Why Science? | dreamlet, Pseudobulk Mixed Models and Complex Single-Cell Cohorts

Section 1 of 36

1. Begin with a cohort question

Large studies may profile many cell types across hundreds of people, visits or conditions. The useful question names a trait, cell population and biological sampling structure.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Next section

Section 2 of 36

2. Recognise two scales

Cells reveal within-sample detail, while subjects or biosamples support population inference. The workflow must respect both levels without treating cells as independent people.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 3 of 36

3. Define pseudobulk

For each sample and cell cluster, raw counts are summed gene by gene. The resulting profiles resemble bulk RNA-seq observations tied to real experimental units.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 4 of 36

4. Define precision

Pseudobulk profiles based on few cells or low counts may be noisier. Precision weights express that unequal reliability instead of pretending every observation is equally informative.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 5 of 36

5. Define repeated measures

Several tissues, visits or conditions from one person are correlated. A random effect can model that shared subject structure when replication supports estimation.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 6 of 36

6. Write the contrast first

State the coefficient or comparison in words before writing a formula. This prevents a technically valid model from answering the wrong biological question.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.

Contents · Previous section · Next section

Section 7 of 36

7. Read the current evidence in scope

The peer-reviewed dreamlet study and official documentation describe scalable pseudobulk differential-expression workflows using precision-weighted linear mixed models for complex multi-sample designs.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 8 of 36

8. Build the sample manifest

Create one row per biosample with subject, visit, tissue, condition, batch and planned covariates. Verify unique identifiers and legal combinations.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 9 of 36

9. Validate cell annotations

Cell-cluster labels should have marker support and consistent meaning across samples. A mislabeled cluster corrupts every pseudobulk profile built from it.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 10 of 36

10. Aggregate from raw counts

Sum integer-like counts within each sample–cluster combination. Record whether cells were excluded and ensure aggregation does not mix subjects or conditions.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 11 of 36

11. Inspect cluster coverage

Tabulate cells and total counts per sample and cluster. Sparse combinations may require filtering or a narrower claim rather than imputation.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 12 of 36

12. Normalise each assay appropriately

Library sizes and composition vary across pseudobulk profiles. Use documented normalisation and inspect whether extreme genes dominate scaling.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.

Contents · Previous section · Next section

Section 13 of 36

13. Estimate precision weights

Weights combine mean–variance trends and observation quality for the fitted linear-model framework. Plot them and investigate samples with unexpectedly low precision.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 14 of 36

14. Build fixed effects

Condition, batch, age or other justified variables may enter the model. Every term consumes information and needs a clear scientific or design rationale.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with a fictional dreamlet table

This classroom table is invented and is not a result from the dreamlet paper.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 16 of 36

16. Add random effects carefully

Use subject or other grouping effects to represent correlation. Too few groups or redundant effects can produce singular fits and unstable variance estimates.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 17 of 36

17. Fit genes within cell types

dreamlet fits a regression for each gene and cluster assay. Preserve identical contrast definitions while allowing assay-specific missing samples and precision.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 18 of 36

18. Use empirical-Bayes moderation

Borrowing information across genes can stabilise variance estimates. Report the framework and do not confuse moderated statistics with biological replication.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.

Contents · Previous section · Next section

Section 19 of 36

19. Test the declared contrast

A coefficient may represent a group difference, continuous trait or interaction. Interpret its sign and scale in the exact coding used.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 20 of 36

20. Correct for multiplicity

Define whether false-discovery control applies within cell types, across all gene–cell-type tests or another family. Report that family visibly.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 21 of 36

21. Plot sample-level values

Show pseudobulk expression, fitted effects and residuals for each subject. A strong statistic may still be driven by a small subgroup.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 22 of 36

22. Inspect variance components

Random-effect estimates reveal where variation lies, but boundary values and wide uncertainty need caution. Do not force a narrative from every component.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 23 of 36

23. Check residual behaviour

Review mean–variance fit, extreme residuals, leverage and heteroscedasticity. Weighted models still require diagnostics.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 24 of 36

24. Test formula sensitivity

Compare a predeclared main model with plausible reduced or expanded alternatives. Explain which results change and why.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.

Contents · Previous section · Next section

Section 25 of 36

25. Audit missing assays

A cell type absent from some samples creates an unbalanced analysis set. Verify that missingness does not track the condition being tested.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 26 of 36

26. Compare with muscat

muscat provides multi-sample differential-state workflows and pseudobulk or mixed-model engines. dreamlet emphasises scalable precision-weighted linear mixed models for complex cohort designs.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 27 of 36

27. Compare with NEBULA

NEBULA fits cell-level negative-binomial mixed models with subject effects. dreamlet aggregates to pseudobulk and fits weighted linear mixed models across samples.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 28 of 36

28. Build a cohort evidence card

Record cluster, subjects, visits, aggregation, normalisation, weights, formula, contrast, diagnostics, adjusted evidence, effect and validation.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 29 of 36

29. Connect dreamlet to school science

Repeated measurements from the same student are not the same as measurements from new students. Grouped designs make that familiar distinction explicit.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 30 of 36

30. Try a repeated-measures activity

Measure pulse before and after exercise for several students. Compare an analysis that ignores pairing with one that keeps each student’s two readings connected.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

Write association, representation or distributional difference when that is what was tested. Reserve cause, diagnosis and benefit for designs that support stronger claims.

Contents · Previous section · Next section

Section 31 of 36

31. Connect to mathematics

Weighted regression, variance, contrasts, matrices and random effects show how algebra represents both comparison and correlation.

dreamlet aggregates raw counts within each sample and cell cluster, keeping individuals or biosamples as the independent units for the main comparison.

Archive the sample manifest, count matrix provenance, formula, contrasts, filters, software versions, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 32 of 36

32. Connect to computing and AI literacy

Fast parallel software can analyse thousands of genes, but only a clear design explains what each coefficient means.

Precision weights let more reliable gene–sample observations contribute appropriately while representing uncertainty from variable depth and cell yield.

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.

Contents · Previous section · Next section

Section 33 of 36

33. Connect to school and career pathways

Curiosity about cohorts can lead toward biology, statistics, public health, medicine, computing or bioinformatics through varied educational routes.

Linear mixed models can represent repeated measures or nested structure through random effects, provided the design contains enough independent levels to estimate them.

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.

Contents · Previous section · Next section

Section 34 of 36

34. Create a family evidence habit

When a result cites millions of cells, ask how many people contributed them, whether visits were repeated and whether most subjects support the effect.

Each cell type forms an assay with its own missingness, depth and variance; one global quality judgement cannot replace cell-type-specific diagnostics.

Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful embedding is not a complete result.

Contents · Previous section · Next section

Section 35 of 36

35. Use precise dreamlet language

Say a cell-type-specific pseudobulk expression association under the fitted mixed model. Avoid turning it into individual prognosis or causation.

A complex formula is useful only if its terms are identifiable. Confounded batches, tiny groups and singular random effects must be exposed rather than hidden.

Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a cohort-to-claim bundle

Deliver manifests, cluster checks, pseudobulk matrices, weights, formulae, contrasts, diagnostics, sample plots, sensitivity analyses and independent validation.

Parallel and on-disk computation make large studies practical, but computational scale does not enlarge the biological sample size.

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