Batch effects in an expression study

A batch is any shared handling step that can nudge every sample that passed through it. In an expression study the nudge is often larger than the biology, and a principal-component plot will say so without embarrassment. This glossary walks through the words people use when that happens, and through the design habit that keeps a batch from swallowing the contrast. The study those words belong to is from cells to a gene expression result.

The decision the vocabulary supports

You are deciding whether a difference in the data can be discussed as biology, as handling, or as a mixture you can model. The decision happens twice: once when you lay out the extractions, and again when a plot comes back coloured by date. Learning the nouns now saves you from treating a kit lot as a pathway.

Words that keep the argument straight

A batch effect is a systematic shift tied to a processing group rather than to the biological factor you meant to study. The group might be a library-preparation date, an operator, a sequencer, a lane, a flow cell, an extraction-kit lot, a shipment, or a plate of reverse transcriptions. The shift can be a global change in counts, a change in the genes that look variable, or a quiet tilt that only appears after normalisation.

Confounding means two explanations are tied together so the data cannot separate them. If all the knockout cultures were extracted from one kit lot and all the wild-type cultures from another, genotype and lot are confounded. Any gene that differs has two stories. Balance means each batch contains each biological group, so the contrast is visible inside the batch. A blocked design is one way to arrange that balance on purpose.

A covariate is a column in the sample sheet you are willing to put in the model: batch, sex, rinse, donor, processing day. If you did not record it, the model cannot use it. Principal component analysis, or PCA, is an ordination that spreads samples along axes of high variance. It is a picture of structure. It is not a test of your contrast. People say "the batch separated on PC1" when the first axis lines up with processing group. That sentence is a warning, and it is often true that this axis explains more spread than the treatment.

Correction is a model that adjusts for a recorded batch factor. ComBat-style empirical Bayes adjustments and linear-model terms, including those used in RNA-seq workflows that include batch in the design, are examples of the class. They change the data or the test under assumptions. They are not a time machine. They cannot recover a contrast that was perfectly tied to the batch, because there is no residual comparison left to estimate. When batch and biology are partly crossed, a model can often estimate the biology while accounting for the batch. When they are identical, the model is singular and the honest output is a redesign.

Overcorrection is what happens when the adjustment is flexible enough to eat the biological signal, especially if batch and group are unbalanced or if you correct on the same variable you later pretend you discovered. Inspect plots before and after. Keep the biological groups identifiable.

The unit you replicate still matters. A batch of technical reruns is not a batch of new biological units. That distinction is biological versus technical replicates.

How a batch gets into the measurement

RNA-seq is sensitive to library composition. A different fragmentation time, a different depletion lot, or a different number of PCR cycles changes which fragments are abundant. Those chemistry shifts survive into the count matrix and can look like coordinated gene changes. qPCR has its own version: a plate, a calibration, a seal, an operator's pipetting. Either method will also record a real freezer failure as if it were expression.

Public records in the Sequence Read Archive and the European Nucleotide Archive sometimes include preparation dates and centre names, and sometimes they do not. When you combine studies, missing batch labels are a reason to be modest. Genome context for the genes that then light up can be checked on the UCSC Genome Browser, which does not tell you whether the signal was a lot number.

Design first, then a model

Randomise or balance biological groups across days, operators and kit lots. If you must process in two weeks, put every condition into each week. Record the columns at the bench, not afterwards from memory. When the data return, colour a PCA or a similar ordination by every column you recorded. Treatment, date, operator, sequencer, sex, lane. The column that paints the clearest split is your leading suspect.

If the suspect is not your biological factor, include it in the differential model when the design allows, or apply a correction whose assumptions you can state. Then check that the treatment contrast is still estimated from samples that co-occurred inside batches. If the suspect column is identical to the treatment column, stop. No heatmap theme and no software flag repairs perfect confounding. Repeat a balanced subset or narrow the claim to within-batch comparisons that exist.

The analysis conversation can be framed with the differential expression analysis enquiry reference. Bring the sample sheet. An analyst can fit a model to columns you recorded. They cannot reconstruct a column you never wrote down.

When the picture is more convincing than the biology

Clusters on a heatmap often repeat the PCA split. Sorting samples by similarity will group Monday's libraries together and paint a block of genes that are really a block of dates. That picture is discussed as a view in a heatmap is a picture, not a conclusion. The troubleshooting step is the same: put the date under the columns. If the block matches the date, you have found the batch. If a gene also matches a treatment that was balanced across dates, you have a candidate that survived the harder test.

Do not delete samples until the clouds look biological. Removing the inconvenient batch can leave a one-group experiment. Do not normalise the batch away in a figure you show while testing on the uncorrected matrix, or the reverse, without saying so. The reader should be able to see the same adjustment the p-values used.

Research limits

A batch-aware analysis supports a research contrast that the design identified. It does not create clinical validation, and it does not override institutional rules for the samples. Combining other people's studies from a public archive adds their batches to yours. Treat that as a new design problem, not as free replication.

Lots, seasons, and a freezer log

Kit lots that arrive months apart, operators on different shifts, and a sequencer booking that slips into the next season are ordinary sources of batch in a working laboratory. So is the power cut that warms one ultra-low rack overnight. Humidity during extraction days can change yield and inhibition together for every sample processed that afternoon. Write the lot, the date, the operator and any temperature excursion into the sample sheet while they are still known. A balanced book-out across those conditions is the protection. A correction model later can use only the columns that survived.

Specifying this in an enquiry

Tell the person who will analyse, or who will sequence, how many batches you already have, whether groups are crossed with them, and which columns exist. Reagent and instrument classes are in the genomics and sequencing catalogue. The laboratory path is the nucleic acid analysis pathway.

The mRNA sequencing enquiry reference and the differential expression analysis enquiry reference are enquiry references for discussing library work and the model. Put the sample sheet structure on the quote request and ask whether a quotation is possible. A batch factor can be discussed only if you can name it. The pages are there so that specification can be written clearly.