How many replicates a simple design needs to discuss

Three wells filled from one flask are one biological sample that you happened to pipette three times. The design question "how many replicates?" is a question about which unit you intend to make a claim about, not a number you can lift from a neighbouring paper and still mean the same thing. This page says what to write down in that discussion. It does not print a sample size. The place of replication in an expression result is sketched in from cells to a gene expression result.

Count the unit you intend to talk about

A replicate is a repeat of the unit in the sentence. If the sentence is about a treatment of mice, the unit is the mouse. If the sentence is about independently grown cultures, the unit is the culture started from a separate seeding or a separate thaw, not a well split from its neighbour after treatment. If the sentence is about donors, the unit is the donor. Cells inside one donor are not extra donors. Aliquots of one RNA extraction are not extra extractions.

Write that unit in the first line of the design note. Then count how many independent units you will have in each group. That count is the n a treatment claim can discuss. Everything else is a nested measurement: technical replicates, sequencing lanes, qPCR wells. Nested measurements can be worth doing. They are not a substitute for the unit.

Pseudoreplication is the name for analysing the nested measurements as if they were the unit. The variance looks small because the wells share their history. The p value looks strong. The next real unit is unimpressed. A simple design avoids this by averaging technical wells, or otherwise accounting for them, and by testing at the biological unit. This page will not pretend a particular software contrast fixes a design that never had independent units.

One unit cannot carry a treatment sentence

With a single treated unit and a single control unit, you have observed two histories. They differ by the treatment and by everything else that differed: the passage, the animal's litter, the hour the extraction started, the column that clogged. No statistical ritual separates those. You may describe the two samples. You may decide the assay is worth repeating. You may not write that the treatment changed expression in this system. The system has not been observed more than once.

People sometimes keep one biological unit and sequence it deeply, or run many primer pairs, and feel the dataset is large. The dataset is wide. It is still n equals 1 on the treatment. Thousands of genes do not create thousands of biological replicates. They create thousands of measurements on the same unreplicated contrast, and a false discovery procedure does not invent the missing units.

A balanced caution applies to a pile of technical replicates. Three qPCR wells on one cDNA estimate noise in the relative expression assay. Report them as technical. If one well is an outlier, you have learned something about the plate. You have not learned the biological spread. The MIQE guidelines ask you to be plain about this distinction. They do not hand you a mandatory n.

Pilots exist to learn the measurement

A pilot is a small run whose job is operational and descriptive. Did the extraction yield RNA you can use? Does the assay see the transcript above the no-template control? Did the sample identifiers survive the journey from freezer to file? What did the spread look like among the few units you could spare? Those answers change the design note. They do not silently become the study.

A pilot is not a power calculator with the output torn off. This article will not invent an effect size, a variance, or a sample size that "achieves" a percentage of power. Those inputs have to come from your system, and a person who owns the calculation should state them. If you have no variance estimate yet, the pilot's job is to start one, not to skip the arithmetic by copying n from a different tissue.

After a pilot, the branch is explicit. If the assay cannot see the transcript, more replicates of a blind assay will not help. Change the assay or the RNA handling. If the assay works and the units are obtainable, design the real contrast at the unit you wrote down. If the units are not obtainable, shrink the claim to what a description of few units can say, or change the question. Do not relabel the pilot as definitive because a heatmap looked tidy.

Public methods on protocols.io show designs of many shapes. The existence of a protocol with a certain n is not evidence that your unit is the same as theirs.

Sentences to write into the design note

A discussion of replicates is a short text, not a vibe. Include these sentences.

The biological unit is this. Examples that force honesty: one mouse, one patient, one plant, one differentiation batch, one flask seeded on a different day from a different vial. "One well" is rarely the unit you mean.

The contrast is this against that. Paired or unpaired. A before-and-after on the same animal is not the same n as two independent groups. Say which one you have.

Batches that could confound the contrast are these: extraction day, operator, library batch, sequencing lane, qPCR plate. Say how units from each group will be spread across those batches. If every treated unit will be processed on Monday, the calendar is not a nuisance. It is the design.

Technical replicates will be handled like this. Averaged before the test, or kept in a model that does not pretend they are biological. State it.

The claim we want to be allowed to write is this. The claim we will not write if the unit count stays at one is the treatment claim. Putting the forbidden sentence in the note is what keeps it out of the abstract.

How many units we can actually obtain is this. Ethics, cost of the animal work, rarity of the tissue, and calendar limits are real. A design that needs units you cannot get is not a design. It is a wish. Discuss a question that fits the units, or gather the units before you sequence.

If a statistician will compute a sample size, the note should already contain the unit, a variance you are willing to defend, and the effect you care about. Without those, the calculation has nothing honest to chew. Do not ask the calculation to supply them.

Units people mix up

A deposited study in the Sequence Read Archive with three files that are splits of one library will look like n equals 3 to a hurried reader. Your sample sheet should make that misreading impossible.

Designs that only look replicated

A common simple design fails in one of four ways. All biological variation is actually one parent culture split after treatment. The groups are processed on different days. The "replicates" are different primers on one sample. Or a pilot with two units is written up with the grammar of a confirmatory study. Each of these can produce a polished figure. The design note is what a reader should be allowed to check against the figure.

If a control fails, the branch is not to drop the awkward unit and keep the word n. A dropped unit changes the design. Record it. If too many units fail, the study is under-replicated relative to its own note, and the claim shrinks. Replacing a failed mouse with a technical rerun of a successful mouse does not fill the gap.

Sequencing depth does not fill it either. Depth improves the measurement of the units you have. It is not a replicate.

No medical sample-size approval

This discussion is about research claims. It does not approve a clinical trial size, a diagnostic study, or a biosafety level. Work with animals or human material follows the institutional and legal review that applies to you. A clear n in a research note is still not an ethics approval. Do not describe a design discussion as a calculation of medical benefit.

Agree the unit before the tubes are split

Two laboratories sharing a study need the unit written before anyone aliquots. A freezer full of tubes from one animal can be labelled as many samples and still be one animal. Humidity, frost and power cuts threaten the labels. They do not create independence. The naming habits that keep those tubes honest are a separate problem. The scientific point is earlier: identity and replication are decided at collection.

If one site grows the treated cultures and another grows the controls, the site is confounded with the treatment. Interleaving is a design sentence, not a statistical apology after the fact. Write where each unit will be grown and processed. A later size factor will not discover that all the controls share a water bath.

What a design discussion should contain

Bring the unit, the contrast, the obtainable number of units, the batch plan, and the sentence you will not write if n stays small. If the measurement is RNA-seq, the library discussion can use the mRNA sequencing enquiry reference. If the end is a differential table, the analysis discussion can use the differential expression analysis enquiry reference. Arrive with the design note. A method page is a prompt for how those units would be assayed, not a source of a default triplicate.

Consumables for the work are grouped in the genomics and sequencing catalogue. The sample-to-result path is nucleic acid analysis. Put the unit and the contrast in the quote request.