WHY MATHEMATICS?

Why Mathematics? | SpaceFlow for Spatial Trajectories: Deep Graph Infomax, Spatial Regularization, Embedding Geometry and Pseudo-Spatiotemporal Maps

How SpaceFlow uses an information-maximizing graph network and spatial continuity constraints to learn domains and trace pseudo-spatiotemporal organization in tissue.

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

Represent locations and expression as a spatial graph
Contrast genuine node-summary pairs with corrupted pairs
Regularize latent geometry with physical proximity
Combine embedding order and space into pseudo-spatiotemporal maps

Chapter map

1. Node feature matrix
2. Spatial graph
3. Graph convolution encoder
4. Local embeddings
5. Global readout
6. Summary nonlinearity
7. Corrupted negatives
8. Discriminator score
9. Binary contrastive loss
10. Information-maximizing interpretation
11. Spatial regularization
12. Joint objective
13. Embedding geometry
14. Spatial domains
15. Pseudo-spatiotemporal map
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. Node feature matrix

A small example makes the mechanism visible. Every spot or cell has a selected normalized expression vector.

A concrete situation

Fifty principal components summarize a profile.

What the mathematics does not license

Dimension reduction can remove rare signals. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect explained variance and markers. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

2. Spatial graph

This is where modelling choices begin to affect evidence. Coordinates determine edges between nearby locations.

A concrete situation

A k-nearest graph connects each node to ten neighbours.

What the mathematics does not license

Density changes neighbourhood scale. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

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

3. Graph convolution encoder

The central mathematical idea is concrete. Node features are aggregated with spatial neighbours into latent vectors.

A concrete situation

A cell receives a 32-dimensional embedding.

What the mathematics does not license

Several layers can over-smooth. A trustworthy workflow records the choice before seeing the final test result.

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. Local embeddings

Here the geometry becomes operational. Each node vector encodes expression and neighbourhood context.

A concrete situation

Similar epithelial regions cluster in latent space.

What the mathematics does not license

Similarity is model-dependent. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Repeat seeds 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.

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. Global readout

This step deserves a careful pause. A pooling function turns all node vectors into one graph summary.

A concrete situation

The mean embedding becomes a summary vector.

What the mathematics does not license

A global average can hide rare regions. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare robust and regional readouts. 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. Summary nonlinearity

The useful question is not only how to calculate, but what the calculation assumes. A bounded activation produces the context vector used by the discriminator.

A concrete situation

A sigmoid maps summary coordinates to zero-one range.

What the mathematics does not license

Saturation weakens gradients. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Monitor activation 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. Corrupted negatives

A small example makes the mechanism visible. Feature rows are permuted so expression no longer matches graph location.

A concrete situation

A tumour profile is assigned to a distant normal node.

What the mathematics does not license

Easy negatives teach shortcuts. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use several corruption schemes. 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. Discriminator score

This is where modelling choices begin to affect evidence. A bilinear or similar function scores compatibility between a node and summary.

A concrete situation

A real pair scores 3 while a corrupted pair scores -1.

What the mathematics does not license

High scores can reflect nuisance structure. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Test batch predictability. 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. Binary contrastive loss

The central mathematical idea is concrete. Log loss rewards real pairs and penalizes corrupted pairs.

A concrete situation

Sigmoid scores become probabilities for two labels.

What the mathematics does not license

The objective is a surrogate for information. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Report negatives and balance. 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. Information-maximizing interpretation

Here the geometry becomes operational. Successful discrimination encourages local representations that preserve global graph information.

A concrete situation

Domain structure becomes encoded without labels.

What the mathematics does not license

Mutual information is not directly measured exactly. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Describe the bound and estimator. 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. Spatial regularization

This step deserves a careful pause. Nearby physical nodes are penalized when their embeddings separate excessively.

A concrete situation

Two adjacent cells incur a squared-distance penalty.

What the mathematics does not license

Strong regularization blurs a real boundary. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Sweep the spatial weight. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

12. Joint objective

The useful question is not only how to calculate, but what the calculation assumes. Information and spatial terms are combined with a declared coefficient.

A concrete situation

Loss equals contrastive loss plus lambda times spatial loss.

What the mathematics does not license

Lambda controls the scientific answer. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Publish a sensitivity curve. 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. Embedding geometry

A small example makes the mechanism visible. Distances and neighbourhoods in the learned space support downstream analysis.

A concrete situation

A continuous arc connects developmental states.

What the mathematics does not license

Curves can be optimization artifacts. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare methods and held-out 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.

14. Spatial domains

This is where modelling choices begin to affect evidence. Clustering groups latent vectors into coherent tissue regions.

A concrete situation

A cortex section separates layers.

What the mathematics does not license

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

Try it, check it, transfer it

Use anatomy and stability. 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. Pseudo-spatiotemporal map

The central mathematical idea is concrete. A diffusion or pseudotime ordering is combined with physical relationships.

A concrete situation

Cells progress from a source region along a spatial route.

What the mathematics does not license

Root and graph choices change direction. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Repeat plausible roots. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

16. Governance and learning

Here the geometry becomes operational. Version graph, corruption, encoder, spatial weight, clustering and root.

A concrete situation

A student calculates one discriminator and spatial-loss term.

What the mathematics does not license

A trajectory is conditional evidence. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Write a spatiotemporal 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

SpaceFlow turns spatial transcriptomics into a graph and learns a low-dimensional map with Deep Graph Infomax. Nearby cells or spots form nodes and edges; a graph encoder creates local vectors; a readout summarizes the whole graph; and a discriminator learns to distinguish genuine node-summary pairs from corrupted pairs. This information-maximizing objective avoids manual domain labels.

The embedding is also regularized to respect spatial continuity. Neighbouring locations should usually remain close in the latent space unless the data support a boundary. After training, clustering identifies spatial domains, and a pseudo-spatiotemporal map combines an expression-derived ordering with spatial relationships to trace gradual tissue organization. The mathematics connects graph theory, information objectives, geometry and trajectory inference.

Validation should challenge corruption rules, graph density and spatial regularization strength. Compare with non-spatial embeddings, reserve entire regions and sections, and repeat cluster and trajectory choices. A pseudo-spatiotemporal path is an inferred ordering from static measurements. It can guide experiments, but direct time-course, lineage or perturbation data are stronger evidence of temporal development.

Frequently asked questions

What is SpaceFlow?

SpaceFlow is an unsupervised graph-learning method that creates spatially consistent embeddings for spatial-domain and pseudo-spatiotemporal analysis.

What is Deep Graph Infomax?

It trains a graph encoder to make real node embeddings informative about a graph-level summary by distinguishing true node-summary pairs from corrupted negative pairs.

What is the corrupted graph view?

It is a deliberately broken association, commonly created by permuting node features, used to construct negatives for the discriminator.

Why add spatial regularization?

The information objective alone can place physically adjacent, similar locations far apart. A spatial term encourages local continuity while still allowing data-supported boundaries.

What is a pseudo-spatiotemporal map?

It is an inferred ordering that combines learned expression geometry and spatial relationships to describe gradual tissue organization.

Does a PSM measure real elapsed time?

No. It is a relative model-derived order. Direct time-course, lineage and perturbation evidence are needed for temporal claims.

Primary source and fact-check note

Honglei Ren, Benjamin L. Walker, Zixuan Cang and Qing Nie, “Identifying multicellular spatiotemporal organization of cells with SpaceFlow,” Nature Communications 13, 4076 (2022), DOI 10.1038/s41467-022-31739-w. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: topic-specific validation and responsible use

Graph audit

Compare k, radius and density-aware edges.

Corruption audit

Test whether negatives are too easy.

Information audit

Describe the optimized lower-bound objective.

Spatial audit

Sweep regularization without blurring boundaries.

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 and resolutions.

Trajectory audit

Vary roots and graph choices.

External check

Use time-course or lineage evidence.

Final claim

A PSM is a relative inferred organization.

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: Node feature matrix

Begin with this mechanism: Every spot or cell has a selected normalized expression vector. Work through the concrete case: Fifty principal components summarize a profile. Make the limitation visible: Dimension reduction can remove rare signals. Complete the evidence step: Inspect explained variance and markers. 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 graph

Begin with this mechanism: Coordinates determine edges between nearby locations. Work through the concrete case: A k-nearest graph connects each node to ten neighbours. Make the limitation visible: Density changes neighbourhood scale. Complete the evidence step: Compare k and radius graphs. 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: Graph convolution encoder

Begin with this mechanism: Node features are aggregated with spatial neighbours into latent vectors. Work through the concrete case: A cell receives a 32-dimensional embedding. Make the limitation visible: Several layers can over-smooth. Complete the evidence step: Compare receptive fields. 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: Local embeddings

Begin with this mechanism: Each node vector encodes expression and neighbourhood context. Work through the concrete case: Similar epithelial regions cluster in latent space. Make the limitation visible: Similarity is model-dependent. Complete the evidence step: Repeat seeds 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.

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: Global readout

Begin with this mechanism: A pooling function turns all node vectors into one graph summary. Work through the concrete case: The mean embedding becomes a summary vector. Make the limitation visible: A global average can hide rare regions. Complete the evidence step: Compare robust and regional readouts. 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: Summary nonlinearity

Begin with this mechanism: A bounded activation produces the context vector used by the discriminator. Work through the concrete case: A sigmoid maps summary coordinates to zero-one range. Make the limitation visible: Saturation weakens gradients. Complete the evidence step: Monitor activation 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: Corrupted negatives

Begin with this mechanism: Feature rows are permuted so expression no longer matches graph location. Work through the concrete case: A tumour profile is assigned to a distant normal node. Make the limitation visible: Easy negatives teach shortcuts. Complete the evidence step: Use several corruption schemes. 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: Discriminator score

Begin with this mechanism: A bilinear or similar function scores compatibility between a node and summary. Work through the concrete case: A real pair scores 3 while a corrupted pair scores -1. Make the limitation visible: High scores can reflect nuisance structure. Complete the evidence step: Test batch predictability. 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: Binary contrastive loss

Begin with this mechanism: Log loss rewards real pairs and penalizes corrupted pairs. Work through the concrete case: Sigmoid scores become probabilities for two labels. Make the limitation visible: The objective is a surrogate for information. Complete the evidence step: Report negatives and balance. 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: Information-maximizing interpretation

Begin with this mechanism: Successful discrimination encourages local representations that preserve global graph information. Work through the concrete case: Domain structure becomes encoded without labels. Make the limitation visible: Mutual information is not directly measured exactly. Complete the evidence step: Describe the bound and estimator. 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: Spatial regularization

Begin with this mechanism: Nearby physical nodes are penalized when their embeddings separate excessively. Work through the concrete case: Two adjacent cells incur a squared-distance penalty. Make the limitation visible: Strong regularization blurs a real boundary. Complete the evidence step: Sweep the spatial weight. 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: Joint objective

Begin with this mechanism: Information and spatial terms are combined with a declared coefficient. Work through the concrete case: Loss equals contrastive loss plus lambda times spatial loss. Make the limitation visible: Lambda controls the scientific answer. Complete the evidence step: Publish a sensitivity curve. 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: Embedding geometry

Begin with this mechanism: Distances and neighbourhoods in the learned space support downstream analysis. Work through the concrete case: A continuous arc connects developmental states. Make the limitation visible: Curves can be optimization artifacts. Complete the evidence step: Compare methods and held-out 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 14: Spatial domains

Begin with this mechanism: Clustering groups latent vectors into coherent tissue regions. Work through the concrete case: A cortex section separates layers. Make the limitation visible: Cluster count changes boundaries. Complete the evidence step: Use anatomy and stability. 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: Pseudo-spatiotemporal map

Begin with this mechanism: A diffusion or pseudotime ordering is combined with physical relationships. Work through the concrete case: Cells progress from a source region along a spatial route. Make the limitation visible: Root and graph choices change direction. Complete the evidence step: Repeat plausible roots. 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, corruption, encoder, spatial weight, clustering and root. Work through the concrete case: A student calculates one discriminator and spatial-loss term. Make the limitation visible: A trajectory is conditional evidence. Complete the evidence step: Write a spatiotemporal 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.