eduKateSG · Why Science?
Condition a generative cell-state model on molecular structure or gene function to forecast unseen perturbations—while keeping novelty and uncertainty visible
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 Gears Knowledge Graphs Multigene Perturbation Prediction; Why Science Cpa Compositional Perturbations Dose Response Prediction; Why Science Scgen Latent Space Arithmetic Cross Condition Perturbation Prediction; Why Science Celloracle Gene Regulatory Networks In Silico Perturbation 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: PerturbNet primary study; PerturbNet PubMed record; Official PerturbNet 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.
PerturbNet, published by Yu and colleagues in Molecular Systems Biology in 2025, is a deep generative model for predicting distributions of single-cell states after unseen chemical or genetic perturbations. It uses information such as chemical structure for small molecules and gene functional annotations for genetic interventions, while accounting for covariates including dose and cell type. For science learning, it joins chemistry, genetics and AI in one optimistic question: can prior knowledge help us choose a better experiment? The answer is a testable forecast, never automatic proof of mechanism, safety or clinical benefit.
Inside this guide
1–12 · Foundations and models
- 1. Begin with unseen perturbations
- 2. Represent chemicals
- 3. Represent genetic interventions
- 4. Condition a generative model
- 5. Use invertible components
- 6. Decode cell states
- 7. Read the 2025 evidence in scope
- 8. Define unseen precisely
- 9. Build a perturbation manifest
- 10. Inspect descriptor coverage
- 11. Choose a strict chemical split
- 12. Choose a strict genetic split
13–24 · Evidence, testing and applications
- 13. Protect biological experiments
- 14. Predeclare covariates
- 15. Practise with a fictional PerturbNet table
- 16. Separate novelty from difficulty
- 17. Evaluate full distributions
- 18. Audit chemical representations
- 19. Audit gene annotations
- 20. Audit dose and cell context
- 21. Run seed ensembles
- 22. Use strong baselines
- 23. Test integrated explanations
- 24. Construct novelty-aware nulls
25–36 · Learning, decisions and pathways
- 25. Compare with GEARS
- 26. Build a novelty evidence card
- 27. Keep health claims separate
- 28. Design a prospective experiment
- 29. Connect PerturbNet to school science
- 30. Try a descriptor activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family novelty habit
- 35. Use precise PerturbNet language
- 36. Finish with a novelty-to-lab bundle
Section 1 of 36
1. Begin with unseen perturbations
Experiments cannot test every chemical or genetic intervention. PerturbNet uses descriptors of a new perturbation to generate a predicted distribution of single-cell responses.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 2 of 36
2. Represent chemicals
Molecular structure carries information about atoms, bonds and substructures. A learned representation can relate an unseen molecule to known compounds without making it biologically identical.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 3 of 36
3. Represent genetic interventions
Gene functional annotations provide prior information for unmeasured genetic perturbations. Missing or broad annotations limit what can be transferred.
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 4 of 36
4. Condition a generative model
Perturbation representations and covariates guide a deep generative mapping toward likely cell-state distributions.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 5 of 36
5. Use invertible components
Conditional invertible neural-network modules support flexible distribution transformation in latent space. Invertibility is mathematical structure, not proof of reversible biology.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 6 of 36
6. Decode cell states
A single-cell expression model converts latent samples into predicted gene profiles. Generated cells are simulations rather than new measurements.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Section 7 of 36
7. Read the 2025 evidence in scope
Yu and colleagues evaluated PerturbNet for unseen chemical and genetic perturbations using multiple benchmark datasets and interpretation analyses.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 8 of 36
8. Define unseen precisely
State whether novelty means a new molecule, chemical scaffold, gene, variant, dose, cell type or combination. These are different tests.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 9 of 36
9. Build a perturbation manifest
Record structure identifier or gene annotation, specimen, cell type, dose, time, batch, assay, intervention quality and cell count.
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 10 of 36
10. Inspect descriptor coverage
Find nearest training compounds or genes and report distances. An unseen label can still be almost identical to training examples.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 11 of 36
11. Choose a strict chemical split
Keep related scaffolds together when the goal is chemical novelty. Random compounds may overstate generalisation through close analogues.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 12 of 36
12. Choose a strict genetic split
Remove all examples for held-out genes or variants. Confirm that functional annotations were not derived from test outcomes.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Section 13 of 36
13. Protect biological experiments
Hold out plates, cultures or donors so predictions are not rewarded for shared technical structure.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 14 of 36
14. Predeclare covariates
Specify dose, cell type and other context given to the model. A forecast is conditional on those values and their training support.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 15 of 36
15. Practise with a fictional PerturbNet table
This classroom table is invented and is not output from the PerturbNet paper.
| Fictional target | Descriptor support | PerturbNet error | Baseline error | Next check |
|---|---|---|---|---|
| Molecule A | Close scaffold | 0.19 | 0.30 | New plate |
| Gene B | Rich annotation | 0.25 | 0.38 | Independent guide |
| Molecule C | Distant scaffold | 0.69 | 0.57 | Measure first |
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 16 of 36
16. Separate novelty from difficulty
Unseen does not always mean distant. Stratify targets by scaffold, embedding or annotation distance from training perturbations.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 17 of 36
17. Evaluate full distributions
Compare means, variances, correlations, quantiles and subpopulations. A strong pseudobulk score can hide incorrect heterogeneity.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 18 of 36
18. Audit chemical representations
Test alternative descriptors and scaffold splits. If performance disappears on distant chemistry, state the support boundary.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Section 19 of 36
19. Audit gene annotations
Remove or perturb annotation features and compare. Useful prior knowledge may also import bias from well-studied genes.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 20 of 36
20. Audit dose and cell context
A structure alone cannot determine response. Check predictions across doses and cell types without pooling away hard cases.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 21 of 36
21. Run seed ensembles
Repeat training and generation. Report uncertainty for each unseen perturbation rather than one average model score.
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 22 of 36
22. Use strong baselines
Compare nearest known perturbation, descriptor regression, condition means and no change. Generative complexity should earn its place.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 23 of 36
23. Test integrated explanations
Feature-attribution scores can suggest influential chemical features or annotations, but they are model explanations rather than mechanisms.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 24 of 36
24. Construct novelty-aware nulls
Randomise descriptors or annotations while preserving similarity structure. This tests whether prior information genuinely helps.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Section 25 of 36
25. Compare with GEARS
GEARS uses gene knowledge graphs for genetic perturbations; PerturbNet handles both chemicals and genes through learned condition representations.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 26 of 36
26. Build a novelty evidence card
Record holdout type, descriptor source, nearest neighbours, biological split, baseline gains, subgroup errors, seed spread and validation.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 27 of 36
27. Keep health claims separate
Cultured-cell transcription does not establish organism toxicity, therapeutic benefit, pharmacokinetics or an appropriate dose.
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 28 of 36
28. Design a prospective experiment
Choose targets across near and distant novelty, lock predictions first, then measure new experiments with matched controls.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 29 of 36
29. Connect PerturbNet to school science
Chemical structure, genes, variables and prediction form a happy bridge between biology, chemistry and computing.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 30 of 36
30. Try a descriptor activity
Describe unfamiliar objects using features, predict responses from neighbours and test where similarity helps or misleads.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Section 31 of 36
31. Connect to mathematics
Embeddings, conditional probability, invertible maps, distances and distributions organise unseen-response prediction.
PerturbNet is a generative framework designed to predict distributions of cell states under chemical or genetic perturbations not seen during training.
Archive molecular structures or gene annotations, doses, cell types, specimens, preprocessing, network settings, seeds, checkpoints and repository version.
Section 32 of 36
32. Connect to computing and AI literacy
AI can combine prior descriptors with examples, but performance depends on what unseen means and how data are split.
Chemical structure and gene functional annotations provide condition information, so quality and coverage of those representations limit novelty claims.
Use scaffold-aware chemical splits or complete unseen-gene splits, plus untouched biological experiments, before claiming novelty generalisation.
Section 33 of 36
33. Connect to school and career choices
Students excited by molecules and models can explore biology, chemistry, computing, statistics, medicine or biotechnology.
Conditional invertible neural-network components map perturbation descriptions toward latent cell-state distributions while an expression decoder generates profiles.
Compare no change, nearest known perturbation, descriptor regression and appropriate CPA or GEARS-style baselines on identical evidence.
Section 34 of 36
34. Create a family novelty habit
When AI predicts something new, ask how far it is from training examples and whether the split kept close relatives apart.
Dose and cell-type covariates help define context, but unmeasured biology and experimental batch can still confound response.
Report descriptor distance, training-neighbour support, distributional metrics, marker recovery, subgroup errors and seed uncertainty.
Section 35 of 36
35. Use precise PerturbNet language
Say descriptor-conditioned generated distribution and unseen perturbation. Avoid discovered drug, proven target or clinical outcome.
The 2025 study benchmarked chemical and genetic settings and demonstrated interpretation analyses; new compounds and genes need prospective tests.
Build a novelty card with representation source, holdout definition, nearest training neighbours, baselines, uncertainty and experimental status.
Section 36 of 36
36. Finish with a novelty-to-lab bundle
Deliver manifests, strict splits, representation audits, baselines, seed ensembles, novelty cards and prospective experiments.
Predicted expression distributions prioritise experiments and do not establish mechanism, toxicity, therapeutic benefit or appropriate human use.
Say predicted unseen-perturbation response; reserve mechanism, safety, efficacy and clinical relevance for matched independent evidence.
Did you know? A randomly held-out molecule may still share a scaffold with training compounds. Scaffold-aware evaluation asks a tougher and more useful novelty question.
Gene annotations are richer for well-studied genes, so performance should be stratified by annotation coverage to expose knowledge bias.
Prospective validation is the clearest finish: lock the prediction, run a new experiment and publish both successes and failures.
Keep chemical and genetic benchmarks separate before summarising. The model can use different descriptors and face different novelty gaps in each setting.
A useful final dashboard shows nearest training neighbours, descriptor distance, biological replication, baseline gain, distribution error and model uncertainty for every proposed experiment.
Contents · Previous section · Continue to the Science Learning Hub