Why Science? | MELD, Graph Signal Processing and Continuous Perturbation Likelihood

Section 1 of 36

1. Begin with overlapping states

Perturbed and control cells may occupy the same broad clusters while differing gradually along activation or developmental continua. A cluster count can therefore average away the very region where the experiment matters.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Next section

Section 2 of 36

2. Ask a distribution question

Instead of asking only which cluster contains more treated cells, ask where each experimental condition is relatively concentrated across the shared phenotypic landscape.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 3 of 36

3. Build a shared representation

Combine relevant cells in a feature space that reflects measured biology. The representation should not use labels in a way that guarantees the separation later reported.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 4 of 36

4. Connect local neighbours

Construct a cell graph so nearby profiles share information. A graph formalises local similarity; it is neither physical contact nor proof of lineage.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 5 of 36

5. Encode sample labels

Represent condition membership as signals on graph vertices. The labels describe observed experimental assignment or sample origin, not a molecular mechanism.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 6 of 36

6. Smooth with discipline

Graph signal processing reduces label noise by borrowing information from nearby cells. Excessive smoothing can erase narrow states, while too little smoothing can leave sampling noise dominant.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

Contents · Previous section · Next section

Section 7 of 36

7. Read the 2021 study in scope

Burkhardt and colleagues described MELD as a continuous measure of experimental perturbation effects across transcriptomic state space and released an open-source Python implementation.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Previous section · Next section

Section 8 of 36

8. Define the experimental unit

State whether labels come from donors, animals, cultures, time points or technical batches. The design determines what can generalise beyond collected cells.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 9 of 36

9. Audit sample balance

Large differences in captured cell numbers can affect density estimates. Record downsampling, normalisation and whether rare conditions remain adequately represented.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 10 of 36

10. Choose features before outcomes

Highly variable genes, principal components or proteins define neighbourhoods. Select them through a reproducible rule rather than tuning until a preferred condition gradient appears.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 11 of 36

11. Check batch structure

If batch aligns with condition, density differences may be technical. Balanced processing or a justified correction strategy is stronger than post-hoc reassurance.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 12 of 36

12. Inspect graph connectivity

Disconnected components, hubs and bridges can reshape smoothing. Compare neighbour numbers and dimensions, and explain whether important states remain locally coherent.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

Contents · Previous section · Next section

Section 13 of 36

13. Estimate sample-associated density

MELD turns condition labels into smooth density estimates over the graph, offering a cell-level view of where each condition is represented.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Previous section · Next section

Section 14 of 36

14. Normalise to relative likelihood

For two or more conditions, compare sample-associated densities so each cell receives a relative-likelihood profile. The denominator and represented conditions must be stated.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with a fictional MELD table

This classroom example is invented, not a result from the MELD paper.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 16 of 36

16. Read a high relative likelihood

A high treated likelihood means similar observed cells are relatively enriched for the treated label under the model. It does not say treatment created the state or that every such cell responded.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 17 of 36

17. Read a gradient

A smooth rise can suggest a progressive shift across state space. Validate it against markers, time, replicates and preprocessing before giving it biological language.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 18 of 36

18. Inspect rare regions

Sparse neighbourhoods can produce unstable densities. Show cell counts and sample support, not only extreme colours at the edge of an embedding.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

Contents · Previous section · Next section

Section 19 of 36

19. Compare biological replicates

Recalculate patterns by donor or experiment. A pooled gradient driven by one specimen should be described as specimen-specific until replicated.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Previous section · Next section

Section 20 of 36

20. Quantify uncertainty

Resample biological units where possible and examine stability across graph settings. A beautiful deterministic map can conceal fragile sampling.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 21 of 36

21. Run label permutations

Shuffle labels within appropriate design blocks. The permuted maps show how much apparent enrichment the workflow produces when condition association is absent.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 22 of 36

22. Compare cluster proportions

Traditional cluster counts answer a coarser question. Agreement is reassuring; disagreement can reveal a continuous subpopulation that deserves targeted validation.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 23 of 36

23. Compare with Milo

Milo formally tests sample-level counts in overlapping neighbourhoods. MELD supplies a continuous relative-likelihood landscape and should not be described as the same inferential procedure.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 24 of 36

24. Compare with Augur

Augur ranks cell-type responsiveness using classifier separability. MELD maps where condition-associated density changes occur across a shared manifold.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

Contents · Previous section · Next section

Section 25 of 36

25. Compare with Mixscape

Mixscape classifies perturbed and non-perturbed cells in pooled screens. MELD estimates relative condition likelihood without asserting successful editing in each cell.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Previous section · Next section

Section 26 of 36

26. Build a likelihood evidence card

Place the graph, conditions, sample balance, relative-likelihood range, replicate stability, marker interpretation and next validation on one compact record.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 27 of 36

27. Design the next experiment

Collect independent samples with balanced processing, preselect markers for high- and low-likelihood regions and add a time point that could distinguish selection from transition.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 28 of 36

28. Choose claim-matched validation

Flow cytometry, targeted sequencing, imaging or functional assays should directly test the biological interpretation, not merely recreate the same graph from the same data.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 29 of 36

29. Connect MELD to school science

The method makes variables, fair comparisons and continuous measurement tangible: the answer is a gradient of evidence rather than a forced yes-or-no label.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 30 of 36

30. Try a neighbour-smoothing activity

Give students condition-labelled dots on a map, average labels within local neighbourhoods and discuss when smoothing reveals structure or hides small islands.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

Contents · Previous section · Next section

Section 31 of 36

31. Connect to mathematics

Graphs, weighted averages, probability densities and ratios turn local patterns into quantities whose assumptions can be inspected.

MELD treats each condition as a distribution over a shared cell-state manifold and uses a graph to smooth sample-associated signals while preserving local structure.

Archive samples, labels, preprocessing, feature choices, graph parameters, random seeds, software versions and every exclusion so another analyst can reproduce the decision path.

Contents · Previous section · Next section

Section 32 of 36

32. Connect to computing and AI literacy

Algorithms inherit feature choices and neighbourhood rules. A colour scale can organise complexity, but it cannot decide what caused the pattern.

The sample-associated density estimates where cells carrying each label occur; normalising across conditions yields a relative likelihood for each cell state.

Keep biological samples as the unit of replication. Thousands of cells improve resolution, but they do not turn one donor, animal or culture into thousands of independent experiments.

Contents · Previous section · Next section

Section 33 of 36

33. Connect to school and career pathways

Students who enjoy biological patterns and quantitative maps can explore biology, statistics, computing, bioinformatics, medicine or laboratory science through several routes.

A continuous score can show gradients and transitions that disappear when cells are first forced into discrete clusters, but resolution still depends on the graph.

Use negative controls, label permutations and reasonable alternative settings. A trustworthy signal should exceed a design-matched null and survive choices that were not selected for attractiveness.

Contents · Previous section · Next section

Section 34 of 36

34. Create a family evidence habit

When a headline says treatment changed cells, ask where the change appeared, how conditions were compared and whether independent samples repeated it.

Relative likelihood compares represented conditions under the observed sampling scheme. It is not an absolute population frequency and not a posterior probability of a biological mechanism.

Report effect direction, uncertainty, sample support and sensitivity—not only a colourful embedding or adjusted value. The display is a map of evidence, not the evidence itself.

Contents · Previous section · Next section

Section 35 of 36

35. Use precise MELD language

Say sample-associated density, relative likelihood and condition-enriched region. Avoid causal effect, response probability or clinical benefit unless separate evidence supports those phrases.

Graph construction, feature selection, batch handling and condition balance influence the result, so sensitivity analyses belong beside the main figure.

Create an evidence card naming the contrast, measured modality, statistical unit, result, validation status, unresolved alternatives and the narrowest defensible claim.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a graph-to-validation bundle

Deliver manifests, graph settings, balance checks, likelihood maps, replicate analyses, permutations, sensitivity results, evidence cards and the next discriminating experiment.

The original paper paired MELD with vector-field methods, but this guide owns the density-and-relative-likelihood question rather than claiming every companion analysis.

Write association, enrichment or differential abundance when that is what was tested. Reserve cause, mechanism, diagnosis and treatment benefit for experiments that directly support those claims.

## Source date and scope note Primary papers and official software documentation were checked for this article on 10 October 2026. Software interfaces and recommendations can change, so readers should consult the linked current documentation before reproducing an analysis.

## Final reader checklist Before accepting a claim, confirm the biological sample count, condition definition, measured modality, neighbourhood construction, statistical unit, uncertainty, software version, negative controls, sensitivity checks, independent validation and the exact boundary between association and causation.

Contents · Previous section · Continue to the Science Learning Hub