Home | Blog | Software | Publications | GitHub

Visualize mean signals in row clusters by ComplexHeatmap

In this post, we will demonstrate how to visualize mean signals in row clusters by ComplexHeatmap package.

First we load the packages and generate the random matrix.



mat = cbind(rbind(matrix(rnorm(64, -1), 8), matrix(rnorm(64, 1), 8), matrix(rnorm(64, -1), 8)),
            rbind(matrix(rnorm(64, 1), 8), matrix(rnorm(64, -2), 8), matrix(rnorm(64, -1), 8)),
            rbind(matrix(rnorm(64, -0.5), 8), matrix(rnorm(64, -1), 8), matrix(rnorm(64, 1), 8)))

The random matrix contains three distinct subgroups by rows and following is how it looks by means of heatmap.

Heatmap(mat, name = "foo")

plot of chunk unnamed-chunk-3

Next we are going to do little bit more to enhance the visual effect on the row clusters:

  1. split the heatmap by rows to separate row clusters.
  2. show and compare mean signals in the three row clusters.

We apply k-means clustering on rows. Although you can specify it by km in Heatmap() function, since the partitioning information wil be used in several places, we calculate it in the first place.

km = 3
colors = brewer.pal(km, "Set1")
partition = kmeans(mat, centers = km)$cluster

We will put the mean signals in row clusters as a column annotation put on top of the heatmap. This can be done by constructing a self-defined annotation function. The only input of this function is index which is the index of columns that will be automatically adjusted by column clustering or column reordering.

In following code, the basic logic is:

  1. calculate mean value in different row cluster,
  2. push a viewport to put graphics,
  3. add polygons which show the mean signals,
  4. add y-axis.
anno_col_mean = function(index) {
    col_means = lapply(1:km, function(i) colMeans(mat[partition == i, index]))
    n = length(index)
    rg = range(unlist(col_means))
    pushViewport(viewport(xscale = c(0.5, n + 0.5), yscale = rg))
    grid.rect(gp = gpar(fill = "transparent"))
    for(i in seq_along(col_means)) {
        grid.polygon(c(1:n, n:1), c(col_means[[i]], rep(rg[1], n)), 
            gp = gpar(fill = paste0(colors[i], "80"), col = NA), default.units = "native")
    grid.yaxis(gp = gpar(fontsize = 8))

ha = HeatmapAnnotation(col_mean = anno_col_mean)

Now we can put everything to make the new heatmap. In the heatmap, we specified split = partition to split the heatmap by k-means partition which has already be calculated. To make the correspondance between signal lines and row clusters, we add a color bar on the right of the main heatmap.

Heatmap(mat, name = "foo", column_dend_side = "bottom", top_annotation = ha, 
    top_annotation_height = unit(2, "cm"), split = partition, show_row_names = FALSE) +
Heatmap(partition, col = structure(colors, names = as.character(1:km)), show_row_names = FALSE, 
    show_heatmap_legend = FALSE, name = "", width = unit(5, "mm"))

decorate_annotation("col_mean", {
    grid.text("mean\nsignal", unit(-10, "mm"), rot = 90, just = "bottom")

plot of chunk unnamed-chunk-6