Why Science? | PRESCIENT, Potential Landscapes and Physical-Time Cell Trajectories

Section 1 of 36

1. Begin with a landscape

Imagine a marble moving across hills and valleys. PRESCIENT learns a high-dimensional version of that picture so cell states may drift toward likely regions while noise produces different possible journeys.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Next section

Section 2 of 36

2. Understand the potential function

The model assigns a scalar potential to latent cell states. Its gradient helps define a drift direction, so fitted valleys and ridges organise simulated development.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 3 of 36

3. Add stochastic movement

Cells with similar starts need not follow identical paths. A diffusion term represents randomness and unresolved influences, producing a distribution of trajectories rather than one promised route.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 4 of 36

4. Use physical time

Training samples come from experimentally recorded times. Those labels constrain how population distributions change, even though individual cells are destroyed and are not followed continuously.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 5 of 36

5. Learn population dynamics

The objective compares simulated and measured cell populations over time. Success means distributions align under chosen metrics; it does not mean every real lineage was recovered.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 6 of 36

6. Query new initial states

After fitting, researchers can initialise simulations in selected regions and explore possible futures. Results are most credible where similar states and times were represented in training.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Contents · Previous section · Next section

Section 7 of 36

7. Read the 2021 evidence in scope

Yeo, Saksena and Gifford introduced PRESCIENT for learning and interrogating differentiation landscapes. The paper supports studied datasets and modelling tasks, not universal biological forecasts.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Previous section · Next section

Section 8 of 36

8. Define the biological question

Name organism, tissue, process, intervention and time horizon. Decide whether the goal is interpolation, fate-probability comparison or experiment design.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 9 of 36

9. Build a time manifest

Record specimen, physical time, batch, treatment, cell count, assay and quality filters. Separate biological time from processing order and pseudotime.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 10 of 36

10. Inspect population coverage

Plot specimens and cell states at every time. A gap may be a real transition or simply an unsampled population, so uncertainty must remain visible.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 11 of 36

11. Protect independent specimens

Hold out donors, cultures or replicate series. Random cell splits can leak near-identical specimen structure across training and testing.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 12 of 36

12. Choose representation carefully

Document genes, normalisation and latent embedding. A compressed space can smooth noise while erasing rare states or clinically important variation.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Contents · Previous section · Next section

Section 13 of 36

13. Separate interpolation and extrapolation

Predictions between observed times differ from those beyond the final measurement. Label both clearly and test them separately.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Previous section · Next section

Section 14 of 36

14. Predeclare starting populations

Choose initial cells before reading outcomes. Cherry-picked starts can make one attractive valley look inevitable.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with a fictional landscape table

This classroom table is invented and is not output from the PRESCIENT paper.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 16 of 36

16. Separate probability and fate

A simulated fraction entering a basin is conditional on the fitted model, initialisation and time horizon. It is not an individual’s predetermined destiny.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 17 of 36

17. Map training support

For every prediction, report distance from observed states and density of nearby training cells. Attractive extrapolations should carry explicit warnings.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 18 of 36

18. Audit stochasticity

Run many trajectories per start and several model seeds. Monte Carlo variation and model uncertainty are different and should not be collapsed.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Contents · Previous section · Next section

Section 19 of 36

19. Test landscape smoothness

Change plausible regularisation and representation settings. Valleys that appear only under one convenient configuration are fragile findings.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Previous section · Next section

Section 20 of 36

20. Compare simple dynamics

Use no-change, linear movement and nearest-time population baselines. A learned landscape should demonstrate gains on untouched evidence.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 21 of 36

21. Hide a measured time

Train without one real timepoint and predict its population. This gives an honest test before attempting a truly unobserved time.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 22 of 36

22. Audit rare branches

Average distribution accuracy can hide a lost lineage. Evaluate known rare states, branch proportions and marker recovery separately.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 23 of 36

23. Construct temporal nulls

Shuffle time labels within valid blocks or simulate stationary populations with matching noise. The pipeline should not invent direction from batch structure.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 24 of 36

24. Compare with scNODE

scNODE learns continuous latent dynamics through a neural ODE, while PRESCIENT uses a potential landscape with stochastic diffusion. Agreement and disagreement reveal different assumptions.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Contents · Previous section · Next section

Section 25 of 36

25. Connect networks and landscapes

Scribe and TENET prioritise directional gene relationships; landscapes describe population movement. Neither automatically validates the other, but together they can sharpen experiments.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Previous section · Next section

Section 26 of 36

26. Build a trajectory evidence card

Place start support, predicted basin probability, seed spread, held-out-time accuracy, specimen replication and experimental markers together.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 27 of 36

27. Simulate interventions cautiously

Changing a feature or potential term explores a fitted counterfactual. Call it an in-silico intervention until a matched experiment tests it.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 28 of 36

28. Design the next measurement

Sample where predicted paths diverge, uncertainty peaks or a rare branch emerges. New evidence should discriminate models rather than decorate one story.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 29 of 36

29. Connect to school science

Potential energy, variables, fair comparisons and repeated measurements make the method approachable within secondary science and science enrichment.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 30 of 36

30. Try a marble-landscape activity

Roll marbles across several surfaces, repeat starts and record endpoints. Then change one ridge and distinguish observed outcomes from predicted ones.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Contents · Previous section · Next section

Section 31 of 36

31. Connect to mathematics

Gradients, probability distributions, stochastic differential equations and distances turn a landscape metaphor into a quantitative model.

PRESCIENT represents differentiation with a learned potential function and stochastic dynamics, so both landscape shape and diffusion assumptions influence trajectories.

Archive raw and processed matrices, specimen identifiers, conditions, perturbations, doses or times, gene mapping, seeds, checkpoints and software versions.

Contents · Previous section · Next section

Section 32 of 36

32. Connect to computing and AI literacy

PRESCIENT shows how optimisation, simulation and validation form a pipeline. Each output inherits assumptions from preprocessing and architecture.

Physical-time sampling constrains direction better than an unlabelled snapshot, but destructive assays still observe different cells at each time.

Use biological holdouts—whole specimens, cell types or perturbations—and keep the final test evidence untouched during model selection.

Contents · Previous section · Next section

Section 33 of 36

33. Connect to school and career choices

Learners who enjoy dynamic systems can explore biology, physics, mathematics, computing, engineering, medicine or bioinformatics.

A low-potential basin is an attractor in the fitted landscape, not automatic proof of a stable biological fate.

Compare with a no-change predictor, condition mean and an appropriate simple linear or nearest-neighbour baseline on identical splits.

Contents · Previous section · Next section

Section 34 of 36

34. Create a family prediction habit

When a model draws a smooth path, ask which points were measured, which were generated and where training support becomes thin.

Simulating from arbitrary initial states makes the model queryable, while predictions far from training support deserve visibly wider uncertainty.

Repeat training across seeds and report intervals, subgroup errors and failure cases instead of selecting one attractive run.

Contents · Previous section · Next section

Section 35 of 36

35. Use precise landscape language

Say fitted potential, simulated trajectory and predicted basin probability. Do not call generated paths observed lineages.

Intervention simulations alter a fitted system; they prioritise experiments but cannot establish treatment effects without real perturbation data.

Create a prediction card with training support, holdout type, baseline comparison, uncertainty, evidence source and planned experimental check.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a landscape-to-lab bundle

Deliver manifests, representations, holdouts, baselines, seed ensembles, support maps, nulls, evidence cards and a discriminating sampling plan.

The 2021 study demonstrated the framework on population-level time-series scRNA-seq; every new organism and protocol needs local validation.

Say model-predicted response or simulated trajectory; reserve observed, causal, safe and effective for evidence that actually supports those words.

Did you know? A potential landscape can be queried from many starts, but it does not prove that real cells traverse the model’s smoothest route. Show measured and simulated states with different symbols.

For human data, preserve donor-aware validation and consent-aware governance. Share code and derived diagnostics without assuming raw profiles may be released.

A strong conclusion names the physical times, held-out specimens, support region and next measurement that could disprove the preferred landscape.

Contents · Previous section · Continue to the Science Learning Hub