Hi community,

I used sva on DESeq2 normalized count.

mod=model.matrix(~group, colData(dds))

mod0=model.matrix(~1, colData(dds))

sv=svaseq(counts(dds, normalized=TRUE), mod, mod0)

Is it recommended to use these SVs in removeBatchEffect(), and how should it be done if it is recommended?

Thanks.

I feel like this has been asked and answered many times before. Generally, the recommendations for DESeq2 (and edgeR/limma, see their manuals and many posts on this here on the support site) is to include covariates (like batch, surrogate variables, anything...) into the design rather than using direct adjustment methods.

That means, you would add the SVs to colData(dds) and then do ~SV1+SV2+SV(n)+your_actual_design. You could use removeBatchEffect() on the output of vst or rlog to visualize / approximate the impact of adding these variables, e.g. by doing PCA with and without correction. But this is not input for DESeq2 as it uses raw counts. The removeBatchEffect does not return raw or integer counts, so it's not compatible.

See for example https://github.com/mikelove/preNivolumabOnNivolumab/blob/main/preNivolumabOnNivolumab.knit.md which is a little case study from Mike Love using RUVseq (same principles apply for sva) with DESeq2.