eduKateSG · Why Science?
Rank which cell types carry the clearest perturbation signal—without mistaking classifier separability for effect size or cause
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 Mixscape Perturbation Detection Crispr Response Signatures; Why Science Scpram Attention Transfer Unseen Cell Type Responses; Why Science Cinema Ot Confounder Separation Single Cell Treatment Effects; Why Science Gears Knowledge Graphs Multigene Perturbation Prediction; 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: Augur primary study; Augur PubMed record; Augur Nature Protocols workflow; Official Augur repository; 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.
Augur, reported by Skinnider and colleagues in Nature Biotechnology, is a machine-learning framework that prioritises cell types according to how readily perturbed and unperturbed cells can be distinguished within each type. Its central output is based on cross-validated classification performance, commonly summarised by area under the receiver-operating-characteristic curve. The framework has been applied to single-cell RNA sequencing, chromatin-accessibility and imaging-transcriptomics data. A high score means strong separability in the analysed features; it is not automatically the largest biological effect, a causal mechanism or a clinical priority.
Inside this guide
1–12 · Foundations and models
- 1. Begin with too many cell types
- 2. Define responsiveness operationally
- 3. Keep biological importance separate
- 4. Choose the contrast
- 5. Subsample cells repeatedly
- 6. Train and test a classifier
- 7. Read the evidence in scope
- 8. Define cell types before ranking
- 9. Protect sample independence
- 10. Balance the condition classes
- 11. Freeze the feature space
- 12. Choose a classifier transparently
13–24 · Evidence, testing and applications
- 13. Predeclare subsample size
- 14. Repeat across seeds
- 15. Practise with a fictional Augur table
- 16. Read AUC correctly
- 17. Inspect the score distribution
- 18. Check replicate consistency
- 19. Inspect informative features
- 20. Compare across modalities carefully
- 21. Run label permutations
- 22. Test annotation sensitivity
- 23. Use pseudobulk companions
- 24. Avoid circular follow-up
25–36 · Learning, decisions and pathways
- 25. Compare with Mixscape
- 26. Compare with Milo
- 27. Build a prioritisation evidence card
- 28. Design the next experiment
- 29. Connect Augur to school science
- 30. Try a blinded-ranking activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family ranking habit
- 35. Use precise Augur language
- 36. Finish with a contrast-to-follow-up bundle
Section 1 of 36
1. Begin with too many cell types
A complex tissue can contain many annotated cell types. Augur helps prioritise which types carry the most distinguishable perturbation signal for deeper follow-up.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 2 of 36
2. Define responsiveness operationally
In Augur, responsiveness means that a classifier can separate perturbed and unperturbed cells within a cell type using measured features.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 3 of 36
3. Keep biological importance separate
A high classification score does not automatically mean a cell type drives disease, benefits most from treatment or has the largest causal role.
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 4 of 36
4. Choose the contrast
Name the intervention, control, dose, time, organism, tissue and modality. A ranking belongs to that specific contrast.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 5 of 36
5. Subsample cells repeatedly
Augur repeatedly draws cells within each type so very abundant types do not win merely because they provide more training examples.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 6 of 36
6. Train and test a classifier
A supervised model learns the condition label from features and is evaluated on held-out observations under a declared resampling scheme.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Section 7 of 36
7. Read the evidence in scope
Skinnider and colleagues presented Augur as a way to prioritise cell types in diverse single-cell experiments, including expression, chromatin and imaging measurements.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 8 of 36
8. Define cell types before ranking
Annotations can come from markers, reference mapping or clustering. Record the method because merged or split labels can change the competition.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 9 of 36
9. Protect sample independence
Cells from the same donor or specimen share biology and technical effects. Grouped validation is stronger than random cell-level folds.
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 10 of 36
10. Balance the condition classes
Unequal perturbed and control counts can alter learning and precision. Document sampling and use metrics appropriate to the task.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 11 of 36
11. Freeze the feature space
Select variable genes, peaks or image features without consulting test labels. Leakage makes the ranking look more certain than it is.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 12 of 36
12. Choose a classifier transparently
The original framework supports supervised learning choices. Report hyperparameters, tuning and whether rankings persist with a reasonable alternative.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Section 13 of 36
13. Predeclare subsample size
Small subsamples increase uncertainty; large ones may exclude rare types or amplify dependence. Show the trade-off instead of tuning for an attractive ordering.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 14 of 36
14. Repeat across seeds
Resampling and model fitting are stochastic. A robust prioritisation should survive different random seeds and reasonable analysis settings.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 15 of 36
15. Practise with a fictional Augur table
This classroom table is invented and is not output from the Augur paper.
| Fictional cell type | Median AUC | AUC interval | Replicates agree | Next action |
|---|---|---|---|---|
| Type A | 0.84 | 0.79–0.88 | Yes | Validate markers |
| Type B | 0.68 | 0.54–0.78 | Mixed | Add samples |
| Type C | 0.51 | 0.47–0.56 | Yes | Low priority |
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 16 of 36
16. Read AUC correctly
AUC is the probability that a randomly chosen perturbed observation receives a higher score than a randomly chosen control under the model’s ranking.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 17 of 36
17. Inspect the score distribution
A single median hides unstable subsamples. Plot the full distribution and report uncertainty for each cell type.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 18 of 36
18. Check replicate consistency
Compute or visualise responsiveness separately by donor, culture or experiment. One high-signal replicate should not define a general claim.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Section 19 of 36
19. Inspect informative features
Feature importance can suggest follow-up measurements, but correlated features and model choice limit mechanistic interpretation.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 20 of 36
20. Compare across modalities carefully
RNA, chromatin accessibility and imaging features have different noise and dimensionality. Similar AUC values do not mean identical biological effects.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 21 of 36
21. Run label permutations
Shuffle condition labels within valid experimental blocks. The null distribution reveals whether apparent separation exceeds chance under the design.
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 22 of 36
22. Test annotation sensitivity
Merge or refine uncertain cell types and repeat the analysis. A ranking that depends on one fragile taxonomy deserves a narrow conclusion.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 23 of 36
23. Use pseudobulk companions
Sample-level summaries and differential tests can show effect direction and replication, complementing cell-level separability.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 24 of 36
24. Avoid circular follow-up
If the same features both chose the top cell type and confirm its pathway, use new data or a predeclared validation assay for confirmation.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Section 25 of 36
25. Compare with Mixscape
Mixscape classifies whether guide-positive cells show a perturbation signature; Augur compares responsiveness across annotated cell types in a perturbation contrast.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 26 of 36
26. Compare with Milo
Milo locates differential abundance over graph neighbourhoods. Augur ranks condition separability within existing cell-type labels.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 27 of 36
27. Build a prioritisation evidence card
Place cell definition, sample split, AUC distribution, null result, replicate agreement, informative features and follow-up plan together.
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 28 of 36
28. Design the next experiment
Collect independent samples and measure markers chosen before data collection in both highly ranked and lower-ranked cell types.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 29 of 36
29. Connect Augur to school science
Classification, controlled comparisons and repeatability show why scientific ranking needs a clearly defined measurement rule.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 30 of 36
30. Try a blinded-ranking activity
Give teams noisy feature cards from two conditions, let them train a simple rule on one set and score an untouched set before comparing rankings.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Section 31 of 36
31. Connect to mathematics
ROC curves, AUC, resampling, distributions and uncertainty intervals make responsiveness a quantitative rather than rhetorical idea.
Augur asks how well a classifier can distinguish perturbed from unperturbed cells within each annotated cell type after repeated subsampling.
Archive labels, cell-type annotations, specimens, features, subsampling size, classifier, folds, seeds, version and every exclusion.
Section 32 of 36
32. Connect to computing and AI literacy
High predictive accuracy answers a discrimination question. It does not automatically answer why the groups differ or what should be done.
The area under the ROC curve summarises ranking discrimination across thresholds; chance-level separability is near 0.5 under a balanced binary task.
Split biological samples before feature selection and tuning so cells from one donor or experiment cannot appear on both sides of evaluation.
Section 33 of 36
33. Connect to school and career choices
Learners drawn to ranking and patterns can explore biology, statistics, computing, neuroscience, imaging or data science through many education routes.
AUC depends on features, sample balance, replicate structure and preprocessing, so a responsiveness ranking is conditional on the analysis design.
Compare label permutations, simple distance measures, pseudobulk differential tests and alternative classifiers on the same evidence.
Section 34 of 36
34. Create a family ranking habit
When something is called the most responsive, ask how responsiveness was defined, what was held out and whether independent samples agreed.
A cell type can show subtle coordinated changes that classify well without having the largest mean fold change in any single feature.
Report AUC distributions across subsamples, per-replicate results, class balance, calibration, ranking stability and feature sensitivity.
Section 35 of 36
35. Use precise Augur language
Say cross-validated perturbation separability and prioritisation ranking. Avoid strongest effect, causal cell type or treatment target unless other evidence supports it.
The framework prioritises where to look next across modalities; it does not itself identify the causal pathway or prove biological importance.
Build a responsiveness card with contrast, modality, unit, classifier, AUC uncertainty, replication and follow-up status.
Section 36 of 36
36. Finish with a contrast-to-follow-up bundle
Deliver manifests, grouped splits, subsampling distributions, nulls, annotation sensitivity, sample-level companions, evidence cards and independent validation.
Repeated subsampling reduces domination by abundant cell types, but biological replication—not repeated reuse of the same cells—supports generalisation.
Say classifier separability or Augur prioritisation score; avoid largest effect, most important cell or causal responder without separate evidence.
Did you know? Two cell types can have similar average fold changes yet very different classifier separability if one has a coordinated multifeature response. That is why Augur’s score answers a pattern-recognition question.
A useful final figure pairs the AUC distribution with per-sample effect plots and a label-permutation null. Readers can then see priority, replication and chance performance together.
If cell-type ranking changes after a reasonable annotation merge or donor holdout, report the instability instead of choosing the most dramatic list.
Contents · Previous section · Continue to the Science Learning Hub