Choosing reference genes instead of hoping

GAPDH is a gene with a job in glycolysis. That job is a poor reason to treat the transcript as a constant. ACTB, B2M and 18S ribosomal RNA have jobs of their own, and those jobs change in some experiments. A reference gene, sometimes still called a housekeeping gene, is only a reference if its abundance stays steady across the samples you intend to compare. This page explains how to choose one, or several, for the experiment in front of you. The path from cells to a result is in from cells to a gene expression result. How a relative assay uses the choice is in RT-qPCR for relative expression.

The decision the normaliser is part of

Anyone comparing transcript abundance by RT-qPCR has to divide the target by something. The decision is which transcripts may sit in that denominator for this contrast, this tissue and this time course. Pick badly, and a real change in the reference is reported as a change in the target. Pick well, and the target comparison is less tangled with loading, reverse-transcription yield and ordinary pipetting.

The MIQE guidelines ask for a justification of the number and choice of reference genes, and for a description of the normalisation. That justification is experimental, not traditional. A methods sentence that only says "normalised to GAPDH" has not yet supplied it.

Stability is a property of this sample set

A useful reference transcript is expressed high enough to measure cleanly, and it does not respond to the treatment, the genotype, the developmental stage or the batch you are studying. "High enough" means the Cq sits in the range where your assay is specific and the standard curve, if you ran one, is still linear. "Does not respond" means you measured it across the real groups, including the vehicle, the time-zero flask and the stressed cells. A panel validated in untreated liver does not automatically travel to drug-treated hepatocyte cultures.

Candidates often fail for biological reasons that look, in hindsight, obvious. GAPDH tracks glycolytic flux. ACTB tracks cytoskeletal remodelling and can move with confluence. 18S is abundant, which can push it onto a different dilution than a rare mRNA, and it is not polyadenylated, so an oligo-dT primed reverse transcription treats it differently from an mRNA target. A reference and a target that were not copied under the same priming strategy are a shaky pair.

Confirm you are talking about the intended gene. Ensembl and GenBank are public places to match a symbol to a transcript model in the species you actually have. A primer pair aimed at the wrong paralogue is a stable mistake.

How published stability methods rank candidates

Two method names come up as soon as a laboratory stops hoping. geNorm ranks candidate reference genes by their pairwise stability across the samples: genes that keep a steady ratio to the other candidates rise to the top, and genes that wander fall. NormFinder uses a model of variation within and between groups, which is helpful when your design has clear treatment arms and you need a gene that is stable inside those arms. Other tools exist. They share a habit. They need a panel of candidates measured on the same cDNA set, and they return a ranking for that set.

The ranking is evidence, and it is local. A gene that wins in a 24-sample mouse spleen series can lose in a cell-line drug screen. Read the group structure you give the tool. If you hide the treatment labels from a method that can use them, you may keep a gene that moves in opposite directions inside two groups and looks flat overall. If your groups are batches rather than biology, a "stable" gene may simply be stable to the batch and still useless for the contrast.

When the claim is quantitative, use more than one reference gene. A single transcript is one biological variable dressed as a constant. Two or three genes that agree give the denominator somewhere to stand if one of them twitches. The stability software can help you see whether adding another gene still changes the normalisation. The cutoff you adopt should be written down as your rule for this study, tied to the method you ran, rather than borrowed as a universal pass mark from a different sample type.

A workflow with a failed candidate

List four to eight candidates that are expressed in this tissue and that you do not already believe are part of the response. Include the famous ones if you wish, and include at least a few that are less famous. Design intron-spanning assays where the gene structure allows it, and check specificity the way you would for a target. Measure every candidate on every sample that will enter the comparison, using the same cDNA and the same plate rules.

Rank them with a stated method. Drop any gene that tracks the treatment. If the top gene still shows a clear group difference by eye, or the ranking method flags it, it is a target, not a reference. Replace it and rank again. If the surviving set is stable, those are your reference genes for this experiment, and the target assays are interpreted against them. The arithmetic that then builds a fold change is discussed in delta-delta Cq and what it assumes. That arithmetic assumes the references really are stable. The ranking is how you earn the assumption.

Branch when the panel collapses. If every candidate moves, look hard at RNA integrity and at whether the treatment is so globally disruptive that no transcript is a fair normaliser. You may need an exogenous spike, a different assay, or a claim that stays at the level of "these transcripts changed together" without a single-gene fold change. Do not rescue the analysis by switching to whichever candidate makes the target story neatest. That is hoping with extra steps.

When a famous gene moves, the target story moves with it

Suppose treated cells induce glycolysis and GAPDH rises, while your transcript of interest is unchanged. Normalising to GAPDH manufactures a decrease. The figure looks decisive. The biology you reported was the reference. The defence is the stability plot you made before locking the denominator, plus a second reference that did not rise. If both references rise together, you may be looking at a global RNA-content change or a loading difference, and the target claim waits.

The same trap appears in differentiation time courses, serum-starvation experiments, and comparisons of tumours with adjacent tissue. Proliferation, necrosis and cell-type composition all rearrange "housekeeping" transcripts. A tissue comparison is a composition comparison. A reference gene expressed only in one cell type will track the abundance of that cell type.

RNA quality can imitate instability. If one group is more degraded, long transcripts can look lower, and a short amplicon on a sturdy reference can look stable for the wrong reason. Pair the ranking with the integrity notes from extraction. Where samples were delayed, protecting RNA during extraction is the upstream control, and the reference-gene panel will not repair a group that was left warm.

What the choice does not authorise

A stable reference set supports a relative comparison of the transcripts you assayed, in these samples, under the efficiency assumptions of the calculation you chose. It does not prove a protein change. It does not make the assay diagnostic. It does not license a single gene chosen after the target result was already known. Human and animal material remains an institutional biosafety and ethics decision. This explainer does not set a containment level.

A warm transfer is another condition

A candidate panel validated on RNA that was lysed at the bench can look different on aliquots that travelled across campus in a warm afternoon and arrived partly thawed. Treat the transfer as a condition and either include those samples in the stability test or keep them out of the quantitative set. Humidity and power cuts belong in the sample sheet for the same reason a drug dose does: they can move RNA abundance before any normaliser is applied. Write them down, then decide whether the reference ranking still describes the tubes you will report.

What an enquiry should say about normalisation

If you are sourcing primers, a reverse-transcriptase class or a real-time mix, say how many reference assays you plan to validate and in which species. Those reagent classes sit in the molecular biology catalogue. If the question is transcriptome-wide, reference genes are the wrong sole normaliser, and the design belongs with people who can discuss library type and a differential model. The mRNA sequencing enquiry reference and the differential expression analysis enquiry reference are there for that discussion. The surrounding sample path is the nucleic acid analysis pathway.

Put species, tissue, the contrast, the candidate reference list and whether stability data already exist into the quote request. Ask whether a quotation is possible. Those solution pages are enquiry references a laboratory can use to specify the work. They describe a conversation about method, which you start by writing the contrast clearly.