eduKateSG · Why Science?
Keep the subject hierarchy inside a scalable count model—so a sea of cells does not masquerade as a sea of independent people
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 Mast Hurdle Models Single Cell Differential Expression; Why Science Scdd Distributional Changes Single Cell Expression; 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: NEBULA primary study; NEBULA PubMed record; NEBULA open full text; Official nebula package guide; 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.
NEBULA—NEgative Binomial mixed model Using a Large-sample Approximation—was presented by He and colleagues in Communications Biology in 2021. It fits subject-aware count models for large multi-subject single-cell datasets, separating between-subject and within-subject overdispersion while accelerating likelihood calculations through analytical approximations. It can support differential-expression or co-expression analysis, but subject count, predictor structure, convergence and model fit remain more important than the number of captured cells.
Inside this guide
1–12 · Foundations and models
- 1. Begin with nested counts
- 2. Define between-subject variation
- 3. Define within-subject variation
- 4. Choose differential or co-expression
- 5. Protect subject count
- 6. Write a predictor map
- 7. Read the 2021 evidence in scope
- 8. Prepare raw counts
- 9. Map every cell to a subject
- 10. Create the predictor matrix
- 11. Choose offsets thoughtfully
- 12. Filter for information
13–24 · Evidence, testing and applications
- 13. Choose NBGMM or PGMM
- 14. Set parameter bounds visibly
- 15. Practise with a fictional NEBULA table
- 16. Understand the approximation
- 17. Fit gene-wise models
- 18. Read fixed-effect coefficients
- 19. Read dispersion estimates
- 20. Control multiple testing
- 21. Inspect subject-level support
- 22. Check convergence and boundaries
- 23. Compare model forms
- 24. Run subject leave-one-out checks
25–36 · Learning, decisions and pathways
- 25. Run cell-subsampling checks
- 26. Compare with MAST
- 27. Compare with dreamlet
- 28. Build a mixed-model evidence card
- 29. Connect NEBULA to school science
- 30. Try a nested-measurement 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 NEBULA language
- 36. Finish with a hierarchy-to-claim bundle
Section 1 of 36
1. Begin with nested counts
A multi-subject experiment observes genes in cells, and cells belong to subjects. That membership creates correlation that a flat analysis can mistake for replication.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 2 of 36
2. Define between-subject variation
People or specimens differ in baseline expression for biological and technical reasons. A subject-level random effect represents that shared influence.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 3 of 36
3. Define within-subject variation
Cells from the same subject still vary beyond their mean. Cell-level overdispersion captures heterogeneity and measurement variability at the cellular scale.
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 4 of 36
4. Choose differential or co-expression
Differential expression links a predictor to one gene’s counts; co-expression examines relationships between genes. These are different questions with different interpretations.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 5 of 36
5. Protect subject count
A thousand cells from each of two subjects do not support the same population claim as cells from fifty subjects. State both counts prominently.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 6 of 36
6. Write a predictor map
Label every predictor as subject-level, sample-level or cell-level and note interactions. This exposes impossible comparisons and duplicated information.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
Section 7 of 36
7. Read the 2021 evidence in scope
He and colleagues introduced NEBULA as a fast negative-binomial mixed-model method for differential or co-expression analysis in large multi-subject single-cell datasets.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 8 of 36
8. Prepare raw counts
Provide a genes-by-cells matrix with documented quantification and filtering. Do not silently substitute log-normalised values for the count likelihood.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 9 of 36
9. Map every cell to a subject
The subject identifier vector must align exactly with count-matrix columns. Validate lengths, ordering and missing IDs before modelling.
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 10 of 36
10. Create the predictor matrix
Encode the intercept, condition and covariates explicitly. Inspect rank, reference levels and within-subject versus between-subject variation.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 11 of 36
11. Choose offsets thoughtfully
Offsets can represent library size or exposure. Plot them by subject and condition and record the calculation.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 12 of 36
12. Filter for information
Genes need enough non-zero counts and cells per group for stable estimation. Apply an outcome-blind rule before reviewing associations.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
Section 13 of 36
13. Choose NBGMM or PGMM
The negative-binomial gamma mixed model includes cell-level overdispersion; the Poisson-gamma alternative omits it. Let fit and scientific context guide the choice.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 14 of 36
14. Set parameter bounds visibly
Optimisation bounds can affect dispersion estimates, especially near edges. Report bounds and flag genes whose estimates press against them.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 15 of 36
15. Practise with a fictional NEBULA table
This classroom table is invented and is not a result from the NEBULA paper.
| Fictional gene | Subjects | Cells | Subject dispersion | Cell dispersion | Reading |
|---|---|---|---|---|---|
| GENE-N | 40 | 80,000 | 0.20 | 0.70 | Both levels vary |
| GENE-P | 40 | 75,000 | 0.01 | 0.55 | Check boundary |
| GENE-Q | 8 | 30,000 | 0.48 | 0.32 | Few subjects |
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 16 of 36
16. Understand the approximation
NEBULA gains speed by analytically approximating high-dimensional integration rather than relying on slower generic procedures. Speed claims belong to tested settings, not every future dataset.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 17 of 36
17. Fit gene-wise models
Each gene receives parameter estimates, evidence values and diagnostics. Track failures and convergence instead of dropping them silently.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 18 of 36
18. Read fixed-effect coefficients
The coefficient describes a conditional association on the model’s link scale. Convert it carefully when discussing fold changes or expected counts.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
Section 19 of 36
19. Read dispersion estimates
Subject and cell components describe modelled heterogeneity. Near-zero or extreme values may indicate a simpler model, poor information or mismatch.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 20 of 36
20. Control multiple testing
Correct across the declared gene or gene-pair family. Co-expression screens can create especially large testing burdens.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 21 of 36
21. Inspect subject-level support
Plot observed and fitted summaries per subject. A cell-rich subject should not be allowed to dominate the population story unnoticed.
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 22 of 36
22. Check convergence and boundaries
Review optimiser messages, finite estimates and parameter bounds. Refit or simplify questionable genes rather than ranking unstable results.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 23 of 36
23. Compare model forms
Contrast NBGMM with PGMM or a sample-level alternative for key genes. Disagreement is evidence about assumptions, not an inconvenience to hide.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 24 of 36
24. Run subject leave-one-out checks
Refit headline signals while removing one subject at a time when feasible. Large swings reveal fragile generalisation.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
Section 25 of 36
25. Run cell-subsampling checks
Balance cell counts across subjects and repeat selected models. This tests whether unequal cellular depth drives the finding.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 26 of 36
26. Compare with MAST
MAST uses a hurdle model for detection and positive expression and can include random effects. NEBULA models raw counts with subject- and cell-level overdispersion.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 27 of 36
27. Compare with dreamlet
dreamlet aggregates counts into sample–cluster pseudobulk profiles. NEBULA keeps cells in a hierarchical count model while accounting for subjects.
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 28 of 36
28. Build a mixed-model evidence card
Record subjects, cells, filters, predictors, offsets, model form, bounds, convergence, dispersions, effect, multiplicity and sensitivity.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 29 of 36
29. Connect NEBULA to school science
Several readings from one student describe that student well, while several students support a class-level conclusion. Nested models formalise this distinction.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 30 of 36
30. Try a nested-measurement activity
Measure three leaves on each of several plants. Compare treating leaves as independent with modelling plant identity and discuss how confidence changes.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
Section 31 of 36
31. Connect to mathematics
Count distributions, logarithms, variance components, matrices and approximation show how mathematical structure makes large analyses possible.
NEBULA models raw single-cell counts while using subject identifiers to represent the hierarchy of cells nested within people or other biological units.
Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.
Section 32 of 36
32. Connect to computing and AI literacy
Computational efficiency expands what can be fitted, but a fast answer is trustworthy only when inputs, hierarchy and diagnostics are correct.
Its negative-binomial gamma mixed model separates cell-level overdispersion from subject-level variation; the Poisson-gamma option is a simpler special case.
Keep the biological sample as the unit of replication. More cells can improve cellular description, but they do not create more independent people, animals or cultures.
Section 33 of 36
33. Connect to school and career pathways
Interest in nested data can lead toward biology, statistics, medicine, public health, computing or quantitative genetics.
Large-sample approximations and analytical treatment of integrals speed computation, but approximations and boundary parameters still require diagnostics.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state which false signal each check can reveal.
Section 34 of 36
34. Create a family evidence habit
When a study reports hundreds of thousands of cells, ask how many subjects, how balanced their cell counts were and whether one subject drove the result.
Predictors operate at cell or subject level according to the design. Their coding and variation determine which association the coefficient can identify.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.
Section 35 of 36
35. Use precise NEBULA language
Say the predictor is associated with modelled counts or co-expression under a subject-aware mixed model. Avoid causal regulation or clinical prediction.
Co-expression analysis asks whether genes covary after accounting for the model; it does not by itself establish a regulatory interaction or direct molecular contact.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 36 of 36
36. Finish with a hierarchy-to-claim bundle
Deliver counts, subject mapping, predictors, offsets, settings, convergence records, dispersion estimates, subject plots, sensitivity checks and validation.
Many cells per subject can strengthen within-subject estimation, while the number and balance of subjects still govern population-level generalisation.
Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.
## Source date and scope note Primary papers, open records and 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 modality, preprocessing, statistical unit, model formula, uncertainty, software version, negative controls, sensitivity checks, independent validation and the exact boundary between association and causation.
Contents · Previous section · Continue to the Science Learning Hub