Biological World Models0%
03 · Projects

Four projects, one virtual cell.

Each project is a thread from week one to the capstone. Track milestones and deliverables, capture notes, and ship a working artifact.

PROJECT 01 / 04

Pathway World Model

Hybrid graph + dynamics + perturbation response.

Build a hybrid pathway model using a curated graph prior, simple dynamics, and a learned perturbation residual. The model takes a baseline state plus an intervention and predicts a pathway-level response with calibrated uncertainty.

Status
Not started
Milestones
0/6
Inputs
  • →Pathway graph
  • →Baseline state
  • →Intervention
  • →Dose / time metadata (if available)
Outputs
  • ←Predicted pathway activity
  • ←Predicted cell-state shift
  • ←Uncertainty estimate
Milestones
Deliverable checklist
Model assumptions
  • A1The chosen pathway prior captures the dominant edges relevant to the intervention class.
  • A2Mass-action / Hill-like kinetics are an adequate approximation at the timescale of interest.
  • A3The perturbation acts (primarily) at one or a few annotated nodes; off-target activity is small or filtered.
  • A4Single-cell measurements are an unbiased sample of the cell-state distribution after batch correction.
  • A5The neural residual is regularized enough that it does not silently overwrite the mechanistic prior.
Falsifiable prediction

On a held-out intervention not seen during training, the model's predicted pathway-activity vector correlates with the measured response at r ≥ 0.5 with calibrated 80% intervals (empirical coverage within ±10%). If coverage collapses or correlation drops to baseline, the model is falsified for this pathway/cell context.

If this prediction fails on the stated test, the project's central claim is refuted in this scope.

Notes
PROJECT 02 / 04

Disease-State Reversal

Find perturbations that move disease toward health.

Define a disease signature and a healthy reference signature. Use the world model to search for interventions whose predicted effect maximally reverses the disease signature, subject to toxicity and context filters.

Status
Not started
Milestones
0/6
Milestones
Deliverable checklist
Model assumptions
  • A1The chosen signature representation captures clinically meaningful disease–healthy differences (not just batch).
  • A2Cell-context priors (cell type, tissue) are correctly conditioned during scoring.
  • A3Toxicity and selectivity filters are calibrated for the relevant cell type.
  • A4Linear superposition of perturbation effects is at least roughly valid for ranking.
  • A5Healthy reference is genuinely 'healthier' in the relevant axis — not just a different diseased state.
Falsifiable prediction

At least one of the top-5 ranked novel reversal interventions, when tested in the chosen cell-type model system, will produce a measurable shift along the disease–healthy axis larger than its predicted uncertainty band, without exceeding the toxicity threshold. If all 5 fail, the ranking is falsified for this context.

If this prediction fails on the stated test, the project's central claim is refuted in this scope.

Notes
PROJECT 03 / 04

Structure-to-Cell Bridge

From binding pocket to phenotype.

Pick one target. Use structure to reason about intervention mechanism. Connect target inhibition (or activation) to your pathway model and predict the downstream cell-state consequences.

Status
Not started
Milestones
0/6
Milestones
Deliverable checklist
Model assumptions
  • A1The relevant binding mode is well-represented by the structural model used (PDB or AlphaFold variant).
  • A2Target engagement at the modeled pocket maps monotonically to functional inhibition / activation.
  • A3Selectivity vs annotated off-targets has been estimated and is acceptable for the intended context.
  • A4Pathway compensation and feedback are bounded enough that target perturbation produces a predictable downstream shift.
  • A5The chosen cell context expresses the target and the relevant downstream pathway components.
Falsifiable prediction

At a target-engagement level exceeding 80% (predicted from dose and binding model), the downstream pathway-activity readout will shift in the predicted direction with effect size ≥ 1 standard deviation of baseline variability. If engagement is achieved without the predicted downstream shift, the structure-to-cell linkage is falsified for this target.

If this prediction fails on the stated test, the project's central claim is refuted in this scope.

Notes
PROJECT 04 / 04

Active-Learning Experiment Planner

Choose the next experiment to maximize information.

Build an algorithm that scores candidate perturbation experiments by expected information gain (or another acquisition function), and recommends the next experiment to run.

Status
Not started
Milestones
0/6
Milestones
Deliverable checklist
Model assumptions
  • A1Posterior uncertainty is approximately well-calibrated, so EIG is meaningful.
  • A2Experiment cost (reagents, time, cells) can be expressed numerically and compared across actions.
  • A3The action space is a useful approximation of what is actually executable in the lab.
  • A4The downstream task / decision the planner optimizes for is the right objective.
  • A5Sequential decisions can be approximated as one-step myopic without large regret.
Falsifiable prediction

Across simulated rollouts on held-out data, the active-learning policy reaches a target predictive performance with at least 30% fewer experiments than uniform random sampling, at matched cost. If the gap is below 10% across multiple seeds, the acquisition strategy is falsified for this regime.

If this prediction fails on the stated test, the project's central claim is refuted in this scope.

Notes