Early dynamics of chronic myeloid leukemia on nilotinib predicts deep molecular response

Abstract

Chronic myeloid leukemia (CML) is a myeloproliferative disorder caused by the BCR-ABL1 tyrosine kinase. Although ABL1-specific tyrosine kinase inhibitors (TKIs) including nilotinib have dramatically improved the prognosis of patients with CML, the TKI efficacy depends on the individual patient. In this work, we found that the patients with different nilotinib responses can be classified by using the estimated parameters of our simple dynamical model with two common laboratory findings. Furthermore, our proposed method identified patients who failed to achieve a treatment goal with high fidelity according to the data collected only at three initial time points during nilotinib therapy. Since our model relies on the general properties of TKI response, our framework would be applicable to CML patients who receive frontline nilotinib or other TKIs.

Fig. 1. Several criteria of the current scoring systems and guidelines are not suitable for DMR classification.

Each dot indicates each patient value within the N-road dataset. The top and bottom bars of the boxes indicate the maximum and minimum values, respectively. The top, center, and bottom lines are the upper quartiles, medians, and lower quartiles, respectively. Patients are divided into two groups: a DMR patient set and a non-DMR patient set. a The EUTOS score did not predict non-DMR patients. The patient is considered “High risk” if the EUTOS score of a CML patient is >87 and “Low risk” otherwise. b The ELTS score did not also predict non-DMR patients. A CML patient is considered as “High risk” if the patient’s ELTS score is >2.2185, “Low risk” if the score is 1.5680 or lower, and “Intermediate risk” otherwise. c–e According to the ELN guideline, we plotted patient IS and separated them into three groups: “Optimal”, “Warning”, and “Failure” for the data at (c) 3, (d) 6, and (e) 12 months after the initiation of TKI administration. In c, d many non-DMR patients were predicted as “Optimal”, indicating the failure of these predictions.

Fig. 2. Key features for DMR prediction: the decreasing rate and convergence value of IS, and the reduction rate and convergence value of CML cells.

a, b The time series of IS are plotted in (a) for a non-DMR patient (patient # 19) and in (b) for a DMR patient (patient # 9). The decreasing rate and convergence value of IS are two key features for DMR prediction. A detailed information of the parameter estimation can be found in Methods. c The distributions of the IS decreasing rate for non-DMR and that for DMR patients were different. d The distributions of the IS convergence value for non-DMR and that for DMR patients were distinct. The definitions of the bars of the boxes, lines, and dots used in the boxplots (c, d) are the same as those in Fig. 1. e, f We show the estimated dynamics of normal WBCs and CML cells counts in (e) for a non-DMR patient (patient # 19) and in (f) for a DMR patient (patient # 9). These figures indicate that the reduction rate and convergence value of CML cells play a key role in the distinction between DMR and non-DMR patients.

Fig. 3. DMR and non-DMR patients were separated on the plane of the reduction rate and convergence value of CML cells.

a The estimated time series of CML dynamics (red dashed lines) sufficiently approximated the measured data for IS [%] and for WBC counts [number / μl] (black solid lines). The patients shown in the colored boxes were chosen as examples from the patients within the same color regions shown in (c). b The distribution of the convergence value and recovery rate [1/month] of normal WBCs did not show a correlation in terms of nilotinib response. The number indicated on each dot is patient #. c The distribution of the convergence value and reduction rate of CML cells was apparently classified into several subsets: a set including most DMR patients (the blue region), non-DMR patients (the orange region), and outlier patients (other color regions). d By choosing optimal threshold values, the distribution of DMR patients shown in (c) was perfectly separated from that of non-DMR patients. Notably, this result was obtained by using all patient data points.

Fig. 4. Our prediction method achieved favorable performance in comparison with the criteria of the current scoring systems and guidelines.

We summarize the performance for accuracy (blue bars), sensitivity (green bars), specificity (orange bars), and F1 scores (purple bars) of the EUTOS score, the ELTS score, the ELN guideline for the cases of 3, 6, 12 months, and our proposed method (presented with the three rightmost labels). In the current study, “positive” and “negative” indicate DMR and non-DMR patients, respectively. In the cases of the EUTOS scores, the ELTS scores, and the ELN guidelines, we decided to estimate a target patient as “positive”, if the patient was classified as “Low risk” or “Optimal”. Under this estimation, we obtained accuracy, sensitivity (the true positive rate), and specificity (the true negative rate) for each case. In our setting, the prediction of non-DMR patients is highly important, because the administrated TKI should be promptly changed for them. Thus, our method was superior to the EUTOS score and the ELN guideline because our method showed higher specificity than them. The ELTS score cannot distinguish between non-DMR and DMR patients. Notably, each of our three results was optimized in terms of accuracy, sensitivity, and specificity through the training period. See also “DMR prediction criteria” and “Performance of DMR prediction” in Methods.