Why Science? | SpaOTsc, Optimal Transport and Spatial-Signalling Relationship Evidence

Section 1 of 36

1. Begin with the lost-location problem

Single-cell RNA sequencing can profile many genes while dissociation removes each cell’s original position. SpaOTsc asks whether a spatial assay measuring fewer genes can help reconstruct relationships among those dissociated cells. The output is an inferred spatial arrangement, not a recovered photograph.

Contents · Next section

Section 2 of 36

2. Think in distributions, not fixed pairs

Optimal transport moves probability mass between source and target distributions while minimising a cost. Here the source and target represent cells or spatial measurements. A coupling spreads probability across alternatives rather than declaring one certain match.

Contents · Previous section · Next section

Section 3 of 36

3. Use structured optimal transport

SpaOTsc incorporates relational structure so the mapping considers both expression similarity and spatial organisation. The structure helps when exact one-to-one labels are absent. Its cost definitions and regularisation determine which arrangements are favoured.

Contents · Previous section · Next section

Section 4 of 36

4. Map single cells to spatial measurements

Shared genes link dissociated profiles with spatial reference measurements. A cell may map to several positions with different weights. Reference resolution, missing states and platform differences bound the mapping.

Contents · Previous section · Next section

Section 5 of 36

5. Infer cell–cell spatial distances

Once the model creates a spatial relationship among cells, it can estimate pairwise distances for downstream questions. These distances are derived from the model and reference, not measured physical separations between the sequenced cells.

Contents · Previous section · Next section

Section 6 of 36

6. Constrain communication by inferred space

Candidate signalling relationships can be filtered or weighted by spatial proximity. This can remove implausibly distant pairs and highlight local hypotheses. It still does not observe ligand release, receptor binding or response.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the demonstrated scope specific

The 2020 Nature Communications study described mapping, distance reconstruction, spatial subclustering, space-constrained communication and gene–gene information-flow analysis. Results from its datasets establish feasibility, not universal accuracy across tissues, panels or reference designs.

Contents · Previous section · Next section

Section 8 of 36

8. Match cell and spatial gene identifiers

Harmonise gene symbols, versions, duplicates and species mapping before transport. Preserve the complete overlap list. A small or biased shared panel can make unrelated cells look similar because the features that distinguish them were never measured spatially.

Contents · Previous section · Next section

Section 9 of 36

9. Quality-control both modalities separately

For single-cell data, map counts, detected genes, doublets and batch. For spatial data, map coverage, segmentation, background and tissue boundaries. Integration cannot rescue poor inputs and may spread one modality’s artefacts into the other.

Contents · Previous section · Next section

Section 10 of 36

10. Choose expression transformations transparently

Record normalisation, scaling, feature selection and any imputation. Transport costs operate on the processed features. If reconstructed or denoised values enter the cost, they should not later be treated as independent validation.

Contents · Previous section · Next section

Section 11 of 36

11. Define the cost matrix

Explain how molecular dissimilarity and spatial structure contribute to cost. Report weights, metric and scaling. A distance dominated by highly variable genes may ignore subtle markers, while a strong structural term can force overly smooth arrangements.

Contents · Previous section · Next section

Section 12 of 36

12. Represent uncertainty in the coupling

Keep the full transport plan or calibrated summaries rather than only the highest-weight match. Entropy, alternative locations and sensitivity reveal ambiguity. Hard labels hide information precisely where the data do not identify a unique solution.

Contents · Previous section · Next section

Section 13 of 36

13. Separate prior knowledge from evaluation

If known anatomy, cell labels or marker genes tune costs, they cannot independently validate the same mapping. Reserve held-out markers, images or specimens for evaluation and label tuned analyses honestly.

Contents · Previous section · Next section

Section 14 of 36

14. Archive solver details

Save regularisation, convergence tolerances, iterations, random seeds, numerical diagnostics, source and target weights, software commit and hardware. A returned coupling is not automatically a well-converged or unique solution.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented transport table

This fictional table teaches uncertainty, not method performance.

Contents · Previous section · Next section

Section 16 of 36

16. Read a mapped location probabilistically

A high transport weight means one location is favoured under the chosen inputs and costs. It is not proof that the dissociated cell physically occupied that exact coordinate. Report regional resolution appropriate to uncertainty.

Contents · Previous section · Next section

Section 17 of 36

17. Use inferred distances as model outputs

Distance estimates can support clustering or communication hypotheses. Compare them with any available known neighbours, spatial markers or synthetic truth. Do not quote derived distances as microscope measurements.

Contents · Previous section · Next section

Section 18 of 36

18. Distinguish proximity from communication

Cells can be near without signalling and communicate over distance through diffusible factors. Ligand and receptor expression are prerequisites, not evidence of transfer. Protein, perturbation and response data are needed for stronger mechanism.

Contents · Previous section · Next section

Section 19 of 36

19. Treat gene–gene flow as a candidate network

The framework can propose intercellular regulatory relationships. Directional arrows summarise a model. They should guide experimental tests rather than become a finished pathway diagram.

Contents · Previous section · Next section

Section 20 of 36

20. Check missing states and reference bias

A cell state absent from the spatial reference may be forced into familiar regions. Hold out known populations or add an unresolved category to test how the method behaves when the target space is incomplete.

Contents · Previous section · Next section

Section 21 of 36

21. Replicate across specimens and panels

Repeat mapping with independent tissue and, where possible, a different spatial gene panel. One reference can encode specimen-specific anatomy. Generalisation requires more than many cells from one organism.

Contents · Previous section · Next section

Section 22 of 36

22. Audit feature-overlap sensitivity

Recompute mappings with core shared genes, markers removed and one broader set. Track regional probabilities and signalling conclusions. A result that depends on a handful of markers should say so.

Contents · Previous section · Next section

Section 23 of 36

23. Audit cost-weight sensitivity

Vary molecular and structural weights over a prespecified range. Stable cells are easier to interpret; moving cells identify weakly constrained regions. Avoid choosing the setting solely for resemblance to expected anatomy.

Contents · Previous section · Next section

Section 24 of 36

24. Audit mass and regularisation choices

Cell and region weights plus transport regularisation affect diffuse versus concentrated couplings. Publish a small sensitivity grid and convergence information. Visual sharpness is not synonymous with accuracy.

Contents · Previous section · Next section

Section 25 of 36

25. Validate with held-out spatial evidence

Use genes, protein staining or imaging excluded from fitting. Evaluate clear, ambiguous and negative examples. Validation should reflect the spatial resolution of the claim rather than demand exact coordinates the assay cannot support.

Contents · Previous section · Next section

Section 26 of 36

26. Compare with simpler label transfer

Nearest-centroid or correlation mapping offers a baseline under the same shared genes. Structured transport should add reproducible spatial organisation or communication constraints, not merely a more sophisticated display.

Contents · Previous section · Next section

Section 27 of 36

27. Report non-identifiability openly

Several couplings may explain the same observed panel. If alternatives persist across settings, report a set of plausible regions or networks. Scientific honesty sometimes means refusing a single map.

Contents · Previous section · Next section

Section 28 of 36

28. Design perturbations from the inferred network

Select a ligand–receptor or region pair whose predicted response differs from a competing model. Perturb the sender, block the receptor or measure time-resolved response. Such tests can move an optimal-transport hypothesis toward mechanism.

Contents · Previous section · Next section

Section 29 of 36

29. Build a classroom model before code

Give students four coloured source cards and three target boxes with partial clues. Ask them to distribute probability rather than force one card into one box, then change the cost rule. Keep measured values, derived quantities, assumptions and conclusions in separate columns. The goal is not to imitate specialist software; it is to make the evidence chain visible enough to question.

Contents · Previous section · Next section

Section 30 of 36

30. Start with Primary Science habits

Primary Science already teaches the habits beneath SpaOTsc: observe carefully, compare fairly, record consistently and avoid a conclusion larger than the test. A simple map of plant height, shade or water can show that location changes interpretation without introducing advanced statistics.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science to practise claim limits

PSLE Science rewards links between observation, process and explanation. SpaOTsc offers a modern reminder that a pattern is not automatically a cause. Students can ask what was changed, what was measured, what stayed uncontrolled and what extra observation would distinguish competing explanations.

Contents · Previous section · Next section

Section 32 of 36

32. Connect Secondary and O-Level Science

Secondary Science and O-Level Biology develop cells, organisation, variation, experimental design and evaluation. SpaOTsc adds a contemporary case where molecular measurements and tissue position meet. The useful learning outcome is controlled reasoning, not memorising a package name.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing support the science

Computing contributes data structures, algorithms, visualisation, optimisation and reproducibility. Biology supplies the specimen, mechanism and validation. SpaOTsc shows why reliable code is necessary but insufficient: an algorithm can execute perfectly on mismatched coordinates or a biased sample.

Contents · Previous section · Next section

Section 34 of 36

34. Choose school opportunities by fit

Families should verify Biology laboratories, data projects, Computing, mathematics enrichment and research mentoring on current official school and MOE pages. A school mention of genomics, AI or data science does not guarantee depth, admission or a career outcome. Look for sustained inquiry, careful teaching, accessible mentoring and time to explain evidence.

Contents · Previous section · Next section

Section 35 of 36

35. See the connected career families

Reasoning used in SpaOTsc appears in single-cell genomics, systems biology, optimal transport, bioinformatics, tissue imaging and quantitative medicine. Routes can pass through polytechnic, junior college, university or continuing education with different blends of biology, mathematics, statistics, computing and communication. Current course requirements must be checked directly; one project cannot guarantee entry or employment.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a bounded scientific claim

A defensible conclusion says exactly what SpaOTsc estimated and under which inputs, settings and specimens. SpaOTsc infers mappings, distances and communication candidates under a structured transport model; it does not recover exact lost coordinates or directly observe signalling. Report uncertainty, sensitivity, replication, negative results and orthogonal validation together. That boundary is not weakness; it is what makes the result testable and reusable.

A defensible SpaOTsc project begins with two linked data contracts. For the dissociated single-cell matrix, record specimens, batches, filtering, doublet handling, gene identifiers and cell labels. For the spatial reference, record technology, measured genes, coordinates, resolution, masks and segmentation. Preserve the exact shared-feature matrix supplied to the transport problem.

Pre-register the mapping question and expected resolution. A region-level reconstruction, a cell-distance estimate and a communication network are different claims. Define the primary output, evaluation set and acceptable uncertainty before tuning the transport to resemble known anatomy.

The cost function deserves a visible audit. State molecular distance, structural distance, scaling, feature weights and regularisation. Plot representative cost distributions. If one modality dominates numerically, the combined model may act like expression-only transfer or geometry-only placement despite a balanced verbal description.

Keep soft transport probabilities through analysis. Publish entropy, alternative regions and calibration where possible. A hard maximum-probability map is convenient for a figure but can erase the central fact that several spatial arrangements explain a limited gene panel.

Test reference incompleteness by withholding a known state or region. Observe whether those cells become diffuse, map to the nearest familiar location or receive confident but wrong assignments. This stress test shows how the model behaves when real tissue contains biology absent from the reference.

Feature sensitivity is essential. Repeat representative mappings with highly expressed genes removed, marker genes withheld and one broader shared set. If a region depends on two known markers, report that narrow support and validate them independently instead of calling the whole transcriptome spatially recovered.

Communication analysis needs separate evidence. Publish ligand, receptor and response-gene assumptions, spatial-range settings and negative controls. Compare against a non-spatial network and a distance-shuffled network. A communication edge is a prioritised experiment, not a detected molecular transmission.

Validate at the correct scale. If the spatial reference contains spots larger than cells, do not demand or advertise exact single-cell coordinates. Use held-out spatial genes, protein imaging or anatomy to evaluate regions and neighbourhoods at the resolution the assay can actually support.

Replicate at the specimen level and report heterogeneity. Mapping thousands of cells to one reference does not establish population generality. A second specimen, a different spatial panel or a held-out tissue section provides a stronger test than another random split of the same cells.

Reproducibility materials include raw and processed matrices, shared gene list, source and target weights, cost matrices, transport plan, regularisation, convergence diagnostics, seeds, mapping summaries, inferred distances, communication networks, software commit, environment and checksums. Save alternative solutions and negative results.

Did you know? Optimal transport does not literally move cells back into tissue. It moves probability mass between mathematical distributions under a cost rule. That distinction is exactly why the method can integrate different assays—and why its output needs uncertainty and biological validation.

A bounded conclusion might say: ‘SpaOTsc produced region-level cell mappings and space-constrained communication hypotheses that remained stable across shared-feature, cost and regularisation checks and agreed with held-out spatial evidence.’ It should not describe the mapping as recovered ground-truth coordinates or inferred signalling as observed communication.

A revealing classroom demonstration is to map the same cells against two references: one complete and one missing a known region. Students can compare transport entropy, alternative destinations and downstream communication networks. The demonstration explains why an algorithm can always return an answer while the scientific question may remain underdetermined. Confidence should fall when the reference cannot represent the biology being queried.

Mass balance also needs biological interpretation. Standard transport formulations distribute source and target mass according to supplied weights, yet tissues can differ in cell abundance, sampling efficiency and missing regions. Publish the weights, explain whether they represent cells, spots or normalised mass, and test an alternative weighting scheme. A visually plausible map that depends entirely on arbitrary mass choices is a sensitivity result, not a robust reconstruction.

When evaluating spatial signalling, separate three links: ligand expression in a putative sender, receptor expression in a putative receiver and a compatible response in the receiver. Distance can prioritise these links, but it cannot establish secretion, binding or causal response. Orthogonal protein localisation, receptor perturbation or a time course provides a much stronger bridge from computational network to biological claim.

An evidence ladder helps families and students understand the result. A soft mapping supported by held-out genes is stronger than a colourful projection alone; replication in a second specimen is stronger still; independent imaging or perturbation can move a hypothesis toward mechanism. This is the optimistic power of data integration: not instant certainty, but a clear route to the next decisive experiment.

Before publication, include a compact decision log: why this spatial reference was selected, which cell states were excluded, how shared genes were chosen, what regularisation range was tested and which validation result could have changed the conclusion. Decision logs make analytical judgement visible. They also help a future team reproduce the reasoning, not just rerun the code, when a new reference panel or tissue atlas becomes available.

Contents · Previous section · Continue to the Science Learning Hub