Clustering RNA-Seq RPKM/ReadsCount issues

[Amative]

Hello all,

I have an RNA-Seq expression data, both RPKM/ReadsCount. Number of records in each dataset is ~35,000 gene.

The samples are organised as follows:
Tissue #1: 5 stages , 2 replicates for each stage. Total 10 samples
Tissue #2: 5 stages , 2 replicates for each stage. Total 10 samples

What I want to do with the expression data is basically clustering. I need to cluster each tissue separate from the other for now.

A couple of issues I came across:

• What is the correct way to deal with missing values as I have many of them scattered all over the dataset?

• I read that it good to get rid of those genes that have zero ReadsCount in more than a specific percentage (30%, 40%, 50% ... ) of the samples? makes sense to me, but I dunno if it is correct!

• What kind of data should go into the clustering process (RPKM, ReadsCount, add 0.00001 to all RPKM values, take the log2 , then calculate the standard normalized value of each RPKM

• Should replicates go into the clustering, or values from both replicates of each samples need to normalized in some way to produce one value (this way I will end up with 5 columns for each sample)?

I would really appreciate any kind of help with each of the above issues

[Amative]

Any help, please!

[dpryan]

The shortest answer would be to simply refer you to section 7 of the DESeq vignette as an example.

I should add, that there are multiple ways to do this sort of thing. Also, biological replicates should be kept separate, otherwise you'd might as well not have bothered running them.

[muthu545]

Hi,

For Read counts data set matrix (35000 genes * 20 samples),
First filter all the genes that is zero across all the samples.

Second, load the filtered dataset in R using DESeq package --> perform size factor normalization -->
and apply 20-40% (your choice, whichever works better) quantile filtering to filter out the low expressed genes.

Third, use the filtered dataset from above to calculate the distance matrix
and then used pclust function to perform clustering of the samples.

I'm not sure the same could be applied for RPKMs.... you could use the cummeRbund
package to cluster the samples based on RPKM values if the RPKMs was derived from cufflinks suite with differential expression data from cuffidiff.

I hope this will work for you and cluster the 2 tissue samples separately.

Thanks
--
Muthu