WHY MATHEMATICS?

Why Mathematics? | GraphST for Spatial Representation: Graph Neural Networks, Contrastive Losses, Mapping Matrices and Multi-Slice Integration

How GraphST uses spatial graphs, denoising encoders and contrastive objectives to learn tissue domains, integrate slices and map single-cell references into spatial measurements.

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 spatial neighbours and expression with a graph neural network
Use reconstruction and contrastive objectives without manual labels
Learn a constrained cell-to-spot mapping matrix for deconvolution
Test batch removal and biological conservation across tissue slices

Chapter map

1. Spot feature matrix
2. Spatial graph A
3. Graph neural encoder
4. Feature corruption view
5. Shared encoder
6. Positive similarity
7. Negative separation
8. Contrastive objective
9. Self-reconstruction loss
10. Domain clustering
11. Block graph for slices
12. Batch-mixing evidence
13. Single-cell autoencoder
14. Mapping matrix M
15. Augmentation-free mapping contrast
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. Spot feature matrix

This step deserves a careful pause. Normalized gene expression supplies node features for each spatial location.

A concrete situation

Three thousand genes describe 2,500 spots.

What the mathematics does not license

Batch effects enter the features. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Audit 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. Spatial graph A

The useful question is not only how to calculate, but what the calculation assumes. Coordinates define edges between nearby spots or cells.

A concrete situation

A k-nearest graph links every spot to six neighbours.

What the mathematics does not license

One k may not fit different densities. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Run graph sensitivity. 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. Graph neural encoder

A small example makes the mechanism visible. Neighbour features are aggregated through trainable layers into a low-dimensional representation.

A concrete situation

A spot embedding includes local tissue context.

What the mathematics does not license

Deep layers can erase boundaries. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare receptive fields. 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. Feature corruption view

This is where modelling choices begin to affect evidence. A data augmentation perturbs expression while retaining a related graph view for self-supervision.

A concrete situation

Selected features are shuffled or masked for a contrastive view.

What the mathematics does not license

Augmentation can destroy biology. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Test perturbation strength. 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. Shared encoder

The central mathematical idea is concrete. Original and augmented views pass through the same network so representations are comparable.

A concrete situation

Two views produce vectors z and z-prime.

What the mathematics does not license

Shared weights do not guarantee invariance. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Measure agreement by region. 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. Positive similarity

Here the geometry becomes operational. Matched representations are encouraged to have high cosine or score similarity.

A concrete situation

One spot and its augmented copy form a positive pair.

What the mathematics does not license

A poor augmentation makes a false positive. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Inspect examples. 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. Negative separation

This step deserves a careful pause. Unmatched nodes provide contrast that prevents embedding collapse.

A concrete situation

A distant unrelated spot is treated as a negative.

What the mathematics does not license

Distant can still mean same cell type. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Use spatial and biological 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.

8. Contrastive objective

The useful question is not only how to calculate, but what the calculation assumes. Log-softmax ratios balance positive similarity against many negatives.

A concrete situation

Raising the positive score lowers the loss.

What the mathematics does not license

Temperature controls sharpness. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report temperature and batch 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.

9. Self-reconstruction loss

A small example makes the mechanism visible. A decoder reconstructs informative spot features from the embedding.

A concrete situation

The model predicts a denoised expression representation.

What the mathematics does not license

Reconstruction may preserve technical noise. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use gene-wise 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.

10. Domain clustering

This is where modelling choices begin to affect evidence. The learned representation is clustered into spatial domains after training.

A concrete situation

A breast tissue section separates tumour and stromal niches.

What the mathematics does not license

Cluster count changes the map. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare stability and pathology. 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. Block graph for slices

The central mathematical idea is concrete. Multiple tissue sections can be represented jointly while preserving within-slice spatial edges.

A concrete situation

Two adjacent sections contribute separate graph blocks.

What the mathematics does not license

No cross-slice edge may limit direct alignment. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Document integration design. 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. Batch-mixing evidence

Here the geometry becomes operational. Integrated embeddings should mix technical batches only where biology is comparable.

A concrete situation

Replicate cortex layers align across slices.

What the mathematics does not license

Excess mixing erases condition effects. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

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

13. Single-cell autoencoder

This step deserves a careful pause. Reference cells are compressed and reconstructed to reduce raw expression noise before mapping.

A concrete situation

A cell receives a denoised 32-dimensional feature.

What the mathematics does not license

Reference states absent in tissue can still map. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Add novelty controls. 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. Mapping matrix M

The useful question is not only how to calculate, but what the calculation assumes. Nonnegative weights assign cells or cell states across spatial spots.

A concrete situation

M has cells as rows and spots as columns.

What the mathematics does not license

Constraints and normalization define meaning. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Publish row and column 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.

15. Augmentation-free mapping contrast

A small example makes the mechanism visible. Mapped single-cell features are aligned with reconstructed spatial features while neighbour relations shape similarities.

A concrete situation

A predicted spot representation approaches its observed counterpart.

What the mathematics does not license

Alignment can exploit shortcuts. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use held-out genes and mixtures. 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 graphs, augmentations, losses, temperature, mapping constraints and slices.

A concrete situation

A student traces one composition from cells through M to a spot.

What the mathematics does not license

An embedding is conditional evidence. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

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

A seven-day practice lab

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

GraphST learns spatial representations with graph neural networks and self-supervised contrastive objectives. It uses gene expression as node features and spatial proximity as graph structure, reconstructs informative spot features and teaches adjacent or matched representations to agree more than inappropriate comparisons. The same framework supports clustering, multi-slice integration and cell-type deconvolution.

For deconvolution, GraphST first denoises single-cell and spatial profiles in learned feature spaces, then estimates a nonnegative cell-to-spot mapping matrix. An augmentation-free contrastive objective aligns mapped single-cell features with reconstructed spatial features while using spatial neighbourhoods to define helpful similarities. Summed mapping weights support cell-type composition estimates.

Validation must match the downstream question. Domain clustering needs held-out anatomy and stability; multi-slice integration needs biological conservation plus batch-removal diagnostics; deconvolution needs known mixtures, reference swaps and imaging. Contrastive separation can produce attractive embeddings even when its positive and negative pair rules are biologically wrong.

Frequently asked questions

What is GraphST?

GraphST is a graph self-supervised contrastive-learning framework for spatial clustering, integration of multiple tissue slices and cell-type deconvolution.

What is the positive-pair idea?

Representations that correspond to the same spot under augmentation, mapped and observed versions of a spot, or spatially related nodes are encouraged to be similar according to the task design.

Why are negative pairs important?

They provide separation pressure so all embeddings do not collapse to one point. A biologically poor negative rule can also push genuinely related tissue states apart.

What is the mapping matrix?

It is a learned nonnegative matrix that assigns single-cell profiles across spatial spots. Aggregating rows by cell label yields estimated cell-type composition.

How can GraphST integrate slices?

A shared graph representation is learned across tissue sections so analogous biological structure aligns while technical differences are reduced. Alignment must be tested for biological over-correction.

Does better contrastive loss prove better biology?

No. The loss measures agreement with chosen pairs and objectives. Independent anatomy, held-out sections and known mixtures are required.

Primary source and fact-check note

Yahui Long, Kok Siong Ang, Mengwei Li et al., “Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST,” Nature Communications 14, 1155 (2023), DOI 10.1038/s41467-023-36796-3. 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

Graph audit

Test neighbour rules and tissue gaps.

Augmentation audit

Show which biological information perturbations preserve.

Contrast audit

Report positives, negatives, temperature and batches.

Reconstruction audit

Inspect gene- and region-level residuals.

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.

Integration audit

Measure batch removal and biology conservation.

Mapping audit

Test known mixtures and absent reference states.

Section holdout

Reserve complete slices from tuning.

Final claim

A contrastive embedding or mapping is model evidence.

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: Spot feature matrix

Begin with this mechanism: Normalized gene expression supplies node features for each spatial location. Work through the concrete case: Three thousand genes describe 2,500 spots. Make the limitation visible: Batch effects enter the features. Complete the evidence step: Audit each slice separately. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 2: Spatial graph A

Begin with this mechanism: Coordinates define edges between nearby spots or cells. Work through the concrete case: A k-nearest graph links every spot to six neighbours. Make the limitation visible: One k may not fit different densities. Complete the evidence step: Run graph sensitivity. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 3: Graph neural encoder

Begin with this mechanism: Neighbour features are aggregated through trainable layers into a low-dimensional representation. Work through the concrete case: A spot embedding includes local tissue context. Make the limitation visible: Deep layers can erase boundaries. Complete the evidence step: Compare receptive fields. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 4: Feature corruption view

Begin with this mechanism: A data augmentation perturbs expression while retaining a related graph view for self-supervision. Work through the concrete case: Selected features are shuffled or masked for a contrastive view. Make the limitation visible: Augmentation can destroy biology. Complete the evidence step: Test perturbation strength. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

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: Shared encoder

Begin with this mechanism: Original and augmented views pass through the same network so representations are comparable. Work through the concrete case: Two views produce vectors z and z-prime. Make the limitation visible: Shared weights do not guarantee invariance. Complete the evidence step: Measure agreement by region. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 6: Positive similarity

Begin with this mechanism: Matched representations are encouraged to have high cosine or score similarity. Work through the concrete case: One spot and its augmented copy form a positive pair. Make the limitation visible: A poor augmentation makes a false positive. Complete the evidence step: Inspect examples. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 7: Negative separation

Begin with this mechanism: Unmatched nodes provide contrast that prevents embedding collapse. Work through the concrete case: A distant unrelated spot is treated as a negative. Make the limitation visible: Distant can still mean same cell type. Complete the evidence step: Use spatial and biological checks. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 8: Contrastive objective

Begin with this mechanism: Log-softmax ratios balance positive similarity against many negatives. Work through the concrete case: Raising the positive score lowers the loss. Make the limitation visible: Temperature controls sharpness. Complete the evidence step: Report temperature and batch size. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

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: Self-reconstruction loss

Begin with this mechanism: A decoder reconstructs informative spot features from the embedding. Work through the concrete case: The model predicts a denoised expression representation. Make the limitation visible: Reconstruction may preserve technical noise. Complete the evidence step: Use gene-wise residuals. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 10: Domain clustering

Begin with this mechanism: The learned representation is clustered into spatial domains after training. Work through the concrete case: A breast tissue section separates tumour and stromal niches. Make the limitation visible: Cluster count changes the map. Complete the evidence step: Compare stability and pathology. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 11: Block graph for slices

Begin with this mechanism: Multiple tissue sections can be represented jointly while preserving within-slice spatial edges. Work through the concrete case: Two adjacent sections contribute separate graph blocks. Make the limitation visible: No cross-slice edge may limit direct alignment. Complete the evidence step: Document integration design. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 12: Batch-mixing evidence

Begin with this mechanism: Integrated embeddings should mix technical batches only where biology is comparable. Work through the concrete case: Replicate cortex layers align across slices. Make the limitation visible: Excess mixing erases condition effects. Complete the evidence step: Report conservation and removal metrics. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

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: Single-cell autoencoder

Begin with this mechanism: Reference cells are compressed and reconstructed to reduce raw expression noise before mapping. Work through the concrete case: A cell receives a denoised 32-dimensional feature. Make the limitation visible: Reference states absent in tissue can still map. Complete the evidence step: Add novelty controls. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 14: Mapping matrix M

Begin with this mechanism: Nonnegative weights assign cells or cell states across spatial spots. Work through the concrete case: M has cells as rows and spots as columns. Make the limitation visible: Constraints and normalization define meaning. Complete the evidence step: Publish row and column rules. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 15: Augmentation-free mapping contrast

Begin with this mechanism: Mapped single-cell features are aligned with reconstructed spatial features while neighbour relations shape similarities. Work through the concrete case: A predicted spot representation approaches its observed counterpart. Make the limitation visible: Alignment can exploit shortcuts. Complete the evidence step: Use held-out genes and mixtures. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 16: Governance and learning

Begin with this mechanism: Version graphs, augmentations, losses, temperature, mapping constraints and slices. Work through the concrete case: A student traces one composition from cells through M to a spot. Make the limitation visible: An embedding is conditional evidence. Complete the evidence step: Write a contrastive-learning ledger. Write the matrices, graph, similarity, probability distribution, loss term or permutation rule 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, tissue sections, spatial regions, image batches or gene sets outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.