Why Science? | scDiffCom, Permutation Tests and Differential Cell–Cell Communication

Section 1 of 36

1. Start with a contrast, not a catalogue

scDiffCom is designed for a pairwise comparison: which candidate cell–cell interactions differ between two declared conditions? That is more specific than drawing two networks and comparing them by eye. The conditions, experimental units and eligible cell populations must be defined before testing.

Contents · Next section

Section 2 of 36

2. Define a communication candidate

A candidate combines an emitter cell type, receiver cell type and ligand–receptor interaction. This triplet is called a cell–cell interaction in the workflow. It represents expression compatibility under a resource, not observed transport or receptor activation.

Contents · Previous section · Next section

Section 3 of 36

3. Distinguish detection from difference

Detection asks whether a candidate has more expression support than expected in a condition. Differential analysis asks whether support changes between conditions. A candidate can be detectable in both conditions yet differ modestly, or appear condition-specific because power differs.

Contents · Previous section · Next section

Section 4 of 36

4. Understand the permutation logic

scDiffCom uses permutation procedures to build null expectations for detection and differential statistics. The exchangeability rule matters: labels should be permuted only where the experimental design makes that comparison valid. More permutations refine tail estimates but cannot repair a confounded design.

Contents · Previous section · Next section

Section 5 of 36

5. Know simple and complex interactions

The resource distinguishes simple interactions from heteromeric ligand or receptor complexes. For a complex, all required subunits matter. Reporting only the most abundant subunit can create a biologically incomplete route.

Contents · Previous section · Next section

Section 6 of 36

6. Use enrichment after primary testing

Overrepresentation analysis can summarise significant results by ligands, receptors, cell types, Gene Ontology or KEGG categories. Enrichment helps navigate many candidates. It does not rescue an unstable primary contrast or prove pathway activation.

Contents · Previous section · Next section

Section 7 of 36

7. Keep the ageing atlas boundary

The 2023 paper introduced scAgeCom, an atlas of age-related communication changes in mouse tissues, alongside the method. Those reported patterns belong to the analysed datasets and validations. A new tissue, age range or species requires its own evidence.

Contents · Previous section · Next section

Section 8 of 36

8. Prepare annotated Seurat objects

The official workflow expects appropriately processed single-cell data with cell-type labels and condition information. Preserve raw counts, normalisation, filtering, integration choices and object versions. Hidden preprocessing differences can masquerade as biological communication changes.

Contents · Previous section · Next section

Section 9 of 36

9. Name the biological replicate

State whether the independent unit is a donor, animal, culture or batch. Cells from one specimen are not independent condition replicates. Summarise candidate support within specimens and show prevalence across them.

Contents · Previous section · Next section

Section 10 of 36

10. Balance cell-type opportunity

Conditions often differ in cell abundance or detection depth. Report numbers of cells and specimens for every emitter–receiver combination. A missing interaction may reflect missing opportunity rather than altered signalling.

Contents · Previous section · Next section

Section 11 of 36

11. Freeze the interaction resource

Archive the exact scDiffCom package and ligand–receptor database version. The paper’s resource integrated seven named databases and approximately five thousand curated interactions at that snapshot. Updates can change both the tested universe and enrichment background.

Contents · Previous section · Next section

Section 12 of 36

12. Declare species and identifiers

Human, mouse and rat resources require correct gene identifiers and mapping. Record synonym resolution and discarded genes. Orthology conversion can merge or split relationships and should never be hidden inside preprocessing.

Contents · Previous section · Next section

Section 13 of 36

13. Plan permutation counts

The official vignette uses 1,000 permutations for a preliminary run and recommends 10,000 otherwise. Treat those values as implementation guidance, not universal guarantees. Report the number used, random seed, attainable P-value resolution and computational shortcuts.

Contents · Previous section · Next section

Section 14 of 36

14. Archive complete outputs

Keep all tested candidates, statistics, raw and adjusted P values, expression summaries and enrichment denominators. A filtered network alone prevents readers from distinguishing untested, unsupported and threshold-excluded routes.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented contrast table

This fictional table teaches interpretation; it is not scDiffCom output.

Contents · Previous section · Next section

Section 16 of 36

16. Read effect size beside significance

A small but precise shift and a large uncertain shift answer different questions. Report the differential statistic, uncertainty and specimen distribution with adjusted significance. Do not rank solely by P value.

Contents · Previous section · Next section

Section 17 of 36

17. Inspect emitter and receiver expression

Show ligand expression in the emitter and receptor or complex subunits in the receiver for each condition. A differential score should not hide whether the change arises from one side, both sides or altered population composition.

Contents · Previous section · Next section

Section 18 of 36

18. Interpret non-detection carefully

Failure to detect a candidate is not proof that communication is absent. Low counts, dropout, small cell populations or a missing database edge can cause non-detection. Use ‘not supported under this analysis’ rather than ‘no signalling’.

Contents · Previous section · Next section

Section 19 of 36

19. Compare condition labels honestly

A two-condition analysis is clearest when the contrast is biologically interpretable and sample processing is balanced. If condition is confounded with batch, technology or collection time, no permutation count can identify the biological cause.

Contents · Previous section · Next section

Section 20 of 36

20. Read enrichment denominators

An enriched ligand family or pathway depends on the universe of tested candidates. Publish that universe and correction method. Broadly annotated categories can appear frequently because they contain many eligible interactions.

Contents · Previous section · Next section

Section 21 of 36

21. Look for coherent evidence

A persuasive result combines a stable effect, replicated specimen support, complete complex expression and a plausible receiver response. Several related candidates can form a coherent hypothesis without implying every edge is active.

Contents · Previous section · Next section

Section 22 of 36

22. Audit label exchangeability

Write down exactly what was permuted. Condition labels across independent specimens may be valid; arbitrary shuffling across correlated cells often is not. Use sample-aware designs or complementary pseudobulk analyses where appropriate.

Contents · Previous section · Next section

Section 23 of 36

23. Audit cell annotations

Repeat key contrasts after merging uncertain subtypes and after excluding doublets or low-quality populations. A differential interaction driven by one unstable annotation should be labelled exploratory.

Contents · Previous section · Next section

Section 24 of 36

24. Audit composition

Downsample or model opportunity to examine whether unequal cell counts explain the result. Preserve sample balance. Composition sensitivity is not merely technical; it tells readers whether the claimed change concerns cell presence or per-cell communication evidence.

Contents · Previous section · Next section

Section 25 of 36

25. Audit the database ensemble

Because the resource aggregates multiple databases, inspect provenance for leading pairs and compare high-confidence subsets. A route supported only by one uncertain source should be prioritised differently from a repeatedly curated interaction.

Contents · Previous section · Next section

Section 26 of 36

26. Audit multiple testing

Thousands of emitter–receiver–interaction combinations create a large search space. Report the correction family and adjusted values. Post-hoc filtering should not silently redefine the hypothesis family after seeing the results.

Contents · Previous section · Next section

Section 27 of 36

27. Design condition-specific validation

Confirm ligand and receptor proteins in the relevant populations, then perturb the route in both conditions and measure an early receiver response. A differential claim needs evidence that the same intervention has different consequences, not merely different RNA abundance.

Contents · Previous section · Next section

Section 28 of 36

28. Write the bounded conclusion

Prefer: ‘scDiffCom identified this candidate as differentially supported between the two conditions under the declared annotations, resource and permutation design.’ Avoid claiming that altered physical signalling was directly measured.

Contents · Previous section · Next section

Section 29 of 36

29. Begin with Primary Science comparisons

Children learn by comparing like with like. Two plants, two temperatures or two materials make sense only when other factors are controlled. That is the same fairness scDiffCom needs when comparing conditions.

Contents · Previous section · Next section

Section 30 of 36

30. Use PSLE Science to teach controls

Students can list variables to keep constant and decide what counts as repeated evidence. A fictional communication table makes ‘fair test’ reasoning visible without requiring advanced statistics.

Contents · Previous section · Next section

Section 31 of 36

31. Connect Secondary and O-Level Biology

Cell receptors, coordination, ageing and gene expression provide the biological context. Differential communication shows how the same molecular vocabulary can be organised differently across conditions.

Contents · Previous section · Next section

Section 32 of 36

32. Let Computing explain permutations

A permutation test creates a reference distribution by rearranging labels under a justified rule. Students can simulate this with cards or code, then see why invalid shuffling produces misleading certainty.

Contents · Previous section · Next section

Section 33 of 36

33. Use science tuition for interpretation

Effective science tuition helps learners read axes, denominators and uncertainty, not just memorise conclusions. scDiffCom is a strong example of why ‘significant’ and ‘important’ are not synonyms.

Contents · Previous section · Next section

Section 34 of 36

34. Explore interdisciplinary careers

Single-cell studies involve biologists, clinicians, statisticians, software developers and data stewards. Career pathways grow from transferable reasoning and collaboration rather than mastery of one R package.

Contents · Previous section · Next section

Section 35 of 36

35. Evaluate school choices broadly

Families can look for verified school opportunities in experimentation, computing, statistics and research communication. Current programme details should come from official school sources; this method does not support claims about any school.

Contents · Previous section · Next section

Section 36 of 36

36. End with the useful question

The happiest scientific question is not ‘Which condition has more arrows?’ It is ‘Which difference survives fair comparison and tells us what to test next?’ That question keeps the network small enough to understand and strong enough to investigate.

The strongest scDiffCom report treats two conditions as a designed experiment rather than two folders of cells. Show specimen counts, cell-type coverage, sequencing depth and collection protocol side by side. If an emitter or receiver is scarce in one condition, state whether the candidate had enough opportunity to be tested fairly.

The paper’s curated resource drew from CellChat, CellPhoneDB, CellTalkDB, connectomeDB2020, ICELLNET, NicheNet and SingleCellSignalR. That breadth is useful, yet different databases encode complexes, direction and evidence differently. Preserve provenance so a reader can distinguish repeatedly supported biology from an interaction supplied by one source.

Permutation results should be reproducible. Save the seed and test statistic, and report the smallest attainable P value. With 1,000 permutations, tail resolution is coarser than with 10,000. The larger run improves numerical resolution; it does not make cell labels, resource edges or biological replicates more accurate.

Overrepresentation can turn thousands of significant triplets into an understandable theme, but it must use an honest background. A pathway enriched among significant interactions may simply contain many curated pairs. Report the tested universe, foreground size, correction and the individual candidates carrying the enrichment.

Did you know? The official workflow can also run detection without a differential comparison. That is useful for characterising one condition, but a detection-only result must not be described as condition-specific. The wording of the claim should match the question the statistical test actually asked.

A good validation ladder begins with protein localisation, proceeds to pathway-proximal receiver readouts, and ends with perturbation and rescue across both conditions. When the differential effect persists at each rung, the computational contrast becomes a sharper mechanistic explanation. When it does not, the negative result helps revise the network.

Before testing communication, audit differential expression within each cell type. If nearly every ligand and receptor changes because of a global sequencing-depth shift, communication statistics may simply repackage that technical contrast. Use balanced preprocessing, inspect expression distributions and verify that the candidate change remains after specimen-aware normalisation or pseudobulk checks.

Complexes need explicit completeness rules. For a heteromeric receptor, list every required subunit and show its detection by condition and specimen. A strong signal from one chain cannot compensate biologically for an absent essential chain, even if an averaging rule produces a large score. Sensitivity to minimum, mean or geometric combination should be reported.

Condition differences can also arise from altered opportunity. Suppose a receiver population becomes rare in condition B: there are fewer cells in which receptor expression can be measured, and the variance of the score changes. Report eligible emitter–receiver counts, use matched subsampling where informative and avoid treating a sparse comparison as evidence of biological disappearance.

Permutation exchangeability should be demonstrated, not assumed. If donors contribute both conditions, permutations can respect the pairing; if donors are condition-specific, the specimen is the unit that carries condition. Shuffling condition labels cell by cell destroys dependence and creates an unrealistically large sample. State the permutation block and justify it from the experiment.

Multiple-testing correction should follow the declared family of questions. If testing every cell-pair and interaction combination, correct over that full family or present a principled hierarchical strategy. Filtering after seeing P values can inflate discovery. Predefined expression and cell-count filters are easier to defend and reproduce.

Compare scDiffCom results with transparent baselines. Show condition-specific ligand and receptor pseudobulk changes, a simple product score and, where relevant, another differential network method such as CrossTalkeR. Agreement strengthens prioritisation; disagreement identifies whether topology, resource choice or permutation logic drives the claim.

For longitudinal ageing or treatment studies, two-condition contrasts may be only one slice of a trend. Repeat adjacent contrasts or model ordered time where design permits. A route that changes smoothly across ages differs biologically from one that spikes at a single collection point. Pairwise software should not force a continuous process into an oversimplified story.

Publish a candidate-level audit sheet with emitter, receiver, interaction provenance, complex composition, specimen effect, cell opportunity, statistic, adjusted P value, enrichment membership and validation status. The sheet lets another team reproduce why a route was prioritised and prevents an attractive network diagram from becoming the only record of analysis.

Do not compare raw counts of significant interactions when the tested opportunity differs. A condition with more cells, more detectable genes or more eligible emitter–receiver combinations can produce more findings without stronger biology. Use opportunity-aware proportions, effect distributions and specimen recurrence, then state precisely what denominator supports a network-level comparison.

Sensitivity results should be planned around plausible choices: minimum cells per group, expression thresholds, simple versus complex interaction handling, database subsets, permutation blocks and multiple-testing correction. Summarise which claims survive all reasonable settings and which depend on one decision. That stability map is often more informative than a single final network.

For teaching, let students act as condition labels and interaction candidates. Shuffle only whole specimen envelopes, never individual cell cards, and compare the null distribution. The physical exercise makes dependence visible and shows why a huge number of cells cannot replace independent biological replication.

Use careful public wording. Say ‘differentially supported candidate’, ‘permutation-based evidence’ and ‘enriched category’; reserve ‘changed signalling’ for experiments that directly address molecular function. This vocabulary is not timid. It keeps the computation and biology connected while preserving the value of both.

Before release, check every link and source date, confirm that the article distinguishes the 2023 paper from current package documentation, and label all toy values as invented. Then read each conclusion backwards: can every phrase be traced to a displayed statistic, a primary source or an explicit experimental result? This final provenance pass protects both readers and researchers.

The method’s real educational gift is disciplined comparison. It invites learners to ask whether two conditions had equal opportunity, whether the null model was fair, whether the difference repeated across specimens and which experiment could separate molecular change from altered composition. Those questions travel far beyond single-cell biology.

Contents · Previous section · Continue to the Science Learning Hub