WHY MATHEMATICS?
Why Mathematics? | Symphony for Atlas Mapping: Reference Compression, PCA Projection, Soft Clusters and Query Correction
How Symphony compresses an integrated atlas into a portable mathematical reference, maps new cells into a stable embedding and transfers annotations—without rebuilding the atlas for every query.
Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.
Choose a reading route
- Build and validate a stable integrated reference
- Compress the atlas into reusable mathematical summaries
- Project query cells through reference PCA and soft clusters
- Correct, place, annotate and test unsupported states
Chapter map
- 1. Why atlas mapping needs a stable frame
- 2. Building the integrated reference
- 3. Reference PCA
- 4. Soft reference clusters
- 5. Harmony-derived correction terms
- 6. Reference compression
- 7. Query normalisation and scaling
- 8. PCA projection of queries
- 9. Query soft-cluster assignment
- 10. Query batch correction
- 11. Stable placement in the atlas
- 12. Annotation and feature transfer
- 13. Trajectory localisation
- 14. Held-out and novel-state validation
- 15. A student compressed-reference project
- 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 atlas mapping needs a stable frame
The useful question is not only how to calculate, but what the calculation assumes. A large reference should remain fixed so queries from different times can be compared in the same coordinate system.
A concrete situation
Many laboratories map cohorts to one immune atlas.
What the mathematics does not license
A stable frame can preserve reference bias. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Version the atlas and benchmark set. 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. Building the integrated reference
A small example makes the mechanism visible. The reference is first harmonised across donors, studies or platforms with a declared biological target.
A concrete situation
Shared cell states form a common embedding.
What the mathematics does not license
Confounded design can erase condition biology. That distinction matters because a correct formula can still answer the wrong operational question.
Try it, check it, transfer it
Audit reference integration before compression. 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. Reference PCA
This is where modelling choices begin to affect evidence. Principal components define linear directions that compress reference variation. Loadings become part of the reusable mapping operator.
A concrete situation
Genes multiply by stored loadings to yield scores.
What the mathematics does not license
A query with missing or shifted genes projects poorly. The habit of checking assumptions is one of the practical benefits of learning mathematics.
Try it, check it, transfer it
Check feature order 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.
4. Soft reference clusters
The central mathematical idea is concrete. Reference cells have graded memberships in multiple clusters rather than one hard partition.
A concrete situation
A transitional cell belongs partly to two clusters.
What the mathematics does not license
Softness is not calibrated biological probability. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Plot membership entropy and 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.
Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.
5. Harmony-derived correction terms
Here the geometry becomes operational. Cluster-specific batch terms describe nuisance shifts learned during reference construction.
A concrete situation
Donor offsets are represented within soft clusters.
What the mathematics does not license
The model inherits Harmony assumptions and settings. That distinction matters because a correct formula can still answer the wrong operational question.
Try it, check it, transfer it
Save objectives, penalties and convergence. 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. Reference compression
This step deserves a careful pause. Symphony stores sufficient summaries for mapping instead of every reference count.
A concrete situation
A compact object is portable and fast.
What the mathematics does not license
Compression can omit context needed for later audits. The habit of checking assumptions is one of the practical benefits of learning mathematics.
Try it, check it, transfer it
Publish a manifest of retained objects. 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. Query normalisation and scaling
The useful question is not only how to calculate, but what the calculation assumes. Query cells must follow the reference feature and scaling contract before projection.
A concrete situation
A new study uses the same genes and transforms.
What the mathematics does not license
Recomputing query-specific scales changes meaning. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Apply locked reference parameters. 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. PCA projection of queries
A small example makes the mechanism visible. A query vector is multiplied by reference loadings to enter the reference PC space.
A concrete situation
New cells obtain coordinates without refitting PCA.
What the mathematics does not license
Out-of-range values still receive coordinates. That distinction matters because a correct formula can still answer the wrong operational question.
Try it, check it, transfer it
Measure leverage and reconstruction. 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.
Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.
9. Query soft-cluster assignment
This is where modelling choices begin to affect evidence. Projected cells receive memberships relative to reference centroids.
A concrete situation
An immune query distributes across neighbouring reference states.
What the mathematics does not license
Novel states may be split among familiar clusters. The habit of checking assumptions is one of the practical benefits of learning mathematics.
Try it, check it, transfer it
Track maximum membership and entropy. 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. Query batch correction
The central mathematical idea is concrete. Reference-derived cluster terms adjust query coordinates while leaving the reference embedding unchanged.
A concrete situation
A platform shift is removed conditionally on cluster membership.
What the mathematics does not license
Disease variation can be mistaken for batch. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Compare raw and corrected displacement. 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. Stable placement in the atlas
Here the geometry becomes operational. Corrected query coordinates are placed beside reference cells for neighbour search and visualisation.
A concrete situation
Repeated queries share one coordinate frame.
What the mathematics does not license
Two-dimensional plots can hide unsupported distance. That distinction matters because a correct formula can still answer the wrong operational question.
Try it, check it, transfer it
Audit high-dimensional neighbours. 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. Annotation and feature transfer
This step deserves a careful pause. Reference labels or measured features can be predicted from nearby atlas cells.
A concrete situation
A query receives a cell-type label or surface-protein estimate.
What the mathematics does not license
Transfer is inference, not measurement. The habit of checking assumptions is one of the practical benefits of learning mathematics.
Try it, check it, transfer it
Report sources, weights and uncertainty. 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.
Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.
13. Trajectory localisation
The useful question is not only how to calculate, but what the calculation assumes. Queries can be localised along a reference developmental path when coverage supports it.
A concrete situation
Fetal liver cells map along haematopoiesis.
What the mathematics does not license
Projection can shortcut or force branches. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Check markers and within-query order. 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 and novel-state validation
A small example makes the mechanism visible. Reserve studies and known absent populations to test mapping and rejection.
A concrete situation
A new species or donor tests portability.
What the mathematics does not license
Good speed can coexist with wrong placement. That distinction matters because a correct formula can still answer the wrong operational question.
Try it, check it, transfer it
Lock thresholds before final 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.
15. A student compressed-reference project
This is where modelling choices begin to affect evidence. Store PCA loadings, centroids and correction offsets for a small dataset, then map a third batch without refitting.
A concrete situation
Known coordinates reveal mapping error.
What the mathematics does not license
Real atlases add soft clusters and millions of cells. The habit of checking assumptions is one of the practical benefits of learning mathematics.
Try it, check it, transfer it
Measure size, speed and drift. 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 genes, loadings, clusters, correction terms, labels and confidence rules. Atlas engineering joins linear algebra, optimisation and stewardship.
A concrete situation
Students can begin with matrix projection.
What the mathematics does not license
Mapping guarantees no identity or clinical outcome. The same reasoning helps students move from an exercise to a real dataset.
Try it, check it, transfer it
Write a compact-reference model 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.
Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.
A seven-day practice lab
- Day 1: define the decision, observation unit, X variables and Y or reference condition.
- Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
- Day 3: build a transparent baseline and write what it assumes.
- Day 4: fit the main method inside a training-only pipeline; record every parameter.
- Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
- Day 6: compare validation metrics with the practical workload created by errors.
- 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 Symphony audit begins with a versioned integrated reference containing PCA loadings, Harmony-derived cluster information, batch-correction terms, centroids and reference annotations. The compact object lets query cells be projected into the same stable low-dimensional space without rebuilding the full atlas.
Record genes, scaling, reference PCA, soft clusters, compression choices, query projection, cluster assignment, correction, neighbour search, label transfer and confidence. Reference compression improves portability, but every omitted detail can limit reproducibility or hide unsupported query states.
Validate with held-out studies, distant technologies, rare and novel states, reference-geometry stability, mapping speed, label accuracy, trajectory localisation and original markers. Efficient mapping is strongest when coverage limits and rejection rules travel with the reference object.
Frequently asked questions
What is Symphony?
It is a method for building compact integrated single-cell references and efficiently mapping query cells into their stable embedding.
What does the compact reference contain?
It stores mathematical summaries such as genes, scaling information, PCA loadings, soft-cluster terms and annotation support rather than the full atlas matrix.
How is a query mapped?
Query expression is normalised and projected through reference PCA, assigned to reference soft clusters, corrected with learned terms and placed in the reference space.
Does Symphony change the reference?
Query mapping is designed to keep the reference embedding stable while localising new cells within it.
Can labels or protein values be transferred?
Reference-defined annotations or features can be inferred through neighbours, but they remain predictions requiring validation.
How should Symphony be validated?
Use held-out studies, novel states, mapping accuracy, reference stability, trajectory checks, marker evidence, rejection rules and runtime measurements.
Primary source and fact-check note
Joyce B. Kang, Aparna Nathan, Kathryn Weinand et al., “Efficient and precise single-cell reference atlas mapping with Symphony,” Nature Communications 12, 5890 (2021), DOI 10.1038/s41467-021-25957-x. 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
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
Compression manifest
List every object retained in the compact reference: genes, scaling values, PCA loadings, centroids, soft-cluster summaries, correction terms, labels and software versions. Also list what was omitted. Portability depends on knowing which future audits remain possible.
Reference drift test
Map a fixed benchmark query to successive reference versions. Compare coordinates, neighbours, transferred labels and feature estimates. A stable coordinate frame should have explicit tolerances and rollback points.
Projection leverage
Measure whether query PC scores fall within the range and density of reference scores. Extreme leverage can indicate feature mismatch or novel biology even when a coordinate is returned. Combine this with reconstruction and neighbour support.
Soft-cluster uncertainty
Store membership vectors, maximum membership and entropy for each query. A diffuse membership may indicate a transition, low quality, novelty or model mismatch. Check original markers before interpretation.
Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.
Correction displacement
Calculate the vector from uncorrected to corrected query coordinates and compare its magnitude with local reference distances. Very large corrections deserve review, especially when condition is confounded with batch.
Transfer validation
Hide known labels or protein measurements in a validation query and test transfer against truth. Report class- or feature-specific error and interval coverage. A transferred value remains a prediction.
Speed and fidelity
Measure reference size, loading time and query runtime beside mapping accuracy and rare-state recall. Efficiency is useful only when compression and approximation preserve the decisions needed by readers.
Final claim
A precise result states the reference version, compact object, mapping error, coverage and unsupported queries. It does not say that fast placement proves identity, lineage or disease mechanism. Stable reuse and visible limits are the achievement.
Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.
Mechanism workshop: calculate, challenge and transfer
Workshop 1: Why atlas mapping needs a stable frame
Start with the mechanism: A large reference should remain fixed so queries from different times can be compared in the same coordinate system. The worked case is concrete: Many laboratories map cohorts to one immune atlas. Expose the boundary immediately: A stable frame can preserve reference bias. Complete this evidence step: Version the atlas and benchmark set. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 2: Building the integrated reference
Start with the mechanism: The reference is first harmonised across donors, studies or platforms with a declared biological target. The worked case is concrete: Shared cell states form a common embedding. Expose the boundary immediately: Confounded design can erase condition biology. Complete this evidence step: Audit reference integration before compression. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 3: Reference PCA
Start with the mechanism: Principal components define linear directions that compress reference variation. Loadings become part of the reusable mapping operator. The worked case is concrete: Genes multiply by stored loadings to yield scores. Expose the boundary immediately: A query with missing or shifted genes projects poorly. Complete this evidence step: Check feature order and scaling. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 4: Soft reference clusters
Start with the mechanism: Reference cells have graded memberships in multiple clusters rather than one hard partition. The worked case is concrete: A transitional cell belongs partly to two clusters. Expose the boundary immediately: Softness is not calibrated biological probability. Complete this evidence step: Plot membership entropy and stability. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.
Workshop 5: Harmony-derived correction terms
Start with the mechanism: Cluster-specific batch terms describe nuisance shifts learned during reference construction. The worked case is concrete: Donor offsets are represented within soft clusters. Expose the boundary immediately: The model inherits Harmony assumptions and settings. Complete this evidence step: Save objectives, penalties and convergence. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 6: Reference compression
Start with the mechanism: Symphony stores sufficient summaries for mapping instead of every reference count. The worked case is concrete: A compact object is portable and fast. Expose the boundary immediately: Compression can omit context needed for later audits. Complete this evidence step: Publish a manifest of retained objects. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 7: Query normalisation and scaling
Start with the mechanism: Query cells must follow the reference feature and scaling contract before projection. The worked case is concrete: A new study uses the same genes and transforms. Expose the boundary immediately: Recomputing query-specific scales changes meaning. Complete this evidence step: Apply locked reference parameters. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 8: PCA projection of queries
Start with the mechanism: A query vector is multiplied by reference loadings to enter the reference PC space. The worked case is concrete: New cells obtain coordinates without refitting PCA. Expose the boundary immediately: Out-of-range values still receive coordinates. Complete this evidence step: Measure leverage and reconstruction. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.
Workshop 9: Query soft-cluster assignment
Start with the mechanism: Projected cells receive memberships relative to reference centroids. The worked case is concrete: An immune query distributes across neighbouring reference states. Expose the boundary immediately: Novel states may be split among familiar clusters. Complete this evidence step: Track maximum membership and entropy. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 10: Query batch correction
Start with the mechanism: Reference-derived cluster terms adjust query coordinates while leaving the reference embedding unchanged. The worked case is concrete: A platform shift is removed conditionally on cluster membership. Expose the boundary immediately: Disease variation can be mistaken for batch. Complete this evidence step: Compare raw and corrected displacement. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 11: Stable placement in the atlas
Start with the mechanism: Corrected query coordinates are placed beside reference cells for neighbour search and visualisation. The worked case is concrete: Repeated queries share one coordinate frame. Expose the boundary immediately: Two-dimensional plots can hide unsupported distance. Complete this evidence step: Audit high-dimensional neighbours. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 12: Annotation and feature transfer
Start with the mechanism: Reference labels or measured features can be predicted from nearby atlas cells. The worked case is concrete: A query receives a cell-type label or surface-protein estimate. Expose the boundary immediately: Transfer is inference, not measurement. Complete this evidence step: Report sources, weights and uncertainty. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.
Workshop 13: Trajectory localisation
Start with the mechanism: Queries can be localised along a reference developmental path when coverage supports it. The worked case is concrete: Fetal liver cells map along haematopoiesis. Expose the boundary immediately: Projection can shortcut or force branches. Complete this evidence step: Check markers and within-query order. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 14: Held-out and novel-state validation
Start with the mechanism: Reserve studies and known absent populations to test mapping and rejection. The worked case is concrete: A new species or donor tests portability. Expose the boundary immediately: Good speed can coexist with wrong placement. Complete this evidence step: Lock thresholds before final queries. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 15: A student compressed-reference project
Start with the mechanism: Store PCA loadings, centroids and correction offsets for a small dataset, then map a third batch without refitting. The worked case is concrete: Known coordinates reveal mapping error. Expose the boundary immediately: Real atlases add soft clusters and millions of cells. Complete this evidence step: Measure size, speed and drift. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.
Workshop 16: Governance and careers
Start with the mechanism: Version genes, loadings, clusters, correction terms, labels and confidence rules. Atlas engineering joins linear algebra, optimisation and stewardship. The worked case is concrete: Students can begin with matrix projection. Expose the boundary immediately: Mapping guarantees no identity or clinical outcome. Complete this evidence step: Write a compact-reference model card. Write the input vectors or matrices, the operation, the expected output, uncertainty and one observation that would contradict the claim. Repeat the calculation under a nearby defensible setting and compare assignments, neighbours or corrections. Use untouched donors, studies or simulated truth where possible because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so every conclusion remains auditable.