WHY MATHEMATICS?

Why Mathematics? | SCINA for Semi-Supervised Cell Typing: Marker Signatures, Bimodal Expression, Expectation–Maximisation and Unknown Cells

How SCINA turns expert marker sets into a transparent semi-supervised model: marker expression is represented by high and low modes, EM alternates between probabilistic assignment and parameter fitting, and cells unsupported by every declared signature can remain unknown.

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

Translate marker knowledge into a statistical input
Model high and low expression with two modes
Follow expectation–maximisation and posterior assignment
Test unknown cells, noisy markers and transfer limits

Chapter map

1. Why marker knowledge needs mathematics
2. Expression-matrix contract
3. Positive marker sets
4. Inverting low markers
5. Bimodal assumption
6. Latent categorical identity
7. Complete-data likelihood
8. Expectation step
9. Maximisation step
10. Convergence and initialisation
11. Overlapping signatures
12. Unknown category
13. Marker-noise tolerance
14. Worked four-cell example
15. Validation beyond random cells
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 marker knowledge needs mathematics

The central mathematical idea is concrete. A marker list says what should be informative; a probability model explains how several imperfect markers combine across many cells.

A concrete situation

Three markers support one immune type with different strengths.

What the mathematics does not license

Markers can be tissue- or state-dependent. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Cite and version every signature. 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. Expression-matrix contract

Here the geometry becomes operational. SCINA accepts rows of genes and columns of cells or samples after suitable preprocessing chosen by the analyst.

A concrete situation

Counts are normalised and log-transformed before input.

What the mathematics does not license

Mixing raw and transformed scales changes distribution shape. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Store units and transformations. 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. Positive marker sets

This step deserves a careful pause. For each candidate type, one or more genes are declared characteristically high relative to other types.

A concrete situation

Marker A and B support type 1.

What the mathematics does not license

One marker may support several related types. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Draw a gene-by-type signature matrix. 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. Inverting low markers

The useful question is not only how to calculate, but what the calculation assumes. A low-expression marker can be multiplied by minus one so that a larger pseudo-value supports the intended type.

A concrete situation

Values 5, 6, 4 and 0 become −5, −6, −4 and 0.

What the mathematics does not license

Sign inversion does not repair measurement dropout. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Check the biological direction before inversion. 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. Bimodal assumption

A small example makes the mechanism visible. For each signature gene, a low mode represents nonsupported cells and a high mode represents supported types.

A concrete situation

A histogram shows two expression peaks.

What the mathematics does not license

Some genes are unimodal or have more than two states. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Plot distributions by validated 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.

6. Latent categorical identity

This is where modelling choices begin to affect evidence. Each cell has an unobserved category among the declared types and, when enabled, the unknown category.

A concrete situation

Cell j may be 0.75 type A and 0.20 type B.

What the mathematics does not license

Candidate categories bound the named outcomes. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Map the offered label space. 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. Complete-data likelihood

The central mathematical idea is concrete. The model links each latent category to the probability of the observed marker values under high or low modes.

A concrete situation

A cell high on two A markers fits A better.

What the mathematics does not license

Likelihood depends on distribution assumptions. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Calculate a two-marker toy likelihood. 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. Expectation step

Here the geometry becomes operational. Given current mode parameters and type proportions, EM computes posterior responsibility for every cell–type pair.

A concrete situation

Responsibilities across candidates sum to one.

What the mathematics does not license

A posterior can be sharp under a wrong model. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Preserve full responsibility 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.

9. Maximisation step

This step deserves a careful pause. Given responsibilities, EM updates type proportions and expression-distribution parameters to improve the expected objective.

A concrete situation

High-responsibility A cells pull A parameters.

What the mathematics does not license

Small classes can have unstable estimates. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot parameters and objective by iteration. 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. Convergence and initialisation

The useful question is not only how to calculate, but what the calculation assumes. Expectation and maximisation alternate until the declared assignments or objective stabilise.

A concrete situation

Two starts reach the same fit.

What the mathematics does not license

EM can settle at a local optimum. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat seeds and compare solutions. 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. Overlapping signatures

A small example makes the mechanism visible. Related types may share markers, so evidence from the distinguishing genes and all other signatures matters.

A concrete situation

Two T-cell types share a lineage marker.

What the mathematics does not license

Near-identical sets create unavoidable ambiguity. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Measure signature overlap before fitting. 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. Unknown category

This is where modelling choices begin to affect evidence. Cells with low overall support for all supplied signatures can be designated unknown in the default mode.

A concrete situation

A stromal cell appears in an immune-only panel.

What the mathematics does not license

Unknown does not identify the missing type. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Cluster unknown cells only as follow-up. 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. Marker-noise tolerance

The central mathematical idea is concrete. The published evaluation perturbed prior signatures, illustrating why several markers and sensitivity analysis matter.

A concrete situation

One incorrect gene enters a five-gene set.

What the mathematics does not license

Correlated errors can still dominate. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Remove and swap markers systematically. 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. Worked four-cell example

Here the geometry becomes operational. Create two types, two markers and one unsupported cell; hand-calculate high/low evidence and one responsibility update.

A concrete situation

The unsupported cell is low on both markers.

What the mathematics does not license

Toy separation is cleaner than real data. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Repeat with one ambiguous 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.

15. Validation beyond random cells

This step deserves a careful pause. Whole donors, tissues and platforms must remain outside signature tuning and model fitting.

A concrete situation

A donor-held-out score falls below random-cell CV.

What the mathematics does not license

Shared batch structure inflates random splits. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report grouped external metrics. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

16. Governance and careers

The useful question is not only how to calculate, but what the calculation assumes. Version preprocessing, signatures, modes, responsibilities, convergence and unknown settings. This links mixture models, statistics and bioinformatics.

A concrete situation

Students can begin with two-mode mixtures.

What the mathematics does not license

An assignment proves no mechanism or diagnosis. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Write a semi-supervised annotation 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

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 SCINA audit starts with a preprocessed expression matrix and a declared set of marker genes for every candidate type. The algorithm treats marker expression as a two-mode pattern, combines that prior knowledge with an otherwise unsupervised estimation process and iteratively updates cell-type responsibilities and distribution parameters through expectation–maximisation.

Record the exact matrix transformation, marker source, positive and inverted negative markers, candidate types, unknown-cell switch, initial values, convergence trace, posterior assignments and cells with low support across every signature. The fitted answer is conditional on the offered marker sets and on the bimodal assumption; it is not an independent discovery of cellular identity.

Validate with donor-, tissue- and platform-held-out studies, corrupted marker lists, overlapping signatures, weak-marker cases and deliberately omitted types. Report assignment coverage, per-type precision and recall, posterior margins, unknown-cell recovery and stability across seeds. A useful semi-supervised method remains honest about uncertainty and reference coverage.

Frequently asked questions

What is SCINA?

SCINA, or Semi-supervised Category Identification and Assignment, combines user-supplied marker-gene signatures with an expectation–maximisation model to cluster and assign cells or samples. The paper also applies it to cytometry and bulk-expression data.

Why is SCINA semi-supervised?

The cell labels are not supplied, but prior knowledge states which markers should be high for which types. That weak supervision guides an otherwise unsupervised estimation process.

What data does SCINA expect?

It takes an expression matrix plus one marker set per candidate type. The paper states that users should perform suitable logarithmic transformation or normalisation first when needed.

How are negative markers handled?

A gene known to be low for a type can be sign-inverted to create a pseudo-marker whose high value supports that type.

What does bimodal expression mean here?

Each signature gene is assumed to have a lower mode outside its supported type and a higher mode in the type or types for which it is a marker.

Can SCINA leave a cell unknown?

Yes. In its default unknown-search mode, cells with overall low expression across the supplied signatures can be assigned to an unknown category; the software also provides a switch to turn this search off.

Primary source and fact-check note

Ze Zhang, Di Wu, Arthur L. Robinson et al., “SCINA: A Semi-Supervised Subtyping Algorithm of Single Cells and Bulk Samples,” Genes 10(7), 531 (2019), DOI 10.3390/genes10070531. 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

Signature-provenance audit

Record source tissue, species, evidence and date for every positive and inverted negative marker. A statistical fit cannot repair a false signature.

Bimodality audit

Plot each marker distribution and compare one- and two-component descriptions. Flag markers without credible high and low modes.

EM stability test

Repeat initial values and seeds. Compare objective traces, fitted modes, class proportions, posteriors and unknown outcomes.

Marker-corruption challenge

Remove, swap and invert markers in a held-out validation study. Report when redundancy protects the result and when it does not.

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.

Overlap audit

Calculate pairwise overlap among type signatures and identify cells whose posterior is driven by only one distinguishing marker.

Unknown-type challenge

Withhold a known type, keep unknown search on and measure unknown recall, false unknowns and absorption by related classes.

Domain holdout

Leave complete donors, platforms and tissues outside all signature tuning. Random cells share hidden context.

Final claim

Name preprocessing, marker version, bimodal fit, EM convergence and unknown setting. The result is conditional on signatures and model assumptions.

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 marker knowledge needs mathematics

Start with the mechanism: A marker list says what should be informative; a probability model explains how several imperfect markers combine across many cells. The worked case is concrete: Three markers support one immune type with different strengths. Expose the boundary: Markers can be tissue- or state-dependent. Complete this evidence step: Cite and version every signature. 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: Expression-matrix contract

Start with the mechanism: SCINA accepts rows of genes and columns of cells or samples after suitable preprocessing chosen by the analyst. The worked case is concrete: Counts are normalised and log-transformed before input. Expose the boundary: Mixing raw and transformed scales changes distribution shape. Complete this evidence step: Store units and transformations. 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: Positive marker sets

Start with the mechanism: For each candidate type, one or more genes are declared characteristically high relative to other types. The worked case is concrete: Marker A and B support type 1. Expose the boundary: One marker may support several related types. Complete this evidence step: Draw a gene-by-type signature matrix. 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: Inverting low markers

Start with the mechanism: A low-expression marker can be multiplied by minus one so that a larger pseudo-value supports the intended type. The worked case is concrete: Values 5, 6, 4 and 0 become −5, −6, −4 and 0. Expose the boundary: Sign inversion does not repair measurement dropout. Complete this evidence step: Check the biological direction before inversion. 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.

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: Bimodal assumption

Start with the mechanism: For each signature gene, a low mode represents nonsupported cells and a high mode represents supported types. The worked case is concrete: A histogram shows two expression peaks. Expose the boundary: Some genes are unimodal or have more than two states. Complete this evidence step: Plot distributions by validated 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 6: Latent categorical identity

Start with the mechanism: Each cell has an unobserved category among the declared types and, when enabled, the unknown category. The worked case is concrete: Cell j may be 0.75 type A and 0.20 type B. Expose the boundary: Candidate categories bound the named outcomes. Complete this evidence step: Map the offered label space. 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: Complete-data likelihood

Start with the mechanism: The model links each latent category to the probability of the observed marker values under high or low modes. The worked case is concrete: A cell high on two A markers fits A better. Expose the boundary: Likelihood depends on distribution assumptions. Complete this evidence step: Calculate a two-marker toy likelihood. 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: Expectation step

Start with the mechanism: Given current mode parameters and type proportions, EM computes posterior responsibility for every cell–type pair. The worked case is concrete: Responsibilities across candidates sum to one. Expose the boundary: A posterior can be sharp under a wrong model. Complete this evidence step: Preserve full responsibility vectors. 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.

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: Maximisation step

Start with the mechanism: Given responsibilities, EM updates type proportions and expression-distribution parameters to improve the expected objective. The worked case is concrete: High-responsibility A cells pull A parameters. Expose the boundary: Small classes can have unstable estimates. Complete this evidence step: Plot parameters and objective by iteration. 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: Convergence and initialisation

Start with the mechanism: Expectation and maximisation alternate until the declared assignments or objective stabilise. The worked case is concrete: Two starts reach the same fit. Expose the boundary: EM can settle at a local optimum. Complete this evidence step: Repeat seeds and compare solutions. 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: Overlapping signatures

Start with the mechanism: Related types may share markers, so evidence from the distinguishing genes and all other signatures matters. The worked case is concrete: Two T-cell types share a lineage marker. Expose the boundary: Near-identical sets create unavoidable ambiguity. Complete this evidence step: Measure signature overlap before fitting. 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: Unknown category

Start with the mechanism: Cells with low overall support for all supplied signatures can be designated unknown in the default mode. The worked case is concrete: A stromal cell appears in an immune-only panel. Expose the boundary: Unknown does not identify the missing type. Complete this evidence step: Cluster unknown cells only as follow-up. 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.

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: Marker-noise tolerance

Start with the mechanism: The published evaluation perturbed prior signatures, illustrating why several markers and sensitivity analysis matter. The worked case is concrete: One incorrect gene enters a five-gene set. Expose the boundary: Correlated errors can still dominate. Complete this evidence step: Remove and swap markers systematically. 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: Worked four-cell example

Start with the mechanism: Create two types, two markers and one unsupported cell; hand-calculate high/low evidence and one responsibility update. The worked case is concrete: The unsupported cell is low on both markers. Expose the boundary: Toy separation is cleaner than real data. Complete this evidence step: Repeat with one ambiguous 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 15: Validation beyond random cells

Start with the mechanism: Whole donors, tissues and platforms must remain outside signature tuning and model fitting. The worked case is concrete: A donor-held-out score falls below random-cell CV. Expose the boundary: Shared batch structure inflates random splits. Complete this evidence step: Report grouped external metrics. 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 preprocessing, signatures, modes, responsibilities, convergence and unknown settings. This links mixture models, statistics and bioinformatics. The worked case is concrete: Students can begin with two-mode mixtures. Expose the boundary: An assignment proves no mechanism or diagnosis. Complete this evidence step: Write a semi-supervised annotation 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.