Experimental design & controls
Biological replicates are independent material and tell you about the population; technical replicates are the same material measured again and tell you about the measurement. Only the first is n, which is why three wells from one flask are n = 1 however tight they look.
Our ruleSimulations prepare you for the bench but don't replace hands-on practice.
Biological vs technical replicates
Biological replicates are independent material and tell you about the population; technical replicates are the same material measured again and tell you about the measurement. Only the first is n, which is why three wells from one flask are n = 1 however tight they look.
How it differs from the standard method
- Spend extra wells on more biological replicates once the technical spread is a third of the biological spread or less.
- Average technical replicates into one value per biological replicate before any statistics; do not feed them in separately.
- Nested designs (cells within animals, sections within blocks) need a model that respects the nesting, or the units must be averaged first.
- Say in the legend what the n is: 'n = 4 independent cultures, each the mean of 2 wells' answers the question before it is asked.
1The design decides what the data can mean
An experiment is a comparison, and a comparison is only as good as the things that are held constant across it. Every experiment therefore has two parts: the one thing you vary on purpose, and the machinery that stops everything else varying with it.
2Biological and technical replicates answer different questions
A biological replicate is independent biological material: a separate culture thawed and passaged on its own, a separate animal, a separate donor. The spread between biological replicates tells you how much the thing you are measuring varies in the population you want to talk about, and that spread is what a comparison between groups has to be judged against.
Independent material
Separate thaws, animals or donors, treated and processed separately. Spread ≈ 20–50% in a cell line, more in primary material. This is the n.
The same material, measured twice
Repeat wells, repeat injections, repeat scans. Spread of a few per cent. Averages them away; never counts as n.
Sub-samples counted as n
Wells split from one flask, cells within one animal, sections from one block. Independence has to hold at the level you want to generalise to.
3What each control rules out
- Positive control
- Something that must respond. It rules out a silent assay failure, and it is the only thing that lets a flat result be reported as a negative rather than as nothing.
- Negative control
- Something that must not respond: a blank, a no-template, an isotype antibody, an untransfected well. It rules out background, carry-over and contamination being read as signal.
- Vehicle control
- The solvent at the same final concentration, from the same bottle, handled identically. It rules out the solvent. DMSO at 0.1% changes proliferation and expression in many cell types, so without it the solvent's effect is credited to the compound.
- Reference or loading control
- A measurement that should not change, used to normalise. It rules out differences in input amount: provided the reference really is unaffected by the treatment, which is worth checking rather than assuming.
4Batch effects, and why the order matters
Running all the controls first and all the treated samples afterwards is faster, tidier and the most common way to ruin an experiment. Reagents drift across a run, plates warm up, operators settle down, a new lot of medium arrives on Tuesday, the outer wells of a plate evaporate and read 10–20% high.
5Blinding, and deciding the analysis first
Wherever the measurement involves judgement (scoring a phenotype, drawing a gate, choosing a field, calling a band) knowing which group a sample came from shifts the number towards what you expect. This is not dishonesty and it is not available to introspection, which is why the control for it is procedural: a colleague relabels the samples with random codes, holds the key, and opens it after the analysis is fixed.
6How many is enough
- The smallest difference you can reliably detect depends on the biological spread and on n, not on how carefully you pipette.
- At a biological SD of 0.30 on a log₂ scale (≈ 22% CV), three replicates per group resolve about a 1.6-fold difference; five resolve about 1.45-fold.
- Technical replicates help only until the measurement error is small compared with the biological spread, which usually happens at two.
- An underpowered experiment that comes out negative tells you nothing, and it costs as much material as a powered one.
- Fix n in advance. Adding samples until the test passes, and stopping when it does, inflates the false-positive rate several-fold.
Key terms
Practise these as flashcards →The smallest thing that could independently have received a different treatment. It is what n counts, and it is often a flask, a cage or an animal rather than a well.
Counting sub-samples of one unit as independent replicates. It shrinks the error bars and inflates the degrees of freedom without adding information.
A nuisance variable that varies with the treatment (day, plate position, operator, reagent lot) so that its effect and the treatment's cannot be separated.
A systematic offset shared by everything processed together. Harmless if both groups are present in every batch; fatal if they are not.
Allocation in which every block (plate, day, cage, litter) contains all treatments, with the order randomised inside the block.
Arranging that the person taking the measurement does not know the group. Most valuable where the readout involves judgement.
Writing down the primary readout, comparison, normalisation and exclusion rule before the data exist.
The chance of detecting an effect of a given size if it is real. Conventionally 80%, which still means one run in five misses a true effect.
The solvent a compound is delivered in. The comparison group gets it at the same final concentration and nothing else.
When to use it
- Before any experiment whose result will be shown to somebody else, which in practice means before you order the reagents.
- When a result does not replicate and you need to work out which of the things that differed could explain it.
- When a reviewer or a supervisor asks what n means in a figure, and the honest answer is not the number of data points.
- When work has to be split across more than one plate, day, operator or reagent lot: that is, almost always.
Compared with alternatives
A design only survives if the plan, the allocation and the deviations were written down at the time.
Where the replicate argument usually surfaces: three wells of one cDNA are not three replicates.
A plate assay in which the controls and the layout are the design: standards, blanks and where the samples sit.
Vehicle controls and plate position decide whether a dose–response curve means anything at all.
Sub-techniques
Choosing one here switches the whole page, including the protocol and the lab, to it.
Biological replicates are independent material and tell you about the population; technical replicates are the same material measured again and tell you about the measurement. Only the first is n, which is why three wells from one flask are n = 1 however tight they look.
- Spend extra wells on more biological replicates once the technical spread is a third of the biological spread or less.
- Average technical replicates into one value per biological replicate before any statistics; do not feed them in separately.
- Nested designs (cells within animals, sections within blocks) need a model that respects the nesting, or the units must be averaged first.
- Say in the legend what the n is: 'n = 4 independent cultures, each the mean of 2 wells' answers the question before it is asked.