heatmaps vignette

Malcolm Perry

2019-05-03

heatmaps

Flexible plotting of Functional Genomics data + Sequence features

This package provides simple functions for plotting heatmaps over sets of genomic windows.

This vignette is an example workflow using ChIP-seq data in zebrafish: the User Guide contains detailed information on the package internals if you want fine-grained control of your plots or to develop tools which use parts of the package.

First things first, load the heatmaps package:

Reading in Data

heatmaps is written using core Bioconductor packages, so reading in data can be easily accomplished with the standard tools.

Here we read in a set of zebrafish promoters from a 30% Epiboly (30p) embryo, defined using CAGE data, and corresponding H3K4me3 ChIP-seq data:

Plotting Functional Genomics Data

Many kinds of functional genomics data, such as ChIP-seq, RNA-seq or DNase-seq can be visualised as ‘coverage’ tracks. In UCSC, these would be wig, bigWig or bedGraph files.

First, we need to create our windows. We can create another GRanges object which contains 500bp either side of our promoters, using the promoters function in GenomicRanges. Unfortunately, some of the resulting ranges go off the end of a chromosome and so must be dropped: this is done by testing the width of the trimmed object.

The CoverageHeatmap function creates a heatmap object from a GRanges object or an RleList. If we are using a GRanges object then the weight can be specified. Internally, this is passed to the coverage function from GenomicRanges. In the example we are working with RleLists, which are returned by the coverage function.

All heatmaps contain a coords slot, which lets plotHeatmap know how to plot the co-ordinates on the x-axis: very often, our plots will be centered on some feature rather than starting from zero on the x axis. The label slot is optional, and is displayed in the top left-hand corner of the plot by default, if present.

## Warning in valid.GenomicRanges.seqinfo(x, suggest.trim = TRUE): GRanges object contains 1 out-of-bound range located on sequence
##   chr21. Note that ranges located on a sequence whose length is
##   unknown (NA) or on a circular sequence are not considered
##   out-of-bound (use seqlengths() and isCircular() to get the lengths
##   and circularity flags of the underlying sequences). You can use
##   trim() to trim these ranges. See ?`trim,GenomicRanges-method` for
##   more information.

The plotHeatmapList function will plot the returned heatmap object to the active device. This function also allows multiple plots to be plotted at the same time, and sets the device margins. It is usually easier to plotHeatmapList rather than plotHeatmap directly.

Options for plotting can be passed to plotHeatmapList function. Here, we set the label text size (cex.label) to be smaller than the default, and use the default color scheme from RColorBrewer. A complete list of color schemes is available using the command RColorBrewer::display.brewer.all(), or on the ColorBrewer website.

## plotting heatmap H3K4me3 30p

Another way of visualising this signal is using a meta-region plot. This is effectively just a sum over the ‘columns’ of a heatmap.

We can see from this picture that there is an enrichment of H3K4me3 signal downstream of the promoters. It appears to have some kind of phase, but it’s not very clear what’s happening.

If we subtract negative strand reads from positive strand reads, a better picture starts to emerge.

It is very easy to specify custom color schemes in heatmaps. If a vector of colors is supplied (in any format R understands), then they are interpolated by colorRamp.

When we are using non-obvious color schemes, it can help to plot a legend describing the value of the colors. This is handled automatically by plotHeatmapList if the option legend=TRUE is set.

## plotting heatmap Phase

It can also be helpful to cluster heatmaps. The heatmaps package does not provide methods for clustering, but can display a partition defined by the user. This is to make sure any method can be used, and that the clusters can be recovered after plotting (particularly using non-deterministic methods like k-means).

We can use a simple k-means approach from within R to partition the rows of our image matrix, then re-order the rows, remembering the clustering:

## plotting heatmap kmeans

Plotting Sequence Features

heatmaps also contains convenient functions to plot sequence features, such as kmer content or PWM matches, or genomic windows.

First we extract the sequence associated with our windows:

Now we can use the function PatternHeatmap to extract patterns from our sequence. We can specify either kmers, including ambiguity codes, or using PWMs.

## plotting heatmap TA

This heatmap is difficult to see patterns in because the points are binary and the data is sparse. heatmaps provides a function to smooth this data. It also lets us resize the image so that our plots don’t take ages to plot. If we plotted every point individually, the result would be much higher resolution than could possibly fit onscreen. output.size specifies the dimensions of the output image matrix.

The algorithm argument specifies the smoothing method. Specifying “kernel” uses the bkde2D function from the package KernSmooth. In this case, because we are using binary data, this would be chosen automatically.

## 
## Calculating kernel density...
## Warning in bkde2D(cbind(df$i, df$j), bandwidth = sigma, gridsize =
## output.size, : Binning grid too coarse for current (small) bandwidth:
## consider increasing 'gridsize'
## plotting heatmap TA

Using PWMs instead of kmers is very similar, except we also have to specify a minimum match score. This can either be absolute or expressed as a percentage (see ?Biostrings::matchPWM for details).

## 
## Calculating kernel density...
## plotting heatmap TATA

An alternative way to visualise PWMs is to plot the score at every point, which is what the function PWMScanHeatmap does. It’s also useful to smooth the output of this function, except this time, because we have continuous rather than binary data, a Gaussian blur is used (EBImage::blur). Again, this would be chosen automatically in this particular case.

Because PWMScanHeatmap can produce some very high and very low values, it’s visually often better to centre the scale around the mean value (as defined in percentages) before plotting, rather than from 0 to 100, or just the min/max values in the heatmap.

## 
## Applying Gaussian blur...
## plotting heatmap TATA

Plotting lists of plots

We have so been using plotHeatmapList to plot individual plots, because it automatically controls the device for use. As its name suggests, we can also plot lists of plots together using this function.

In order to normalise signals between heatmaps, we can specify groups of related plots to plotHeatmapList which will normalise the scales and display settings. In this example we normalise our “AT” and “CG” plots, because these occur at different frequencies. The groups parameter takes anything interpretable as a factor - just specifying numbers is usually the easiest option.

We can specify options for all plots at once, or on a per-group basis. This works by passing a list of options, rather than a vector. Note that for colors (among others), a list (e.g. list("red", "white", "blue") has a very different meaning to the vector (c("red", "white", "blue")) that we used in an earlier example.

The resulting plots shows how “TA” and “CG” content contribute to “TATA” binding potential, as well as promoter H3K4me3, around promoters.

## 
## Calculating kernel density...
## Warning in bkde2D(cbind(df$i, df$j), bandwidth = sigma, gridsize =
## output.size, : Binning grid too coarse for current (small) bandwidth:
## consider increasing 'gridsize'
## 
## Applying Gaussian blur...
## plotting heatmap TA
## plotting heatmap CG
## plotting heatmap TATA
## plotting heatmap H3K4me3 30p

Plotting to file

Heatmaps can take a long time to plot, so it is usually best to plot straight to a file rather than to the R graphics device (although this works fine, and is reasonably quick if you smooth the plots). The default settings for margin sizes, text size etc. are aimed at creating plots which are around 10cm x 20cm (per heatmap), or 4 in by 8 in, as this also looks good on the R graphics device.

PNG is recommended, since PDFs can end up being massive files if care is not taken reducing the image size. When using the PNG device to produce high quality images (which would be suitable for printing), it’s helpful to set the size in real-world units (rather than pixels) and then increase the resolution above the default 72ppi, since this produced high-res plots without the scaling issues than come with specifying pixel sizes.

More Complex Plots

The examples above should give an indication of how publication quality figures can be made from most data types and easily plotted in any format. However, heatmaps was also designed to be more flexible than that, so complex, publication-ready figures can be generated programmatically rather than painstakingly edited in Illustrator or Inkscape. This improves redproducibility, and save a lot of time in cases where the data change, or the same operation is carried out repeatedly.

The following example is taken from Haberle et al, 2014, Nature.

## Warning in valid.GenomicRanges.seqinfo(x, suggest.trim = TRUE): GRanges object contains 1 out-of-bound range located on sequence
##   chr21. Note that ranges located on a sequence whose length is
##   unknown (NA) or on a circular sequence are not considered
##   out-of-bound (use seqlengths() and isCircular() to get the lengths
##   and circularity flags of the underlying sequences). You can use
##   trim() to trim these ranges. See ?`trim,GenomicRanges-method` for
##   more information.

Since we’re going to be making several PatternHeatmaps and smoothing them, we first create a function to do this quickly.

## 
## Calculating kernel density...
## 
## Calculating kernel density...
## 
## Calculating kernel density...
## 
## Calculating kernel density...

We’re not using plotHeatmapList, so our scales won’t be normalised automatically.

We can set options for top and bottom plots separately. If we want to pass options to a plotHeatmap from a list (rather than typing them out in the function call), we must pass all available options as a list. (The function only looks at dots (...) if the options argument is empty). To save us setting all the default options manually, heatmapOptions creates a full list with the specified changes.

Here we want the top plots to have white labels to stand out from the background, and no x ticks since they will be present in the bottom plot. We specify slightly larger x-axis labels than the default.

We also need to specify the margins for our plots, which will be different depending on which part of the final image they occupy. the total margins for each plot are the same so that each heatmap will be the same size.

Finally we can get to the actual plotting. The layout is specified to have a narrow column on the right for a legend.

We have to set the parameters before each call to plotHeatmap.

Plotting additional features to the canvas is easy. The coordinates in use are calculated as follows:

1 sequence or window in the original heatmap is one unit along the y axis.

1 bp in the original sequence or windows is one unit along the x axis.

The bottom left corner is (0, 0), so to label a particular window on the y axis the reverse index (nseq - index) is used, and for bp along the x axis are calculated from -coords(hm)[1].

par(xpd=TRUE/FALSE) is used to allow plotting of the colored triangles outside the normal plotting regions, and has to be reset otherwise the reference lines at x=0 will be plotted outside the canvas as well.

Et voila! A full figure for a paper produced entirely from R.