Why Science? | MESSI, Spatial Neighbours, Mixture-of-Experts and Response-Gene Evidence

Section 1 of 36

1. Begin with prediction

MESSI asks whether signalling information inside a cell and in its spatial neighbours helps predict response-gene expression. Predictive performance gives an objective test of usefulness. A predictive coefficient is not automatically a causal biochemical effect.

Contents · Next section

Section 2 of 36

2. Keep the 2021 paper boundary

The Bioinformatics study evaluated MERFISH hypothalamus, MERFISH U-2 OS and STARmap medial-prefrontal-cortex datasets. Its accuracy and biological examples belong to those platforms, gene panels, specimens and validation designs. New applications need new held-out evaluation.

Contents · Previous section · Next section

Section 3 of 36

3. Define intracellular signalling genes

For a focal cell, MESSI uses selected ligand or receptor genes expressed within that cell as intracellular signalling features. These features describe possible local state. They do not directly measure protein modification or receptor activation.

Contents · Previous section · Next section

Section 4 of 36

4. Define intercellular signalling genes

Ligand expression in neighbouring cells supplies external signalling features for the focal cell. The neighbourhood is defined from spatial coordinates. This adds location to the model, but a nearby ligand transcript is still only a candidate influence.

Contents · Previous section · Next section

Section 5 of 36

5. Predict response genes

The model predicts expression of selected response genes in the focal cell. Accurate out-of-sample prediction indicates that the features contain information about the responses. It does not reveal which unmeasured mechanism carries that information.

Contents · Previous section · Next section

Section 6 of 36

6. Use a mixture of experts

Instead of one regression for every cell of a type, MESSI probabilistically assigns cells to latent experts representing subtypes. Different experts can learn different relationships. The subtype is model-derived and requires biological interpretation.

Contents · Previous section · Next section

Section 7 of 36

7. Share information through multi-task learning

Response genes are learned jointly so their covariance can share statistical strength. This may improve prediction when responses are related. The covariance model is an analytical structure, not a regulatory network measured in the tissue.

Contents · Previous section · Next section

Section 8 of 36

8. Create a spatial-data manifest

Record platform, species, tissue, condition, specimen, coordinates, gene panel, cell type, preprocessing, neighbour rule, signalling-gene resource, response-gene definition, train-test split and MESSI version. Every item affects interpretation.

Contents · Previous section · Next section

Section 9 of 36

9. Protect coordinate integrity

Verify units, orientation, transformed coordinates, tissue masks and removed cells. A geometry error changes neighbourhoods and therefore the external features. Archive raw and processed coordinates.

Contents · Previous section · Next section

Section 10 of 36

10. Define neighbours biologically

The published analyses used Delaunay triangulation with dataset-specific distance cut-offs. Choose a rule justified by resolution and mediator range. Repeat key conclusions across reasonable alternatives rather than treating one graph as anatomical truth.

Contents · Previous section · Next section

Section 11 of 36

11. Audit gene-panel coverage

MERFISH and STARmap panels may measure hundreds of selected genes, while other datasets measure thousands. State which ligands, receptors and responses could be observed. An absent gene outside the panel is not a biological negative.

Contents · Previous section · Next section

Section 12 of 36

12. Freeze signalling-gene resources

Archive the ligand and receptor definitions and any analyst-selected features. The study drew on established ligand/receptor information. Resource choice determines which molecular explanations the model can consider.

Contents · Previous section · Next section

Section 13 of 36

13. Protect held-out specimens

The study used cross-validation that learned on some animals and tested on others where applicable. Preserve biological independence in new evaluation. Randomly splitting cells from the same animal can overstate generalisation.

Contents · Previous section · Next section

Section 14 of 36

14. Archive model objects

Retain feature matrices, neighbour graph, expert assignments, coefficients, covariance estimates, regularisation settings, seeds, folds, predictions and residuals. A coefficient plot alone cannot reproduce the model.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented expert table

This fictional table is not MESSI output. Decide whether each result supports prediction, subtype structure or mechanism.

Contents · Previous section · Next section

Section 16 of 36

16. Read expert membership

Report assignment probabilities and cell counts, not only hard labels. Cells near an expert boundary may share programmes. Compare experts with known markers, anatomy and specimens before naming a biological subtype.

Contents · Previous section · Next section

Section 17 of 36

17. Read coefficients conditionally

A large coefficient means the feature contributes under the model, scaling, expert and correlated-feature set. It is not an effect size from a controlled intervention. Show uncertainty and stability across folds.

Contents · Previous section · Next section

Section 18 of 36

18. Inspect collinearity

The paper notes that highly correlated signalling genes can divide importance and cause key features to be missed. Examine correlated groups, regularisation sensitivity and grouped interpretations. Do not over-rank one member of an interchangeable set.

Contents · Previous section · Next section

Section 19 of 36

19. Inspect residuals

Plot prediction error by response gene, expert, specimen and region. Good average accuracy can hide a poorly modelled rare subtype. Residual spatial structure suggests missing neighbourhood or anatomical information.

Contents · Previous section · Next section

Section 20 of 36

20. Compare internal and neighbour features

Fit transparent baselines using intracellular features alone and neighbourhood features alone. MESSI adds intercellular value when neighbour information improves held-out prediction beyond the cell’s own state.

Contents · Previous section · Next section

Section 21 of 36

21. Compare across specimens

Ask whether the same expert and leading features recur in independent animals or batches. A subtype confined to one specimen may be real, but it should be labelled specimen-specific until replicated.

Contents · Previous section · Next section

Section 22 of 36

22. Audit neighbour geometry

Repeat Delaunay and distance choices within a planned range. Contact signals and diffusible ligands should not share an unexamined neighbourhood rule. Report which coefficients and experts remain stable.

Contents · Previous section · Next section

Section 23 of 36

23. Audit feature selection

Response-gene and signalling-gene choices shape the prediction task. Compare preregistered or biologically justified sets. Choosing features after observing results can make the model appear more explanatory than it is.

Contents · Previous section · Next section

Section 24 of 36

24. Audit regularisation and experts

Vary regularisation and number of experts using cross-validation rather than visual preference. Too few experts merge biology; too many can fit noise or specimen identity. Report selection criteria.

Contents · Previous section · Next section

Section 25 of 36

25. Audit platform limitations

Gene coverage, detection efficiency and spatial resolution differ between MERFISH, STARmap and whole-transcriptome technologies. A method may accept several platforms without making their evidence equivalent.

Contents · Previous section · Next section

Section 26 of 36

26. Add proteins and timing

Confirm leading ligands and receptors at the protein level and test whether source availability precedes receiver response. Spatial RNA at one time point cannot establish order or molecular activity.

Contents · Previous section · Next section

Section 27 of 36

27. Design a predictive perturbation

Choose two experts predicted to respond differently to the same neighbour ligand. Perturb the ligand or receptor and measure early and late response genes. A subtype-specific outcome tests more than a global association.

Contents · Previous section · Next section

Section 28 of 36

28. Write a bounded MESSI claim

Prefer: ‘MESSI used neighbouring-cell and intracellular features to improve held-out response-gene prediction and prioritised this ligand within a stable expert.’ Avoid saying the model proved the ligand caused the response.

Contents · Previous section · Next section

Section 29 of 36

29. Turn neighbours into a classroom model

Place students as cells on a grid and give each internal and neighbour feature cards. Ask them to predict a response, then change the neighbourhood. The activity separates prediction from causation.

Contents · Previous section · Next section

Section 30 of 36

30. Begin with Primary Science patterns

Observing surroundings, grouping similar cases and predicting outcomes are foundational skills. Learners can see that a nearby object may help prediction without being the cause.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science for fair comparisons

Compare a model with and without neighbour information while holding other variables steady. This is a simple fair test of added predictive value.

Contents · Previous section · Next section

Section 32 of 36

32. Connect Secondary and O-Level Science

Cells, signalling, response, variables and experimental design provide the biological frame. MESSI adds spatial data and statistical prediction while keeping the need for experiments visible.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing explain the model

Computing contributes graphs, mixtures, probabilities, regression, covariance, regularisation and cross-validation. Biology defines meaningful features and held-out units. Both are required for reliable interpretation.

Contents · Previous section · Next section

Section 34 of 36

34. Choose opportunities by verified fit

Check current official school or enrichment information for sustained work in science, mathematics, computing and research. Do not infer programme access or outcomes from a broad STEM description.

Contents · Previous section · Next section

Section 35 of 36

35. See connected careers

Spatial modelling connects to bioinformatics, neuroscience, statistics, machine learning, imaging, software engineering and experimental biology. Verify current course and role requirements directly.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with prediction and mechanism separate

A defensible MESSI conclusion names the neighbour graph, features, experts, folds, specimens, prediction metrics and stability, then separates predictive importance from protein, temporal and perturbation evidence. That distinction makes the model more useful, not less.

MESSI begins with a prediction problem rather than an arrow-drawing problem. For a chosen cell type, it uses signalling-related features inside a focal cell and ligand information from spatial neighbours to predict response-gene expression. A mixture-of-experts model can represent latent subtypes, and multi-task learning shares structure across response genes. Predictive success is valuable; it is not the same as identifying a biochemical cause.

Did you know? In the published workflow, spatial neighbourhoods can be constructed with Delaunay triangulation and a distance cutoff. Triangulation connects nearby points geometrically, while the cutoff removes implausibly long edges. Both choices affect which cells can contribute neighbouring ligands, so geometry is part of the scientific model.

Start with coordinates and units. Preserve the original spatial reference, image orientation, scale and transformations. Plot the neighbour graph on the tissue before fitting anything. Boundaries, tears and sparse regions can create geometrically valid but biologically implausible edges. A tissue mask or compartment rule may be needed.

Define the focal population with multiple markers and specimen support. MESSI’s experts may reveal response heterogeneity within that population, but they cannot repair a fundamentally mixed or mislabelled cell class. Compare the inferred subtypes with known anatomy, cell state and technical covariates before giving them biological names.

Feature construction deserves a ledger. Separate intracellular signalling features from neighbour-derived ligand features. Report the gene resource, species mapping, normalisation, missing-gene policy and aggregation across neighbours. A sum, mean and distance-weighted mean answer different questions when a cell has many neighbours.

The mixture-of-experts structure should be stress-tested. Repeat fitting across initialisations and component counts, align experts by their prediction profiles and report assignment uncertainty. An expert that appears only under one seed should not become a newly named biological subtype. Stable recurrence across specimens is more persuasive.

Multi-task learning can borrow strength across response genes, which is helpful when outputs share structure. It can also spread bias when a group of highly correlated genes dominates. Inspect task-specific performance and residuals, not only an average score. A model that predicts abundant response genes well but fails on the headline target needs a narrow conclusion.

The paper evaluated the method on MERFISH hypothalamus, MERFISH U-2 OS and STARmap mouse medial-prefrontal-cortex datasets. Those platforms and tissues differ in biology, resolution and gene panels. Their use demonstrates breadth in the study; it does not make measurements from different technologies interchangeable.

Cross-validation must respect specimens and spatial autocorrelation. Randomly splitting neighbouring cells can leak local state from training into test data. Where multiple animals or samples exist, hold out entire biological units. Where only one field exists, use spatial blocks and describe the remaining limitations. The evaluation design determines what predictive performance means.

Feature importance and coefficients require cautious verbs. A neighbour ligand that improves prediction is associated with response information under the fitted model. It may proxy anatomy, cell density or an unmeasured programme. Correlated signalling genes can divide importance or substitute for one another, a limitation discussed in the method’s context.

Use ablations that match the claim. Remove neighbour ligands to test whether spatial context adds value. Remove intracellular features to test receiver-state contribution. Shuffle coordinates within compartments, alter the distance cutoff and compare a simple regularised regression. MESSI earns its complexity when these comparisons reveal stable, out-of-sample structure.

Validation should target the prediction. Confirm the ligand and receptor proteins in the appropriate compartments, perturb the proposed route and measure the response gene over time. Compare near and distant cells where geometry makes that meaningful. Rescue or receptor blockade can separate the proposed pathway from a broad environmental association.

For students, MESSI brings together maps, neighbours and prediction. Primary and PSLE Science provide observation and fair testing; Secondary and O-Level Biology add cell signalling and response; Computing supplies features, training, cross-validation and error analysis. The central lesson is cheerful and practical: a useful model is one whose mistakes are measured and whose claims stay within the evidence.

A report should include predictive metrics per response gene, uncertainty across specimens, expert stability, neighbour sensitivity and a table of leading features with correlation groups. Show at least one successful prediction and one failure. Failure analysis reveals whether the model misses a cell state, a region or a class of responses.

Archive the exact repository revision, environment, gene lists, neighbour graph, distance rule, expert count, regularisation and train/test partitions. The official MESSI repository supports reproducibility; the peer-reviewed paper anchors the published method. Future software updates should be cited as current implementations, not silently attributed to the 2021 version.

The bounded conclusion is useful: MESSI can predict spatial response-gene patterns and prioritise intracellular and neighbour-associated signalling features while representing latent heterogeneity. It does not directly observe ligand release, binding, receptor activation or causal regulation. Prediction supplies a map for the next experiment—not the final molecular verdict.

Before interpreting a coefficient, inspect the local feature neighbourhood. Several correlated ligands or signalling genes may carry nearly the same information, causing the model to divide weight unpredictably. Group correlated features, refit after removing one member and describe a stable feature family when individual rankings are not reproducible.

Calibrate predictions as well as ranking them. Plot predicted against observed response values by specimen, expert and anatomical region. Check whether residual variance grows with expression and whether one area dominates the fit. A good average correlation can coexist with systematic failure in the region that matters biologically.

Choose the next experiment to break a predictive tie. If neighbouring ligand abundance and receiver state both explain the response, perturb the ligand while controlling for state, or compare matched receivers with different neighbourhoods. The most informative experiment is the one that makes rival explanations diverge. In this way, MESSI’s prediction becomes a disciplined guide to mechanism.

Spatial negatives deserve equal attention. A model may predict well in one compartment and fail at a boundary, around a vessel or in a sparse region. Map those residuals back to the tissue. They can reveal missing cell types, an unsuitable neighbourhood definition or biology outside the feature set. A residual map is often more instructive than another global performance number.

When comparing conditions, fit and evaluate the contrast at the specimen level. Report whether the same expert structure and feature families recur before interpreting differences. If a new expert appears only in one batch, check technical covariates and cell coverage first. Prediction remains most persuasive when it generalises to held-out biological units under the same declared task.

Contents · Previous section · Continue to the Science Learning Hub