Reordering parent dendrogram slices in k-means split heatmap

[rbutleriii]

Hi @jokergoo, this is very similar to #864 , but instead when using column_km to split groups with k-means clustering. I get 6 groups, and would like the swing the groups on their parent branches to re-order from c(6, 5, 4, 3, 2, 1) to c(6, 1, 2, 4, 3, 5).

In general, to make this reproducible I am setting the seed first, but the issue is related to the structure of the list object I get from column_dend(drawn)

Error: `cluster_columns` should be a logical value, a clustering function or a
clustering object.

Example below

library(ComplexHeatmap)

m = t(USArrests[c(1, 6, 13, 20, 23), ])
# set.seed(3) 
hm <- Heatmap(m, column_km = 2)
drawn <- draw(hm)
#  col_dend <- as.dendrogram(column_dend(drawn))
col_dend <- column_dend(drawn)
# set.seed(3)
hm <- Heatmap(m, cluster_columns = col_dend[c(2, 1)])

It is easy to reorder the list, but I cannot figure out a way to feed it back to the constructor or drawn object.

[jokergoo]

It is not possible under current implementation. The reason is the dendrogram on the column is actually combination of several dendrograms: dendrograms for individual column groups and a "global" dendrogram connecting all column dendrograms.

If the column dendrogram is not necessary, you can precisely control the order of the column groups by the following steps:

1. obtain the groups and columns and orders after clustering within each group:

hm <- Heatmap(m, column_km = 2)
hm <- draw(hm)
column_order = column_order(hm)

2. generate the corresponding "group" variable.

group = rep("1", ncol(m))
group[column_order[[2]]] = "2"

3. make the heatmap again, but using the pre-calculated group and column order

# then make it again
Heatmap(m, column_split = group, column_order = unlist(column_order), cluster_columns = FALSE)

Now you can control the order of column groups by e.g. column_order[c(6, 1, 2, ...)].

[rbutleriii]

Hmm, if I do want the within slice clustering to remain, can I still set the seed and specify column_split = group and cluster_column_slices = FALSE to get the same clustering, or do the random start points depend on the column_order?

[jokergoo]

If you look at these two lines in the source code:

ComplexHeatmap/R/Heatmap-class.R

Lines 1569 to 1570 in b608461

	dend_slice = as.dendrogram(hclust(dist(t(slice_mean))))
	dend_slice = reorder(dend_slice, slice_mean, mean)

The group-level dendrogram is reordered by the group means, and currently users cannot modify this setting. What I suggest you do is

1. clone this github repo
2. modify line 1570. I think the easiest way is to use the dendextend package to "rotation branches" under a specific node (https://cran.r-project.org/web/packages/dendextend/vignettes/dendextend.html#rotation), something like:

dend_slice = dendextend::rotate(dend_slices, c(6, 1, 2, 4, 3, 5))

3. use devtools::load_all() to load this package.

I might add arguments to control the slice-level dendrogram reordering in the future, but currently it is not easy to design an easy-to-use interface.