eduKateSG · Why Science?
Fit overdispersed count models at modern single-cell scale—without mistaking computational speed for independent biological evidence
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 Nebula Subject Aware Count Models Scalable Single Cell Inference; Why Science Zinb Wave Zero Inflated Counts Low Dimensional Signal; Why Science Dreamlet Pseudobulk Mixed Models Complex Single Cell Cohorts; 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: glmGamPoi primary study; Official Bioconductor glmGamPoi package; Official glmGamPoi quickstart; 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.
glmGamPoi is an open-source Bioconductor package introduced by Ahlmann-Eltze and Huber for fitting Gamma–Poisson generalised linear models to count data, particularly large single-cell datasets with many small counts. Its primary paper describes efficient overdispersion estimation and support for on-disk matrices. Current package documentation also includes a pseudobulk differential-expression workflow. Efficiency expands the analyses that can be attempted; it does not turn cells from the same person or sample into independent biological replicates.
Inside this guide
1–12 · Foundations and models
- 1. Begin with many small counts
- 2. Define the Gamma–Poisson distribution
- 3. Use a log-linked mean
- 4. Understand overdispersion
- 5. Separate algorithmic scale from evidence scale
- 6. Choose cell-level or pseudobulk purpose
- 7. Read the 2020 evidence in scope
- 8. Prepare integer count inputs
- 9. Attach sample and cell metadata
- 10. Filter with a declared rule
- 11. Build the model matrix
- 12. Choose size factors or offsets
13–24 · Evidence, testing and applications
- 13. Select in-memory or on-disk storage
- 14. Fit Gamma–Poisson parameters
- 15. Practise with a fictional glmGamPoi table
- 16. Compare full and reduced models
- 17. Inspect deviance and fitted means
- 18. Interpret coefficients on the right scale
- 19. Control the testing family
- 20. Build pseudobulk profiles when appropriate
- 21. Inspect subject-level support
- 22. Audit sparse-feature behaviour
- 23. Audit size-factor sensitivity
- 24. Audit model-form sensitivity
25–36 · Learning, decisions and pathways
- 25. Audit sample influence
- 26. Benchmark without overclaiming
- 27. Compare with NEBULA
- 28. Build a glmGamPoi evidence card
- 29. Connect glmGamPoi to school science
- 30. Try an overdispersion 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 glmGamPoi language
- 36. Finish with a scale-to-claim bundle
Section 1 of 36
1. Begin with many small counts
Single-cell matrices contain many zeros and small integers across thousands of cells. The scientific challenge is to model this variation efficiently without forgetting the experiment’s sample hierarchy.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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. Define the Gamma–Poisson distribution
A Poisson rate that varies according to a gamma distribution produces an overdispersed count model equivalent to a negative-binomial form. Its parameters describe mean and extra variation.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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. Use a log-linked mean
Predictors combine through a design matrix on the log scale. Offsets can represent differing exposure or library size while coefficients encode comparisons.
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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. Understand overdispersion
When dispersion is near zero, variation approaches Poisson behaviour; larger dispersion permits wider counts. Estimate quality matters especially for sparse genes.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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. Separate algorithmic scale from evidence scale
Millions of cell rows can create a large computation, but biological inference may still rest on only a handful of donors or samples.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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 cell-level or pseudobulk purpose
Cell-level modelling can support representation or conditional comparisons, while multi-sample population claims often need sample-aware aggregation or hierarchical structure.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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 2020 evidence in scope
Ahlmann-Eltze and Huber described faster and more accurate Gamma–Poisson fitting for modern single-cell count matrices, including on-disk computation. Benchmarks do not erase downstream design requirements.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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. Prepare integer count inputs
Confirm that rows and columns match the intended genes and cells or pseudobulk samples. Preserve the assay layer containing counts rather than transformed values.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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. Attach sample and cell metadata
Map cells to donor, sample, condition, batch and cell type. Missing or duplicated identifiers can silently change the unit of analysis.
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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. Filter with a declared rule
Remove features or groups lacking enough information for the intended fit. Filtering should reduce unstable computation without being tuned to attractive outcomes.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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. Build the model matrix
Encode predictors explicitly and check rank, reference levels and interactions. A large matrix does not solve confounding among columns.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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. Choose size factors or offsets
Scaling accounts for different total counts or exposures. Inspect extreme factors and verify that the normalisation assumption matches the analysis level.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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. Select in-memory or on-disk storage
On-disk arrays reduce memory pressure at the cost of input/output patterns. Record chunking, storage format and hardware because reproducibility includes computational context.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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. Fit Gamma–Poisson parameters
Estimate coefficients and dispersion for each feature. Save convergence, timing and any genes for which estimates hit boundaries or fail.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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 glmGamPoi table
This classroom table is invented and is not a result from the glmGamPoi paper.
| Fictional gene | Dispersion | Log fold change | Samples agree | Reading |
|---|---|---|---|---|
| GENE-G | 0.08 | 1.0 | Yes | Stable candidate |
| GENE-H | 1.20 | 1.8 | No | Highly variable |
| GENE-I | 0.15 | 0.1 | Yes | Little difference |
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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. Compare full and reduced models
A likelihood-ratio test compares nested models that differ by the tested terms. Write both formulas and confirm that their difference is the intended hypothesis.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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. Inspect deviance and fitted means
Large residual deviance or systematic fitted-versus-observed patterns can signal model mismatch, unmodelled groups or problematic features.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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. Interpret coefficients on the right scale
Exponentiated coefficients relate to multiplicative changes in the expected count under the design. State reference levels and offsets before giving a direction.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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
Adjust across features, cell types and contrasts according to the planned family. Running millions of fast tests does not reduce multiplicity.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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. Build pseudobulk profiles when appropriate
Aggregate raw counts within each biological sample and cell population before sample-level differential testing. Retain cell counts as quality information rather than replication.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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. Inspect subject-level support
Plot pseudobulk counts or fitted summaries for every subject. A result driven by one cell-rich donor requires a narrow conclusion.
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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. Audit sparse-feature behaviour
Very rare genes can produce unstable coefficients or boundary dispersions. Examine detection across samples and repeat after reasonable filters.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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. Audit size-factor sensitivity
Use documented alternative scaling when a few features dominate counts. Differences reveal dependence on composition assumptions.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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. Audit model-form sensitivity
Compare Gamma–Poisson fits with a defensible alternative for key genes. Agreement supports robustness; disagreement identifies assumptions to discuss.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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. Audit sample influence
Leave out one donor or pseudobulk sample at a time for headline results. More cells within the remaining samples cannot hide a fragile donor effect.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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. Benchmark without overclaiming
Runtime and memory benchmarks depend on matrix shape, sparsity, storage and hardware. Report the tested environment rather than presenting one speed ratio as universal.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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 NEBULA
NEBULA adds explicit subject-aware random effects for cell-level count inference; glmGamPoi focuses on efficient Gamma–Poisson fitting and can support sample-aware pseudobulk workflows.
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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 glmGamPoi evidence card
Record hierarchy, counts, filters, storage, scaling, design, full and reduced models, dispersion, diagnostics, runtime, multiplicity, sensitivity and validation.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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 glmGamPoi to school science
A model can be both scientifically principled and computationally efficient. Good science checks whether the faster calculation preserves the question and evidence structure.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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 an overdispersion activity
Generate repeated count sets with the same mean but different spreads. Compare them with a Poisson expectation and discuss what extra variation might represent.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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
Probability mixtures, logarithms, matrices, likelihood and dispersion show how mathematics links mechanisms of variability to observable counts.
The Gamma–Poisson distribution is another parameterisation of the negative binomial and represents counts whose variance exceeds their mean.
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
Sparse matrices, on-disk arrays and efficient algorithms determine what can be computed. Scale is a systems property, not a guarantee of truth.
glmGamPoi uses efficient algorithms for many repeated count-model fits and can work with on-disk representations when matrices exceed memory.
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
Large biological datasets connect life science with software engineering, statistics, medicine, biotechnology and high-performance computing.
A log link connects the expected count to predictors and offsets. Coefficients remain conditional on the chosen design and reference levels.
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 study celebrates millions of cells, ask how many independent samples were measured, how the hierarchy was modelled and whether donor-level plots agree.
Likelihood-ratio comparisons ask whether a larger model explains evidence beyond a nested reduced model; the two formulas must encode the intended question.
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 glmGamPoi language
Say the Gamma–Poisson model supports a count difference under the stated design. Avoid turning model fit into a causal pathway or clinical prediction.
Many zeros and small counts are common in single-cell data, but a Gamma–Poisson model does not label each zero as biological or technical.
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 scale-to-claim bundle
Deliver matrices, metadata, hierarchy, code, environment, storage details, formulas, fits, diagnostics, full tests, donor plots, sensitivity and bounded conclusions.
For multi-sample single-cell inference, pseudobulk aggregation can preserve samples as replicates; direct cell-level fits require a hierarchy that matches the claim.
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