Expression microarrays
Measure thousands of transcripts by hybridization; largely replaced by RNA-seq.
Our ruleSimulations prepare you for the bench but don't replace hands-on practice.
1Twenty thousand hybridisations at once, on a fixed list
A microarray is a glass slide or chip carrying hundreds of thousands of short DNA probes, printed or synthesised in known positions, each one complementary to a chosen transcript. You label the RNA from your sample, pour it over the array, let it hybridise overnight, wash off what did not stick, and scan the fluorescence.
The probe set
You see what was printed. One probe set per gene, usually biased to the 3′ end, so isoform switches are invisible and anything unannotated does not exist as far as the experiment is concerned.
Specificity
A probe reports a sequence neighbourhood, not a gene. Permissive hybridisation and washing let a paralogue's transcript be counted as yours; stringent conditions remove it and take signal off the dimmest spots with it.
Dynamic range
A detection floor from background and read noise, a 16-bit ceiling at the top, and a saturating hybridisation isotherm in between. Two to three decades of usable range, against five or six in a transcriptome.
Comparability
No two arrays take up the same amount of label or scan at the same brightness, so raw intensities are not comparable. Normalisation is part of the measurement, not a tidying step afterwards.
2Why the dynamic range is the real limitation
Each feature carries a finite number of probe molecules, so the relationship between transcript concentration and signal is a saturating binding curve: at low concentration the signal is proportional to the amount present, and as the spot fills up, each extra molecule of target buys less signal. At the bottom, non-specific binding and the detector's own read noise set a floor, and anything below it is simply not detected.
3One colour or two, and why a dye swap is a design and not a calculation
In a one-colour design each sample is labelled with the same dye and hybridised to its own array, so comparisons are made between images. In a two-colour design two samples are labelled with Cy3 and Cy5 and hybridised to the same array, so the ratio is measured spot by spot and anything that varies between arrays (spot size, local hybridisation, printing) cancels.
- Probe / feature
- An immobilised oligonucleotide (or set of them) at a known position on the chip, complementary to one target. Several probes summarised together make a probe set.
- MA plot
- Log ratio (M) against mean log intensity (A), one point per feature. The diagnostic plot for an array: every characteristic artefact has a recognisable shape on it.
- Loess normalisation
- Fitting M as a smooth function of A and subtracting the fit. The only common normalisation that removes a bias which depends on intensity.
- Quantile normalisation
- Forcing every array to the same distribution of intensities. Removes global differences completely; cannot touch an intensity-dependent bias within an array.
- Dye swap
- Repeating a two-colour hybridisation with the dye assignments exchanged. Averaging the two log ratios cancels dye bias, including the per-sequence part.
- Spike-in controls
- Synthetic transcripts at known concentrations added before labelling. They locate the detection floor, the top of the linear range and whether labelling worked.
- Fold-change compression
- The systematic understatement of large changes caused by the saturating hybridisation curve. Worse for abundant transcripts and worse after two rounds of amplification.
- False discovery rate
- The expected proportion of false positives among the genes called significant. The necessary replacement for p < 0.05 when twenty thousand hypotheses are tested together.
4Twenty-two thousand tests in one experiment
Testing 22,000 probe sets at p < 0.05 produces about 1,100 significant results when nothing at all is happening. That is not a subtlety; it is the dominant feature of the statistics, and the reason the false discovery rate was adopted so quickly in this field.
Key terms
Practise these as flashcards →The saturating relationship between target concentration and spot signal, set by the finite number of probe molecules on a feature.
The signal below which a spot cannot be distinguished from background, set by non-specific binding and the detector's fixed read noise.
The top of the detector's digital range, usually 65,535 counts. A clipped spot reports that number whatever is on it.
A probe binding a related sequence, typically a paralogue. Produces a confident, reproducible, entirely false change.
Putting arrays or channels on a common scale before anything is compared. Part of the measurement, and different methods remove different artefacts.
A systematic, gene-specific difference between samples processed at different times. Removable only if it is not aligned with the comparison of interest.
An independent sample taken through the whole workflow. The only kind of replicate that supports a statement about biology.
Combining the several probes of a probe set into one value (RMA, MAS5 and relatives), with its own background and normalisation assumptions.
When to use it
- Screening a fixed, well-annotated list of genes across many samples, where the price per sample and a standardised protocol matter more than finding anything unannotated.
- Large cohort or clinical studies with locked-down content, where the assay has to behave identically for years and be comparable to earlier batches.
- Genotyping and methylation work, where array formats remain the standard tool because the list of positions genuinely is fixed in advance.
- Re-analysing or extending the very large public archive of array experiments, which needs an understanding of the method's blind spots to read honestly.
- Teaching and demonstrating hybridisation-based measurement, where the array makes the trade-offs between sensitivity, specificity and dynamic range unusually visible.
Compared with alternatives
No probe list, a dynamic range set by depth rather than optics, isoform-level resolution and the ability to find the unexpected. The default choice now for expression work; arrays win only on cost per sample at high throughput with fixed content.
The standard way to confirm array hits: far better dynamic range and precision for a handful of genes, and specificity you can check. An array ranks candidates; qPCR measures them.
The same hybridisation logic with one probe instead of twenty thousand, and the thing an array fundamentally cannot do: report transcript size and isoform state directly.
Answers a question arrays never could, since an array measures the average of whatever cells went into the tube. Worth considering whenever a change might be a change in cell composition.
Where array formats are still routine: a fixed list of CpG positions is exactly what an array is good at, and the content being locked down is an advantage rather than a limitation.
SNP arrays remain the standard tool for large cohorts for the same reason: the positions are known in advance, so nothing is lost by fixing the list.