Radiation World Model V1.0 is the renamed Cell World Model V1.5. Existing figures may retain historical version labels. No weights, equations, or evaluation results changed in this release.
Model overviewCancer World Model · V1.0 · November 14, 2026
End-to-end radiation-response training
A trainable RNA-to-kinetics cell subsystem, ten recorded improvement cycles, and the three established accuracy plots.
Research release · evaluated model CW15-57df53d3db170d4c · 274 verified cell lines × 279 RNA features
Result: V1.0 successfully trains through its internal radiation-response states. It is not an overall accuracy improvement: reused-test CCC increased from 0.158 to 0.197, while MAE worsened from 0.1296 to 0.1411. Nested-CV CCC is only 0.155. This is an experimental end-to-end model, not a validated universal cell or a clinical predictor.
1. Data and scope
The exact matrix recovered from V1.4 contains 274 matched cell lines, not the recalled 278. Its 279 gene columns are baseline log₂(TPM + 1) expression and its target is measured mSF2. All 274 lines were retained; there was no outlier deletion or substitution of missing cells.
2. What changed from V1.4
V1.4 fitted a separate RNA→mSF2 regression and imposed its survival estimate at a radiation pulse. In V1.0, the loss is backpropagated through the exact solution of a coupled kinetic system. RNA-to-rate weights and shared kinetic parameters are fitted together, and the live core uses the same equations.
279 RNA inputs→Trainable gene / kinetic network→Damage + stress states→Integrated injury→mSF2
Radiation pulse D: F ← F + 0.8 D; S ← S + 0.2 D
dF/dt = −k_fast F
dS/dt = −k_slow S
dA/dt = u(F + S) − v A
dH/dt = a(F + ρS) + cA + b(F + S)²
Predicted surviving fraction at time t = exp[−H(t)]
Training prediction = mean over 5 independently initialized models at t = 72 h
Training dose = 2 Gy; baseline sham = 1
F and S are fast- and slow-repairing lesion proxies. A is an inferred delayed stress signal, not a measured p53/BAX concentration. H is cumulative reproductive injury, not the probability that the displayed cell is acutely dead. All lesion/stress quantities have model-relative units. The 72-hour window is an assumed damage-resolution horizon—not an inferred assay collection time.
The tested families included direct gene-to-rate encoders, one- and two-hidden-layer networks, pathway bottlenecks, recurrent module connections, hybrid networks and radial-basis encoders. The retained model is the simpler, heavily regularized all-279-gene encoder, followed by nonlinear positive rate transforms and state integration. Extra hidden layers did not outperform it on development CCC.
The decoder has five latent slots; the optional quadratic-injury channel is inactive in the retained model (b = 0). Four RNA-conditioned channels remain active: fast repair, slow repair, injury susceptibility and stress activation. Each ensemble member stores 1,408 parameter slots, including the inactive optional channel. Five seeds are averaged equally, not selected by test performance.
Training objective:
MSE / (training-target variance + 0.002)
+ 2 × (1 − CCC)
+ 20 × sum of squared network weights
+ 0.015 × sum of squared shared kinetic offsets
180 full-batch Adam epochs; learning rate 0.012
RNA training noise: standard deviation 0.08 in standardized units
No best-seed selection, test-set calibration or output stretching
Gradients were checked to reach the RNA inputs, both repair-rate connections, injury susceptibility and stress activation. Closed-form state propagation was independently compared with SciPy integration. Exact integration avoids making training quality depend on an arbitrary numerical step size.
Integration with the existing simulator: the fitted radiation core now produces a live survival trajectory rather than an instantaneous regression-score drop. Learned mean fast/slow repair rates also feed the legacy damage equations. The legacy cell’s p53, RNA/protein, metabolism, acute fate and drug responses remain approximate; those functions were not all calibrated by mSF2.
3. Evaluation protocol and final performance
Candidate selection maximized development out-of-fold CCC, with MAE as a tie-breaker. Ten-cycle model design was restricted to development data. The nested audit used 5 outer folds and 3 inner folds, rerunning the adaptive search independently inside each outer training set. Only then were five models fitted to all 219 development lines and evaluated on the unchanged 55 historical test lines.
There were 29 development configurations and 145 configurations across the five inner searches, totaling 640 neural model fits including ensemble refits and the separately labeled all-data fit. The historical test was reused from earlier project work: it is not new external validation. Repeated project-level development and inherited features still limit independence.
On the same 55 test cells, the CCC gain is 0.040, but MAE is 8.9% worse. Both Pearson correlation and RMSE also worsen. Higher CCC alone is not enough to call this a better predictor.
The test slope is 0.139, far from the identity slope of 1. Predictions remain compressed, and extreme radiosensitivity/resistance is not reliably recovered. Negative nested/test R² values further show the generalization problem.
Descriptive paired bootstrap (2,000 resamples): test CCC 95% range [-0.068, 0.421]; test MAE range [0.1109, 0.1730]. These intervals condition on the fitted model and are not corrected for historical selection/reuse.
4. The three established plots
Measured mSF2 is on the horizontal axis and prediction on the vertical axis. Every plot uses the full 0–1 range, an identity line and an ordinary regression line. All cell lines are retained.
Training fit: 219 development lines. This measures fit to data used for optimization, not generalization.
Nested CV: 219 development lines, each predicted only by an outer-fold model not fitted or selected using that line’s outcome.
Reused test: 55 legacy test cell lines excluded from V1.0 fitting and model selection. Historical reuse prevents treating this as independent external validation.5. Ten actual cycles: changes, reasons and outcomes
Development selection scores are optimistic compared with nested validation. A cycle that failed to improve the retained model is shown as such; the record does not assume ten consecutive gains.
6. What the results diagnose
Overfitting was the dominant observed failure. In the weakly regularized first kinetic model, average training CCC reached 0.976 but validation CCC was only 0.101. Stronger shrinkage improved validation. A 16-node hidden network could also fit training data strongly without corresponding validation improvement. The regularized 32→8 model collapsed toward a constant prediction.
Increasing output variation did not solve the association problem. Concordance-focused loss and input-noise regularization raised selection-stage CCC, but nested validation substantially reduced the apparent gain. The low slope is not fixed by adding more layers or stretching predictions; the input-to-outcome relationship must generalize.
These data do not uniquely identify intracellular dynamics. They provide one baseline RNA profile and one 2-Gy survival label per line, not matched molecular time courses. Different internal kinetics can reproduce the same endpoint.
The example changes the fast-repair rate and compensates with an injury coefficient. It is a mathematical identifiability demonstration, not a second biological measurement. Therefore learned parameters should be called effective endpoint-calibrated rates, not verified gene-specific biochemical constants.
Other limits: RNA does not automatically specify functional mutation status or active protein levels; the medium is clamped; the image is synthetic; acute death and reproductive survival differ; the 24-lineage cohort remains illustrative; the trained core has no fitted drug/environment interaction input. Non-2-Gy doses and fractionated schedules are extrapolations. Named-drug endpoint predictors are retained from V1.4, not improved by this radiation-only training.
7. Software verification
Testing used Chromium with software-rendered graphics. These are software/numerical checks, not experimental validation of cell fate or molecular biology.
8. Run and reproduce
Run the app: extract the application ZIP and open Cell_World_Lab.html in a desktop browser. It needs no installation or Internet connection. Start with A549, choose “2 Gy at hour 4,” and advance time. The V1.0 trained-core panel tracks lesions, stress, injury and surviving fraction. The Training results button opens the three plots and iteration record.
Evaluation model: models/trained_ensemble.json is fitted to the 219 development lines only. The downloadable app uses this model so its predictions correspond to the evaluation table. Optional full-data model: models/trained_all274_refit.json was fitted to all 274 lines using the locked configuration; it has no held-out subset.
# From the complete source-package folder:
python -m pip install -r requirements.txt
python training/audit_data.py
python training/run_cycles.py
python training/nested_and_final.py
python training/refit_all.py
python tests/test_core.py
node tests/test_js.js
python app/build.py
# Predict from an RNA matrix (gene-symbol columns, log2(TPM+1)):
python training/predict.py --rna data/RNA_274_Cell_Lines_279_Genes_Log2TPMplus1.csv --out predictions.csv
# Optional all-data refit; do NOT interpret included lines as held out:
python training/predict.py --rna your_RNA.csv --model models/trained_all274_refit.json --out full_fit_predictions.csv
Python numerical dependencies are pinned in requirements.txt. CPU PyTorch may require its CPU wheel index on some systems. Browser tests additionally need Chromium; its location is configurable in the supplied test script. All fitted coefficients, raw input subset, split manifests, candidate scores, outer-fold choices, predictions, losses, gene weights, inferred rates and tests are included.
9. Source and method references
The requested basis is the supplied V1.4 source package and its RNA–mSF2 matrix. Public data references inherited from that package are listed below; these do not validate the newly inferred molecular trajectories.
- Oike et al., CLARIFi radiosensitivity database, Cancer Science, DOI 10.1111/cas.16334; author data deposit.
- Broad DepMap, Public 24Q4 release. This release reuses the project’s processed 279-gene subset.
- PyTorch, automatic differentiation mechanics.
- scikit-learn, nested versus non-nested cross-validation.
Bottom line: V1.0 replaces the disconnected radiation endpoint with an end-to-end-trained kinetic response core, but it does not establish good predictive accuracy. Further model claims should depend on better out-of-sample performance and independent molecular/functional measurements, not training fit alone.