Quantitative prediction of long-term molecular response in TKI-treated CML - Lessons from an imatinib versus dasatinib comparison

Abstract

Longitudinal monitoring of BCR-ABL transcript levels in peripheral blood of CML patients treated with tyrosine kinase inhibitors (TKI) revealed a typical biphasic response. Although second generation TKIs like dasatinib proved more efficient in achieving molecular remission compared to first generation TKI imatinib, it is unclear how individual responses differ between the drugs and whether mechanisms of drug action can be deduced from the dynamic data. We use time courses from the DASISION trial to address statistical differences in the dynamic response between first line imatinib vs. dasatinib treatment cohorts and we analyze differences between the cohorts by fitting an established mathematical model of functional CML treatment to individual time courses. On average, dasatinib-treated patients show a steeper initial response, while the long-term response only marginally differed between the treatments. Supplementing each patient time course with a corresponding confidence region, we illustrate the consequences of the uncertainty estimate for the underlying mechanisms of CML remission. Our model suggests that the observed BCR-ABL dynamics may result from different, underlying stem cell dynamics. These results illustrate that the perception and description of CML treatment response as a dynamic process on the level of individual patients is a prerequisite for reliable patient-specific response predictions and treatment optimizations.

Trial registration: ClinicalTrials.gov NCT00481247.

Figure 1

Data flowchart. The flowchart illustrates the selection process of patient data for statistical and model analyses.

Figure 2

Statistical und mechanistic model. (A) Bi-exponential model for the description of time course data, parameterized by the intercepts A, B, and the initial slope α and the secondary slope β. (B) Model setup of the mechanistic, single-cell based clonal competition model of CML pathogenesis and treatment. Leukemic cells are shown in blue, normal cells in grey. Both cell types change between a state of proliferative inactivity (A) and a proliferative state (Ω) before cells differentiate into peripheral blood. TKI activity is indicated by the cytotoxic effect and the prolonged quiescence of leukemic stem cell in state A. The five parameters modified for the simulation screen are identified by roman numerals (i)–(v).

Figure 3

Dynamics of treatment response. (A) Time course data of a random subset of 8 patients per treatment cohort. BCR-ABL levels below detection threshold are indicated by open triangles. (B) Time-course of mean BCR-ABL levels (±SD) are shown for intervals of 2 months. The lines correspond to the fixed-effect predictions of the mean for the imatinib and the dasatinib cohorts. (C) Scatter plot illustrating the missing correlation between initial slope α, and secondary slope β for all available, individually fitted patient time courses (Spearman correlation [95% confidence interval]: imatinib r = 0.27 [0.13; 0.40], dasatinib: r = 0.17 [0.03; 0.31], all: r = 0.24 [0.14; 0.33]).

Figure 4

Estimation of model parameters and prediction accuracy. (A) The optimal fit of the bi-exponential regression model ${\hat{\rho }}_i^{\tau }$ is shown along with a point-wise confidence interval for one patient i. (B) Identification of all parameter configurations ${\hat{\mathrm{\Theta }}}_i^{\tau }$, for which the bi-exponential fit of the resulting model simulation is contained within the confidence region of the patient’s kinetic. (C) Scatter plot relating the initial slope α of each patient’s response with the rate of gradual TKI-effect onset (r_trans) obtained for the most suitable model simulation ${\hat{\vartheta }}_i^{\tau }$. (D) Scatter plot relating the long-term decline β of each patient’s response with the specific activation rate of the residual LSC (${\mathrm{f}}_{\mathrm{\omega }}^{\mathrm{CML}})$ obtained for the most suitable model simulation ${\hat{\vartheta }}_i^{\tau }$. (E) False positives (FP) and false negatives (FN) rates for predictions of 5 year outcomes as a function of shorter observation periods (n = 234).

Figure 5

Estimating residual stem cell numbers. (A) Variability for predictions of residual leukemic stem cell (LSC) numbers for the model simulations ${\hat{\mathrm{\Theta }}}_i^{\tau }$in Fig. 4A,B. Every green line corresponds to one suitable parameter configuration within ${\hat{\mathrm{\Theta }}}_i^{\tau }$indicated in blue in Fig. 4B. The green coloring scheme indicates the distance of the simulation results to the bi-exponential approximation ${\hat{\rho }}_i^{\tau }$of patient i. lsc_max and lsc_min refer to the maximal and minimal number of predicted LSCs at 5 years. (B) Correlation of the residual variance ${\hat{\sigma }}_i^2$ of the model fit (as a measure of data quality) to the variability of the number of predicted LSCs. (C) Correlation of the width of the prediction interval in the peripheral blood and the variability of the number of predicted LSCs. (D) Correlation of the second slope β of the patient fit ${\hat{\rho }}_i^{\tau }$with the maximal number of predicted LSCs.