eduKateSG · Why Science?
Learn a perturbation direction in a latent space and ask whether it transfers to another cell type—without treating vector arithmetic as biological proof
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 Scnode Neural Odes Missing Timepoint Single Cell Prediction; Why Science Deepsem Structural Equation Models Single Cell Regulatory Networks; Why Science Celloracle Gene Regulatory Networks In Silico Perturbation Evidence; Why Science Genie3 Tree Ensembles Gene Regulatory Network Inference Evidence; 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: scGen primary study; scGen PubMed record; Official scGen 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.
scGen, reported by Lotfollahi, Wolf and Theis in Nature Methods in 2019, combines a variational autoencoder with latent-space vector arithmetic to predict how cells may respond to perturbation. The study explored transfer across cell types, studies and species. That makes scGen a lively secondary science bridge between experimental controls and AI representation learning, but its generated response remains a hypothesis that must be checked on genuinely held-out cells and, ultimately, in the laboratory.
Inside this guide
1–12 · Foundations and models
- 1. Begin with vector arithmetic
- 2. Understand the encoder
- 3. Estimate the perturbation vector
- 4. Transfer to a target cell type
- 5. Keep linearity conditional
- 6. Separate prediction from mechanism
- 7. Read the 2019 evidence in scope
- 8. Define the transfer question
- 9. Build a condition manifest
- 10. Balance biological units
- 11. Hold out the target response
- 12. Choose genes transparently
13–24 · Evidence, testing and applications
- 13. Preserve rare subtypes
- 14. Predeclare comparison metrics
- 15. Practise with a fictional scGen table
- 16. Separate mean and distribution
- 17. Plot the latent operation
- 18. Audit vector stability
- 19. Test reference choice
- 20. Audit dose and time mismatch
- 21. Use strong baselines
- 22. Protect against leakage
- 23. Test cross-study transfer honestly
- 24. Audit cross-species mapping
25–36 · Learning, decisions and pathways
- 25. Compare with CPA
- 26. Build a transfer evidence card
- 27. Treat differential genes carefully
- 28. Design the laboratory check
- 29. Connect to school science
- 30. Try a classroom translation activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family AI habit
- 35. Use precise scGen language
- 36. Finish with a vector-to-validation bundle
Section 1 of 36
1. Begin with vector arithmetic
scGen asks whether the difference between perturbed and control cells can be represented as a direction in a learned latent space, then added to a different target population.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 2 of 36
2. Understand the encoder
A variational autoencoder compresses gene expression into a probability distribution over latent coordinates. Compression may retain dominant programmes and lose rare signals.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 3 of 36
3. Estimate the perturbation vector
Average latent positions for perturbed and control reference cells are subtracted. That vector is treated as a reusable response direction under the model.
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 4 of 36
4. Transfer to a target cell type
The vector is added to latent control cells from a target group, and the decoder produces predicted perturbed expression.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 5 of 36
5. Keep linearity conditional
Vector arithmetic assumes a useful local regularity in the representation. Interactions, saturation and target-specific pathways can violate that assumption.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 6 of 36
6. Separate prediction from mechanism
A correct expression forecast does not prove which molecular route caused it. Different mechanisms can generate similar endpoint profiles.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Section 7 of 36
7. Read the 2019 evidence in scope
Lotfollahi, Wolf and Theis reported scGen and evaluated cross-cell-type, cross-study and cross-species prediction tasks. New datasets still need local benchmarks.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 8 of 36
8. Define the transfer question
Name reference cells, target cells, perturbation, duration and desired output. Avoid vague claims that one vector represents every context.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 9 of 36
9. Build a condition manifest
Record specimen, cell type, condition, dose, time, batch, assay and cell count. Check whether reference and target contrasts are comparable.
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 10 of 36
10. Balance biological units
A perturbation vector should not merely separate donors or processing days. Use paired or balanced designs where possible.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 11 of 36
11. Hold out the target response
Keep all truly perturbed target cells outside training and model selection. Reveal them only for the final transfer test.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 12 of 36
12. Choose genes transparently
Document normalisation, highly variable genes and identifier mapping. The official repository recommends attention to normalised input and gene selection.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Section 13 of 36
13. Preserve rare subtypes
Analyse whether target control cells contain the subpopulations expected after perturbation. A global vector can obscure heterogeneous response.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 14 of 36
14. Predeclare comparison metrics
Use distributional distances, correlation, differential-expression recovery and subgroup errors. No single score captures every failure.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 15 of 36
15. Practise with a fictional scGen table
This classroom table is invented and is not output from the scGen paper.
| Fictional target | Baseline error | scGen error | Marker recovery | Next check |
|---|---|---|---|---|
| Cell type A | 0.42 | 0.21 | Strong | New donor |
| Cell type B | 0.31 | 0.29 | Mixed | Audit vector |
| Rare subtype C | 0.55 | 0.61 | Weak | Collect cells |
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 16 of 36
16. Separate mean and distribution
A predicted mean can be accurate while variation and subtypes are wrong. Compare full populations as well as pseudobulk summaries.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 17 of 36
17. Plot the latent operation
Show control, translated and measured perturbed cells with provenance labels. A neat arrow is an explanation aid, not validation.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 18 of 36
18. Audit vector stability
Re-estimate the perturbation direction across donors, seeds and balanced subsamples. Report angle, magnitude and downstream prediction variation.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Section 19 of 36
19. Test reference choice
A vector learned from one cell type may not transfer equally from another. Compare credible references without choosing retrospectively.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 20 of 36
20. Audit dose and time mismatch
Responses measured at different exposures may encode incomparable biology. Match them or explicitly limit the claim.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 21 of 36
21. Use strong baselines
Compare no change, target-control mean plus reference fold changes and simple regression. Latent arithmetic should earn its complexity.
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 22 of 36
22. Protect against leakage
Gene selection, scaling and model tuning must use training data only. Target perturbed cells cannot influence preprocessing decisions.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 23 of 36
23. Test cross-study transfer honestly
Hold out entire studies or laboratories when claiming portability. Random cells from the same experiment are not a cross-study test.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 24 of 36
24. Audit cross-species mapping
Orthologue selection and duplicated genes influence transfer. Report mapping coverage and failures rather than only shared successes.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Section 25 of 36
25. Compare with CPA
CPA explicitly composes perturbations, covariates and doses, while scGen centres a latent difference vector. Their strengths and assumptions differ.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 26 of 36
26. Build a transfer evidence card
Record vector source, target support, held-out unit, baseline, global error, marker recovery, rare-state error and uncertainty.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 27 of 36
27. Treat differential genes carefully
Predicted differential expression is a prioritisation list. Control false discovery and confirm important genes in independent measurements.
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 28 of 36
28. Design the laboratory check
Collect perturbed target cells from new specimens, match dose and time, and include positive, negative and rescue controls where appropriate.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 29 of 36
29. Connect to school science
scGen links control groups, variables and fair tests with vectors and AI. It offers a concrete way to discuss how models generalise.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 30 of 36
30. Try a classroom translation activity
Represent objects with measured features, learn the change caused by one condition and test whether the same change transfers to a new group.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Section 31 of 36
31. Connect to mathematics
Vector subtraction, addition, averages, distributions and distance metrics make latent-space prediction understandable.
scGen encodes cells with a variational autoencoder, estimates a perturbation vector in latent space and transfers that vector to a target population.
Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.
Section 32 of 36
32. Connect to computing and AI literacy
An encoder, arithmetic step and decoder can fail separately. Reproducible preprocessing and untouched tests are essential.
Latent subtraction and addition are modelling operations; biological effects need not be perfectly linear, shared or context-independent.
Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.
Section 33 of 36
33. Connect to school and career choices
Students drawn to experiments and AI can explore biology, mathematics, computing, statistics, medicine or biomedical engineering.
The published work explored transfer across cell types, studies and species, while each new transfer requires held-out evidence.
Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.
Section 34 of 36
34. Create a family AI habit
When an AI transfers a pattern, ask where it learned the pattern, how different the target is and which real result was hidden for testing.
Two-condition designs make the perturbation contrast easy to state, but donor, batch and cell-state imbalance can masquerade as that contrast.
Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.
Section 35 of 36
35. Use precise scGen language
Say latent perturbation vector and predicted target response. Avoid claiming that arithmetic discovered the biological mechanism.
Generated expression profiles can support hypothesis formation; they do not replace measurements or demonstrate safety, efficacy or mechanism.
Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.
Section 36 of 36
36. Finish with a vector-to-validation bundle
Deliver manifests, held-out targets, gene mapping, baseline suite, vector stability, subgroup metrics, prediction cards and new experiments.
A useful result beats simple condition and cell-type baselines on unseen biological units and preserves both global and marker-level structure.
Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.
Did you know? The same latent vector can improve an average prediction yet fail for a rare subtype. Plot subgroup distributions and marker-level errors.
Cross-species transfer requires transparent orthologue mapping and should never erase species-specific biology. Keep mapping losses in the evidence card.
The most useful prediction is one that leads to a new target-cell experiment whose outcome can confirm, refine or reject the transferred vector.
Contents · Previous section · Continue to the Science Learning Hub