1 INTRODUCTION

Acute myeloid leukemia (AML) with mutated NPM1 is recognized as a distinct entity in the 2016 revision to the WHO classification of myeloid neoplasms.¹ Mutations in the NPM1 gene are common in AML, occurring in approximately 30% of all adult cases and in 50%-60% of those with a normal karyotype.^{1, 2} More than 55 different NPM1 mutations have been described, of which type A is the most common, constituting 80% of the variants in adult AML.^{3, 4}

Assessment of measurable residual disease (MRD) has two major clinical purposes: to evaluate treatment response after induction and consolidation therapy and to detect pending relapse allowing for preemptive treatment.^5-7 NPM1 mutations are particularly suitable as leukemia-specific targets for MRD detection since they are common, usually occur de novo, are present in the whole leukemic cell population, and are stable over the course of disease in most cases.^{2, 3} Numerous studies have shown the importance of NPM1 MRD for prognosis.^8-13 In bone marrow (BM), a less than 3 log₁₀ reduction of NPM1 transcripts as compared to the diagnostic level after two cycles of chemotherapy or a failure to reduce NPM1-mutated transcript levels to <200/10⁴ ABL copies post treatment entail a higher risk of relapse and death.^{9, 11, 14} In peripheral blood (PB), the persistence of NPM1-mutated transcripts after two courses of cytoreductive therapy is an independent prognostic factor for death.¹⁵ Yet, there is no single threshold, time point or material (BM vs PB), that consistently is the most predictive among clinical trials.^{9, 12-16} Consequently, there is a lack of consensus regarding optimal timing and relevant cutoffs for clinical assessment of MRD. In 2018, the European LeukemiaNet (ELN) published the first consensus guidelines for MRD in AML.⁷ In these, molecular MRD is recommended over the use of multiparameter flow cytometry (MFC) for NPM1-mutated AML. Among the molecular MRD techniques, reverse transcriptase quantitative PCR (RT-qPCR), performed on RNA, is currently preferred because of its high sensitivity, reflecting numerous copies of mutated mRNA for each copy of DNA. It is also the most commonly used method in large clinical studies.^{6, 9, 11, 12, 14, 15}

With the evolution of high-throughput sequencing techniques (next generation sequencing; NGS), new molecular targets have been identified, resulting in the possibility and need for DNA-based MRD analysis.^17-20 We and others have demonstrated the power of the DNA-based methods targeted deep sequencing (deep seq) by NGS, quantitative PCR (qPCR), and droplet digital PCR (ddPCR) for MRD detection.^{8, 21-30} Compared with RT-qPCR, DNA-based methods offer certain advantages. They allow for assessment of other mutations than RT-qPCR, which can only be used for highly expressed genes/gene fusions, and are not affected by variations in gene expression, thus more accurately measuring the fraction of leukemic cells. qPCR is accurate and highly sensitive ^{8, 29, 31} and neither ddPCR nor deep seq require standard curves or reference genes.^{25, 28} In addition, deep seq can readily identify all types of NPM1 mutations without the need for mutation-specific primers or probes.²¹ From a practical point of view, DNA preparation allows for more flexible protocols than those for RNA preparation, and the stability of DNA simplifies the logistics of sample transport.

There is a shortage of comparative studies between RNA and genomic DNA for MRD detection and consequently a lack of knowledge if DNA-based methods can add relevant MRD information or even replace RT-qPCR. Therefore, the aim of this study was to compare MRD results from NPM1 mutations measured by three different genomic DNA MRD methods; qPCR, ddPCR, and deep seq, with the gold standard RT-qPCR performed on RNA. The results show strong correlation between the three different DNA methods with RT-qPCR standing out as the most sensitive. As compared to a defined RT-qPCR BM cutoff of prognostic relevance for assessment of treatment response, the DNA-based MRD methods offer high negative and positive predictive values.

2 MATERIALS AND METHODS

3 RESULTS

3.1 Comparison between DNA-based methods and RT-qPCR

Because of its high sensitivity, RT-qPCR is the recommended method for MRD analysis in NPM1-mutated AML. In order to determine the performance of different DNA-based MRD detection methods as compared to RT-qPCR, we analyzed 110 follow-up samples from 30 patients with NPM1 type A mutation with RT-qPCR, qPCR, ddPCR and deep seq. There was a statistically significant correlation between all three DNA-based methods and RT-qPCR (Figure 1A-C). Among the DNA-based methods, qPCR showed the highest correlation with RT-qPCR (R_s = 0.903 in BM, R_s = 0.936 in PB, Figure 1A), followed by ddPCR (R_s = 0.881 in BM, R_s = 0.774 in PB, Figure 1B) and deep seq (R_s = 0.826 in BM, R_s = 0.743 in PB, Figure 1C), all P < .001. In BM, the DNA results agreed with those from the RT-qPCR analysis in 84% (56/67) of the samples analyzed by qPCR (Cohen's κ 0.640, 95% CI; 0.448-0.832), 81% (54/67) by ddPCR (κ 0.618; 0.445-0.790), and 73% (49/67) by deep seq (κ 0.496; 0.324-0.668) (Table 1a). In PB, the agreement with RT-qPCR for MRD detection was higher, with 95% (41/43) for qPCR with κ 0.890 (0.741-1.000), 88% (38/43) for ddPCR with κ 0.705 (0.468-0.943), and 86% (37/43) for deep seq with κ 0.638 (0.379-0.896) (Table 1a). To assess diagnostic accuracy of the DNA-based MRD methods, sensitivity, specificity, PPV and NPV were calculated across all of them (Table 1b). In BM, the DNA methods showed good specificity and PPV. In PB, all DNA methods proved highly specific (>96%), but qPCR showed superior sensitivity, PPV and NPV (Table 1b). The DNA-based methods failed to detect leukemic signals in a considerable proportion of RT-qPCR-positive BM samples (Table 1c), demonstrating superior sensitivity of RT-qPCR. On the other hand, leukemic DNA was detected by qPCR (and to a lesser extent ddPCR/deep seq) in a minor fraction (~10%) of samples without detectable transcripts with RT-qPCR, despite high RNA quality (Table 1c). To further dissect the relationship between expression levels and the amount of leukemic DNA, that is, leukemic cells, we analyzed the ratio between RT-qPCR and qPCR results in six patients with several (≥3) follow-up samples. The ratio between transcripts and mutated DNA varied both between individuals and within individuals (Figure 1D). Together, these results suggest that transcript analysis is more complex than just a simple measurement of leukemic cell burden.

TABLE 1. Detectability of leukemic signals and diagnostic accuracy of the different DNA-based MRD methods (qPCR, ddPCR or deep seq) as compared to RT-qPCR in NPM1-mutated AML. a: MRD results (detectability) for the NPM1 type A mutation in BM and PB follow-up samples comparing RT-qPCR with the different DNA-based methods. b: Diagnostic accuracy of the DNA-based methods with respect to detectable leukemic transcripts by RT-qPCR. c: False negative rates of the DNA-based MRD methods and RT-qPCR with respect to detectable leukemia in PB or BM with either RT-qPCR or any of the DNA-based methods

a)
RT-qPCR		qPCR		ddPCR		Deep seq
		Detectable	Undetectable	Detectable	Undetectable	Detectable	Undetectable
BM (n = 67)	Detectable	38	7	32	13	27	18
	Undetectable	4	18	0	22	0	22
PB (n = 43)	Detectable	12	1	9	4	8	5
	Undetectable	1	29	1a	29	1a	29

b)
DNA-based MRD method	Sensitivity / PPA (%)	Specificity / NPA (%)	PPV (%)	NPV (%)
BM Detectable leukemic transcripts by RT-qPCR
qPCR (95% CI)	84.4 (71.2-92.3)	81.8 (61.5-92.7)	90.5 (77.9-96.2)	72.0 (52.4-85.7)
ddPCR (95% CI)	71.1 (56.6-82.3)	100 (93.1-100)	100 (89.3-100)	62.9 (46.3-76.8)
Deep seq (95% CI)	60.0 (45.5-73.0)	100 (93.1-100)	100 (87.6-100)	55.0 (39.8-69.3)
PB Detectable leukemic transcripts by RT-qPCR
qPCR (95% CI)	92.3 (66.7-98.6)	96.7 (83.3-99.4)	92.3 (66.7-98.6)	96.7 (83.3-99.4)
ddPCR (95% CI)	69.2 (42.4-87.3)	96.7 (83.3-99.4)	90.0 (59.6-98.2)	87.9 (72.7-95.1)
Deep seq (95% CI)	61.5 (35.5-82.3)	96.7 (83.3-99.4)	88.9 (56.5-98.0)	85.3 1(69.9-93.6)

c)
	qPCR	ddPCR	Deep seq	RT-qPCR
	False neg. rate (%)	False neg. rate (%)	False neg. rate (%)	False neg. rate (%)
Detectable leukemic cDNA (n = 58) (95% CI)	13.8% (7.2-24.9)	29.3% (19.2-42.0)	39.7% (28.1-52.5)	n/a
Detectable leukemic DNA (n = 58) (95% CI)	n/a	n/a	n/a	10.3% (4.8-20.8)

• Abbreviations: BM, bone marrow; cDNA, complementary DNA; CI, confidence interval; n/a, not applicable; Neg., negative; NPA, negative percent agreement; NPV, negative predictive value; PB, peripheral blood; PPA, positive percent agreement; PPV, positive predictive value.
• ^a Same sample.

4 DISCUSSION

Monitoring MRD in AML patients has strong clinical relevance since accurate MRD detection can identify those at high risk of relapse. In AML with mutated NPM1, RNA-based MRD assessment using RT-qPCR is most established although several DNA-based methods, which all display certain advantages, are available. In this study, we examined the correlation for mutated NPM1 detection in both PB and BM follow-up samples between the gold standard method RT-qPCR and the DNA-based methods qPCR, ddPCR and deep seq. The results show that although the RNA-based method is more sensitive, DNA-based techniques may add relevant information that can be missed by RT-qPCR. Also, for detection of BM levels above an RT-qPCR cutoff of relevance for prognosis, the DNA-based methods showed positive and negative predictive values above 90%.

Since most laboratories have only one method for each MRD marker, published comparisons between RNA and DNA are limited, but support our findings of strong associations between results of RT-qPCR and DNA-based MRD methods. In an early study on NGS-based MRD assessment, including 38 samples from 10 patients, Thol et al demonstrated a 95% concordance between MRD results obtained by sequencing and RT-qPCR of type A mutated NPM1.³⁰ Also, Patkar et al demonstrated a significant correlation between RT-qPCR and deep seq targeting the type A NPM1 mutation in 71 samples, taken at the end of induction and consolidation therapy.³⁷ In AML with t(8;21)(q22;q22), Duployez et al showed a strong correlation between RUNX1-RUNX1T1 levels measured as fusion transcripts by RT-qPCR and as fusion gene copies by break-point specific qPCR.³⁸ In children with AML, we have shown good agreement between MRD levels obtained with patient-tailored deep seq of leukemia-specific mutations and RT-qPCR of RUNX1-RUNX1T1 and KMT2A-MLLT10 transcripts, respectively.²² When we compared detectability of MRD between the DNA-based methods and RT-qPCR in BM, the agreement was rather low, except for qPCR. In PB, the agreement was better. Taken together, these results imply that a proportion of transcript-positive samples will be missed with DNA-based techniques. This is probably of highest relevance when the aim of monitoring is early detection of pending relapse, a situation where RT-qPCR may be superior, at least in cases with fast relapse kinetics. However, in patients with NPM1 mutation without FLT3-ITD, molecular progression often occur over several months before hematological relapse.^{40, 41}

Despite a good correlation between RT-qPCR and DNA-based methods in this study, we noticed that the relation between leukemic transcripts and mutated DNA (RNA copy number per leukemic DNA molecule) fluctuated considerably between different follow-up samples for a number of patients. This indicates that transcript analysis is more complex than a frank translation of residual leukemic cell burden. A high leukemic transcript load may reflect either many leukemic cells or a few cells with high expression. Whether the expressing cells have leukemia-initiating potential or constitute differentiated cells can actually not be determined. Interestingly, we found 10% samples with convincing leukemic DNA signals, but devoid of leukemic transcripts. Thus, interpreting results from MRD assessment may be more complicated than meets the eye, and sensitive DNA-based analyses have the potential to add important information for some patients.

When comparing the three different DNA-based methods with each other, we found a high correlation. Previous studies comparing DNA-based methods for NPM1 MRD detection have also demonstrated a good correlation between the methods. However, those studies were all based on significantly fewer samples compared to this study, often with the aim to verify a new analysis method. We have in different such studies on NPM1 mutation type A shown significant correlations between qPCR and ddPCR in 34 samples²⁸ and between qPCR and deep seq in 8 samples.²⁵ Similar results have been obtained by others in analyses of other mutations in AML, for example, between NGS and ddPCR for mutated IDH1/2.²⁷ Here, qPCR showed detectable levels of mutated NPM1 in more samples than ddPCR and deep seq and detection of diluted OCI-AML3 at lower levels, probably reflecting the higher DNA input in the qPCR assay (DNA inputs were most often 500 ng for qPCR, 100 ng for ddPCR, and 100 ng for deep seq). It cannot be ruled out that the performance of the methods would be similar if equal DNA input is used. One drawback with qPCR is its requirement of both a standard curve and a reference gene – as in RT-qPCR – factors not needed in ddPCR or deep seq. Both qPCR and ddPCR, like RT-qPCR, require separate primers/probes for each specific mutation of interest but have simple workflows and short turn-around-times. For AML with mutated NPM1, most clinical laboratories using one of these methods will therefore not offer MRD analysis for patients with the rarest NPM1 mutations. Deep seq on the other hand simultaneously covers all different mutation types of NPM1 in a single assay, without the need for standard curve, mutation-specific primer/probe or reference gene, but to a higher cost.²¹ Thus, the different MRD methods all have their advantages and disadvantages, and the appropriate choice of method will depend on the specific mutation, the locally available technique and the purpose of the MRD analysis.

To shed light on the potential of DNA-based techniques for determination of MRD, we compared results from qPCR, ddPCR and deep seq with an RT-qPCR cutoff of high relevance for response to treatment; 3 log₁₀ reduction of NPM1 transcripts as compared to the diagnostic level and 200 mutant NPM1 copies/10⁴ ABL copies.^{9, 11, 14, 15, 32} Hubmann et al showed that a 3 log₁₀ reduction in NPM1/ABL1 mutation ratio in BM after induction therapy was the most appropriate threshold to identify patients at high risk of relapse.⁹ The 3 log₁₀ reduction in BM was confirmed by Ivey et al, who also demonstrated that an even stronger predictor of relapse and death is persistence of mutated NPM1 transcripts in PB after two cycles of chemotherapy.¹⁵ We aimed especially for high specificity for calculation of relevant DNA cutoffs, in order not to inflict a risk of overtreatment when risk stratifying with DNA dependent MRD methods. Based on the ROC analyses and available literature, we chose 0.1% leukemic cells for qPCR⁸ and consequently for a heterozygous mutation 0.05% VAF for ddPCR and deep seq. The high specificity at these cutoffs was confirmed for all three methods. Considering sensitivity, the DNA-based methods correctly classified most, but not all samples defined as MRD^high. Conversely, almost all MRD results above the cutoff level for the DNA-based methods were also classified as MRD^high for RT-qPCR, reflecting their high ability to predict clinically relevant MRD. The proposed cutoffs are also in line with the ELN recommendations for flow cytometry, where 0.1% leukemic cells is considered critical for risk stratification of AML patients.⁷ A limitation of our approach is that we were not able to investigate whether our RT-qPCR assay gave similar results to the assay originally described by Kronke et al,¹⁴ which may bias the interpretation. The DNA assays will require validation of their prognostic power in an independent cohort of uniformly treated patients before they can be used to inform treatment decisions.

So far there is no single assay, threshold or time point where MRD analysis can confidently predict the course for the individual patient. Relying on single measurements of MRD with whatever method confers a risk of oversimplification, since some MRD-negative patients will relapse anyway, and conversely, some patients will remain in remission despite detectable levels of MRD.^{19, 39} Molecular MRD methods are still unavailable for the majority of AML patients that lack NPM1 mutation, RUNX1-RUNX1T1 or CBFB-MYH11 but harbor other mutations. Whether our results on DNA-based methods can be translated to other mutations remains to be tested with the advent of more extensive DNA-based MRD analyses. Important aspects to consider for such translation is whether the mutations of interest reflect clonal hematopoiesis.⁴²

In conclusion, this study shows that DNA-based methods offer high positive and negative predictive values with respect to residual leukemic NPM1 transcripts at levels of importance for response to treatment. This suggests that DNA-based MRD techniques have the potential for determination of clinically relevant MRD, although offering lower sensitivity than RT-qPCR.

5 ETHICS APPROVAL STATEMENT

The study was approved by the Regional Ethical Review Boards at Lund University (diary number 2014/505 and diary number 2017/850) and at the University of Gothenburg (diary number 2017/1138).

CONFLICT OF INTEREST

YC and LHS have ownership interest (including stock, patents, etc) in SAGA Diagnostics AB. There are no other conflicts of interest to disclose.

AUTHOR CONTRIBUTIONS

LP, SJA, ME and LF conceived and designed the study. VL, GJ, ME and LF provided samples and clinical information. LP, SJA, GSB, HK, KV, LZ, JA, YC and LHS performed the analyses. AA and GSB created the script for deep seq. LP, SJA, ME and LF analyzed data and wrote the report. All authors approved the final version of the report for publication.

PATIENT CONSENT STATEMENT

Patients consented to biobanking of samples for research purposes.