Cancer World Model · reproducible research release

Cell World Lab V1.4
Real-RNA response calibration

A working offline simulator with shared RNA-to-response predictors, measured cell-line profiles, named-drug inputs, and an auditable ten-cycle training record.

Built November 7, 2026 · Fitted model CW14-6800e2fef3ce0db5

Conclusion first: V1.4 is not yet a reliable universal cell model. Nutlin-3a and palbociclib show exploratory signal on a small held-out set. Radiation and MK-2206 do not beat the no-RNA MAE baseline. The intracellular kinetics, death timing, combination responses, and environmental interactions have not been fitted by these endpoint data.

1. Run the model

Extract the package and open Cell_World_Lab.html in Chrome or Edge. The initial profile is real A549 baseline RNA. Open Real cells & fitted responses to choose another measured profile and inspect observations, estimates and held-out results. Select 2 Gy at hour 4 to watch the response. Drug concentrations remain drafts until Add named drug to dish is pressed.

The application contains 274 measured baseline profiles, all three named-drug predictors, the radiation predictor, coefficients, and normalization. It works offline; no file is sent to a remote service. The current appearance is a synthetic adherent-cell schematic, not a measured image of the selected line—particularly important for suspension leukemia lines.

2. What data were actually used?

DataUsed in this releaseLimitations
Baseline RNA274 cell lines × 279 genes, log₂(TPM+1), recovered from the project DepMap Public 24Q4 matrixProcessed expression subset; no new full-transcriptome download, FASTQ processing, or matched post-treatment RNA. Prior panel-selection history was not re-derived.
Radiation274 matched mSF2 values from the existing CLARIFi-linked matrixLiterature medians of clonogenic SF2 at 2 Gy; RNA and survival from separate experimental cultures.
Drug response103 unambiguously matched lines × 3 compounds; historical GDSC1001 public subsetCurated numerical text rows with COSMIC IDs and source line numbers; convenience subset, not full current GDSC. Screening limits/censoring unavailable.
CompoundsNutlin-3a (1047), MK-2206 (1053), PD-0332991/palbociclib (1054)Fitted natural-log IC50 in µM. High fitted IC50s may be extrapolated beyond measured concentrations.

CLARIFi defines mSF2 as the median surviving fraction after 2 Gy from published clonogenic assays (Oike et al.). The Sanger documentation distinguishes 72-hour viability dose-response endpoints and natural-log IC50 values. These are not acute death fractions or direct target-binding affinities. The source subset README dates its historical data to November 2015. The RNA release is Broad DepMap Public 24Q4.

3. Representative cell lines: actual contrasts and model limitations

Below, radiation values are fractions; drug pairs are observed / predicted IC50 in µM. This descriptive set spans substantial measured variation. It is not a selected validation set. Rows marked development were used in fitting and must not be interpreted as held-out accuracy.

Cell lineSplitObserved mSF2Predicted mSF2Nutlin-3a µM: obs / predMK-2206 µM: obs / predPalbociclib µM: obs / pred
A549development0.5750.7064.36 / 8.121.48 / 6.900.69 / 2.68
MCF7development0.5200.5095.59 / 13.662.43 / 4.301.49 / 4.01
BT-549development0.7020.609200.17 / 90.6514.55 / 7.2799.94 / 11.96
T-47Ddevelopment0.5870.54360.52 / 61.830.15 / 4.010.60 / 4.41
HT-144development0.0800.5232.82 / 6.0010.01 / 6.533.10 / 3.11
KYSE-140development0.8030.587144.02 / 116.597.76 / 6.077.23 / 4.86
KNS-42test0.1250.529147.77 / 16.8382.61 / 6.172.70 / 2.89
MDA-MB-468test0.5100.58991.24 / 155.713.34 / 7.0739.35 / 9.52
NALM-6development0.0900.4061.96 / 6.450.81 / 5.041.00 / 2.52
Important failure example: HT-144 has observed mSF2 0.080, but the fitted model predicts approximately 0.523; KYSE-140 has observed 0.803 but predicted approximately 0.587. Therefore the release does not yet recover these large radiation differences, even though the application can load their real RNA. Do not confuse a working RNA input route with validated response fidelity.

4. Training and locked evaluation

The development/test split is frozen by ModelID: all 14 legacy test IDs were retained and 41 others added using a fixed salted SHA256 rule. Radiation has 219 development and 55 test cells; each drug has 83 development and 20 test cells. The same global held-out identities apply across targets. Names, tissue labels, response measurements and IDs are excluded from input features. Scaling is fitted inside each development fold. Final selected models are trained only on development rows.

Model selection minimized 5-fold development out-of-fold MAE. Pearson, Spearman, CCC, RMSE and R² are also reported. Candidate configurations and ensembles were selected before a single final test readout. No tuning or split changes were made after that readout. Development model selection makes its OOF scores optimistic relative to independent validation; the reused 279-gene panel may have an earlier supervised-selection history. Tissue-exclusion tests are stress audits, not an independent nested evaluation.

EndpointTotalDevTestTest MAENo-RNA MAEPearson rCCCR²Assessment
Radiation mSF2274219550.12960.12890.2510.1580.010Weak / no demonstrated held-out advantage; do not trust individual-cell predictions
Nutlin-3a10383200.84470.98630.4760.4160.156Exploratory signal; internal test only
MK-220610383200.89280.8926-0.064-0.013-0.038Weak / no demonstrated held-out advantage; do not trust individual-cell predictions
Palbociclib10383201.00591.10180.6830.3150.213Exploratory signal; internal test only

Drug errors use ln(IC50 / µM), not µM. Radiation MAE is on the 0–1 mSF2 scale. Nutlin-3a improves test MAE by 14.4%; palbociclib by 8.7%. Radiation is 0.56% worse and MK-2206 is 0.03% worse than the mean baseline. Palbociclib’s correlation is higher than its CCC because the estimates remain compressed toward the mean; good ranking is not equivalent to accurate absolute agreement.

5. The ten actual improvement cycles

These were sequential training and audit cycles, not independent AI-agent workers. A new cycle did not necessarily improve the best development score. The best retained configurations and every candidate score are included in the package.

CycleWork performedBest dev MAE: radiationNutlin-3aMK-2206Palbociclib
1Data integrity and frozen identity split————
2No-RNA control baseline0.14311.19921.18621.2417
3Biology-only module model0.14111.19671.18621.2260
4All-gene shrinkage vs canonical 45-gene subset0.13541.11641.17691.2051
5Sparse learned RNA-response edges0.13541.06771.17691.2051
6Nonlinear RNA response mapping0.13541.06771.17691.2038
7Biology-prior plus residual-gene model0.13541.06771.17691.2038
8Regularized multi-family ensemble selection0.13541.06771.17601.1834
9Coverage, error bands and lineage-shift audit————
10Locked held-out evaluation and deployment export————

197 candidate records were retained across tasks. Final selections: radiation—all-gene ridge; Nutlin-3a—all-gene elastic net; MK-2206—two ridge models averaged; palbociclib—pathway-kernel and canonical-gene ridge averaged.

6. What is learned, and how it changes the simulation

Baseline RNA (279 measured input features)
  ├─ Frozen trained scaling + coefficients / kernels
  │    ├─ Estimated 2-Gy mSF2 (weak)
  │    └─ Three drug-specific ln(IC50) values
  │          → concentration / IC50 effective-potency terms
  └─ Existing gene → function prior network (45 of 98 genes present)
       → 13 biological capacities → molecular ODEs / cell cycle / fate

Changing only a cell name or its displayed observed outcome does not change predictions. Changing its RNA does. All cells use the same fitted coefficients; no measured-outcome lookup is used as a shortcut.

The 13 pathway-prior module weights remain unfitted; only the empirical endpoint routes were trained. Those statistical gene-response associations should not be interpreted as causal biochemical reaction-rate constants. The named-drug bridge uses RNA-predicted phenotypic IC50 as an effective potency scale. Intracellular exposure obeys dC/dt=(Cmedium−Cinside)/1.2h. Palbociclib reduces the cycle gate; MK-2206 reduces the AKT target; Nutlin-3a attenuates MDM2-dependent p53 removal and adds a small p53-dependent stabilization drive. These kinetic couplings are assumptions, not measured or fitted dose/time curves. p53 function remains a separate input.

Radiation retains the previous damage/repair dynamics. In addition, H ← H − ln(predicted mSF2) × dose/2 tracks a separate empirical reproductive-survival readout, exp(−H). At a single 2-Gy pulse it equals the fitted mSF2. Other doses/fractions use an assumed exponential extrapolation; no α/β, interval-dependent recovery or drug synergy was learned. This readout neither forces acute death nor calibrates the independent lineage-regrowth assay.

7. Import new RNA and use predictions responsibly

Use examples/Import_Template_279_Genes.csv: gene rows, one fixed Development_reference column, and one or more samples. Declare TPM or log₂(TPM+1) correctly. At least 266/279 trained features must be present; otherwise the fitted response layer is disabled. Missing values are not expression zero. Above the threshold, missing inputs are imputed from the development mean, with coverage displayed.

The default module reference is the frozen development median. Empirical predictors always use their frozen development normalization, regardless of the prior-network reference selector or weight edits. tools/prepare_rna_v14.py handles common local gene-row or cell-row matrices and creates the correct reference. Raw counts, Ensembl-only IDs, batch correction and cross-platform harmonization require additional preparation.

Error bands are 90th-percentile absolute errors from development OOF predictions of the selected model. They are selection-influenced exploratory bands, not guaranteed confidence or conformal intervals. A nearest-standardized-RNA-distance warning flags some extrapolations; absence of the warning does not establish validity.

8. Software verification

CheckResult
Model/integration/input checks27 passed; 0 failed
Real-profile stress cases45 scenarios, 76 hours each, finite/nonnegative states
Python ↔ JavaScript inference parity583 measured rows; max absolute error 3.78e-09
Browser checks34 passed; 0 failed
RegressionAll original unprofiled V1.3 presets retained their 48-hour trajectories
Offline behaviorNo HTTP requests during browser operation
InterfaceReal-profile loading, dose draft/apply, washout, radiation, worker assay, exports, replay, view switching, mobile widths

These are software/numerical checks, not evidence that the model predicts real cell fate correctly. The source package includes exact test scripts and logs. Browser tests ran in Chromium with a software-rendered graphics path; this is not a claim of validation in every browser/GPU configuration.

9. Reproduce the release

python training/train.py
python tools/export_assets.py
python build.py
node tests/test_v14.js
python tests/test_browser_v14.py

Install the Python requirements and Chromium/Playwright for the browser tests. Pure inference and the app require no installations. Exact data hashes, ModelID splits, all CV candidates, frozen model selection, train/OOF/test predictions and fitted coefficients are bundled. base/V1_3/src/model.js is retained for the regression check.

10. Next scientific calibration priority

Expand to the full matched RNA panel and richer drug-screen coverage, include independently measured genotype/function and baseline growth, and fit perturbation time courses as separate targets. Then test on new cell lines, doses, compounds and experimental batches. Endpoint fitting alone does not identify the internal dynamic network or guarantee correct behavior under an unseen trigger.

Research software only. Third-party data retain their original terms. See DATA_SOURCES.txt and data/provenance.json. The application is independently written, not SPARCED.