A two-compartment mechanistic model of adult T-cell acute lymphoblastic leukaemia: oncogenetic profile, MRD kinetics and relapse (working note and code)

A two-compartment mechanistic model of adult T-cell acute lymphoblastic leukaemia (T-ALL), built entirely from published aggregate data, addressing an open question: why the adverse oncogenetic profile predicts relapse independently of post-induction measurable residual disease (MRD). The model couples a cycling, chemosensitive population that MRD measures with a residual compartment that it does not, and treats relapse as stochastic escape from that compartment. Principal findings. Calibrated on seventeen published targets from GRAALL-2003/2005 (Beldjord et al., Blood 2014), the model yields three results that were not imposed on it. First, the chemosensitivity of the two oncogenetic profiles is near-identical, with a multiplier between 0.94 and 1.00 that is stable across five independent calibrations and interior to its bounds when those are relaxed above 1; a likelihood profile bounds the difference, placing the adverse profile at most twelve per cent less chemosensitive. Second, a Sobol variance decomposition (3,200 evaluations, Saltelli sampling, Jansen estimators) attributes 0.94 to 0.97 of the variance of the prognostic gap to the residual compartment, 0.02 to 0.07 to chemosensitivity, with an interaction share indistinguishable from zero. Third, the two mechanisms leave signatures separated by an order of magnitude in the plane of (MRD difference, relapse ratio), which allows any published genotype to be classified from two numbers; the resulting ordering of genotypes reproduces, by an unrelated route, the single-cell ordering reported by Costea et al. (Nat Commun 2025). The conclusion is negative. Among patients who achieve complete remission, chemosensitivity cannot account for the residual prognostic gap between oncogenetic profiles, because any difference large enough to do so would be visible in the MRD and is not. A compartment that MRD does not measure is what remains. What is not established, stated with equal force: the size of that compartment is not identifiable, the ratio between profiles varying from 44 to 1184 with optimiser convergence; the exclusion of large reservoirs was never established and those conclusions are withdrawn; no formal external validation was possible, and the single pre-registered external prediction (PETHEMA ALL-2019) failed on its point estimate and escapes refutation only for lack of power; and the scope is restricted to patients in complete remission, since two independent Spanish cohorts show that the complete remission rate itself differs markedly by oncogenetic profile, a mechanism the model does not contain. The deposit contains the full working note, the simulation and calibration code, the fitted parameters, the outputs of each analysis, and the pre-registered external prediction with its SHA-256 hashes. No individual patient data and no unpublished clinical trial data were used. This is a dated working record, not a submitted manuscript. Version 1.1.0 adds an external validation and does not modify the model. The four-gene oncogenetic classifier is reconstructed patient by patient on the 1,309 sequenced cases of the Children's Oncology Group trial AALL0434 (Pölönen et al., Nature 2024) and crossed with day-29 minimal residual disease, induction failure and relapse. Because event-free survival in that cohort is counted from enrolment and induction failure is counted as an event, the cohort does not condition on complete remission. The differential induction failure rate by oncogenetic profile — the quantity version 1.0.0 explicitly declined to estimate — is therefore measurable, and is reported: 1.8% against 4.4%. The MRD gap is 15 to 19 points depending on threshold and on how the classifier is defined, against 15.1 points observed in GRAALL adults, and the relapse ratio is 2.3 against 3.6. The direction and magnitude agree with an independent analysis of induction failure in UKALL2003 and UKALL2011 (O'Connor et al., J Clin Oncol 2023). Added files: NOTE_v1.1.pdf, cog_external_validation.py, rdata_reader.py, cog_patients.csv, cog_resultats.txt. Note under CC BY 4.0, code under MIT, as in version 1.0.0. Version 1.2.0 adds two analyses of the same cohort and does not modify the model. It is a dated record of work in progress: what is established and what is not are separated explicitly. ESTABLISHED. Conditioning on complete remission -- which every published comparison of oncogenetics and minimal residual disease does by design -- inflates the between-profile gap in relapse incidence by +0.5 points [-3.8 ; +5.0] in the Children's Oncology Group cohort AALL0434 (n = 1,309), measured by comparing two Aalen-Johansen analyses that differ only by that conditioning. The mechanism is a denominator effect governed by the DIFFERENCE in remission rates between groups, not by their level: the same arithmetic gives +3.9 points under the remission rates of PETHEMA ALL-2019 and +13.3 points under those of the earlier PETHEMA trials. In GRAALL-like conditions the inflation is therefore a few points at most, and does not explain a relapse ratio of 3.6. EXPLORED, NOT ESTABLISHED. The classifier's prognostic weight is spent in the first two to three years: among patients still relapse-free at three years the two profiles are indistinguishable (+0.3 points [-2.1 ; +2.7]). The calibrated model, run forward to ten years, reproduces this. We state explicitly why that is a weak test: the adult relapse calendar is itself a calibration target, so recovering it in a second population tests the invariance of that calendar rather than the mechanism. That invariance is itself tested here, and holds at two and three years while the one-year point diverges in the adverse profile. Added files: NOTE_v1.2.pdf, five analysis scripts, their verbatim output, one figure, and the 1,309-row patient table carried over from version 1.1.0. Note under CC BY 4.0, code under MIT, as in versions 1.0.0 and 1.1.0.