eduKateSG · Why Science?
Keep transcript-quantification uncertainty in the analysis while using flexible, rank-based comparisons across biological samples
Reading routes
Science learning becomes useful when a familiar object or observation is turned into a system of quantities, mechanisms and claim limits. This guide owns one applied evidence-reading job inside eduKateSG’s wider Science estate. It connects naturally to Why Science Sleuth Bootstrap Quantification Uncertainty Rna Seq; Why Science Ebseq Empirical Bayes Gene Isoform Rna Seq; Why Science Limma Voom Precision Weights Linear Models Rna Seq; Why Science Single Cell Rna Sequencing Barcodes Transcriptome Heterogeneity Evidence; Education Hub; Singapore Secondary School Directory; Career Adulthood Hub; Science Learning Hub. It also keeps current school and public claims traceable to visible primary sources: Swish primary study; Official Bioconductor fishpond package; Official Swish vignette, 28 April 2026; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science syllabus; 2026 MOE G2 Computing syllabus. The sources describe the scientific scope; this article translates that scope into a calm route for Primary Science, PSLE Science, Secondary Science, O-Level Science, STEM exploration, school choices and career pathways without inventing admission or employment outcomes.
Swish is a nonparametric differential-expression method in the Bioconductor fishpond package. It extends ideas from SAMseq while integrating inferential replicates—such as bootstrap or Gibbs samples from transcript quantifiers—so uncertainty from ambiguous read assignment is carried into the comparison. The current official vignette documents two-group, paired, batch-aware and interaction workflows and suggests that roughly 20 to 30 inferential replicates are generally sufficient. Those replicates quantify uncertainty; they never replace independently sampled biological units.
Inside this guide
1–12 · Foundations and models
- 1. Begin with the scientific unit
- 2. Separate measurement from biology
- 3. Write the comparison first
- 4. Meet the uncertainty
- 5. Identify the central assumption
- 6. Keep the claim bounded
- 7. Read the primary evidence in scope
- 8. Build a trustworthy input matrix
- 9. Create a sample sheet
- 10. Inspect library composition
- 11. Check sample relationships
- 12. Filter by information
13–24 · Evidence, testing and applications
- 13. Confirm the design is estimable
- 14. Choose normalisation deliberately
- 15. Practise with a fictional evidence table
- 16. Run the method with recorded settings
- 17. Interpret the central statistic
- 18. Correct for many features
- 19. Read effects beside probabilities
- 20. Plot every biological sample
- 21. Inspect the method-specific diagnostic
- 22. Challenge the key assumption
- 23. Test filtering sensitivity
- 24. Test normalisation sensitivity
25–36 · Learning, decisions and pathways
- 25. Test sample influence
- 26. Test parameter sensitivity
- 27. Compare an adjacent method
- 28. Build an evidence card
- 29. Connect to fair testing
- 30. Try a classroom activity
- 31. Connect to mathematics
- 32. Connect to computing literacy
- 33. Connect to education pathways
- 34. Create a family question habit
- 35. Use precise scientific language
- 36. Finish with an evidence bundle
Section 1 of 36
1. Begin with the scientific unit
Begin with independently collected biological samples. Inferential replicates describe quantification uncertainty inside each sample and cannot increase biological sample size.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 2 of 36
2. Separate measurement from biology
Transcript abundance is estimated from fragments that may match several isoforms. The same reads can support multiple plausible allocations.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 3 of 36
3. Write the comparison first
State whether the comparison is gene- or transcript-level, paired or unpaired, and whether the target is a main effect or interaction.
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 4 of 36
4. Meet the uncertainty
Inferential replicates quantify how abundance estimates vary under the quantifier’s model. Swish carries that variability into nonparametric testing.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 5 of 36
5. Identify the central assumption
Label permutations must match the experimental design. Exchangeability fails when paired, blocked or batch-structured samples are permuted as independent.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 6 of 36
6. Keep the claim bounded
A Swish result supports a measured differential-expression pattern under the quantification and permutation scheme, not direct regulatory mechanism.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
Section 7 of 36
7. Read the primary evidence in scope
Zhu and colleagues described Swish and evaluated inferential-replicate-aware nonparametric inference. Read accuracy findings within those simulations and datasets.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 8 of 36
8. Build a trustworthy input matrix
Import compatible quantifications and inferential replicates through a documented pipeline. Record index, annotation, quantifier, tximport or tximeta settings and aggregation.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 9 of 36
9. Create a sample sheet
Store condition, pair, batch and interaction variables with exact sample identifiers. Verify levels before running label-preserving transformations.
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 10 of 36
10. Inspect library composition
Inspect totals, detection, transcript ambiguity and scaling factors. Very low information can produce unstable ranks regardless of the test.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 11 of 36
11. Check sample relationships
Use sample ordination after appropriate transformation to spot unexpected groups, then trace any anomaly back to metadata and raw quality information.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 12 of 36
12. Filter by information
Filter low-information features with a declared rule before testing. Retain the rule and feature universe for reproducibility.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
Section 13 of 36
13. Confirm the design is estimable
Ensure every factor combination has independent support. An interaction with an empty or single-sample cell cannot deliver a robust biological inference.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 14 of 36
14. Choose normalisation deliberately
Apply the documented scaleInfReps workflow and inspect scale factors. Normalisation assumptions remain part of the evidence.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 15 of 36
15. Practise with a fictional evidence table
This classroom table is invented and is not a result from the Swish study.
| Fictional transcript | Log2 fold change | Swish statistic | q-value | Reading |
|---|---|---|---|---|
| TX-W1 | 1.2 | 3.7 | 0.012 | Supported increase |
| TX-W2 | -0.8 | -1.1 | 0.310 | Weak evidence |
| TX-W3 | 1.9 | 2.4 | 0.064 | Uncertain |
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 16 of 36
16. Run the method with recorded settings
Run labelKeep or the documented preparation for the design, set a reproducible seed, then call swish with the correct covariates and interaction specification.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 17 of 36
17. Interpret the central statistic
Interpret the signed statistic, fold change and q-value together. Confirm the sign convention and factor reference levels before writing a biological direction.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 18 of 36
18. Correct for many features
Adjust across all tested features and intended contrasts. A small permutation-derived value still belongs to a large decision family.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
Section 19 of 36
19. Read effects beside probabilities
Show effect magnitude and inferential uncertainty. A large estimated fold change can remain poorly determined when reads map ambiguously.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 20 of 36
20. Plot every biological sample
Plot abundance for every biological sample and, where useful, its inferential-replicate spread. Keep those two levels visually distinct.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 21 of 36
21. Inspect the method-specific diagnostic
Inspect feature-level inferential relative variance or related uncertainty summaries. High uncertainty should narrow claims, not be silently filtered after results.
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 22 of 36
22. Challenge the key assumption
Challenge whether inferential replicates came from a compatible quantifier and preserve the ambiguity relevant to the selected transcript or gene aggregation.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 23 of 36
23. Test filtering sensitivity
Repeat defensible low-count filters and record rank, direction and q-value changes among headline features.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 24 of 36
24. Test normalisation sensitivity
Compare justified scaling or aggregation choices, especially when a few transcripts dominate or gene-level summarisation changes uncertainty.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
Section 25 of 36
25. Test sample influence
Leave out one biological sample or pair at a time when possible. Record features whose direction or discovery status depends on one unit.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 26 of 36
26. Test parameter sensitivity
Compare roughly 20 and 30 inferential replicates when available, as the vignette says this range is generally sufficient; archive the exact subset and seed.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 27 of 36
27. Compare an adjacent method
Compare with sleuth only after aligning quantifier, inferential replicates, feature level and design. Nonparametric and model-based outputs need not rank identically.
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 28 of 36
28. Build an evidence card
Record quantifier, inferential-replicate type and count, imports, scaling, labels, pairing, seed, tests, plots, sensitivities, versions and validation.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 29 of 36
29. Connect to fair testing
Repeated measurement is not repeated sampling. Swish offers a modern lesson in distinguishing uncertainty about a measurement from variation among organisms.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 30 of 36
30. Try a classroom activity
Give groups ambiguous fragments and several plausible allocations per sample, then compare whether a condition ranking survives those allocations.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
Section 31 of 36
31. Connect to mathematics
Ranks, permutations, logarithms, variance and false-discovery adjustment turn uncertain transcript estimates into a disciplined comparison.
Inferential replicates are repeated abundance estimates produced by a quantifier to represent uncertainty from finite reads and ambiguous transcript assignment.
Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.
Section 32 of 36
32. Connect to computing literacy
Object classes, seeds, identifiers and parallel settings must be recorded. A reproducible analysis is an executable scientific explanation.
Swish combines biological-sample permutations with inferential-replicate information. Quantification resamples remain nested uncertainty estimates, not new biological samples.
Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.
Section 33 of 36
33. Connect to education pathways
Transcriptomics connects genetics, probability and software, with pathways into biotechnology, biomedical research, statistics and computational biology.
The official workflow scales counts, prepares inferential data and runs swish with a recorded seed; exact object structure and function arguments should be archived.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.
Section 34 of 36
34. Create a family question habit
When many inferential replicates are shown, ask how many independent biological samples were collected and whether the design constrained permutations correctly.
Matched or paired samples require restricted label permutations so within-pair structure is preserved. Ignoring pairing wastes design information and can distort the null comparison.
Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.
Section 35 of 36
35. Use precise scientific language
Say that Swish supports differential expression while accounting for quantification uncertainty from stated inferential replicates and design restrictions.
Interaction analyses ask whether a condition effect changes across another factor. They need adequate samples in every relevant combination and precise contrast language.
Build an evidence card naming the question, measured material, statistical unit, model, contrast, result, alternatives, validation status and narrowest defensible claim.
Section 36 of 36
36. Finish with an evidence bundle
Deliver raw-data references, annotations, quantifications, inferential replicates, metadata, code, seeds, complete tests, plots, sensitivities, versions and bounded claims.
Rank-based robustness does not remove the need for filtering, sample independence, correct transcript identifiers, multiple-testing control and biological validation.
Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.
## Source date and scope note Primary studies and current official software documentation were checked for this article on 11 October 2026. Package interfaces, recommendations and dependencies can change; readers should consult the linked official package page and vignette before reproduction.
## Final reader checklist Before accepting a claim, confirm the biological sample count, condition definition, measured material, annotation, preprocessing, statistical unit, exact comparison, effect size, uncertainty, software versions, controls, sensitivity checks, independent validation and the boundary between association and causation.
Contents · Previous section · Continue to the Science Learning Hub