Why Science? | scTensor, Many-to-Many Cell Interactions and Hypergraph Evidence

Section 1 of 36

1. Think beyond pairwise arrows

scTensor represents candidate cell–cell interactions as relationships among ligand-expressing cell types, receptor-expressing cell types and ligand–receptor pairs. That three-way structure can reveal many-to-many patterns that a list of independent arrows may hide. It remains an expression-derived representation, not observed molecular traffic.

Contents · Next section

Section 2 of 36

2. Keep the 2023 paper boundary

The method was reported in BMC Bioinformatics in 2023 and evaluated on simulated and empirical datasets. Performance statements belong to the datasets, comparisons and settings in that study. New tissues, labels and interaction resources require new validation rather than inherited certainty.

Contents · Previous section · Next section

Section 3 of 36

3. Build the three tensor axes

One axis describes ligand-producing cell types, another receptor-bearing cell types and the third ligand–receptor pair identity. Every tensor entry therefore depends on cell annotation, expression preprocessing and the eligible pair database. State these dependencies before interpreting a component.

Contents · Previous section · Next section

Section 4 of 36

4. Use hypergraphs for many-to-many structure

A hyperedge can connect more than two vertices, matching biological situations in which several cell types and molecular pairs participate in a shared pattern. A detected hypergraph summarises coordinated support; it does not mean every listed cell physically contacted every other cell.

Contents · Previous section · Next section

Section 5 of 36

5. Understand Tucker2 decomposition

The published workflow uses non-negative Tucker2 decomposition to identify latent structures in the interaction tensor. Non-negativity aids interpretation, while the fitted factors and core describe a compressed approximation. Rank choice, initialisation and scaling influence which structures emerge.

Contents · Previous section · Next section

Section 6 of 36

6. Respect unsupervised discovery

scTensor does not need a pre-labelled condition to find patterns. That is useful for exploration, but an unsupervised component is not automatically meaningful. Biological labels should be assigned only after examining its cells, ligand–receptor pairs, specimens and robustness.

Contents · Previous section · Next section

Section 7 of 36

7. Separate structure from mechanism

Coordinated ligand and receptor transcripts under a curated pairing support a communication hypothesis. Secretion, protein complex assembly, receptor binding and receiver response are separate biological steps. Keep those unmeasured steps visible when naming a hypergraph.

Contents · Previous section · Next section

Section 8 of 36

8. Create a tensor manifest

Record species, specimen, assay, cell labels, filtering, normalisation, ligand–receptor resource, scTensor and R versions, ranks, initialisations and random seeds. A tensor cannot be interpreted without knowing what each axis and zero actually means.

Contents · Previous section · Next section

Section 9 of 36

9. Audit cell-type labels

Check marker coherence, cell counts, doublets, transitional populations and specimen balance for every sending and receiving group. Repeat key findings at a broader label resolution. A narrow or unstable label can create a distinctive hyperedge that disappears after sensible relabelling.

Contents · Previous section · Next section

Section 10 of 36

10. Document expression normalisation

The paper describes its own preprocessing choices, including count normalisation and transformation. Preserve the exact method used in a new analysis. Scaling changes the relative influence of abundant and moderate signals and can reshape the decomposition.

Contents · Previous section · Next section

Section 11 of 36

11. Freeze the species-specific pair set

The package provides ligand–receptor resources spanning many organisms. Select the correct species and archive the exact table. If orthology is added, label it as a new inference rather than direct same-species evidence. Record complex and identifier handling.

Contents · Previous section · Next section

Section 12 of 36

12. Archive the decomposition

Save the input tensor, axis labels, missingness mask, fitted factors, core tensor, ranks, initialisations, seeds and complete hyperedge tables. A rendered network cannot reproduce a decomposition or show whether factors swapped order between runs.

Contents · Previous section · Next section

Section 13 of 36

13. Protect the specimen unit

Cells belong to donors, animals or samples. Hyperedges should be checked within independent specimens or through resampling that respects nesting. A large number of cells improves resolution but does not replace biological replication.

Contents · Previous section · Next section

Section 14 of 36

14. Plan null models honestly

A useful null preserves expression prevalence, cell composition and specimen structure while disrupting the particular interaction relationship under test. A shuffle that destroys all biological organisation is too easy to beat and can exaggerate significance.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented hypergraph table

This fictional table is not scTensor output. It shows how similar-looking hypergraphs can require different next steps. Read across all axes before giving a component a biological name.

Contents · Previous section · Next section

Section 16 of 36

16. Read factor loadings together

Publish the leading ligand cell types, receptor cell types and ligand–receptor pairs for each component. A component name based on one familiar molecule hides the multi-axis evidence. Show loadings and uncertainty, not only a colourful hypergraph.

Contents · Previous section · Next section

Section 17 of 36

17. Inspect the core tensor

The core describes how latent cell and molecular factors combine. Large entries identify fitted relationships under the selected ranks. They are not probabilities of communication. Compare core structure across nearby ranks and random starts before interpreting it.

Contents · Previous section · Next section

Section 18 of 36

18. Compare hypergraphs across specimens

Ask whether a similar pattern appears in each independent biological unit and whether its member loadings agree. Report prevalence and heterogeneity. A pooled hypergraph driven by one specimen should be labelled specimen-specific rather than general.

Contents · Previous section · Next section

Section 19 of 36

19. Audit broad all-to-all patterns

The paper discusses limits of label-permutation approaches when interactions are broadly shared. In an extreme all-to-all pattern, swapping labels changes little, so genuine broad structure may be hard to detect. Absence of significance is not always absence of biology.

Contents · Previous section · Next section

Section 20 of 36

20. Audit many-to-many bias

Conversely, broad expression and large cell groups can make a dense pattern look important. Compare abundance-matched nulls, detection fractions and specimen prevalence. Use controls that distinguish coordinated interaction structure from ubiquitous transcripts.

Contents · Previous section · Next section

Section 21 of 36

21. Compare interaction databases

Harmonise species, direction, identifiers and complex rules before comparing resources. Shared hyperedges support robustness; resource-specific edges reveal curation differences. Database breadth affects the tensor dimension and can change rank selection as well as biology.

Contents · Previous section · Next section

Section 22 of 36

22. Audit rank selection

Too few components merge distinct patterns; too many split stable structure or fit noise. Report reconstruction error, stability, interpretability and conclusions across neighbouring ranks. Do not choose rank only because it produces the prettiest separation.

Contents · Previous section · Next section

Section 23 of 36

23. Audit random initialisation

Non-negative factorisation can converge to different local solutions. Run multiple seeds, align similar components by their loadings and report stability. A claim attached only to ‘component two’ without alignment is fragile because component numbers can swap.

Contents · Previous section · Next section

Section 24 of 36

24. Audit reconstruction and residuals

Inspect which interactions the model reconstructs well and where residuals remain. A small average error can hide failure in a rare but biologically important group. Publish residual diagnostics for headline hypergraphs.

Contents · Previous section · Next section

Section 25 of 36

25. Add spatial evidence

If a hypergraph implies local communication, test whether relevant cell types occupy compatible compartments and distances. Pairwise proximity alone may not support a many-to-many arrangement. Spatial evidence should match the mediator’s range and the hypergraph’s members.

Contents · Previous section · Next section

Section 26 of 36

26. Add protein evidence

Confirm key ligands, receptors and required complex subunits at the protein level. RNA co-expression across several cell types cannot establish secretion or assembled receptors. Choose assays that match soluble, contact-dependent or matrix-associated signalling.

Contents · Previous section · Next section

Section 27 of 36

27. Design a joint perturbation

Select a hypergraph prediction that differs from simpler pairwise models. Perturb one key ligand or receptor and measure responses across the several nominated receiver groups. The best experiment tests the many-to-many claim, not merely one familiar edge.

Contents · Previous section · Next section

Section 28 of 36

28. Write a bounded scTensor claim

Prefer: ‘scTensor identified a stable hypergraph linking these cell types and pairs under two ranks and three specimens.’ Avoid: ‘All cells communicated through this pathway.’ The bounded version states exactly what the decomposition contributed.

Contents · Previous section · Next section

Section 29 of 36

29. Teach with a three-dimensional model

Students can build a cube whose axes are ligand cells, receptor cells and molecular pairs, then mark supported entries. Rotating the cube shows why one flat table can hide coordinated structure. They can then ask which entries are measurements and which are inferences.

Contents · Previous section · Next section

Section 30 of 36

30. Begin with Primary Science grouping

Classification and pattern recognition begin early. Grouping observations by source, receiver and signal shows that categories must be defined consistently. Accurate labels matter before any advanced mathematics enters the story.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science for fair comparisons

Learners can compare two hypothetical hypergraphs while changing one factor at a time: cell number, label, database or protein evidence. This practises fair testing and prevents a larger graph from being mistaken for a better explanation.

Contents · Previous section · Next section

Section 32 of 36

32. Connect Secondary and O-Level Science

Cells, transport, coordination and experimental design provide the biological foundation. scTensor adds a data representation for complex systems. Students can evaluate whether evidence supports a molecular mechanism or only a pattern.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing explain tensors

Computing introduces multidimensional arrays, factorisation, random seeds and reproducible code. Biology explains what the axes mean. Neither discipline can be skipped: a valid decomposition with mislabelled cell types is still scientifically wrong.

Contents · Previous section · Next section

Section 34 of 36

34. Choose opportunities by evidence

For school choices or enrichment, verify current programmes on official pages and look for sustained work in biology, mathematics, computing, data handling and research. Do not infer admissions, facilities or outcomes from a broad STEM label.

Contents · Previous section · Next section

Section 35 of 36

35. See connected careers

Hypergraph reasoning appears in bioinformatics, network science, statistics, computational biology, pharmaceutical research and software engineering. Educational pathways vary and role requirements change, so current institutional and employer information should guide decisions.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with all three axes visible

A defensible scTensor conclusion names the tensor axes, cell labels, pair resource, normalisation, rank, seed stability and specimen support, then separates hypergraph structure from spatial, protein and perturbation validation. The method is strongest when complexity stays inspectable.

scTensor treats a communication landscape as a genuinely multi-way object. A ligand cell type, a receptor cell type and a ligand–receptor pair jointly locate an entry in the tensor. Decomposition then searches for recurring structures across those axes. The benefit is not simply a larger graph; it is a representation in which many-to-many organisation can be examined directly.

Did you know? A hypergraph edge can join more than two vertices. That makes hypergraphs a natural language for patterns involving several sending cell types, several receiving cell types and a family of molecular pairs—relationships that are awkward to describe as isolated pairwise arrows.

The published paper uses non-negative Tucker2 decomposition. Non-negativity helps because positive contributions cannot cancel through opposite signs, but factorisation remains a model. Run several initialisations, compare neighbouring ranks and align components by loading similarity rather than component number. A stable scientific conclusion should survive ordinary numerical ambiguity.

The three axes need strong definitions. Cell types require defensible markers and specimen support. Ligand–receptor pairs require a frozen, species-appropriate database and transparent complex handling. Tensor entries require documented normalisation. If any axis changes, the fitted hypergraphs can change even when the underlying count matrix does not.

Permutation testing needs special care for broad patterns. The scTensor study discusses how label permutations can struggle when interactions approach an all-to-all structure: swapping labels barely changes a ubiquitous pattern. That is a useful reminder that a negative statistical result can sometimes reflect the null design rather than biological absence.

The opposite risk is also real. Highly prevalent genes and large cell groups can populate many tensor entries and dominate a component. Compare abundance-matched controls, detection fractions and specimen prevalence. Inspect whether a hypergraph remains after downsampling large groups or excluding ubiquitous housekeeping-like signals.

Validate the joint prediction, not only one edge. If a hypergraph links two sender groups, three receiver groups and a molecular family, choose a perturbation expected to affect a specific subset. Measure the receiver responses across all nominated groups. A result matching that pattern supports the many-to-many model more strongly than validating one familiar ligand–receptor pair.

Archive the input tensor and missingness mask. A zero can mean low expression, no eligible interaction, an absent cell type or a missing measurement. Those meanings are not interchangeable. Explicit missingness prevents structural absence from being misread as a biological negative.

For students, the tensor can begin as a cube made from paper or blocks. Each face offers a different view, but the object stays the same. This connects classification, fair comparison and systems thinking with multidimensional arrays and factorisation. It also shows why visualisation never replaces the underlying definitions.

The bounded conclusion is that scTensor can reveal stable, context-aware hypergraphs of candidate cell–cell interaction under a declared representation. The hypergraphs remain conditional on labels, preprocessing, pair resources, rank and null models. Protein, spatial and perturbation evidence are what move them towards mechanism.

Report component uncertainty rather than treating the fitted decomposition as unique. Bootstrap cells within specimens, refit the model and match components by similarity of their three loading profiles. Summarise which cell types and molecular pairs recur and which are unstable. This distinguishes a robust hypergraph core from decorative members that appear only in a small fraction of runs.

Compare the tensor result with pairwise summaries. A many-to-many method earns its complexity when the joint structure makes a prediction that separate pair scores miss—for example, a molecular family coordinating several receiver groups only in a defined context. If the same conclusion is obvious from one pair, the simpler representation may be easier to communicate and validate.

Denominators matter. State how many cell types, candidate pairs and tensor entries were eligible before filtering, how many were observed and how many contributed to each component. A dense hypergraph drawn from a very small eligible universe has a different interpretation from the same picture drawn after broad screening.

Keep informative absences distinct: an absent cell type, an unmeasured gene, an ineligible pair, a zero after normalisation and a low fitted loading are not the same. Encoding these states explicitly protects the tensor from turning missing opportunity into evidence of no communication.

When comparing conditions, avoid fitting an attractive global decomposition and then treating cell-level loadings as independent replicates. Define the biological unit first. Estimate condition contrasts across specimens, with uncertainty that respects pairing and batch. The tensor organises the evidence; it does not change the experimental unit.

A balanced results page should show one stable many-to-many pattern, one rank-sensitive component, one database-sensitive edge and one realistic negative control. Readers can then see how reliability was judged. This is more educational and scientifically useful than presenting only the most colourful hypergraph.

The best final figure links the hypergraph back to inspectable tables. Every cell type and ligand–receptor pair should be traceable to its loading, specimen support, input expression and database record. Navigable evidence is what turns a complex decomposition into a practical scientific tool.

That traceability also gives future experiments a precise starting point.

Contents · Previous section · Continue to the Science Learning Hub