Why Science? | NicheCompass, Signalling Gene Programs and Spatial-Niche Atlas Evidence

Section 1 of 36

1. Start with the tissue niche

A niche is a spatially organised cellular environment whose neighbours, molecular programmes and tissue context matter together. NicheCompass learns a low-dimensional representation intended to capture such structure across one or many spatial samples.

Contents · Next section

Section 2 of 36

2. Expand the method’s name

NicheCompass stands for Niche Identification based on Cellular grapH Embeddings of COMmunication Programs Aligned across Spatial Samples. The name captures three commitments: graph structure, interpretable signalling programmes and cross-sample alignment.

Contents · Previous section · Next section

Section 3 of 36

3. Build the spatial graph

Cells or spots become nodes and spatial proximity creates edges. Gene expression supplies node features, with optional chromatin accessibility and sample covariates. The graph defines who can influence a neighbourhood; it is a model of opportunity.

Contents · Previous section · Next section

Section 4 of 36

4. Align latent features with gene programs

Latent dimensions are linked to prior signalling gene programs. This gives the embedding biological handles instead of leaving every dimension anonymous. Prior alignment can also import database bias, so programme genes and provenance must be visible.

Contents · Previous section · Next section

Section 5 of 36

5. Use a conditional variational graph autoencoder

The model combines graph neural-network encoding with a multimodal conditional variational framework. It learns probabilistic latent representations while accounting for covariates. A latent coordinate is a compressed model feature, not a directly measured biological quantity.

Contents · Previous section · Next section

Section 6 of 36

6. Reconstruct neighbours and omics

A graph decoder reconstructs neighbourhood adjacency, while masked linear decoders reconstruct a cell’s own and aggregated-neighbourhood molecular features. For a signalling programme, ligand information can be associated with neighbours while receptor and target information belongs to the focal node.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the atlas claims grounded

The paper evaluated mouse organogenesis, human breast and lung cancer, multimodal mouse brain and large atlas integrations. These demonstrations establish scope in reported datasets. They do not make every transferred niche label correct in a new laboratory or tissue.

Contents · Previous section · Next section

Section 8 of 36

8. Build a sample manifest

Record organism, organ, section, condition, donor, assay, resolution, coordinates, preprocessing, batch and covariates. Multi-sample integration is only interpretable when these differences are declared rather than silently removed.

Contents · Previous section · Next section

Section 9 of 36

9. Choose the spatial graph deliberately

Document neighbour number or radius, coordinate units, boundaries and graph symmetrisation. Compare plausible settings. A spot-level graph may mix cells, while a single-cell graph may still inherit segmentation error.

Contents · Previous section · Next section

Section 10 of 36

10. Freeze signalling gene programs

Archive ligands, receptors, targets, species mapping and database versions. Show which genes are observed in the assay and which are missing. A programme cannot be interpreted from its label alone.

Contents · Previous section · Next section

Section 11 of 36

11. Add chromatin data cautiously

Optional accessibility can strengthen multimodal interpretation when matched and quality controlled. State whether modalities come from the same cells or aligned samples, how peaks map to genes and how missing modalities are handled.

Contents · Previous section · Next section

Section 12 of 36

12. Encode covariates honestly

Sample covariates help distinguish shared niche biology from technical or condition-specific variation. Overcorrection can remove real differences; undercorrection can turn batch into a niche. Show embeddings coloured by both biology and technical factors.

Contents · Previous section · Next section

Section 13 of 36

13. Protect validation units

Split by specimen or section, not random neighbouring nodes, when testing generalisation. Spatial autocorrelation makes random holdouts optimistic. A mapped query sample is a stronger test than rediscovering the training atlas.

Contents · Previous section · Next section

Section 14 of 36

14. Archive optimisation and pruning

Preserve code version, latent size, graph layers, losses, regularisation, seeds, pruning rules and stopping criteria. The method can prune redundant programmes; report which were removed and why.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented niche table

This table is fictional and not NicheCompass output.

Contents · Previous section · Next section

Section 16 of 36

16. Read embeddings as maps

Nearby points in latent space have similar model representations. Distance is not automatically a biological effect size. Inspect which programmes, genes and neighbourhood features explain a separation before naming a niche.

Contents · Previous section · Next section

Section 17 of 36

17. Interpret programme activity

A high latent programme activity indicates that the fitted model uses the aligned gene programme strongly at that location. It is not proof that each ligand bound its receptor or that every target changed causally.

Contents · Previous section · Next section

Section 18 of 36

18. Inspect programme genes

For each headline programme, publish positive and negative gene weights, observed expression and spatial maps. A label such as ‘inflammatory’ can hide several distinct molecular patterns.

Contents · Previous section · Next section

Section 19 of 36

19. Read niche boundaries

Compare inferred boundaries with histology, known anatomy and independent markers. Smooth latent transitions may reflect biology; sharp boundaries may reflect sectioning or graph construction. Keep both possibilities in view.

Contents · Previous section · Next section

Section 20 of 36

20. Map query samples carefully

Query mapping can fine-tune a new sample against a reference and support label transfer. Report mapping uncertainty, reference coverage and out-of-distribution regions. A forced nearest label is not evidence that the query truly matches.

Contents · Previous section · Next section

Section 21 of 36

21. Use uncertainty during transfer

Classifier probabilities or related uncertainty measures can flag ambiguous locations. Publish them rather than only the winning label. Low confidence may identify transitional biology or a missing reference niche.

Contents · Previous section · Next section

Section 22 of 36

22. Audit prior-program dependence

Repeat with high-confidence programme subsets, alternate databases or selected ablations. If a niche disappears when one uncertain programme is removed, label it prior-sensitive and plan targeted validation.

Contents · Previous section · Next section

Section 23 of 36

23. Audit graph sensitivity

Vary neighbour definitions and boundary constraints. Stable niches should not depend on one arbitrary radius unless there is a clear biological reason. Show adjusted agreement metrics and spatial differences.

Contents · Previous section · Next section

Section 24 of 36

24. Audit sample alignment

Examine whether alignment mixes corresponding biological states while preserving genuine condition differences. Compare with unintegrated data and simpler integration baselines. Integration is successful when it clarifies comparable structure without erasing the question.

Contents · Previous section · Next section

Section 25 of 36

25. Audit modality contribution

Fit expression-only and multimodal variants. Report where accessibility adds reproducible information and where it changes labels without support. More modalities do not guarantee more truth.

Contents · Previous section · Next section

Section 26 of 36

26. Audit scale claims

Large demonstrations—including million-cell and multi-million-cell integrations—show computational reach under reported hardware and data. Scalability is not the same as biological validity; a very large atlas still needs specimen-aware evaluation.

Contents · Previous section · Next section

Section 27 of 36

27. Design a niche-specific experiment

Choose a recurring programme at a defined spatial boundary. Confirm ligand, receptor and target proteins in the correct cells, perturb the route and measure whether the niche marker or response changes locally across time.

Contents · Previous section · Next section

Section 28 of 36

28. Write the bounded conclusion

Prefer: ‘NicheCompass learned a recurring spatial niche associated with this signalling gene programme across these samples.’ Avoid saying the latent programme proved cell–cell signalling.

Contents · Previous section · Next section

Section 29 of 36

29. Begin with Primary Science habitats

A habitat lesson already asks how an organism’s surroundings affect it. A tissue niche is a microscopic version of that systems question, with careful limits on what proximity can prove.

Contents · Previous section · Next section

Section 30 of 36

30. Use PSLE Science for classification

Students can group observations by shared features, then test whether a new example belongs. This introduces clustering, reference mapping and uncertainty in an accessible way.

Contents · Previous section · Next section

Section 31 of 36

31. Connect Secondary and O-Level Biology

Tissues, cell specialisation, receptors, gene expression and homeostasis supply the biological language. Spatial niches show how cell identity and location interact.

Contents · Previous section · Next section

Section 32 of 36

32. Let Computing explain latent space

Graphs, encoders, decoders, dimensionality reduction, classification and validation are rich Computing topics. Students should learn that a compact representation is useful because it preserves chosen information, not because it is reality itself.

Contents · Previous section · Next section

Section 33 of 36

33. Use science tuition for source literacy

A strong science lesson traces a claim to a primary paper, code repository, database and experiment. NicheCompass makes source provenance especially important because priors are built into the representation.

Contents · Previous section · Next section

Section 34 of 36

34. Explore atlas careers

Spatial atlas projects need molecular biologists, imaging specialists, computational scientists, software engineers and data curators. Career pathways are broad and collaborative, with communication as important as coding.

Contents · Previous section · Next section

Section 35 of 36

35. Consider school opportunities accurately

Families can seek verified opportunities in biology, computing, mathematics, research and teamwork. Current programmes and admissions information must be checked on official school sources; this article makes no school-specific promise.

Contents · Previous section · Next section

Section 36 of 36

36. End with the atlas question

The most useful question is ‘Where does the reference stop being a good explanation?’ Ambiguous query regions and unstable programmes are not failures—they are invitations to discover a new or transitional niche.

NicheCompass combines two ambitions: scalable atlas integration and interpretable latent biology. The signalling-program alignment is meant to make the embedding readable, while the variational graph architecture can integrate many samples. These strengths should be evaluated separately: an embedding can scale well yet attach an unstable programme label, or interpret well in one tissue yet generalise poorly.

The masked omics decoders are central to interpretability. By constraining how genes connect to latent programmes, they expose which ligand, receptor and target sets carry a feature. Publish the relevant weights and spatial expression maps. A programme name without its genes is too easy to overinterpret.

The paper demonstrated integration at very large scale, including a central-nervous-system atlas of roughly 1.09 million cells and a larger integration reported across 239 tissue sections. Those figures describe specific demonstrations in the study. They should not be turned into a blanket guarantee about every hardware setup, dataset or biological conclusion.

Query mapping should include an abstention mindset. If the reference lacks the query’s biology, forcing a label can hide novelty. Use uncertainty, distance to reference support and independent markers to identify regions that should remain unassigned. A new niche often first appears as an honest failure to map.

Did you know? Pruning redundant gene programmes can improve clarity, but it also changes the vocabulary available to the model. Archive both initial and retained programmes, and show why a programme was removed. Otherwise, readers cannot tell whether absence from the result reflects biology, low data support or a pruning decision.

For students, an atlas is a wonderful metaphor with a caution: maps are designed representations. They preserve some relationships and omit others. NicheCompass makes that lesson concrete—coordinates, graph neighbours, latent dimensions and labels are all choices whose quality is tested by whether they support repeatable observations and useful experiments.

Quality control should be spatial as well as molecular. Map library size, detected genes, mitochondrial signal, segmentation area and batch across every section. A technical gradient can look like a niche because neighbouring locations share preparation artefacts. Include these variables in diagnostic plots and, where appropriate, as covariates or exclusion criteria.

Gene-program construction requires a visible audit trail. Record the databases and versions, ligand–receptor direction, target-gene evidence, species mapping and any manual curation. Summarise how many programme genes are present in each assay. A programme with only a few observed genes should not receive the same confidence as one measured comprehensively.

Cross-sample alignment must preserve meaningful asymmetry. A tumour niche that genuinely occurs in one condition should not be forced to overlap a healthy reference. Inspect sample mixing within known shared cell states while separately testing retention of condition-specific structure. Alignment quality is a balance, not maximum mixing.

Latent dimensions can correlate with cell-type composition, tissue layer, pathology and technical depth simultaneously. Regressing out every association may erase biology, whereas naming a dimension from one correlation may oversimplify it. Use programme weights, spatial maps, independent markers and specimen recurrence together before attaching a niche label.

Reference mapping needs out-of-distribution checks. Calculate whether query representations fall inside the support of reference samples and identify programmes absent from the reference. Allow uncertain regions to remain unlabelled. An honest unknown category is scientifically richer than a confident but forced match.

Multimodal analyses should report modality-specific missingness. Accessibility and expression may differ in coverage, noise and spatial resolution. Show whether a programme remains when each modality is removed and whether one modality dominates the embedding. Concordance can strengthen interpretation; disagreement can reveal temporal or technical differences.

Large atlases require hierarchical validation. First test local neighbourhood reconstruction, then specimen-level niche recurrence, then organ or condition-level alignment, and finally query mapping. A single global embedding metric cannot verify every scale. Match each claim to a metric and an independent biological check.

Software versions matter because scalable frameworks evolve. Cite the peer-reviewed 2025 method for its reported architecture and demonstrations, and cite the exact repository or documentation version for current implementation behaviour. Do not project later defaults backward into the paper or assume a tutorial setting is universally optimal.

End with a niche evidence card: spatial boundary, programme genes, modality support, specimen prevalence, graph sensitivity, mapping uncertainty, independent markers and planned perturbation. This compact record keeps the atlas navigable and ensures that every colourful region points toward a testable biological question.

Niche naming should be conservative and compositional. A label such as ‘vascular-inflammatory niche’ should be supported by independent vascular markers, inflammatory programme genes, spatial co-occurrence and recurrence across specimens. If only one component is secure, name the observed component and keep the broader interpretation provisional.

Benchmark against methods that answer neighbouring questions. BANKSY emphasises neighbourhood features for spatial domains, GraphST learns spatial graph representations, STAligner aligns sections and CellNEST prioritises communication edges. Agreement can strengthen a niche boundary, while disagreement often reveals whether the conclusion depends on clustering, alignment or signalling priors.

Atlas evaluation should include rare and transitional regions, not only large easy clusters. Report whether small niches are recovered across seeds and samples, and whether query mapping preserves them. Rare structures are where forced integration and pruning can do the most damage, so they deserve explicit diagnostics.

Spatial cross-validation can use held-out sections, donors or anatomically coherent blocks. Each choice asks a different question. State whether the goal is interpolation within a section, transfer across specimens or mapping into a reference atlas. Performance should never be presented without naming that generalisation task.

For a classroom model, let students arrange cell cards on a grid, assign neighbouring signals and compress the grid into two map coordinates. Then add a new card that does not fit the reference. Deciding whether to label it, leave it unknown or revise the atlas captures the heart of responsible reference mapping.

The publication-ready conclusion should mention data modality, spatial unit, sample count, programme evidence, graph sensitivity and experimental status. These details protect the article from becoming a promise that latent spaces can replace biology. The joy of NicheCompass lies in organising complex evidence so that new niches and better experiments become visible.

Before sharing an atlas, verify every primary-source link, pin the repository revision and retain the exact programme files used. A future database update can change the apparent biological meaning without changing the raw tissue data. Versioned priors make that distinction auditable.

The final figure should visually separate measured expression, graph-defined neighbourhoods, latent programme activity, transferred labels and experimentally validated markers. Readers can then see which layer supports each statement. Layered visual honesty is especially important when a polished embedding makes all points look equally certain.

Contents · Previous section · Continue to the Science Learning Hub