Skip to content

Lecture notes. Written from the material of StatOmics Sga21, licensed CC BY-NC-SA 4.0. These are notes, not a transcript: the material has been reorganised and rewritten. This adaptation carries the same licence, and the original is linked above.

13. RNA-seq versus Microarray Reproducibility

What this covers

This chapter works from a single figure and a fragment of surrounding text — Figure 1 and part of the abstract of Marioni et al.'s study comparing RNA sequencing to microarrays, kept in the course materials as a standalone image. It answers one question: how do you build an experiment so that a claim like "sequencing finds more differentially expressed genes than microarrays" is a fair comparison rather than an artefact of the samples being different, and what did that comparison find? It assumes only that the reader knows what a microarray and RNA sequencing are as expression-profiling technologies, and what a technical replicate is.

Why the comparison needed a paper

By the time high-throughput sequencing became cheap enough to use for expression profiling, microarrays were the established, well-characterised technology. Before switching, two things need checking: is the new measurement reproducible across repeated runs of the same sample, and does it actually do better at the task it would replace the array for — finding genes whose expression differs between conditions? Marioni et al. built one experiment to answer both questions together, by using the same extracted RNA to feed both platforms, so that any difference in the outcome could not be blamed on the two platforms having measured different biological material.

The experimental design

Liver, total RNA Kidney, total RNA mRNA purification mRNA purification Affymetrix array, 3 technical replicates Illumina sequencing, 7 lanes across 2 runs Compare differentially expressed genes between technologies
Figure 1A of Marioni et al.: the same liver and kidney RNA is purified and then split between the array and sequencing platforms, so that both technologies measure the same biological starting material before their results are compared.

Two tissue samples — liver and kidney — were each carried through total RNA extraction and mRNA purification, and then each split between two arms:

  • an array arm: each mRNA sample hybridized to Affymetrix microarrays, in three technical replicates;
  • a sequencing arm: each mRNA sample sequenced on the Illumina platform across seven lanes, spread over two sequencing runs.

Both arms feed the same downstream step: finding differentially expressed genes, and comparing the two technologies' results directly, on the same underlying RNA.

The layout of the sequencing runs (Figure 1B) does not put all of one tissue's lanes together. Each run's eight lanes alternate kidney and liver samples, with the fifth lane of each run shown in a third colour, and most of Run 2's lanes — marked with an asterisk — were sequenced at a lower loading concentration (1.5 pM) than the rest. Spreading each tissue across lanes and across both runs keeps a difference between the tissues from being confounded with a difference between lanes or between runs.

What the comparison found

Three findings, reported from the same dataset:

  • Reproducibility. The sequencing data were highly reproducible, with few systematic differences among the technical replicates.
  • A Poisson model fits the technical noise. For a gene measured by counting reads, the natural model for pure counting noise is the Poisson distribution, in which the variance of a count equals its mean. The variation seen across technical replicates was well captured by exactly this model: only about 0.5% of genes showed counts more variable across replicates than a Poisson model would predict. That gives a statistical basis for treating a departure from Poisson variation as a signal of real, biological differential expression, rather than of measurement noise.
  • More power to detect differential expression. Using a Poisson-based test built on that model, the sequence data identified 30% more differentially expressed genes than a standard analysis of the array data did, at the same false discovery rate. The source also notes the sequencing data's potential to detect alternative splicing — something a fixed microarray probe design cannot do, since a microarray can only measure what its probes were manufactured to look for.

Sources

  • Text and figure: docs/omics-statistics/statomics/sga21/images_sequencing/marioni_fig1.md and its associated image marioni_fig1/figures/p001-1.jpeg, in the knowledge-base-library repository. This is a model-reconstructed conversion (fidelity: reconstructed) of images_sequencing/marioni_fig1.pdf from the statOmics/SGA21 course repository (commit 0ad787d4cc2bb2f4636440840a8a923cf6c09839), licensed CC BY-NC-SA 4.0, converted 2026-09-18. The source PDF has no text layer; the extracted prose follows the note's paraphrase of what appears to be the abstract of Marioni, Mason, Mane, Stephens & Gilad, "RNA-seq: an assessment of technical reproducibility and comparison with gene expression arrays" — the paper the figure is drawn from. The description of the two-panel figure (study workflow and lane layout) is this writer's reading of the image itself; the original PDF carried no machine-readable caption.
  • No slides, transcript, or exercises were supplied for this item — the note above is the entirety of the source material for this chapter. The full paper is referred to by the figure but was not itself part of the supplied material, and neither was whatever lecture originally showed this slide alongside it.

← 12. Genomics: A Molecular Biology Primer · Contents · 14. Detecting Lane Effects in RNA-seq →