1 Introduction

This document explains the functionalities available in the esetVis package.

This package contains wrapper functions for three types of visualization: spectral map(P.J. 1976), tsne(van der Maaten 2008) and linear discriminant analysis(Fisher 1936) for data contained in a expressionSet Bioconductor (or an SummarizedExperiment) object. The visualizations are available in two types: static, via the ggplot2 package or interactive, via the ggvis or rbokeh packages.

2 Example dataset

2.1 ExpressionSet object

The ALL dataset contains microarray results from 128 patients with acute lymphoblastic leukemia (ALL). The data is contained in a Bioconductor ExpressionSet object. Extra gene annotation is added to the object, via the annotation package hgu95av2.db.

    library(ALL)
    data(ALL)
    
    # to get gene annotation from probe IDs (from the paper HGU95aV2 gene chip was used)
    library("hgu95av2.db")
    library("AnnotationDbi")
    probeIDs <- featureNames(ALL)
    geneInfo <- AnnotationDbi::select(hgu95av2.db, probeIDs, 
        c("ENTREZID", "SYMBOL", "GENENAME"), "PROBEID")
    # 482 on the 12625 probe IDs don't have ENTREZ ID/SYMBOL/GENENAME

    # remove genes with duplicated annotation: 1214
    geneInfoWthtDuplicates <- geneInfo[!duplicated(geneInfo$PROBEID), ]

    # remove genes without annotation: 482
    genesWthtAnnotation <- rowSums(is.na(geneInfoWthtDuplicates)) > 0
    geneInfoWthtDuplicatesAndWithAnnotation <- geneInfoWthtDuplicates[!genesWthtAnnotation, ]
    
    probeIDsWithAnnotation <- featureNames(ALL)[featureNames(ALL) %in% 
        geneInfoWthtDuplicatesAndWithAnnotation$PROBEID]
    ALL <- ALL[probeIDsWithAnnotation, ]
    
    fData <- geneInfoWthtDuplicatesAndWithAnnotation[
        match(probeIDsWithAnnotation, geneInfoWthtDuplicatesAndWithAnnotation$PROBEID), ]
    rownames(fData) <- probeIDsWithAnnotation
    fData(ALL) <- fData

    # grouping variable: B = B-cell, T = T-cell
    groupingVariable <- pData(ALL)$BT
    
    # create custom palette
    colorPalette <- c("dodgerblue", 
        colorRampPalette(c("white","dodgerblue2", "darkblue"))(5)[-1], 
        "red", colorRampPalette(c("white", "red3", "darkred"))(5)[-1])
    color <- groupingVariable; levels(color) <- colorPalette
    
    # reformat type of the remission
    remissionType <- ifelse(is.na(ALL$remission), "unknown", as.character(ALL$remission))
    ALL$remissionType <- factor(remissionType,
        levels = c("unknown", "CR", "REF"), 
        labels = c("unknown", "achieved", "refractory"))

Following tables detail some sample and gene annotation contained in the ALL ExpressionSet used in the vignette.

subset of the phenoData of the ALL dataset for the first genes
  cod sex age BT remissionType
01005 1005 M 53 B2 achieved
01010 1010 M 19 B2 achieved
03002 3002 F 52 B4 achieved
04006 4006 M 38 B1 achieved
04007 4007 M 57 B2 achieved
04008 4008 M 17 B1 achieved
featureData of the ALL dataset for the first genes
  PROBEID ENTREZID SYMBOL GENENAME
1000_at 1000_at 5595 MAPK3 mitogen-activated protein kinase 3
1001_at 1001_at 7075 TIE1 tyrosine kinase with immunoglobulin like and EGF like domains 1
1002_f_at 1002_f_at 1557 CYP2C19 cytochrome P450 family 2 subfamily C member 19
1003_s_at 1003_s_at 643 CXCR5 C-X-C motif chemokine receptor 5
1004_at 1004_at 643 CXCR5 C-X-C motif chemokine receptor 5
1005_at 1005_at 1843 DUSP1 dual specificity phosphatase 1

2.2 SummarizedExperiment object

The functions of the package also supports object of class: SummarizedExperiment. Note: In this case, the functions fData, pData, exprs should be replaced by their corresponding functions rowData, colData and assay.

3 Spectral map: esetSpectralMap

3.1 Simple spectral map

The function esetSpectralMap creates a spectral map(P.J. 1976) for the dataset. Some default parameters are set, e.g. to print the top 10 samples and top 10 genes, to display the first two dimensions of the analysis…

The resulting biplot contains two components:

  • in the background, a cloud of the genes coordinates (plotted with the hexbin package)
  • in the foreground, each sample of the data is represented as a single point/symbol

Here is an example for the ALL dataset, with the default parameters.

    print(esetSpectralMap(eset = ALL))

3.2 Additional sample information

Several annotation variables are available in the eSet.

3.2.1 General

Four different aesthetics [aes] can be used to display these variables:

  • color, with the tag color
  • transparency, with the tag alpha
  • size, with the tag size
  • shape, with the tag shape

For each of this aesthetic [aes], several parameters are available:

  • [aes]Var: name of the column of the phenoData of the eSet used for the aesthetic, i.e. colorVar
  • [aes]: palette/specified shape/size used for the aesthetic, i.e. color

3.2.2 Custom size and transparency

Custom size and the transparency (variables sizeVar and alphaVar) can be specified:

  • if the size/transparency is a numerical variable (numeric or integer), the range of the size/transparency can be specified respectively with the arguments sizeRange and alphaRange
  • in the other cases (factor or character), custom size/transparency can be specified directly respectively via the size and alpha arguments

In the example, the type and stage of the disease (variable BT) is used for coloring, the remission type for the transparency, the sex for the shape and the age for the size of the points. A custom color palette is specified via the color argument.

    print(esetSpectralMap(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Spectral map \n Sample annotation (1)",
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex", 
        sizeVar = "age", sizeRange = c(0.1, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamples = 0, topGenes = 0, cloudGenes = FALSE))

Just for the demonstration, another visualization of the same dataset, using this time a continuous variable: age for coloring and transparency, a factor for the size and BT for the shape.

    print(esetSpectralMap(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Spectral map \n Sample annotation (2)",
        colorVar = "age",
        shapeVar = "BT", shape = 15+1:nlevels(ALL$BT),
        sizeVar = "remissionType", size = c(2, 4, 6),
        alphaVar = "age", alphaRange = c(0.2, 0.8),
        topSamples = 0, topGenes = 0, cloudGenes = TRUE))

3.3 Custom gene representation

Several parameters related to the gene visualization are available:

  • gene subset: only a subset of the genes (at least two) can be displayed, via the argument psids
  • gene cloud:
  • inclusion/removal of the gene cloud via cloudGenes
  • number of bins specification via cloudGenesNBins
  • cloud color via cloudGenesColor
  • legend:
    • inclusion/removal of the gene legend via cloudGenesIncludeLegend
    • title legend specified via cloudGenesTitleLegend

The spectral map is done only on the subset of the genes, with the number of bins, color, and legend title modified.

    print(esetSpectralMap(eset = ALL,
        psids = 1:1000,
        title = "Acute lymphoblastic leukemia dataset \n Spectral map \n Custom cloud genes",
        topSamples = 0, topGenes = 0, 
        cloudGenes = TRUE, cloudGenesColor = "red", cloudGenesNBins = 50,
        cloudGenesIncludeLegend = TRUE, cloudGenesTitleLegend = "number of features"))

3.4 Label outlying elements

3.4.1 Parameters

Three different kind of elements can be labelled in the plot: genes, samples and gene sets.

For each [element], several parameters are available:

  • top[Elements]: number of elements to label
  • top[Elements]Var (not available for gene sets): column of the corresponding element in the eSet used for labelling, in phenoData for sample and featureData for gene. If not specified, the feature/sample names of the eSet are used
  • top[Elements]Just: label justification
  • top[Elements]Cex: label size
  • top[Elements]Color: label color

3.4.2 Method to select top elements

The distance (sum of squared coordinates) of the gene/sample/gene set to the origin of the plot is used to rank the elements, and to extract the top ‘outlying’ ones.

3.4.3 Package used for static plot

By default (and if installed), the package ggrepel is used for text labelling (as in this vignette), to avoid overlapping labels. The text labels can also be displayed with the ggplot2 by setting the parameter packageTextLabel to ggplot2 (default ggrepel).

3.4.4 Example

In the example, the top genes are labelled with gene symbols (SYMBOL column of the phenoData of the eSet), and the top samples with patient identifiers (cod column of the phenoData of the eSet).

    print(esetSpectralMap(eset = ALL, 
        title = paste("Acute lymphoblastic leukemia dataset \n",
            "Spectral map \n Label outlying samples and genes"),
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topGenes = 10, topGenesVar = "SYMBOL",
        topGenesCex = 2, topGenesColor = "darkgrey",
        topSamples = 15, topSamplesVar = "cod", topSamplesColor = "chocolate4",
        topSamplesCex = 3
    ))

3.5 Gene sets annotation

Genes can be grouped into biologically meaningful gene sets, which can be labelled in the biplot.

Compared to previous section, two additional parameters are available:

  • geneSets (required): list of gene sets. Each element in the list should contain genes identifiers, and the list should be named
  • geneSetsVar: column of the featureData of the eSet used to map the gene identifiers contained in the geneSets object
  • geneSetsMaxNChar: number of characters used in gene sets labels

The geneSets can be created with the getGeneSetsForPlot function (wrapper around the getGeneSets function of the MLP package), which can extract gene sets available in the Gene Ontology (Biological Process, Molecular Function and Cellular Component) and KEGG databases. Custom gene set lists can also be provided.

In the following example, only the pathways from the GOBP database are used.

    geneSets <- getGeneSetsForPlot(
        entrezIdentifiers = fData(ALL)$ENTREZID, 
        species = "Human", 
        geneSetSource = 'GOBP',
        useDescription = TRUE)

Then this list of gene sets is provided to the esetSpectralMap function.

    print(esetSpectralMap(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Spectral map \n Gene set annotation",
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topGenes = 0,
        topGeneSets = 4, geneSets = geneSets, geneSetsVar = "ENTREZID", geneSetsMaxNChar = 30,
        topGeneSetsCex = 2, topGeneSetsColor = "purple"))

Note: because of the inherent hierarchical structure of the Gene Ontology database, sets of genes can be very similar, which can result in close (even overlapping) labels in the visualization.

3.6 Dimensions of the biplot

In all previous plots, only the first dimensions of the principal component analysis were visualized, this can be specified via the dim parameter. The third and fourth dimensions are visualized in the next Figure. This parameter is only available for the spectral map visualization.

    print(esetSpectralMap(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Spectral map \n Dimensions of the PCA",
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(0.5, 4),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        dim = c(3, 4)))

3.7 Implementation

The function uses the mpm and plot.mpm function from the mpm package. Some default parameters are set for these two functions, they can be changed via the mpm.args and plot.mpm.args arguments. For further details, refer to the documentation of these two functions.

Note: One important argument is logtrans in mpm function, which indicates if the data should be first log-transformed before the computation. It is set by default to FALSE, assuming that the data are already in the log scale (it is the case for the ALL dataset).

3.8 Interactive spectral map

All plots available in the esetVis package can be static or interactive.

The argument typePlot can be set respectively to static (by default), in this case ggplot2 is used, or interactive, in this case either the ggvis or rbokeh package is used.

Two functionalities of the interactive plot can be of interest for such high-dimensional data:

  • hoover to check sample annotation. By default, only the sample variables used for the aesthetics are displayed when hoovering on a specific sample dot. Additional sample variables (contained in phenoData) displayed in the hoover can be specified via the interactiveTooltipExtraVars parameter.
  • zoom to focus on specific sample in high-dimensional dataset

The same spectral map than in previous section is used, this time in an interactive plot.

    # wrapper for rbokeh/ggvis example
    esetSPMInteractive <- function(...)
        esetSpectralMap(eset = ALL, 
            title = paste("Acute lymphoblastic leukemia dataset - spectral map"),
            colorVar = "BT", color = colorPalette,
            shapeVar = "sex",
            alphaVar = "remissionType",
            typePlot = "interactive", ...
    )

3.8.1 rbokeh

    esetSPMInteractive(
        packageInteractivity = "rbokeh",
        figInteractiveSize = c(700, 600),
        size = 6,
        # use all phenoData variables for hoover
        interactiveTooltipExtraVars = varLabels(ALL))

3.8.2 ggvis

Note: as ggvis plot requires to have a R session running, only a static version of the plot is included.

    # embed a static version of the plot
    library(ggvis)
    knit_print.ggvis(
        esetSPMInteractive(
            packageInteractivity = "ggvis",
            sizeVar = "age", sizeRange = c(2, 6),
            figInteractiveSize = c(700, 600)
        )
    )

4 Tsne: esetTsne

4.1 General

Another unsupervised visualization is available in the package: t-Distributed Stochastic Neighbor Embedding (tsne). The function esetTnse uses the Rtsne function of the same package.

Most of the previous parameters discussed for the spectral map are available for this visualization, at the exception of:

  • parameter linked to gene annotation/labelling. The gene annotation is not (yet) mapped to the sample coordinated obtained as output from the Rtsne function
  • parameter specific to the spectral map, i.e. dim

Arguments to the Rtsne function can be specified via the Rtsne.args argument.

Here is an example of tsne, for the same dataset and same annotation/labelling.

    print(esetTsne(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Tsne",
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamplesVar = "cod"
    ))

4.2 Additional pre-processing step

The tsne can be quite time-consuming, especially for big data. As the Rtsne function used in the background can also uses an object of class dist, the data can be pre-transformed before running the tsne analysis. The argument fctTransformDataForInputTsne enables to specify a custom function to pre-transform the data.

    print(esetTsne(eset = ALL, 
        title = "Acute lymphoblastic leukemia dataset \n Tsne",
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamplesVar = "cod",
        fctTransformDataForInputTsne = 
            function(mat)   as.dist((1 - cor(mat))/2)
    ))

5 Linear discriminant analysis: esetLda

Another visualization, this time semi-supervised (as a variable is used for the computation), is included: Linear Discriminant Analysis. This uses the lda function from the MASS package.

This function maximizes the variance between levels of a factor, here describing some sample annotation, specified via the ldaVar argument.

As this analysis can be quite time consuming, for the demonstration, the analysis is run only a random feature subset of the data.

5.1 All samples

The returnAnalysis parameter can be used, to extract the analysis which will be used as input for the esetPlotWrapper function, used in the background. It is available also for the esetSpectralMap and esetTnse functions.

    # extract random features, because analysis is quite time consuming
    retainedFeatures <- sample(featureNames(ALL), size = floor(nrow(ALL)/5))
    
    # run the analysis
    outputEsetLda <- esetLda(eset = ALL[retainedFeatures, ], ldaVar = "BT",
        title = paste("Acute lymphoblastic leukemia dataset \n",
            "Linear discriminant analysis on BT variable \n",
            "(Subset of the feature data)"),
        colorVar = "BT", color = colorPalette,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamplesVar = "cod", topGenesVar = "SYMBOL",
        returnAnalysis = TRUE
    )

    # extract and print the ggplot object
    print(outputEsetLda$plot)

The top elements (here genes and samples) labelled in the plot can be accessed via the topElements slot of the object. These are labelled with the identifiers used in the plot and named with sample/gene identifiers of the eset.

    # extract top elements labelled in the plot
    pander(t(data.frame(topGenes = outputEsetLda$topElements[["topGenes"]])))
Table continues below
  35016_at 38095_i_at 1096_g_at 39389_at 37988_at
topGenes CD74 HLA-DPB1 CD19 CD9 CD79B
  38242_at 40688_at 36941_at 34033_s_at 39318_at
topGenes BLNK LAT MLLT11 LILRA2 TCL1A
    pander(t(data.frame(topSamples = outputEsetLda$topElements[["topSamples"]])))
Table continues below
  63001 01005 12007 04008 24005 26008 04010 04018
topSamples 63001 1005 12007 4008 24005 26008 4010 4018
  28008 18001
topSamples 28008 18001

When returnAnalysis is set to TRUE, the output of the function can be used as input to the esetPlotWrapper function, and extra parameters can then be modified.

Here the variable used for the shape of the points for the samples is changed to type of remission (remissionType column).

    # to change some plot parameters
    esetPlotWrapper(
        dataPlotSamples = outputEsetLda$analysis$dataPlotSamples,
        dataPlotGenes = outputEsetLda$analysis$dataPlotGenes,
        esetUsed = outputEsetLda$analysis$esetUsed,
        title = paste("Acute lymphoblastic leukemia dataset \n",
            "Linear discriminant analysis on BT variable (2) \n",
            "(Subset of the feature data)"),
        colorVar = "BT", color = colorPalette,
        shapeVar = "remissionType", 
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamplesVar = "cod", topGenesVar = "SYMBOL"
    )

5.2 Data sample subset

From the previous visualization (obtained on a subset of the genes), the biggest difference between all levels of the type/stage of the disease seems to reside between all B-cells (tagged B) and T-cells (tagged T). It might be interesting to focus on a subset of the data, e.g. only one cell type.

    # keep only 'B-cell' samples
    ALLBCell <- ALL[, grep("^B", ALL$BT)]
    ALLBCell$BT <- factor(ALLBCell$BT)
    colorPaletteBCell <- colorPalette[1:nlevels(ALLBCell$BT )]
    
    print(esetLda(eset = ALLBCell[retainedFeatures, ], ldaVar = "BT",
        title = paste("Acute lymphoblastic leukemia dataset \n",
            "Linear discriminant analysis on BT variable \n B-cell only",
            "(Subset of the feature data)"
        ),
        colorVar = "BT", color = colorPaletteBCell,
        shapeVar = "sex",
        sizeVar = "age", sizeRange = c(2, 6),
        alphaVar = "remissionType", alpha = c(0.3, 0.6, 0.9),
        topSamplesVar = "cod", topGenesVar = "SYMBOL",
    ))

The subcell type which seems to differ the most within all B-cell type is the first one: B1 (most of these samples at the right side of the plot).

References

Fisher, R. A. 1936. “The Use of Multiple Measurements in Taxonomic Problems” 7 (2). Annals of Eugenics: 179–88.

P.J., Lewi. 1976. “Spectral Mapping, a Technique for Classifying Biological Activity Profiles of Chemical Compounds” 26. Arzneimittel Forschung (Drug Research): 1295–1300.

van der Maaten, L.J.P. 2008. “Visualizing High-Dimensional Data Using T-SNE” 26. Journal of Machine Learning Research: 2579–2605.