eduKateSG · Why Science?
Find local population shifts along a continuous cell-state landscape—without forcing every biological transition into a hard cluster
Reading routes
Science learning becomes useful when a familiar object or observation is turned into a system of quantities, mechanisms and claim limits. This guide owns one applied evidence-reading job inside eduKateSG’s wider Science estate. It connects naturally to Why Science Sccoda Compositional Models Credible Cell Type Abundance Changes; Why Science Augur Cell Type Responsiveness Perturbation Priority Evidence; Why Science Cinema Ot Confounder Separation Single Cell Treatment Effects; Why Science Scpram Attention Transfer Unseen Cell Type Responses; Education Hub; Singapore Secondary School Directory; Career Adulthood Hub; Science Learning Hub. It also keeps current school and public claims traceable to visible primary sources: Milo primary study; Milo PubMed record; Official MiloR repository; Official MiloR demonstration; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science syllabus; 2026 MOE G2 Computing syllabus. The sources describe the scientific scope; this article translates that scope into a calm route for Primary Science, PSLE Science, Secondary Science, O-Level Science, STEM exploration, school choices and career pathways without inventing admission or employment outcomes.
Milo, developed by Dann and colleagues and published in Nature Biotechnology in 2022, tests differential abundance over overlapping neighbourhoods on a k-nearest-neighbour graph. Instead of requiring one fixed set of discrete cell clusters, it can locate condition-associated population changes along continuous state landscapes. The method counts cells from each biological sample in each neighbourhood and uses a negative-binomial framework with graph-aware false-discovery control. Its discoveries remain local, design-dependent statistical evidence: they do not by themselves explain why abundance changed.
Inside this guide
1–12 · Foundations and models
- 1. Begin with a moving landscape
- 2. Represent cells as neighbours
- 3. Sample overlapping neighbourhoods
- 4. Count by biological sample
- 5. Model the experimental design
- 6. Control local discoveries
- 7. Read the 2022 evidence in scope
- 8. Define abundance carefully
- 9. Build a sample manifest
- 10. Require replication
- 11. Choose integration cautiously
- 12. Choose graph features
13–24 · Evidence, testing and applications
- 13. Set neighbourhood size
- 14. Specify the design formula
- 15. Practise with a fictional Milo table
- 16. Read a local log-fold change
- 17. Map results back to cells
- 18. Check sample support
- 19. Interpret continuous transitions
- 20. Separate abundance and state
- 21. Run label permutations
- 22. Audit graph sensitivity
- 23. Inspect zero and sparse counts
- 24. Validate with an orthogonal measurement
25–36 · Learning, decisions and pathways
- 25. Compare with scCODA
- 26. Compare with Augur
- 27. Build a neighbourhood evidence card
- 28. Design the next experiment
- 29. Connect Milo to school science
- 30. Try a neighbourhood-map activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family population habit
- 35. Use precise Milo language
- 36. Finish with a graph-to-validation bundle
Section 1 of 36
1. Begin with a moving landscape
Cells often vary continuously through activation, development or disease. Hard clusters can cut one transition into arbitrary boxes and dilute a local change.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 2 of 36
2. Represent cells as neighbours
A k-nearest-neighbour graph connects cells with similar measured profiles. The graph is an analysis model of local similarity, not a physical tissue map.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 3 of 36
3. Sample overlapping neighbourhoods
Milo chooses representative index cells and gathers nearby cells around each, producing overlapping local regions across the graph.
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 4 of 36
4. Count by biological sample
For every neighbourhood, count how many cells came from each donor, animal, culture or other independent sample.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 5 of 36
5. Model the experimental design
A negative-binomial generalised linear model links neighbourhood counts to condition and covariates while allowing overdispersion.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 6 of 36
6. Control local discoveries
Graph-aware false-discovery weighting uses neighbourhood relationships to address the many dependent tests created across the landscape.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Section 7 of 36
7. Read the 2022 evidence in scope
Dann and colleagues introduced Milo for differential-abundance testing in single-cell data and demonstrated sensitivity to populations missed by discrete clustering.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 8 of 36
8. Define abundance carefully
Observed cell counts reflect biology plus sampling, dissociation, viability, capture and filtering. A relative shift is not automatically in-vivo expansion or loss.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 9 of 36
9. Build a sample manifest
Record condition, donor, batch, tissue, processing time, total cells, quality metrics and relevant pairing or repeated measures.
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 10 of 36
10. Require replication
A comparison with one sample per condition cannot separate condition from sample identity, regardless of how many cells were captured.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 11 of 36
11. Choose integration cautiously
Batch correction can improve shared geometry but may erase condition structure or align states that should remain distinct. Preserve an unintegrated audit view.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 12 of 36
12. Choose graph features
Document genes, dimensions, distance, number of neighbours and scaling. Each choice changes which cells count as local.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Section 13 of 36
13. Set neighbourhood size
Too-small regions are noisy; too-large regions blur local effects. Explore a scientifically reasonable range and report stability.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 14 of 36
14. Specify the design formula
Include condition, batch, donor pairing and other predeclared covariates where identifiable. A formula cannot rescue a confounded experiment.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 15 of 36
15. Practise with a fictional Milo table
This classroom table is invented and is not output from the Milo paper.
| Fictional neighbourhood | Log fold change | Adjusted evidence | Samples represented | Next action |
|---|---|---|---|---|
| N12 | 1.1 | Strong | 8 | Marker panel |
| N28 | -0.7 | Moderate | 7 | Replicate |
| N41 | 0.2 | Weak | 8 | No claim |
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 16 of 36
16. Read a local log-fold change
The sign and magnitude describe a condition association for neighbourhood counts under the fitted model and chosen reference level.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 17 of 36
17. Map results back to cells
Colour neighbourhood effects on the graph and inspect marker patterns. Overlap means nearby significant regions may describe one broader transition.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 18 of 36
18. Check sample support
A neighbourhood dominated by one donor can look dramatic but lacks general support. Display counts for every biological sample.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Section 19 of 36
19. Interpret continuous transitions
A run of related neighbourhoods can reveal where along a trajectory abundance diverges, without declaring an artificial boundary.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 20 of 36
20. Separate abundance and state
Milo tests how many sampled cells occupy a region. Differential expression asks how molecular state changes within comparable populations.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 21 of 36
21. Run label permutations
Permute condition at the sample level within valid blocks and rerun the full analysis to test whether the design produces excess discoveries.
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 22 of 36
22. Audit graph sensitivity
Vary neighbours, dimensions and sampling density. Record which regions persist and which move or disappear.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 23 of 36
23. Inspect zero and sparse counts
Rare neighbourhoods can have unstable estimates. Report filtering, effective sample support and uncertainty rather than celebrating extreme fold changes.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 24 of 36
24. Validate with an orthogonal measurement
Flow cytometry, histology, targeted panels or another cohort can test whether the local abundance pattern repeats outside the discovery pipeline.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Section 25 of 36
25. Compare with scCODA
scCODA models a whole vector of cell-type counts compositionally; Milo tests overlapping graph neighbourhoods at finer local resolution.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 26 of 36
26. Compare with Augur
Augur ranks condition separability within cell types. Milo asks whether the relative occupancy of local cell-state regions differs across samples.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 27 of 36
27. Build a neighbourhood evidence card
Place graph definition, local markers, per-sample counts, effect, multiple-testing result, sensitivity and validation status together.
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 28 of 36
28. Design the next experiment
Increase biological samples, standardise tissue handling and preselect orthogonal markers for regions that persist across graph settings.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 29 of 36
29. Connect Milo to school science
Population sampling, variation, graphs and fair comparisons connect ecology-style counting ideas to modern cell biology.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 30 of 36
30. Try a neighbourhood-map activity
Place coloured observations on a coordinate map, count overlapping circles by group and discuss how circle size changes discovery resolution.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Section 31 of 36
31. Connect to mathematics
Graphs, nearest neighbours, count distributions, regression, logarithms and multiple testing organise local evidence.
Milo represents cells on a k-nearest-neighbour graph and samples overlapping neighbourhoods so local abundance changes can be tested without fixed clusters.
Archive specimens, condition covariates, cell counts, integration, features, neighbours, dimensions, neighbourhood sampling, design formula and MiloR version.
Section 32 of 36
32. Connect to computing and AI literacy
A graph is a designed representation. Its neighbourhoods reveal patterns only within the features, distance and resolution supplied.
The response for statistical testing is a neighbourhood-by-sample count matrix, keeping biological samples—not individual cells—as replicates.
Use several independent biological samples per group and retain donor, batch and pairing structure in the design matrix.
Section 33 of 36
33. Connect to school and career choices
Students interested in populations and maps can explore biology, ecology, statistics, computing, medicine or bioinformatics through varied routes.
A negative-binomial generalised linear model relates counts to the experimental design while accounting for library-size-like sampling differences.
Compare cluster-level proportions, sample-level compositional tools, label permutations and graph-resolution alternatives on the same study.
Section 34 of 36
34. Create a family population habit
When a subgroup is said to increase, ask relative to what total, across how many independent samples and under which measurement process.
Overlapping neighbourhoods create spatial dependence; Milo uses graph-aware weighting for multiple-testing control rather than treating every test as unrelated.
Report log-fold changes, adjusted values, sample counts, neighbourhood overlap, graph locations and sensitivity to parameters.
Section 35 of 36
35. Use precise Milo language
Say condition-associated differentially abundant neighbourhood. Avoid expanded lineage, depleted tissue or causal transition without additional evidence.
A significant neighbourhood marks a local shift in sampled population abundance, not a newly discovered cell type or a causal mechanism.
Build a neighbourhood card with cell-state markers, sample support, effect, uncertainty, design covariates and validation plan.
Section 36 of 36
36. Finish with a graph-to-validation bundle
Deliver sample manifests, graph specifications, count matrices, design formulas, sensitivity maps, evidence cards and orthogonal replication.
Graph construction, neighbourhood size and sampling density set the resolution, so sensitivity analysis is part of the scientific result.
Say differentially abundant neighbourhood or local population shift; do not say new cell type, lineage conversion or mechanism from abundance alone.
Did you know? Overlapping neighbourhoods let Milo follow a smooth biological transition without pretending that every point belongs to one crisp cluster.
A strong result map shows neighbourhood effects, local markers and per-sample counts together. Colour alone can hide whether a region is supported by many specimens or one unusually large sample.
When tissue processing may change recovery, add an orthogonal counting method before interpreting a sampled proportion as biological expansion or depletion.
Contents · Previous section · Continue to the Science Learning Hub