WHY MATHEMATICS?

Why Mathematics? | STACI for Spatial Transcriptomics and Chromatin Images: Graph Autoencoders, Overparameterization, Latent Alignment and Biomarkers

How STACI combines gene expression, cell neighbourhoods and chromatin images in one learned representation, and why translated expression remains a prediction rather than a measurement.

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 expression together with a cell-adjacency graph
Reconstruct several modalities from one latent representation
Study overparameterization as an empirical integration strategy
Test cross-modal predictions and combined biomarkers on held-out tissue

Chapter map

1. Expression features
2. Cell adjacency
3. Graph convolution
4. Expression decoder
5. Adjacency decoder
6. Joint graph loss
7. Overparameterized latent space
8. Batch diagnostics
9. Chromatin image patches
10. Image autoencoder
11. Latent alignment loss
12. Cross-modal decoder
13. Tissue clustering
14. Joint biomarkers
15. Modality limits
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 features

The central mathematical idea is concrete. Each cell or spot carries a molecular vector.

A concrete situation

STARmap measures a selected gene panel.

What the mathematics does not license

Panel limits the recoverable biology. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Document feature coverage. 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. Cell adjacency

Here the geometry becomes operational. Coordinates become a graph through k-nearest neighbours or a distance rule.

A concrete situation

Every nucleus connects to nearby cells.

What the mathematics does not license

Neighbour choice changes tissue context. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Plot physical edge lengths. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

3. Graph convolution

This step deserves a careful pause. A node representation combines its features with neighbour information.

A concrete situation

One update averages transformed local signals.

What the mathematics does not license

Repeated aggregation can oversmooth. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare layer depth. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

4. Expression decoder

The useful question is not only how to calculate, but what the calculation assumes. Latent coordinates reconstruct gene-expression parameters.

A concrete situation

A cell receives a predicted expression vector.

What the mathematics does not license

Reconstruction can favour abundant genes. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use gene-wise 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.

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. Adjacency decoder

A small example makes the mechanism visible. Latent coordinates also reconstruct graph links.

A concrete situation

Nearby cells receive high link scores.

What the mathematics does not license

Negative-edge sampling changes loss. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Publish sampling 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.

6. Joint graph loss

This is where modelling choices begin to affect evidence. Expression and adjacency losses train one representation.

A concrete situation

A weight balances molecular and graph errors.

What the mathematics does not license

Loss scales can dominate each other. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Normalize and sweep weights. 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. Overparameterized latent space

The central mathematical idea is concrete. Latent width exceeds the input feature count.

A concrete situation

A wide embedding precedes principal-component analysis.

What the mathematics does not license

Capacity can memorize noise. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Use held-out samples. 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. Batch diagnostics

Here the geometry becomes operational. Sample silhouette and entropy of mixing assess residual separation.

A concrete situation

Equivalent regions mix across tissue slices.

What the mathematics does not license

Mixing alone can reward overcorrection. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Pair with anatomical conservation. 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. Chromatin image patches

This step deserves a careful pause. Nucleus-centred pixels represent morphology and condensation.

A concrete situation

A patch covers the full nucleus.

What the mathematics does not license

Segmentation and stain shifts matter. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Audit image quality by sample. 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. Image autoencoder

The useful question is not only how to calculate, but what the calculation assumes. A convolutional encoder and decoder reconstruct each patch.

A concrete situation

The latent code preserves visible chromatin structure.

What the mathematics does not license

Good pixels do not guarantee biological meaning. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Validate morphometric associations. 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. Latent alignment loss

A small example makes the mechanism visible. Paired image and transcriptomic codes are brought closer.

A concrete situation

The same cell supplies two aligned embeddings.

What the mathematics does not license

A strong loss can collapse modality-specific detail. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Inspect cross-modal and private signal. 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. Cross-modal decoder

This is where modelling choices begin to affect evidence. An image code can pass through the expression decoder.

A concrete situation

A held-out image produces an expression prediction.

What the mathematics does not license

Domain shift can make translation unreliable. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Hold out entire conditions. 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. Tissue clustering

The central mathematical idea is concrete. Downstream clustering partitions the learned representation.

A concrete situation

Leiden identifies recurring anatomical regions.

What the mathematics does not license

Resolution adds a separate choice. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Publish stability across resolutions. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

14. Joint biomarkers

Here the geometry becomes operational. Latent gradients can connect expression, neighbourhood and chromatin features.

A concrete situation

A region associates plaque size with combined features.

What the mathematics does not license

Association is not causal mechanism. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Use orthogonal validation. 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. Modality limits

This step deserves a careful pause. Missing or noisy modalities constrain what the model can learn.

A concrete situation

Image-only deployment relies on paired training.

What the mathematics does not license

Unseen tissue states may be out of distribution. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Measure distance to training data. 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

The useful question is not only how to calculate, but what the calculation assumes. Version graphs, patches, widths, loss weights, splits and decoders.

A concrete situation

A student computes one graph step and loss.

What the mathematics does not license

Prediction remains conditional. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Write a STACI 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

STACI learns one latent representation from gene expression, cell adjacency and chromatin images. A graph-convolutional autoencoder encodes expression together with neighbouring cells, while separate decoders reconstruct gene expression and adjacency. A convolutional image autoencoder is aligned with the transcriptomic representation for paired cells.

The method deliberately uses an overparameterized latent space whose dimension is larger than the input feature dimension. In the published experiments this architecture emphasized leading biological variation and reduced sample separation. That is an empirical property of the studied settings, not a universal theorem that wider networks always remove batch effects.

Once paired modalities share a latent space, an image from a held-out tissue can pass through the image encoder and the gene-expression decoder. The output is a model prediction conditional on training data, preprocessing, pairing and domain similarity. It must never be described as a measured transcriptome.

A useful calculation builds a four-node adjacency matrix, performs one normalized graph-aggregation step, compares expression and adjacency reconstruction losses, and adds an alignment distance between image and transcriptomic embeddings. Changing the neighbourhood rule or alignment weight shows which evidence controls the joint representation.

Frequently asked questions

What is STACI?

STACI is a framework that learns a joint representation of spatial gene expression, cellular adjacency and chromatin images using graph and image autoencoders.

What is reconstructed from the graph representation?

Separate decoders reconstruct gene expression and cell adjacency so the latent coordinates must retain molecular and neighbourhood information.

What does overparameterization mean here?

The latent dimension is made larger than the input feature dimension; the paper studies this as a way to reduce sample separation in its datasets.

How are images aligned?

An image autoencoder reconstructs chromatin patches while an alignment loss brings paired image and transcriptomic representations together.

Can STACI infer gene expression from images?

It can produce model-based expression predictions for unseen image-only tissue through the image encoder and expression decoder.

Does that prediction replace an assay?

No. It depends on training domain, pairing, preprocessing and model fit, and requires held-out validation before scientific use.

Primary source and fact-check note

Xinyi Zhang, Yixuan Liu, Zhicheng Dai et al., “Graph-based autoencoder integrates spatial transcriptomics with chromatin images and identifies joint biomarkers for Alzheimer’s disease,” Nature Communications 13, 6003 (2022), DOI 10.1038/s41467-022-35233-1. 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

Feature audit

Document molecular panel and transformations.

Graph audit

Compare kNN and physical-radius adjacency.

Loss audit

Publish reconstruction and alignment scales.

Width audit

Compare under- and overparameterized models.

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.

Batch audit

Use mixing plus anatomical conservation.

Image audit

Check segmentation, patch size and stain shifts.

Prediction audit

Hold out whole samples and conditions.

Final claim

Image-derived expression is a model prediction.

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 features

Begin with this mechanism: Each cell or spot carries a molecular vector. Work through the concrete case: STARmap measures a selected gene panel. Make the limitation visible: Panel limits the recoverable biology. Complete the evidence step: Document feature coverage. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Cell adjacency

Begin with this mechanism: Coordinates become a graph through k-nearest neighbours or a distance rule. Work through the concrete case: Every nucleus connects to nearby cells. Make the limitation visible: Neighbour choice changes tissue context. Complete the evidence step: Plot physical edge lengths. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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 convolution

Begin with this mechanism: A node representation combines its features with neighbour information. Work through the concrete case: One update averages transformed local signals. Make the limitation visible: Repeated aggregation can oversmooth. Complete the evidence step: Compare layer depth. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Expression decoder

Begin with this mechanism: Latent coordinates reconstruct gene-expression parameters. Work through the concrete case: A cell receives a predicted expression vector. Make the limitation visible: Reconstruction can favour abundant genes. Complete the evidence step: Use gene-wise errors. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Adjacency decoder

Begin with this mechanism: Latent coordinates also reconstruct graph links. Work through the concrete case: Nearby cells receive high link scores. Make the limitation visible: Negative-edge sampling changes loss. Complete the evidence step: Publish sampling rules. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Joint graph loss

Begin with this mechanism: Expression and adjacency losses train one representation. Work through the concrete case: A weight balances molecular and graph errors. Make the limitation visible: Loss scales can dominate each other. Complete the evidence step: Normalize and sweep weights. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Overparameterized latent space

Begin with this mechanism: Latent width exceeds the input feature count. Work through the concrete case: A wide embedding precedes principal-component analysis. Make the limitation visible: Capacity can memorize noise. Complete the evidence step: Use held-out samples. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Batch diagnostics

Begin with this mechanism: Sample silhouette and entropy of mixing assess residual separation. Work through the concrete case: Equivalent regions mix across tissue slices. Make the limitation visible: Mixing alone can reward overcorrection. Complete the evidence step: Pair with anatomical conservation. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Chromatin image patches

Begin with this mechanism: Nucleus-centred pixels represent morphology and condensation. Work through the concrete case: A patch covers the full nucleus. Make the limitation visible: Segmentation and stain shifts matter. Complete the evidence step: Audit image quality by sample. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Image autoencoder

Begin with this mechanism: A convolutional encoder and decoder reconstruct each patch. Work through the concrete case: The latent code preserves visible chromatin structure. Make the limitation visible: Good pixels do not guarantee biological meaning. Complete the evidence step: Validate morphometric associations. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Latent alignment loss

Begin with this mechanism: Paired image and transcriptomic codes are brought closer. Work through the concrete case: The same cell supplies two aligned embeddings. Make the limitation visible: A strong loss can collapse modality-specific detail. Complete the evidence step: Inspect cross-modal and private signal. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Cross-modal decoder

Begin with this mechanism: An image code can pass through the expression decoder. Work through the concrete case: A held-out image produces an expression prediction. Make the limitation visible: Domain shift can make translation unreliable. Complete the evidence step: Hold out entire conditions. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Tissue clustering

Begin with this mechanism: Downstream clustering partitions the learned representation. Work through the concrete case: Leiden identifies recurring anatomical regions. Make the limitation visible: Resolution adds a separate choice. Complete the evidence step: Publish stability across resolutions. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Joint biomarkers

Begin with this mechanism: Latent gradients can connect expression, neighbourhood and chromatin features. Work through the concrete case: A region associates plaque size with combined features. Make the limitation visible: Association is not causal mechanism. Complete the evidence step: Use orthogonal validation. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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: Modality limits

Begin with this mechanism: Missing or noisy modalities constrain what the model can learn. Work through the concrete case: Image-only deployment relies on paired training. Make the limitation visible: Unseen tissue states may be out of distribution. Complete the evidence step: Measure distance to training data. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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, patches, widths, loss weights, splits and decoders. Work through the concrete case: A student computes one graph step and loss. Make the limitation visible: Prediction remains conditional. Complete the evidence step: Write a STACI model card. Write the dimensions, counts, coordinates, mixture parameters, graph edges, priors, losses, dispersions, topic proportions or update rules with named units. Calculate one small example by hand, then vary one assumption and explain which result moves and why. Reserve complete sections, donors, tissue regions, genes or image 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.