mappings

The following shows how to obtain gene-to-GO mappings from biomaRt (here for A. thaliana) and how to organize them for the downstream GO term enrichment analysis. Alternatively, the gene-to-GO mappings can be obtained for many organisms from Bioconductor’s *.db genome annotation packages or GO annotation files provided by various genome databases. For each annotation this relatively slow preprocessing step needs to be performed only once. Subsequently, the preprocessed data can be loaded with the load function as shown in the next subsection.

appendStep(sal) <- LineWise(code = {
    # listMarts() # To choose BioMart database
    # listMarts(host='plants.ensembl.org') m <-
    # useMart('plants_mart',
    # host='https://plants.ensembl.org') listDatasets(m)
    m <- useMart("plants_mart", dataset = "athaliana_eg_gene",
        host = "https://plants.ensembl.org")
    # listAttributes(m) # Choose data types you want to
    # download
    go <- getBM(attributes = c("go_id", "tair_locus", "namespace_1003"),
        mart = m)
    go <- go[go[, 3] != "", ]
    go[, 3] <- as.character(go[, 3])
    go[go[, 3] == "molecular_function", 3] <- "F"
    go[go[, 3] == "biological_process", 3] <- "P"
    go[go[, 3] == "cellular_component", 3] <- "C"
    go[1:4, ]
    if (!dir.exists("./data/GO"))
        dir.create("./data/GO")
    write.table(go, "data/GO/GOannotationsBiomart_mod.txt", quote = FALSE,
        row.names = FALSE, col.names = FALSE, sep = "\t")
    catdb <- makeCATdb(myfile = "data/GO/GOannotationsBiomart_mod.txt",
        lib = NULL, org = "", colno = c(1, 2, 3), idconv = NULL)
    save(catdb, file = "data/GO/catdb.RData")
}, step_name = "get_go_annot", dependency = "filter_degs")

3.11.2 Batch GO term enrichment analysis

Apply the enrichment analysis to the DEG sets obtained the above differential expression analysis. Note, in the following example the FDR filter is set here to an unreasonably high value, simply because of the small size of the toy data set used in this vignette. Batch enrichment analysis of many gene sets is performed with the function. When method=all, it returns all GO terms passing the p-value cutoff specified under the cutoff arguments. When method=slim, it returns only the GO terms specified under the myslimv argument. The given example shows how a GO slim vector for a specific organism can be obtained from BioMart.

appendStep(sal) <- LineWise(code = {
    load("data/GO/catdb.RData")
    DEG_list <- filterDEGs(degDF = edgeDF, filter = c(Fold = 2,
        FDR = 50), plot = FALSE)
    up_down <- DEG_list$UporDown
    names(up_down) <- paste(names(up_down), "_up_down", sep = "")
    up <- DEG_list$Up
    names(up) <- paste(names(up), "_up", sep = "")
    down <- DEG_list$Down
    names(down) <- paste(names(down), "_down", sep = "")
    DEGlist <- c(up_down, up, down)
    DEGlist <- DEGlist[sapply(DEGlist, length) > 0]
    BatchResult <- GOCluster_Report(catdb = catdb, setlist = DEGlist,
        method = "all", id_type = "gene", CLSZ = 2, cutoff = 0.9,
        gocats = c("MF", "BP", "CC"), recordSpecGO = NULL)
    m <- useMart("plants_mart", dataset = "athaliana_eg_gene",
        host = "https://plants.ensembl.org")
    goslimvec <- as.character(getBM(attributes = c("goslim_goa_accession"),
        mart = m)[, 1])
    BatchResultslim <- GOCluster_Report(catdb = catdb, setlist = DEGlist,
        method = "slim", id_type = "gene", myslimv = goslimvec,
        CLSZ = 10, cutoff = 0.01, gocats = c("MF", "BP", "CC"),
        recordSpecGO = NULL)
}, step_name = "go_enrich", dependency = "get_go_annot")

3.11.3 Plot batch GO term results

The data.frame generated by GOCluster can be plotted with the goBarplot function. Because of the variable size of the sample sets, it may not always be desirable to show the results from different DEG sets in the same bar plot. Plotting single sample sets is achieved by subsetting the input data frame as shown in the first line of the following example.

appendStep(sal) <- LineWise(code = {
    gos <- BatchResultslim[grep("M6-V6_up_down", BatchResultslim$CLID),
        ]
    gos <- BatchResultslim
    png("results/GOslimbarplotMF.png", height = 8, width = 10)
    goBarplot(gos, gocat = "MF")
    goBarplot(gos, gocat = "BP")
    goBarplot(gos, gocat = "CC")
    dev.off()
}, step_name = "go_plot", dependency = "go_enrich")
Figure 8: GO Slim Barplot for MF Ontology.


3.12 Differential ribosome loading analysis (translational efficiency)

Combined with mRNA-Seq data, Ribo-Seq or polyRibo-Seq experiments can be used to study changes in translational efficiencies of genes and/or transcripts for different treatments. For test purposes the following generates a small test data set from the sample data used in this vignette, where two types of RNA samples (assays) are considered: polyribosomal mRNA (Ribo) and total mRNA (mRNA). In addition, there are two treatments (conditions): M1 and A1.

appendStep(sal) <- LineWise(code = {
    countDFeByg <- read.delim("results/countDFeByg.xls", row.names = 1,
        check.names = FALSE)
    coldata <- S4Vectors::DataFrame(assay = factor(rep(c("Ribo",
        "mRNA"), each = 4)), condition = factor(rep(as.character(targetsWF(sal)[["hisat2_mapping"]]$Factor[1:4]),
        2)), row.names = as.character(targetsWF(sal)[["hisat2_mapping"]]$SampleName)[1:8])
    coldata
}, step_name = "diff_loading", dependency = "go_plot")

Differences in translational efficiencies can be calculated by ratios of ratios for the two conditions:

\[(Ribo\_A1 / mRNA\_A1) / (Ribo\_M1 / mRNA\_M1)\]

The latter can be modeled with the DESeq2 package using the design \(\sim assay + condition + assay:condition\), where the interaction term \(assay:condition\) represents the ratio of ratios. Using the likelihood ratio test of DESeq2, which removes the interaction term in the reduced model, one can test whether the translational efficiency (ribosome loading) is different in condition A1 than in M1.

appendStep(sal) <- LineWise(code = {
    dds <- DESeq2::DESeqDataSetFromMatrix(countData = as.matrix(countDFeByg[,
        rownames(coldata)]), colData = coldata, design = ~assay +
        condition + assay:condition)
    # model.matrix(~ assay + condition + assay:condition,
    # coldata) # Corresponding design matrix
    dds <- DESeq2::DESeq(dds, test = "LRT", reduced = ~assay +
        condition)
    res <- DESeq2::results(dds)
    head(res[order(res$padj), ], 4)
    write.table(res, file = "transleff.xls", quote = FALSE, col.names = NA,
        sep = "\t")
}, step_name = "diff_translational_eff", dependency = "diff_loading")

3.13 Clustering and heat maps

The following example performs hierarchical clustering on the rlog transformed expression matrix subsetted by the DEGs identified in the above differential expression analysis. It uses a Pearson correlation-based distance measure and complete linkage for cluster joining.

appendStep(sal) <- LineWise(code = {
    geneids <- unique(as.character(unlist(DEG_list[[1]])))
    y <- assay(rlog(dds))[geneids, ]
    y <- y[rowSums(y[]) > 0, ]
    png("results/heatmap1.png")
    pheatmap(y, scale = "row", clustering_distance_rows = "correlation",
        clustering_distance_cols = "correlation")
    dev.off()
}, step_name = "heatmap", dependency = "diff_translational_eff")
Figure 9: Heat map with hierarchical clustering dendrograms of DEGs


3.14 Version Information

appendStep(sal) <- LineWise(code = {
    sessionInfo()
}, step_name = "sessionInfo", dependency = "heatmap")

4 Running workflow

4.1 Interactive job submissions in a single machine

For running the workflow, runWF function will execute all the steps store in the workflow container. The execution will be on a single machine without submitting to a queuing system of a computer cluster.

sal <- runWF(sal, run_step = "mandatory")

4.2 Parallelization on clusters

Alternatively, the computation can be greatly accelerated by processing many files in parallel using several compute nodes of a cluster, where a scheduling/queuing system is used for load balancing.

The resources list object provides the number of independent parallel cluster processes defined under the Njobs element in the list. The following example will run 18 processes in parallel using each 4 CPU cores. If the resources available on a cluster allow running all 18 processes at the same time, then the shown sample submission will utilize in a total of 72 CPU cores.

Note, runWF can be used with most queueing systems as it is based on utilities from the batchtools package, which supports the use of template files (*.tmpl) for defining the run parameters of different schedulers. To run the following code, one needs to have both a conffile (see .batchtools.conf.R samples here) and a template file (see *.tmpl samples here) for the queueing available on a system. The following example uses the sample conffile and template files for the Slurm scheduler provided by this package.

The resources can be appended when the step is generated, or it is possible to add these resources later, as the following example using the addResources function:

# wall time in mins, memory in MB
resources <- list(conffile = ".batchtools.conf.R", template = "batchtools.slurm.tmpl",
    Njobs = 18, walltime = 120, ntasks = 1, ncpus = 4, memory = 1024,
    partition = "short")
sal <- addResources(sal, c("hisat2_mapping"), resources = resources)
sal <- runWF(sal, run_step = "mandatory")

4.3 Visualize workflow

systemPipeR workflows instances can be visualized with the plotWF function.

plotWF(sal, rstudio = TRUE)

4.4 Checking workflow status

To check the summary of the workflow, we can use:

sal
statusWF(sal)

4.5 Accessing logs report

systemPipeR compiles all the workflow execution logs in one central location, making it easier to check any standard output (stdout) or standard error (stderr) for any command-line tools used on the workflow or the R code stdout.

sal <- renderLogs(sal)

4.6 Tools used

To check command-line tools used in this workflow, use listCmdTools, and use listCmdModules to check if you have a modular system.

The following code will print out tools required in your custom SPR project in the report. In case you are running the workflow for the first and do not have a project yet, or you just want to browser this workflow, following code displays the tools required by default.

if (file.exists(file.path(".SPRproject", "SYSargsList.yml"))) {
    local({
        sal <- systemPipeR::SPRproject(resume = TRUE)
        systemPipeR::listCmdTools(sal)
        systemPipeR::listCmdModules(sal)
    })
} else {
    cat(crayon::blue$bold("Tools and modules required by this workflow are:\n"))
    cat(c("hisat2/2.1.0", "samtools/1.14"), sep = "\n")
}
## Tools and modules required by this workflow are:
## hisat2/2.1.0
## samtools/1.14

4.7 Report Information

This is the session information for rendering this report. To access the session information of workflow running, check HTML report of renderLogs.

sessionInfo()
## R version 4.4.0 beta (2024-04-15 r86425)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 22.04.4 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.19-bioc/R/lib/libRblas.so 
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.10.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_GB              LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: America/New_York
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats4    stats     graphics  grDevices utils    
## [6] datasets  methods   base     
## 
## other attached packages:
##  [1] systemPipeR_2.10.0          ShortRead_1.62.0           
##  [3] GenomicAlignments_1.40.0    SummarizedExperiment_1.34.0
##  [5] Biobase_2.64.0              MatrixGenerics_1.16.0      
##  [7] matrixStats_1.3.0           BiocParallel_1.38.0        
##  [9] Rsamtools_2.20.0            Biostrings_2.72.0          
## [11] XVector_0.44.0              GenomicRanges_1.56.0       
## [13] GenomeInfoDb_1.40.0         IRanges_2.38.0             
## [15] S4Vectors_0.42.0            BiocGenerics_0.50.0        
## [17] BiocStyle_2.32.0           
## 
## loaded via a namespace (and not attached):
##  [1] gtable_0.3.5            xfun_0.43              
##  [3] bslib_0.7.0             hwriter_1.3.2.1        
##  [5] ggplot2_3.5.1           htmlwidgets_1.6.4      
##  [7] latticeExtra_0.6-30     lattice_0.22-6         
##  [9] generics_0.1.3          vctrs_0.6.5            
## [11] tools_4.4.0             bitops_1.0-7           
## [13] parallel_4.4.0          fansi_1.0.6            
## [15] tibble_3.2.1            pkgconfig_2.0.3        
## [17] Matrix_1.7-0            RColorBrewer_1.1-3     
## [19] lifecycle_1.0.4         GenomeInfoDbData_1.2.12
## [21] stringr_1.5.1           compiler_4.4.0         
## [23] deldir_2.0-4            munsell_0.5.1          
## [25] codetools_0.2-20        htmltools_0.5.8.1      
## [27] sass_0.4.9              yaml_2.3.8             
## [29] pillar_1.9.0            crayon_1.5.2           
## [31] jquerylib_0.1.4         DelayedArray_0.30.0    
## [33] cachem_1.0.8            abind_1.4-5            
## [35] tidyselect_1.2.1        digest_0.6.35          
## [37] stringi_1.8.3           dplyr_1.1.4            
## [39] bookdown_0.39           fastmap_1.1.1          
## [41] grid_4.4.0              colorspace_2.1-0       
## [43] cli_3.6.2               SparseArray_1.4.0      
## [45] magrittr_2.0.3          S4Arrays_1.4.0         
## [47] utf8_1.2.4              UCSC.utils_1.0.0       
## [49] scales_1.3.0            rmarkdown_2.26         
## [51] pwalign_1.0.0           httr_1.4.7             
## [53] jpeg_0.1-10             interp_1.1-6           
## [55] png_0.1-8               evaluate_0.23          
## [57] knitr_1.46              rlang_1.1.3            
## [59] Rcpp_1.0.12             glue_1.7.0             
## [61] BiocManager_1.30.22     formatR_1.14           
## [63] jsonlite_1.8.8          R6_2.5.1               
## [65] zlibbioc_1.50.0

5 Funding

This research was funded by National Science Foundation Grants IOS-0750811 and MCB-1021969, and a Marie Curie European Economic Community Fellowship PIOF-GA-2012-327954.

References

Aspden, Julie L, Ying Chen Eyre-Walker, Rose J Phillips, Unum Amin, Muhammad Ali S Mumtaz, Michele Brocard, and Juan-Pablo Couso. 2014. “Extensive Translation of Small Open Reading Frames Revealed by Poly-Ribo-Seq.” Elife 3: e03528. https://doi.org/10.7554/eLife.03528.

H Backman, Tyler W, and Thomas Girke. 2016. “systemPipeR: NGS workflow and report generation environment.” BMC Bioinformatics 17 (1): 388. https://doi.org/10.1186/s12859-016-1241-0.

Ingolia, N T, S Ghaemmaghami, J R Newman, and J S Weissman. 2009. “Genome-Wide Analysis in Vivo of Translation with Nucleotide Resolution Using Ribosome Profiling.” Science 324 (5924): 218–23. https://doi.org/10.1016/j.ymeth.2009.03.016.

Ingolia, N T, L F Lareau, and J S Weissman. 2011. “Ribosome Profiling of Mouse Embryonic Stem Cells Reveals the Complexity and Dynamics of Mammalian Proteomes.” Cell 147 (4): 789–802. https://doi.org/10.1016/j.cell.2011.10.002.

Juntawong, Piyada, Thomas Girke, Jérémie Bazin, and Julia Bailey-Serres. 2014. “Translational Dynamics Revealed by Genome-Wide Profiling of Ribosome Footprints in Arabidopsis.” Proc. Natl. Acad. Sci. U. S. A. 111 (1): E203–12. https://doi.org/10.1073/pnas.1317811111.

Juntawong, Piyada, Maureen Hummel, Jeremie Bazin, and Julia Bailey-Serres. 2015. “Ribosome Profiling: A Tool for Quantitative Evaluation of Dynamics in mRNA Translation.” Methods Mol. Biol. 1284: 139–73. https://doi.org/10.1007/978-1-4939-2444-8\_7.

Kim, Daehwan, Ben Langmead, and Steven L Salzberg. 2015. “HISAT: A Fast Spliced Aligner with Low Memory Requirements.” Nat. Methods 12 (4): 357–60.

Lawrence, Michael, Wolfgang Huber, Hervé Pagès, Patrick Aboyoun, Marc Carlson, Robert Gentleman, Martin T Morgan, and Vincent J Carey. 2013. “Software for Computing and Annotating Genomic Ranges.” PLoS Comput. Biol. 9 (8): e1003118. https://doi.org/10.1371/journal.pcbi.1003118.

Robinson, M D, D J McCarthy, and G K Smyth. 2010. “EdgeR: A Bioconductor Package for Differential Expression Analysis of Digital Gene Expression Data.” Bioinformatics 26 (1): 139–40. https://doi.org/10.1093/bioinformatics/btp616.