how to interpret of "pvclust" dendrogram and finding height for cutting dendrogram?

Question

I study on RNA-seq expression dataset about one cancer in TCGA. I downloaded FPKM dataset and removed batch effect by ComBat() function. Now, I used pvclust for clustering 72 samples and the result of that can access in this link.my code for running pvclust is as below:

fit_LMS <-pvclust(datForpvclust,method.hclust = "average",
                          method.dist = "euclidean",nboot=10000, parallel=TRUE ) 
pvrect(fit_LMS, alpha=0.95)

As you see, by alpha=0.95 just 3 clusters are significant although in right cluster a few clusters have been included. really I can't analysis this result. So, I have 2 questions: 1- What is the interpretation of pvclust dendrogram? Does my dataset has meaning full clusters?

2- I am interested in the height of tree cut in the dendrogram, which height is better for this dataset based on pvclust result? H=105 or H=110 or another height?

I appreciate it if anybody shares his/her comment with me.

Answer 1

• You do not need to give links to your images, you can embed images in your questions using the "image" icon.
• FPKM (and RPKM) might not be the best RNA-seq normalization method, see this and this.

pvclust providing p-values for the clusters computed by hclust. To do so, bootstrap resampling is used. From their documentation:

Values on the edges of the clustering are p-values (%). Red values are AU p-values, and green values are BP values. Clusters with AU larger than 95% are highlighted by rectangles, which are strongly supported by data.

Basically you are getting a clustering pattern and a statistic associated with each of the clusters giving you an idea about the "statistical significance" of the clusters you have obtained (i.e. clusters most probably not being a result of strong batch effetcs, ...).

To be able to comment on the resulting clusters making sense or not, you will need to consult somebody with domain knowledge, it is not a number or two that would determine the quality of the output but also how good is the fit between your result and the (expected) biology. For example in your case with TCGA data:

• Do the samples cluster by batch?
• Do tumor and adjacent normal tissue samples aggregate into different clusters?
• Do the samples in the same cluster share the same mutation/methylation pattern/disease subgroup?