This is a test version of Biostars. For the public version, visit https://www.biostars.org.
How To Adjust Data For Confounding Covariates Before Pca?

I have the model:

~group+batch+age+BMI

which I have been using in LIMMA, where "group" is the biologically interesting thing and the others are confounders. What is the best method/package to use to remove the effects of the confounding covariates prior to PCA, so that I know I am only witnessing the biologically relevant clustering?

I have noticed several of the packages available in R either assume only one confounding covariate or only deal with categorical variables.

pca
• 13,046 views • link
updated by Ram Ram • written by scottwilliamrobinson scottwilliamrobinson

2 answers

Have you found an answer?

Actually, I think you could use the partial residual as described here.

Basically, I think you can take partial residual relative to the confounders for each variable and use the partial residual matrix to perform PCA. Since I've never heard people doing so, I am not 100% sure this is correct. But I really cannot think of a problem with that.

• 1 views • link
written by fansili2013 • fansili2013

The sva/ComBat dynamic duo is likely a good place to start. Reading through the vignette you'll find that you can use both categorical and continuous variables, too.

• 0 views • link
written by Steve Lianoglou Steve Lianoglou

The only bit I could find talking about continuous variables:

"By default, all adjustment variables will be treated as factor variables by the ComBat function. If you would like to include continuous adjustment variables, also create a vector containing the column numbers of the continuous covariates in the model matrix. This vector must then be input into ComBat via the numCovs option.

We now apply the ComBat function to the data, using parametric empirical Bayesian adjustments.

...

This returns an expression matrix, with the same dimensions as your original dataset. This new expression matrix has been adjusted for batch. "

So this is talking about including them in the model to make the batch effect removal more precise - not about removing the effect of the continuous variables. Please correct me if there is something else I missed.

• 0 views • link
written by scottwilliamrobinson scottwilliamrobinson

Hi Scott, I'm not sure if this will be helpful, but I just stumbled on this cbcbSEQ package, which might be of use to use. In particular, check out the vignette (PDF) where they are using a modified ComBat function (combatMod). Maybe this is closer to what you are after?

• 0 views • link
written by Steve Lianoglou Steve Lianoglou

hey ..? did you find a solution, I am wondering the same thing..??

:)

• 0 views • link
updated by Ram Ram • written by cbourdon • cbourdon

Sorry for the late response, haven't been on here in a while and the alert must have been filtered out of my inbox. Unfortunately I couldn't find a solution so I had to just go ahead and use comBat to remove the categorical batch effect and allow the effect of the continuous variable to persist.

• 0 views • link
written by scottwilliamrobinson scottwilliamrobinson

Log in to answer this question.

More posts like this