CAGEr: an R package for CAGE (Cap Analysis of Gene Expression) data analysis and promoterome mining

In CAGEr: Analysis of CAGE (Cap Analysis of Gene Expression) sequencing data for precise mapping of transcription start sites and promoterome mining

options(signif = 3, digits = 3)
knitr::opts_chunk$set(tidy = FALSE, cache = TRUE, autodep = TRUE, fig.height = 5.5,
                      message = FALSE, error = FALSE, warning = TRUE)
set.seed(0xdada)

Introduction

This document briefly how to use r Biocpkg("CAGEr") CAGEr, a Bioconductor package designed to process, analyse and visualise Cap Analysis of Gene Expression (CAGE)sequencing data. CAGE [@Kodzius:2006] is a high-throughput method for transcriptome analysis that utilizes cap trapping [@Carninci:1996], a technique based on the biotinylation of the 7-methylguanosine cap of Pol II transcripts, to pulldown the 5′-complete cDNAs reversely transcribed from the captured transcripts. A linker sequence is ligated to the 5′ end of the cDNA and a specific restriction enzyme is used to cleave off a short fragment from the 5′ end. Resulting fragments are then amplified and sequenced using massive parallel high-throughput sequencing technology, which results in a large number of short sequenced tags that can be mapped back to the referent genome to infer the exact position of the transcription start sites (TSSs) used for transcription of captured RNAs (Figure \@ref(fig:CAGEprotocol)). The number of CAGE tags supporting each TSS gives the information on the relative frequency of its usage and can be used as a measure of expression from that specific TSS. Thus, CAGE provides information on two aspects of capped transcriptome: genome-wide 1bp-resolution map of TSSs and transcript expression levels. This information can be used for various analyses, from 5′ centered expression profiling [@Takahashi:2012] to studying promoter architecture [@Carninci:2006].

knitr::include_graphics("images/CAGEprotocol.svg")

CAGE samples derived from various organisms (genomes) can be analysed by CAGEr and the only limitation is the availability of the referent genome as a r Biocpkg("BSgenome") package in case when raw mapped CAGE tags are processed. CAGEr provides a comprehensive workflow that starts from mapped CAGE tags and includes reconstruction of TSSs and promoters and their visualisation, as well as more specialized downstream analyses like promoter width, expression profiling and differential TSS usage. It can use both Binary Sequence Alignment Map (BAM) files of aligned CAGE tags or files with genomic locations of TSSs and number of supporting CAGE tags as input. If BAM files are provided CAGEr constructs TSSs from aligned CAGE tags and counts the number of tags supporting each TSS, while allowing filtering out low-quality tags and removing technology-specific bias. It further performs normalization of raw CAGE tag count, clustering of TSSs into tag clusters (TC) and their aggregation across multiple CAGE experiments into promoters to construct the promoterome. Various methods for normalization and clustering of TSSs are supported. Exporting data into different types of track files allows various visualisations of TSSs and clusters (promoters) in the UCSC Genome Browser, which facilitate generation of hypotheses. CAGEr manipulates multiple CAGE experiments at once and performs analyses across datasets, including expression profiling and detection of differential TSS usage (promoter shifting). Multicore option for parallel processing is supported on Unix-like platforms, which significantly reduces computing time.

Here are some of the functionalities provided in this package:

• Reading in multiple CAGE datasets from various sources; user provided BAM or TSS input files, public CAGE datasets from accompanying data package.

• Correcting systematic G nucleotide addition bias at the 5′ end of CAGE tags.

• Plotting pairwise scatter plots, calculating correlation between datasets and merging datasets.

• Normalizing raw CAGE tag count: simple tag per million (tpm) or power-law based normalization [@Balwierz:2009].

• Clustering individual TSSs into tag clusters (TCs) and aggregating clusters across multiple CAGE datasets to create a set of consensus promoters.

• Making bedGraph or BED files of individual TSSs or clusters for visualisation in the genome browser.

• Expression clustering of individual TSSs or consensus promoters into distinct expression profiles using common clustering algorithms.

• Calculating promoter width based on the cumulative distribution of CAGE signal along the promoter.

• Scoring and statistically testing differential TSS usage (promoter shifting) and detecting promoters that shift between two samples.

Several data packages are accompanying CAGEr package. They contain majority of the up-to-date publicly available CAGE data produced by major consortia including FANTOM and ENCODE. These include r Biocpkg("FANTOM3and4CAGE") package available from Bioconductor, as well as r Biocpkg("ENCODEprojectCAGE") and r Biocpkg("ZebrafishDevelopmentalCAGE") packages available from http://promshift.genereg.net/CAGEr/. In addition, direct fetching of TSS data from FANTOM5 web resource (the largest collection of TSS data for human and mouse) from within CAGEr is also available. These are all valuable resources of genome-wide TSSs in various tissue/cell types for various model organisms that can be used directly in R. Section 5 in this vignette describes how these public datasets can be included into a workflow provided by CAGEr. For further information on the content of the data packages and the list of available CAGE datasets please refer to the vignette of the corresponding data package.

For further details on the implemented methods and for citing the CAGEr package in your work please refer to [@Haberle:2015].

Input data for CAGEr {#input-formats}

CAGEr package supports three types of CAGE data input:

• Sequenced CAGE tags mapped to the genome: either BAM (Binary Sequence Alignment Map) files of sequenced CAGE tags aligned to the referent genome (including the paired-end data such as CAGEscan) or BED files of CAGE tags (fragments).

• CAGE detected TSSs (CTSSs): tab separated files with genomic coordinates of TSSs and number of tags supporting each TSS. The file should not contain a header and the data must be organized in four columns:

  • name of the chromosome: names must match the names of chromosomes in the corresponding BSgenome package.
  • 1-based coordinate of the TSS on the chromosome
  • genomic strand: should be either + or -
  • number of CAGE tags supporting that TSS

• Publicly available CAGE datasets from R data package: Several data packages containing CAGE data for various organisms produced by major consortia are accompanying this package. Selected subset of these data can be used as input for \Rpackage{CAGEr}.

The type and the format of the input files is specified at the beginning of the workflow, when the CAGEset object is created (section 4.2). This is done by setting the inputFilesType argument, which accepts the following self-explanatory options referring to formats mentioned above: "bam", "bamPairedEnd", "bed", "ctss", "CTSStable".

In addition, the package provides a method for coercing a data.frame object containing single base-pair TSS information into a CAGEset object (as described in section 4.2.1), which can be further used in the workflow described below.

The CAGEr workflow

Getting started

We start the workflow by creating a CAGEexp object, which is a container for storing CAGE datasets and all the results that will be generated by applying specific functions. The CAGEexp objects are an extension of the r Biocpkg("MultiAssayExperiment") class, and therefore can use all their methods. The expression data is stored in CAGEexp using r Biocpkg("SummarizedExperiment") objects, and can also access their methods.

To load the CAGEr package and the other libraries into your R envirnoment type:

library(CAGEr)

Creating a CAGEexp object {#create-CAGEexp}

Specifying a genome assembly

In this tutorial we will be using data from zebrafish Danio rerio that was mapped to the danRer7 assembly of the genome. Therefore, the corresponding genome package r Biocpkg("BSgenome.Drerio.UCSC.danRer7") has to be installed. It will be automatically loaded by CAGEr commands when needed.

In case the data is mapped to a genome that is not readily available through BSgenome package (not in the list returned by BSgenome::available.genomes() function), a custom BSgenome package has to be build and installed first. (See the vignette within the BSgenome package for instructions on how to build a custom genome package). The genomeName argument can then be set to the name of the build genome package when creating a CAGEexp object (see the section Creating CAGEexp object below).

The BSgenome package is used to access information about the chromosomes (sequence, length, circularity). By default, CAGEr will discard alignments that are not on chromosomes named in the BSgenome package. The genomeName argument can be set to NULL in order to suppress this behaviour, or as a last resort when no BSgenome package is available. However, CAGEr functions that need access to the genome sequence, for instance for G-correction will not work in that case.

Specifying input files

The subset of zebrafish (Danio rerio) developmental time-series CAGE data generated by [@Nepal:2013] will be used in the following demonstration of the CAGEr workflow.

Files with genomic coordinates of TSSs detected by CAGE in 4 zebrafish developmental stages are included in this package in the extdata subdirectory. The files contain TSSs from a part of chromosome 17 (26,000,000-46,000,000), and there are two files for one of the developmental stages (two independent replicas). The data in files is organized in four tab-separated columns as described above in section \@ref(input-formats).

inputFiles = list.files( system.file("extdata", package = "CAGEr")
                       , "ctss$"
                       , full.names = TRUE)

Creating the object

The CAGEexp object is crated with the CAGEexp contstructor, that requries information on file path an type, sample names and reference genome name.

ce <- CAGEexp( genomeName     = "BSgenome.Drerio.UCSC.danRer7"
             , inputFiles     = inputFiles
             , inputFilesType = "ctss"
             , sampleLabels   = sub( ".chr17.ctss", "", basename(inputFiles))
)

To display the created object type:

ce

The supplied information can be seen in the Input data information section, whereas all other slots are still empty, since no data has been read yet and no analysis conducted.

Reading in the data

In case when the CAGE / TSS data is to be read from input files, an empty CAGEexp object with information about the files is first created as described above in section \@ref(create-CAGEexp). To actually read in the data into the object we use getCTSS() function, that will add an experiment called tagCountMatrix to the CAGEexp object.

getCTSS(ce)
ce

This function reads the provided files in the order they were specified in the inputFiles argument. It creates a single set of all TSSs detected across all input datasets (union of TSSs) and a table with counts of CAGE tags supporting each TSS in every dataset. (Note that in case when a CAGEr object is created by coercion from an existing expression table there is no need to call getCTSS()).

Genomic coordinates of all TSSs and numbers of supporting CAGE tags in every input sample can be retrieved using the CTSStagCountSE() function. CTSScoordinatesGR() accesses the CTSS coordinates and CTSStagCountDF() accesses the CTSS expression values.^[Data can also be accessed directly using the native methods of the MultiAssayExperiment and SummarizedExperiment classes, for example ce[["tagCountMatrix"]], rowRanges(ce[["tagCountMatrix"]]) and assay(ce[["tagCountMatrix"]]).]

CTSStagCountSE(ce)
CTSScoordinatesGR(ce)
CTSStagCountDF(ce)

For compatiblity with earlier CAGEr works using CAGEset objects, and to provide simpler data formats, the coordinates and expression values can also be accessed as simple data.frames. Note howerver that with large data sets it can cause extreme performance issues.

head(CTSScoordinates(ce))
head(CTSStagCountDf(ce))
head(CTSStagCount(ce))

Note that the samples are ordered in the way they were supplied when creating the CAGEexp object and will be presented in that order in all the results and plots. To check sample labels and their ordering type:

sampleLabels(ce)

In addition, a colour is assigned to each sample, which is consistently used to depict that sample in all the plots. By default a rainbow palette of colours is used and the hexadecimal format of the assigned colours can be seen as names attribute of sample labels shown above. The colours can be changed to taste at any point in the workflow using the setColors() function.

Quality controls and preliminary analyses

Genome annotations

By design, CAGE tags map transcription start sites and therefore detect promoters. Quantitatively, the proportion of tags that map to promoter regions will depend both on the quality of the libraries and the quality of the genome annotation, which may be incomplete. Nevertheless, strong variations between libraries prepared in the same experiment may be used for quality controls.

CAGEr can intersect CTSSes with reference transcript models and annotate them with the name(s) of the models, and the region categories promoter, exon, intron and unknown, by using the annotateCTSS function. The reference models can be GENCODE loaded with the import.gff function of the r Biocpkg("rtracklayer") package, or any other input that has the same structure, see help("annotateCTSS") for details. In this example, we will use a sample annotation for zebrafish (see help("exampleZv9_annot")).

annotateCTSS(ce, exampleZv9_annot)

The annotation results are stored as tag counts in the sample metadata, and as new columns in the CTSS genomic ranges

colData(ce)[,c("librarySizes", "promoter", "exon", "intron", "unknown")]
CTSScoordinatesGR(ce)

A function plotAnnot is provided to plot the annotations as stacked bar plots. Here, all the CAGE libraries look very promoter-specific.

plotAnnot(ce, "counts")

Correlation between samples

As part of the basic sanity checks, we can explore the data by looking at the correlation between the samples. The plotCorrelation2() function will plot pairwise scatter plots of expression scores per TSS or consensus cluster and calculate correlation coefficients between all possible pairs of samples^[Alternatively, the plotCorrelation() function does the same and colors the scatterplots according to point density, but is much slower.]. A threshold can be set, so that only regions with an expression score (raw or normalized) above the threshold (either in one or both samples) are considered when calculating correlation. Three different correlation measures are supported: Pearson's, Spearman's and Kendall's correlation coefficients. Note that while the scatterplots are on a logarithmic scale with pseudocount added to the zero values, the correlation coefficients are calculated on untransformed (but thresholded) data.

corr.m <- plotCorrelation2( ce, samples = "all"
                          , tagCountThreshold = 1, applyThresholdBoth = FALSE
                          , method = "pearson")

Merging of replicates

Based on calculated correlation we might want to merge and/or rearrange some of the datasets. To rearrange the samples in the temporal order of the zebrafish development (unfertilized egg -> high -> 30 percent dome -> prim6) and to merge the two replicas for the prim6 developmental stage we use the mergeSamples() function:

mergeSamples(ce, mergeIndex = c(3,2,4,4,1), 
            mergedSampleLabels = c("zf_unfertilized_egg", "zf_high", "zf_30p_dome", "zf_prim6"))
annotateCTSS(ce, exampleZv9_annot)

The mergeIndex argument controls which samples will be merged and how the final dataset will be ordered. Samples labeled by the same number (in our case samples three and four) will be merged together by summing number of CAGE tags per TSS. The final set of samples will be ordered in the ascending order of values provided in mergeIndex and will be labeled by the labels provided in the mergedSampleLabels argument. Note that mergeSamples function resets all slots with results of downstream analyses, so in case there were any results in the CAGEexp object prior to merging, they will be removed. Thus, annotation has to be redone.

Normalization

Library sizes (number of total sequenced tags) of individual experiments differ, thus normalization is required to make them comparable. The librarySizes function returns the total number of CAGE tags in each sample:

librarySizes(ce)

The CAGEr package supports both simple tags per million normalization and power-law based normalization. It has been shown that many CAGE datasets follow a power-law distribution [@Balwierz:2009]. Plotting the number of CAGE tags (X-axis) against the number of TSSs that are supported by <= of that number of tags (Y-axis) results in a distribution that can be approximated by a power-law. On a log-log scale this reverse cumulative distribution will manifest as a monotonically decreasing linear function, which can be defined as

$$y = -1 * \alpha * x + \beta$$

and is fully determined by the slope $\alpha$ and total number of tags T (which together with $\alpha$ determines the value of $\beta$).

To check whether our CAGE datasets follow power-law distribution and in which range of values, we can use the plotReverseCumulatives function:

plotReverseCumulatives(ce, fitInRange = c(5, 1000), onePlot = TRUE)

In addition to the reverse cumulative plots (Figure \@ref(fig:ReverseCumulatives)), a power-law distribution will be fitted to each reverse cumulative using values in the specified range (denoted with dashed lines in Figure \@ref(fig:ReverseCumulatives)) and the value of $\alpha$ will be reported for each sample (shown in the brackets in the Figure \@ref(fig:ReverseCumulatives) legend). The plots can help in choosing the optimal parameters for power-law based normalization. We can see that the reverse cumulative distributions look similar and follow the power-law in the central part of the CAGE tag counts values with a slope between -1.1 and -1.3. Thus, we choose a range from 5 to 1000 tags to fit a power-law, and we normalize all samples to a referent power-law distribution with a total of 50,000 tags and slope of -1.2 ($\alpha = 1.2$).^[Note that since this example dataset contains only data from one part of chromosome 17 and the total number of tags is very small, we normalize to a referent distribution with a similarly small number of tags. When analyzing full datasets it is reasonable to set total number of tags for referent distribution to one million to get normalized tags per million values.]

To perform normalization we pass these parameters to the normalizeTagCount function.

normalizeTagCount(ce, method = "powerLaw", fitInRange = c(5, 1000), alpha = 1.2, T = 5*10^4)
ce[["tagCountMatrix"]]

The normalization is performed as described in [@Balwierz:2009]:

• Power-law is fitted to the reverse cumulative distribution in the specified range of CAGE tags values to each sample separately.
• A referent power-law distribution is defined based on the provided alpha (slope in the log-log representation) and T (total number of tags) parameters. Setting T to 1 million results in normalized tags per million (tpm) values.
• Every sample is normalized to the defined referent distribution, i.e. given the parameters that approximate its own power-law distribution it is calculated how many tags would each TSS have in the referent power-law distribution.

In addition to the two provided normalization methods, a pass-through option none can be set as method parameter to keep using raw tag counts in all downstream steps. Note that normalizeTagCount() has to be applied to CAGEr object before moving to next steps. Thus, in order to keep using raw tag counts run the function with method="none". In that case, all results and parameters in the further steps that would normally refer to normalized CAGE signal (denoted as tpm), will actually be raw tag counts.

CTSS clustering

Transcription start sites are found in the promoter region of a gene and reflect the transcriptional activity of that promoter (Figure \@ref(fig:CTSSbedGraph)). TSSs in the close proximity of each other give rise to a functionally equivalent set of transcripts and are likely regulated by the same promoter elements. Thus, TSSs can be spatially clustered into larger transcriptional units, called tag clusters (TCs) that correspond to individual promoters. CAGEr supports three methods for spatial clustering of TSSs along the genome, two ab initio methods driven by the data itself, as well as assigning TSSs to predefined genomic regions:

• Simple distance-based clustering in which two neighbouring TSSs are joined together if they are closer than some specified distance (greedy algorithm);

• Parametric clustering of data attached to sequences based on the density of the signal [@Frith:2007], http://www.cbrc.jp/paraclu/;

• Counting TSSs and their signal in a set of user supplied genomic regions (e.g. annotation derived promoter regions or other regions of interest).

These functionalities are provided in the function clusterCTSS(), which accepts additional arguments for controlling which CTSSs will be included in the clustering as well as for refining the final set of tag clusters.

We will perform a simple distance-based clustering using 20 bp as a maximal allowed distance between two neighbouring TSSs. Prior to clustering we will filter out low-fidelity TSSs - the ones supported by less than 1 normalized tag counts in all of the samples.

clusterCTSS( object = ce
           , threshold = 1
           , thresholdIsTpm = TRUE
           , nrPassThreshold = 1
           , method = "distclu"
           , maxDist = 20
           , removeSingletons = TRUE
           , keepSingletonsAbove = 5)

Our final set of tag clusters will not include singletons (clusters with only one TSS), unless the normalized signal is above 5, \emph{i.e.} it is a reasonably supported TSS. The clusterCTSS function creates a set of clusters for each sample separately; for each cluster it returns the genomic coordinates, counts the number of TSSs within the cluster, determines the position of the most frequently used (dominant) TSS, calculates the total CAGE signal within the cluster and CAGE signal supporting the dominant TSS only. We can extract tag clusters for a desired sample from CAGEexp object by calling the tagClusters function:

head(tagClusters(ce, sample = "zf_unfertilized_egg"))

Promoter width

Genome-wide mapping of TSSs using CAGE has initially revealed two major classes of promoters in mammals [@Carninci:2006], with respect to the number and distribution of TSSs within the promoter. They have been further correlated with differences in the underlying sequence and the functional classes of the genes they regulate, as well as the organization of the chromatin around them.

• "broad" promoters with multiple TSSs characterized by a high GC content and overlap with a CpG island, which are associated with widely expressed or developmentally regulated genes;
• "sharp" promoters with one dominant TSS often associated with a TATA-box at a fixed upstream distance, which often regulate tissue-specific transcription.

Thus, the width of the promoter is an important characteristic that distinguishes different functional classes of promoters. CAGEr analyzes promoter width across all samples present in the CAGEexp object. It defines promoter width by taking into account both the positions and the CAGE signal at TSSs along the tag cluster, thus making it more robust with respect to total expression and local level of noise at the promoter. Width of every tag cluster is calculated as following:

1. Cumulative distribution of CAGE signal along the cluster is calculated.
2. A "lower" (qLow) and an "upper" (qUp) quantile are selected by the user.
3. From the 5′ end the position, the position of a quantile $q$ is detemined as the first base in which of the cumulative expression is higher or equal to $q\%$ of the total CAGE signal of that cluster.
4. Promoter interquantile width is defined as the distance (in base pairs) between the two quantile positions.

The procedure is schematically shown in Figure \@ref(fig:CumulativeDistribution).

knitr::include_graphics("images/CumulativeDistributionAndQuantiles.svg")

Required computations are done using cumulativeCTSSdistribution and quantilePositions functions, which calculate cumulative distribution for every tag cluster in each of the samples and determine the positions of selected quantiles, respectively:

cumulativeCTSSdistribution(ce, clusters = "tagClusters", useMulticore = T)
quantilePositions(ce, clusters = "tagClusters", qLow = 0.1, qUp = 0.9)

Tag clusters and their interquantile width can be retrieved by calling tagClusters function:

tagClustersGR( ce, "zf_unfertilized_egg"
             , returnInterquantileWidth = TRUE,  qLow = 0.1, qUp = 0.9)

Once the cumulative distributions and the positions of quantiles have been calculated, the histograms of interquantile width can be plotted to globally compare the promoter width across different samples (Figure \@ref(fig:TagClustersInterquantileWidth):

plotInterquantileWidth(ce, clusters = "tagClusters", tpmThreshold = 3, qLow = 0.1, qUp = 0.9)

Significant difference in the promoter width might indicate global differences in the modes of gene regulation between the two samples. The histograms can also help in choosing an appropriate width threshold for separating sharp and broad promoters.

Creating consensus promoters across samples

Tag clusters are created for each sample individually and they are often sample-specific, thus can be present in one sample but absent in another. In addition, in many cases tag clusters do not coincide perfectly within the same promoter region, or there might be two clusters in one sample and only one larger in the other. To be able to compare genome-wide transcriptional activity across samples and to perform expression profiling, a single set of consensus clusters needs to be created. This is done using the aggregateTagClusters function, which aggregates tag clusters from all samples into a single set of non-overlapping consensus clusters:

aggregateTagClusters(ce, tpmThreshold = 5, qLow = 0.1, qUp = 0.9, maxDist = 100)
ce$outOfClusters / ce$librarySizes *100

Tag clusters can be aggregated using their full span (from start to end) or using positions of previously calculated quantiles as their boundaries. Only tag clusters above given tag count threshold will be considered and two clusters will be aggregated together if their boundaries (i.e. either starts and ends or positions of quantiles) are <= maxDist apart. Final set of consensus clusters can be retrieved by:

consensusClustersGR(ce)

which will return genomic coordinates and sum of CAGE signal across all samples for each consensus cluster (the tpm column).

Analogously to tag clusters, analysis of promoter width can be performed for consensus clusters as well, using the same cumulativeCTSSdistribution, quantilePositions and plotInterquantileWidth functions as described above, but by setting the clusters parameter to "consensusClusters". Like the CTSS, the consensus clusters can also be annotated:

annotateConsensusClusters(ce, exampleZv9_annot)
cumulativeCTSSdistribution(ce, clusters = "consensusClusters", useMulticore = TRUE)
quantilePositions(ce, clusters = "consensusClusters", qLow = 0.1, qUp = 0.9, useMulticore = TRUE)

Although consensus clusters are created to represent consensus across all samples, they obviously have different CAGE signal and can have different width or position of the dominant TSS in the different samples. Sample-specific information on consensus clusters can be retrieved with the \Rfunction{consensusClusters} function, by specifying desired sample name (analogous to retrieving tag clusters):

consensusClustersGR( ce, sample = "zf_unfertilized_egg"
                       , returnInterquantileWidth = TRUE,  qLow = 0.1, qUp = 0.9)

This will, in addition to genomic coordinates of the consensus clusters, which are constant across all samples, also return the position of the dominant TSS, the CAGE signal (tpm) and the interquantile width specific for a given sample. Note that when specifying individual sample, only the consensus clusters that have some CAGE signal in that sample will be returned (which will be a subset of all consensus clusters).

Data export for genome browsers

CAGE data can be visualized in the genomic context by exporting raw or normalized CAGE tag counts to a bedGraph (or BigWig) file and uploading (or linking) the file to a genome browser. Positions of TSSs and tag counts supporting them are exported using exportCTSStoBedGraph()^[Note that the ZENBU genome browser can display natively data from BAM or BED files as coverage tracks.]:

exportCTSStoBedGraph(ce, values = "normalized", format = "bedGraph", oneFile = TRUE)

This will produce a single bedGraph file with multiple annotated tracks that can be directly visualized as custom tracks in the genome browser (Figure \@ref(fig:CTSSbedGraph)).

There are two tracks per sample; one for TSSs on the plus strand and the other for the minus strand. Values for TSSs on minus strand are shown as negative and are pointing downwards in the browser.

Alternatively, the tracks can be exported to a binary BigWig format:

exportCTSStoBedGraph(ce, values = "normalized", format = "BigWig")

which will produce two BigWig files per sample (one for TSSs on the plus strand and the other for the minus strand) and an accompanying text file with track headers.

knitr::include_graphics("images/CTSSbedGraph.svg")

Interquantile width can also be visualized in a gene-like representation in the UCSC genome browser by exporting the data into a BED file:

exportToBed(object = ce, what = "tagClusters", qLow = 0.1, qUp = 0.9, oneFile = TRUE)

In this gene-like representation (Figure \@ref(fig:tagClustersBed)), the oriented line shows the full span of the cluster, filled block marks the interquantile width and a single base-pair thick block denotes the position of the dominant TSS.

knitr::include_graphics("images/TagClustersBed.svg")

Expression profiling

Since CAGE signal reflects level of transcription from a given TSS or promoter it can be used for 5′ centered expression profiling. Expression clustering can be done at level of individual CTSSs or at level of entire promoters (consensus clusters). In the former case, feature vector containing log transformed and scaled normalized CAGE signal at individual TSS across multiple samples is used as input for clustering algorithm, whereas in the latter case CAGE signal within the entire consensus cluster is used. CAGEr supports two unsupervised clustering algorithms: kmeans and self-organizing maps (SOM). Both algorithms require to specify a number of clusters in advance.

We will perform expression clustering at the level of entire promoter using SOM algorithm and applying it only to promoters with normalized CAGE signal $>= 15$ in at least one sample.

# Not supported yet for CAGEexp objects, sorry.
# getExpressionProfiles(ce, what = "consensusClusters", tpmThreshold = 10, 
#       nrPassThreshold = 1, method = "som", xDim = 4, yDim = 2)

Distribution of expression across samples for 8 clusters returned by SOM (4 $\times$ 2 map) can be visualized using the plotExpressionProfiles function as shown in Figure \@ref(fig:ConsensusClustersExpressionProfiles):

# Not supported yet for CAGEexp objects, sorry.
# plotExpressionProfiles(ce, what = "consensusClusters")

knitr::include_graphics("images/ConsensusClustersExpressionProfiles.svg")

Each cluster is shown in different color and is marked by its label and the number of elements (promoters) in the cluster. We can extract promoters belonging to a specific cluster by typing:

# Not supported yet for CAGEexp objects, sorry.
# class3_1 <- extractExpressionClass(ce, 
#       what = "consensusClusters", which = "3_1")
# head(class3_1)

Consensus clusters and information on their expression profile can be exported to a BED file, which allows visualization of the promoters in the genome browser colored in the color of the expression cluster they belong to (Figure \@ref(fig:ConsensusClustersBed):

# Not supported yet for CAGEexp objects, sorry.
# exportToBed(ce, what = "consensusClusters", 
#       colorByExpressionProfile = TRUE)

knitr::include_graphics("images/ConsensusClustersBed.svg")

Expression profiling of individual TSSs is done using the same procedure as described above for consensus clusters, only by setting wha = "CTSS" in all of the functions.

Differential expression analysis

The raw expression table for the consensus clusters can be exported to the r Biocpkg("DESeq2") package for differential expression analysis. For this, the column data needs to contain factors that can group the samples. They can have any name.

ce$group <- factor(c("a", "a", "b", "b"))
dds <- consensusClustersDESeq2(ce, ~group)

Shifting promoters

As shown in Figure \@ref(fig:tagClustersBed), TSSs within the same promoter region can be used differently in different samples. Thus, although the overall transcription level from a promoter does not change between the samples, the differential usage of TSSs or promoter shifting may indicate changes in the regulation of transcription from that promoter, which cannot be detected by expression profiling. To detect this promoter shifting, a method described in @[Haberle:2014] has been implemented in CAGEr. Shifting can be detected between two individual samples or between two groups of samples. In the latter case, samples are first merged into groups and then compared in the same way as two individual samples. For all promoters a shifting score is calculated based on the difference in the cumulative distribution of CAGE signal along that promoter in the two samples. In addition, a more general assessment of differential TSS usage is obtained by performing Kolmogorov-Smirnov test on the cumulative distributions of CAGE signal, as described below. Thus, prior to shifting score calculation and statistical testing, we have to calculate cumulative distribution along all consensus clusters:

cumulativeCTSSdistribution(ce, clusters = "consensusClusters")

Next, we calculate a shifting score and P-value of Kolmogorov-Smirnov test for all promoters comparing two specified samples:

# Not supported yet for CAGEexp objects, sorry.
# scoreShift(ce, groupX = "zf_unfertilized_egg", groupY = "zf_prim6",
#       testKS = TRUE, useTpmKS = FALSE)

This function will calculate shifting score as illustrated in Figure \@ref(fig:ShiftingScore). Values of shifting score are in range between -Inf and 1. Positive values can be interpreted as the proportion of transcription initiation in the sample with lower expression that is happening "outside" (either upstream or downstream) of the region used for transcription initiation in the other sample. In contrast, negative values indicate no physical separation, i.e. the region used for transcription initiation in the sample with lower expression is completely contained within the region used for transcription initiation in the other sample. Thus, shifting score detects only the degree of upstream or downstream shifting, but does not detect more general changes in TSS rearrangement in the region, e.g. narrowing or broadening of the region used for transcription.

\ To assess any general change in the TSS usage within the promoter region, a two-sample Kolmogorov-Smirnov (K-S) test on cumulative sums of CAGE signal along the consensus cluster is performed. Cumulative sums in both samples are scaled to range between 0 and 1 and are considered to be empirical cumulative distribution functions (ECDF) reflecting sampling of TSS positions during transcription initiation. K-S test is performed to assess whether the two underlying probability distributions differ. To obtain a P-value i.e. the level at which the null-hypothesis can be rejected), sample sizes that generated the ECDFs are required, in addition to actual K-S statistics calculated from ECDFs. These are derived either from raw tag counts, i.e. exact number of times each TSS in the cluster was sampled during sequencing (when useTpmKS = FALSE), or from normalized tpm values (when useTpmKS = TRUE). P-values obtained from K-S tests are further corrected for multiple testing using Benjamini and Hochenberg (BH) method and for each P-value a corresponding false-discovery rate (FDR) is also reported.

knitr::include_graphics("images/ShiftingScore.svg")

We can select a subset of promoters with shifting score and/or FDR above specified threshold:

# Not supported yet for CAGEexp objects, sorry.
# shifting.promoters <- getShiftingPromoters(ce, 
#       tpmThreshold = 5, scoreThreshold = 0.6,
#       fdrThreshold = 0.01)
# head(shifting.promoters)

The getShiftingPromoters function returns genomic coordinates, shifting score and P-value (FDR) of the promoters, as well as the value of CAGE signal and position of the dominant TSS in the two compared (groups of) samples. Figure \@ref(fig:ShiftingPromoter) shows the difference in the CAGE signal between the two compared samples for one of the selected high-scoring shifting promoters.

knitr::include_graphics("images/ShiftingPromoter.svg")

Appendix

Creating a CAGEexp object by coercing a data frame {#coerce-CAGEexp}

A CAGEexp object can also be created directly by coercing a data frame containing single base-pair TSS information. To be able to do the coercion into a CAGEexp, the data frame must conform with the following:

• The data frame must have at least 4 columns;

• the first three columns must be named chr, pos and strand, and contain chromosome name, 1-based genomic coordinate of the TSS (positive integer) and TSS strand information (+ or -), respectively;

• these first three columns must be of the class character, integer and character, respectively;

• all additional columns must be of the class integer and should contain raw CAGE tag counts (non-negative integer) supporting each TSS in different samples (columns). At least one such column with tag counts must be present;

• the names of the columns containing tag counts must begin with a letter, and these column names are used as sample labels in the resulting CAGEexp object.

An example of such data frame is shown below:

TSS.df <- read.table(system.file( "extdata/Zf.unfertilized.egg.chr17.ctss"
                                , package = "CAGEr"))
# make sure the column names are as required
colnames(TSS.df) <- c("chr", "pos", "strand", "zf_unfertilized_egg")
# make sure the column classes are as required
TSS.df$chr <- as.character(TSS.df$chr)
TSS.df$pos <- as.integer(TSS.df$pos)
TSS.df$strand <- as.character(TSS.df$strand)
TSS.df$zf_unfertilized_egg <- as.integer(TSS.df$zf_unfertilized_egg)
head(TSS.df)

This data.frame can now be coerced to a CAGEexp object, which will fill the corresponding slots of the object with provided TSS information:

ce.coerced <- as(TSS.df, "CAGEexp")
ce.coerced

Summary of the CAGEr accessor functions

Originally there was one accessor per slot in CAGEset objects (the original CAGEr format), but now that I added the CAGEexp class, that uses different underlying formats, the number of accessors increased because a) I provide the old ones for backwards compatibility and b) there multiple possible output formats.

Before releasing this CAGEr update in Bioconductor, I would like to be sure that the number of accessors and the name scheme are good enough.

Please let me know your opinion about the names and scope of the accessors below:

CTSS data

Name | Output --------------------|------------------------------------------------------ CTSScoordinates | Coordinate table in ad-hoc data.frame format. CTSScoordinatesGR | Coordinate table in GRanges format. CTSStagCount | Raw CTSS counts in ad-hoc data.frame format (with coordinates). CTSStagCountDA | Raw CTSS counts in DelayedArray format wrapping a integer Rle DataFrame. CTSStagCountDF | Raw CTSS counts in integer Rle DataFrame format. CTSStagCountDf | Raw CTSS counts in data.frame format (without coordinates). CTSStagCountGR | Raw CTSS counts in GRanges format (single samples). CTSStagCountSE | RangedSummarizedExperiment containing an assay for the Raw CTSS counts. CTSStagCountTable | Returns CTSStagCount for CAGEsets and CTSStagCountDF for CAGEexps. CTSSnormalizedTpm | Normalised CTSS values in ad-hoc data.frame format (with coordinates). CTSSnormalizedTpmDF | Normalised CTSS values in Rle DataFrame format. CTSSnormalizedTpmDf | Normalised CTSS values in ad-hoc data.frame format (without coordinates). CTSSnormalizedTpmGR | Normalised CTSS values in GRanges format (single samples).

Cluster data

Name | Output ------------------------|------------------------------------------------------ consensusClusters | Consensus cluster coordinates in ad-hoc data.frame format. consensusClustersDESeq2 | Consensus cluster coordinates and expression matrix in DESeq2 format. consensusClustersGR | Consensus cluster coordinates in GRanges format. consensusClustersSE | Consensus cluster coordinates and expression matrix in RangedSummarizedExperiment format. consensusClustersTpm | Consensus cluster coordinates and raw expression matrix. tagClusters | Per-sample list of tag cluster coordinates in ad-hoc data.frame format. tagClustersGR | Per-sample GRangesList of tag cluster coordinates.

Gene data

Name | Output ----------------|------------------------------------------------------ GeneExpDESeq2 | Gene expression data in DESeq2 format. GeneExpSE | Gene expression data in SummarizedExperiment format.

Summary of the CAGEexp experiment slots and assays

A CAGEexp object can contain the following experiments.

Please let me know your opinion about the names

Name | Assays | Comment ------------------|------------------------------|--------------------------- tagCountMatrix | counts, normalizedTpmMatrix | RangedSummarizedExperiment seqNameTotals | counts, norm | SummarizedExperiment consensusClusters | counts, normalized, q_x, q_y | RangedSummarizedExperiment geneExpMatrix | counts | SummarizedExperiment

CAGEexp assays

Name | Experiment | Comment --------------------|-------------------|------------------------------------------------------- counts | tagCountMatrix | Integer Rle DataFrame of CTSS raw counts. counts | seqNameTotals | Numeric matrix of total counts per chromosome. counts | consensusClusters | Integer matrix of consensus cluster expression counts. counts | geneExpMatrix | Integer matrix of gene expression counts. normalizedTpmMatrix | tagCountMatrix | Numeric matrix of normalised CTSS expression scores. norm | seqNameTotals | Numeric matrix of percent total counts per chromosome. normalized | consensusClusters | Numeric matrix of normalised CC expression scores. q_x, q_y, q_z, ... | consensusClusters | Interger Rle DataFrame of relative quantile positions

Summary of the CAGEr classes

The CTSS, CTSS.chr, TagCluster and ConsensusClsuters are mostly used internally or type safety and preventing me (Charles) from mixing up inputs. They are visible from the outside. Should they be used more extensively ? Can they be replaced by more "core" Bioconductor classes ?

Name | Comment ------------------|------------------------------------------------------------------------- CAGEset | The original CAGEr class, based on data frames and matrices. CAGEexp | The new CAGEr class, based on GRanges, DataFrames, etc. CAGEr | Union classs for functions that accept both CAGEset and CAGEexp. CTSS | Wraps GRanges and guarantees that width equals 1. CTSS.chr | Same as CTSS but also guaranteers the there is only one chromosome (useful in some loops) TagClusters | Wraps GRanges, represents the fact that each sample has their own tag clusters. ConsensusClusters | Wraps GRanges, represents the fact that they are valid for all the samples. CAGErCluster | Union class for functions that accept both TagClusters and ConsensusClusters.

Session info {.unnumbered}

sessionInfo()

References

CAGEr documentation built on Jan. 17, 2021, 2 a.m.