Chapter map

1. 1. Low-dimensional input
2. 2. Gaussian cluster model
3. 3. Hidden labels
4. 4. Neighbour graph
5. 5. HMRF energy
6. 6. Smoothness parameter
7. 7. Empirical-Bayes estimation
8. 8. Expectation step
9. 9. Maximization step
10. 10. Iterative conditional modes
11. 11. Cluster-count candidates
12. 12. Modified BIC
13. 13. Spatial boundaries
14. 14. Scalability
15. 15. Biological validation
16. 16. Governance and learning

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

---

1. Low-dimensional input

The central mathematical idea is concrete. A reduced feature vector represents each spatial location.

A concrete situation

Ten components summarize normalized expression.

What the mathematics does not license

Feature selection can hide rare programs. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat several gene and rank 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.

---

2. Gaussian cluster model

Here the geometry becomes operational. Each latent cluster has a mean and covariance in feature space.

A concrete situation

Two tissue regions form overlapping ellipses.

What the mathematics does not license

Non-Gaussian groups may split or merge. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Inspect posterior residuals. 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. Hidden labels

This step deserves a careful pause. Every spot has an unobserved categorical domain label.

A concrete situation

A boundary spot has competing cluster scores.

What the mathematics does not license

Hard maps conceal ambiguity. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report local evidence margins. 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. Neighbour graph

The useful question is not only how to calculate, but what the calculation assumes. Coordinates define which label pairs interact.

A concrete situation

A Visium lattice supplies adjacent spots.

What the mathematics does not license

Missing spots change node degrees. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Map all edges and lengths. 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. HMRF energy

A small example makes the mechanism visible. Molecular likelihood and neighbour agreement form a joint objective.

A concrete situation

One label fits expression but disagrees with neighbours.

What the mathematics does not license

Terms use different scales. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Show both contributions. 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. Smoothness parameter

This is where modelling choices begin to affect evidence. A coefficient sets the reward for equal neighbouring labels.

A concrete situation

Larger smoothness removes isolated assignments.

What the mathematics does not license

Too much smoothness erases thin regions. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot maps across a grid. 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. Empirical-Bayes estimation

The central mathematical idea is concrete. Unknown parameters are fitted from the observed section.

A concrete situation

Means, variances and smoothness update iteratively.

What the mathematics does not license

Data reuse can overstate certainty. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use held-out regions. 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. Conditional expectations summarize hidden quantities under current parameters.

A concrete situation

Responsibilities change after a mean update.

What the mathematics does not license

Spatial dependence complicates exact inference. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

State every approximation. 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. Maximization step

This step deserves a careful pause. Parameters update to improve the expected objective.

A concrete situation

Cluster centres move toward weighted features.

What the mathematics does not license

Degenerate covariance can inflate fit. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Regularize and check conditioning. 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. Iterative conditional modes

The useful question is not only how to calculate, but what the calculation assumes. Each label changes to its locally best conditional state.

A concrete situation

A chain spot switches after its neighbours agree.

What the mathematics does not license

ICM can stop at a local mode. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Run diverse initializations. 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. Cluster-count candidates

A small example makes the mechanism visible. Separate fits compare plausible values of K.

A concrete situation

Models with four through eight domains are evaluated.

What the mathematics does not license

The candidate range constrains selection. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Include neighbouring K maps. 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. Modified BIC

This is where modelling choices begin to affect evidence. A penalty balances fit improvement against added complexity.

A concrete situation

An extra cluster must justify more parameters.

What the mathematics does not license

Penalty formulas rely on assumptions. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report raw fit and penalty. 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. Spatial boundaries

The central mathematical idea is concrete. A boundary is where molecular evidence outweighs the continuity reward.

A concrete situation

A cortical layer remains narrow at moderate smoothness.

What the mathematics does not license

Sampling gaps can mimic boundaries. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Overlay morphology and missingness. 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. Scalability

Here the geometry becomes operational. EM-ICM avoids costly full posterior sampling.

A concrete situation

A large spatial dataset fits in practical time.

What the mathematics does not license

Fast convergence is not global optimality. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Benchmark restarts and memory. 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. Biological validation

This step deserves a careful pause. Markers and anatomy test whether clusters are useful.

A concrete situation

Layer genes agree with a domain boundary.

What the mathematics does not license

Circular marker selection exaggerates agreement. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use independent annotations. 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 learning

The useful question is not only how to calculate, but what the calculation assumes. Version features, graph, K range, smoothness, starts and stopping rules.

A concrete situation

A student completes one EM-ICM update.

What the mathematics does not license

A selected model remains conditional. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Write an SC-MEB evidence ledger. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

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

Questions for a student, parent or mentor

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

A method-specific audit before deployment

SC-MEB joins a low-dimensional Gaussian mixture to a hidden Markov random field. The mixture asks which cluster best explains each spot in feature space; the spatial field asks whether neighbouring labels form a plausible tissue map. An empirical-Bayes fit estimates parameters from the observed section instead of fixing every smoothness choice in advance.

Its expectation-maximization procedure alternates between latent quantities and model parameters. Iterative conditional modes updates each label to its locally best conditional value. That combination is computationally efficient, but a local optimum can depend on initialization. Multiple starts, objective traces and sensitivity checks remain part of responsible use.

The smoothness coefficient controls how much an agreeing edge contributes. SC-MEB estimates this coefficient and can compare candidate cluster counts using a modified Bayesian information criterion. A selected value is evidence under the fitted model, not proof that tissue has exactly that many biological compartments.

A small worked case uses six spots on a chain, two latent features and two candidate clusters. Calculate two Gaussian log scores, add the neighbour-agreement contribution, update one label by ICM and compare the penalized fit for two versus three clusters. The arithmetic makes the fit-smoothness trade-off visible.

---

Frequently asked questions

What is SC-MEB?

SC-MEB is an empirical-Bayes spatial-clustering method that combines low-dimensional expression features with a hidden Markov random field on neighbouring locations.

What does the spatial field contribute?

It rewards equal labels at adjacent locations according to a smoothness parameter, while molecular likelihood can still support a boundary.

Why use empirical Bayes?

Model parameters, including spatial smoothness, are estimated from the observed dataset rather than being fully fixed or sampled from a complete Bayesian hierarchy.

How is SC-MEB fitted?

An expectation-maximization algorithm updates model quantities, and iterative conditional modes supplies efficient local label updates.

How is the number of domains chosen?

Candidate values can be compared with a modified Bayesian information criterion, subject to the chosen features, graph and candidate range.

Is an inferred domain an anatomical fact?

No. It is a model-based partition that requires marker, morphology, replication and sensitivity validation.

---

Primary source and fact-check note

Yi Yang, Xingjie Shi, Wei Liu et al., “SC-MEB: spatial clustering with hidden Markov random field using empirical Bayes,” Briefings in Bioinformatics 23 (2022), bbab466, DOI 10.1093/bib/bbab466. 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

• DR-SC joint representation and clustering
• BASS Bayesian tissue hierarchy
• SpatialPCA spatial factors
• 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

Feature audit

Repeat gene sets and latent ranks.

Graph audit

Map degrees, components and physical edge lengths.

Smoothness audit

Show maps across fitted and neighbouring values.

Optimization audit

Use multiple EM-ICM starts.

Count audit

Compare modified-BIC neighbours.

Boundary audit

Overlay morphology and missing spots.

Validation audit

Use independent markers and annotations.

Final claim

A fitted domain is a conditional statistical partition.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Low-dimensional input

Begin with this mechanism: A reduced feature vector represents each spatial location. Work through the concrete case: Ten components summarize normalized expression. Make the limitation visible: Feature selection can hide rare programs. Complete the evidence step: Repeat several gene and rank choices. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 2: Gaussian cluster model

Begin with this mechanism: Each latent cluster has a mean and covariance in feature space. Work through the concrete case: Two tissue regions form overlapping ellipses. Make the limitation visible: Non-Gaussian groups may split or merge. Complete the evidence step: Inspect posterior residuals. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 3: Hidden labels

Begin with this mechanism: Every spot has an unobserved categorical domain label. Work through the concrete case: A boundary spot has competing cluster scores. Make the limitation visible: Hard maps conceal ambiguity. Complete the evidence step: Report local evidence margins. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 4: Neighbour graph

Begin with this mechanism: Coordinates define which label pairs interact. Work through the concrete case: A Visium lattice supplies adjacent spots. Make the limitation visible: Missing spots change node degrees. Complete the evidence step: Map all edges and lengths. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 5: HMRF energy

Begin with this mechanism: Molecular likelihood and neighbour agreement form a joint objective. Work through the concrete case: One label fits expression but disagrees with neighbours. Make the limitation visible: Terms use different scales. Complete the evidence step: Show both contributions. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 6: Smoothness parameter

Begin with this mechanism: A coefficient sets the reward for equal neighbouring labels. Work through the concrete case: Larger smoothness removes isolated assignments. Make the limitation visible: Too much smoothness erases thin regions. Complete the evidence step: Plot maps across a grid. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 7: Empirical-Bayes estimation

Begin with this mechanism: Unknown parameters are fitted from the observed section. Work through the concrete case: Means, variances and smoothness update iteratively. Make the limitation visible: Data reuse can overstate certainty. Complete the evidence step: Use held-out regions. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 8: Expectation step

Begin with this mechanism: Conditional expectations summarize hidden quantities under current parameters. Work through the concrete case: Responsibilities change after a mean update. Make the limitation visible: Spatial dependence complicates exact inference. Complete the evidence step: State every approximation. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 9: Maximization step

Begin with this mechanism: Parameters update to improve the expected objective. Work through the concrete case: Cluster centres move toward weighted features. Make the limitation visible: Degenerate covariance can inflate fit. Complete the evidence step: Regularize and check conditioning. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 10: Iterative conditional modes

Begin with this mechanism: Each label changes to its locally best conditional state. Work through the concrete case: A chain spot switches after its neighbours agree. Make the limitation visible: ICM can stop at a local mode. Complete the evidence step: Run diverse initializations. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 11: Cluster-count candidates

Begin with this mechanism: Separate fits compare plausible values of K. Work through the concrete case: Models with four through eight domains are evaluated. Make the limitation visible: The candidate range constrains selection. Complete the evidence step: Include neighbouring K maps. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 12: Modified BIC

Begin with this mechanism: A penalty balances fit improvement against added complexity. Work through the concrete case: An extra cluster must justify more parameters. Make the limitation visible: Penalty formulas rely on assumptions. Complete the evidence step: Report raw fit and penalty. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 13: Spatial boundaries

Begin with this mechanism: A boundary is where molecular evidence outweighs the continuity reward. Work through the concrete case: A cortical layer remains narrow at moderate smoothness. Make the limitation visible: Sampling gaps can mimic boundaries. Complete the evidence step: Overlay morphology and missingness. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 14: Scalability

Begin with this mechanism: EM-ICM avoids costly full posterior sampling. Work through the concrete case: A large spatial dataset fits in practical time. Make the limitation visible: Fast convergence is not global optimality. Complete the evidence step: Benchmark restarts and memory. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 15: Biological validation

Begin with this mechanism: Markers and anatomy test whether clusters are useful. Work through the concrete case: Layer genes agree with a domain boundary. Make the limitation visible: Circular marker selection exaggerates agreement. Complete the evidence step: Use independent annotations. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 16: Governance and learning

Begin with this mechanism: Version features, graph, K range, smoothness, starts and stopping rules. Work through the concrete case: A student completes one EM-ICM update. Make the limitation visible: A selected model remains conditional. Complete the evidence step: Write an SC-MEB evidence ledger. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.