BCR::ABL1 digital PCR for treatment‐free remission prediction in chronic myeloid leukemia patients: An individual participant data meta‐analysis

Approximately half of chronic phase (CP) CML patients in a stable deep molecular response (DMR; MR4.0; BCR::ABL1<0.01% on the International Scale [IS]) can safely and durably cease their therapy, entering a treatment-free remission (TFR).1-3 The alternative half of patients will show a molecular recurrence (MolR) requiring restarting TKI therapy.3 A longer total TKI treatment duration and DMR duration have been associated with higher probability of TFR.1, 2, 4 The identification of additional parameters to predict TFR success is clinically important. The depth of molecular remission beyond MR4.0 quantified by standard real-time PCR (RT-PCR) for BCR::ABL1 does not consistently correlate with the chance of sustained TFR.5 However, the depth of molecular response measured by BCR::ABL1 digital PCR, an alternative and more precise quantification technique, has been reported to be associated with TFR outcome in several studies. We aimed to establish the role of BCR::ABL1 digital PCR (dPCR) in TFR prediction, and to assess its relation to other clinical markers. This individual participant data (IPD) meta-analysis was done in adherence with the PRISMA-IPD flow diagram and checklist.6 Peer-reviewed studies were eligible if the predictive value of RNA-based BCR::ABL1 dPCR was assessed for CML patients attempting a first TKI discontinuation (Figure S1). PubMed, EMBASE, and the Cochrane Library databases were searched (Appendix S1). The outcome of interest was time to MolR. MolR was defined as BCR::ABL1>0.1%IS in one assessment or a 1-log BCR::ABL1 increase in two consecutive assessments. The primary objective was to assess MolR prediction with BCR::ABL1 dPCR. The secondary objectives were to assess other parameters for MolR prediction and to build a risk stratification model for MolR. A one-stage statistical approach was used, and the following parameters were analyzed in a Cox proportional hazards model: dPCR, age, sex, treatment duration, DMR duration, Sokal score, ELTS score, BCR::ABL1 transcript type, and TKI generation (see Supplemental Files for full methodology and references of included studies). Tables S1–S3 show the characteristics of the included studies and the final cohort of 635 patients. After TKI discontinuation, 41% of patients experienced MolR. MolR patients had a shorter duration of treatment (6.8 vs. 7.8 years, p = 0.012) and had a numerically, but not statistically significant shorter DMR duration (3.2 vs. 3.6 years, p = 0.245; studies BE, LU, AT) and UDMR duration (2.8 vs. 3.2 years, p = 0.098; studies MO, NI, CO). A total of 459 patients (72%) had a dPCR result below the prediction cut-off (dPCR-low). In these patients, 33% experienced MolR, versus 62% in patients with a dPCR result above the prediction cut-off (dPCR-high), p < 0.001. The e13a2 transcript, e14a2 transcript or both transcripts were detected in 108 (30%), 206 (58%) and 42 (12%) patients, respectively. MolR patients more often had the e13a2 transcript (41% vs. 22%, p < 0.001). No differences were noted in age, sex, TKI type, ELTS, or Sokal score. The 24-month probability of MolR in dPCR-low versus dPCR-high patients was 33% (95% CI, 28%–37%) and 61% (95% CI, 53%–68%), respectively (Figure S2A; p < 0.001). Using the cutoff of 6 years of TKI treatment, patients with a short versus long TKI treatment duration had 24-month MolR probability of 48% (95% CI, 41%–55%) and 36% (95% CI, 31%–41%), respectively (Figure S2B; p = 0.012). The HR of dPCR-high and TKI treatment duration (numeric) for MolR was 3.154 (95% confidence interval [CI], 1.952–5.096, p < 0.001) and 0.956 (95% CI, 0.916–0.999, p = 0.044; Figure S3A), respectively. UDMR duration (in three studies MO, NI, CO) was also significantly associated with TFR outcome with an HR of 0.898 (95% CI, 0.812–0.992, p = 0.034). Harboring the e14a2 transcript had a favorable impact on TFR, associated with a 24-month MolR probability of 34% (95% CI, 28%–41%) compared to 59% (95% CI, 48%–67%) in patients with the e13a2 transcript, and 48% (95% CI, 30%–61%) in patients harboring both transcripts (Figure S2C, p < 0.001). For the subsequent regression analysis, patients harboring both transcripts were grouped with patients harboring the e13a2 transcript alone, in order to clearly distinguish patients having a low MolR (i.e., e14a2 patients). The HR of the e13a2 transcript (or both) for MolR was 1.582 (95% CI, 1.119–2.236, p = 0.009). Both dPCR-high and BCR::ABL1 transcript type remained significantly associated with MolR probability in the multivariable analysis (Figure S3B). The HR of dPCR-high for MolR corrected for TKI treatment duration was 3.211 (95% CI, 2.001–5.154, p < 0.001). In a multivariable model combining dPCR-high, TKI treatment duration <6 years and BCR::ABL1 transcript type (e13a2 or both), their respective HR for MolR were 2.327 (95% CI, 1.548–3.497, p < 0.001), 1.278 (95% CI, 0.921–1.771; p = 0.141) and 1.480 (95% CI, 1.037–2.113; p = 0.031). The Cox proportional hazard assumption was not violated in the presented models. No significant interactions were observed between dPCR-high and other included variables. A sensitivity analysis of the first multivariable model was done using a two-stage random effects Cox regression analysis (Figure S4). The pooled HR of dPCR-low for MolR was 3.120 (95% CI, 1.876–5.189, p < 0.001). A low between-study heterogeneity was observed for treatment duration (I2 = 35%), and a moderate heterogeneity was observed for dPCR-high (I2 = 66%). An explorative analysis of molecular recurrence prediction in various clinical scenarios was described in the Supplemental Files and Figure S5, for example, exploring various subgroups based on time to recurrence, treatment duration, TKI type, and BCR::ABL1 transcript type. Based on our results and prior knowledge of relevant clinical predictive parameters, the parameters BCR::ABL1 dPCR, treatment duration, and BCR::ABL1 transcript type were combined in a composite risk score (Figure 1). The risk parameters were weighted based on the rounding of their HR for MolR (presented in model 2), that is, +2, +1, and +1, respectively (Figure 1). Patients were divided into low (score 0), intermediate (score 1–2), and high (score 3–4) MolR risk category. The model was fitted in three studies (Figure 1A) and validated in a fourth study (Figure 1B). All studies together, this resulted in a 24-month MolR probability of 25% (95% CI, 16%–32%), 45% (95% CI, 37%–52%), and 68% (95% CI, 55%–78%), respectively (see Supplemental Files for details on model validation). In this patient-level meta-analysis, depth of molecular response measured by BCR::ABL1 dPCR was a robust and consistent predictor for MolR, independent of treatment duration, TKI type, and BCR::ABL1 transcript type. BCR::ABL1 dPCR could aid in the timing of a TFR attempt, especially in the context of an "early" attempt in patients treated with a TKI for less than 6 years. Patients who attempted an early discontinuation had a favorable 24-month MolR probability of 39% if their dPCR result was low, compared to 67% if their dPCR result was high. A longer treatment duration in patients with a low dPCR result adds modestly to limiting MolR probability (29%). Patients with the e14a2 transcript more often maintained TFR and conversely, patients with the e13a2 transcript had an inferior probability of sustained TFR, which is in line with previous reports.7 Differential outcomes for both transcripts have already been described in terms of TKI treatment response.7 A technical quantification artifact cannot be excluded completely.8 However, in our study the association of BCR::ABL1 transcript type with a probability of MolR was independent of the dPCR result, which suggests a transcript-specific disease phenotype. One potential explanation might be a differential immunogenicity of e14a2 BCR::ABL1 positive cells as these provoke a cytotoxic T-lymphocyte response, in line with the theory that successful TFR involves immunological mechanisms.9 Based on our results and previous reports, we combined the three most relevant parameters in one risk stratification model, that is, depth of molecular response assessed by BCR::ABL1 dPCR, TKI treatment duration, and BCR::ABL1 transcript type. Patients were classified into low, intermediate, or high-risk categories, having a probability of MolR of 25%, 45%, and 68% at 24 months, respectively. This proposed model could deliver a risk-adapted approach in the decision of TKI discontinuation, as low-risk patients might confidently discontinue their TKI with a good chance of TFR success, and high-risk patients should preferentially continue their TKI treatment. The primary strength of this study is the intercontinental collaboration and assembling of patient-level data thus creating one of the largest discontinuation cohorts assessed to date. A limitation of this study is the heterogeneity of dPCR assays and therefore a dependency on study-specific prediction cutoffs. It was thus not possible to align dPCR results as a numeric variable in the models. We addressed this issue by the addition of a stratification per study and a random effect term in the Cox regression analysis. For further implementation in clinical practice, dPCR assays should undergo standardization for which alignment with the IS would be preferable. Secondly, the optimal TFR prediction cutoff on the IS should still be further validated. Two studies reported quite similar cutoffs on the IS, i.e. 0.0023%IS (NI) and 0.0030%IS (LU). Ongoing research will provide more clarity on this matter. In conclusion, depth of molecular response measured by BCR::ABL1 dPCR is a valuable and robust predictive parameter for successful TKI discontinuation and may be used to identify TKI stop candidates, including patients treated for less than 6 years aiming for an early discontinuation attempt. The combination of BCR::ABL1 dPCR with TKI treatment duration and BCR::ABL1 transcript type, further improved risk stratification for MolR. PW, PV, and CK conceptualized the research idea. PW and CK coordinated the study. Individual Patient Data were provided by SD, FEN, FXM, EA, MM, JR, SBe, DR, MF, SM, CGP, IC, LL, DY, SBr, GC, and MB. CK collected and pooled the data. CK and JR performed the statistical data-analysis. CK visualized the data. PW and CK had full access to the study data and take responsibility for the integrity of the data and the accuracy of the data analysis. PW and CK wrote the first draft of the manuscript. All co-authors revised the manuscript, figures, and tables. All co-authors approved the final version of the manuscript, figures, and tables. This meta-analysis was academically funded by an institutional grant from the Albert Schweitzer Hospital and a grant from ZonMw, a governmental funding agency in the Netherlands. The funders of this meta-analysis had no role in study design, data collection, data analysis, data interpretation, or writing of the report. SD is a speaker for Novartis and Incyte. FEN is a consultant for Novartis, SPARC; board entity for BMS-Celgene, Novartis, Pfizer, Incyte Biosciences, Institutional grants from Novartis, Incyte Biosciences. DY received research funding from Novartis and BMS; and received honoraria from Novartis, Takeda, Pfizer and Amgen. SBr is a member of the advisory boards of Qiagen, Novartis and Cepheid, received honoraria from Qiagen, Novartis, and Cepheid, and research funding from Novartis and Cepheid. PW has received institutional grants from Novartis and BMS-Celgene; and has received speaker fees from BMS-Celgene, Incyte, Novartis, and Pfizer. This individual participant data meta-analysis was conducted using data from previously published studies. Each of these studies had appropriate ethical approvals, and patient consent was obtained according to local and national regulations at the time of data collection. The data used in the current meta-analysis were de-identified and anonymized to protect participant confidentiality. Deidentified Individual Participant Data will be made available upon request to investigators whose proposed use of the data has been approved by our Study Committee with a representative of all included study cohorts. Proposals for access should be sent to [email protected]. Text S1. Methodology. Text S2. exploratory analyses of TFR prediction in various clinical scenarios. Text S3. risk stratification model. Appendix S1. search strategy, collected item list, R script main one-stage model. Table S1. study and BCR::ABL1 digital PCR characteristics. Table S2. baseline patient characteristics and molecular recurrence characteristics, grouped by study. Table S3. baseline characteristics in the pooled patient cohort, with or without molecular recurrence. Figure S1. flow diagram of study identification, screening, and inclusion. Figure S2. Kaplan Meier estimates based on different subgroups. Figure S3. one-stage univariable (A) and multivariable (B) Cox regression analysis presenting the hazard ratios for molecular recurrence. Figure S4. sensitivity analysis of multivariable model 1 using a two-stage approach. Figure S5. one stage multivariable Cox regression analysis in various clinical scenarios. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.