Why Science? | DcjComm, Joint Cell Clustering and Communication-Network Evidence

Section 1 of 36

1. See the joined workflow

DcjComm was designed to connect functional-module discovery, dimension reduction, cell clustering and communication inference within one analytical framework. That coherence can reduce arbitrary hand-offs. It does not remove uncertainty from preprocessing, factorisation, clustering, prior networks or sample design.

Contents · Next section

Section 2 of 36

2. Keep the 2024 paper boundary

The method appeared in Genome Biology in 2024 and was evaluated across public datasets and task-specific comparisons. Performance claims belong to those datasets, metrics and settings. A new study should validate both clustering and communication outputs independently.

Contents · Previous section · Next section

Section 3 of 36

3. Use NMF-based joint learning

The first component uses non-negative-matrix-factorisation-based learning so gene modules, low-dimensional representation and clusters inform one another. Non-negativity can aid interpretation. The fitted factors remain an approximation influenced by rank, scale and initialisation.

Contents · Previous section · Next section

Section 4 of 36

4. Discover functional modules

Modules group genes with coordinated contributions under the model. Inspect member genes, stability and enrichment rather than naming a module from one marker. A module may reflect cell identity, treatment, cell cycle, stress or technical variation.

Contents · Previous section · Next section

Section 5 of 36

5. Reduce dimensions transparently

The learned representation compresses expression for clustering and visualisation. A two-dimensional plot is a view, not the full evidence. Preserve factor loadings, distances and diagnostics so attractive separation does not substitute for quantitative cluster evaluation.

Contents · Previous section · Next section

Section 6 of 36

6. Cluster cells with uncertainty

Clusters provide candidate cell groups for communication analysis. Check marker coherence, stability, doublets, transitional states and specimen balance. A cluster boundary created by rank or resolution can manufacture a sender or receiver group.

Contents · Previous section · Next section

Section 7 of 36

7. Score communication with two layers

The communication component combines ligand–receptor expression with downstream transcription-factor activity and targets. Joining extracellular and intracellular evidence can prioritise functional routes. It also means an edge inherits uncertainty from both layers.

Contents · Previous section · Next section

Section 8 of 36

8. Create an end-to-end manifest

Record species, specimens, assay, filtering, normalisation, gene universe, NMF ranks, initialisations, cluster settings, ligand–receptor resource, receptor–TF and TF–target networks, scoring thresholds and DcjComm version. Coherence is reproducible only when the whole chain is versioned.

Contents · Previous section · Next section

Section 9 of 36

9. Audit preprocessing

Filtering, normalisation and variable-gene selection influence modules and clusters before communication scoring begins. Repeat key findings under justified alternatives. A later network score cannot repair a sender group that exists only under one preprocessing choice.

Contents · Previous section · Next section

Section 10 of 36

10. Audit rank and cluster number

Too few factors or clusters merge distinct biology; too many split stable populations and inflate possible cell pairs. Report stability across neighbouring choices and show whether headline communication routes persist after sensible merges or splits.

Contents · Previous section · Next section

Section 11 of 36

11. Freeze prior networks

The published framework draws receptor-to-TF and TF-to-target knowledge from curated pathway resources such as Reactome and KEGG. Archive the exact edges and versions. Network updates can change intracellular evidence while sample expression remains the same.

Contents · Previous section · Next section

Section 12 of 36

12. Archive every joint output

Retain modules, factor matrices, embeddings, cluster labels, marker summaries, ligand–receptor scores, TF activities, target tests and complete route tables. A final network figure cannot reveal which upstream decision created an edge.

Contents · Previous section · Next section

Section 13 of 36

13. Protect specimen independence

Cells are nested within biological specimens. Check modules, clusters and routes within donors or animals and report prevalence across them. Pooled cell counts do not replace independent replication, even when the joint model is mathematically sophisticated.

Contents · Previous section · Next section

Section 14 of 36

14. Plan task-specific controls

Use clustering controls for cell identity, network controls for molecular routes and specimen-aware nulls for group comparisons. One global shuffle rarely tests every stage fairly. A coherent pipeline still needs controls matched to each scientific claim.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented joint-evidence table

This fictional table is not DcjComm output. It shows how a strong final score can hide different upstream weaknesses. Inspect every layer before choosing a validation experiment.

Contents · Previous section · Next section

Section 16 of 36

16. Read functional modules first

Before examining communication, identify which modules define the sender and receiver clusters. Ask whether they reflect biology relevant to the route or a confounder such as cell cycle. This connects cell identity to the later network claim.

Contents · Previous section · Next section

Section 17 of 36

17. Read the embedding cautiously

Neighbourhoods in a reduced space depend on the learned representation. Quantify separation and stability rather than relying on visual distance. Cells plotted nearby are not necessarily spatial neighbours in tissue or direct communication partners.

Contents · Previous section · Next section

Section 18 of 36

18. Read cluster labels as hypotheses

Assign cell types with multiple markers and, where possible, external references. Preserve uncertain and transitional labels. Communication between two named clusters sounds mechanistic, so the names must be better supported than a single marker.

Contents · Previous section · Next section

Section 19 of 36

19. Decompose the communication score

The method combines an intercellular ligand–receptor component with intracellular TF evidence. Publish both values and their ingredients. A product can be small because either layer is weak and can appear impressive when one layer dominates.

Contents · Previous section · Next section

Section 20 of 36

20. Inspect transcription-factor tests

The published model uses target information and statistical testing to assess TF activity. Report the target set, background, correction and specimen support. Enrichment is not direct TF protein activation and should be validated when central to the claim.

Contents · Previous section · Next section

Section 21 of 36

21. Compare specimens before conditions

Estimate route support within each independent specimen, then compare groups. Check whether the same sender and receiver clusters exist in all specimens. A condition effect can otherwise be a cell-composition or label difference.

Contents · Previous section · Next section

Section 22 of 36

22. Audit clustering sensitivity

Repeat communication scoring under stable alternative ranks, resolutions or broader labels. Track whether leading routes and their molecular members persist. A route that vanishes after a small clustering change should be labelled boundary-sensitive.

Contents · Previous section · Next section

Section 23 of 36

23. Audit network sensitivity

Recompute headline routes with high-confidence edges or a compatible alternative knowledge resource. Shared conclusions are stronger; network-specific routes reveal prior dependence. Do not choose a pathway database after seeing which one supports the preferred story.

Contents · Previous section · Next section

Section 24 of 36

24. Compare simpler baselines

Compare DcjComm with expression-only ligand–receptor scoring, independent clustering and direct target associations. The joint workflow adds value when it produces a stable, experimentally discriminating prediction that simpler approaches miss.

Contents · Previous section · Next section

Section 25 of 36

25. Publish negative results

Show routes that fail because clusters are unstable, ligand or receptor expression is absent, TF evidence is weak, specimens disagree or prior networks lack support. These negatives reveal what the joint model actually filters.

Contents · Previous section · Next section

Section 26 of 36

26. Add protein, space and time

Confirm relevant proteins, anatomical opportunity and temporal order. RNA-based joint evidence does not measure secretion, surface receptor, pathway activation or signal travel. Choose orthogonal assays for the weakest step rather than repeating another transcriptomic analysis.

Contents · Previous section · Next section

Section 27 of 36

27. Design an end-to-end perturbation

Perturb a leading ligand or receptor and measure an early receiver event, the nominated TF programme and later target response in the predicted cluster. Include an alternative route and a cluster identity check. This tests both the network and the cell grouping.

Contents · Previous section · Next section

Section 28 of 36

28. Write a bounded DcjComm claim

Prefer: ‘DcjComm identified a cluster-stable route with ligand–receptor and TF-target support across these specimens.’ Avoid: ‘The joint model proved causal communication.’ The bounded wording is both more accurate and easier to test.

Contents · Previous section · Next section

Section 29 of 36

29. Build a classroom pipeline

Give students stages—clean data, group observations, identify a possible interaction, trace a response and design a test. Let one error enter at an early stage and ask how it affects later conclusions. This demonstrates why coherence does not erase upstream uncertainty.

Contents · Previous section · Next section

Section 30 of 36

30. Begin with Primary Science systems

Primary learners already group observations, compare systems and record evidence. They can see that a conclusion depends on the quality of each earlier step. Joint reasoning is simply careful scientific bookkeeping at a larger scale.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science for cause and effect

Students can identify which evidence supports cell identity, interaction or response, then design a fair test that distinguishes two routes. They learn that several correlated observations do not automatically establish causation.

Contents · Previous section · Next section

Section 32 of 36

32. Connect Secondary and O-Level Science

Cells, coordination, gene expression, variables and experimental design provide the biological foundation. DcjComm offers a modern example in which classification and mechanism are analysed together but still require separate evidence.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing make the pipeline auditable

Computing contributes matrix factorisation, clustering, graphs, statistical tests, version control and reproducible workflows. Biology determines whether modules, cell labels and molecular routes make sense. Both are necessary for a defensible result.

Contents · Previous section · Next section

Section 34 of 36

34. Choose programmes by verified fit

When considering science enrichment or school choices, check current official details and look for sustained opportunities in experimental science, data analysis, mathematics, computing and research practice. Avoid promises about admission or outcomes.

Contents · Previous section · Next section

Section 35 of 36

35. See connected careers

The workflow connects to bioinformatics, systems biology, statistics, data science, research software, immunology and drug discovery. Education routes vary and requirements evolve, so learners should verify current institutional and employer sources.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with uncertainty at every layer

A defensible DcjComm conclusion names preprocessing, modules, ranks, clusters, prior networks, component scores, specimens and sensitivity checks, then separates joint computational evidence from protein, spatial, temporal and perturbation validation. Coherence helps most when each uncertainty remains visible.

DcjComm’s appeal is that it does not treat cell clustering and communication analysis as unrelated chores. Functional modules, a reduced representation and cluster assignments are learned coherently before ligand–receptor and transcription-factor evidence are combined. The joined workflow can reduce hand-off inconsistency while making upstream sensitivity especially important.

Did you know? A final communication score can depend on decisions made before any ligand–receptor pair is considered. If factor rank changes a cluster boundary, the sender and receiver expression summaries change, which can rerank the network. End-to-end coherence therefore increases the value of end-to-end auditing.

Start with the non-negative factorisation. Examine reconstruction, stability across seeds and the gene loadings that define each module. Distinguish modules reflecting identity from those driven by cell cycle, stress or sequencing depth. A cluster derived from a technical module should not become a biological actor in a communication story.

Cell labels deserve independent support. Use several markers, reference mapping where suitable and specimen prevalence. Preserve transitional or uncertain groups rather than forcing every cell into a crisp category. Then repeat leading routes after sensible cluster merges or splits to see whether conclusions depend on a narrow boundary.

The communication component combines extracellular and intracellular evidence. The ligand and receptor should be expressed in the nominated groups; the downstream transcription factor should have supported targets in the receiver. Publish both components and the target set. A product score is concise, but it can hide one weak side.

Prior networks from resources such as Reactome and KEGG provide valuable receptor-to-TF and TF-to-target links. They are versioned knowledge, enriched for well-studied pathways and not tissue-specific measurements. Compare high-confidence or alternative networks and label routes that depend on one uncertain edge.

Study design remains outside the algorithm. Cells are nested within specimens, treatments may be paired and cell abundance can differ by condition. Estimate modules, cluster stability and route support at the specimen level. A model cannot manufacture independent replication from more cells.

Use task-matched baselines. Compare cluster quality with established clustering approaches, and compare communication routes with expression-only or independent-network methods. DcjComm adds practical value when its joined evidence produces a stable, discriminating hypothesis that simpler baselines do not reveal.

Validation should traverse the chain. Confirm sender and receiver identities, ligand and receptor proteins, a receptor-proximal event, the nominated TF programme and later targets. Perturb the route under matched timing and include a competing pathway. This identifies whether failure lies in the cell grouping, extracellular edge or intracellular mechanism.

Archive all intermediate objects. If a future analyst changes normalisation, rank, cluster number or knowledge networks, they should be able to determine which downstream routes changed and why. A single network image cannot provide that audit trail.

For students, DcjComm is a lesson in pipelines and systems. Primary Science supplies grouping and observation; PSLE Science develops fair tests; Secondary and O-Level Biology add cells and regulation; Computing makes factorisation, clustering, graphs and versioning explicit. The sophisticated part is not avoiding uncertainty but carrying it honestly.

The bounded conclusion is that DcjComm can coordinate functional modules, clustering and multi-layer communication evidence into a coherent set of hypotheses. It does not turn coherence into causation. Stable cell identities, specimen replication and orthogonal molecular experiments remain essential.

An end-to-end sensitivity map can make the method especially informative. Change one justified decision at a time—normalisation, NMF rank, cluster number, prior-network confidence or communication threshold—and record which modules, labels and routes change. The map reveals whether uncertainty enters early and propagates or appears only at the final score.

Keep discovery and evaluation separate. If the same data choose the factors, define clusters, select network routes and demonstrate separation, apparent performance can be optimistic. Use held-out specimens, external labels or preregistered criteria where feasible. A coherent workflow benefits from an equally coherent evaluation design.

Publish component denominators: genes entering modules, cells and specimens in each cluster, eligible ligand–receptor pairs, TF target-set sizes and routes surviving correction. These counts help readers recognise when a striking score is supported by broad evidence or by a very small corner of the data.

Finally, compare whether the communication conclusion persists after replacing the joint clusters with independently validated labels. Agreement supports robustness; disagreement does not automatically invalidate DcjComm, but it shows that the network interpretation is coupled to its clustering solution and should be tested accordingly.

Do not let an integrated score erase direction. Preserve which cluster supplies the ligand, which supplies the receptor, which TF is evaluated in the receiver and which targets support that TF. A reversed cell label or network edge can produce a convincing but biologically wrong story, so hand-audit every headline route end to end.

A balanced report should show one route stable across the complete pipeline, one cluster-sensitive route, one prior-network-sensitive route and one negative control. This makes the advantages and limits of joint modelling visible instead of presenting only successful edges.

Use informative failures to improve the workflow. If a route fails protein validation, revisit extracellular evidence; if the cells are mislabelled, revisit factors and modules; if the TF response is absent, revisit the intracellular network. Because the pipeline is joined, a failed experiment can identify which analytical layer needs correction.

The result is a workflow that can learn responsibly from both scientific success, analytical uncertainty and experimental disagreement.

Contents · Previous section · Continue to the Science Learning Hub