Cancer World Model · Radiation researchPublic release V1.0 · formerly Cell World Model V1.5 · unchanged fitted weights

Radiation World Model

RNA-conditioned radiation response · trained dynamic core
Version 1.0
EXPLORATORY MODEL  ·  Radiation core trained end-to-end. Test CCC 0.197, MAE 0.1411: not an overall accuracy improvement. Other cellular dynamics and drug interactions remain exploratory.
PAUSED
0.00 hours
Experiment
LIVE CELL · ADHERENT CULTURE
Generic mammalian cell · generation 0
Cell status
Cycling
Viable · no terminal fate commitment
Attached · spreading

Cell cycle

2.00 C DNA
G1SG2M
G1 · 0% of phase
Cell volume1.40 pL*1.00 × birth volume
Drag to rotate · scroll to zoom
Synthetic microscopy, not an experimental image
Cell footprint follows volume; shape follows the cell cycle. Surface texture and geometry are illustrative.Scale / pL are assumed

V1.0 · trained radiation dynamics

RNA → rates → states → survival
Single 2 Gy, 72-hour modeled repair window is the training condition. Live curves are inferred; cell-line reproductive survival is not acute death.

Lesions and delayed stress

Integrated injury → surviving fraction

Response dashboard

Damage → checkpoint → recovery / death · time zero stays fixed
Volume change—
DNA replicated0%
Completed divisions0
Last divisionNot yet
All plots use time 0 to the current simulated time. No invented oscillations are added.

Experiment timeline

0 events
No scheduled perturbations. Cell evolves in a constant environment.

Model & methods

RNA → cellular functions → dynamic response

Mechanistic prior workspace. The separate V1.0 empirical layer uses the 279-gene feature panel and frozen fitted coefficients. Each gene's normalized RNA value feeds signed gene-to-function edges. Thirteen module capacities then enter the time-dependent model. These weights are hypotheses, not fitted causal coefficients. Per-gene inputs remain fixed; aggregate RNA/protein programs evolve.

1 · Load expression

First column: gene_symbol. Other columns: cells or samples, including an explicit reference. Blanks are missing, not zero. Counts, FPKM, z scores and cross-batch values need preprocessing. Large DepMap files can be reduced with the supplied prepare_rna.py script.

2 · Select identity & reference

p53 functional status is a separate input. RNA does not determine mutation, enzyme activity or drug-target affinity. Reference normalization is not recomputed from test-cell outcomes.

3 · Inspect and apply

No expression matrix loaded.

Apply restarts time at zero and preserves the current experiment protocol. Changing the sample or weights is a draft until Apply. With at least 266 of the 279 trained genes, the empirical layer is also enabled. Molecular prior edges remain unfitted. Endpoints use development-only normalization, independent of this reference selector.

Measured baseline RNA → centered gene inputs (0–1) → signed prior weights → bounded module capacities → ODEs / cell fate

Function capacities and data coverage

Editable directed gene → function edges

GeneFunctionSample log₂(TPM+1)ReferenceInput node (0–1)Signed weightWeighted log difference
For each function: score = Σ(w × clipped gene log difference) / Σ|w|; multiplier = 2^(strength × score), bounded 0.25–4. Missing pairs contribute zero to the score (not RNA zero). Below 50% weighted coverage, the module remains neutral. At least 30% overall mapped-gene coverage is required. These editable mechanistic edges are untrained. They do not change the frozen empirical endpoint coefficients or their validation results.
Real data · frozen coefficients · held-out audit

RNA-conditioned response laboratory

Not a universal validated cell. Trained: radiation mSF2 and three drug ln(IC50) endpoints. Not trained: full signaling trajectories, apoptosis timing, drug combinations or environmental interactions. Radiation / MK-2206 have no held-out MAE advantage; individual predictions are unreliable.

1 · Choose measured RNA

274 real baseline profiles recovered from the project matrix; not newly sequenced or re-downloaded full transcriptomes. No cell-name lookup is used to predict responses.

2 · Inspect response estimates

Bands are development out-of-fold absolute-error bands, not guaranteed confidence intervals. Missing features are mean-imputed only with ≥95% coverage; RNA-distance warnings mark potentially out-of-domain profiles. Observations below are comparison labels, never simulator inputs.

Measured contrast across representative lines

This descriptive panel was selected to illustrate measured contrasts, not to estimate test performance. Development predictions are fitted values; only the locked test results below evaluate held-out cell lines. RNA is baseline; response assays are separate experiments.

Locked held-out evaluation

Same held-out ModelIDs across targets. Radiation: 219 development / 55 test; drugs: 83 / 20. Hyperparameters selected with 5-fold development-only validation. The existing 279-gene panel may have earlier selection history, so this is retrospective internal evaluation—not an independent prospective test.

Training scope, data provenance and ten cycles

V1.0 · end-to-end radiation training

Evaluated fit: 219 development lines; 55 reused test lines remain excluded.

The architecture is now end-to-end trained through its kinetic states. Prediction accuracy remains limited: CCC improved slightly, but MAE worsened. No outliers were removed.
279 measured RNA inputs → nonlinear positive kinetic rates → fast / slow lesions and stress → integrated reproductive injury → exp(−injury)

Ten actual iteration cycles

Nested CV reruns adaptive model selection inside each outer training fold. All figures retain every cell line. The inherited gene panel and historical test reuse limit independence. The optional all-274 refit in the source package is not used for these evaluation plots.

What was and was not trained

The radiation encoder, effective repair/stress kinetics and injury readout were fitted using measured mSF2. The other V1.4 cell functions, acute death thresholds and drug models were retained, not retrained. Internal kinetic trajectories are not uniquely identified by these endpoint labels. Mutations, proteins, culture effects and combination responses are not independently calibrated.