WHY MATHEMATICS?

Why Mathematics? | STAligner for Multi-Slice Tissue Maps: Graph Attention, Mutual Nearest Neighbours, Triplet Loss and 3D Alignment

How STAligner iteratively learns spatial embeddings and cross-slice correspondences, integrates comparable tissue states and uses landmark-domain pairs for physical 3D reconstruction.

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

Preserve each slice through its own spatial neighbour graph
Learn spatially aware embeddings with graph attention reconstruction
Update cross-slice mutual-neighbour triplets iteratively
Guide physical slice alignment with validated landmark pairs

Chapter map

1. Slice-specific expression
2. Within-slice coordinates
3. Spatial neighbour networks
4. Block-diagonal graph
5. Graph attention encoder
6. Autoencoder decoder
7. Initial embedding
8. Cross-slice neighbour search
9. Mutual nearest neighbours
10. Negative sampling
11. Triplet distance
12. Margin loss
13. Iterative refinement
14. Shared spatial domains
15. ICP coordinate 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. Slice-specific expression

A small example makes the mechanism visible. Every section contributes normalized features with dataset labels retained for audit.

A concrete situation

Three serial slices each contain 2,000 spots.

What the mathematics does not license

Preprocessing differences mimic batch. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Version each slice separately. 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. Within-slice coordinates

This is where modelling choices begin to affect evidence. Locations remain in their own two-dimensional coordinate systems.

A concrete situation

Slice one and slice two have different origins.

What the mathematics does not license

Raw coordinates are not directly comparable. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record scale, orientation and units. 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. Spatial neighbour networks

The central mathematical idea is concrete. A graph is built inside each slice from local coordinates.

A concrete situation

Every interior spot connects to nearby peers.

What the mathematics does not license

Edges can cross tissue holes. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Overlay every slice graph. 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. Block-diagonal graph

Here the geometry becomes operational. Slice graphs form disconnected blocks before cross-slice alignment.

A concrete situation

No physical edge directly joins slice one to two.

What the mathematics does not license

Disconnected blocks rely on expression for initial comparability. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Check feature overlap. 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. Graph attention encoder

This step deserves a careful pause. Adaptive neighbour weights transform expression into spatially aware codes.

A concrete situation

A boundary node emphasizes same-domain neighbours.

What the mathematics does not license

Attention can learn batch-specific cues. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Inspect weights by slice. 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. Autoencoder decoder

The useful question is not only how to calculate, but what the calculation assumes. Latent codes reconstruct the normalized input features.

A concrete situation

The decoder predicts the original feature vector.

What the mathematics does not license

Low error can preserve technical noise. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Audit residuals by batch. 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. Initial embedding

A small example makes the mechanism visible. Pretraining supplies a shared coordinate system for provisional correspondence search.

A concrete situation

Analogous cortical layers approach one another.

What the mathematics does not license

Initial geometry can contain batch separation. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare without alignment. 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. Cross-slice neighbour search

This is where modelling choices begin to affect evidence. Distances in embedding space identify candidate matches between different slices.

A concrete situation

A spot in layer four searches slice two neighbours.

What the mathematics does not license

Missing states force bad matches. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Allow unmatched spots. 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. Mutual nearest neighbours

The central mathematical idea is concrete. Reciprocal nearest relationships produce anchor-positive candidates.

A concrete situation

Spot a chooses b and b chooses a.

What the mathematics does not license

Mutuality does not ensure biological identity. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Validate markers and domains. 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. Negative sampling

Here the geometry becomes operational. A dissimilar spot supplies the third member of each triplet.

A concrete situation

The negative comes from a distinct location and expression state.

What the mathematics does not license

False negatives push related biology apart. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect negative rules. 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. Triplet distance

This step deserves a careful pause. Euclidean distances compare anchor-positive and anchor-negative embeddings.

A concrete situation

d(a,p)=0.4 and d(a,n)=1.2.

What the mathematics does not license

Scale affects margin meaning. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Standardize or document norms. 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. Margin loss

The useful question is not only how to calculate, but what the calculation assumes. The hinge term max(dap minus dan plus margin, zero) enforces separation.

A concrete situation

With margin 1, the example loss is 0.2.

What the mathematics does not license

Too large a margin distorts geometry. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Sweep the margin. 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. Iterative refinement

A small example makes the mechanism visible. Training alternates with rebuilding triplets from current embeddings.

A concrete situation

Pairs improve after several update rounds.

What the mathematics does not license

Wrong early pairs can self-reinforce. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Track pair turnover and precision. 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. Shared spatial domains

This is where modelling choices begin to affect evidence. Clustering the corrected embedding identifies comparable and condition-specific regions.

A concrete situation

Normal and diseased hippocampus share broad anatomy but retain a disease substructure.

What the mathematics does not license

Over-correction erases the difference. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report conservation and batch removal. 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. ICP coordinate alignment

The central mathematical idea is concrete. MNN pairs within chosen landmark domains guide rigid or permitted coordinate transformations.

A concrete situation

A moving slice rotates and translates toward a reference.

What the mathematics does not license

Sparse or uneven anchors misregister tissue. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Report overlap and landmark 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.

16. Governance and learning

Here the geometry becomes operational. Version graphs, embeddings, MNN rules, negatives, margin, iterations and landmarks.

A concrete situation

A student traces one pair from graph code to 3D coordinate.

What the mathematics does not license

Alignment is a model-supported correspondence. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Write a multi-slice integration 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

STAligner integrates several spatial transcriptomics slices with a graph attention autoencoder and iterative triplet learning. Each slice first keeps its own spatial neighbour graph. The autoencoder learns a spatially aware embedding, then mutual nearest neighbours across slices become candidate anchor-positive pairs. A dissimilar spot supplies the negative example, and a margin-based triplet loss pulls the positive closer while pushing the negative away.

Triplets are rebuilt as the embedding improves, so training alternates between learning geometry and updating cross-slice correspondences. This supports joint spatial domains across samples, technologies, conditions and developmental stages. For physical 3D reconstruction, mutual-nearest-neighbour pairs within selected corresponding domains can guide an iterative closest point transformation of slice coordinates.

The method depends on the quality and coverage of cross-slice correspondences. Missing anatomy, different cell states or weak overlap can create too few reliable mutual neighbours. Validate anchor precision, hold out slices, report biological conservation and batch removal, inspect disease-specific structure and test 3D landmark choices. Alignment should reveal comparable structure without forcing every slice to look the same.

Frequently asked questions

What is STAligner?

STAligner is a graph-attention method for integrating multiple spatial transcriptomics datasets while retaining spatial context and identifying shared domains.

Why keep a graph for each slice?

Within-slice graphs encode local tissue relationships without inventing physical edges between sections. Cross-slice relationships are learned later from embeddings.

What is a mutual nearest neighbour?

Two spots in different slices form an MNN pair when each is among the other’s nearest candidates in the shared embedding under the selected rule.

What is triplet loss?

It compares an anchor, a positive and a negative, penalizing cases where the positive is not closer than the negative by a chosen margin.

Why update triplets iteratively?

The first embedding gives provisional pairs. Recomputing pairs after training can refine correspondences, though early mistakes may also reinforce themselves.

How does 3D alignment work?

Within selected corresponding landmark domains, cross-slice MNN pairs can guide an iterative closest point coordinate transformation. The result depends on overlap and anchor coverage.

Primary source and fact-check note

Xiang Zhou, Kangning Dong and Shihua Zhang, “Integrating spatial transcriptomics data across different conditions, technologies and developmental stages,” Nature Computational Science 3, 894–906 (2023), DOI 10.1038/s43588-023-00528-w. 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

Slice audit

Keep preprocessing and quality metrics per section.

Graph audit

Inspect every within-slice network.

MNN audit

Report coverage, reciprocity and marker agreement.

Negative audit

Check false-negative rates.

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.

Loss audit

Sweep margin and triplet weight.

Integration audit

Report conservation and batch removal.

3D audit

Check landmark coverage and ICP residuals.

Final claim

A correspondence is conditional on overlap and representation.

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: Slice-specific expression

Begin with this mechanism: Every section contributes normalized features with dataset labels retained for audit. Work through the concrete case: Three serial slices each contain 2,000 spots. Make the limitation visible: Preprocessing differences mimic batch. Complete the evidence step: Version each slice separately. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Within-slice coordinates

Begin with this mechanism: Locations remain in their own two-dimensional coordinate systems. Work through the concrete case: Slice one and slice two have different origins. Make the limitation visible: Raw coordinates are not directly comparable. Complete the evidence step: Record scale, orientation and units. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Spatial neighbour networks

Begin with this mechanism: A graph is built inside each slice from local coordinates. Work through the concrete case: Every interior spot connects to nearby peers. Make the limitation visible: Edges can cross tissue holes. Complete the evidence step: Overlay every slice graph. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Block-diagonal graph

Begin with this mechanism: Slice graphs form disconnected blocks before cross-slice alignment. Work through the concrete case: No physical edge directly joins slice one to two. Make the limitation visible: Disconnected blocks rely on expression for initial comparability. Complete the evidence step: Check feature overlap. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Graph attention encoder

Begin with this mechanism: Adaptive neighbour weights transform expression into spatially aware codes. Work through the concrete case: A boundary node emphasizes same-domain neighbours. Make the limitation visible: Attention can learn batch-specific cues. Complete the evidence step: Inspect weights by slice. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Autoencoder decoder

Begin with this mechanism: Latent codes reconstruct the normalized input features. Work through the concrete case: The decoder predicts the original feature vector. Make the limitation visible: Low error can preserve technical noise. Complete the evidence step: Audit residuals by batch. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Initial embedding

Begin with this mechanism: Pretraining supplies a shared coordinate system for provisional correspondence search. Work through the concrete case: Analogous cortical layers approach one another. Make the limitation visible: Initial geometry can contain batch separation. Complete the evidence step: Compare without alignment. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Cross-slice neighbour search

Begin with this mechanism: Distances in embedding space identify candidate matches between different slices. Work through the concrete case: A spot in layer four searches slice two neighbours. Make the limitation visible: Missing states force bad matches. Complete the evidence step: Allow unmatched spots. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Mutual nearest neighbours

Begin with this mechanism: Reciprocal nearest relationships produce anchor-positive candidates. Work through the concrete case: Spot a chooses b and b chooses a. Make the limitation visible: Mutuality does not ensure biological identity. Complete the evidence step: Validate markers and domains. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Negative sampling

Begin with this mechanism: A dissimilar spot supplies the third member of each triplet. Work through the concrete case: The negative comes from a distinct location and expression state. Make the limitation visible: False negatives push related biology apart. Complete the evidence step: Inspect negative rules. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Triplet distance

Begin with this mechanism: Euclidean distances compare anchor-positive and anchor-negative embeddings. Work through the concrete case: d(a,p)=0.4 and d(a,n)=1.2. Make the limitation visible: Scale affects margin meaning. Complete the evidence step: Standardize or document norms. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Margin loss

Begin with this mechanism: The hinge term max(dap minus dan plus margin, zero) enforces separation. Work through the concrete case: With margin 1, the example loss is 0.2. Make the limitation visible: Too large a margin distorts geometry. Complete the evidence step: Sweep the margin. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Iterative refinement

Begin with this mechanism: Training alternates with rebuilding triplets from current embeddings. Work through the concrete case: Pairs improve after several update rounds. Make the limitation visible: Wrong early pairs can self-reinforce. Complete the evidence step: Track pair turnover and precision. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: Shared spatial domains

Begin with this mechanism: Clustering the corrected embedding identifies comparable and condition-specific regions. Work through the concrete case: Normal and diseased hippocampus share broad anatomy but retain a disease substructure. Make the limitation visible: Over-correction erases the difference. Complete the evidence step: Report conservation and batch removal. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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: ICP coordinate alignment

Begin with this mechanism: MNN pairs within chosen landmark domains guide rigid or permitted coordinate transformations. Work through the concrete case: A moving slice rotates and translates toward a reference. Make the limitation visible: Sparse or uneven anchors misregister tissue. Complete the evidence step: Report overlap and landmark residuals. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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, embeddings, MNN rules, negatives, margin, iterations and landmarks. Work through the concrete case: A student traces one pair from graph code to 3D coordinate. Make the limitation visible: Alignment is a model-supported correspondence. Complete the evidence step: Write a multi-slice integration card. Write every matrix, graph, attention weight, posterior distribution, discriminator score, triplet distance or transformation with named dimensions and units. Calculate one small example by hand, then change one assumption and explain which output moves and why. Use complete donors, slides, tissue regions, image batches or gene panels 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.