Model-Based Inference and Classification of Immunologic Control Mechanisms from TKI Cessation and Dose Reduction in Patients with CML

Chronic myeloid leukemia (CML) is a myeloproliferative disorder, which is characterized by the unregulated proliferation of immature myeloid cells in the bone marrow. CML is caused by a chromosomal translocation between chromosomes 9 and 22. The resulting BCR-ABL1 fusion protein acts as constitutively activated tyrosine kinase triggering a cascade of protein phosphorylation, which deregulate cell cycle, apoptosis regulation, cell adhesion, and genetic stability. Because of their unregulated growth and their distorted differentiation, immature leukemic cells accumulate and impair normal hematopoiesis in the bone marrow, leading to the patient's death if left untreated.

Tyrosine kinase inhibitors (TKI) specifically target the kinase activity of the BCR-ABL1 protein with high efficiency and have been established as the first line treatment for patients with CML (1). Individual treatment responses are monitored by measuring the proportion of BCR-ABL1 transcripts relative to a reference gene, for example, ABL1 or GUS, in blood cell samples by using reverse transcription and quantitative real-time PCR (qRT-PCR; refs. 2–4). Most patients show a typical biexponential treatment response with a rapid, initial decline (α slope), followed by a moderate, second decline (β slope; refs. 5–7). Whereas the initial decline can be attributed to the eradication of proliferating leukemic cells, the second decline has been suggested to result from a slower eradication of quiescent leukemic stem cells (3, 4, 8, 9). Within five years of treatment, about two thirds of the patients achieve a major molecular remission (MMR), that is, a BCR-ABL1 reduction of three logs from the baseline (MR3), while at least one third of these additionally achieve a deep molecular remission (DMR, i.e., MR4 or lower; refs. 4, 7, 10).

TKI discontinuation has been established as an experimental option for well responding patients with DMR for at least one year (11, 12). Different studies independently confirmed that about half of the patients show a molecular recurrence, while the others stay in sustained treatment-free remission (TFR) after TKI stop. Consistently, most patients present with a recurrence within 6 months, while only a few cases are observed thereafter (11–14). The overall good response of those patients after restarting treatment with the previously administered TKI indicates that clonal transformation and resistance occurrence is not a primary problem in CML. As it appears unlikely that even a sustained remission truly indicates a complete eradication of the leukemic cells, other factors have to account for a continuing control of a minimal, potentially undetectable residual leukemic load. Although treatment discontinuation is highly desirable to reduce treatment-related side-effects and lower financial expenditures (15, 16), it is still not possible to prospectively identify those patients that are at risk for a molecular recurrence. Investigations of clinical markers and scores to predict the recurrence behavior of patients after the treatment cessation revealed that both TKI treatment duration and the duration of a DMR were also associated with a higher probability of TFR (11, 13, 17, 18). However, it is still unclear whether the dynamics of the initial TKI treatment response (e.g., the initial slope of decline) correlate with the remission occurrence after treatment discontinuation.

The underlying mechanisms of the recurrence behavior after TKI stop are still controversial. While fewer recurrences for patients with longer treatment suggest that a leukemic stem cell exhaustion is an important determinant, it is not a sufficient criteria to prospectively identify nonrecurring patients (13, 17). Favorable outcomes of treatment discontinuation for patients that were previously treated with immune-modulating drugs, such as IFNα, suggest that immunologic factors might play an additional and important role (11, 18, 19). In this context, it has been demonstrated that specific subpopulations of dendritic cells and natural killer cells, as well as the cytokine secretion rate of natural killer cells are associated with higher probabilities of a treatment-free remission (20, 21). Furthermore, there are several reports about patients with low but detectable BCR-ABL1 levels over longer time periods after therapy discontinuation that do not relapse (14, 22). This is a strong indicator that also other control mechanisms, such as the patient's immune response, are important determinants of a TFR.

Mathematical oncology has been established as a complementary effort to obtain insights into cancer biology and treatment. At the same time, model-based understanding of individual patient records is developing into a key method for devising adaptive therapies in the coming era of personalized medicine (23–26). CML is a show-case example, where several models have propelled the conceptual understanding of CML treatment dynamics (5–7, 9, 27–33) and are considered for the design of new clinical trials (34). Especially the long-term effect of TKI treatment on residual stem cell numbers and the effect of combination therapies were in focus. In a recent publication, we provided evidence that TKI dose reduction is a safe strategy for many patients in sustained remission while preserving the antileukemic effect (9). Complementary efforts also accounted for interactions between leukemic and immune cells (35–39). In a prominent approach, Clapp and colleagues used a CML–immune interaction to explain fluctuations of BCR-ABL1 transcripts in TKI-treated patients with CML (37). However, it remains elusive to which extent an immunologic control is a crucial mediator to distinguish patients that maintain TFR from those that will eventually relapse.

Here, we used BCR-ABL1 time courses of TKI-treated patients with CML that were enrolled in previously published TKI discontinuation studies from different centers in Europe. In particular, we focused on patients for which complete time courses during the initial TKI therapy and after treatment cessation are available. Therefore, potential correlations between response dynamics, remission occurrences, and timings after cessation become accessible. Motivated by the observation that the initial treatment response before TKI cessation does not show obvious correlations with remission occurrences, we aim to explain the resulting dynamics in terms of an ordinary differential equation (ODE) model of TKI-treated CML. Explicitly including a patient-specific, CML-dependent immune component, we are able to demonstrate that three different immunologic configurations can determine the overall outcome after treatment cessation. We further investigate how this patient-specific configuration can be estimated from system perturbations, such as TKI dose reduction scenarios prior to treatment cessation. Our predictions closely resemble recent clinical findings substantiating our conclusion that treatment response during TKI dose reduction is indeed informative to predict a patient's future outcome after stopping therapy (40).

We analyzed time courses of 60 TKI-treated patients with CML, for whom TKI therapy had been stopped as a clinical intervention. Informed written consent was obtained from each subject according to the local regulations of the participating centers. Corresponding clinical trials were conducted in accordance with the Declaration of Helsinki and applicable regulatory requirements. The protocols were approved by the Institutional Review Board or ethics committee of each participating center. Detailed information on the patient cohort is available in the Supplementary Materials. For all 60 patients, serial BCR-ABL1/ABL1 measurements before as well as after cessation are available. For the purpose of this analysis, the date of a molecular recurrence after cessation was defined as the first detected BCR-ABL1/ABL1-ratio above 0.1%, indicating a loss of MR3, or the reinitiation of TKI treatment, whatever was reported first.

Furthermore, we selected patients who received TKI monotherapy before stopping, who were monitored at a sufficient number of time points to estimate the initial and secondary slopes, and who presented with the typical biexponential response dynamic (Fig. 1A, Supplementary Materials). The 21 selected patients, fulfilling those criteria, were compared with the full patient cohort (n = 60) and showed no obvious differences for the initial BCR-ABL1 levels, treatment duration, recurrence behavior, follow-up duration, recurrence times and used TKI, and are, therefore, considered to be representative examples (Supplementary Fig. S1). Moreover, the overall recurrence behavior of the selected patient cohort is comparable with larger clinical studies (11, 14).

For our analysis, we apply an ODE model, which we proposed earlier in a methodologic article qualitatively comparing a set of CML models with different functional interaction terms between leukemic cells and immune cells (39).

This model is sketched in Fig. 1B and formally described by:

The model distinguishes between a population of quiescent leukemic cells (⁠|X$|⁠) and a population of actively cycling leukemic cells (⁠|Y$|⁠), which proliferate with the rate |{p_Y}$|⁠, whereas the growth is limited by a carrying capacity |{K_Y}$|⁠. Leukemic cells can switch reversibly between the active and the quiescent state with transition rates |{p_{XY}}$| and |{p_{YX}}$|⁠. Apoptosis is negligible for the quiescent population |X$| and can be efficiently integrated in the proliferation term for the activated cells |Y$|⁠. TKI treatment is modeled by a kill rate |TKI$|⁠, which acts on proliferating cells |Y,$| but does not affect quiescent cells |X$|⁠. Furthermore, we do not explicitly include resistance occurrence in the current model as it does not present a major challenge in CML treatment. A complete eradication of leukemic cells is defined as a decrease of leukemic cells in |X$| and |Y$| below the threshold of one cell. The corresponding BCR-ABL1/ABL1 ratio in the peripheral blood is calculated as the ratio of proliferating leukemic cells to the carrying capacity |{K_Y}$| (see Supplementary Materials for details).

Furthermore, the model integrates a population of CML-specific immune effector cells (⁠|Z),$| which are generated at a constant, low production rate |{r_Z}$| and undergo apoptosis with rate |a$|⁠. They eliminate proliferating leukemic cells |Y$| with the kill rate |m$|⁠. The leukemia-dependent recruitment of immune cells follows a nonlinear functional response where |{p_Z}$| and |{K_Z}$| are positive constants. This functional response leads to an optimal immune cell recruitment for intermediate leukemic cell levels. Low numbers of proliferating leukemic cells |(Y \lt \, {K_Z}$|⁠), the immune cell recruitment increases, and the immune cells |Z$| are stimulated to replicate in presence of proliferating leukemic cells |Y$|⁠, reaching a maximum |{p_Z}/( {2{K_Z}} )$| when |Y\ = \ {K_Z}$|⁠. For higher leukemic cell numbers (⁠|Y \gt \, {K_Z}$|⁠), the immune cell recruitment decreases with |Y$|⁠, reflecting the assumption that the proliferation of immune cells is decreased for high levels of proliferating leukemic cells |Y$|⁠. This assumption follows recent findings, suggesting that a high load of CML cells inhibits the immune effector cells' function and number (41). As a result, we obtain an immune window for which the recruitment exceeds the degradation rate |a$| of the immune cells and leads to an optimal immune response (see Supplementary Materials).

For all patients, we use fixed, universal values for the immune-mediated killing rate |m$|⁠, the proliferation rate |{p_Y}$|⁠, the carrying capacity |{K_Y}$|⁠, the immune cells natural influx |{r_Z},$| and the immune cells apoptosis rate |a$|⁠. In contrast, the transition rates |{p_{XY}}$| and |{p_{YX}}$|⁠, the TKI kill rate |TKI$|⁠, and the immune parameters |{K_Z}$| and |{p_Z}\ $|are considered patient-specific parameters and are estimated with different strategies (see Supplementary Material).

Comparing the BCR-ABL1 kinetics of the 21 TKI-treated patients with CML before treatment cessation, we detected no obvious differences between the recurring and nonrecurring patient groups, that is, we found no markers in the patient data, which could potentially serve as a predictive measure to prospectively identify patients that show a treatment-free remission after treatment cessation (see Supplementary Materials and Supplementary Figs. S2 and S3). Motivated by these results, we developed an ODE model of CML treatment to investigate which part of a patient's individual therapy response confers the relevant information to reliably distinguish recurrence from nonrecurrence patients. To do so, we investigated which level of model complexity and what type of patient data are necessary as inputs to obtain model fits that sufficiently represent the available BCR-ABL1 data before and after treatment stop and that would allow to anticipate the response dynamics to TKI cessation. The models and input data used in each fitting strategy are presented below, with an increasing level of complexity.

As a reference model, we use a reduced version of the suggested ODE model without an immunologic component, that is, all immunologic parameters values are set to zero (Fig. 1B; see Materials and Methods, |{p_Z}$| = |{K_Z}\ = \ a\ = \ {r_Z}\ = \ m\ = \ 0$|⁠). This model predicts a complete eradication of residual disease levels only for very long treatment times. Thus, treatment cessation at any earlier time point will eventually lead to recurrence. Adapting this model to each available, individual patient time course by estimating the patient-specific model parameters |{p_{XY}}$|⁠, |{p_{YX}},$| and |TKI$| from the precessation BCR-ABL1 data, we confirm that relapse is predicted for all patients (Fig. 2A, example time courses in Fig. 2B and C), which is in contrast to the clinical observations. In summary, the reference model without immune system is not suitable to describe the nonrecurrence cases and thereby opposes clinical findings of TFR (11–14).

Clinical studies suggest that immunologic components can potentially control minimal residual disease levels and, therefore, might prevent (molecular) recurrences after TKI stop (20, 21). Therefore, we use the ODE model (Materials and Methods, Eqs. A–C) that explicitly considers an immune component (39). Because measurements of individual antileukemic immune conditions are not available, we investigate three different approaches (fitting strategy I–III) for estimating the relevant immune parameters |{K_Z}$| and |{p_Z}$| and compare the corresponding simulation results with the clinical data.

In fitting strategy I, we consider a generic immune system configuration with identical immune parameters |{K_Z}$| and |{p_Z}$| for all patients. The remaining parameters |{p_{XY}}$|⁠, |{p_{YX}},$| and |TKI$| are estimated by individually fitting the model to the precessation BCR-ABL1 time courses. A grid-based search in the |({K_Z}$|⁠, |{p_Z})$| space of immune parameters only identifies configurations in which the overall rates and timings of recurrence are not sufficiently met (Fig. 2D). Furthermore, the model predictions fail on the individual level, as neither the time courses nor the recurrence behavior could be predicted reliably (Fig. 2D, inset). These findings as well as the recognition of immunologic differences between different patients argues in favor of patient-specific immune parameters.

In fitting strategy II, besides parameters |{p_{XY}}$|⁠, |{p_{YX,}}$| and |TKI$|⁠, we also estimate patient-specific values for immune parameter |{K_Z}$| and |{p_Z}$|⁠; however, we only apply the fitting routine to the precessation BCR-ABL1 time courses. We observe no statistically significant difference between recurring and nonrecurring patients with respect to the fitted immune parameter values (Supplementary Fig. S4). Furthermore, the optimal fits fail to correctly predict the outcomes for individual patients (Fig. 2E). This indicates that the configuration of the immune response is most likely not imprinted in the patient response under TKI treatment, in which the drug mediated leukemia reduction is the dominating process.

In fitting strategy III, we provide pre- and postcessation data to fit patient specific model parameters |{p_{XY}}$|⁠, |{p_{YX}}$|⁠, |TKI$|⁠, |{K_{Z,}}$| and |{p_Z}$|⁠. We demonstrate that a patient-specific immune configuration is sufficient to consistently explain the clinical data (example time courses in Fig. 2B and C; complete data in Supplementary Fig. S5; Supplementary Table S1), and that it can be obtained from patient's response after TKI stopping. The model correctly describes the behavior on the population level (Fig. 2F), as well as on the individual patients (Fig. 2F, inset).

Having a univariate look at the individually estimated parameters of the immune model using fitting strategy III (Fig. 3A–E), we observed only minor differences between the recurring and the nonrecurring patients, that do not allow to clearly distinguish the patient groups. However, a bivariate analysis of the immune parameters |{K_Z}$| and |{p_Z}$| reveals a distinction between recurrence and nonrecurrence cases (Fig. 3F). In particular, a lower value for the location of the immune window |{K_Z}$| together with a higher proliferation of the immune cells |{p_Z}$| convey a favorable outcome after therapy stop. This pattern is also confirmed at the level of individual patients in which we studied the predicted outcome for optimal fits with systematically varying immune parameters (Fig. 3G; Supplementary Fig. S6). This analysis reveals distinct parameter regions for which either remission or recurrence is predicted, although the precise location of those regions further depends on all model parameters.

From these results, we conclude that an individual immunologic component (or another TKI-independent antileukemic effect) is necessary to quantitatively explain the individual BCR-ABL1 time courses of CML patients before and after stopping the TKI treatment. Our results also suggest that the correct estimation of the parameters describing such immunologic component for each patient is not possible based on the BCR-ABL1 dynamics under constant TKI treatment alone and cannot be used for the prospective prediction of the molecular recurrence after TKI stop.

Dynamical models, as the one suggested here for the interaction between leukemic and immunologic cells (Fig. 4A), are characterized by steady states that describe configurations in which the model quantities (in our case, cell populations) have reached an equilibrium. Stable steady states and their basins of attractions are conveniently depicted in a state space representation, which mimics a physical landscape (Fig. 4B; ref. 42). Typical steady states in our model refer to a fully developed leukemia (“disease steady state”, |Y \approx {K_Z}$|⁠) or an immunologic control of residual leukemic levels (“remission steady state”, |Y \ll {K_Z}$|⁠), while trajectories represent dynamical changes of the system state along time. The existence and the precise location of the steady states and their basins of attraction depend on the particular leukemic and immunologic model parameters and thereby determines the range of possible steady states that can be achieved after treatment cessation (Fig. 4C). As these parameters, obtained from fitting strategy III, differ between individual patients, they also describe “patient-specific immunologic landscapes.”

A detailed mathematical analysis suggests that the available patients can be grouped in three general classes that correspond to structurally different underlying landscapes of the ODE model:

• Class A: For certain parameter configurations, the immunologic landscape has only one stable steady state, namely the recurrence steady state |{E_H}$|⁠. This means that the patient will always present with recurring disease after treatment cessation in this model due to an insufficient immune response, if CML is not completely eradicated, irrespective of the degree of tumor load reduction. The corresponding immunological landscape is visualized in Fig. 5A and depicts the recurrence behaviour depending on the number of immune cells and leukemic cells at treatment cessation. According to our estimates, 6 of 21 patients fall into class A and ultimately present with recurring disease after treatment cessation (example in Fig. 5B).

• Class B: For other parameter configurations, the immunologic landscape has two stable steady states: the disease steady state |{E_H},$| and the remission steady state |{E_L}$|⁠. In this case, there is a distinct remission level of BCR-ABL1 abundance, below which a strong immune system can further diminish the leukemia without TKI support. The corresponding immunologic landscape is divided into these two basins of attraction and is visualized in Fig. 5C. We estimate 8 of 21 patients in this class, which all maintain TFR (example in Fig. 5D).

• Class C: The third class has the same stable steady states as class B, but in this case a small disturbance from the cure steady state |{E_0}$| leads to the attraction basin of the recurrence steady state |{E_H}$| instead of the remission steady state |{E_L}$|⁠. Only for a small range of CML abundance and a sufficiently high level of immune cells, the immune system is appropriately activated to keep the leukemia under sustained control. Figure 5E and F illustrates this control region as an isolated attractor basin. In the ideal case, the TKI therapy only reduces the leukemic load to a level that is sufficient to still activate the immune system to achieve this balance. However, if TKI treatment reduces tumor load to a very deep level, the CML cells regrow after therapy cessation as the immune response was also reduced too much. This represents a patient with CML that may potentially achieve TFR but has a weak immune response. We estimate that 7 of 21 patients fall in this class, of which four have a recurrence and three remain in TFR (two examples in Fig. 5G and H).

For completeness, there is a fourth class, in which only a cure steady state exists. In this case, CML would not develop at all due to a strongly suppressive immune system. Naturally, those individuals do not appear in the patient cohort at all.

We showed that the immunologic configuration of each patient determines which steady states can be reached using TKI treatment. It should be pointed out that the resulting conclusion does not depend only on the model fits to the data, but also on the particular mechanisms assumed by the model structure. Within those restrictions, it appears that patients in class A can only stop TKI treatment in the case that the disease is completely eradicated. This would require a median treatment time of 29 years in our simulations and was not achieved in any of the considered patients. However, even if treatment cessation is not an option for these patients, our previously published results suggests that TKI dose reduction could be considered as a long-term treatment alternative (9).

From a perspective of treatment optimization, patients in classes B and C are most interesting as they present an immune window, in which a TKI-based reduction of the leukemic cells can sufficiently stimulate the expansion of the immune cell population (Supplementary Fig. S7; Supplementary Materials). Our model suggests that patients in class B are characterized by an immunologic response that is sufficient to control the leukemia once the leukemic load has initially been reduced below a certain threshold. This remission allows for an activation of the immune system to further control the leukemia eradication even in the absence of TKI treatment. It is essential that the initial remission and the immunologic activation surpass a certain threshold, which is indicated by the line separating the different basins of attraction in Figs. 4B and 5C (separatrix). Clinically, this can be achieved by a sufficiently long TKI therapy, although we predict that this necessary time span was already reached much earlier for the respective six patients, in comparison with their actual treatment times (Supplementary Table S2).

In contrast to class B, the model analysis implies that patients in class C can also present with recurrence if a long TKI treatment is applied. Only in a narrow region of CML abundance the immune system is sufficiently stimulated: if leukemic load is too high, the immunologic component is still suppressed, while for too low levels the stimulation is not strong enough. In this respect, TFR can only be achieved if treatment keeps the patient within his individualized immune window for a sufficient time thereby supporting the adequate proliferation of immune cells, such that the patient reaches the basin of attraction of the remission steady state |{E_L}$| (Fig. 6A and B). If the treatment intensity is too high or the treatment duration is too long, this might lead to an “overtreatment” where the inherent immunologic defense is not quickly and sufficiently activated to control a recurrence once TKI is stopped (Fig. 6C–H). We show with a hypothetical treatment protocol that an adjustment of the necessary balance between leukemia abundance and immunologic activation can be achieved within this model by detailed assessment of both cell populations and a narrowly adapted TKI administration (Supplementary Figs. S8 and S9).

Detailed information about a patient's response to TKI treatment cessation (according to fitting strategy III) can only be obtained if the complete data (including postcessation measurements) is available. Thus, this approach can obviously not serve as a prediction strategy before therapy stop. However, we show in the following that response to dose reduction, prior to therapy stop, will also provide information to identify the patient specific immunologic landscape and is, therefore, likely to provide important information about the disease dynamics after treatment cessation. Both, clinical and modeling evidence support the strategy to use information from intermediate dose reduction as this appears as a safe treatment option for almost all well-responding patients with CML (9, 43).

Specifically, individual fits for all patients according to the immune model and fitting strategy III allow to mathematically simulate how the patients would have responded if they were treated with a reduced TKI dose instead of stopping TKI completely. We use these model simulations to derive information about the predicted BCR-ABL1 ratio during a 50% dose reduction within a 12-month period. Figure 7A and B illustrates two typical time courses.

Quantitatively, we estimate the linear slopes of the individual BCR-ABL1/ABL1 response during dose reduction and correlate it to the final remission status after treatment cessation (Fig. 7C). A logistic regression analysis reveals that a 0.01 increase in the estimated slope increases the chance of recurrence by 21% (OR: 1.21; 95% CI: 1.07–1.51), thereby indicating that recurring patients are predicted to present with higher (positive) slopes of the BCR-ABL1 ratio during the dose reduction period. Moreover, complementing this plot with the association of each patient with its predicted particular response class A, B, or C, we observe that class A patients have higher positive slopes and always have a recurrence, while most of class B patients show constant BCR-ABL1 levels, therefore, staying in TFR. Class C patients show both, constant or increasing BCR-ABL1 levels. However, higher positive slopes are more often observed in recurring patients. We suggest that patients with pronounced increases in BCR-ABL1 levels after dose reduction should not stop TKI treatment as this increase points toward an insufficient immune control and conveys an increased risk for molecular recurrence.

Our results are in qualitative and quantitative agreement with a recent reanalysis of clinical data from the DESTINY trial (NCT01804985; refs. 43, 44), which differs from other TKI stop studies as in this trial the TKI treatment is reduced to 50% of the standard dose for 12 months prior to cessation. On the basis of a dataset of 171 patients, we could demonstrate that the patient-individual slope of BCR-ABL1/ABL1 ratios monitored during TKI dose reduction strongly correlates with the risk of individual recurrence after TKI stop (OR: 1.28; 95% CI: 1.17–1.42) and can serve as a promising indicator for high-risk patients (40). Although time courses prior to dose reduction are not available from this study and preclude fitting of the complete ODE model, the overall conclusion of both, the presented conceptual approach and a paralleling data analysis, suggest that dose alterations are a valid means to probe the immunologic configuration of leukemic remission.

Here we present an ODE model for CML treatment that explicitly includes an immunologic component and apply it to describe the therapy response and recurrence behavior of a cohort of 21 patients with CML with detailed BCR-ABL1 follow-up over their whole patient history. We demonstrate that an antileukemic immunologic mechanism is necessary to account for a TKI-independent disease control, which prevents molecular recurrence emerging from residual leukemic cell levels after TKI cessation. Without such a mechanism, a long-term TFR can only be achieved if a complete eradication of leukemic cells is assumed. However, the presence of detectable MRD levels in many patients after therapy cessation (14) is not consistent with this assumption, which strongly suggests an additional control instance, which others (20, 21, 45, 46) and we (39) interpret as a set of immunologic factors. Including these aspects into our modeling approach, the available clinical data can be sufficiently described on the level of individual patients.

On the basis of our simulation results we classify patients into three different groups regarding their predicted immune system configuration (“immunologic landscape”): insufficient immune response (class A), strong immune response (class B), and weak immune response (class C). Class A patients are not able to control residual leukemic cells and would always present with CML recurrence as long as the disease is not completely eradicated. Consistent with the results of Horn and colleagues (47), this is only accomplished on very long timescales in our simulations and would, therefore, result in a lifelong therapy for most affected patients. However, as we suggested earlier, those patients might be eligible for substantial TKI dose reductions during long-term maintenance therapy (9). In contrast, class B patients are predicted to have a strong immune response and to control the leukemia once the leukemic load has been reduced below a certain threshold and thus, are predicted to require only a minimal treatment time (less than 5 years for the studied patients, see Supplementary Table S2) to achieve TFR. For class C patients with a weak immune response, our model predicts that TFR achievement depends on an optimal balance between leukemia abundance and immunologic activation before treatment cessation and could be accomplished by a narrowly adapted TKI administration. These results are in line with those from a recent modeling study that suggested the existence a “Goldilocks Window” in which treatment is required to optimize the balance between maximal tumor reduction and preservation of patient immune function (26).

We also show that the information required to classify the patients according to their immune response and to predict their recurrence behavior cannot be obtained from BCR-ABL1 measurements before treatment cessation only. A different fitting strategy (III) assessing also BCR-ABL1 measurements after treatment cessation shows that the BCR-ABL1 changes resulting from this system perturbation (i.e., TKI stop) yields the necessary information. Interestingly, our simulation results demonstrate that also a less drastic system perturbation, that is, a TKI dose reduction, can provide similar information and can be used to predict the individual outcome after treatment cessation. The feasibility of such an approach has been complemented by a recent reanalysis of the DESTINY trial (NCT01804985), which evaluated a beneficial effect of a 12-month dose reduction treatment prior to TKI stop. We could confirm based on the clinical data of 171 patients that the patient's response dynamic during TKI dose reduction is indeed predictive for the individual risk of CML recurrence after TKI stop (40, 43).

Direct measurements of the individual immune compartments and their activation states represent another road to better understand the configuration of the antileukemic immune response in patients with CML. Several studies identified different immunologic markers in patients with CML that correlate with the probability of treatment-free remission after therapy cessation (20, 21). Learning from the behavior of these populations under continuing TKI treatment and with lowered leukemic load could further contribute to identify a patient's “immunologic landscape” and be informative for the prediction of individual outcomes after treatment stopping. However, as it is not clear which immunologic subset provides the suggested observed anti-CML response (48, 49), corresponding measurements are currently not feasible and strongly argue in favor of our indirect modeling approach, suggesting retrieval of similar information from BCR-ABL1 dynamics after TKI dose reduction.

Our analysis is based on a rather small cohort of patients. Although our results do not depend on the study size, we can derive the strongest conclusions with respect to illustrating the conceptual approach of inferring immune responses from treatment alterations and demonstrating its predictive power. Our results are further based on a set of simplifications and assumptions. As such, we do not consider resistance mutations as almost no such events have been reported during TKI cessation studies in CML and almost all patients respond well to reinitiation of TKI treatment with their previous drug (11). This might be different for other disease entities in which tumor evolution imposes serious challenges to long-term disease control. Focusing on a related aspect it has been shown that different immune cell types are associated with recurrence behavior of patients with CML (20, 21). However, for simplicity, we restricted our analysis to a unified antileukemic immune compartment in the model and did not distinguish between different immune cell populations and interactions between them. Furthermore, the model is based on an interaction between leukemic and immune cells, in which the immune cell population is only activated for intermediate levels of leukemic burden, reflecting the assumption that immune cells are not efficiently activated for small numbers of leukemic cells and are additionally suppressed by high tumor load. Similar assumptions have been discussed recently (37) while we also illustrated the suitability of other mechanisms of interaction (39).

In summary, our results support the notion of immunologic mechanisms as an important factor to determine the success of TFR in patients with CML. Importantly, we show that besides the direct measurement of the immune response, also system perturbations, such as a TKI dose reduction, can (indirectly) provide information about the individual disease dynamics and, therefore, allow to predict the risk of CML recurrence for individual patients after TKI stop. Such results demonstrate the potential of mathematical models in providing insights on the mechanisms underlying cancer treatment as well in delineating different treatment strategies. Applications to other cancer entities, in which the endogenous immune system can support the control or even the eradication of residual tumor cells, are a natural continuation of this work and will become even more important with the availability of cancer immunotherapies that allow modulation of individual immune responses (50).

S. Mustjoki reports receiving a commercial research grant from Pfizer, Novartis, and Bristol Myers Squibb and has received speakers bureau honoraria from Bristol Myers Squibb, Incyte, and Novartis. P.J. Jost reports receiving a commercial research grant from Böhringer and Abbvie, has received speakers bureau honoraria from Novartis, BMS/Celgene, Abbvie, Böhringer, Servier, Pfizer, and has done expert testimony for WEHI. F.-X. Mahon reports receiving speakers bureau honoraria from Novartis and is a consultant/advisory board member for Novartis. I. Roeder reports receiving a commercial research grant from Bristol-Myers Squibb and has received speakers bureau honoraria from Bristol-Myers Squibb and Janssen-Cilag. I. Glauche reports receiving a commercial research grant from Bristol-Myers Squibb. No potential conflicts of interest were disclosed by the other authors.

Conception and design: T. Hähnel, C. Baldow, I. Roeder, A.C. Fassoni, I. Glauche

Development of methodology: T. Hähnel, C. Baldow, I. Roeder, A.C. Fassoni, I. Glauche

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Guilhot, F. Guilhot, S. Saussele, S. Mustjoki, S. Jilg, P.J. Jost, S. Dulucq, F.-X. Mahon

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Hähnel, C. Baldow, I. Roeder, A.C. Fassoni, I. Glauche

Writing, review, and/or revision of the manuscript: T. Hähnel, C. Baldow, J. Guilhot, S. Saussele, S. Mustjoki, S. Dulucq, F.-X. Mahon, I. Roeder, A.C. Fassoni, I. Glauche

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S. Mustjoki, I. Glauche

Study supervision: C. Baldow, I. Roeder, A.C. Fassoni, I. Glauche

We thank all patients and hospital staff for providing this valuable data for scientific assessment. This work was supported by the German Federal Ministry of Education and Research (www.bmbf.de/en/), grant number 031A424 “HaematoOpt” (to I. Roeder) and grant number 031A315 “MessAge” (to I. Glauche), as well as the ERA-Net ERACoSysMed JTC-2 project “prediCt” (project number 031L0136A to I. Roeder). The research of A.C. Fassoni was supported by the Excellence Initiative of the German Federal and State Governments (Dresden Junior Fellowship) and by CAPES/Pós-Doutorado no Exterior Grant number 88881.119037/2016-01. S. Mustjoki was supported by Finnish Cancer Organizations, Sigrid Juselius Foundation, and Gyllenberg Foundation.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.