The ecology in the hematopoietic stem cell niche determines the clinical outcome in chronic myeloid leukemia

Abstract

Chronic myeloid leukemia (CML) is a blood disease that disrupts normal function of the hematopoietic system. Despite the great progress made in terms of molecular therapies for CML, there remain large gaps in our understanding. By comparing mathematical models that describe CML progression and etiology we sought to identify those models that provide the best description of disease dynamics and their underlying mechanisms. Data for two clinical outcomes--disease remission or relapse--are considered, and we investigate these using Bayesian inference techniques throughout. We find that it is not possible to choose between the models based on fits to the data alone; however, by studying model predictions we can discard models that fail to take niche effects into account. More detailed analysis of the remaining models reveals mechanistic differences: for one model, leukemia stem cell dynamics determine the disease outcome; and for the other model disease progression is determined at the stage of progenitor cells, in particular by differences in progenitor death rates. This analysis also reveals distinct transient dynamics that will be experimentally accessible, but are currently at the limits of what is possible to measure. To resolve these differences we need to be able to probe the hematopoietic stem cell niche directly. Our analysis highlights the importance of further mapping of the bone marrow hematopoietic niche microenvironment as the "ecological" interactions between cells in this niche appear to be intricately linked to disease outcome.

Keywords: cancer progression; competition; model selection; niche dynamics.

Fig. 1.

Mechanistic models for hematopoiesis in CML and experimental measurements of leukemia progression in patients. (A) Graphical description of models M1, M2, and M3. The hematopoietic niche is represented by a red rectangle: every species inside the rectangle influences the niche, whereas only species marked by red inhibitory arrows are affected by it. Cell death is illustrated by light gray arrows. (B) The change in the BCR-ABL level over time (in months) is represented by gray lines for each of the 69 patients under CML treatment. The average change over all patients is illustrated by a dashed green line; the change in the BCR-ABL level for a patient who exhibits disease relapse is illustrated by a dashed red line. Our model comparison analysis is based on these two trajectories. (C) Description of how niche competition is treated by each model. Each species can impact the niche (effector), be affected by the niche (niche reliant), and/or be fully unconstrained by the niche.

Fig. 2.

Model comparison based on the fit to the data and the importance of niche competition. (A) Fits to the data (yellow diamonds) for the three models and the two outcomes. All three models are able to explain the observed disease progression for both outcomes. (B) Predicted trajectories for stem cell species for each model. The lines represent the median of the evolution and the error bars designate the 5% and 95% percentiles for a set of 1,000 parameters sampled from the posterior distribution. For model M2, LSCs exhibit unbounded exponential growth for all sampled parameter values. (C) Models are ranked by the probability of each to explain the observed datasets. For each model and each outcome, the logarithm of the model probability is computed using a SMC sampler; the average over 10 independent runs is displayed and the number in brackets is the SD.

Fig. 3.

The predicted change over time of each species for a timespan of 6 y (the total period of the trial), for (A and C) model M1 and (B and D) model M3 and (A and B) remission and (C and D) relapse. For each model and each outcome, 1,000 parameter sets were sampled from the posterior distribution and each line corresponds to the simulated trajectory for one of these parameter sets.

Fig. 4.

Linear discriminant analysis was used to predict the key differences between remission and relapse. For (Left) model M1 and (Right) model M3, clear separation between disease outcomes can be made (Upper). The parameter contributions (coefficients) to the linear discriminant are shown on a horizontal axis. For model M1, the healthy parameters are represented by blue dots and the leukemic parameters by orange dots; for model M3, the pink dots designate the death rates. (Lower) Shown for model M1 is a density plot of against with the location of remission outcomes marked in blue and the relapse outcomes in red: using only these two parameters a good separation is already achieved. We also show the posterior density distribution of parameter alone in case of remission (blue) and relapse (red)—moderate separability is observed. For M3 we show the posterior density distribution for each of the death rates and we see that for , remission (blue) and relapse (red) can be well distinguished.