Why Science? | BayesSpace, Bayesian Spatial Clustering and Subspot-Resolution Evidence

Section 1 of 36

1. Separate measured spots from enhanced subspots

Spatial assays measure molecules at physical capture spots. BayesSpace can subdivide those spots computationally and estimate finer patterns, but the subspots are not newly measured locations. This distinction is the article’s anchor: enhanced resolution is inference supported by neighbours and expression, not additional molecules collected from tissue.

Contents · Next section

Section 2 of 36

2. Represent expression in fewer dimensions

The method works with a low-dimensional representation, commonly principal components, that summarises gene-expression variation. This reduces noise and computation. It also means feature selection, normalisation and component number determine which biological signals remain available for clustering.

Contents · Previous section · Next section

Section 3 of 36

3. Use spatial neighbours as prior information

Adjacent spots often share tissue structure. BayesSpace encodes that expectation so neighbouring labels tend to be coherent. The prior can recover weak boundaries, yet it can also smooth across real sharp transitions or small rare regions if given too much influence.

Contents · Previous section · Next section

Section 4 of 36

4. Cluster within a Bayesian model

Bayesian spatial clustering combines expression evidence with neighbourhood structure and uncertainty. The fitted labels describe domains under the chosen model and number of clusters. They are not automatically cell types, anatomical diagnoses or developmental lineages.

Contents · Previous section · Next section

Section 5 of 36

5. Sample with Markov chain Monte Carlo

BayesSpace uses Markov chain Monte Carlo to explore model states. Chains require enough iterations, burn-in and diagnostic checking. A returned label map is not proof of convergence; repeated chains and stable summaries are essential before interpreting subtle boundaries.

Contents · Previous section · Next section

Section 6 of 36

6. Enhance resolution computationally

For platforms such as Visium, the method can divide spots into subspots and infer finer patterns from the surrounding evidence. Enhancement can sharpen known structure, but several fine arrangements may explain the same coarse measurements. Validation must match the claimed spatial scale.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the primary demonstrations specific

The 2021 Nature Biotechnology paper introduced Bayesian spatial clustering and resolution enhancement and evaluated the method on spatial transcriptomics datasets. Those results show useful examples of coherent domains and enhanced patterns. Every new tissue still needs cluster, prior, convergence and validation checks.

Contents · Previous section · Next section

Section 8 of 36

8. Quality-control the expression matrix

Inspect counts, detected genes, mitochondrial signal, tissue coverage, background and batch. Technical gradients can become spatially smooth clusters. Plot quality metrics beside inferred domains and prespecify exclusions so damaged tissue does not define biology.

Contents · Previous section · Next section

Section 9 of 36

9. Verify coordinates and neighbourhoods

Check coordinate units, grid geometry, tissue holes, edges, rotation and missing spots. Neighbour relations should respect the physical array and tissue mask. Connecting across a tear or empty lumen can pull unrelated regions together.

Contents · Previous section · Next section

Section 10 of 36

10. Choose informative features

Select genes and principal components with a documented rationale. Too few components may erase subtle biology; too many can carry noise and batch. Compare plausible choices and report which boundaries persist. Feature stability is part of evidence, not merely preprocessing.

Contents · Previous section · Next section

Section 11 of 36

11. Choose the number of clusters

The requested cluster count strongly shapes the map. Examine a defensible range, hierarchy, stability and external landmarks. A larger number always creates more regions, but those extra colours do not automatically represent new tissue compartments.

Contents · Previous section · Next section

Section 12 of 36

12. Specify spatial priors

Report neighbourhood definition and parameters controlling spatial coherence. Perform sensitivity analyses with weaker and stronger smoothing. If a region appears only under a narrow prior, it should remain provisional until supported by independent markers or morphology.

Contents · Previous section · Next section

Section 13 of 36

13. Run multiple chains and seeds

Use several chains or runs, inspect traces or relevant diagnostics and compare label agreement. Save configurations and seeds. Consensus can identify stable cores, while disagreement may reveal poor mixing, weak data or a genuine transition zone.

Contents · Previous section · Next section

Section 14 of 36

14. Map uncertainty, not only labels

Posterior summaries or repeated-run stability should accompany the hard cluster map. Edge spots and low-count regions often carry greater uncertainty. A confidence layer helps readers distinguish a stable domain centre from an attractive but fragile boundary.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented clustering table

This fictional table is for interpretation only.

Contents · Previous section · Next section

Section 16 of 36

16. Interpret spatial clusters cautiously

A cluster can correspond to a tissue layer, tumour compartment, compositional zone or technical region. Describe marker genes, morphology and sample context before naming it. Avoid using cell-type language unless cell composition is independently established.

Contents · Previous section · Next section

Section 17 of 36

17. Interpret enhanced subspots cautiously

Subspot patterns can suggest where fine structure lies inside a capture spot. They should be presented with the original spot measurements and sensitivity maps. A sharply coloured subspot is a model allocation, not direct subcellular or single-cell observation.

Contents · Previous section · Next section

Section 18 of 36

18. Impute expression with clear labels

BayesSpace can estimate gene expression at enhanced resolution. Label these values as imputed or enhanced in figures and files. Validate predictions for held-out or orthogonally measured genes before using them to support a novel boundary.

Contents · Previous section · Next section

Section 19 of 36

19. Estimate composition only with supporting models

Cell-type composition can be analysed alongside enhanced structure, but each extra model adds assumptions. Propagate uncertainty rather than treating one inferred matrix as raw truth for the next step. Validate key populations with independent markers.

Contents · Previous section · Next section

Section 20 of 36

20. Replicate domains across specimens

A domain recurring across independent sections and subjects is stronger than a cluster with thousands of correlated spots in one sample. Align only comparable anatomy, show all specimens and use specimen-level statistics for generalisation.

Contents · Previous section · Next section

Section 21 of 36

21. Validate at the claimed resolution

Broad domains can be checked with histology and regional markers. Fine subspot claims require higher-resolution imaging, segmentation or another spatial assay. Validation at the original spot scale cannot prove a within-spot arrangement.

Contents · Previous section · Next section

Section 22 of 36

22. Challenge over-smoothing

Insert synthetic small regions or examine known sharp boundaries to see whether the prior erases them. Compare weaker spatial influence and expression-only clustering. A smooth map can be visually appealing while missing biologically important islands.

Contents · Previous section · Next section

Section 23 of 36

23. Challenge cluster-number dependence

Track how domains split or merge across candidate cluster counts. Report stable parent regions and uncertain subdivisions. Selecting the count that best matches a desired picture without independent evidence is circular.

Contents · Previous section · Next section

Section 24 of 36

24. Challenge edges and tissue gaps

Spots at edges have fewer neighbours, and holes can distort local support. Map degree, uncertainty and enhanced assignments near gaps. Use tissue masks so the model does not borrow evidence through empty space.

Contents · Previous section · Next section

Section 25 of 36

25. Challenge MCMC convergence

Longer runs, multiple chains and different starting states should give compatible broad conclusions. If labels switch or traces drift, increase sampling or simplify the claim. Computation ending successfully is not the same as statistical convergence.

Contents · Previous section · Next section

Section 26 of 36

26. State that imputation is not measurement

Enhanced values may be useful for hypothesis generation and visualisation, but they do not increase the number of captured molecules. Preserve original counts, disclose which values were imputed and avoid treating subspot sample size as independent experimental replication.

Contents · Previous section · Next section

Section 27 of 36

27. Compare BayesSpace with SpaGCN

BayesSpace uses Bayesian spatial clustering and MCMC, whereas SpaGCN combines expression, coordinates and histology in a graph-convolutional framework. Agreement can support broad domains. Differences may expose image influence, prior smoothing, feature choice or cluster resolution.

Contents · Previous section · Next section

Section 28 of 36

28. Keep negative spatial claims bounded

Failure to see a small region may reflect coarse spots, low counts or smoothing. Simulate regions of different size and contrast to estimate the detection envelope. State what broad structures were excluded and which fine alternatives remain unresolved.

Contents · Previous section · Next section

Section 29 of 36

29. Try a classroom model

Give students a coarse grid with noisy numbers and ask them to label regions using both values and neighbours. Then add one tiny island to test whether smoothing preserves or erases it. Ask students to state what was observed, what was inferred and what information remains unresolved. That three-column habit transfers directly to laboratory reports and prevents a plausible calculation from being mistaken for a new measurement.

Contents · Previous section · Next section

Section 30 of 36

30. Start the habit in Primary Science

Primary Science builds careful observation, comparison and fair testing. A child can learn the foundation of BayesSpace without advanced algebra by sorting evidence from explanation, checking repeated measurements and asking whether a model has enough information. Those habits matter in PSLE Science and much later in computational biology.

Contents · Previous section · Next section

Section 31 of 36

31. Strengthen PSLE Science reasoning

For PSLE Science, the useful link is not memorising software names. It is answering with evidence: identify the variable, read the table, connect cause and effect, and avoid a conclusion that exceeds the results. BayesSpace makes that familiar discipline visible in a modern research setting.

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, graphs and evaluation. BayesSpace adds a contemporary example in which gene counts, tissue position and computation meet. Students can practise interpreting axes, controls, uncertainty and alternative explanations while keeping syllabus fundamentals central.

Contents · Previous section · Next section

Section 33 of 36

33. Use Computing as a partner

Computing contributes data structures, algorithms, iteration, visualisation and reproducibility. Science contributes the question, samples, controls and biological interpretation. The productive boundary is clear: code can organise and test evidence, but it cannot rescue a weak reference or turn an assumption into an observation.

Contents · Previous section · Next section

Section 34 of 36

34. Choose schools by fit, not a single buzzword

Families comparing Singapore schools should verify current subject combinations, laboratories, applied learning, CCAs, mentoring and timetable fit on official school and MOE pages. A fashionable term such as BayesSpace is not evidence of programme quality. Look for sustained inquiry, sound teaching and opportunities to explain results.

Contents · Previous section · Next section

Section 35 of 36

35. See the career pathways

The reasoning behind BayesSpace appears in spatial genomics, biostatistics, medical imaging, computational pathology and scientific computing. Pathways can include polytechnic, junior college, university and continuing education, with different mixtures of biology, statistics, computing, engineering and communication. Careers depend on qualifications and experience; one school project does not promise an outcome.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a bounded claim

A careful conclusion keeps the central distinction visible: BayesSpace computationally clusters and enhances measured spots; subspots and imputed expression are not new physical measurements. Report the data, model, uncertainty, sensitivity and validation together. That measured language is not timid—it is how science remains useful when the result moves from a colourful figure toward a decision.

A publication-ready BayesSpace analysis starts with a spatial data contract. Record platform, spot geometry, physical dimensions, tissue mask, coordinate system, specimens, processing batches and image registration. The neighbourhood graph must respect the array and actual tissue. Display edges around holes, tears and borders so readers can see where spatial support is sparse or potentially misleading.

Pre-register quality filters, normalisation, feature genes, principal components, candidate cluster counts, spatial parameters, MCMC iterations, burn-in, chains, seeds, primary domain comparison and validation. Resolution enhancement adds another layer: specify subspot geometry, enhanced outputs and the assay that will test fine-scale claims. Separate exploratory maps from locked confirmation.

Publish a funnel from raw spots to tissue spots, quality-controlled counts, selected genes, principal components, modelled spots and enhanced subspots. Map every exclusion and count metric. Technical gradients are often spatially coherent; a Bayesian prior can reinforce them unless raw quality and batch are inspected beside the cluster map.

Feature sensitivity deserves formal analysis. Repeat clustering across defensible gene and principal-component choices. Track which broad domains remain stable and where boundaries move. Principal components compress biology and noise together, so the chosen representation is part of the evidence. Save component loadings and explained variation, not only the final labels.

Cluster number should be reported as a hierarchy. Examine several plausible values, quantify agreement and compare with independent tissue landmarks. More clusters inevitably produce more colours. A fine subdivision is credible only when its marker, morphology and replication evidence also becomes more specific, not merely because the algorithm can partition the data.

Spatial-prior sensitivity belongs in the main results. Compare weaker and stronger coherence and, where useful, expression-only clustering. Stable domain cores may persist while transition zones move. Report that layered outcome. Selecting the smoothest map risks erasing rare islands or sharp biological interfaces that do not conform to neighbouring majority labels.

MCMC diagnostics are not optional. Run multiple chains or independent starts, retain enough post-burn-in samples and inspect relevant traces or stability summaries. Label switching and poor mixing can produce changing maps even when code completes. Archive samples or sufficient summaries so uncertainty can be re-examined rather than reduced to one hard label.

Resolution enhancement must remain visually tethered to original measurements. Pair every subspot panel with the measured spot grid, counts and uncertainty. Use a different legend or explicit label for imputed expression. A fine-scale prediction can guide targeted imaging, but it should not be counted as new molecules, new independent observations or direct single-cell localisation.

Validate at matching scale. Broad clusters can be checked with anatomy, histology and regional markers. A within-spot arrangement needs higher-resolution imaging, segmentation or another spatial assay. Validation performed only at spot scale cannot establish which subspot held the signal. Include expected negatives and challenging edges, not only stable centres.

Biological replication remains specimen-based. Neighbouring spots, subspots and MCMC samples are not independent animals or patients. Show all specimens, balance processing and use specimen-level summaries for general claims. Adjacent sections help test technical repeatability, while independent subjects test biological generalisation.

Reproducibility materials include raw and processed matrices, coordinates, tissue masks, quality maps, features, component loadings, cluster candidates, neighbourhood settings, priors, chains, seeds, diagnostics, posterior summaries, enhanced outputs, validation data, package version, environment, code and checksums. Preserve original and imputed values in separate objects to prevent accidental mixing downstream.

Negative results need size-and-contrast calibration. Simulate spatial islands and gradients at different spot scales, expression contrasts and count depths, then estimate which structures survive smoothing and enhancement. Failure to detect a tiny region does not prove absence. It may show that the coarse assay and spatial prior could not resolve that configuration.

A careful final sentence might say: ‘Prespecified BayesSpace analyses identified broad domains reproducible across independent sections, while computational subspot enhancement generated testable fine-scale hypotheses supported for selected targets by higher-resolution assays.’ It should not describe imputed subspots as additional measured locations or let smoothness stand in for truth.

The most useful enhanced map therefore behaves like a hypothesis generator with a visible audit trail. Readers should be able to move backward from each subspot claim to its neighbouring measured spots, feature representation, prior, chain diagnostics and independent validation. That traceability is what turns computational detail into scientific evidence rather than decorative precision.

Contents · Previous section · Continue to the Science Learning Hub