Chapter map

1. 1. Why fixed marker rules fail
2. 2. Marker indicator matrix
3. 3. Known-type clusters
4. 4. Base expression level
5. 5. Elevated marker increment
6. 6. Per-cell marker flexibility
7. 7. Marker confidence weights
8. 8. Non-marker centroid term
9. 9. Joint objective
10. 10. Alternating optimisation
11. 11. Multiple random starts
12. 12. Elevated-marker proportion
13. 13. Decided and undecided groups
14. 14. Unknown-cell similarity test
15. 15. Worked constrained example
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 fixed marker rules fail

The useful question is not only how to calculate, but what the calculation assumes. Dropout and biological heterogeneity mean a true cell may not overexpress every declared marker.

A concrete situation

One of four B-cell markers is low.

What the mathematics does not license

Relaxation can also admit incorrect cells. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Separate flexibility from evidence 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.

---

2. Marker indicator matrix

A small example makes the mechanism visible. A binary matrix gamma records whether gene i is a marker for type k.

A concrete situation

Gamma equals one for marker A and type 2.

What the mathematics does not license

The matrix may contain outdated relationships. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Cite each nonzero entry. 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. Known-type clusters

This is where modelling choices begin to affect evidence. With K known types, the first stage partitions every cell into K provisional clusters.

A concrete situation

Four known types give four clusters.

What the mathematics does not license

Unknown cells are forced somewhere initially. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Treat stage-one labels as provisional. 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. Base expression level

The central mathematical idea is concrete. For a marker gene, mu represents its ordinary level outside the supported type.

A concrete situation

Most nonsupported cells sit near 0.2.

What the mathematics does not license

A shared marker may not have one common baseline. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Inspect residuals by candidate 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.

---

5. Elevated marker increment

Here the geometry becomes operational. A nonnegative delta raises the supported-type level from mu to mu plus delta.

A concrete situation

Base 0.2 and delta 0.7 give 0.9.

What the mathematics does not license

A near-zero delta makes the marker uninformative. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Plot fitted increments. 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. Per-cell marker flexibility

This step deserves a careful pause. Within the supported cluster, the objective uses whichever squared error is smaller: distance to base or to elevated level.

A concrete situation

A dropout-like cell chooses the base term.

What the mathematics does not license

Too much flexibility can weaken discrimination. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Count base and elevated choices per marker. 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. Marker confidence weights

The useful question is not only how to calculate, but what the calculation assumes. Weights scale marker errors and can express stronger prior confidence; the paper balances their total contribution against non-markers.

A concrete situation

Doubling w emphasises signature fit.

What the mathematics does not license

Weights chosen after seeing truth cause leakage. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Lock weights in training only. 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. Non-marker centroid term

A small example makes the mechanism visible. For non-marker genes, the objective is ordinary K-means within-cluster sum of squares around each centroid.

A concrete situation

A cell joins the closest broad-expression profile.

What the mathematics does not license

High-dimensional distance can be noisy. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Standardise and test feature choices. 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. Joint objective

This is where modelling choices begin to affect evidence. Marker and non-marker terms are added, so the solution reflects both prior biological guidance and data-wide cluster compactness.

A concrete situation

A weak marker is rescued by coherent non-marker profile.

What the mathematics does not license

The balance is a scientific assumption. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Plot outcomes across w values. 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. Alternating optimisation

The central mathematical idea is concrete. Holding parameters fixed, assign each cell to the lowest-cost cluster; holding assignments fixed, update baselines, increments and centroids.

A concrete situation

Costs fall across iterations.

What the mathematics does not license

Alternation ensures a local, not global, result. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Save every objective value. 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. Multiple random starts

Here the geometry becomes operational. The published workflow repeats optimisation from different cluster initialisations and chooses the best final objective.

A concrete situation

Ten starts reach three distinct local fits.

What the mathematics does not license

The lowest cost may still be biologically poor. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Compare stability as well as cost. 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. Elevated-marker proportion

This step deserves a careful pause. For a provisionally assigned cell, alpha is the fraction of that type’s markers fitted at the elevated level.

A concrete situation

Three of four markers are elevated, so alpha is 0.75.

What the mathematics does not license

Few-marker types make fractions coarse. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Report numerator and denominator. 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. Decided and undecided groups

The useful question is not only how to calculate, but what the calculation assumes. Cells above a chosen alpha threshold form a trusted decided set; weaker cells enter an undecided set for further testing.

A concrete situation

Alpha 0.75 passes a 0.5 rule.

What the mathematics does not license

A hard boundary can be unstable. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Plot coverage across alpha 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.

---

14. Unknown-cell similarity test

A small example makes the mechanism visible. Undecided cells are compared with trusted cells through non-marker expression; insufficient similarity leads to an unknown outcome.

A concrete situation

A weak-marker cell still resembles decided neighbours.

What the mathematics does not license

Unknown populations need not form one cluster. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Test heterogeneous withheld types. 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. Worked constrained example

This is where modelling choices begin to affect evidence. Use two types, two markers and two non-markers to compare base/elevated choices and centroid costs.

A concrete situation

One true type-A cell has a low A marker.

What the mathematics does not license

A four-gene toy understates sparsity. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat with a target-only cell. 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

The central mathematical idea is concrete. Version gamma, weights, objectives, starts, parameters, alpha and unknown thresholds. This links optimisation, clustering and bioinformatics.

A concrete situation

Students can begin with weighted squared distances.

What the mathematics does not license

A label remains conditional evidence. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Write an optimisation audit card. 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 scSorter audit begins with a gene-by-type indicator matrix, expression measurements and a declared balance between marker and non-marker evidence. Its first stage solves a constrained clustering problem that lets each marker be either at a base or elevated level while non-marker genes contribute ordinary within-cluster squared distance.

Record marker weights, base levels, elevated increments, non-marker centroids, cluster assignments, objective values, random starts and convergence. Then preserve the second-stage unknown-cell calculation: the fraction of elevated markers, the decided and undecided groups, the non-marker similarity test and the chosen threshold for every known cluster.

Validate under marker dropout, incorrect markers, rare types, target-only types, donor and technology holdouts and repeated initialisations. Report both known-type assignment and unknown detection. A local optimum with a familiar label remains a mathematical candidate, not proof that the cell has the claimed lineage or function.

---

Frequently asked questions

What is scSorter?

scSorter is a marker-guided method that assigns cells to known types without requiring a labelled reference dataset. It solves a constrained clustering problem and then performs a separate unknown-cell calling step.

Why use non-marker genes?

Non-marker genes contribute within-cluster squared distances, so the broader expression profile can support structure when individual marker genes are inconsistently high.

Can a marker be low in a true cell?

Yes. For its corresponding type, scSorter lets each marker observation choose the closer of a base level and an elevated level.

What does the marker weight mean?

A positive weight expresses confidence in a marker–type relationship and controls its contribution relative to the non-marker part of the objective.

Does optimisation guarantee the global best solution?

No. The paper describes a combinatorial problem and an alternating algorithm that reaches a local optimum, so it uses multiple random starts and keeps the best final cost.

How are unknown cells called?

After all cells are provisionally clustered into known types, scSorter examines the proportion of elevated markers and uses non-marker similarity to separate sufficiently supported known cells from unknown cells.

---

Primary source and fact-check note

Hongyu Guo and Jun Li, “scSorter: assigning cells to known cell types according to marker genes,” Genome Biology 22, 69 (2021), DOI 10.1186/s13059-021-02281-7. 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

• SCINA semi-supervised marker assignment
• SciBet portable probability models
• scLearn learned comparison geometry
• How Mathematics Works

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

Indicator-matrix audit

Store gamma, marker weights and biological evidence. Separate missing markers from explicitly non-marker genes.

Base/elevated audit

For every marker–type pair, report mu, delta and the fraction of assigned cells choosing each level.

Weight-sensitivity study

Vary the overall marker-to-non-marker balance without consulting final labels. Plot accuracy, stability and unknown-cell recovery.

Local-optimum audit

Run many initial partitions and compare cost, adjusted Rand agreement, cluster parameters and biological plausibility.

Alpha-threshold curve

Plot decided coverage, known-cell error and unknown recovery across elevated-marker proportion thresholds.

Non-marker test audit

Record trusted and undecided sets, distance or test statistics, selected thresholds and sensitivity to feature filtering.

Heterogeneous-unknown challenge

Mix several target-only populations that do not form one cluster. Quantify whether the two-stage design rejects each one.

Final claim

Name gamma, weights, objective, starts, fitted levels, alpha and unknown rule. A local optimum remains conditional evidence.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Why fixed marker rules fail

Start with the mechanism: Dropout and biological heterogeneity mean a true cell may not overexpress every declared marker. The worked case is concrete: One of four B-cell markers is low. Expose the boundary: Relaxation can also admit incorrect cells. Complete this evidence step: Separate flexibility from evidence quality. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 2: Marker indicator matrix

Start with the mechanism: A binary matrix gamma records whether gene i is a marker for type k. The worked case is concrete: Gamma equals one for marker A and type 2. Expose the boundary: The matrix may contain outdated relationships. Complete this evidence step: Cite each nonzero entry. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 3: Known-type clusters

Start with the mechanism: With K known types, the first stage partitions every cell into K provisional clusters. The worked case is concrete: Four known types give four clusters. Expose the boundary: Unknown cells are forced somewhere initially. Complete this evidence step: Treat stage-one labels as provisional. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 4: Base expression level

Start with the mechanism: For a marker gene, mu represents its ordinary level outside the supported type. The worked case is concrete: Most nonsupported cells sit near 0.2. Expose the boundary: A shared marker may not have one common baseline. Complete this evidence step: Inspect residuals by candidate type. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 5: Elevated marker increment

Start with the mechanism: A nonnegative delta raises the supported-type level from mu to mu plus delta. The worked case is concrete: Base 0.2 and delta 0.7 give 0.9. Expose the boundary: A near-zero delta makes the marker uninformative. Complete this evidence step: Plot fitted increments. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 6: Per-cell marker flexibility

Start with the mechanism: Within the supported cluster, the objective uses whichever squared error is smaller: distance to base or to elevated level. The worked case is concrete: A dropout-like cell chooses the base term. Expose the boundary: Too much flexibility can weaken discrimination. Complete this evidence step: Count base and elevated choices per marker. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 7: Marker confidence weights

Start with the mechanism: Weights scale marker errors and can express stronger prior confidence; the paper balances their total contribution against non-markers. The worked case is concrete: Doubling w emphasises signature fit. Expose the boundary: Weights chosen after seeing truth cause leakage. Complete this evidence step: Lock weights in training only. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 8: Non-marker centroid term

Start with the mechanism: For non-marker genes, the objective is ordinary K-means within-cluster sum of squares around each centroid. The worked case is concrete: A cell joins the closest broad-expression profile. Expose the boundary: High-dimensional distance can be noisy. Complete this evidence step: Standardise and test feature choices. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 9: Joint objective

Start with the mechanism: Marker and non-marker terms are added, so the solution reflects both prior biological guidance and data-wide cluster compactness. The worked case is concrete: A weak marker is rescued by coherent non-marker profile. Expose the boundary: The balance is a scientific assumption. Complete this evidence step: Plot outcomes across w values. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 10: Alternating optimisation

Start with the mechanism: Holding parameters fixed, assign each cell to the lowest-cost cluster; holding assignments fixed, update baselines, increments and centroids. The worked case is concrete: Costs fall across iterations. Expose the boundary: Alternation ensures a local, not global, result. Complete this evidence step: Save every objective value. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 11: Multiple random starts

Start with the mechanism: The published workflow repeats optimisation from different cluster initialisations and chooses the best final objective. The worked case is concrete: Ten starts reach three distinct local fits. Expose the boundary: The lowest cost may still be biologically poor. Complete this evidence step: Compare stability as well as cost. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 12: Elevated-marker proportion

Start with the mechanism: For a provisionally assigned cell, alpha is the fraction of that type’s markers fitted at the elevated level. The worked case is concrete: Three of four markers are elevated, so alpha is 0.75. Expose the boundary: Few-marker types make fractions coarse. Complete this evidence step: Report numerator and denominator. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 13: Decided and undecided groups

Start with the mechanism: Cells above a chosen alpha threshold form a trusted decided set; weaker cells enter an undecided set for further testing. The worked case is concrete: Alpha 0.75 passes a 0.5 rule. Expose the boundary: A hard boundary can be unstable. Complete this evidence step: Plot coverage across alpha thresholds. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 14: Unknown-cell similarity test

Start with the mechanism: Undecided cells are compared with trusted cells through non-marker expression; insufficient similarity leads to an unknown outcome. The worked case is concrete: A weak-marker cell still resembles decided neighbours. Expose the boundary: Unknown populations need not form one cluster. Complete this evidence step: Test heterogeneous withheld types. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 15: Worked constrained example

Start with the mechanism: Use two types, two markers and two non-markers to compare base/elevated choices and centroid costs. The worked case is concrete: One true type-A cell has a low A marker. Expose the boundary: A four-gene toy understates sparsity. Complete this evidence step: Repeat with a target-only cell. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 16: Governance and careers

Start with the mechanism: Version gamma, weights, objectives, starts, parameters, alpha and unknown thresholds. This links optimisation, clustering and bioinformatics. The worked case is concrete: Students can begin with weighted squared distances. Expose the boundary: A label remains conditional evidence. Complete this evidence step: Write an optimisation audit card. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare marker modes, responsibilities, objective terms, centroids, landmarks, soft assignments, losses or unknown outcomes. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.