Search
Question: PCA of voom weight values
0
gravatar for Lorena Pantano
3.8 years ago by
Boston
Lorena Pantano90 wrote:

Hi,

I have a big rnaseq data set (146 samples), I did the voom transformation with a design matrix that have my factor with my 4 groups of interest and 4 more factors that could affect the expression as well, but I am not interested in them,only added to catch the variation they may introduce.

I did a PCA of the weights that voom returns, and I saw my samples clustered in the 4 groups I am interested to do the DE, so suddenly I had the question what this means? Is something that we should expect, or that means that are some bias, or is not important?

I tried to think about it, I thought that weight won't be correlated with anything, but some genes are doing the separation of the samples because of the weight values. Since weights are used in the glm, and they are correlated with my groups, don't know if results will be correct.

thanks in advance

ADD COMMENTlink modified 3.7 years ago by Gordon Smyth35k • written 3.8 years ago by Lorena Pantano90

on a side note, I am looking to the density plots, and may be the expression distribution can explain this. but still don't know if I need to be more carefully in the DE step, or try to normalize better.

ADD REPLYlink written 3.8 years ago by Lorena Pantano90

At first blush, I don't think I'd expect weights to correlate with anything, but if you accept that higher expressed genes will also have a higher weight, then the result you observe isn't so surprising, no?

ADD REPLYlink written 3.8 years ago by Steve Lianoglou12k
2
gravatar for Gordon Smyth
3.7 years ago by
Gordon Smyth35k
Walter and Eliza Hall Institute of Medical Research, Melbourne, Australia
Gordon Smyth35k wrote:

I am not sure what you hope to learn from doing a PCA of the voom weights.

The voom weights are higher for larger counts. If some of your samples have higher sequencing depths than other samples, then the weights be will be systematically higher for those samples. If the samples in one of your groups or factors have higher than average sequencing depths then the voom weights will also reflect that distinction.

In general, the voom weights may well be associated with factors in your experiment because they are a function of the fitted count sizes, however this doesn't introduce bias. The weights just make the DE analysis more powerful by reflecting the variance trend that larger counts have smaller coefficients of variation.

ADD COMMENTlink written 3.7 years ago by Gordon Smyth35k

Thank you Gordon. After looking at the code that makes sense. I didn't understand very well the concept of weight until that.

So thank you (and the rest) to take your time to answer here.  I didn't get notification for all your answer, so I thought I hadn't any. 

really appreciated.

ADD REPLYlink written 3.7 years ago by Lorena Pantano90
1
gravatar for Ryan C. Thompson
3.8 years ago by
The Scripps Research Institute, La Jolla, CA
Ryan C. Thompson7.0k wrote:

I don't think that different groups having different weights is necessarily a concern. That might just indicate that some groups are more variable than others.

ADD COMMENTlink written 3.8 years ago by Ryan C. Thompson7.0k
Please log in to add an answer.

Help
Access

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 2.2.0
Traffic: 363 users visited in the last hour