Biological versus technical replicates

Three wells from one flask are three measurements of one culture. They tell you whether the pipette and the instrument agree. They do not tell you whether the next flask, the next mouse, or the next donor would agree. That distinction is the whole of biological versus technical replication. This selection guide helps you choose the unit that matches the sentence you want to write, for RT-qPCR and for RNA-seq. The study frame is from cells to a gene expression result.

Start from the sentence, then count the units

Write the claim in plain language and underline the noun that would have to be true again. "This drug changes this transcript in this cell line" needs independently treated cultures, not three holes of one plate split after lysis. "This genotype changes expression in this mouse strain" needs mice. "This tumour type differs from matched adjacent tissue" needs donors, and the matching is part of the design. The underlined noun is the biological replicate. Everything you repeat inside that noun is technical.

People get into trouble when the noun is vague. "We had nine replicates" might mean nine mice, or three mice with three libraries each, or one RNA tube aliquoted nine times. Those designs support different claims. Public archives such as the Sequence Read Archive are full of experiments whose replicate labels only make sense if the methods section defines them. Copy that discipline into your own sample sheet before anyone builds a count matrix.

What each kind of replicate is allowed to capture

A biological replicate captures the variation of the unit you care about: how cultures diverge after independent seeding and treatment, how animals differ under the same nominal dose, how donors differ inside a clinical category. That variation is usually larger than instrument noise, and it is the variation a treatment claim has to beat.

A technical replicate captures the measurement process. Splitting one cDNA across three qPCR wells estimates pipetting and well noise. Building two libraries from one RNA aliquot estimates library-prep noise. Running the same library on two lanes estimates a slice of sequencing noise. Those numbers are useful when a method is new or a sample is precious and you need to know the assay is stable. They become a problem only when they are silently promoted to biological n.

There is a middle layer that deserves its own name. Two pieces of one tumour, extracted separately, are not two patients. They are closer to subsampling of one biological unit. Two passages of one cell line started on different days can be biological replicates for a claim about that line under your protocol, and they are still one genetic background. Say the layer out loud. A nested design, technical inside biological, is a respectable design. Pretending the inner layer is the outer one is pseudoreplication.

The differential expression analysis enquiry reference is an enquiry page for the modelling side of this hierarchy. Read it as a prompt for what to specify. The biological unit has to be decided by the laboratory that owns the question. An analysis cannot invent mice that were never treated.

Choosing the set you can actually defend

Match the budget to biological units first. If you can prepare twelve RNA samples, twelve independent cultures spread across the groups teach you more about a treatment than three cultures sequenced four times each. Add technical replication where the measurement is the fragile step: a new extraction method, a low-input library, a qPCR assay you have not yet watched for well-to-well scatter. Once that scatter is known and small, further technical copies add less than another biological unit.

For a qPCR plate, technical triplicates are a common way to catch a sealing failure. Average them, or model them as nested, and let the biological cultures be the n in the test. For RNA-seq, a second library from the same RNA is rarely the best use of a limited run if you still have only two animals per group. Depth and library type still matter. They are covered, for the selection chemistry, in RNA-seq library types: poly(A) and ribodepletion and, as a technology class, in the Illumina sequencing overview. Neither one multiplies mice.

Branch when the design collapses. If a contamination event destroys two of three biological replicates in one group, you no longer have the n you planned. Do not fill the gap with technical re-sequencing of the survivor and keep the original statistical sentence. Report the loss, or repeat the biological unit. If batch and treatment are tied together because all replicates of one group were grown on one day, you have a confounding problem. That is batch effects in an expression study, and extra technical wells will not untie it.

Pseudoreplication that looks like a strong result

The failure mode is a tiny p-value. Technical replicates sit close together, so a test that pretends they are independent finds a treatment effect with false confidence. The figure looks clean because the noise you included is the small noise. Biological scatter, the noise the claim must survive, never entered the model. Reviewers who ask "were these independent cultures?" are asking you to redraw the unit.

A second failure is averaging too early and losing the hierarchy. If you collapse technical wells without recording how much they differed, you cannot show the assay was stable. Keep the technical spread in the supplement or the notebook, and keep the biological points in the plot that carries the claim. A third failure is mixing units across methods. Single-cell counts from one well are many cells and one capture. They are not many donors. Bulk replicates and single-cell captures answer different n questions even when the tissue is the same.

Method collections on protocols.io show how workflows are written. They do not define your unit of inference. That definition belongs in the protocol you sign before the harvest.

Limits of what replication can promise

No replicate scheme in this page is a sample-size calculation for your effect size. A statistician, a pilot variance and the size of difference you care about belong in that calculation. This guide only stops the common category error. Replication also does not make an assay diagnostic, and it does not replace biosafety review for the cells or animals you will use. Institutional rules decide containment and ethics. Extra tubes do not.

One rack and an overnight outage

If a power cut warms the freezer and only the treated group's box was in the affected rack, you have not lost "some technical backup". You have damaged a biological group in a way that confounds treatment with thaw. The remaining frozen controls are not a fair pair. Note the excursion, and do not restore balance by re-measuring the surviving RNA three times. Replace the biological units, or analyse only what the damaged set can still support. Heat and humidity during harvest belong in the same log, because a culture left on a warm bench is a different treatment from the one in the protocol.

What to put in the enquiry

Name the biological unit, how many you have per group, whether technical replicates exist, the species, the library preference and the contrast. Those facts decide whether a sequencing conversation is even the right shape. Instrument and reagent classes sit in the genomics and sequencing catalogue. The wider path is the nucleic acid analysis pathway.

The mRNA sequencing enquiry reference and the differential expression analysis enquiry reference are enquiry references you can use to frame the work. Send the replicate hierarchy with the quote request and ask whether a quotation is possible. A design can be discussed from that note. The pages help you specify the unit of inference. They do not add biological replicates you have not prepared.