Correcting Batch Effects in RNA-seq

What a batch effect is

A batch effect is a systematic technical difference between groups of samples: they were processed on different days, run on different lanes, or prepared with different kits or by different people. The problem is that these differences can shift expression in ways that look exactly like real biology. If you do not account for them, a batch effect can either invent differences that are not real or hide ones that are.

1The best defense is design

The cheapest fix happens before any sequencing. Spread your conditions across batches so that each processing day, lane, or kit contains a mix of treated and control samples. Then a batch difference washes out instead of lining up with your comparison. This is the heart of good experimental design, and no statistical trick later is as good as getting the layout right up front.

2When a batch is unavoidable: model it

Sometimes you cannot avoid a batch, but as long as you recorded which batch each sample belongs to, you can handle it. Put the batch in the DESeq2 design: ~ batch + condition. The model then accounts for the batch differences while still testing the condition. Order matters by convention: the variable of interest goes last, so for the airway data with batch added it would read ~ batch + dex, where dex is the treatment column. You are not removing the batch from the data; you are letting the model subtract its effect.

3Batch-corrected values are for plots only

There is a legitimate place for explicitly removing a batch effect: visualization. For a clean PCA or heatmap you can apply limma::removeBatchEffect or ComBat from the sva package to your transformed (vst or rlog) values, so the plot shows biology rather than batch. But these adjusted values are for the eye only. Do not feed batch-corrected values back into DESeq2. The correct workflow is to model the batch in the design for testing, and to remove the batch only on a copy of the transformed data that you use for figures.

When the batch cannot be separated

One hard limit: if treatment is fully confounded with batch, meaning every treated sample is in one batch and every control in another, then no method can untangle them. The model cannot tell whether a difference came from the treatment or the batch, because they move together perfectly. There is nothing to subtract that does not also subtract the biology. This is why design comes first: confounding is a planning failure that statistics cannot repair.

Situation	Can it be fixed?	How
Conditions spread across batches	Best case	Optionally model batch; design already protects you
Batch unavoidable but recorded	Yes	Add batch to the design, e.g. ~ batch + dex
Need a clean PCA or heatmap	Yes, for plots only	limma::removeBatchEffect or ComBat on transformed data
Treatment fully confounded with batch	No	Cannot be separated; only a new design fixes it

A worked example

The example shows both halves: the batch goes into the design for the actual test, and a separate batch-corrected matrix is made only for plotting.

library(DESeq2)
library(limma)

# --- Correct way to TEST: model the batch in the design ---
dds <- DESeqDataSetFromMatrix(countData = counts_raw,   # raw integers
                              colData   = sample_info,  # has 'batch' and 'dex'
                              design    = ~ batch + dex) # variable of interest last
dds <- DESeq(dds)
res <- results(dds)        # tests dex while accounting for batch

# --- Visualization ONLY: remove batch on transformed data ---
vsd <- vst(dds, blind = FALSE)
plotPCA(vsd, intgroup = c("batch", "dex"))   # batch still visible here

mat <- assay(vsd)
# subtract batch for a cleaner plot; do NOT feed 'mat' back into DESeq2
mat <- limma::removeBatchEffect(mat, batch = vsd$batch)
assay(vsd) <- mat
plotPCA(vsd, intgroup = "dex")               # now split by condition

Check your understanding

Next in Track 3

Differential expression with DESeq2 →