Why Science? | SpaGCN, Graph Convolution, Histology and Spatial-Domain Evidence

Section 1 of 36

1. Represent tissue as a graph

SpaGCN treats each spatial measurement as a node and connects nodes through weighted edges. Nearby spots can influence one another, with histology helping define similarity. The graph is a scientific choice about neighbourhood, not a neutral copy of tissue, so its construction belongs in the evidence record.

Contents · Next section

Section 2 of 36

2. Combine three information streams

Gene-expression counts describe molecular state, spatial coordinates describe geometry and histology supplies image context. SpaGCN integrates them to seek coherent domains. Each stream can help the others, yet errors can also propagate: a stain artefact may alter graph weights and then sharpen a molecular boundary.

Contents · Previous section · Next section

Section 3 of 36

3. Use graph convolution

Graph convolution updates a node’s representation by combining its features with information from connected neighbours. This can strengthen signals that are spatially coherent and reduce isolated noise. It can also blur a real narrow transition if the neighbourhood is too broad or the graph crosses an anatomical barrier.

Contents · Previous section · Next section

Section 4 of 36

4. Learn a low-dimensional representation

The model compresses high-dimensional expression and graph context into features useful for clustering. A compact representation is easier to analyse, but biological meaning is not guaranteed by mathematical separation. Marker genes, morphology and independent specimens must interpret the clusters.

Contents · Previous section · Next section

Section 5 of 36

5. Call clusters spatial domains cautiously

SpaGCN groups spots into domains expected to be spatially coherent. A domain can correspond to a cortical layer, tumour compartment or other tissue region. It is not automatically a cell type, histological diagnosis or developmental lineage; its meaning depends on measured composition and validation.

Contents · Previous section · Next section

Section 6 of 36

6. Find domain-enriched genes

After domains are identified, the workflow can test genes associated with one region or with subpatterns inside it. Circular analysis is a risk if the same genes define and validate a domain. Use independent markers, held-out data or separate specimens for stronger confirmation.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the seven-dataset demonstration in scope

The 2021 Nature Methods paper evaluated SpaGCN across seven spatially resolved transcriptomics datasets and reported domain and spatial-gene findings. Breadth across datasets is encouraging, but every new platform and tissue still needs scale, image and parameter checks.

Contents · Previous section · Next section

Section 8 of 36

8. Prepare expression counts

Inspect library size, detected genes, mitochondrial signal, background and tissue coverage. Normalise and select features with a documented rationale. Technical gradients can become graph-smooth patterns, so compare candidate domains with raw quality maps before giving them biological names.

Contents · Previous section · Next section

Section 9 of 36

9. Register coordinates and histology

Spatial coordinates and the tissue image must share a reliable frame. Check rotation, scale, crop, distortion and missing tissue. An image pixel attached to the wrong spot can redirect graph weights. Publish registration landmarks and residuals rather than assuming vendor alignment is exact.

Contents · Previous section · Next section

Section 10 of 36

10. Inspect stain quality

Histology intensity reflects tissue structure but also illumination, section thickness, folds, tears and staining batch. Correct broad imaging artefacts carefully and keep the raw image. If a domain disappears when image information is reduced, ask whether histology added biology or noise.

Contents · Previous section · Next section

Section 11 of 36

11. Choose the histology scale parameter

SpaGCN uses spatial and histological information to define effective distances. The scale controlling histology’s contribution should be justified and tested. A setting that maximises cluster appearance may not maximise truth; evaluate against external landmarks and stability.

Contents · Previous section · Next section

Section 12 of 36

12. Construct adjacency weights

Edges may depend on physical distance and image-derived similarity. Report neighbour counts, distance rule, weight distribution and disconnected spots. Tissue holes and edges need special attention because an algorithm may connect across empty space or leave legitimate border regions isolated.

Contents · Previous section · Next section

Section 13 of 36

13. Train with repeated seeds

Neural models and clustering can depend on initialisation. Run several seeds, quantify agreement and map unstable spots. Consensus can be useful, but it should not hide genuine transitional regions that vary because the tissue lacks a sharp boundary.

Contents · Previous section · Next section

Section 14 of 36

14. Select domain number responsibly

More clusters produce finer maps but can split continuous variation; fewer clusters merge distinct anatomy. Use prior tissue knowledge cautiously, compare plausible resolutions and report the hierarchy. A result is stronger when the same broad compartments persist across settings.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented graph table

This fictional table separates molecular and image evidence.

Values are invented and are not performance benchmarks.

Contents · Previous section · Next section

Section 16 of 36

16. Interpret domain-guided differential expression

Genes enriched in one inferred domain can help describe tissue biology. Control for library size, multiple testing and specimen structure. Validation is strongest when enrichment appears in independent tissue and agrees with an assay not used to build the graph.

Contents · Previous section · Next section

Section 17 of 36

17. Read the pancreatic-cancer example carefully

The paper included spatial analysis of pancreatic cancer tissue, where molecular and histological structure can reveal compartments. Tumours are heterogeneous and clinical interpretation requires representative patients and pathology review. A domain in one section is not a diagnostic category by itself.

Contents · Previous section · Next section

Section 18 of 36

18. Read human dorsolateral prefrontal cortex layers

Cortical layers offer anatomical structure for evaluating spatial domains. Agreement with expert annotation supports the method in that dataset. Differences may reflect algorithmic error, ambiguous boundaries or biological gradients, so disagreement deserves inspection rather than automatic correction.

Contents · Previous section · Next section

Section 19 of 36

19. Read mouse-brain domains

Mouse brain contains recognisable regions and marker patterns useful for validation. Reproducing broad anatomy tests orientation and coherence. It does not establish accuracy for rare cells or subcellular signals, which require different evidence and resolution.

Contents · Previous section · Next section

Section 20 of 36

20. Seek recurring domains across sections

The paper also considered information across tissue sections. Repeated patterns can strengthen evidence when sections are independently processed and properly aligned. Section-to-section agreement should be quantified, while missing or damaged regions remain visible.

Contents · Previous section · Next section

Section 21 of 36

21. Transfer spatial genes with restraint

Genes associated with a domain in one dataset may be examined in another. Transfer requires matched anatomy, condition and measurement. A marker that follows one boundary in mouse brain may not define the same compartment in human disease tissue.

Contents · Previous section · Next section

Section 22 of 36

22. Challenge neighbour definitions

Vary radius, number of neighbours and whether edges cross tissue gaps. Broad conclusions should survive reasonable choices. If a narrow domain exists only under one graph, report it as parameter-dependent rather than selecting the most attractive map.

Contents · Previous section · Next section

Section 23 of 36

23. Challenge image artefacts

Mask folds, tears, dust and saturated regions, then rerun the analysis. Compare expression-only and image-informed graphs. A domain aligned with an artefact should fail this challenge; a robust biological region should retain molecular support.

Contents · Previous section · Next section

Section 24 of 36

24. Challenge spot mixtures

Many spatial spots contain multiple cells. Graph smoothing can make mixtures look like homogeneous tissue. Use deconvolution, marker co-expression or higher-resolution validation to determine whether a domain represents one population, a stable mixture or a transition.

Contents · Previous section · Next section

Section 25 of 36

25. Challenge over-smoothing

Monitor whether narrow boundaries or small regions disappear as graph influence increases. Simulations with known domains can estimate the smallest feature retained. A smooth map is not automatically a faithful map when biology is discontinuous.

Contents · Previous section · Next section

Section 26 of 36

26. Compare SpaGCN with Tangram

SpaGCN discovers domains from spatial expression, coordinates and histology; Tangram aligns external single-cell profiles to spatial data. One finds graph-coherent regions, the other transfers cell and gene information. Combining them can help interpretation but also compounds assumptions.

Contents · Previous section · Next section

Section 27 of 36

27. Compare SpaGCN with Squidpy

SpaGCN is a specific graph-convolutional modelling approach. Squidpy is a broader analysis framework with tools for graphs, neighbourhoods, spatial statistics and images. The choice depends on question; software breadth and model specificity are different virtues.

Contents · Previous section · Next section

Section 28 of 36

28. Treat null findings honestly

Failure to find a stable domain can mean gradual biology, insufficient capture, low resolution, poor histology or unsuitable graph settings. Report sensitivity and the smallest simulated region recoverable. Do not manufacture clusters merely because a pipeline expects them.

Contents · Previous section · Next section

Section 29 of 36

29. Build a paper-grid graph

Students place coloured gene counts on a grid, connect neighbours and compare clusters before and after adding a shaded ‘histology’ layer. An intentionally misleading stain patch shows how another data source can improve or misdirect a model.

Contents · Previous section · Next section

Section 30 of 36

30. Connect to Primary Science patterns

Young learners can identify neighbours, observe colour patterns and distinguish a measured square from a grouped region. They can explain why edge squares have fewer neighbours and why one odd square should be checked before a new conclusion.

Contents · Previous section · Next section

Section 31 of 36

31. Strengthen PSLE Science reasoning

A fictional graph table supports comparison, reliability and fair testing. Pupils can identify the unstable spot, propose repeating the model without the stain and state a conclusion that recognises a transition instead of forcing a boundary.

Contents · Previous section · Next section

Section 32 of 36

32. Extend into Secondary and O-Level Science

Biology contributes tissues and gene expression; Mathematics contributes coordinates and graphs; Computing contributes algorithms and parameter tests; Chemistry contributes staining. The example makes interdisciplinary reasoning concrete while official syllabuses define assessed content.

Contents · Previous section · Next section

Section 33 of 36

33. Use the topic for school choices

Students interested in science, mathematics or computing can verify current school offerings and facilities through official sources. Ask how projects are supervised and evaluated. This article makes no unverified programme, admission or outcome claim.

Contents · Previous section · Next section

Section 34 of 36

34. See the career pathways

Spatial graph analysis brings together histotechnologists, molecular scientists, pathologists, computer scientists and statisticians. Each protects a different part of the evidence chain. Career routes require current qualifications and experience; one method does not guarantee entry.

Contents · Previous section · Next section

Section 35 of 36

35. Did You Know? Neighbours can change a spot’s identity

In graph convolution, a spot is represented partly through connected neighbours. That is powerful when tissue is coherent, but it means graph design can move a boundary. Testing several graphs is a scientific control, not merely a software preference.

Contents · Previous section · Next section

Section 36 of 36

36. Keep the final claim bounded

A defensible SpaGCN conclusion names expression quality, coordinate registration, histology processing, graph definition, scale parameters, cluster stability, spatial resolution, validation and independent specimens. It may report a reproducible spatial domain. It should not turn graph smoothness into cellular purity or one tumour section into clinical truth.

A strong SpaGCN study treats graph construction as part of the experiment. Archive spot coordinates, physical units, tissue masks, image registration, histology preprocessing, neighbour rules, distance distributions and the scale controlling image contribution. A graph should be inspectable on top of the tissue, with edges across gaps or damaged regions visibly flagged.

Pre-register quality filters, normalisation, features, graph parameters, target number or range of domains, random seeds, stability metric, primary anatomical comparison and validation. Exploratory resolutions can reveal structure, but a confirmatory boundary should be tested on an independent section with the chosen settings frozen.

Publish a funnel from raw spots through tissue filtering, molecular quality, image quality, graph connectivity, model fitting and final domain analysis. Map every excluded spot. Report isolated nodes, unusually high-degree nodes and regions where histology was unavailable. Domain interpretation is unsafe when missingness silently outlines the cluster.

Histology needs its own quality report: stain protocol, scanner, magnification, illumination, colour normalisation, artefact masks and registration residuals. Compare raw and processed images. Tissue colour can encode real structure, but it can also encode batch, folds or fading. Image information should earn influence rather than receiving it automatically.

Graph sensitivity should vary neighbour scale, histology weight and tissue-gap rules. Quantify domain agreement and map unstable spots. Core regions may remain consistent while boundaries move; report both. Selecting the graph that most closely resembles an expected atlas without independent testing is circular validation.

Model stability should include repeated seeds and plausible cluster resolutions. Use consensus carefully: disagreement can signal a continuous transition rather than algorithmic failure. Report domain hierarchies when broad compartments are stable but fine subdivisions are not. A single hard label per spot is often less informative than stability and membership evidence.

Spatially variable gene analysis must avoid double use of data. If domains were learned from the same genes, enrichment partly reflects the construction process. Confirm leading genes in held-out sections, external datasets or orthogonal assays. Correct multiple testing and report effect sizes, not only significance produced by many spots.

Biological replication should operate at animal or patient level. Spatial spots are correlated within sections, and several sections from one specimen are not independent people. Balance tissue processing across groups and show domain prevalence and gene effects for every specimen. A domain recurring across independent sections and subjects is much stronger than one with thousands of internal spots.

Validation should combine anatomy, markers and perturbation where relevant. Known layers or pathology can test broad domains, while independent RNA or protein assays test boundaries and genes. If histology informed the graph, do not use the identical image feature as the only validation. Choose evidence not already embedded in the model.

Figures should show raw histology, expression quality, graph edges, domain maps, stability, boundary uncertainty, spatial genes and specimen-level summaries. Include expression-only and image-informed comparisons. Pair a representative image with all-specimen metrics so readers can separate visual illustration from general evidence.

Reproducibility materials include raw counts, coordinates, full-resolution images, tissue masks, transformations, graph matrices, preprocessing, software version, environment, seeds, embeddings, domain labels, gene tests, code and checksums. Preserve alternative graph results and failed runs. Future reviewers need to know whether a conclusion was robust or selected from many variants.

Negative findings require simulations at the observed spot density and noise. Insert domains of varying size, contrast and boundary sharpness to estimate what the workflow could recover. A null result may exclude a broad high-contrast compartment while leaving a thin transition or rare microdomain unresolved. Report the scale explicitly.

Transfer to a new tissue demands new image and graph validation. Cortical layers, tumour nests and glandular tissue have different topology and stain behaviour. Re-establish neighbourhood scale, artefact handling, resolution and marker evidence rather than carrying over convenient defaults. The same code can produce technically valid but biologically inappropriate clusters.

Quality monitoring should continue across processing dates. Track molecular yield, image brightness, registration residuals, graph degree and domain stability using a reference section or controlled sample. If one treatment group was processed later, repeat a balanced subset before interpreting a new domain. Temporal drift can be spatially smooth enough to look biological.

A careful final sentence might say: ‘With registered histology, quality-controlled expression and a prespecified spatial graph, SpaGCN identified domains that were stable across parameter settings and independent specimens and agreed with orthogonal markers.’ It should not equate a graph cluster with a pure cell type, a smooth boundary with direct anatomy or one patient section with clinical generalisation.

Contents · Previous section · Continue to the Science Learning Hub