eduKateSG · Why Science?
Turn replicated sequencing counts into an auditable comparison—while keeping normalisation, dispersion and biological design in full view
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 Single Cell Rna Sequencing Barcodes Transcriptome Heterogeneity Evidence; Why Science Muscat Multi Sample Multi Group Differential State Analysis; Why Science Dreamlet Pseudobulk Mixed Models Complex Single Cell Cohorts; Why Science Zinb Wave Zero Inflated Counts Low Dimensional Signal; 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: DESeq2 primary study; Official DESeq2 workflow; Official Bioconductor DESeq2 package; 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.
DESeq2, introduced by Love, Huber and Anders in Genome Biology in 2014, models read counts with negative-binomial generalised linear models. It estimates sample-size factors, learns a mean–dispersion relationship, moderates gene-wise dispersion estimates and tests declared model coefficients or contrasts. Current Bioconductor documentation also separates statistical testing from optional log2 fold-change shrinkage. The result is not a list of genes that are automatically important: sample design, count provenance, independent replication, coefficient meaning and diagnostic checks define the claim.
Inside this guide
1–12 · Foundations and models
- 1. Begin with a replicated count question
- 2. Separate reads from expression
- 3. Understand the negative-binomial model
- 4. Define the design matrix
- 5. Distinguish coefficient from contrast
- 6. Choose the independent unit
- 7. Read the 2014 evidence in scope
- 8. Build a trustworthy count matrix
- 9. Build sample metadata beside counts
- 10. Inspect library and detection patterns
- 11. Filter without peeking at outcomes
- 12. Check design rank and confounding
13–24 · Evidence, testing and applications
- 13. Estimate size factors
- 14. Estimate gene-wise dispersion
- 15. Practise with a fictional DESeq2 table
- 16. Moderate dispersion estimates
- 17. Fit the generalised linear models
- 18. Test the declared comparison
- 19. Control the testing family
- 20. Shrink fold changes for suitable tasks
- 21. Read MA plots carefully
- 22. Inspect sample-level expression
- 23. Review Cook's-distance flags
- 24. Check transformed data for exploration
25–36 · Learning, decisions and pathways
- 25. Run leave-one-sample sensitivity
- 26. Compare defensible formulas
- 27. Compare with edgeR
- 28. Build a DESeq2 evidence card
- 29. Connect DESeq2 to school science
- 30. Try a library-size activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career pathways
- 34. Create a family evidence habit
- 35. Use precise DESeq2 language
- 36. Finish with a counts-to-claim bundle
Section 1 of 36
1. Begin with a replicated count question
Differential expression starts with a comparison among independent biological samples. Name the organism, tissue or cell population, condition, time point and experimental unit before opening any software.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 2 of 36
2. Separate reads from expression
A read count is a measurement produced by library preparation, sequencing, alignment or quantification and feature summarisation. It is related to abundance, but it is not a direct count of molecules in the original organism.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 3 of 36
3. Understand the negative-binomial model
Counts vary because of sampling and biological differences. A negative-binomial model allows the variance to exceed the mean, giving replication a formal place in the analysis.
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 4 of 36
4. Define the design matrix
The design formula translates the experiment into model columns. Terms for batch, condition, pairing or interaction must be estimable from the available samples.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 5 of 36
5. Distinguish coefficient from contrast
A coefficient follows the factor coding; a contrast combines coefficients into a specified comparison. Results can be numerically correct yet scientifically wrong when that mapping is misunderstood.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 6 of 36
6. Choose the independent unit
Replicate people, animals, cultures or other independently assigned units support inference. Technical lanes, repeated reads and thousands of cells from one sample do not replace them.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
Section 7 of 36
7. Read the 2014 evidence in scope
Love, Huber and Anders described moderated estimation of fold changes and dispersions for RNA-seq count data. The study evaluates a statistical workflow, not a universal guarantee for every assay or design.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 8 of 36
8. Build a trustworthy count matrix
Use integer-like feature counts from a documented quantification path. Confirm row identifiers, column order, strandedness, annotation version and whether summaries represent genes or transcripts.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 9 of 36
9. Build sample metadata beside counts
Match every count column to one and only one metadata row. Check condition labels, batches, subjects, time points and missing values before constructing the DESeq2 object.
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 10 of 36
10. Inspect library and detection patterns
Plot library totals, detected features and sample relationships. Extreme depth, low complexity or unexpected clustering should be explained before formal testing.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 11 of 36
11. Filter without peeking at outcomes
Remove features with almost no information using a rule chosen before inspecting group significance. Record the retained universe because multiple-testing correction depends on it.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 12 of 36
12. Check design rank and confounding
A batch perfectly aligned with condition cannot be separated statistically. Resolve such problems in experimental design or narrow the question rather than adding an impossible formula.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
Section 13 of 36
13. Estimate size factors
DESeq2 estimates relative scaling factors so samples can be compared despite different effective library depths. Examine composition extremes and consider documented alternatives when the default assumption is strained.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 14 of 36
14. Estimate gene-wise dispersion
Replicate variation is estimated for each gene, then related to an overall mean–dispersion trend. Sparse replication makes that shared structure especially important.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 15 of 36
15. Practise with a fictional DESeq2 table
This classroom table is invented and is not a result from the DESeq2 paper.
| Fictional gene | Log2 fold change | Adjusted evidence | Replicates agree | Reading |
|---|---|---|---|---|
| GENE-A | 1.4 | 0.004 | Yes | Follow up |
| GENE-B | -0.8 | 0.310 | Yes | Uncertain |
| GENE-C | 2.2 | 0.020 | No | Fragile |
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 16 of 36
16. Moderate dispersion estimates
DESeq2 pulls noisy gene-wise estimates toward a fitted trend by an amount informed by the data. Moderation improves stability but should still be checked against residual and sample-level behaviour.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 17 of 36
17. Fit the generalised linear models
For each retained gene, the expected count is linked to the design through coefficients. Save convergence messages and the exact package version used for the fit.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 18 of 36
18. Test the declared comparison
Use a named coefficient or an explicit contrast that matches the research question. State the null hypothesis, direction convention and whether a thresholded effect was tested.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
Section 19 of 36
19. Control the testing family
Adjusted values depend on the set of hypotheses tested. Define whether the family spans genes, cell types, contrasts or all of these together.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 20 of 36
20. Shrink fold changes for suitable tasks
Shrinkage can reduce unstable large estimates from low-information genes and improve ranking or display. Report the estimator and keep it conceptually separate from significance testing.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 21 of 36
21. Read MA plots carefully
An MA plot places effect estimates against mean abundance. Look for intensity-dependent patterns, outliers and unstable low-count estimates rather than treating colour alone as evidence.
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 22 of 36
22. Inspect sample-level expression
Plot normalised counts for headline genes by biological sample. Verify that direction is shared and not created by one extreme library.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 23 of 36
23. Review Cook's-distance flags
Influential observations may trigger outlier handling or missing results under documented rules. Investigate sample quality and biology before deciding whether exclusion is justified.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 24 of 36
24. Check transformed data for exploration
Variance-stabilising or regularised-log transformations support visualisation and clustering; they are not the inputs for the standard count test. Keep exploration and inference distinct.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
Section 25 of 36
25. Run leave-one-sample sensitivity
Refit important results after removing one sample at a time when the design permits. Large sign or rank changes reveal dependence on individual libraries.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 26 of 36
26. Compare defensible formulas
Test whether adding a batch term, interaction or pairing structure changes the conclusion for scientific reasons. Do not choose the formula that merely produces the shortest list.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 27 of 36
27. Compare with edgeR
Both frameworks model replicated counts with negative-binomial ideas and information sharing. Their normalisation, dispersion and testing implementations differ, so agreement should be evaluated on the same design and inputs.
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 28 of 36
28. Build a DESeq2 evidence card
Summarise input provenance, sample count, formula, size-factor method, dispersion fit, contrast, outlier rules, multiplicity, effect estimate, sensitivity and validation in one auditable record.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 29 of 36
29. Connect DESeq2 to school science
A fair comparison controls known differences while leaving the tested factor visible. That is the same logic students use when separating variables in an experiment.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 30 of 36
30. Try a library-size activity
Give teams bead bags of different total sizes but similar colour proportions. Compare raw colour counts with proportions and discuss when one colour dominating a bag complicates scaling.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
Section 31 of 36
31. Connect to mathematics
Ratios, logarithms, distributions, regression coefficients and uncertainty show how mathematics turns noisy counts into bounded comparisons.
DESeq2 represents each gene with a negative-binomial count model whose mean depends on sample-specific scaling and the design matrix.
Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 32 of 36
32. Connect to computing and AI literacy
The result depends on data structures, factor levels, software versions and defaults. A reproducible script should expose those choices instead of presenting a gene list as machine certainty.
Size factors address differences in effective sequencing depth under a composition assumption; they are not a universal correction for every technical or biological imbalance.
Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.
Section 33 of 36
33. Connect to school and career pathways
Interest in gene-expression evidence can lead toward biology, biotechnology, medicine, statistics, computer science or bioinformatics through different educational routes.
Dispersion describes extra variability beyond a Poisson model. Moderation borrows strength across genes while retaining gene-specific behaviour.
Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.
Section 34 of 36
34. Create a family evidence habit
When a headline says a gene changed, ask how many independent samples were compared, what counts meant, how large the effect was and whether sample plots agree.
A model coefficient has meaning only through factor coding and reference levels. Inspect resultsNames before deciding that a sign means increase or decrease.
Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.
Section 35 of 36
35. Use precise DESeq2 language
Say a gene showed differential count evidence under the stated model and contrast. Avoid saying it causes a trait or proves a mechanism without experiments designed for that claim.
Log2 fold-change shrinkage can stabilise noisy estimates for ranking and visualisation, but it does not manufacture replication or replace the original statistical test.
Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.
Section 36 of 36
36. Finish with a counts-to-claim bundle
Deliver counts, metadata, scripts, session information, diagnostics, coefficient definitions, complete results, plots, sensitivity analyses and a short statement of what remains uncertain.
Independent filtering and outlier handling affect the reported testing set. Record those decisions and inspect why any result received a missing adjusted value.
Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.
## Source date and scope note Primary studies and current official software documentation were checked for this article on 11 October 2026. Software interfaces and recommendations can change, so readers should consult the linked current documentation before reproducing an analysis.
## Final reader checklist Before accepting a claim, confirm the biological sample count, condition definition, measured material, preprocessing, statistical unit, design matrix, exact contrast, effect size, uncertainty, software version, negative controls, sensitivity checks, independent validation and the boundary between association and causation.
Contents · Previous section · Continue to the Science Learning Hub