Leukemic fitness predicts treatment-free remission and guides adaptive therapy in chronic myeloid leukemia

Chronic myeloid leukemia is controlled by tyrosine kinase inhibitors, yet only about half of well-responding patients who stop therapy remain in long-term remission, and currently it cannot be predicted who will relapse. Mechanistic models reproduce both outcomes but rely on many parameters and longitudinal immune measurements, limiting clinical use. Here we derive a minimal two-equation model in which the immune compartment is eliminated by a quasi-steady-state argument and retained implicitly as a tumor load-dependent kill rate. This reduction lowers the parameter count from fourteen to eight, requires no direct immune measurements, and preserves the two key features of immune-leukemia interaction: a switch-like per-capita kill rate and a bell-shaped immune window peaking at deep remission. A complete analytical treatment shows that a patient's long-term outcome depends on a single quantity, the leukemic fitness, defined as the ratio of leukemic proliferation to maximal immune-mediated killing. Values above one render the cure steady state unstable, whereas values below one make it stable and therefore attainable; the rate-limiting step is depletion of the quiescent stem cell reservoir below a separatrix, reproducing the clinical link between treatment duration and remission level with durable remission after cessation. Fitted to 75 patients of the DESTINY trial, the minimal model matches the fit quality of its more detailed predecessors. Leukemic fitness estimated from time courses up to treatment cessation, including the dose-reduction period, is associated with outcome, with discrimination improving further when six months of post-cessation data are added. Guided by this single parameter, we design two personalized strategies: extended full-dose therapy for low-fitness patients who have not yet crossed the separatrix, and maintenance dosing for high-fitness patients. Leukemic fitness is thus a clinical predictor that can be estimated from routine monitoring data, and a minimally parametrized model is enough to turn patient trajectories into personalized treatment decisions.