eduKateSG · Why Science?
Learn a transport map between unpaired control and perturbed cell populations—without pretending a mathematical pairing is an observed before-and-after cell
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 Prescient Potential Landscapes Physical Time Cell Trajectories; Why Science Scgen Latent Space Arithmetic Cross Condition Perturbation Prediction; Why Science Cpa Compositional Perturbations Dose Response Prediction; Why Science Scnode Neural Odes Missing Timepoint Single Cell 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: CellOT primary study; CellOT PubMed record; Official CellOT 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.
CellOT, reported by Bunne and colleagues in Nature Methods in 2023, learns maps between unpaired control and perturbed single-cell distributions using neural optimal transport and input-convex architectures. The study evaluated drug responses measured with single-cell RNA sequencing and multiplexed protein imaging, then examined generalisation to held-out patients, cross-species responses and developmental trajectories. For science learning, CellOT makes distributions, fair comparisons and model assumptions wonderfully concrete—while the inferred pairing between cells remains a mathematical construction, not a directly observed lineage.
Inside this guide
1–12 · Foundations and models
- 1. Begin with unpaired cells
- 2. Understand optimal transport
- 3. See the transport map
- 4. Use a convex potential
- 5. Keep minimal effort conditional
- 6. Predict heterogeneous responses
- 7. Read the 2023 evidence in scope
- 8. Define the transport question
- 9. Build a specimen manifest
- 10. Check source–target support
- 11. Split by biological unit
- 12. Choose features transparently
13–24 · Evidence, testing and applications
- 13. Predeclare the cost
- 14. Protect rare populations
- 15. Practise with a fictional CellOT table
- 16. Separate map and observation
- 17. Evaluate full distributions
- 18. Audit map smoothness
- 19. Run seed ensembles
- 20. Test feature sensitivity
- 21. Use strong simple baselines
- 22. Audit proliferation and death
- 23. Construct matched nulls
- 24. Test held-out patients
25–36 · Learning, decisions and pathways
- 25. Test cross-species transfer
- 26. Build a transport evidence card
- 27. Treat trajectories carefully
- 28. Design the next experiment
- 29. Connect CellOT to school science
- 30. Try a classroom transport activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family evidence habit
- 35. Use precise transport language
- 36. Finish with a map-to-measurement bundle
Section 1 of 36
1. Begin with unpaired cells
A sequencing experiment destroys each measured cell, so control and treated populations cannot be matched cell by cell. CellOT learns a plausible map between those two clouds.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 2 of 36
2. Understand optimal transport
Optimal transport asks how to move probability mass from one distribution to another with minimum declared cost. The answer depends on features, geometry and cost.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 3 of 36
3. See the transport map
For each control state, the model generates a predicted treated state. Applying the map to a population should reproduce important properties of the measured target distribution.
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 4 of 36
4. Use a convex potential
CellOT uses input-convex neural networks to learn a potential whose gradient defines the map. Architectural constraints help optimisation without making the map biologically unique.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 5 of 36
5. Keep minimal effort conditional
The selected coupling reflects a mathematical minimal-effort principle in the chosen representation. Biology may follow routes the cost does not encode.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 6 of 36
6. Predict heterogeneous responses
Different control cells can move differently under the same perturbation. This is richer than adding one population-average vector to every cell.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Section 7 of 36
7. Read the 2023 evidence in scope
Bunne and colleagues evaluated scRNA-seq and multiplexed protein-imaging responses, plus held-out-patient, cross-species and developmental settings.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 8 of 36
8. Define the transport question
Name source population, perturbation, duration, target population and feature space. Decide whether the claim concerns distributions, subgroups or individual cells.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 9 of 36
9. Build a specimen manifest
Record patient or culture, condition, dose, time, batch, assay, cell count and quality filters. Many cells do not replace independent specimens.
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 10 of 36
10. Check source–target support
Plot both distributions before fitting. Regions present only after treatment may represent induction, selection, proliferation or sampling differences.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 11 of 36
11. Split by biological unit
Reserve complete specimens or experiments. Random cell splits leak nearly identical biological and technical structure.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 12 of 36
12. Choose features transparently
Document genes, proteins, normalisation, variable-feature selection and any latent representation. Geometry changes when features or scaling change.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Section 13 of 36
13. Predeclare the cost
State the distance or learned geometry used to move mass. Test whether conclusions survive other scientifically plausible representations.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 14 of 36
14. Protect rare populations
Rare cells can be cheap to ignore in an average distance. Evaluate subgroup mass, markers and response separately.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 15 of 36
15. Practise with a fictional CellOT table
This classroom table is invented and is not output from the CellOT paper.
| Fictional subgroup | Transport error | Baseline error | Source support | Next check |
|---|---|---|---|---|
| State A | 0.18 | 0.31 | High | New donor |
| State B | 0.27 | 0.29 | Medium | Marker panel |
| Rare C | 0.62 | 0.55 | Low | Enrich cells |
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 16 of 36
16. Separate map and observation
A transported cell is a counterfactual prediction. Only the endpoint population was measured; the individual correspondence was inferred.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 17 of 36
17. Evaluate full distributions
Compare means, variances, correlations, quantiles and multimodality. A close average can conceal a missing responder group.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 18 of 36
18. Audit map smoothness
Nearby control cells should not jump unpredictably without evidence. Inspect local continuity while allowing real nonlinear transitions.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Section 19 of 36
19. Run seed ensembles
Repeat optimisation and plot prediction variability. Similar target distances can arise from different individual maps.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 20 of 36
20. Test feature sensitivity
Refit using defensible gene sets, protein panels or latent dimensions. A biological claim should not rest on one convenient geometry.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 21 of 36
21. Use strong simple baselines
Compare no change, mean-shift and linear transport. Complexity should improve untouched biological evidence, not only training fit.
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 22 of 36
22. Audit proliferation and death
Optimal transport redistributes mass, while perturbations may change abundance through division or apoptosis. Check whether assumptions fit the experiment.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 23 of 36
23. Construct matched nulls
Permute condition within specimens or simulate unrelated distributions with similar spread. The method should not find a meaningful map in every contrast.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 24 of 36
24. Test held-out patients
When claiming patient transfer, train and tune without that person’s treated cells. Report variation rather than only pooled accuracy.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Section 25 of 36
25. Test cross-species transfer
Orthologue mapping and species differences are evidence layers. Report mapping coverage and where transport fails.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 26 of 36
26. Build a transport evidence card
Place source support, cost, seed stability, baseline gains, subgroup performance and independent measurements together.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 27 of 36
27. Treat trajectories carefully
A transport map between endpoints does not reveal every intermediate state. Time-series measurements are needed to test the route.
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 28 of 36
28. Design the next experiment
Sample where maps disagree, rare states diverge or target-only populations appear. The goal is to separate transport, selection and induction explanations.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 29 of 36
29. Connect CellOT to school science
The method turns fair comparisons, variables, distributions and model assumptions into a practical secondary science story.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 30 of 36
30. Try a classroom transport activity
Move tokens between two histograms at different costs, compare several valid couplings and ask what extra observation would identify the real pairing.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Section 31 of 36
31. Connect to mathematics
Probability measures, convexity, gradients, distances and optimisation make the unpaired-response problem precise.
CellOT treats control and perturbed measurements as unpaired distributions because destructive assays do not observe the same cell twice.
Archive control and treated specimens, preprocessing, feature space, cost, network architecture, optimiser, seeds, checkpoints and repository version.
Section 32 of 36
32. Connect to computing and AI literacy
A structured neural network can encode a mathematical principle, yet data splits and baselines still decide whether it generalises.
Optimal transport selects a coupling under a cost principle; that mathematical coupling is not a recovered biological lineage.
Hold out entire patients, cultures or experiments and keep their treated cells sealed until transport choices are frozen.
Section 33 of 36
33. Connect to school and career choices
Learners who enjoy distributions can explore biology, mathematics, computing, statistics, medicine or biomedical engineering.
Input-convex neural architectures parameterise a transport potential while imposing useful structure on the learned map.
Compare no change, population shift, nearest-neighbour and a simple linear map on the same biological splits.
Section 34 of 36
34. Create a family evidence habit
When a model pairs unpaired observations, ask what cost chose the pairing and whether any real individual was followed.
The fitted map predicts heterogeneous responses for new control cells, but reliability depends on overlap, representation and perturbation scope.
Report global distances, marker distributions, subgroup errors, map smoothness and seed spread rather than one favourable score.
Section 35 of 36
35. Use precise transport language
Say inferred coupling, transported prediction and held-out distribution. Reserve tracked response for longitudinal measurement.
The 2023 study evaluated RNA and multiplexed protein-imaging data plus held-out-patient and cross-species settings; new systems need local tests.
Build a transport card containing source support, target scope, cost, holdout unit, baselines, uncertainty and experimental validation.
Section 36 of 36
36. Finish with a map-to-measurement bundle
Deliver manifests, representations, costs, biological holdouts, baselines, seed ensembles, subgroup audits, cards and a discriminating experiment.
A model can align population distributions while mispairing individuals or missing rare states, so distribution and subgroup checks belong together.
Say inferred transport map and predicted response; never describe transported cells as directly tracked or clinically validated.
Did you know? Many different individual couplings can reproduce the same final population. That is why CellOT’s map is a principled hypothesis, not a recovered movie.
For human data, preserve patient-level validation and consent-aware governance. Share code and aggregate diagnostics without assuming raw profiles may be released.
A strong final figure labels measured control, transported prediction and measured target separately, then shows rare-state and patient-level uncertainty.
Contents · Previous section · Continue to the Science Learning Hub