WHY MATHEMATICS?

Why Mathematics? | STAGATE for Spatial Domains: Neighbour Graphs, Attention Weights, Autoencoder Reconstruction and Denoised Expression

How STAGATE learns adaptive neighbour influence on a spatial graph, compresses gene expression with a graph attention autoencoder and supports domain discovery and denoising.

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

Build and inspect a spatial neighbour network
Normalize graph attention weights over local candidates
Reconstruct expression through an encoder and decoder
Test domains and denoising with held-out measurements

Chapter map

1. Expression matrix X
2. Spatial coordinates
3. Neighbour rule
4. Adjacency matrix
5. Raw attention score
6. Neighbour softmax
7. Attention aggregation
8. Encoder layers
9. Latent embedding Z
10. Decoder reconstruction
11. Reconstruction loss
12. Domain clustering
13. Denoised expression
14. Domain-related genes
15. Three-dimensional use
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 matrix X

This step deserves a careful pause. Rows are spots and columns are selected genes after declared normalization.

A concrete situation

Two thousand spots each have 3,000 features.

What the mathematics does not license

Transformations alter reconstruction targets. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Version genes and 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 coordinates

The useful question is not only how to calculate, but what the calculation assumes. Each spot has x and y coordinates in a documented reference frame.

A concrete situation

Two adjacent Visium spots are one lattice step apart.

What the mathematics does not license

Pixel and physical units can differ. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record scale and orientation. 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. Neighbour rule

A small example makes the mechanism visible. A radius or platform-aware rule selects candidate spatial edges.

A concrete situation

An interior spot connects to its six immediate neighbours.

What the mathematics does not license

A radius can bridge a tissue void. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Overlay every edge. 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. Adjacency matrix

This is where modelling choices begin to affect evidence. A binary matrix records which node pairs are connected.

A concrete situation

A_ij equals one for chosen neighbours.

What the mathematics does not license

The matrix encodes a modelling assumption. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare radii and graph types. 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. Raw attention score

The central mathematical idea is concrete. Trainable transformations score compatibility between a centre and each neighbour.

A concrete situation

Two similar cortical spots receive a larger score.

What the mathematics does not license

Scores can track technical similarity. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect by batch and 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. Neighbour softmax

Here the geometry becomes operational. Exponentiated scores are divided by their neighbourhood sum.

A concrete situation

Scores 1 and 2 become weights about 0.27 and 0.73.

What the mathematics does not license

Softmax forces relative competition. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report weight distributions. 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. Attention aggregation

This step deserves a careful pause. A centre receives a weighted sum of transformed neighbour features.

A concrete situation

A boundary spot borrows mostly from its same-layer neighbour.

What the mathematics does not license

A false high-weight edge spreads error. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Challenge cross-boundary edges. 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. Encoder layers

The useful question is not only how to calculate, but what the calculation assumes. Successive graph attention layers compress expression into a latent code.

A concrete situation

A 3,000-gene vector becomes 30 latent values.

What the mathematics does not license

Depth can over-smooth domains. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare one and two layers. 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. Latent embedding Z

A small example makes the mechanism visible. Each spot becomes a point in a lower-dimensional geometry.

A concrete situation

Nearby layer spots form a compact cloud.

What the mathematics does not license

Distance depends on training and scale. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat seeds and standardization. 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. Decoder reconstruction

This is where modelling choices begin to affect evidence. A mirrored graph-aware decoder maps latent codes back to gene space.

A concrete situation

The model predicts a full expression profile.

What the mathematics does not license

Good averages can hide rare-gene errors. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Stratify 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.

11. Reconstruction loss

The central mathematical idea is concrete. Squared discrepancies between input and reconstructed normalized expression are minimized.

A concrete situation

An error of 0.2 contributes 0.04 before averaging.

What the mathematics does not license

Loss weights frequent signals heavily. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect gene-level errors. 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. Domain clustering

Here the geometry becomes operational. A conventional clustering method groups the learned latent vectors.

A concrete situation

mclust or Louvain separates candidate cortical layers.

What the mathematics does not license

Cluster count changes the map. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use stability and anatomy. 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. Denoised expression

This step deserves a careful pause. Decoder output can be interpreted as a graph-informed reconstruction.

A concrete situation

A low-count marker is restored in a coherent region.

What the mathematics does not license

Smoothing can fabricate continuity. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Hide measured genes and regions. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

14. Domain-related genes

The useful question is not only how to calculate, but what the calculation assumes. Expression comparisons describe genes associated with learned regions.

A concrete situation

A marker has large fold change across two domains.

What the mathematics does not license

Circular discovery inflates certainty. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Confirm in independent sections. 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. Three-dimensional use

A small example makes the mechanism visible. Embeddings and spatial domains can support reconstruction across serial slices.

A concrete situation

Corresponding layers guide slice comparison.

What the mathematics does not license

Registration errors mimic change. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Validate landmarks and orientation. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

16. Governance and learning

This is where modelling choices begin to affect evidence. Version graph, architecture, weights, losses, seeds and clustering.

A concrete situation

A student traces one spot from coordinates to reconstructed gene.

What the mathematics does not license

Attention is influence, not causal importance. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

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

A seven-day practice lab

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

Questions for a student, parent or mentor

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

A method-specific audit before deployment

STAGATE begins with a spatial neighbour network and asks a focused mathematical question: which nearby spots should influence one another, and by how much? A graph attention autoencoder learns those influence weights from the expression profiles while reconstructing the input. Its latent coordinates can support spatial-domain clustering, and its decoder can provide a denoised expression estimate.

The method separates geometry from learned compatibility. Coordinates determine the candidate edges; attention scores decide how information moves along those edges; an encoder compresses the weighted messages; and a decoder minimizes reconstruction error. This makes the graph inspectable. A close spot may receive little weight when its expression differs sharply, while a boundary-consistent neighbour may receive more.

Validation must challenge both the candidate graph and the attention weights. Inspect edges across tissue gaps, compare fixed averaging with learned attention, reserve genes or complete regions before tuning, and repeat clustering across seeds and graph radii. A reconstructed value is a model estimate rather than a replacement for the original count, and a smooth domain map still requires anatomical evidence.

Frequently asked questions

What is STAGATE?

STAGATE is a graph attention autoencoder for learning low-dimensional spatially aware representations from gene expression and a spatial neighbour network.

What does the spatial neighbour network do?

It defines which nearby spots may exchange information. The graph supplies candidate relationships; the attention mechanism learns their relative influence.

What is an attention coefficient?

It is a normalized weight assigned to a neighbour during message passing. Coefficients for the selected neighbourhood are commonly normalized with a softmax.

Why use an autoencoder?

The encoder compresses expression through graph-aware transformations, while the decoder reconstructs expression. Reconstruction loss teaches a representation that retains information needed to explain the input.

Is denoised expression newly measured data?

No. It is a reconstruction conditional on the trained network, graph and preprocessing choices. Measured holdouts and independent assays are needed.

Do learned domains prove anatomy?

No. They are clusters in a model-derived embedding. Histology, markers, independent sections and sensitivity analysis are required.

Primary source and fact-check note

Kangning Dong and Shihua Zhang, “Deciphering spatial domains from spatially resolved transcriptomics with an adaptive graph attention auto-encoder,” Nature Communications 13, 1739 (2022), DOI 10.1038/s41467-022-29439-6. 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

Graph audit

Overlay radius edges and tissue masks.

Attention audit

Compare learned weights with fixed averaging.

Boundary audit

Challenge cross-domain high-weight edges.

Reconstruction audit

Hide genes and complete spatial regions.

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.

Cluster audit

Repeat seeds, radii and cluster counts.

Denoising audit

Compare predictions with untouched measurements.

External check

Use histology and independent sections.

Final claim

Attention and denoising are conditional model outputs.

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 matrix X

Begin with this mechanism: Rows are spots and columns are selected genes after declared normalization. Work through the concrete case: Two thousand spots each have 3,000 features. Make the limitation visible: Transformations alter reconstruction targets. Complete the evidence step: Version genes and preprocessing. 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: Spatial coordinates

Begin with this mechanism: Each spot has x and y coordinates in a documented reference frame. Work through the concrete case: Two adjacent Visium spots are one lattice step apart. Make the limitation visible: Pixel and physical units can differ. Complete the evidence step: Record scale and orientation. 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: Neighbour rule

Begin with this mechanism: A radius or platform-aware rule selects candidate spatial edges. Work through the concrete case: An interior spot connects to its six immediate neighbours. Make the limitation visible: A radius can bridge a tissue void. Complete the evidence step: Overlay every edge. 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: Adjacency matrix

Begin with this mechanism: A binary matrix records which node pairs are connected. Work through the concrete case: A_ij equals one for chosen neighbours. Make the limitation visible: The matrix encodes a modelling assumption. Complete the evidence step: Compare radii and graph types. 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: Raw attention score

Begin with this mechanism: Trainable transformations score compatibility between a centre and each neighbour. Work through the concrete case: Two similar cortical spots receive a larger score. Make the limitation visible: Scores can track technical similarity. Complete the evidence step: Inspect by batch and region. 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: Neighbour softmax

Begin with this mechanism: Exponentiated scores are divided by their neighbourhood sum. Work through the concrete case: Scores 1 and 2 become weights about 0.27 and 0.73. Make the limitation visible: Softmax forces relative competition. Complete the evidence step: Report weight distributions. 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: Attention aggregation

Begin with this mechanism: A centre receives a weighted sum of transformed neighbour features. Work through the concrete case: A boundary spot borrows mostly from its same-layer neighbour. Make the limitation visible: A false high-weight edge spreads error. Complete the evidence step: Challenge cross-boundary edges. 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: Encoder layers

Begin with this mechanism: Successive graph attention layers compress expression into a latent code. Work through the concrete case: A 3,000-gene vector becomes 30 latent values. Make the limitation visible: Depth can over-smooth domains. Complete the evidence step: Compare one and two layers. 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: Latent embedding Z

Begin with this mechanism: Each spot becomes a point in a lower-dimensional geometry. Work through the concrete case: Nearby layer spots form a compact cloud. Make the limitation visible: Distance depends on training and scale. Complete the evidence step: Repeat seeds and standardization. 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: Decoder reconstruction

Begin with this mechanism: A mirrored graph-aware decoder maps latent codes back to gene space. Work through the concrete case: The model predicts a full expression profile. Make the limitation visible: Good averages can hide rare-gene errors. Complete the evidence step: Stratify 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 11: Reconstruction loss

Begin with this mechanism: Squared discrepancies between input and reconstructed normalized expression are minimized. Work through the concrete case: An error of 0.2 contributes 0.04 before averaging. Make the limitation visible: Loss weights frequent signals heavily. Complete the evidence step: Inspect gene-level errors. 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: Domain clustering

Begin with this mechanism: A conventional clustering method groups the learned latent vectors. Work through the concrete case: mclust or Louvain separates candidate cortical layers. Make the limitation visible: Cluster count changes the map. Complete the evidence step: Use stability and anatomy. 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: Denoised expression

Begin with this mechanism: Decoder output can be interpreted as a graph-informed reconstruction. Work through the concrete case: A low-count marker is restored in a coherent region. Make the limitation visible: Smoothing can fabricate continuity. Complete the evidence step: Hide measured genes and regions. 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: Domain-related genes

Begin with this mechanism: Expression comparisons describe genes associated with learned regions. Work through the concrete case: A marker has large fold change across two domains. Make the limitation visible: Circular discovery inflates certainty. Complete the evidence step: Confirm in independent sections. 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: Three-dimensional use

Begin with this mechanism: Embeddings and spatial domains can support reconstruction across serial slices. Work through the concrete case: Corresponding layers guide slice comparison. Make the limitation visible: Registration errors mimic change. Complete the evidence step: Validate landmarks and orientation. 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 graph, architecture, weights, losses, seeds and clustering. Work through the concrete case: A student traces one spot from coordinates to reconstructed gene. Make the limitation visible: Attention is influence, not causal importance. Complete the evidence step: Write a graph-attention model 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.