Can we forecast induction failure in acute myeloid leukemia?

Standard induction therapy for fit patients with acute myeloid leukemia (AML) consists of a combination therapy with anthracycline and cytarabine. This classical regimen, typically called “7+3”, has not changed for several decades.1 While many patients achieve a complete remission (CR) with standard induction therapy, approximately 10–40% of patients fail to respond to induction treatment.32 These patients are classified as having primary refractory disease (RD) or treatment failure, defined as a failure to achieve CR or incomplete hematologic recovery (Cri) after two courses of induction treatment.4 Unfortunately, treatment of patients with RD is extremely challenging, as even with salvage therapy followed by allogeneic stem cell transplantation, patient outcomes remain poor.3

It is still difficult for hematologists to reliably predict RD in newly diagnosed AML patients prior to initiation of therapy. At time of diagnosis, we typically risk stratify our patients based on their cytogenetic and molecular profile. A very helpful classification was introduced by the European Leukemia Net (ELN) in 2010,5 (revised in 20174) and this currently includes three prognostic groups integrating cytogenetics as well as the mutational status of FLT3-ITD (including mutational load), NPM1, ASXL1, TP53, RUNX1, CEBPA (biallelic mutants). However, this risk stratification is geared towards the estimation of overall survival (OS) and event-free survival (EFS), and not primarily towards forecasting RD.4 Although there is a strong correlation between treatment failure and OS, they still present different outcome measures.64

Several groups have attempted to develop specific scores to predict induction failure in AML. A reliable score primarily focusing on the likelihood of treatment failure rather than OS could improve patient care and treatment in many ways. If we could reliably predict that a patient would not respond to “7+3” treatment prior to induction therapy, we would be compelled to search for alternatives at the time of diagnosis, potentially sparing the patient from the toxicity of treatments that prove to be ineffective. As several new agents are being studied front line (e.g. FLT3 and IDH1/2 inhibitors with intensive chemotherapy, BCL2-inhibitors in combination with low-dose cytarabine or azacitidine, etc.) alternatives for “7+3” might soon become a reality. In addition, a reliable RD score could allow us to identify those patients who require an urgent donor search at the time of diagnosis.107 Furthermore, an RD score could become an important consideration when designing clinical trials that specifically target this high-risk patient group.

In this issue of Haematologica, Herold et al. introduce a 29-gene and cytogenetic score that can help to predict resistance to induction chemotherapy in adult AML patients.11 Importantly, this score was developed on the basis of various categories of prognostic markers, considering clinical characteristics, laboratory variables, cytogenetics, mutational status of 68 genes that are frequently mutated in AML, and the expression profile of 29 genes known to be prognostic for AML. Their score estimates the likelihood of primary RD based on large independent clinical training sets. The first cohort (training set 1) included 407 patients of the AML Cooperative Group (AMLCG trials between 1999–2005), the second cohort (training set 2) consisted of 462 AML patients treated in the Haemato-Oncology Foundation for Adults in the Netherlands (HOVON) trials and the validation cohort was based on 210 AMLCG-2008 trial patients with the addition of 40 patients with RD from the AMLCG 1999 trial. The implementation of a large validation cohort is critical for assessing the reliability of any score, especially for clinical practice. The score was calculated as a weighted linear sum of the individual predictors. Interestingly, the final predictor by Herold et al. (predictive score 29 MRC or PS29MRC) included expression levels of 29 genes and the UK Medical Research Council (MRC) cytogenetic risk classification, while other parameters such as gene mutations were tested but were excluded from the final score.12 Importantly, this predictive classifier proved to be significant for RD, both as a continuous variable as well as a dichotomous variable that divides patients into high and low risk. In the multivariate analysis, only PS29MRC, age and TP53 mutations remained independently significant for RD prediction. While the predictor was primarily designed to be associated with RD on day 16 after induction chemotherapy, the score also proved to be strongly associated with survival. When examining different groups of the current ELN 2017 classification, the predictive power of the score was shown in the intermediate and the unfavorable ELN groups, while it could not be shown in the favorable genetic group (likely related to low RD rate in patients with favorable cytogenetics). The validation cohort nicely reproduced the data of the training cohort. All these aspects are suggestive of a very reliable predictive score.

The area under receiver-operating characteristic curve (AUC) can be used as a measure for the predictive ability of a score, with an AUC of 0.7–0.8 classified as fair and less than we would desire for primary treatment decisions.1413 The classifier by Herold et al. reached an AUC of 0.76 in the validation set. In contrast, Walter et al. developed a model for resistance prediction in AML based on the analysis of 4601 patients treated within European and US AML trials.13 They found that age, performance status, white blood cell count, secondary disease, cytogenetic risk and NPM1/FLT3-ITD mutational status were strongly associated independently with primary resistance. Unlike Herold et al., they did not include a complex mutational and gene expression profile in their analysis (Table 1). However, with their model, they achieved a similar AUC (0.78) to that of Herold et al.

Table 1.Schematic overview of recent studies developing a model for response prediction to induction chemotherapy in intensively treated acute myeloid leukemia (AML) patients.

Krug et al. also developed a model based on a cohort of 1406 patients aged over 60 years diagnosed with AML but otherwise medically fit, and who underwent treatment with two intense induction chemotherapy cycles within the AML-CG.15 The validation cohort consisted of an independent cohort of 801 patients aged over 60 years. Their score was based on body temperature, age, secondary disease, hemoglobin, platelet count, fibrinogen, serum concentration of lactate dehydrogenase and cytogenetics. Instead of RD, the achievement of CR and early death were the primary outcome parameters of this score (Table 1). Using CR prediction, the model of Krug et al. had an AUC of 0.68 in the validation set.15

Gerstung et al. have also developed a prognostic algorithm based on a knowledge bank of 1540 AML patients whose cytogenetic, molecular profile, and clinical data were analyzed in detail.1716 Here, a number of outcome parameters can be obtained (including death without remission, death without and after relapse, alive after relapse, alive in first CR and alive without CR), and RD can be indirectly calculated (Table 1).

Thus, prediction of RD remains complex, and these scoring systems have yet to find their way into routine clinical practice. The questions of when and how we employ them for everyday clinical evaluation and treatment decisions remain. Here, feasibility and predictability must be considered. It will not be feasible to use a score requiring far more laboratory evaluation (e.g. microarray data, etc.) than is routinely performed. For example, gene expression analysis is not routinely performed in clinical practice and the time required might become relevant for patients with a high leukemic burden in need of urgent therapy. Furthermore, unlike sequencing, gene expression analysis is not covered by the healthcare systems of many countries. However, with the advances being made in technologies, such evaluation could quickly become more feasible. Just as important as feasibility is the level of predictability. We can only justify primarily basing our treatment decisions on scoring systems with a sufficiently high predictability. That none of the proposed scoring systems reach an AUC close to 0.9, even when including all parameters currently known to be prognostic, underscores the challenges of reliably predicting patient outcome at the time of diagnosis. This is highlighted by Herold et al., who used all prognostic parameters currently considered relevant, studied these parameters extensively in the context of RD prediction, and thus, rightfully described an “obstacle” to achieving a higher AUC that is difficult to overcome.

Herold et al. describe an innovative approach of how to tackle the pressing question of RD prediction. Independently of its clinical use, it can potentially help us to better understand the biology of primary refractory disease. It is still unknown why some patients with a molecularly more favorable risk profile still fail induction chemotherapy. The gene expression data that predict primary refractory disease might also lead the way to identifying novel targets for AML therapy. Even if the predictive classifier of Herold et al. may not find its way into clinical practice just yet, it carries the potential of becoming a tool for designing clinical trials and developing novel treatment strategies.