A machine learning approach identifies a transcriptomic signature predicting treatment-free remission in chronic myeloid leukemia

Abstract

Abstract Introduction Treatment-free remission (TFR) is currently the ultimate goal for patients with chronic myeloid leukemia (CML). However, nearly 50% of patients relapse after treatment cessation. While some clinical and biological factors have been shown to be associated with TFR, there is currently no validated score to predict TFR in the clinical setting, and the factors underlying molecular relapse are still poorly understood. The main objective of this study was to use transcriptomic data to predict TFR before imatinib (IMA) cessation in patients with CML. Methods Patients enrolled in the STIM2 multicenter trial (ClinicalTrials.gov, NCT01343173) and with available frozen peripheral blood cells (PBC) sample before IMA cessation were included in the main cohort (n=96 patients). In the STIM2 trial, IMA first-line cessation was proposed in CML patients with a sustained deep molecular response (defined as remission lasting more than 2 consecutive years and confirmed on five datapoints of BCR::ABL1 analyses by quantitative RT-PCR during these 2 years). Molecular relapse after TKI cessation was defined as loss of MMR at 1 point. Additional patients with available frozen PBC just before IMA or nilotinib (NIL) cessation from three French academic centers (Centre Hospitalier Lyon Sud, Centre Leon Bérard, Institut Bergonié) were included in the independent validation cohort (n=72 patients, including n=37 IMA and n=35 NIL first-line). High-throughput RNA sequencing was conducted on an Illumina® NovaSeq6000™ platform after polyA selection generating paired-end reads of 2x100bp per sample. The primary outcome was TFR success at 2 years as a binary endpoint. The main cohort was randomly split into a training (n=73) and a testing (n=23) set, with a stratification on TFR at 2 years. Differential gene expression analysis was performed using DESEQ2 on n=500 bootstrap samples with replacement of the training set. A transcriptomic signature was built upon genes showing a stable differential expression (i.e. with an FDR < 0.1 in at least half of the models). Results Ninety-six patients were included in the main cohort from the STIM2 multicenter trial. Median age was 55 [IQR: 45-65] years, with n=56 (58,3%) females. Fifty-two (54,2%) patients presented an e14a2 transcript, n=34 (35.4%) an e13a2 transcript (missing value for n=10). ELTS scores were low in 69 (71.9%), intermediate in 11 (11.5%), high in 8 (8,3%) and missing in 8 (8,3%) patients. The median duration of IMA was 77.3 [IQR: 53.9-107] months, and median MR4 duration was 39.5 [27.4-56.7] months before TKI cessation. Differential gene expression analysis using a custom machine learning approach identified 48 genes with a stable and significant differential expression across bootstrap samples of the training set. A gene expression signature based on these 48 genes demonstrated a good discrimination of TFR at 2 years, with AUROC [95%CI] of 0.84 [0.74-0.94] and 0.73 [0.60-0.97] in the training and testing sets, respectively. Evaluation on the external validation cohort confirmed the signature's discriminative power, with an AUROC of 0.71 [0.58-0.83] in the overall validation cohort, 0.77 [0.61-0.92] among IMA-treated patients and 0.68 [0.48-0.87] among NIL-treated patients. Considering TFR as a time-dependent variable, multivariate cox analysis confirmed a significantly higher probability of TFR associated with the signature, with an adjusted hazard ratio (aHR) of 0.32 [95%CI 0.14-0.74] for the high-median versus low-median expression group, independent of MR4 and TKI duration (p=0.008). Gene-set enrichment analysis of hallmark pathways revealed an enrichment of glycolysis, angiogenesis and TNF-a signaling via NFK-b in the high-median group (associated with TFR) and an enrichment in interferon-a, interferon-g, IL2/STAT5 and inflammatory response pathways in the low-median group (associated with molecular relapse). Summary/Conclusion These data provide evidence that TFR can be predicted in patients with CML in the chronic phase attempting TFR, at TKI cessation, using a transcriptomic approach. A prospective trial is currently ongoing to further validate the signature and establish cutoff that can be used at the individual level.