Identification of Multivariable microRNA and Clinical Biomarkers Panels to Predict Imatinib Response in Chronic Myeloid Leukemia at Diagnosis

Abstract

Imatinib Mesylate (IM) was once hailed as the magic bullet for chronic myeloid leukemia (CML) and remains a front-line therapy for CML to this day alongside other tyrosine kinase inhibitors (TKIs). However, TKI treatments are rarely curative and while they demonstrate efficacy in managing the disease, patients are often required to receive life-long treatment or risk experiencing relapse. Patients also frequently harbor innate TKI-resistance mechanisms to front-line IM therapy and without prior screening, patients can spend precious months on non-optimal treatment before transitioning to other therapeutic strategies. Thus, there is a growing interest in studying biomarkers in patients which can predict TKI response upon diagnosis. CML has unique microRNA (miRNA) expression profiles at different stages of the disease and in response to TKI-treatment. We previously generated global transcriptome profiles on treatment-naïve CD34 + CML cells with known subsequent imatinib (IM) responses and identified several differentially expressed miRNAs in IM-nonresponders as compared to IM-responders. We also reported that IM response could be predicted in treatment-naïve CD34 + cells by an in vitro colony forming cell (CFC) assay. Additionally, we previously analyzed patient data from the ENESTxtnd clinical trial to show that miRNA expression combined with CFC output data could accurately predict Nilotinib response in patients. In this study, we used a Cox Proportional Hazard (CoxPH) analysis to identify 22 out of a pool of 35 pre-validated miRNAs which were associated with IM nonresponse in CD34 + CML cells of 80 patients at diagnosis. These patients were later classified as IM-responders or IM-nonresponders based on the latest European LeukemiaNet (ELN) guidelines. Interestingly, a Welch t-test revealed 16 of these 22 IM-response associated miRNAs to be differentially expressed between IM-responders and IM-nonresponders. Between IM-responders and IM-nonresponders we also found 7 clinical parameters that were associated with IM-response of which 3 matched parameters had significantly different values including CFC assay outputs ( p=0.0015), Sokal scores ( p=0.0070) and white blood cell counts (WBC, p=0.0028). We then trained a machine learning model employing the random-forest (RF), gradient-boosting (GBM) and naïve-bayes (NB) algorithms with different combinations of the 16 miRNAs with and without the clinical parameters of these patients to identify panels with high predictive performance based on area-under-curve (AUC) values of receiver-operating-characteristic (ROC) and precision-recall (PR) curves. Notably, the multivariable panel consisting of both miRNAs and clinical features (AUC-ROC RF=0.83, GBM=0.83, NB=0.84, AUC-PR RF=0.68, GBM=0.67, NB=0.70) performed better than either miRNA (AUC-ROC RF=0.72, GBM=0.73, NB=0.78, AUC-PR RF=0.49, GBM=0.51, NB=0.4) or clinical (AUC-ROC RF=0.82, GBM=0.84, NB=0.84, AUC-PR RF=0.64, GBM=0.67, NB=0.64) panels alone. Interestingly, 2 miRNAs in this panel, miR-185 and miR-145, were also significant classifiers for our Nilotinib predictive study suggesting that expression patterns of these miRNAs may have predictive properties for multiple TKI responses. Thus, we show that predictive accuracy of biomarkers may be supplemented by inclusion of multivariable parameters, and our findings may inform future studies on developing predictive panels for more optimized treatment plans in the clinic.