Adding batch effect as part of the adjustment

[mxdeluca]

I am currently working with TCGA data with known batch effects within the methylation data, and was wondering if it would be possible/advisable from your perspective to add batch effect as part of the adjustment ? I would be very keen to see if there's any additional biological pathways emerging from your approach.

I would basically add batch as part of the clustering step:

model <- stepFlexmix(y2 ~ x + batch, k = 1:nmax, nrep = nrep, verbose = FALSE)

and subsequently also include it as part of the regression :

  b_vs_pur <- lapply(1:nmax, function(z) {
    if(z %in% cl) {
      # Include all batch columns in the regression
      lm(y[cl == z] ~ x[cl == z] + batch[cl == z, ])  # Use batch matrix as covariate
    } else { NA }
  })

  b_vs_1mp <- lapply(1:nmax, function(z) {
    if(z %in% cl) {
      # Include all batch columns in the regression
      lm(y[cl == z] ~ x2[cl == z] + batch[cl == z, ])  # Use batch matrix as covariate
    } else { NA }
  })

curious to hear your thoughts

Thank you

[deborahfnoliveira]

Hi,

Thank you for reaching out and for your interest in PureBeta!

We discussed the point you raised about batch effects and we are not sure how impactful it would be to add a batch covariate to our approach. FlexMix is free to identify the different populations per CpG and small changes to the beta values probably wouldn’t change the final purity estimates that much. In addition, the code changes you proposed here wouldn’t be enough as the other functions in the package don’t expect covariates in the generated output/input.

If you want to account for batches in a specific data set, how about applying the batch correction as a preprocessing step before running our pipeline? Then you wouldn’t have to change the package itself and you could test different/choose your preferred correction method.

Best,

Deborah & Iñaki