WHY MATHEMATICS?

Why Mathematics? | Hotspot for Informative Gene Modules: Neighbour Graphs, Local Autocorrelation, Pairwise Local Correlation and Module Scores

How Hotspot finds genes whose variation follows a chosen cell geometry, then groups locally correlated genes into interpretable modules without claiming automatic regulation.

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

Choose the cell-cell geometry that defines local similarity
Model gene expectations under an appropriate observation model
Rank informative genes by local autocorrelation
Build modules from pairwise local correlations and score them

Chapter map

1. Cell metric
2. Neighbour graph
3. Edge weights
4. UMI depth
5. Depth-adjusted negative binomial
6. Alternative observation models
7. Standardized residuals
8. Local autocorrelation numerator
9. Null variance and Z-score
10. FDR selection
11. Pairwise local correlation
12. Gene correlation matrix
13. Hierarchical clustering
14. Module assignment
15. Module scores
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. Cell metric

This step deserves a careful pause. A distance or graph encodes the notion of cell similarity.

A concrete situation

Physical coordinates define spatial neighbours.

What the mathematics does not license

Different metrics answer different questions. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Name the geometry. 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. Neighbour graph

The useful question is not only how to calculate, but what the calculation assumes. Each cell connects to a limited local set.

A concrete situation

Thirty nearby cells receive edges.

What the mathematics does not license

k changes physical scale. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot edge 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.

3. Edge weights

A small example makes the mechanism visible. Closer or more similar cells receive larger influence.

A concrete situation

A distance kernel downweights remote neighbours.

What the mathematics does not license

Weight decay sets locality. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Publish the formula. 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. UMI depth

This is where modelling choices begin to affect evidence. Total counts affect expected expression and variance.

A concrete situation

A high-depth cell has more expected counts.

What the mathematics does not license

Depth can correlate with biology. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Map depth against modules. 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. Depth-adjusted negative binomial

The central mathematical idea is concrete. Counts receive gene- and cell-specific expectations under a dispersed model.

A concrete situation

One gene has fitted mean and dispersion.

What the mathematics does not license

Model misfit distorts Z-scores. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use residual checks. 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. Alternative observation models

Here the geometry becomes operational. Bernoulli, normal or pre-standardized inputs support other modalities.

A concrete situation

Protein detection uses a different null.

What the mathematics does not license

Model choice changes meaning. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Justify the option. 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. Standardized residuals

This step deserves a careful pause. Observed expression is converted to a comparable deviation from expectation.

A concrete situation

A cell is two standard units above mean.

What the mathematics does not license

Extreme residuals can dominate. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Check influence. 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. Local autocorrelation numerator

The useful question is not only how to calculate, but what the calculation assumes. Weighted products of neighbouring residuals summarize agreement.

A concrete situation

High-high and low-low neighbours add positive evidence.

What the mathematics does not license

A global gradient can dominate local structure. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect 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.

9. Null variance and Z-score

A small example makes the mechanism visible. Weight geometry and observation variance scale the statistic.

A concrete situation

One gene receives an autocorrelation Z of four.

What the mathematics does not license

Normal approximation needs calibration. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use null simulations. 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. FDR selection

This is where modelling choices begin to affect evidence. Autocorrelation p-values are corrected across genes.

A concrete situation

Only significant genes enter pair analysis.

What the mathematics does not license

Thresholding can destabilize borderline modules. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Show nearby cutoffs. 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. Pairwise local correlation

The central mathematical idea is concrete. Two genes are compared across neighbouring cells as well as within cells.

A concrete situation

Gene A near gene B yields positive local association.

What the mathematics does not license

Proximity does not imply interaction. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Avoid signalling claims. 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. Gene correlation matrix

Here the geometry becomes operational. All selected pairs form a symmetric Z-score matrix.

A concrete situation

Blocks of positive local correlation emerge.

What the mathematics does not license

Cost grows with selected genes. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record selection size. 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. Hierarchical clustering

This step deserves a careful pause. Gene distances derived from local association are merged into a tree.

A concrete situation

A branch contains twenty related genes.

What the mathematics does not license

Cut height changes modules. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Publish the dendrogram. 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. Module assignment

The useful question is not only how to calculate, but what the calculation assumes. Minimum size and ambiguity rules determine final membership.

A concrete situation

Weak genes remain unassigned.

What the mathematics does not license

Forced assignment inflates coherence. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Keep ambiguous genes visible. 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. Module scores

A small example makes the mechanism visible. Combined module expression is summarized for every cell.

A concrete situation

A spatial map shows one programme enriched locally.

What the mathematics does not license

Scores depend on normalization and members. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Validate held-out genes. 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

This is where modelling choices begin to affect evidence. Version graph, model, k, weights, FDR and module cut rules.

A concrete situation

A student calculates one weighted autocorrelation.

What the mathematics does not license

Local correlation is not causation. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write a Hotspot 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.

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

Hotspot begins with a similarity graph that may represent physical proximity, a transcriptional manifold or a lineage tree. For each gene, it asks whether standardized deviations agree more among strongly connected cells than expected under a gene-specific null model. This local autocorrelation identifies informative variation aligned with the chosen geometry.

A depth-adjusted negative-binomial model can account for UMI counts, while Bernoulli, normal and pre-standardized options support other measurements. Significant genes are then compared in pairs using local correlations, which capture a gene expressed in one cell near another gene as well as ordinary within-cell association.

Hierarchical clustering of the gene-by-gene local-correlation matrix produces modules, and module scores summarize their activity across cells. The graph defines the meaning of local: a spatial graph yields spatial modules, while a latent-space graph yields transcriptional modules. Modules are statistical co-variation patterns, not proof of one regulatory pathway.

Students can build a five-cell weighted graph, standardize one gene, multiply neighbour deviations by edge weights and form a local-autocorrelation numerator. Repeating for two genes gives a local correlation, after which a tiny gene distance matrix can be clustered into a module.

Frequently asked questions

What is Hotspot?

Hotspot is a graph-based method for identifying informative genes and groups of genes whose variation follows a chosen cell-cell similarity geometry.

What can define the graph?

Physical proximity, a low-dimensional transcriptional embedding, a user-supplied distance matrix or a lineage tree can define cell similarity.

What is local autocorrelation?

It measures whether standardized deviations for one gene tend to agree across heavily weighted neighbouring cell pairs.

Why use a gene observation model?

Depth-adjusted negative-binomial, Bernoulli or normal models provide per-cell expectations and variances under a no-local-pattern null.

How are gene modules formed?

Significant genes receive pairwise local-correlation Z-scores, which are converted into gene distances and hierarchically clustered.

Does a module equal a regulatory pathway?

No. It is a local co-variation pattern whose interpretation requires perturbation, prior biology and independent replication.

Primary source and fact-check note

David DeTomaso and Nir Yosef, “Hotspot identifies informative gene modules across modalities of single-cell genomics,” Cell Systems 12, 446–456.e9 (2021), DOI 10.1016/j.cels.2021.04.005. 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

Geometry audit

Name physical, latent or lineage meaning.

Graph audit

Sweep k and edge weights.

Observation audit

Check negative-binomial or alternative residuals.

Autocorrelation audit

Simulate null Z-scores.

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.

Gene audit

Show FDR and selection stability.

Module audit

Publish dendrogram and cut sensitivity.

Score audit

Validate module maps on held-out genes.

Final claim

Local co-variation is not regulation.

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: Cell metric

Begin with this mechanism: A distance or graph encodes the notion of cell similarity. Work through the concrete case: Physical coordinates define spatial neighbours. Make the limitation visible: Different metrics answer different questions. Complete the evidence step: Name the geometry. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Neighbour graph

Begin with this mechanism: Each cell connects to a limited local set. Work through the concrete case: Thirty nearby cells receive edges. Make the limitation visible: k changes physical scale. Complete the evidence step: Plot edge lengths. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Edge weights

Begin with this mechanism: Closer or more similar cells receive larger influence. Work through the concrete case: A distance kernel downweights remote neighbours. Make the limitation visible: Weight decay sets locality. Complete the evidence step: Publish the formula. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: UMI depth

Begin with this mechanism: Total counts affect expected expression and variance. Work through the concrete case: A high-depth cell has more expected counts. Make the limitation visible: Depth can correlate with biology. Complete the evidence step: Map depth against modules. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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.

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: Depth-adjusted negative binomial

Begin with this mechanism: Counts receive gene- and cell-specific expectations under a dispersed model. Work through the concrete case: One gene has fitted mean and dispersion. Make the limitation visible: Model misfit distorts Z-scores. Complete the evidence step: Use residual checks. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Alternative observation models

Begin with this mechanism: Bernoulli, normal or pre-standardized inputs support other modalities. Work through the concrete case: Protein detection uses a different null. Make the limitation visible: Model choice changes meaning. Complete the evidence step: Justify the option. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Standardized residuals

Begin with this mechanism: Observed expression is converted to a comparable deviation from expectation. Work through the concrete case: A cell is two standard units above mean. Make the limitation visible: Extreme residuals can dominate. Complete the evidence step: Check influence. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Local autocorrelation numerator

Begin with this mechanism: Weighted products of neighbouring residuals summarize agreement. Work through the concrete case: High-high and low-low neighbours add positive evidence. Make the limitation visible: A global gradient can dominate local structure. Complete the evidence step: Inspect maps. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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.

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: Null variance and Z-score

Begin with this mechanism: Weight geometry and observation variance scale the statistic. Work through the concrete case: One gene receives an autocorrelation Z of four. Make the limitation visible: Normal approximation needs calibration. Complete the evidence step: Use null simulations. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: FDR selection

Begin with this mechanism: Autocorrelation p-values are corrected across genes. Work through the concrete case: Only significant genes enter pair analysis. Make the limitation visible: Thresholding can destabilize borderline modules. Complete the evidence step: Show nearby cutoffs. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Pairwise local correlation

Begin with this mechanism: Two genes are compared across neighbouring cells as well as within cells. Work through the concrete case: Gene A near gene B yields positive local association. Make the limitation visible: Proximity does not imply interaction. Complete the evidence step: Avoid signalling claims. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Gene correlation matrix

Begin with this mechanism: All selected pairs form a symmetric Z-score matrix. Work through the concrete case: Blocks of positive local correlation emerge. Make the limitation visible: Cost grows with selected genes. Complete the evidence step: Record selection size. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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.

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: Hierarchical clustering

Begin with this mechanism: Gene distances derived from local association are merged into a tree. Work through the concrete case: A branch contains twenty related genes. Make the limitation visible: Cut height changes modules. Complete the evidence step: Publish the dendrogram. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Module assignment

Begin with this mechanism: Minimum size and ambiguity rules determine final membership. Work through the concrete case: Weak genes remain unassigned. Make the limitation visible: Forced assignment inflates coherence. Complete the evidence step: Keep ambiguous genes visible. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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: Module scores

Begin with this mechanism: Combined module expression is summarized for every cell. Work through the concrete case: A spatial map shows one programme enriched locally. Make the limitation visible: Scores depend on normalization and members. Complete the evidence step: Validate held-out genes. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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 graph, model, k, weights, FDR and module cut rules. Work through the concrete case: A student calculates one weighted autocorrelation. Make the limitation visible: Local correlation is not causation. Complete the evidence step: Write a Hotspot model card. Write the coordinates, pair distances, graph weights, node groups, kernels, covariance matrices, eigenvalues, Fourier coefficients, permutation statistics, module correlations or reconstruction errors with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete tissues, donors, genes, fields of view or regions 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.