BRM5012 Chap.3 RNA-seq Quality Control
RNA-seq Quality Control
RNA-seq quality control combines expression distributions, sample relationships and metadata to decide whether a planned comparison is credible. This chapter explains median-centred normalised-count boxplots, UMAP or related sample maps, outlier investigation, batch effects and complete confounding.
It emphasises plot-to-claim reasoning: describe the visible pattern, test a plausible mechanism, choose a proportionate action and document what changed.
What this chapter covers
- 01
Normalised-count distributions
- 02
Sample relationship maps
- 03
Replicate coherence
- 04
Outlier investigation
- 05
Batch metadata overlays
- 06
Suitability decisions
- 07
Reproducible exclusions
Worked example · free
Rank three synthetic datasets
- +1Rank A first because distributions are comparable and replicate structure follows the intended contrast.
- +1Rank B second but investigate the isolated control before retention or exclusion.
- +1Rank C last because processing day dominates the sample relationship pattern.
- +1Use metadata and raw-read checks to test each proposed mechanism.
- +1State that total detected genes alone does not measure reliability.
Key terms
- Normalised counts
- Expression values adjusted to support meaningful comparison across samples under the chosen method.
- Sample outlier
- A sample whose quality or expression relationship differs enough to require targeted investigation.
- Batch effect
- A systematic technical pattern associated with handling, date, operator, site or platform.
- Suitability judgement
- A documented decision about whether data can support the planned downstream comparison and under what reservations.
RNA-seq Quality Control FAQ
What does a centred boxplot tell me?
Similar median-centred distributions support cross-sample comparability after normalisation. They do not prove that every sample is high quality or that biology is correct, so interpret them with sample relationships, raw-read metrics and metadata.
Should an isolated sample be removed?
Not immediately. Check labels, raw-read quality, mapping, library metrics, contamination and handling metadata. The sample may be mislabeled, technically poor, biologically heterogeneous or a valid response, and each explanation requires a different action.
Can a model repair complete confounding?
No. If every control uses one platform and every case another, platform and condition cannot be estimated separately. Adding both labels to a model does not create information that the design lacks. Narrow the conclusion or collect balanced data.
Assessment move
Practise QC as an evidence sequence rather than a gallery of plots. For each view, write three short lines: the visible pattern, the plausible mechanism and the consequence for analysis. Overlay condition and technical metadata on sample maps and ask which label best explains the strongest structure. When you suspect an outlier, list the checks that would separate mislabelling, low quality and real biology before deciding.
Keep a reproducible QC note with input files, software, parameters, observations and actions. Compare results before and after any exclusion or correction. In workshop discussion, challenge rankings that rely only on detected-gene totals or a dramatic processing anecdote. The final paragraph should state the strongest quality signal, the main concern, the suitability decision and the safeguard used downstream.
Build a QC matrix with plots as rows and candidate explanations as columns. For a count distribution, sample-correlation view and dimension-reduction map, record what is directly visible before assigning a cause. Add condition, donor, extraction batch, library batch and sequencing run as overlays. If one label follows the main separation, say whether that pattern was intended by design or threatens the comparison.
For an apparent outlier, prepare a decision tree that distinguishes a swapped label, low library complexity, contamination, technical failure and genuine extreme biology. Each branch must name another observation that would support or weaken it. Avoid voting by plot count because several plots may reflect the same underlying signal. Instead, combine independent evidence and record uncertainty.
Practise explaining why high detected-gene totals do not by themselves guarantee a useful sample and why a documented processing problem matters only if its consequence is visible. When an exclusion is considered, rerun the relevant summaries with and without the sample, compare effect direction and note whether the design becomes unbalanced.
If correction is applied, verify that the biological signal of interest remains rather than celebrating a visually tighter cluster. Maintain a short audit table containing observation, hypothesis, check, action and downstream sensitivity. Use it to write a bounded QC conclusion: data are suitable, suitable with a safeguard, or unsuitable for the proposed comparison.
Finish by identifying which later result would be most vulnerable if the concern were real. This links quality assessment to differential analysis instead of treating QC as a decorative preliminary section.
Working through RNA-seq Quality Control in BRM5012? Sia is AskSia’s AI Biology tutor — ask any BRM5012 RNA-seq Quality Control question and get a clear, step-by-step explanation grounded in how BRM5012 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
