Chapter map

1. 1. Why integration needs controls
2. 2. Single-cell stable genes
3. 3. Negative-control logic
4. 4. Clustering each dataset
5. 5. Mutual nearest clusters
6. 6. Pseudoreplicate construction
7. 7. Factor model for unwanted variation
8. 8. fastRUVIII adjustment
9. 9. Choosing factor number
10. 10. A small matrix example
11. 11. Control contamination
12. 12. Pseudoreplicate mismatch
13. 13. Preserving trajectories and rare states
14. 14. Held-out validation
15. 15. A student control-gene project
16. 16. Governance and careers

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Why integration needs controls

This step deserves a careful pause. Removing unwanted variation requires information about what should remain unchanged or comparable.

A concrete situation

Stable genes can anchor nuisance estimation.

What the mathematics does not license

A wrong control set removes biology. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Declare nuisance and protected signals. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Single-cell stable genes

The useful question is not only how to calculate, but what the calculation assumes. scSEGs are identified from diverse reference data as genes with relatively stable expression across cell types.

A concrete situation

Stable genes act like negative controls.

What the mathematics does not license

Stability may not transfer to every tissue or perturbation. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Check controls in the target design. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. Negative-control logic

A small example makes the mechanism visible. A true negative-control gene should carry nuisance variation but little signal of interest.

A concrete situation

Its variation helps estimate technical factors.

What the mathematics does not license

Control contamination biases the factor space. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Test association with condition and cell type. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Clustering each dataset

This is where modelling choices begin to affect evidence. Cells are grouped within datasets before cross-dataset matching.

A concrete situation

Clusters summarise candidate cell populations.

What the mathematics does not license

Different resolution can split or merge biology. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Repeat plausible clustering settings. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Mutual nearest clusters

The central mathematical idea is concrete. Comparable clusters across datasets are identified using reciprocal similarity.

A concrete situation

A T-cell cluster in study A matches one in B.

What the mathematics does not license

Reciprocity can fail with missing or shifted populations. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record distances and marker agreement. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Pseudoreplicate construction

Here the geometry becomes operational. Cells sampled from matched clusters form pseudoreplicate groups for estimating unwanted effects.

A concrete situation

Comparable cell groups substitute for missing direct replicates in factor identification.

What the mathematics does not license

They are not independent biological samples. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Store membership and match quality. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Factor model for unwanted variation

This step deserves a careful pause. Expression is represented by biological design plus latent unwanted factors and residual noise.

A concrete situation

Technical variation can be low rank.

What the mathematics does not license

Biology can also be low rank and confounded. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write matrix dimensions and identifiability inputs. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. fastRUVIII adjustment

The useful question is not only how to calculate, but what the calculation assumes. Control genes and pseudoreplicate constraints help estimate unwanted factors, which are then removed under the model.

A concrete situation

Shared technical shifts can be subtracted.

What the mathematics does not license

Too many factors can remove real structure. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot preservation versus mixing across k. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Choosing factor number

A small example makes the mechanism visible. The number of unwanted factors controls correction flexibility.

A concrete situation

One factor may capture library preparation; several may capture platforms.

What the mathematics does not license

Training fit favours complexity. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use held-out donors, controls and biology metrics. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. A small matrix example

This is where modelling choices begin to affect evidence. Let Y equal X beta plus W alpha plus residual, where W holds two unwanted factors.

A concrete situation

Control genes help estimate W without target biology.

What the mathematics does not license

Rotations of W can represent the same subspace. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Calculate a toy projection residual. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Control contamination

The central mathematical idea is concrete. If a stable gene responds to disease in the target study, correction can absorb disease signal.

A concrete situation

Housekeeping is not universal invariance.

What the mathematics does not license

A long control list can hide a few harmful genes. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Run leave-group-out control sensitivity. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Pseudoreplicate mismatch

Here the geometry becomes operational. Wrongly matched clusters make true differences look unwanted.

A concrete situation

A developmental state may resemble but not equal an adult state.

What the mathematics does not license

Marker review and distance thresholds are essential. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Add a rejection route for unsupported matches. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Preserving trajectories and rare states

This step deserves a careful pause. Correction should keep developmental order and populations unique to one study.

A concrete situation

A liver trajectory can be clearer after nuisance removal.

What the mathematics does not license

Overcorrection can flatten branches. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare raw markers and pseudotime. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Held-out validation

The useful question is not only how to calculate, but what the calculation assumes. Reserve donors or studies and evaluate integration without tuning on final labels.

A concrete situation

A new dataset tests transfer of controls and factors.

What the mathematics does not license

Random cells leak structure. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Lock controls and factor number first. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. A student control-gene project

A small example makes the mechanism visible. Simulate biology and one technical factor, add clean and contaminated control genes, then estimate the nuisance direction.

A concrete situation

Known truth shows bias from bad controls.

What the mathematics does not license

Real data add sparse counts and matching. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Vary factor number and match error. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Governance and careers

This is where modelling choices begin to affect evidence. Version controls, clustering, matches, pseudoreplicates, factors and adjusted outputs. Control-based integration joins statistics, biology and audit.

A concrete situation

Students can begin with regression residuals.

What the mathematics does not license

Correction proves no mechanism or biological replication. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write a control and pseudoreplicate ledger. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

A scMerge audit begins with several single-cell datasets, a defensible list of single-cell stably expressed genes and a plan for identifying comparable cell groups across datasets. Stable genes act as negative controls for estimating unwanted factors; pseudoreplicates help identify which factors represent nuisance variation.

Record the reference used to identify stable genes, clustering, mutual nearest clusters, pseudoreplicate construction, number of unwanted factors, fastRUVIII settings, scaling and saved correction matrices. Stable does not mean invariant in every tissue, and pseudoreplicates are algorithmic constructs rather than new biological samples.

Validate with held-out donors and datasets, negative-control stability, pseudoreplicate purity, marker and trajectory preservation, rare-state review and sensitivity to factor number. Use corrected values for their declared purpose while preserving raw counts and biological replication for inference.

---

Frequently asked questions

What are single-cell stably expressed genes?

They are genes selected for stable expression across diverse cell types and used as negative controls for unwanted-factor estimation.

What is a pseudoreplicate?

It is an algorithmically constructed set of comparable cells or summaries across datasets, used to learn unwanted variation.

Are pseudoreplicates biological replicates?

No. They help identify technical factors but do not replace independent donors or experiments.

What is fastRUVIII?

It is the unwanted-variation adjustment framework used by scMerge with control genes and replicate information.

Can stable genes contain biology?

Yes. Stability is context-dependent, so the control set must be audited in the target tissues and conditions.

How should scMerge be validated?

Test control stability, pseudoreplicate purity, factor-number sensitivity, held-out donors, marker and trajectory preservation, rare states and raw-data inference.

---

Primary source and fact-check note

Yingxin Lin, Shila Ghazanfar, Kevin Y. X. Wang et al., “scMerge leverages factor analysis, stable expression, and pseudoreplication to merge multiple single-cell RNA-seq datasets,” Proceedings of the National Academy of Sciences 116(20), 9775–9784 (2019), DOI 10.1073/pnas.1820006116. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• RUV-2 control-gene factor removal
• Mutual nearest neighbours
• BBKNN graph integration

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: topic-specific validation and responsible use

Stable-gene provenance

Record the study, tissues, species and criteria used to identify stably expressed genes. Test whether those genes respond to the biology under study. A gene stable in one reference can vary in another tissue or disease. Remove contaminated controls, repeat the analysis and report how estimated unwanted factors and preserved markers change.

Pseudoreplicate recipe

Document every clustering, matching and mutual-nearest-cluster rule used to build pseudoreplicates. Store source cells, datasets, weights and match scores. Pseudoreplicates help expose shared nuisance factors, but they are computational aggregates rather than new donors. Their number must never inflate the biological sample size in inference.

Purity challenge

Use held-out labels and markers only for evaluation to ask whether matched groups represent comparable biology. Measure within-pseudoreplicate heterogeneity and flag combinations joining distinct states. Repeat correction after removing questionable matches. If conclusions change sharply, narrow the integration claim and improve matching rather than accepting a visually smooth embedding.

Factor-number sensitivity

Fit a justified range of unwanted-factor counts. For each value, report negative-control stability, batch attenuation, marker preservation, trajectory structure and held-out retrieval. Too few factors leave nuisance variation; too many can remove biology. Select the setting inside training data and lock it before examining the final study.

Control-and-match interaction

Bad controls and bad pseudoreplicates can reinforce each other. A control gene affected by biology may direct factor estimation toward a real pathway, while an impure match may make that direction look shared. Cross the two sensitivities: vary the control set and the match set independently, then identify conclusions that survive both changes.

Corrected-value boundary

State which tasks use adjusted values: visualisation, clustering, neighbour construction or another declared purpose. Preserve raw counts and donor structure for biological inference unless a validated model supports a different analysis. Correction produces a derived data product whose uncertainty and provenance belong beside every downstream result.

Rare-state guard

Identify populations present in only one dataset or at low frequency. Inspect whether matching forces them into common states and whether factor removal erases their markers. Separation can be scientifically correct. A robust workflow aligns supported overlap while keeping unique biology visible and labelled as unsupported by cross-dataset matches.

Reproducibility package

Archive control genes, pseudoreplicate matrices, factor estimates, loadings, correction matrices, software versions and random seeds. Include a small worked example that reconstructs one adjusted value from recorded components. This turns an attractive integrated plot into an auditable statistical analysis and makes later atlas updates comparable.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Why integration needs controls

Start with the mechanism: Removing unwanted variation requires information about what should remain unchanged or comparable. The worked case is deliberately concrete: Stable genes can anchor nuisance estimation. Now expose the boundary: A wrong control set removes biology. A student should complete this evidence step: Declare nuisance and protected signals. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 2: Single-cell stable genes

Start with the mechanism: scSEGs are identified from diverse reference data as genes with relatively stable expression across cell types. The worked case is deliberately concrete: Stable genes act like negative controls. Now expose the boundary: Stability may not transfer to every tissue or perturbation. A student should complete this evidence step: Check controls in the target design. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 3: Negative-control logic

Start with the mechanism: A true negative-control gene should carry nuisance variation but little signal of interest. The worked case is deliberately concrete: Its variation helps estimate technical factors. Now expose the boundary: Control contamination biases the factor space. A student should complete this evidence step: Test association with condition and cell type. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 4: Clustering each dataset

Start with the mechanism: Cells are grouped within datasets before cross-dataset matching. The worked case is deliberately concrete: Clusters summarise candidate cell populations. Now expose the boundary: Different resolution can split or merge biology. A student should complete this evidence step: Repeat plausible clustering settings. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 5: Mutual nearest clusters

Start with the mechanism: Comparable clusters across datasets are identified using reciprocal similarity. The worked case is deliberately concrete: A T-cell cluster in study A matches one in B. Now expose the boundary: Reciprocity can fail with missing or shifted populations. A student should complete this evidence step: Record distances and marker agreement. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 6: Pseudoreplicate construction

Start with the mechanism: Cells sampled from matched clusters form pseudoreplicate groups for estimating unwanted effects. The worked case is deliberately concrete: Comparable cell groups substitute for missing direct replicates in factor identification. Now expose the boundary: They are not independent biological samples. A student should complete this evidence step: Store membership and match quality. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 7: Factor model for unwanted variation

Start with the mechanism: Expression is represented by biological design plus latent unwanted factors and residual noise. The worked case is deliberately concrete: Technical variation can be low rank. Now expose the boundary: Biology can also be low rank and confounded. A student should complete this evidence step: Write matrix dimensions and identifiability inputs. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 8: fastRUVIII adjustment

Start with the mechanism: Control genes and pseudoreplicate constraints help estimate unwanted factors, which are then removed under the model. The worked case is deliberately concrete: Shared technical shifts can be subtracted. Now expose the boundary: Too many factors can remove real structure. A student should complete this evidence step: Plot preservation versus mixing across k. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 9: Choosing factor number

Start with the mechanism: The number of unwanted factors controls correction flexibility. The worked case is deliberately concrete: One factor may capture library preparation; several may capture platforms. Now expose the boundary: Training fit favours complexity. A student should complete this evidence step: Use held-out donors, controls and biology metrics. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 10: A small matrix example

Start with the mechanism: Let Y equal X beta plus W alpha plus residual, where W holds two unwanted factors. The worked case is deliberately concrete: Control genes help estimate W without target biology. Now expose the boundary: Rotations of W can represent the same subspace. A student should complete this evidence step: Calculate a toy projection residual. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 11: Control contamination

Start with the mechanism: If a stable gene responds to disease in the target study, correction can absorb disease signal. The worked case is deliberately concrete: Housekeeping is not universal invariance. Now expose the boundary: A long control list can hide a few harmful genes. A student should complete this evidence step: Run leave-group-out control sensitivity. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 12: Pseudoreplicate mismatch

Start with the mechanism: Wrongly matched clusters make true differences look unwanted. The worked case is deliberately concrete: A developmental state may resemble but not equal an adult state. Now expose the boundary: Marker review and distance thresholds are essential. A student should complete this evidence step: Add a rejection route for unsupported matches. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 13: Preserving trajectories and rare states

Start with the mechanism: Correction should keep developmental order and populations unique to one study. The worked case is deliberately concrete: A liver trajectory can be clearer after nuisance removal. Now expose the boundary: Overcorrection can flatten branches. A student should complete this evidence step: Compare raw markers and pseudotime. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 14: Held-out validation

Start with the mechanism: Reserve donors or studies and evaluate integration without tuning on final labels. The worked case is deliberately concrete: A new dataset tests transfer of controls and factors. Now expose the boundary: Random cells leak structure. A student should complete this evidence step: Lock controls and factor number first. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 15: A student control-gene project

Start with the mechanism: Simulate biology and one technical factor, add clean and contaminated control genes, then estimate the nuisance direction. The worked case is deliberately concrete: Known truth shows bias from bad controls. Now expose the boundary: Real data add sparse counts and matching. A student should complete this evidence step: Vary factor number and match error. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.

Workshop 16: Governance and careers

Start with the mechanism: Version controls, clustering, matches, pseudoreplicates, factors and adjusted outputs. Control-based integration joins statistics, biology and audit. The worked case is deliberately concrete: Students can begin with regression residuals. Now expose the boundary: Correction proves no mechanism or biological replication. A student should complete this evidence step: Write a control and pseudoreplicate ledger. Record the inputs, mathematical operation, expected output, uncertainty and one result that would contradict the claim. Then repeat the calculation or model under a nearby defensible setting and compare what changes. This turns a polished output into an auditable argument. Use untouched donors, studies or simulated truth where possible, because random-cell checks can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision as separate lines in the notebook.