Why Science? | LIANA+, Multi-Method Consensus and Cell-Communication Evidence

Section 1 of 36

1. Begin with partial views

Cell–cell communication methods ask related questions with different scores, thresholds and assumptions. LIANA+ provides a common framework so those partial views can be run, compared and integrated. Unification improves auditability; it does not prove that the shared biological premise is correct.

Contents · Next section

Section 2 of 36

2. Separate method from resource

An algorithm determines how expression becomes a score, while a resource determines which interactions are available. Treat these as two experimental factors. If both change together, a different answer cannot be attributed to method or knowledge base.

Contents · Previous section · Next section

Section 3 of 36

3. Standardise without erasing meaning

Common inputs and outputs make methods comparable, but their native statistics still mean different things. A magnitude, specificity score, correlation and rank are not interchangeable units. Preserve method definitions before aggregating them.

Contents · Previous section · Next section

Section 4 of 36

4. Use consensus as evidence synthesis

LIANA+ can combine ranks from several methods into consensus results. Rank aggregation reduces dependence on one algorithm, but methods may share expression summaries and databases. Agreement is not independent replication.

Contents · Previous section · Next section

Section 5 of 36

5. Work across single-cell data

For dissociated data, cell identities and group-level expression define possible senders and receivers. Communication remains a compatibility inference. Specimen structure, abundance and annotation quality should be carried through every method.

Contents · Previous section · Next section

Section 6 of 36

6. Extend to spatial and multi-omics data

LIANA+ includes components for spatially resolved and multi-omics settings, where local or global relationships can be investigated. Coordinates and additional modalities enrich the question while adding alignment, resolution and measurement assumptions.

Contents · Previous section · Next section

Section 7 of 36

7. Compare conditions and programmes

The framework includes strategies for differential communication and unsupervised programme discovery across conditions. These analyses need biological replication and careful factor interpretation. A latent programme is a compact summary, not a named mechanism by default.

Contents · Previous section · Next section

Section 8 of 36

8. Build a standard input manifest

Record AnnData or MuData layers, gene identifiers, raw and transformed matrices, cell and specimen metadata, spatial coordinates, conditions, resources, methods and software environment. Standard objects are helpful only when their semantics are declared.

Contents · Previous section · Next section

Section 9 of 36

9. Audit identities and sample balance

Every method inherits cell labels and sampling. Review rare populations, mixed clusters, donor imbalance and batch effects before building consensus. A consensus of analyses on the same mislabeled cells is still wrong.

Contents · Previous section · Next section

Section 10 of 36

10. Keep raw and transformed layers straight

Document which layer each method consumes and whether it expects counts, normalised values or another representation. Accidentally mixing scales can yield plausible tables with invalid meanings. Preserve the transformation chain in the object and report.

Contents · Previous section · Next section

Section 11 of 36

11. Version interaction resources

LIANA+ can use knowledge bases such as OmniPath-derived resources and custom sets. Save exact files, evidence filters, organism mapping and complex rules. Resource scope is part of the scientific question, not a software footnote.

Contents · Previous section · Next section

Section 12 of 36

12. Predeclare method panels

Choose methods because their assumptions address the research question, not because their consensus flatters a preferred pathway. Explain which families are represented and where methods share inputs or statistics.

Contents · Previous section · Next section

Section 13 of 36

13. Define aggregation before inspection

Specify how ranks, missing interactions and ties are handled. Decide whether the goal is robust screening, resource comparison or method comparison. Changing aggregation after seeing results turns consensus into an aesthetic choice.

Contents · Previous section · Next section

Section 14 of 36

14. Archive interoperable outputs

Save input objects, method-specific long tables, resource records, consensus ranks, spatial metrics, condition models, factor loadings, versions, seeds and plotting code. Do not preserve only the consensus shortlist.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented consensus audit table

This fictional table is for reasoning, not software benchmarking.

Contents · Previous section · Next section

Section 16 of 36

16. Read method-specific scores first

Before inspecting consensus, learn what each method ranks and whether higher or lower is stronger. Retain magnitude and specificity alongside ranks. Consensus should sit above interpretable components, not replace them.

Contents · Previous section · Next section

Section 17 of 36

17. Read rank consensus carefully

A high consensus rank means several included procedures prioritised an interaction relative to others. It does not provide a universal probability or effect size. Report the panel and aggregation rule every time consensus is shown.

Contents · Previous section · Next section

Section 18 of 36

18. Compare resource consensus separately

An interaction supported across independent curated resources addresses knowledge-base robustness, whereas method consensus addresses computational robustness. Crossing these dimensions gives a clearer audit than a single blended score.

Contents · Previous section · Next section

Section 19 of 36

19. Use spatial metrics at the right scale

Spatial relationships can be local, global or multiview. State coordinate units, neighbourhood rules and whether spots contain mixtures. A spatially coherent association still requires molecular and perturbational validation.

Contents · Previous section · Next section

Section 20 of 36

20. Analyse conditions with replication

Differential communication should respect specimen, condition and batch structure. Separate altered cell abundance from altered interaction evidence. Use hierarchical or pseudobulk strategies where appropriate and retain per-specimen estimates.

Contents · Previous section · Next section

Section 21 of 36

21. Interpret latent factors modestly

Matrix or tensor factorizations can reveal coordinated interaction programmes across cell pairs, conditions or samples. Factor labels are interpretations supplied after fitting. Show loadings, stability and alternative factor numbers.

Contents · Previous section · Next section

Section 22 of 36

22. Audit method sensitivity

Remove one method at a time and recompute the shortlist. Compare families with shared versus distinct assumptions. An interaction that collapses after one method leaves is not broad consensus, even if the original rank was high.

Contents · Previous section · Next section

Section 23 of 36

23. Audit resource sensitivity

Repeat with alternative or high-confidence resources under matched methods. Track absent interactions separately from changed scores. Missingness reflects curation scope and should not be confused with negative expression evidence.

Contents · Previous section · Next section

Section 24 of 36

24. Audit aggregation sensitivity

Compare robust rank aggregation with another prespecified consensus rule and sensible handling of ties. Report the stable core. If top candidates reorder completely, the scientific conclusion should be about uncertainty.

Contents · Previous section · Next section

Section 25 of 36

25. Validate outside the framework

Use protein, spatial, perturbational or functional assays that do not reuse the same transcript counts and databases. Independent evidence is what converts multi-method agreement into stronger biology.

Contents · Previous section · Next section

Section 26 of 36

26. Compare a transparent baseline

A simple ligand–receptor expression product or correlation can reveal whether elaborate methods add stable prioritisation. The comparison must use identical cells, resources and filters. Complexity earns trust through validation, not appearance.

Contents · Previous section · Next section

Section 27 of 36

27. Make disagreement visible

A disagreement matrix can be more informative than a consensus list. Trace whether differences arise from magnitude, specificity, complex handling, spatial weighting or resource coverage. Preserve minority methods when they expose a biologically relevant assumption.

Contents · Previous section · Next section

Section 28 of 36

28. Turn consensus into a test portfolio

Choose one high-consensus candidate, one resource-sensitive candidate and one method-sensitive candidate. Design matched follow-ups. This converts computational uncertainty into an efficient experimental portfolio rather than hiding it.

Contents · Previous section · Next section

Section 29 of 36

29. Build a classroom model before code

Let teams rank the same fictional interactions using different transparent rules, then combine ranks and identify which agreements share the same underlying evidence. Label observations, transformations, assumptions and conclusions separately. The goal is not to imitate specialist software; it is to make every evidence hand-off visible enough for a classmate to question.

Contents · Previous section · Next section

Section 30 of 36

30. Begin with Primary Science habits

Primary Science already supplies the foundation for LIANA+: careful observation, fair comparison, consistent records and conclusions that fit the evidence. A simple plant, light or water investigation can show why compatible parts do not prove that an interaction actually occurred.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science to practise claim limits

PSLE Science asks students to connect evidence to process without leaping beyond an experiment. With LIANA+, ask what was measured, what came from a database, what the software estimated and which new observation could separate two explanations.

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, molecular transport, experimental design and evaluation. LIANA+ turns those ideas into a contemporary data problem. The useful lesson is disciplined interpretation, not memorising a software menu.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing support the biology

Computing contributes tables, graphs, algorithms, ranking, version control and reproducibility. Biology supplies specimens, mechanisms and validation. LIANA+ shows why correct code is necessary but insufficient: a flawless program can still analyse confounded samples or unsuitable labels.

Contents · Previous section · Next section

Section 34 of 36

34. Choose opportunities by verified fit

For school choices and science enrichment, verify current programmes on official school and MOE pages. Value opportunities that teach comparison, reproducibility and respectful disagreement between models, alongside biological laboratory reasoning. A fashionable mention of genomics or AI does not guarantee sustained teaching, admission, mentorship or a particular career outcome.

Contents · Previous section · Next section

Section 35 of 36

35. See connected career families

The reasoning in LIANA+ appears in computational biology, multi-omics, data engineering, network science, biostatistics, molecular biology and research software. Routes may pass through polytechnic, junior college, university or continuing education, with different blends of biology, statistics, computing and communication. Requirements change, so check the responsible institution directly.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a bounded scientific claim

A defensible conclusion states exactly what LIANA+ prioritised under which data, database, settings and specimens. LIANA+ standardises and integrates multiple prior-informed communication analyses; consensus cannot replace independent specimens, orthogonal measurements or causal perturbation. Report uncertainty, sensitivity, replication, negative results and independent validation together; those boundaries make a claim more useful.

LIANA+ is most powerful when it makes assumptions comparable. Begin with a method–resource matrix that lists every combination actually run, what each score means and which inputs are shared. This prevents a consensus from appearing to represent more independent evidence than it does.

Standard objects such as AnnData and MuData improve interoperability, but object names do not guarantee correct layers. Record counts, normalised expression, log transformations, batch corrections, cell metadata and spatial coordinates explicitly. A method consuming the wrong layer can return polished but meaningless results.

Select the method panel before inspecting desired pathways. Include methods with complementary ideas—such as magnitude, specificity or correlation—only when their assumptions suit the data. Explain exclusions. Eight procedures that all reward the same group average are not eight independent experiments.

Resource choice deserves a parallel plan. Compare at least a declared high-confidence or alternative interaction set when the main claim depends on curation. Track three outcomes separately: interaction absent from a resource, interaction present but unsupported by expression, and interaction present with a changed rank.

Consensus ranks are relative. They say that an interaction performs well across the included rankings, not that it has a universal effect size. Present method-specific values and ranks beside consensus. Readers should be able to see whether agreement is broad or driven by one family of similar methods.

Missing results require a rule. A method may not score an interaction because the resource excluded it, the expression filter removed it or the statistic was undefined. These cases should not all receive the same worst rank. Document handling and test whether alternatives alter the shortlist.

Did you know? Agreement between algorithms can be correlated because they reuse the same transcript counts, cell labels and ligand–receptor resource. Consensus reduces algorithm-specific noise, but it cannot create new specimens or new molecular measurements.

For spatial analyses, distinguish local metrics, global relationships and multiview models. State coordinate units, neighbourhood masks, tissue boundaries and spot composition. A global spatial association can coexist with weak contact-scale evidence; neither should be silently translated into the other.

For multi-condition data, preserve specimens and use models that respect the experimental design. Differential interaction scores can reflect changes in cell abundance, expression, composition or measurement depth. Decompose those contributors before naming altered communication.

Unsupervised factorisation can reveal recurring programmes across cell pairs and conditions. Test the number of factors, seeds and stability; show loadings rather than only labels. A factor named after one pathway may contain many interactions and should remain descriptive until validated.

Linking extracellular interactions to intracellular pathways and TF activities produces an appealing end-to-end story. Each connection adds prior knowledge and potential ambiguity. Trace the route, display alternative paths and avoid calling the network causal merely because its signs are coherent.

Leave-one-method-out analysis is a practical stress test. Recalculate consensus after removing each method and after removing groups of closely related methods. Report rank ranges. A robust candidate remains near the top; a fragile candidate reveals which procedure owns the conclusion.

Leave-one-resource-out and high-confidence subsets provide a different stress test. Preserve interactions missing by definition, rather than scoring them as biological negatives. Stable conclusions across resources are stronger; unstable ones spotlight curation gaps that may be worth manual review.

Specimen-level replication remains decisive. Run or summarise method support per specimen, then aggregate transparently. Thousands of cells from one donor cannot replace independent samples, and consensus across algorithms cannot compensate for absent replication.

Compare with a transparent baseline under the same resource and filtering. A simple expression product, prevalence rule or correlation makes the incremental benefit of each method visible. If consensus cannot outperform a baseline on held-out or orthogonal evidence, extra complexity has not earned a stronger claim.

Validation must leave the computational family. Use protein localisation, spatial co-occurrence, perturbation, functional readouts or matched modalities. An additional transcript-based algorithm is sensitivity analysis, not orthogonal validation. Name which inferential gap every assay addresses.

Archive the input objects, data dictionary, method–resource matrix, software environment, individual long-format results, consensus procedure, missing-value rules, spatial settings, condition models, factors, sensitivity runs, seeds and figures. Interoperability should make the evidence more portable, not merely the code easier to run.

A defensible sentence might read: “LIANA+ identified a consensus-prioritised interaction that remained stable across the declared method and resource panel, appeared in most specimens and agreed with independent spatial and protein evidence.” It should not describe consensus as proof of binding or causality.

End with a portfolio rather than one winner. Choose a robust candidate, a method-sensitive candidate and a resource-sensitive candidate for matched validation. Success and failure across this trio teach whether the consensus is biologically calibrated and where the framework’s assumptions matter most.

Treat computational scalability as an engineering property rather than biological validation. Fast execution and standard interfaces help teams repeat analyses, but they do not improve specimen design, cell labels or molecular measurement by themselves. Report runtime and memory when operationally useful while keeping biological evidence on a separate axis.

Use a provenance graph for the final shortlist. For each interaction, connect the original expression layer, specimen, cell label, resource record, method-specific score, consensus transformation and validation result. This makes shared dependencies visible. When several methods trace back to the same group average and database entry, their apparent plurality should not be presented as independent confirmation.

Finally, distinguish framework benchmarking from study-specific truth. Published performance demonstrates that components can recover structure in reported settings; it does not guarantee that the top rank in a new tissue is correct. New specimens need their own controls, sensitivity analysis and validation, even when the software has strong documentation and peer-reviewed examples.

Contents · Previous section · Continue to the Science Learning Hub