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 overview
Cell World Model V1.0

Ten-cycle training record

274 cell lines, 279 measured RNA features; 219 development / 55 reused test.

The three evaluation metrics are final. Development model-selection scores are not independent estimates. No outliers were removed; no test-based retuning or prediction stretching was used.

CycleWorkReasonSelection-stage CCCFinding / decision
1Verified cohort, frozen split and no-RNA baselineReplace the recalled count with the verified data; distinguish endpoint fitting from a connected kinetic model.-0.001All 274 matched lines retained. Test identities unchanged; preprocessing and supervised filtering are local to training folds.
2Connect RNA to damage/repair/injury kineticsTrain through exact state integration instead of an isolated mSF2 regression head. Sweep L2 to control overfitting.0.276Large training–validation gap remains; increased capacity is not evidence of better generalization.
3Test hidden-layer capacityThe first models can overfit hundreds of inputs; test one/two bottleneck layers with shrinkage.0.276Large training–validation gap remains; increased capacity is not evidence of better generalization.
4Train for identity-line agreementMSE favors compressed predictions. Add concordance loss inside training rather than stretching outputs after evaluation.0.282Large training–validation gap remains; increased capacity is not evidence of better generalization.
5Fold-local RNA feature selectionReduce irrelevant inputs without using held-out labels; gene ranks are re-estimated inside every training fold.0.282Large training–validation gap remains; increased capacity is not evidence of better generalization.
6Pathway bottlenecks and learned inter-module edgesCompare all-gene connections with the inherited gene/function priors and learn recurrent message-passing edges.0.282Large training–validation gap remains; increased capacity is not evidence of better generalization.
7Nonlinear injury and time-resolution testsAllow pairwise misrepair in the same kinetic injury integral; compare with linear damage and a no-stress ablation.0.282Large training–validation gap remains; increased capacity is not evidence of better generalization.
8Robustness and smooth nonlinear gene mappingTest robustness to RNA noise and outlier loss, and a radial-basis encoder; do not delete difficult cell lines.0.316Large training–validation gap remains; increased capacity is not evidence of better generalization.
9Five-seed ensemble and selection lockReduce initialization variance without selecting the seed that happens to fit a test set.0.295Five seeds averaged equally; ensemble selection-stage CCC 0.295. No best-seed picking.
10Nested adaptive search, locked final fit and reused-test auditSeparate model selection from evaluation; verify whether a strong training fit transfers to unseen cell identities.Nested 0.155 / test 0.197Nested CCC 0.155, test CCC 0.197. Test MAE 0.1411 is worse than V1.4. Kept the locked model and all results.
Final evaluationCCCMAE
Training0.7030.0920
Nested CV0.1550.1492
Reused test0.1970.1411