Why Science? | SPOTlight, Seeded NMF Regression and Spatial Cell-Mixture Evidence

Section 1 of 36

1. Start with spatial mixtures

A spatial transcriptomics spot can collect RNA from several cells. SPOTlight asks which reference cell types, represented through expression topics, combine to explain that mixture. The method keeps weights non-negative, matching the idea that a cell type can contribute zero or more signal rather than negative molecules.

Contents · Next section

Section 2 of 36

2. Seed NMF with marker genes

Non-negative matrix factorisation separates reference expression into topics. SPOTlight seeds those topics with cell-type marker genes so the decomposition begins with biological guidance. Good seeds can stabilise interpretation; weak or circular markers can bake annotation errors into every later spatial estimate.

Contents · Previous section · Next section

Section 3 of 36

3. Understand topics as expression programmes

A topic is a non-negative pattern across genes, not automatically a cell type. Seeded topics are associated with types through markers and reference cells, while shared biological programmes may span multiple types. Inspect topic genes before giving each component a biological name.

Contents · Previous section · Next section

Section 4 of 36

4. Learn cell weights in topic space

The model represents reference cells as combinations of topics. This compresses thousands of genes into structured programmes useful for regression. Compression reduces complexity, but it can merge rare or related states if the chosen topic structure does not preserve their distinctive signals.

Contents · Previous section · Next section

Section 5 of 36

5. Regress each spatial spot non-negatively

Non-negative least squares estimates which topic combinations reconstruct a spatial spot. The resulting weights can then be related back to cell types. A good reconstruction supports the mixture under the model, but several correlated combinations may fit almost equally well.

Contents · Previous section · Next section

Section 6 of 36

6. Retain an unresolved component

Residual expression or an unknown component matters because references are never perfectly complete. Large residuals can indicate missing states, platform effects or poor preprocessing. Forcing every molecule into named cell types creates apparent completeness while hiding model mismatch.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the primary paper’s scope clear

The 2021 Nucleic Acids Research paper introduced SPOTlight for deconvolving spatial spots with single-cell transcriptomes using seeded NMF regression. Demonstrations established the workflow’s usefulness. They do not provide universal marker lists, topic counts or thresholds for every tissue and platform.

Contents · Previous section · Next section

Section 8 of 36

8. Build trustworthy reference annotations

Quality-control cells, remove doublets, inspect batches and document label evidence. Use enough examples per type and consider hierarchical annotations. A seed marker that distinguishes one dataset’s technical batch rather than biology can dominate the factorisation.

Contents · Previous section · Next section

Section 9 of 36

9. Choose markers independently

Select markers with specificity, reproducibility and platform detectability in mind. Avoid using the same spatial pattern as both marker selection and validation. Report the complete marker set and assess results after removing the most influential genes.

Contents · Previous section · Next section

Section 10 of 36

10. Prepare common features carefully

Intersect usable genes across single-cell and spatial assays, then document normalisation and filtering. Highly abundant housekeeping or stress genes can explain counts without resolving types. Feature selection should serve the biological question rather than maximise visual separation.

Contents · Previous section · Next section

Section 11 of 36

11. Choose topic structure deliberately

Too few topics merge distinct programmes; too many can fragment one type or model noise. Compare plausible settings, examine topic coherence and map stability. A visually detailed result is not necessarily more accurate than a stable broader decomposition.

Contents · Previous section · Next section

Section 12 of 36

12. Train seeded NMF reproducibly

Record software version, seed genes, number of topics, random seeds, iterations and convergence. Repeat the factorisation because NMF solutions can depend on initialisation. Major cell-type patterns should remain consistent before small differences are interpreted.

Contents · Previous section · Next section

Section 13 of 36

13. Map topics back to cell types

Use reference-cell topic weights and marker evidence to associate components with types. Shared programmes such as proliferation or interferon response may cross labels. Preserve those as programmes when appropriate rather than forcing every topic into a mutually exclusive identity.

Contents · Previous section · Next section

Section 14 of 36

14. Deconvolve and inspect residuals

For every spatial spot, report reconstructed signal, residual magnitude and cell-type weights. Map high residuals on tissue images. A boundary with poor fit may mark damage or missing biology, and it should not receive the same confidence as a well-explained region.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented NMF table

This fictional table illustrates evidence, not software performance.

Contents · Previous section · Next section

Section 16 of 36

16. Read weights as molecular evidence

SPOTlight weights reflect how reference-derived topics explain each spot. They are not automatically cell fractions or absolute counts. Different cell types produce different amounts of RNA, and platform capture changes scale. Calibration requires matched images or counting evidence.

Contents · Previous section · Next section

Section 17 of 36

17. Use thresholds transparently

If a map declares presence above a threshold, predefine the rule and show sensitivity. A threshold can simplify communication while hiding continuous uncertainty. Publish the unthresholded weights and residuals so readers can evaluate borderline locations.

Contents · Previous section · Next section

Section 18 of 36

18. Separate cell identity from cell state

A seeded topic may represent a lineage, a state or a shared programme. Activated immune cells can express genes absent from a resting reference. Analyse broad identity and condition-responsive programmes separately to avoid turning state changes into false abundance differences.

Contents · Previous section · Next section

Section 19 of 36

19. Compare repeated sections

Run the same prespecified workflow across independent specimens, not only adjacent spots. Quantify overlap of broad distributions and show donor effects. Many spots within one section increase spatial detail but do not create population replication.

Contents · Previous section · Next section

Section 20 of 36

20. Validate spatial mixtures

Use histology, in-situ RNA, protein staining or segmentation on matched tissue. Select abundant, rare and correlated types. Validation should test unexpected mixtures and high-residual regions, not only locations already supported by famous markers.

Contents · Previous section · Next section

Section 21 of 36

21. Report computing and object versions

SPOTlight workflows may use SingleCellExperiment, SpatialExperiment or Seurat objects. Record package versions, object conversions, assay names and transformations. Silent changes in data slots or defaults can alter results even when the same high-level command runs.

Contents · Previous section · Next section

Section 22 of 36

22. Challenge marker dependence

Remove one marker family at a time and rerun the factorisation. If a spatial pattern collapses when one gene disappears, describe it as marker-dependent. Robust cell-type evidence should draw support from multiple independent features where possible.

Contents · Previous section · Next section

Section 23 of 36

23. Challenge collinear cell types

Closely related types produce similar topic profiles and can exchange weight. Examine correlation, merge labels when needed and use targeted markers for confirmation. Fine subtype proportions are not recoverable simply because the reference contains fine labels.

Contents · Previous section · Next section

Section 24 of 36

24. Challenge random seeds and topic number

Repeat NMF with several initialisations and defensible topic counts. Map consensus and disagreement. Stable broad patterns with unstable fine partitions support a hierarchical conclusion, not a single definitive fine map.

Contents · Previous section · Next section

Section 25 of 36

25. Challenge residual thresholds

Residuals are evidence about model adequacy. Explore how conclusions change under different residual cut-offs and avoid discarding poor-fit spots without mapping them. Systematic residuals may reveal a missing programme more interesting than the original decomposition.

Contents · Previous section · Next section

Section 26 of 36

26. Compare SPOTlight with RCTD

SPOTlight uses seeded NMF topics and non-negative regression, while RCTD uses a probabilistic count model with platform-effect correction. Agreement can strengthen broad localisation. Differences can reveal marker, topic, platform or likelihood assumptions that deserve biological testing.

Contents · Previous section · Next section

Section 27 of 36

27. Compare SPOTlight with Stereoscope

Both infer mixtures from single-cell and spatial data, but SPOTlight’s topic construction differs from Stereoscope’s probabilistic count model. Use simulated mixtures and orthogonal measurements to assess the relevant operating range rather than choosing the smoother map.

Contents · Previous section · Next section

Section 28 of 36

28. Keep negative results calibrated

A low weight does not prove absence if the reference lacks the state or its markers are weak. Add synthetic mixtures across abundance levels and spatial sizes to estimate detection limits. Report which populations were detectable and which remained indistinguishable.

Contents · Previous section · Next section

Section 29 of 36

29. Try a classroom model

Mix coloured transparent sheets to create composite squares. Learners factor recurring colours into topics, estimate each mixture and then see what happens when the reference omits yellow. 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 SPOTlight 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. SPOTlight 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. SPOTlight 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 SPOTlight 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 SPOTlight appears in bioinformatics, biomedical data science, tissue biology, statistics and reproducible research software. 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: SPOTlight estimates non-negative reference-topic contributions; it does not directly count cells or make residual signal disappear. 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 strong SPOTlight study begins with marker governance. Record how reference cells were filtered and annotated, how markers were selected, their specificity and their detectability in the spatial platform. Separate markers chosen from prior biology or held-out reference data from markers discovered in the same spatial dataset. Circular marker selection can make a map appear validated before any independent test occurs.

Pre-register the gene intersection, normalisation, number of topics, marker seeds, NMF repetitions, regression settings, residual rule, primary cell-type contrast and validation. Seeded NMF is guided but not uniquely determined. Repeated initialisations and plausible topic counts should be analysed before choosing a headline result. Archive exploratory settings rather than hiding the paths that produced less tidy maps.

Publish a paired funnel for single-cell and spatial data. Show raw cells, quality-controlled cells, doublets removed, donors and labels; then spatial locations, tissue coverage, count filters, genes retained and fitted spots. Map excluded and high-residual locations on histology. Poorly explained areas can reveal damaged tissue, a missing state or a genuine biological programme absent from the reference.

Topic interpretation requires a table of leading genes, marker enrichment, reference-cell distribution and stability across runs. Some topics may encode cell identity, while others capture proliferation, stress or interferon response shared across types. Keep shared programmes as programmes. Forcing every topic into one cell-type name can confuse state with composition.

Assess identifiability among types by comparing their topic profiles. When related types remain collinear, report a broader lineage or interval rather than precise competing weights. Remove one type at a time in controlled tests and observe where its signal moves. This reveals the signature of an incomplete real reference and prevents a complete palette from being mistaken for complete biology.

Synthetic mixtures should match observed library sizes and noise. Vary abundance, relatedness, marker strength and omitted programmes, then report recovery and residuals. Include mixtures below expected detection. Simulation is not proof of performance in tissue, but it provides an honest calibration for interpreting small weights and null results.

Residuals belong beside the deconvolution maps. Plot per-spot reconstruction error, gene-level residuals and spatial structure. A systematic residual region can be more scientifically interesting than the fitted weights because it suggests a missing cell state or programme. Do not discard such spots solely because they make the final map less complete.

Thresholds should be treated as communication choices, not biological laws. Publish continuous weights and show how presence, co-occurrence and region summaries change across defensible cut-offs. Predeclare any threshold used for confirmatory testing. Small numerical changes around a boundary should not create opposite biological stories.

Validation should be independent of the marker seeds wherever possible. Use alternative RNA probes, proteins, morphology or cell segmentation. Test abundant, rare and confusable types and include expected negatives. If the reference marker itself is the only validation signal, the result has not travelled far from its initial assumption.

Statistical inference must operate at the specimen level. Spatial spots and topic weights are correlated within tissue. Show all donors, balance batches and avoid treating thousands of spots as thousands of people. Replication across independently processed specimens is needed before generalising a spatial mixture or disease-associated change.

Reproducibility materials include matrices, metadata, markers, gene intersection, topic number, NMF seeds, convergence results, regression configuration, package versions, object conversions, software environment, weights, residuals, validation data, code and checksums. Bioconductor and Seurat object conventions evolve; record assay and slot names so future users know exactly what entered the model.

Negative findings require a calibrated detection envelope. A low weight can exclude a well-marked, abundant type under a matched reference while leaving a rare or collinear type unresolved. Report mixture simulations and orthogonal assay sensitivity. Do not convert model silence into biological absence without showing that the design could have heard the signal.

A defensible final sentence might say: ‘Marker-seeded SPOTlight analyses supported reproducible broad spatial mixtures across independent specimens, with topic stability, residuals and orthogonal markers reported, while rare and correlated subtypes remained unresolved.’ It should not call topic weights direct cell fractions or treat reference-derived markers as independent confirmation.

Contents · Previous section · Continue to the Science Learning Hub