Microarray and bulk RNA-seq answer the same question — which transcripts changed, and by how much — but they fail in different ways. Microarrays fail through background and probe saturation; RNA-seq fails through composition bias and mapping ambiguity. The analyst's job is to know which failure mode is in play before running any differential model, because normalisation strategy determines every downstream p-value.
Microarray normalisation with RMA
The Robust Multi-array Average (RMA) pipeline is the reference preprocessing for Affymetrix-style arrays, and its three steps each matter. First, background correction removes optical and non-specific signal. Second, quantile normalisation forces the intensity distribution of every array onto a common distribution, which removes array-level scaling artifacts but assumes the overall distribution should be identical across samples — a bad assumption when global expression genuinely shifts. Third, log2 transformation followed by median-polish summarisation aggregates the multiple probes of one probeset into a single value.
- RMA normalises at the probeset level and does not require a separate scaling step.
- Single-channel arrays: express log-ratios relative to a reference or use limma's linear-model framework directly.
- Never mix RMA-processed and MAS5-processed arrays in the same model.
Batch is the dominant confounder in legacy microarray studies. Because array batch is often perfectly confounded with the biological group, no normalisation method can rescue the design; check the sample-to-batch mapping before believing any result.
RNA-seq alignment with STAR and the splice problem
RNA-seq reads spanning exon-exon junctions cannot be aligned by a naive DNA aligner. STAR solves this with a suffix-array seed search followed by stitching and scoring of spliced alignments against a reference genome and annotation, which makes it fast and accurate but memory-hungry — tens of gigabytes of RAM for a mammalian genome index. STAR is the right choice when you need spliced genome alignments for downstream splicing, fusion, or variant analysis.
The alternative is a lightweight path: align against the transcriptome with a pseudo-aligner or a selective-alignment method. This trades the ability to discover novel junctions for speed, low memory, and quantification accuracy for known transcripts.
Quantification with Salmon and its pitfalls
Salmon performs selective alignment and then models the probability that each fragment originated from each transcript, using an expectation-maximisation procedure with the equivalence-class structure of reads. Crucially it corrects for GC bias and sequence-specific bias, and it produces both counts and Transcripts Per Million (TPM). Transcripts that share sequence — isoforms, paralogs, pseudogenes — remain the hard case; equivalence classes may not identify which one the fragment came from, and effective length estimation becomes unstable.
One practical rule: keep counts, not TPM, for differential testing. TPMs are already length- and depth-normalised, and feeding them into a count model violates the model's assumptions and inflates false positives on long transcripts.
Differential expression with DESeq2 and edgeR
Both DESeq2 and edgeR model counts with a negative binomial distribution, and both shrink dispersion estimates toward a fitted trend to borrow strength across genes — the key insight that makes RNA-seq testing stable with few replicates. They differ in the specifics: DESeq2 uses median-of-ratios size factors and shrinks log-fold-changes with an empirical-Bayes prior, while edgeR offers TMM normalisation and quasi-likelihood or likelihood-ratio tests. The default recommendation is to pick one, fit the full model with batch as a covariate, and report the design.
| Method | Model | Normalisation |
|---|---|---|
| DESeq2 | Negative binomial, empirical-Bayes shrinkage | Median of ratios |
| edgeR | Negative binomial, quasi-likelihood | TMM |
| limma-voom | Linear model on log-CPM with precision weights | TMM / none |
Composition bias still bites: if a handful of genes dominate the library, size factors shift for everyone. Interpret a "global" downregulation signature with suspicion, and inspect the raw library sizes and top-expressed genes.
Design matters more than the software
Paired designs, blocking on batch, and interaction terms (for example, genotype by treatment) all belong in the model formula, not in separate comparisons. Because the negative-binomial model uses dispersion estimates shared across genes, adding a covariate that explains real variance improves power across the board. Skipping it does not merely lose power — it manufactures false positives when the covariate is unbalanced.
GO and KEGG enrichment
Enrichment tests ask whether a gene set is over-represented among the significant genes more than chance would predict. Over-representation analysis (ORA) uses a hypergeometric test on a cutoff list; gene set enrichment analysis (GSEA) avoids the arbitrary cutoff by ranking all genes and testing whether a set's members concentrate at the top or bottom of the ranking. Report the background gene universe explicitly — using the whole genome instead of the tested genes is a common and serious error — and correct for multiple testing with the Benjamini-Hochberg false discovery rate.
KEGG and GO annotations are curated, incomplete, and biased toward well-studied genes. Enrichment tells you what the annotation community already knew about your gene list, which is useful but not discovery. Treat a pathway hit as a hypothesis, and be honest that redundancy between overlapping GO terms can inflate a single biological theme into dozens of significant rows.
References
- Irizarry RA, Hobbs B, Collin F, et al. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics. 2003;4(2):249–264.
- Ritchie ME, Phipson B, Wu D, et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015;43(7):e47.
- Dobin A, Davis CA, Schlesinger F, et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29(1):15–21.
- Patro R, Duggal G, Love MI, Irizarry RA, Kingsford C. Salmon provides fast and bias-aware quantification of transcript expression. Nat Methods. 2017;14(4):417–419.
- Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15(12):550.
- Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics. 2010;26(1):139–140.
- Bray NL, Pimentel H, Melsted P, Pachter L. Near-optimal probabilistic RNA-seq quantification. Nat Biotechnol. 2016;34(5):525–527.
- Subramanian A, Tamayo P, Mootha VK, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci USA. 2005;102(43):15545–15550.
- Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Series B. 1995;57(1):289–300.