Chapter map

1. 1. Why references need a mathematical map
2. 2. The shared-feature contract
3. 3. Canonical correlation analysis
4. 4. Multiple canonical dimensions
5. 5. Normalising cell embeddings
6. 6. Mutual nearest neighbours as anchors
7. 7. Anchor filtering
8. 8. Anchor scoring
9. 9. Integration vectors
10. 10. Reference choice and guide trees
11. 11. Reference mapping without rebuilding
12. 12. Label transfer mathematics
13. 13. Preserving condition-specific states
14. 14. Donor-aware validation
15. 15. A student anchor 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 references need a mathematical map

Here the geometry becomes operational. Single-cell datasets can differ by donor, technology and condition. A reference is useful only when query cells can be mapped through a reproducible relation.

A concrete situation

A blood-cell reference may help annotate a new hospital cohort.

What the mathematics does not license

Visual proximity after integration is not proof of identical state. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Write the reference, query, donor and condition roles before modelling. 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. The shared-feature contract

This step deserves a careful pause. Datasets need a defensible common feature space or a validated mapping. Feature selection shapes every later correlation and neighbour.

A concrete situation

Common variable genes can emphasise cell identity across studies.

What the mathematics does not license

Selecting features with final labels can leak the evaluation. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Trace ten genes through mapping, filtering and scaling. 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. Canonical correlation analysis

The useful question is not only how to calculate, but what the calculation assumes. CCA finds vectors a and b so Xa and Yb are highly correlated. In integration, these directions expose variation shared across datasets.

A concrete situation

Two technologies can express one cell-state gradient with different gene weights.

What the mathematics does not license

High canonical correlation can reflect batch-confounded biology or a few outliers. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Compute correlation between two centred score vectors. 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. Multiple canonical dimensions

A small example makes the mechanism visible. Several canonical pairs form a shared low-dimensional space. Dimensions are ordered by the fitted correlation criterion.

A concrete situation

An early dimension may separate immune lineages; a later one may capture activation.

What the mathematics does not license

Training correlation tends to rise with flexible choices. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Assess dimensions on held-out donors and inspect stability. 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. Normalising cell embeddings

This is where modelling choices begin to affect evidence. Cell vectors in the shared space may be normalised before neighbour search so direction and relative structure dominate scale.

A concrete situation

L2 normalisation places score vectors on a unit sphere.

What the mathematics does not license

Normalisation can suppress magnitude that carries biology. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Normalise (3,4) to (0.6,0.8) and explain the change. 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. Mutual nearest neighbours as anchors

The central mathematical idea is concrete. Candidate anchors are reciprocal cross-dataset neighbours in the selected embedding. Reciprocity reduces one-sided matches in dense regions.

A concrete situation

Cell A chooses B and B chooses A across datasets.

What the mathematics does not license

Mutuality cannot rescue a space with no true overlap. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Draw directed neighbour lists and retain reciprocal pairs. 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. Anchor filtering

Here the geometry becomes operational. Candidate pairs can be filtered using local high-dimensional neighbourhood evidence. Filtering removes pairs inconsistent with broader structure.

A concrete situation

A pair close in CCA but far in gene-space neighbours can be rejected.

What the mathematics does not license

A filter can also remove rare valid states. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Report retained and rejected anchors by cell state and donor. 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. Anchor scoring

This step deserves a careful pause. Anchor scores summarise local agreement and help downweight weak correspondences. Scores are relative evidence, not calibrated probabilities of identity.

A concrete situation

Pairs supported by many consistent neighbours receive stronger weight.

What the mathematics does not license

A high score can still join confounded states. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Plot score distributions and inspect extreme pairs. 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. Integration vectors

The useful question is not only how to calculate, but what the calculation assumes. An anchor pair supplies a difference vector between reference and query representations. Nearby query cells receive weighted combinations of such vectors.

A concrete situation

Several local anchors can estimate a smooth batch correction.

What the mathematics does not license

Long-distance anchors can extrapolate an unsupported shift. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Calculate a weighted average of two difference vectors. 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. Reference choice and guide trees

A small example makes the mechanism visible. When integrating many datasets, a reference or merge guide influences computation and geometry. The choice should reflect coverage and intended reuse.

A concrete situation

A diverse healthy reference may anchor several query cohorts.

What the mathematics does not license

A biased reference can make unique query biology look abnormal. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat defensible reference choices and compare neighbourhoods. 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. Reference mapping without rebuilding

This is where modelling choices begin to affect evidence. A stored reference can accept query cells through a defined mapping instead of recomputing the complete atlas. This supports reproducible reuse.

A concrete situation

New cells can inherit reference coordinates and labels.

What the mathematics does not license

A query outside reference coverage still receives a numerical result. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Add novelty and mapping-score thresholds. 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. Label transfer mathematics

The central mathematical idea is concrete. Anchor weights connect reference labels to query cells. Class scores aggregate weighted evidence and the maximum score becomes a prediction with uncertainty.

A concrete situation

A query T cell may receive 0.82 support for a reference label.

What the mathematics does not license

The score is not necessarily a clinical probability. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Build a two-label weighted vote and inspect ties. 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 condition-specific states

Here the geometry becomes operational. Integration should align shared identities while preserving real treatment, disease or developmental differences.

A concrete situation

Activated cells may be unique to one condition.

What the mathematics does not license

Forcing them onto resting references can erase discovery. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Compare raw markers and donor support for large corrections. 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. Donor-aware validation

This step deserves a careful pause. Cells from one donor are related. Hold out donors or studies so near-duplicate biology does not leak into both sides.

A concrete situation

A reference tuned on every donor can overstate transfer.

What the mathematics does not license

Cell count is not the independent-sample count. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Report donors, samples and cells separately. 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 anchor project

The useful question is not only how to calculate, but what the calculation assumes. Create two shifted point clouds with one shared cluster and one unique cluster. Use CCA-like linear projections and reciprocal neighbours.

A concrete situation

Known geometry reveals correct and false anchors.

What the mathematics does not license

Real counts add noise, sparsity and many genes. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Vary overlap, neighbour count and reference choice. 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

A small example makes the mechanism visible. Version reference data, features, CCA, anchors, weights, labels and thresholds. Mapping joins linear algebra, computation and biological review.

A concrete situation

Teams include statisticians, biologists, engineers and data stewards.

What the mathematics does not license

An anchor proves no mechanism, diagnosis or career outcome. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Create a reference model card with supported and rejected queries. 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 Seurat anchor-integration audit begins with explicit datasets, donors, feature mappings and a declared use: joint visualisation, reference mapping or label transfer. The 2019 integration strategy uses canonical correlation analysis to expose shared variation, then finds cross-dataset cell pairs that behave like mutual nearest neighbours. These anchors are candidate correspondences, not ground-truth cell identities.

Record variable features, scaling, CCA dimensions, neighbour parameters, anchor filters, anchor scores, reference choice and integration order. Inspect which donors and cell states contribute anchors. A large anchor count can still be misleading if the pairs join biologically different states or arise from a confounded design.

Validate with donor or study holdouts, raw-marker checks, neighbourhood preservation, rare-state review and a separate label-transfer assessment. Keep the original measurements and embeddings. Integrated values can support comparison and clustering, while differential-expression questions often require the unintegrated assay and an explicit sample-level model.

---

Frequently asked questions

What is an integration anchor?

It is a cross-dataset cell pair judged to represent a shared biological state under the selected feature space and neighbour procedure.

Why use canonical correlation analysis?

CCA seeks linear combinations whose scores correlate across datasets, exposing shared variation that may be obscured by dataset-specific scales.

Are anchors the same as true cell labels?

No. They are algorithmic correspondences. Labels and marker evidence are used separately for validation or transfer.

What is label transfer?

A reference label is predicted for a query cell by weighting information from reference-query anchors and nearby structure, with a transfer score or uncertainty.

Can integration erase real differences?

Yes. Wrong anchors or aggressive correction can merge condition-specific or rare states.

How should anchor integration be validated?

Use donor-held-out tests, anchor audits, neighbourhood and marker preservation, rare-state checks, label-transfer confusion matrices and untouched studies.

---

Primary source and fact-check note

Tim Stuart, Andrew Butler, Paul Hoffman et al., “Comprehensive Integration of Single-Cell Data,” Cell 177(7), 1888–1902.e21 (2019), DOI 10.1016/j.cell.2019.05.031. 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

• Mutual Nearest Neighbours batch correction
• Harmony single-cell integration
• CORAL covariance alignment

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: validation, transfer and responsible use

Reference registry

Archive reference version, donors, conditions, feature space, labels and exclusions. A label-transfer claim is limited to this registry. New technologies or populations require coverage checks rather than automatic reuse.

CCA sensitivity

Repeat justified feature and dimension choices. Compare canonical subspaces, anchor identities and held-out mapping. If small changes produce different anchors, report fragile geometry rather than one definitive map.

Anchor ledger

Save pair identifiers, donors, distances, filter status, scores and labels used only for evaluation. Summarise support by cell state. This reveals whether a reference is broad or anchored by one dense population.

False-anchor challenge

Add a condition-specific cluster with no true counterpart. A sound audit should show low support or rejection. If it maps confidently, revise features, thresholds or the reference.

Weight audit

For selected query cells, store contributing anchors and normalised weights. Large weight on a distant or cross-state anchor deserves review. Reproduce one integrated vector exactly.

Reference bias

Repeat mapping against alternative references where possible. Differences can reveal coverage or label-definition dependence. A consensus is stronger than unexamined reliance on one atlas.

Label transfer metrics

Use held-out confusion matrices, class sensitivity, rejection rates and score distributions. Overall accuracy can hide failure on rare or clinically important states.

Novelty policy

Define when a query is too remote for transfer. Return unknown or review rather than forcing the nearest reference label. Track abstention as a meaningful outcome.

Marker preservation

Check raw marker patterns and donor summaries before and after mapping. Integrated coordinates organise cells; they do not replace measurement-level evidence.

Leakage control

Feature selection, anchor tuning and threshold choice belong inside donor-aware training. Final labels can evaluate but must not shape the mapping they judge.

Student extension

Compare nearest-neighbour anchors before and after L2 normalisation. Learners can see how magnitude and direction affect correspondence.

Reader checkpoint

A complete report names datasets, features, CCA, anchors, filters, weights, reference, rejection, donor holdout and preserved biology. Without those facts, transferred labels are not auditable.

Deep dive: how to stress-test the mathematical claim

Deep dive 1

Inspect anchors with the largest correction influence and return to raw markers. A few high-weight pairs can move many query cells.

Deep dive 2

Compare CCA anchors with RPCA or another justified representation under identical donor holdouts. Disagreement reveals dependence on geometry.

Deep dive 3

Use a reference missing one known cell type. A responsible mapping should abstain or show low support rather than relabel the query type.

Deep dive 4

Check whether label-transfer scores are calibrated by cell type. The same score can have different error rates in dense and rare classes.

Deep dive 5

Repeat feature selection inside outer folds. Global variable features can leak query or test structure.

Deep dive 6

Separate integration success from annotation success. A good joint embedding can still have poor label definitions.

Deep dive 7

Test condition-specific markers after integration using unintegrated expression and donor-level models.

Deep dive 8

Report queries rejected for low coverage. Rejection is evidence of model honesty, not a blemish to remove.

Deep dive 9

Version the reference. New labels or donors change every later transfer claim.

Deep dive 10

End with a claim ladder: anchor correspondence, geometric mapping, label prediction and biological identity require increasing evidence.

Worked evidence notebook

CCA calculation

For centred score vectors u and v, compute their covariance and standard deviations, then correlation. State that CCA optimises weight vectors rather than accepting fixed scores.

Unit-vector check

Normalise (3,4) to (0.6,0.8) and verify length one.

Anchor graph

List directed cross-dataset neighbours and keep reciprocal pairs. Add filter and score columns.

Weighted correction

Combine two anchor difference vectors with normalised weights and reproduce one query correction.

Label vote

Aggregate reference label weights for one query and report the maximum score plus alternatives.

Coverage note

Record distance, anchor count and rejection status for each mapped query.

Practice, misconception and transfer

A parent-friendly starting point

Begin with correlation, unit vectors and nearest neighbours. A student can understand why two datasets need a shared coordinate system before learning CCA software.

A key misconception

Integration anchors are proposed matches, not certified labels. The algorithm supplies evidence that must be checked against markers, donors and experiments.

When to pause

Pause if the reference lacks the query population, feature mapping is weak or condition is confounded with batch. Stronger correction cannot create missing evidence.

A useful extension

Build an unknown category and a rejection threshold. This teaches that responsible AI includes knowing when not to classify.

Career transfer

The same mathematics appears in recommendation, computer vision and domain adaptation, while each field supplies different evidence and risks.

Final learning move

Ask the learner to reproduce one anchor, correction and label vote by hand before trusting a large pipeline.

A final worked scenario

A final reference-mapping scenario

A reference atlas contains four donors and several well-supported immune states. A fifth donor is held out as query. CCA exposes shared dimensions, reciprocal neighbours form anchors and weak pairs are filtered. Common cell types map with stable markers and high transfer accuracy, while a treatment-specific state receives low support and remains unknown. The team reports the rejection rather than forcing a label, uses raw counts for donor-level differential expression and versions the reference. This is a successful mathematical result because the correspondence, correction, vote and uncertainty are all visible. The anchors organise evidence; they do not replace biological judgement.

Mathematical workshop: mechanisms, evidence and limits

Why anchors need more than proximity

An anchor is useful only when closeness in the chosen representation is a reasonable sign of shared state. CCA can expose correlated combinations even when the original variables are noisy, but correlation alone does not establish identity. Density matters too: a common state supplies many candidate neighbours, while a rare state may have only a handful. A strong audit therefore records reciprocal-neighbour status, distance, neighbourhood consistency, marker agreement and donor coverage. It also asks whether a candidate pair survives when one donor is removed. That combination turns an attractive point on a plot into a traceable item of evidence.

From an anchor pair to a correction vector

Suppose a query cell has two supported reference anchors. In the shared space, the reference-minus-query difference for the first pair is (0.20, -0.10), and for the second it is (0.10, 0.05). If normalised weights are 0.75 and 0.25, the weighted correction is 0.75(0.20, -0.10)+0.25(0.10, 0.05)=(0.175, -0.0625). This arithmetic is simple, yet it exposes every assumption: which anchors contributed, how weights were obtained and how far the cell moved. A responsible report stores those quantities rather than treating the integrated coordinate as an unexplained output.

Reference mapping is a transport problem

Reference mapping transports a query into an existing coordinate and label system. That can make a mature atlas reusable, but it also imports the atlas definitions, omissions and biases. If the reference contains only adult tissue, a developmental query may be forced near the least unsuitable adult state. If one ancestry or treatment group is under-represented, anchor support can be uneven. Coverage plots by donor, condition and cell state are therefore part of the result. The question is not merely whether mapping ran; it is where the reference had enough support to justify the transport.

Label transfer as a weighted vote

Label transfer can be explained without mystique. Each supported anchor contributes its reference label with a weight, and the weights for each candidate label are added. A query with weights 0.55 and 0.25 from label A and 0.20 from label B receives scores 0.80 and 0.20. That does not mean an 80% biological probability unless the score has been calibrated. Held-out labels can reveal how often a score range is correct for each class. Rare classes may need different thresholds, and an unknown option is essential when no candidate receives adequate support.

Integration and inference answer different questions

Integrated coordinates are designed to organise shared structure. Differential expression asks whether measured values differ across biological replicates. Performing inference on corrected coordinates can hide the measurement model and reduce real condition effects. A safer workflow uses integration for visualisation, neighbourhood discovery and tentative annotation, then returns to unintegrated counts with donor-aware statistical models for inference. The two stages can inform each other, but their claims remain separate. This distinction helps students see why the most visually pleasing representation is not automatically the correct dataset for every later calculation.

A student experiment with a missing class

Create two small datasets with three shared clusters and add a fourth cluster only to the query. Run the mapping twice: once with forced nearest-label assignment and once with a rejection rule based on distance and anchor support. The forced version will always name the extra cluster, even though no true counterpart exists. The rejection-aware version can preserve it as unknown. Repeat after changing cluster density and noise. This experiment teaches geometry, threshold choice, class imbalance and scientific restraint in one notebook, and it makes the limits of reference mapping visible.

What parents can ask about an atlas result

Parents do not need to inspect code to encourage strong reasoning. Ask: Which cells were in the reference? Was the child dataset represented by independent donors? Which query cells lacked anchors? Were transferred labels checked against raw markers? Did the team keep an unknown category? Could another reference change the result? These questions shift attention from software prestige to evidence. They also model a valuable habit for any data-rich field: identify the comparison group, trace the calculation, examine unsupported cases and match the strength of the conclusion to the strength of the test.

A transparent final claim

A defensible statement might say: across five held-out donors, common immune states formed stable reciprocal anchors to a versioned reference, transferred labels achieved the reported class-specific errors and treatment-specific cells with low support were left unlabelled. That sentence identifies the population, validation unit, result and boundary. It does not say that integration proved cellular identity or removed all technical variation. Clear mathematical language makes the achievement stronger because another team can reproduce the anchor graph, scores and rejection rule and learn exactly where the result applies.

Method checklist and conclusion boundary

Reproducibility checklist

Freeze features, CCA, neighbour settings, anchor filters, scores, weights, reference and label dictionary. Hold out donors, audit influential anchors, test rejection and preserve raw assays. Reproduce one mapping and one label vote.

Careful conclusion

A careful conclusion says that supported query cells mapped to a defined reference through stable anchors and that held-out labels transferred with stated errors. It does not claim universal identity or mechanism.

Final evidence checks

Final evidence checks 1

Preserve reference and query raw assays.

Final evidence checks 2

Version the reference label dictionary.

Final evidence checks 3

Report donor counts and rejected queries.

Final evidence checks 4

Audit influential anchors and weights.

Final evidence checks 5

Separate integrated clustering from differential-expression inference.

Final evidence checks 6

Require independent reproduction of one mapped cell.