using hclust object from FactoMineR::HCPC to force clustering in pheatmap

Question

[komal.rathi]

Cross posted: https://stackoverflow.com/questions/62972420/hclust-object-from-factorminer-does-not-cluster-dendrogram-in-pheatmap

I am trying to create a heatmap of a matrix with 20 rows and 10 columns using pheatmap::pheatmap. To cluster the columns, I am using an hclust object obtained after running FactoMineR::HCPC on the input matrix. However, when I use the hclust object, the

Below is my reproducible code:

library(tidyverse)
library(pheatmap)
library(FactoMineR)

# reproducible df of 20 rows and 10 columns
set.seed(100)
tmp <- matrix(rnorm(10000), nrow = 20, ncol = 10)
tmp <- as.data.frame(tmp)
colnames(tmp) <- paste0('col_', seq(1:ncol(tmp)))
rownames(tmp) <- paste0('row_', seq(1:nrow(tmp)))

# use FactoMineR HCPC for clustering data
res.pcahcpc <- FactoMineR::PCA(X = t(tmp), graph = F)
res.pcahcpc <- FactoMineR::HCPC(res.pcahcpc, nb.clust = 4, graph = F)

# get hclust object from FactoMineR::HCPC
pcahcpc.tree <- res.pcahcpc$call$t$tree

# hclust object
> pcahcpc.tree

Call:
flashClust::hclust(d = dissi, method = method, members = weight)

Cluster method   : ward 
Distance         : euclidean 
Number of objects: 10 

# get cluster information for heatmap annotation
res.pcahcpc <- res.pcahcpc$data.clust # clusters
colnames(res.pcahcpc)[ncol(res.pcahcpc)] <- "PCA_HCPC"
res.pcahcpc <- res.pcahcpc[,'PCA_HCPC', drop = F]

# clusters
> head(res.pcahcpc)
      PCA_HCPC
col_1        1
col_2        1
col_3        4
col_4        2
col_5        3
col_6        1

# create heatmap using PCA HCPC clustering
tmp %>%
  pheatmap(cellwidth = 15, cellheight = 15,
           annotation_col = res.pcahcpc,
           cluster_cols = pcahcpc.tree)

Running the above code gives me the heatmap below and as you can see, it is not clustering the columns by FactoMineR PCA HCPC at all (the top annotation colors should've been grouped). How do I use hclust object from FactoMineR::HCPC to force clustering in pheatmap?

Answer 1

[komal.rathi]

Someone posted an answer to this on stackoverflow which I am sharing here: https://stackoverflow.com/a/63017471/1844024

To get the correct dendrogram, you need first to reorder the columns of tmp according to the information given in pcahcpc.tree$labels.

idx <- as.numeric(gsub("col_","", pcahcpc.tree$labels))
pheatmap(tmp[, idx], cellwidth=15, cellheight=12,
         annotation_col = res.pcahcpc,
         cluster_cols = pcahcpc.tree

Answer 2

[Jean-Karim Heriche]

You need to pass a valid hclust object to cluster_cols and I don't think FactoMineR produces one. Try using as.hclust() or use another heatmap function with a more convenient interface.

[komal.rathi]

The pcahcpc.tree seems to be a valid hclust object:

> class(pcahcpc.tree)
[1] "hclust"