eduKateSG · Why Science?
Align unpaired control and perturbed populations, then use attention to transfer a response into an unseen cell type—while testing every step on sealed biological evidence
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 Cellot Neural Optimal Transport Unpaired Cell Responses; Why Science Scvidr Cell Type Perturbation Vectors Dose Response Prediction; Why Science Scgen Latent Space Arithmetic Cross Condition Perturbation Prediction; Why Science Trvae Distribution Matching Conditional Out Of Sample Generation; 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: scPRAM primary study; scPRAM PubMed record; Official scPRAM 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.
scPRAM, published by Jiang and colleagues in Bioinformatics in 2024, predicts single-cell gene-expression responses for a cell type whose perturbed examples are held out. Its workflow combines variational autoencoding and optimal-transport alignment with an attention mechanism intended to transfer perturbation information from observed cell types. The study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals. That is an exciting AI-learning story, but attention weights and aligned latent states remain model constructs whose biological meaning requires independent evidence.
Inside this guide
1–12 · Foundations and models
- 1. Begin with the missing response
- 2. Encode cell state
- 3. Align unpaired populations
- 4. Learn source responses
- 5. Apply attention
- 6. Decode the target
- 7. Read the 2024 evidence in scope
- 8. Define the target holdout
- 9. Build a transfer manifest
- 10. Protect biological independence
- 11. Inspect source coverage
- 12. Freeze preprocessing
13–24 · Evidence, testing and applications
- 13. Audit transport settings
- 14. Audit attention inputs
- 15. Practise with a fictional scPRAM table
- 16. Separate alignment and prediction
- 17. Evaluate full profiles
- 18. Audit attention stability
- 19. Test target similarity
- 20. Run seed ensembles
- 21. Use strong baselines
- 22. Construct source nulls
- 23. Test cross-species transfer
- 24. Test cross-individual transfer
25–36 · Learning, decisions and pathways
- 25. Compare with scVIDR
- 26. Build an attention-transfer card
- 27. Keep mechanism claims separate
- 28. Design the next experiment
- 29. Connect scPRAM to school science
- 30. Try an attention activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family transfer habit
- 35. Use precise scPRAM language
- 36. Finish with a source-to-target bundle
Section 1 of 36
1. Begin with the missing response
A target cell type may have control measurements but no perturbed examples, while other cell types have both conditions. scPRAM predicts the missing target response.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 2 of 36
2. Encode cell state
A variational autoencoder compresses high-dimensional gene expression into a latent representation that supports comparison and generation.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 3 of 36
3. Align unpaired populations
Optimal transport aligns control and perturbed states among observed cell types. The inferred coupling summarises change without tracking individual cells.
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 4 of 36
4. Learn source responses
For each observed type, the workflow extracts a perturbation-related transformation that can inform another cell context.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 5 of 36
5. Apply attention
Attention assigns context-dependent weights to source information when constructing a target-type response.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 6 of 36
6. Decode the target
The model maps the attended latent prediction back into gene-expression space for evaluation against sealed target measurements.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Section 7 of 36
7. Read the 2024 evidence in scope
Jiang and colleagues reported scPRAM in Bioinformatics and evaluated drug and bacterial perturbation datasets across several transfer settings.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 8 of 36
8. Define the target holdout
State the exact cell type, organism or individual whose perturbed cells are excluded. Do not blur within-study and cross-study prediction.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 9 of 36
9. Build a transfer manifest
Record source and target types, specimen, species, perturbation, dose, time, batch, assay and cell counts.
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 10 of 36
10. Protect biological independence
Hold out donors, cultures or organisms, not only random cells. Shared technical structure can make transfer look easier.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 11 of 36
11. Inspect source coverage
Plot how the target controls relate to each source control. Attention cannot repair the absence of biologically relevant neighbours.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 12 of 36
12. Freeze preprocessing
Choose genes, normalisation and integration without seeing target perturbed data. Target-aware preprocessing can leak the answer.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Section 13 of 36
13. Audit transport settings
Document the cost, regularisation and latent dimension used for alignment. Different geometries can create different source transformations.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 14 of 36
14. Audit attention inputs
State which source summaries, queries and keys determine the weights. A black-box weight is not a biological explanation.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 15 of 36
15. Practise with a fictional scPRAM table
This classroom table is invented and is not output from the scPRAM paper.
| Fictional target | Closest source | Attention stability | scPRAM error | Next check |
|---|---|---|---|---|
| Cell A | Cell B | 0.92 | 0.20 | New donor |
| Cell C | Cell D | 0.71 | 0.31 | Marker panel |
| Cell E | None close | 0.39 | 0.67 | Measure first |
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 16 of 36
16. Separate alignment and prediction
Good distribution alignment does not guarantee correct perturbation transfer. Score both steps on untouched targets.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 17 of 36
17. Evaluate full profiles
Compare gene-wise errors, differential genes, means, variances and subpopulations. A high overall correlation can hide missed response genes.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 18 of 36
18. Audit attention stability
Repeat training and perturb source inputs. If weights change while predictions stay similar, do not over-interpret the weights.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Section 19 of 36
19. Test target similarity
Stratify targets by distance to training controls. Performance should be reported with the support available for each target.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 20 of 36
20. Run seed ensembles
Retrain the complete VAE, transport and attention pipeline. Report prediction and source-weight variation.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 21 of 36
21. Use strong baselines
Compare shared mean response, nearest source, scGen arithmetic and CellOT where appropriate. Attention must improve sealed biological evidence.
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 22 of 36
22. Construct source nulls
Shuffle source labels or pair a target with irrelevant sources while preserving data size. The model should lose transfer accuracy.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 23 of 36
23. Test cross-species transfer
Orthologue mapping, cell definitions and immune contexts differ across species. Report mapping coverage and failures explicitly.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 24 of 36
24. Test cross-individual transfer
Leave complete people out and report participant-level results. Many cells from one person do not replace population diversity.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Section 25 of 36
25. Compare with scVIDR
scVIDR regresses cell-type perturbation vectors and dose relationships; scPRAM transfers aligned source responses with attention.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 26 of 36
26. Build an attention-transfer card
Record sources, target support, transport settings, attention stability, baseline gain, uncertainty and experimental confirmation.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 27 of 36
27. Keep mechanism claims separate
Attention indicates model weighting under the learned representation; it does not prove which cell type or pathway caused the response.
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 28 of 36
28. Design the next experiment
Measure unsupported target types and cases where source weights or competing models disagree most.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 29 of 36
29. Connect scPRAM to school science
Fair tests, comparison groups and transfer across categories become tangible through cell types and perturbations.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 30 of 36
30. Try an attention activity
Predict a hidden object’s change using several reference groups, assign weights before reveal and test whether the weights remain stable.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Section 31 of 36
31. Connect to mathematics
Latent vectors, transport, weighted sums, conditional probability and error measures organise the transfer task.
scPRAM combines latent representation, optimal-transport alignment and attention to transfer a measured perturbation response into an unseen cell type.
Archive target and source cell types, perturbation, specimens, preprocessing, latent settings, transport choices, attention model, seeds and version.
Section 32 of 36
32. Connect to computing and AI literacy
Attention can help a model select information, but its weight is not automatically an explanation of nature.
The target cell type’s perturbed examples are held out while its control state provides context for prediction.
Hold out the complete target-type perturbed group and split donors, species or experiments before feature selection and tuning.
Section 33 of 36
33. Connect to school and career choices
Learners interested in transfer and evidence can explore biology, computing, statistics, medicine or bioinformatics.
Attention weights aggregate information from observed cell types, but weights are not automatically explanations of biological regulation.
Compare no change, shared mean shift, scGen-style arithmetic, CellOT and simple source averaging on identical evidence.
Section 34 of 36
34. Create a family transfer habit
When AI predicts an unseen group, ask which observed groups carried the evidence and how similar the target really was.
Optimal-transport alignment is an inferred relationship between unpaired populations, not a tracked cell lineage.
Report gene-level error, differential-gene recovery, distribution distances, target support, attention stability and seed spread.
Section 35 of 36
35. Use precise scPRAM language
Say attention-based transfer prediction and held-out cell type. Avoid tracked response, causal source or personalised result.
The 2024 study evaluated drug-treatment and bacterial-infection datasets across cell types, species and individuals.
Build a transfer card with source types, target similarity, holdout unit, alignment diagnostics, baselines, uncertainty and validation status.
Section 36 of 36
36. Finish with a source-to-target bundle
Deliver manifests, biological holdouts, transport audits, attention stability tests, baselines, cards and new target measurements.
Transfer depends on source support; a target with no comparable observed state deserves wider uncertainty or no prediction.
Say attention-based predicted response for an unseen cell type; do not call attention weights a mechanism or the output an observed response.
Did you know? Attention can be highly useful for prediction while remaining unstable or biologically ambiguous as an explanation. Test weights before telling a mechanism story.
Keep target perturbed cells sealed until preprocessing, transport settings, attention architecture and baselines are fixed.
A strong final plot shows target-control support, source weights across seeds, predicted and measured profiles, differential genes and simple-baseline performance.
When transferring across species, distinguish mapped orthologues from genes that have no one-to-one correspondence, and report performance on the supported gene set separately.
When transferring across people, score each held-out individual before pooling. A high cell count from a small number of individuals is not broad validation.
If no source cell type resembles the target control population, declining to predict may be more scientific than returning a confident-looking profile.
Contents · Previous section · Continue to the Science Learning Hub