WHY MATHEMATICS?

Why Mathematics? | SPIRAL for Spatial Data Integration: GraphSAGE, Domain Adaptation, Gradient Reversal and Cluster-Aware Optimal Transport

How SPIRAL learns batch-corrected spatial representations and then aligns shared tissue domains into a reference coordinate system without treating all spots as exchangeable.

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

Encode gene expression with spatial-neighbour aggregation
Disentangle batch noise from biology with adversarial learning
Reconstruct corrected expression and classify spatial domains
Align shared clusters with Gromov–Wasserstein transport

Chapter map

1. Expression inputs
2. Spatial neighbour graph
3. Positive and negative pairs
4. GraphSAGE aggregation
5. Low-dimensional embedding
6. Noise classifier
7. Biology discriminator
8. Gradient reversal
9. Expression decoder
10. Total objective
11. Corrected outputs
12. Joint clustering
13. Reference selection
14. Cluster-wise feature cost
15. Gromov–Wasserstein alignment
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. Expression inputs

Here the geometry becomes operational. Every spot has a normalized vector over shared genes.

A concrete situation

Two technologies retain 2,000 common variable genes.

What the mathematics does not license

Normalization may create technology-specific geometry. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat robust preprocessing. 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. Spatial neighbour graph

This step deserves a careful pause. Coordinates define edges within each sample.

A concrete situation

A kNN graph connects local tissue neighbours.

What the mathematics does not license

Density differs across platforms. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Compare k and radius graphs. 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 and negative pairs

The useful question is not only how to calculate, but what the calculation assumes. Random walks identify contextual positives and other nodes supply negatives.

A concrete situation

One-hop co-occurrence forms a positive pair.

What the mathematics does not license

Sampling changes contrastive pressure. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record seeds and pair counts. 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. GraphSAGE aggregation

A small example makes the mechanism visible. A node combines its features with sampled neighbours.

A concrete situation

The mean of neighbour vectors enters the update.

What the mathematics does not license

Averages blur sharp boundaries. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Ablate neighbourhood depth. 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. Low-dimensional embedding

This is where modelling choices begin to affect evidence. Encoder layers compress expression and context.

A concrete situation

A 512-dimensional layer maps to 32 dimensions.

What the mathematics does not license

Compression can lose rare programs. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Test held-out markers. 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. Noise classifier

The central mathematical idea is concrete. A small embedding component predicts sample or batch identity.

A concrete situation

Four dimensions encode batch-discriminating variation.

What the mathematics does not license

Biology may leak into the noise part. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Measure biological content there. 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. Biology discriminator

Here the geometry becomes operational. Another component is challenged to hide batch identity.

A concrete situation

A classifier attempts to recover sample ID.

What the mathematics does not license

Confounded conditions make invariance unsafe. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Use balanced experimental designs. 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. Gradient reversal

This step deserves a careful pause. The encoder receives the negative discriminator gradient.

A concrete situation

The discriminator improves while the encoder removes its clues.

What the mathematics does not license

Minimax training can oscillate. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Track both losses. 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. Expression decoder

The useful question is not only how to calculate, but what the calculation assumes. Latent variables reconstruct high-dimensional expression.

A concrete situation

A 32-dimensional code returns a 512-dimensional profile.

What the mathematics does not license

Reconstruction favors abundant genes. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Weight and inspect rare markers. 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. Total objective

A small example makes the mechanism visible. Reconstruction, graph, classifier and adversarial terms combine.

A concrete situation

Lambda and gamma set equal weights in one documented setup.

What the mathematics does not license

Weights define the biological compromise. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Publish sensitivity curves. 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. Corrected outputs

This is where modelling choices begin to affect evidence. The model returns embeddings and reconstructed expression with reduced batch signal.

A concrete situation

Slices mix by tissue domain rather than platform.

What the mathematics does not license

Mixing metrics can reward overcorrection. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Retain known condition effects. 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. Joint clustering

The central mathematical idea is concrete. Corrected representations produce spatial-domain labels.

A concrete situation

Homologous cortex layers share labels.

What the mathematics does not license

Cluster resolution controls alignment. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Repeat several resolutions. 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. Reference selection

Here the geometry becomes operational. One sample supplies the coordinate system.

A concrete situation

The most complete slice becomes reference.

What the mathematics does not license

Reference-specific distortion propagates. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Try alternative references. 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. Cluster-wise feature cost

This step deserves a careful pause. Biological embeddings compare spots only within shared domains.

A concrete situation

Cluster c has its own cross-sample cost matrix.

What the mathematics does not license

Wrong labels block valid matches. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Audit boundary and novel clusters. 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. Gromov–Wasserstein alignment

The useful question is not only how to calculate, but what the calculation assumes. Feature cost and within-cluster spatial distances combine through alpha.

A concrete situation

A coupling preserves local geometry while matching biology.

What the mathematics does not license

Uniform mass and full overlap may be wrong. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Inspect transport entropy and unmatched structure. 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

A small example makes the mechanism visible. Version graphs, samples, losses, clusters, alpha and reference.

A concrete situation

A student traces one gradient-reversal update.

What the mathematics does not license

Batch invariance is conditional. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Write an integration audit card. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

A seven-day practice lab

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

SPIRAL links two tasks that are often handled separately. Its integration module learns a spatially aware representation with GraphSAGE, a decoder, a noise classifier and a biology discriminator. Neighbour aggregation lets expression be interpreted with local spatial context rather than as isolated rows of a matrix.

A gradient-reversal layer makes batch identity difficult to predict from the biological embedding while a separate noise component retains information useful for identifying batch-specific variation. Reconstruction and classification losses compete with domain confusion, so loss weights and sampling rules define what the model preserves.

Its alignment module operates cluster by cluster. For shared clusters, a Gromov–Wasserstein objective balances biological-embedding cost with preservation of within-sample spatial distances. This reduces cross-domain mixing, but cluster labels, reference choice, alpha and partial biological overlap remain consequential assumptions.

A useful classroom calculation starts with four spots in each sample. Students can compute a mean neighbour aggregation, pass a two-component biological embedding through a simple batch classifier, reverse that gradient for the encoder, and then compare two possible cluster-wise transport plans. The exercise makes the trade-off concrete: stronger batch confusion may improve mixing while weakening a condition-specific signal, and a lower transport cost may still depend on a mistaken cluster label. A responsible report therefore shows retained biology, discriminator performance, corrected-expression residuals, transport entropy and reference sensitivity together.

Frequently asked questions

What is SPIRAL?

SPIRAL combines graph domain adaptation for data integration with cluster-aware optimal transport for coordinate alignment.

What does GraphSAGE contribute?

It samples and aggregates spatial neighbours so a spot representation includes expression and local context.

What is gradient reversal?

It passes an embedding forward unchanged but multiplies its backward gradient by minus one, encouraging the encoder to confuse the batch discriminator.

Why use separate noise and biology components?

The architecture tries to retain batch-identifying variation in one part while making another part useful for shared biology.

What is cluster-aware alignment?

Gromov–Wasserstein transport is solved for shared clusters so distant biological domains are less likely to mix.

Can batch correction erase real biology?

Yes. If condition and batch are confounded, adversarial invariance can remove the signal of interest. Balanced design and holdouts are essential.

Primary source and fact-check note

Tiantian Guo, Zhiyuan Yuan, Yan Pan et al., “SPIRAL: integrating and aligning spatially resolved transcriptomics data across different experiments, conditions, and technologies,” Genome Biology 24, 241 (2023), DOI 10.1186/s13059-023-03078-6. 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

Design audit

Check whether batch and biology are confounded.

Graph audit

Compare neighbour definitions and sampling.

Loss audit

Publish reconstruction and adversarial weight sensitivity.

Discriminator audit

Track batch-prediction accuracy and convergence.

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.

Expression audit

Verify retained spatial and condition-specific signals.

Cluster audit

Repeat labels before transport.

Transport audit

Inspect alpha, coupling entropy and reference choice.

Final claim

Batch removal must preserve the question of interest.

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: Expression inputs

Begin with this mechanism: Every spot has a normalized vector over shared genes. Work through the concrete case: Two technologies retain 2,000 common variable genes. Make the limitation visible: Normalization may create technology-specific geometry. Complete the evidence step: Repeat robust preprocessing. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Spatial neighbour graph

Begin with this mechanism: Coordinates define edges within each sample. Work through the concrete case: A kNN graph connects local tissue neighbours. Make the limitation visible: Density differs across platforms. Complete the evidence step: Compare k and radius graphs. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Positive and negative pairs

Begin with this mechanism: Random walks identify contextual positives and other nodes supply negatives. Work through the concrete case: One-hop co-occurrence forms a positive pair. Make the limitation visible: Sampling changes contrastive pressure. Complete the evidence step: Record seeds and pair counts. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: GraphSAGE aggregation

Begin with this mechanism: A node combines its features with sampled neighbours. Work through the concrete case: The mean of neighbour vectors enters the update. Make the limitation visible: Averages blur sharp boundaries. Complete the evidence step: Ablate neighbourhood depth. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Low-dimensional embedding

Begin with this mechanism: Encoder layers compress expression and context. Work through the concrete case: A 512-dimensional layer maps to 32 dimensions. Make the limitation visible: Compression can lose rare programs. Complete the evidence step: Test held-out markers. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Noise classifier

Begin with this mechanism: A small embedding component predicts sample or batch identity. Work through the concrete case: Four dimensions encode batch-discriminating variation. Make the limitation visible: Biology may leak into the noise part. Complete the evidence step: Measure biological content there. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Biology discriminator

Begin with this mechanism: Another component is challenged to hide batch identity. Work through the concrete case: A classifier attempts to recover sample ID. Make the limitation visible: Confounded conditions make invariance unsafe. Complete the evidence step: Use balanced experimental designs. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Gradient reversal

Begin with this mechanism: The encoder receives the negative discriminator gradient. Work through the concrete case: The discriminator improves while the encoder removes its clues. Make the limitation visible: Minimax training can oscillate. Complete the evidence step: Track both losses. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Expression decoder

Begin with this mechanism: Latent variables reconstruct high-dimensional expression. Work through the concrete case: A 32-dimensional code returns a 512-dimensional profile. Make the limitation visible: Reconstruction favors abundant genes. Complete the evidence step: Weight and inspect rare markers. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Total objective

Begin with this mechanism: Reconstruction, graph, classifier and adversarial terms combine. Work through the concrete case: Lambda and gamma set equal weights in one documented setup. Make the limitation visible: Weights define the biological compromise. Complete the evidence step: Publish sensitivity curves. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Corrected outputs

Begin with this mechanism: The model returns embeddings and reconstructed expression with reduced batch signal. Work through the concrete case: Slices mix by tissue domain rather than platform. Make the limitation visible: Mixing metrics can reward overcorrection. Complete the evidence step: Retain known condition effects. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Joint clustering

Begin with this mechanism: Corrected representations produce spatial-domain labels. Work through the concrete case: Homologous cortex layers share labels. Make the limitation visible: Cluster resolution controls alignment. Complete the evidence step: Repeat several resolutions. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Reference selection

Begin with this mechanism: One sample supplies the coordinate system. Work through the concrete case: The most complete slice becomes reference. Make the limitation visible: Reference-specific distortion propagates. Complete the evidence step: Try alternative references. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Cluster-wise feature cost

Begin with this mechanism: Biological embeddings compare spots only within shared domains. Work through the concrete case: Cluster c has its own cross-sample cost matrix. Make the limitation visible: Wrong labels block valid matches. Complete the evidence step: Audit boundary and novel clusters. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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: Gromov–Wasserstein alignment

Begin with this mechanism: Feature cost and within-cluster spatial distances combine through alpha. Work through the concrete case: A coupling preserves local geometry while matching biology. Make the limitation visible: Uniform mass and full overlap may be wrong. Complete the evidence step: Inspect transport entropy and unmatched structure. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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 graphs, samples, losses, clusters, alpha and reference. Work through the concrete case: A student traces one gradient-reversal update. Make the limitation visible: Batch invariance is conditional. Complete the evidence step: Write an integration audit card. Write the coordinates, graphs, matrices, costs, weights, transformations, losses, couplings or displacement fields with named dimensions and physical units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve complete sections, donors, platforms, regions, genes or landmarks 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.