Predicting minimal residual disease (MRD) levels in tyrosine kinase inhibitor (TKI)-treated chronic myeloid leukemia (CML) patients is of major clinical relevance. The reason is that residual leukemic (stem) cells are the source for both, potential relapses of the leukemicclone but also for its clonal evolution and, therefore, for the occurrence of resistance.

The state-of-the art method for monitoring MRD in TKI-treated CML is the quantification of BCR-ABL levels in the peripheral blood (PB) by PCR. However, the question is whether BCR-ABL levels in the PB can be used as a reliable estimate for residual leukemic cells at the level of hematopoietic stem cells in the bone marrow (BM). Moreover, once the BCR-ABL levels have been reduced to undetectable levels, information on treatment kinetics is censored by the PCR detection limit. Clearly, BCR-ABL negativity in the PB suggests very low levels of residual disease also in the BM, but whether the MRD level remains at a constant level or decreases further cannot be read from the BCR-ABL negativity itself. Thus, also the prediction of a suitable time point for treatment cessation based on residual disease levels cannot be obtained from PCR monitoring in the PB and currently remains a heuristic decision.

To overcome the current lack of a suitable biomarker for residual disease levels in the BM, we propose the application of a computational approach to quantitatively describe and predict long-term BCR-ABL levels. The underlying mathematical model has previously been validated by the comparison to more than 500 long-term BCR-ABL kinetics in the PB from different clinical trials under continuous TKI-treatment [1,2,3].

Here, we present results that show how this computational approach can be used to estimate MRD levels in the BM based on the measurements in the PB. Our results demonstrate that the mathematical model can quantitatively reproduce the cumulative incidence of the loss of deep and major molecular response in a population of patients, as published by Mahon et al. [4] and Rousselot et al. [5]. Furthermore, to demonstrate how the model can be used to predict the BCR-ABL levels and to estimate the molecular relapse probability of individual patients, we compare simulation results with more than 70 individual BCR-ABL-kinetics. For this analysis we use patient data from different clinical studies (e.g. EURO-SKI: NCT01596114, STIM(s): NCT00478985, NCT01343173) where TKI-treatment had been stopped after prolonged deep molecular response periods. Specifically, we propose to combine statistical (non-linear regression) and mechanistic (agent-based) modelling techniques, which allows us to quantify the reliability of model predictions by confidence regions based on the quality (i.e. number and variance) of the clinical measurements and on the particular kinetic response characteristics of individual patients.

The proposed approach has the potential to support clinical decision making because it provides quantitative, patient-specific predictions of the treatment response together with a confidence measure, which allows to judge the amount of information that is provided by the theoretical prediction.