Introduction

Chronic myeloid leukemia (CML) is a myeloproliferative stem cell disease characterized by the fusion oncoprotein BCR-ABL1 [1, 2]. Tyrosine kinase inhibitors (TKIs) that selectively target BCR-ABL1, such as Imatinib Mesylate (IM), and second generation TKIs Dasatinib (DA) and Nilotinib (NL), have demonstrated remarkable clinical efficacy in the treatment of chronic phase (CP) CML [3,4,5,6,7,8,9]. For example, a 10-year follow up of the ENESTnd clinical trial evaluating long-term outcomes of NL and IM, demonstrated that NL had increased rates of treatment-free remission (TFR) eligibility and overall survival compared to IM treatment [10, 11]. The extension of the ENESTnd clinical trial, ENESTxtnd, further supported the use of NL as a front-line therapy, with 81% of CP-CML patients achieving major molecular remission (MMR) by 24 months [12]. These advances in TKI therapy have improved cancer-free survival rates and resulted in faster and deeper levels of molecular response in these patients. However, TKI treatments are not usually curative as some patients develop drug-resistance or are at risk of disease relapse when treatment is discontinued [13,14,15,16,17]. Thus, it has been of interest to develop predictive biomarkers to accurately predict whether patients may achieve a deep molecular response with first-line TKI, require a switch to second generation TKIs or can discontinue treatment safely. Patient material from a variety of sources has been profiled using microarrays, RNA-sequencing and, more recently, single cell RNA-sequencing [18,19,20,21]. Some studies have identified parameters that may indicate poor TKI response in CML patients, such as the detection of additional chromosomal abnormalities in BCR-ABL1+ cells at diagnosis or observing a slow rate of change in BCR-ABL1 transcript levels in the first 3 months of TKI therapy, among others [22,23,24]. It was also reported that IM response could be predicted in treatment-naïve CD34+ stem/progenitor cells by an in vitro CFC assay [25]. Moreover, analysis of differential gene expression between IM-nonresponders and responders has offered predictive potential: a 17-gene panel quantified via qRT-PCR to classify patients as high-risk or low-risk was able to predict early molecular response failure, based on BCR-ABL1 transcript levels at 3 months in CP-CML patients treated with IM or NL [26]. However, differences in the cell types profiled, the technologies used, and the inherent complexity associated with the interpretation of molecular data pose challenges in identifying reliable biomarkers with clinical utility.

MicroRNAs (miRNA) are single-stranded non-coding RNAs of about 18–25 nucleotide length that are aberrantly expressed in various malignancies and regulate numerous biological processes including cell differentiation and survival [27,28,29,30]. Some miRNA tumor suppressors regulating BCR-ABL1 were reported to be downregulated in CML cells whereas oncogenic miRNAs (oncomiRs) modulated by BCR-ABL1 were found to be overexpressed [31,32,33]. Additionally, BCR-ABL1-dependent and -independent miRNA expression profiles were observed in CML patient cells at different stages of the disease [32, 34]. Interestingly, the expression of some of these miRNAs was reported to change in response to TKI-treatment and may therefore be good indicators of TKI-resistance [34,35,36,37,38]. Recently, we generated global transcriptome profiles on treatment-naïve CD34+ CML cells with known subsequent IM responses and identified several differentially expressed miRNAs, including miR-185 and miR-145, in IM-nonresponders as compared to IM-responders [39].

In this study, we aimed to generate an effective, combinational predictive model consisting of miRNAs and in vitro TKI CFC data by analyzing miRNA expression profiles in CD34+ cells from 58 patients (retrospectively classified as NL-responders or NL-nonresponders) and the sensitivity of these cells to TKIs. Through Cox Proportional Hazard (CoxPH) analysis and machine learning approaches, we identified and evaluated two predictive miRNA panels in treatment-naïve or post-NL treatment in CML patients. Interestingly, incorporation of in vitro CFC data into either panel improved the predictive power at each time point. To our knowledge, this is the first study to utilize miRNA expression profiles, CFC data, and clinical response data from matched patient samples to develop a predictive model of response to NL treatment.

Methods

Study design and treatment

Heparin-anticoagulated peripheral blood (PB) samples were obtained from 62 newly diagnosed CML patients enrolled in the Canadian sub-analysis arm of the ENESTxtnd phase IIIb clinical trial (https://clinicaltrials.gov/ct2/show/NCT01254188, Supplementary Table 1). Four patients were excluded due to insufficient enrichment of CD34+ cells (<65%), leaving a total of 58 patient samples. These patients were newly diagnosed (within 6 months) CP-CML patients who had not been previously treated with IM. PB samples were obtained prior to therapy (BL), 1 month (M1) and 3 months (M3) after 300 mg or 400 mg BID nilotinib (NL) therapy [11, 40]. Clinical responses of patients were classified as either NL-responders (R) or NL-nonresponders (NR) at 12 months (46 R, 12 NR) after treatment based on BCR-ABL1 transcript levels following the European Leukemia Net treatment guidelines [41]. Informed consent was obtained in accordance with the Declaration of Helsinki, and all procedures performed were approved by the University of British Columbia Clinical Research Ethics Board (certificate H20-00421).

Colony-forming cell (CFC) assay

CD34+ cells (>90%) were enriched immunomagnetically using EasySep CD34 selection kits (STEMCELL Technologies) and analyzed using a fluorescence-activated cell sorter (FACS). CFC assays were performed as previously described and in a double-blinded approach, with results subsequently linked to each patient’s clinical response to NL [25]. Briefly, 3000 primary CD34+ cells were mixed with MethoCult™ H4230 (STEMCELL Technologies) and growth factor cocktail in the presence or absence of inhibitors: 10 µM NL (Novartis), 150 nM DA (Bristol-Myers Squibb), 5 µM IM (Novartis). Colonies were counted after 14 days.

Microfluidic quantitative PCR

Microfluidic quantitative PCR (qPCR) was conducted as previously described [39]. Briefly, pooled RT primer mix containing 47 distinct miRNA stem-loop and RNU48 primers were prepared and mixed with total RNA extracted from CD34+ CML cells. Amplified cDNA from these reactions was mixed with Taqman master mix and loaded to sample inlets of a Fluidigm 48 ×48 Dynamic array device. High-throughput quantitative real-time PCR was then performed using the BioMark HD system (Fluidigm) to produce image-based fluorescent signals and raw Ct values. The raw Ct values were normalized to the RNU48 endogenous control (ΔCt method).

Statistical analysis

For CFC data, a Student’s two-tailed t-test was used to test for differential colony output between responders and nonresponders. The HTqPCR R package was used to perform data quality control and normalization. Normalization to the endogenous RNU48 control achieved the closest to normal distribution (Supplementary Fig. 1) and was therefore selected as the normalization method in this study. Two samples were excluded as they showed unusual variation in RNU48 Ct values that were >1.5-fold above the interquartile range. Univariate Cox Proportional Hazard (CoxPH) analyses were performed using individual normalized miRNA expression levels to determine the hazard-ratio (HR) of each miRNA. As CoxPH analysis performs best on categorical data, four different cut-offs were applied: median, 1st quartile, 3rd quartile and maximally selected rank statistics [42]. While the maximally selected rank statistics calculated cut-off from the survival (v3.3-1) R package was the method we later adopted in our modeling, only miRNAs that were significantly associated with NL nonresponse for at least two of the four applied cut-offs in the univariate CoxPH analysis were selected for further analysis. Overall, only miRNAs that fulfilled the following criteria were selected for the subsequent multivariate analysis step: significant association with NL nonresponse (Welch t-test), differential expression between NL-responder and NL-nonresponder samples at BL, and unidirectional association with NL nonresponse over time (either increased or decreased hazard at both M1 and M3). Predictive performance of the remaining candidates was evaluated in multivariate CoxPH, random forest (RF) and Naïve-Bayes (NB) machine learning algorithms (R caret v6.0-91 and MLeval v0.3). All models were tenfold cross-validated and predictive performance was assessed based on area-under-curve (AUC) values from receiver operating characteristic (ROC) and precision recall (PR) curves. The R scripts used to generate the presented data are available upon request.

Results

Treatment-naïve CD34+ CML cells from NL-responders and NL-nonresponders generate significantly different CFC outputs during in vitro Nilotinib treatment

We previously demonstrated that CFC assays in IM-supplemented cultures can classify CML patients into IM-responders and IM-nonresponders, based on the colony output from treatment-naïve CD34+ stem/progenitor cells obtained at diagnosis [25]. To determine whether a similar method can be applied to NL, we performed in vitro CFC assays with and without TKI treatments on CD34+ cells from 58 newly diagnosed CML patients enrolled in the ENESTxntd trial. NL-responders (R) had 3-log-fold or greater reduction in BCR-ABL1 transcripts (<0.1%) at month 12 of NL therapy, while those that failed to achieve this response were classified as NL-nonresponders (NR, Supplementary Table 1). As expected, we observed an overall reduction in CFC outputs generated from CD34+ cells upon NL, DA or IM treatment compared to untreated controls. Furthermore, NL-response in patient CFC data correlated with DA-CFC and IM-CFC counts from the same patient sample (DA: Pearson’s R = 0.06831, p < 0.0001; IM: R = 0.7243, p < 0.0001; Fig. 1A). However, while most retrospectively classified responders produced overall fewer colonies upon any TKI treatment, we observed a more diverse pattern for nonresponders, clustering into two response fractions in CFC assays. Some of the nonresponder samples responded poorly and produced high numbers of colonies upon NL or DA or IM treatment, but a few samples produced similar numbers of colonies as CD34+ cells from responders (Fig. 1B). Nevertheless, a significant difference in CFC output was observed between responder and nonresponder samples for NL (p = 0.014) but not for IM or DA (Fig. 1B). To determine if CFC output was significantly associated with nonresponse, each patient was categorized as CFC low or high based on maximally selected rank statistics calculated cut-offs (NL = 46.2%, DA = 26.9%, IM = 34.3%). Univariate CoxPH analysis showed that CFC outputs from all three TKIs were significant classifiers in predicting NL clinical response status (p ≤ 0.019, Table 1) and displayed hazard-ratios (HR) < 1, associating higher colony output with NL nonresponse.

Fig. 1: Comparison of the effect of TKIs in vitro on CD34+ cells from NL-responders and nonresponders.
figure 1

A Correlations of CFC output between NL and IM or DA. B Differential CFC output between NL-NR and NL-R patient samples after treatment with NL, DA, and IM. Dashed line represents the calculated cut-offs determined by maximally selected rank statistics for CoxPH analysis. P values for comparing NL-NR and NL-R groups were calculated using a two-tailed unpaired Student t-test. Data shown are mean ± standard deviation (SD).

Full size image
Table 1 Cox proportional hazard analysis of CFC data.
Full size table

Expression of miR-145 and miR-708 in treatment-naïve CML cells predicts patient response and accuracy is further improved by incorporation of in vitro NL-CFC data

We recently published a miRNA-sequencing study highlighting 47 differentially expressed miRNAs identified in CD34+ stem/progenitor cells between normal bone marrow (NBM) and CML patient samples, particularly between IM-responders and IM-nonresponders [39]. In the present study, we collected microfluidic qRT-PCR expression data for this panel of 47 miRNAs in CD34+ cells obtained at diagnosis (BL), 1-month (M1) and 3-month (M3) post-NL treatment from 58 CML patients. Interestingly, univariate CoxPH analysis identified that expression of 17/47 miRNAs was significantly associated with NL nonresponse (four miRNAs at BL, nine miRNAs at M1, and ten miRNAs at M3; p < 0.05, Fig. 2A and Supplementary Table 2). Some of these miRNAs were significantly associated with nonresponse across all the three time points and overall, 9/47 miRNAs were differentially expressed between NL-responder and NL-nonresponder patients at any time point (BL, M1, M3; p ≤ 0.047, Fig. 2B).

Fig. 2: Study design and univariate analysis of 47 differentially expressed miRNAs between NL-responders and nonresponders.
figure 2

A Workflow of statistical and analytical processes. B Univariate analysis of differentially expressed miRNAs associated with NL-nonresponse across pre- and post-treatment time points as indicated. Box-plots of transcript levels are displayed relative to NL-NR at the treatment-naïve state (BL) on a log scale. P values were calculated using Welch’s t-test.

Full size image

To evaluate their predictive capacity, we next combined miRNAs that demonstrated significant association with NL nonresponse and fulfilled all additional selection criteria (see Materials and Methods) into multivariate models for each individual time point of sample collection. At the treatment-naïve state (BL), increased miR-145 expression (HR 6.95, p = 0.013) and decreased miR-708 expression (HR 0.13, p = 0.009) were associated with NL-nonresponse in multivariate CoxPH modeling (Fig. 3A). ROC-AUC values generated by trained random forest (RF)- and Naïve-Bayes (NB)- machine learning models also indicated predictive performance of miR-145 and miR-708 (ROC-AUC: RF = 0.58, NB = 0.67; Fig. 3C and Supplementary Fig. 2A). We next tested if the model accuracy could be further improved through incorporation of the TKI-CFC response data (NL, IM, and DA). Indeed, multivariate CoxPH models incorporating CFC output (HR 0.15, p = 0.003) but not IM or DA showed superior Concordance in predicting NL nonresponse compared to models based on miRNA expression alone (Ci = 0.8 vs. Ci = 0.89) (Fig. 3B). Most interestingly, results from our trained RF and NB machine learning models showed that inclusion of NL-CFC increased the ROC-AUC and PR-AUC values by 1.2-fold (RF) and 1.9-fold (NB) (Fig. 3D and Supplementary Fig. 2B). Of note, no other clinical factors except for NL-CFC, including WBC count and BCR-ABL1 transcript levels at BL and 12 months, were identified as important features in machine learning or associated with patient response (Supplementary Table 3). Thus, we could identify the combination of miR-145 and miR-708 expression as a predictive indicator of NL response at the treatment-naïve state and we have observed that inclusion of NL-CFC data further enhanced predictive performance.

Fig. 3: Multivariate analysis of miRNAs associated with NL-nonresponse in treatment naïve patients (BL).
figure 3

MiR-145 and miR-708 are associated with and can predict NL-nonresponse as determined by multivariate CoxPH analysis (A) and ROC plots from trained machine learning models (C). Inclusion of NL-CFC data (B) enhances predictive performance based on ROC-AUC (C, D). N represents the number of patient samples that were classified as high or low in miRNA level or CFC output using the calculated cut-offs. The hazard ratios for each variable in the multivariate CoxPH analyses are displayed with 95% confidence and the corresponding p value. ROC plots for each machine learning algorithm, RF and NB, are shown with their corresponding AUC values.

Full size image

MiR-150 and miR-185 expression levels in post-NL CML samples predict treatment response and accuracy is further improved by incorporation of in vitro NL-CFC data

Since our model performed well in treatment-naïve patient cells, we next tested if a similar predictive panel could be generated at the M1 and M3 post-NL treatment. From the nine miRNAs that showed significant association with nonresponse in univariate CoxPH analyses, four miRNAs (miR-145, miR-365, miR-150, miR-185) displayed HR values that remained consistent between M1 and M3 and were therefore selected for further analysis. Overall, combination of miR-150 and miR-185 expression levels into multivariate CoxPH and machine learning models, achieved the best performance (ROC-AUC: RF = 0.76, NB = 0.73; Ci = 0.8; Fig. 4A, C and Supplementary Fig. 2C). Addition of NL-CFC to the M1 model improved Concordance (Ci = 0.88) and overall accuracy (ROC-AUC RF = 0.84, NB = 0.88, Fig. 4B, C and Supplementary Fig. 2C), overall by up to 1.3-fold (Fig. 4D and Supplementary Fig. 2D). Applying the same panel to data collected at M3 validated the predictive potential of miR-150 and miR-185 in multivariate CoxPH analysis (Fig. 5A) but underperformed in the machine learning evaluation with lower-than-expected ROC-AUC values (RF = 0.52, NB = 0.51; Ci = 0.74) (Fig. 5A, C and Supplementary Fig. 2E). However, predictive performance was again greatly improved when NL-CFC was included to the M3 model, enhancing Concordance (Ci = 0.84) and increasing ROC-AUC values by 2-fold (RF = 0.72, NB = 0.68; Fig. 5C, D and Supplementary Fig. 2E, F). After NL treatment, both at M1 and M3, expression levels of miR-150 and miR-185 were identified as potential classifiers of treatment response and incorporation of NL-CFC data yielded significantly improved predictive accuracy, particularly after 1 month of treatment.

Fig. 4: Multivariate analysis of miRNAs associated with NL-nonresponse in 1-Month post-treatment patients (M1).
figure 4

MiR-150 and miR-185 are associated with and can predict NL-nonresponse, as determined by multivariate CoxPH analysis (A) and ROC-plots from trained machine learning models (C). Inclusion of NL-CFC data (B) enhances predictive performance based on improved ROC-AUC (C, D). N represents the number of patient samples that were classified as high or low in miRNA level or CFC output using the calculated cut-offs. The hazard ratios for each variable in the multivariate CoxPH analyses are displayed with 95% confidence and the corresponding p value. ROC plots for each machine learning algorithm, RF and NB, are shown with their corresponding AUC values.

Full size image
Fig. 5: Multivariate analysis of miRNAs associated with NL-nonresponse in 3-month post-treatment patients (M3).
figure 5

MiR-150 and miR-185 are associated with and can predict NL-nonresponse, as determined by multivariate CoxPH analysis (A) and ROC plots from trained machine learning models (C). Inclusion of NL-CFC data (B) enhances predictive performance based on improved ROC-AUC (C, D). N represents the number of patient samples that were classified as high or low in miRNA level or CFC output using the calculated cut-offs. The hazard ratios for each variable in the multivariate CoxPH analyses are displayed with 95% confidence and the corresponding p value. ROC plots for each machine learning algorithm, RF and NB, are shown with their corresponding AUC values.

Full size image

Discussion

TKI therapies induce high rates of initial hematological and molecular responses in CP-CML patients, but TKI resistance and disease progression continue to pose a challenge for some patients. Current clinical scoring systems cannot accurately predict the heterogeneous treatment outcomes that are observed. From our analyses of CML patient samples at the treatment-naïve and post-NL treatment states from 58 patients, we propose two panels of NL-nonresponse predictors: The first panel consists of NL-CFC, miR-145 and miR-708, which predicts NL-nonresponse at diagnosis. The second panel of NL-CFC, miR-150 and miR-185 predicts NL-nonresponse at M1 and M3 post-NL treatment. Although miRNA signatures have alluded to TKI response prediction in CML, this study shows merit for combining matched miRNA expression profiles with in vitro CFC output to predict NL-specific response.

Previously, we demonstrated that IM response could be predicted in treatment naïve CD34+ cells by an in vitro CFC assay in a small cohort study [25]. Here, we show that CFC output data from patient CD34+ cells obtained at diagnosis can also predict NL response in a larger cohort. Since singular predictive variables alone may not be sufficient to predict response due to cellular variations and molecular complexity among patients, we generated a multivariate predictive panel based on a combination of expression of specific miRNAs and patient’s NL-CFC response. This combination improved the predictive value of using miRNA levels alone, as demonstrated by improved Concordance, ROC-AUC and PR-AUC at all time points. It has been well-documented that CML stem cells and their progenitor cells are the least TKI responsive and are responsible for disease recurrence when TKIs are discontinued [2, 43,44,45,46,47]. Notably, BCR-ABL1 transcript and protein levels are significantly elevated in CD34+ cells and even more so in the CD34+CD38− stem cell-enriched population compared to the bulk CD34− population [45, 48]. Therefore, it has been of interest to determine whether distinguishing features of CD34+ leukemic stem/progenitor cells from CML patients might vary amongst patients in correlation with the subsequent clinical response to TKI therapy. Our studies of primitive CD34+ CML cells from individual CP patients show that they indeed display cellular and molecular differences. In particular, clinically defined responders and nonresponders differ significantly from each other with respect to the growth response of their pre-treatment CFC to NL, and their miRNA expression changes. Our two newly identified predictive panels based on stably down- and up-regulated miRNAs whose expression levels differ significantly and associate with treatment response pre- and post-NL therapy might therefore form the basis of prospective tests for predicting early treatment response and ultimately for optimizing CML patient management.

In this study, we highlight miR-145 and miR-708 as potential predictive biomarkers in treatment naïve patients. MiR-145 expression changes were significantly different between IM-responders and IM-nonresponders in CD34+ CML cells obtained at diagnosis [39], adding to previous reporting that miR-145 is differentially expressed between CML peripheral blood patient samples and normal hematopoietic progenitor cells and between AP-CML versus CP-CML cells [38, 39]. NL treatment also seems to be able to increase expression of miR-145 in BCR-ABL1+ cell lines [36]. Although overall miR-145 expression in CD34+ cells differed significantly between NL-responders and NL-nonresponders in this study, we observed that the variation in miR-145 expression level among responders is relatively high with some responder patient cells expressing levels similar to NL-nonresponders; a trend that was also observed in other miRNAs in this study. These observations indicate heterogeneous and differential expression of miRNAs in individual CML patients even before NL treatment. Thus, it is critical to use advanced multivariable statistical and bioinformatics tools to precisely identify useful miRNA predictors among other clinical, biological and molecular parameters. Mechanistically, the role of miR-145 in CML has yet to be fully explored. However, miR-145 was found to be downregulated and was suggested to sensitize resistant cancer cells to treatment by modulating drug efflux and apoptotic pathways in solid tumors [49, 50]. A conflicting report shows that miR-145 is overexpressed in later stage breast cancer and supports cancer cell survival by promoting epithelial to mesenchymal transition and hypermethylating apoptotic genes [51]. However, miRNA expression is highly context-dependent and downstream effects can differ greatly between tumor types. While expression of miR-708 was found to be significantly reduced in CD34+ CML cells compared to normal CD34+ bone marrow cells, there is a lack of understanding of miR-708 in CML and other myeloid leukemias [39]. MiR-708 has been studied most extensively in lymphoblastic leukemias like acute lymphoblastic leukemia (ALL) where it is speculated to have both oncogenic and tumor suppressive functions depending on the subtype of ALL [52]. Clinically, ALL patients with low miR-708 levels were reported to have an increased risk of relapse [52]. Based on our analysis, reduced miR-708 expression in NL-nonresponder cells may also be indicative of more aggressive and drug-resistant properties.

In CD34+ CML cells obtained 1-month and 3-month post-NL treatment, we observed that NL-nonresponder patients had increased levels of miR-150 and decreased levels of miR-185 compared to NL-responders. Combining both miRNAs into multivariate models predicted NL-nonresponse at both time points. Interestingly, both miR-150 and miR-185 have been demonstrated to have tumor suppressive properties in CML [39, 53,54,55]. MiR-150 was suggested to be negatively regulated by BCR-ABL1 via MYC which in turn increases MYB expression and contributes to CML pathogenesis [53]. CML patients who were able to achieve early treatment response (ETR) after IM treatment were observed to have higher miR-150 levels [54, 55]. MiR-185 can be repressed by BCR-ABL1 in IM-nonresponder patients, which contributes to leukemic stem cell survival and TKI-resistance through increased PAK6 and OXPHOS mechanisms [39]. Furthermore, restoration of miR-185 could sensitize IM-resistant cells to TKIs [39]. In summary, these studies suggest that both miR-150 and miR-185 are key players in CML pathogenesis and rationalize their association with TKI-resistance.

The differences in significant miRNA classifiers among our two predictive panels, generated at the treatment naïve and post-NL treatment state, are likely attributed to dynamic changes in miRNA expression in response to NL treatment. Our observation of increased miR-150 expression in NL-nonresponder patients appears to conflict with other studies that reported miR-150 to be downregulated in CML and that its increased expression following IM treatment is an early positive predictor for IM response [54]. This discrepancy may be due to our use of enriched CD34+ stem/progenitor cells as opposed to unpurified peripheral blood and bone marrow cells or because miRNA expression may be modulated differently between NL and IM treatments [21]. Furthermore, it is not uncommon for miRNAs to possess conflicting roles, e.g. serving as oncogenic and tumor suppressive properties in the same cancer type, mainly owed to their ability to regulate expression of multiple target genes, which in turn also include both oncogenes and tumor suppressors [56]. Therefore, independent validation of these proposed predictive panels using greater numbers of patient samples and further exploration of the mechanistic roles of these select miRNAs in CML is warranted to facilitate their utility as prognostic tools in the clinic. Unique to this study, we combined miRNA expression data with patient-matched in vitro CFC outputs. In both BL and M1/M3 models, we found that inclusion of NL-CFC data in the multivariate panels improved predictive performance compared to miRNAs alone. It is widely agreed that treatment response may be affected by multiple factors not just related to biological mechanisms, such as pharmacokinetic variations between CML patients. Thus, prognostic algorithms that rely on genetic signatures alone may not be sufficient to model these extrinsic factors, highlighting the need for multifaceted molecular panels to improve predictive accuracy. The ability to develop rapid and robust tests to predict individual patients’ response to TKI therapy could ultimately have a profound impact on CML patient management, providing a foundation for more effective treatment decisions.