Measurable residual disease (MRD) monitoring by next-generation sequencing before allogeneic hematopoietic cell transplantation in AML

Abstract

Molecular measurable residual disease (MRD) assessment is not established in approximately 60 percent of acute myeloid leukemia (AML) patients due to the lack of suitable markers for quantitative real-time PCR. To overcome this limitation we established an error-corrected next-generation-sequencing (NGS) MRD approach which can be applied to any somatic gene mutation. The clinical significance of this approach was evaluated in 116 AML patients undergoing allogeneic hematopoietic cell transplantation (alloHCT) in complete morphologic remission (CR). Targeted resequencing at the time of diagnosis identified a suitable mutation in 93 percent of the patients covering 24 different genes. MRD was measured in CR samples from peripheral blood or bone marrow before alloHCT and identified 12 patients with persistence of an ancestral clone (variant allele frequency, VAF >5%). The remaining 96 patients formed the final cohort of which 45% were MRD positive (median VAF 0.33, range 0.016-4.91%). In competing risk analysis cumulative incidence of relapse (CIR) was higher in MRD positive than negative patients (HR 5.58, P<0.001, 5-year CIR 66 vs 17%), while non-relapse mortality (NRM) was not significantly different (HR 0.60, P=0.47). In multivariate analysis MRD positivity was an independent negative predictor of CIR (HR 5.68, P<0.001) besides FLT3-ITD and NPM1 mutation status at the time of diagnosis, and of overall survival (OS) (HR 3.0, P=0.004) besides conditioning regimen, TP53 and KRAS mutation status. In conclusion, NGS-based MRD is widely applicable to AML patients, highly predictive of relapse and survival, and may help refining transplant and posttransplant management in AML patients.

Introduction

Molecular predictors for relapse following allogeneic haematopoietic cell transplantation (alloHCT) are urgently needed as 16-51% of AML patients experience recurrence of disease after alloHCT depending on the conditioning intensity and other factors.¹ Measurable residual disease (MRD) monitoring for nucleophosmin 1 (NPM1) mutations has proven to be highly predictive for relapse in AML patients treated with or without alloHCT.^2–6 However, the majority of AML patients undergoing alloHCT do not have the NPM1 marker so that alternative markers and techniques are required.⁷ As detailed in the recently published European LeukemiaNet AML MRD consensus document, no uniform approach to detect these cells has yet been established in AML.⁸ Flow cytometry can be used for MRD detection (Flow-MRD) in the majority of AML patients but it is challenging to standardize this technique.⁸ In over 90% of AML patients molecular aberrations can be identified by next-generation sequencing,⁷ a technology that has found its way into clinical practice for the initial mutational screening at the time of AML diagnosis.⁹ Using NGS for MRD detection is appealing because its flexibility allows using almost every mutated gene as MRD marker. In a pilot study, we have already tested NGS for MRD detection considering mutations in NPM1 and FLT3-ITD.¹⁰ There, we demonstrated the utility of NGS for MRD detection in these non-single nucleotide variants (SNV).¹⁰ The major problem of NGS based MRD detection lies in a sequencing error of up to 1% at each nucleotide position. This challenges the distinction of true mutations with low variant allele frequency (VAF) from sequencing errors. These errors primarily arise during library preparation and sequencing. Thus, for reliably detecting MRD with NGS these problems need to be overcome by an error-corrected sequencing approach in order to detect small mutated clones with a VAF of 1% and less and to discriminate these from sequencing errors.

Mutations in many genes may arise in hematopoietic cells as part of age-related clonal hematopoiesis (ARCH).¹¹ The most frequently mutated gene in clonal hematopoiesis is DNMT3A, which could not predict relapse as MRD marker in previous studies.^12,13 However, the role of other genes less frequently associated with ARCH for measuring MRD and predicting relapse risk is not clear and warrants further investigation.

We adapted the increasingly used concept of error-corrected sequencing in a NGS-based approach of MRD assessment and analysed it for its ability to prognosticate relapse and survival following alloHCT.

Patients, materials and methods

Patients

Patients were included if they were 18 years or older, had a diagnosis of AML excluding acute promyelocytic leukemia, underwent alloHCT in complete morphologic remission (CR) between 1996 and 2016 at Hannover Medical School, and had DNA available at diagnosis and in CR just before alloHCT (median time from diagnosis to CR was 91 days, median time from sample to transplantation 24 days, range 5-71 days). 116 patients were identified and underwent myeloid panel sequencing to identify a suitable molecular MRD marker, which could be a mutation in any gene except DNMT3A and NPM1, which we excluded from MRD markers due to the association of DNMT3A with clonal hematopoiesis^14–17 and the established methodology to measure NPM1 MRD. DNA from the relapse sample was available for 20 patients and was analysed by panel sequencing. Written informed consent was obtained according to the Declaration of Helsinki, and the study was approved by the institutional review board of Hannover Medical School (ethical vote 2179-2014).

NPM1 quantitative real-time polymerase chain reaction (RT-PCR) for MRD assessment

Four adult AML patients with mutated NPM1 were included to quantify and compare NPM1 ^mut transcript levels by quantitative real-time polymerase chain reaction (qRT-PCR) and our NGS based MRD assessment at diagnosis, in CR and at relapse. NPM1 ^mut-specific qRT-PCR was performed as as previously described.²

Cytogenetic and molecular analysis

Pretreatment blood or bone marrow samples were studied centrally by G-and R-banding analysis. Chromosomal abnormalities were described according to the International System for Human Cytogenetic Nomenclature.¹⁸ DNA was extracted as described before using the Allprep DNA/RNA purification kit (Qiagen, Hilden, Germany).¹⁹ DNA sequencing libraries were prepared from samples at diagnosis (n=116) and at relapse (n=20) with a custom TruSight myeloid sequencing panel according to the manufacturers’ instructions (Illumina, San Diego, CA), which included 46 entire genes or hotspots recurrently found in myeloid leukemias (Supplementary Table S1). All samples received individual dual indexes and were pooled at equimolar concentrations. Eighty samples per lane were sequenced on an Illumina HiSeq2500 sequencer using the HiSeq Rapid SBS Kit v2 (Illumina, San Diego, CA) for 251 cycles in both directions. The sequencing data was analysed as described before with modifications as detailed in the supplement.²⁰

Error-corrected sequencing for sensitive MRD detection

We established a custom amplicon sequencing approach for the MiSeq sequencer (Illumina) based on the previously developed method of error corrected sequencing²¹ for sensitive detection of SNVs and indels (Figure 1), which is detailed in Supplementary Methods. To reduce the sequencing error rate we used a proofreading polymerase for PCR, introduced random barcodes to allow bioinformatic error correction, performed the initial PCR with only 5 PCR cycles, avoided identical multiplex identifier (MID)/gene combinations on consecutive MiSeq runs, and established a standardized approach of bioinformatic analysis with error correction (see Supplementary Methods).

Figure 1. Outline of amplicon-based NGS-MRD analysis.

(A) Primers containing the complementary sequence (green), a 16 bp random unique molecular identifier (red) and a so called “common sequence” (blue) are designed around a known mutation (red star) covering a nucleotide sequence of 87-155 bp. The first PCR is run for 5 cycles with primers detailed in the graph and the product is cleaned up and size selected. (B) The second PCR is run for 25 cycles with primers containing a complementary sequence to the “common sequence” (blue), the multiplex identifier (MID, orange) and the Illumina adapter (purple) and the product is cleaned up and size selected. (C) Up to 25 samples are pooled and run on a MiSeq sequencing instrument with 251 cycles in both directions. (D) Sequencing reads are demultiplexed by their MID, aligned to the target region and error corrected by reconstructing read families using the random barcodes introduced in the first PCR and by constructing read1/read2 consensus sequences. The x-axis shows nucleotides around the mutated target region, the y-axis shows the variant allele frequency. The left graph (i) shows the graphical representation of the read family analysis. A blue dot shows the largest variant allele frequency (LVAF) at the respective nucleotide position. The black vertical line indicates the base position of the target mutation and in this example shows a mutation clearly above the background sequencing error. The grey vertical lines indicate +/- 20 basepairs up-and downstream of the target peak. The black horizontal line indicates the mean background error calculated from LVAFs. The green horizontal line indicates the background error + 3 standard deviations of the background error (the target LVAF should be above this line). The red horizontal line indicates the LVAF of the target minus 3 standard deviations of the background error (no other peak should be above this line within +/- 20 basepairs of the target LVAF). The right graph (ii) shows the graphical representation of the R1/R2 corrected analysis. LVAFs are plotted for each nucleotide position. Vertical and horizontal lines are defined as in the left panel.

Abbreviations: PCR, polymerase chain reaction; bp, basepair; MID, multiplex identifier; PE, paired end; VAF, variant allele frequency.

The Illumina Miseq reagent kit v3 (600 cycles, San Diego, USA) was used for sequencing and was run on the MiSeq sequencer aiming for a high coverage per sample (we obtained 75,804 - 937,653 aligned reads per marker with 251 sequencing cycles in both directions).

Bioinformatics and statistical analyses

Bioinformatics analysis of myeloid panel sequencing and of error-corrected sequencing is described in detail in Supplementary Methods. We developed a standardized algorithm for calling SNVs, small and large indels MRD positive or negative based on the number of read families (RF mode, error corrected sequencing) or the number of matching forward (R1) and reverse (R2) reads (R1/R2 mode), considering the background error of the individual sample as limit of detection (see Supplementary methods and Supplementary Figure S1 for details).

Median follow-up time for survival was calculated according to the method of Korn.²² Overall survival (OS) endpoints, measured from the date of alloHCT, were death (failure) and alive at last follow-up (censored). Relapse free survival (RFS) endpoints, measured from the date of alloHCT, were relapse (failure), death in CR (failure) and alive in CR at last follow-up (censored). The Kaplan-Meier method and log-rank tests were used to estimate the distribution of OS and RFS, and to compare differences between survival curves. The Gray test was used to compare and visually represent cumulative incidences of non-relapse mortality (NRM) and relapse (CIR) as competing risks using R package cmprsk. Sixty categorized variables were considered in univariate analysis for OS, CIR, RFS and NRM (Supplementary Table S2 for variables and definition of variable categories). Details of multivariate analysis are described in Supplementary Methods.

Comparisons of variables were performed using the Kolmogorov-Smirnov test and Student’s t-test for continuous variables and the Chi-squared test for categorical variables for exploratory purposes. The positive predictive value was calculated by the ratio of true positive patients / (true positive + false positive patients). The negative predictive value was calculated by the ratio of true negative patients / (true negative + false negative patients).²³

The two-sided level of significance was set at P <0.05. The statistical analyses were performed with the statistical software package SPSS 24.0 (IBM Corporation, Armonk, NY), statistical program R²⁴ using packages “survival”, cmprsk”; Microsoft excel 2010 (Microsoft Corporation, Redmond, WA) and custom linux scripts.

Results

Validation of the NGS-MRD approach

To validate the NGS-MRD approach, four NPM1 mutated AML patients were evaluated for MRD levels at diagnosis, in remission and at molecular relapse by an established RT-PCR method² and by our NGS-MRD approach with primers covering the NPM1 mutation hotspot. MRD positivity/negativity was concordant between both methods for all patients and time points (Supplementary Figure S2) thus supporting sensitive, specific and quantitative measurement of MRD by our NGS-MRD method. In addition, we analysed the detection limit with dilution curves for IDH1 and IDH2 mutated AML samples with a VAF of 50% using DNA from HL60 cells as IDH1/2 wildtype control. By error-corrected sequencing (read family mode) analysis we achieved a linear detection range down to the 10,000-fold dilution, corresponding to a limit of detection of 0.005%. The R1/R2 analysis mode showed a linear detection range down to the 3,000-fold dilution, corresponding to a limit of detection of 0.017% (Figure 2). Furthermore, in one patient with CBFB-MYH11 positive AML we identified with qRT-PCR 56 copies of CBFB-MYH11 per 10,000 ABL copies (0.56%, MRD positive). By NGS we had analysed NRAS in the same patient and found the mutation at a VAF of 0.035% (MRD positive). Thus, both methods identified MRD in this patient, supporting the validity of this NGS assay.

Figure 2. Serial dilution of IDH1 and IDH2 mutated cells with IDH1/IDH2 wildtype DNA.

IDH1 and IDH2 samples were mutated at a VAF of 50%. Mutated DNA was diluted with DNA from IDH1/IDH2 wildtype cells up to 30,000-fold and 8 different dilutions were tested in 3 independent replicates. RF, read family approach (error corrected sequencing analysis). R1/R2, analysis by forward/reverse read correction.

Patients and feasibility of NGS-MRD assessment

At diagnosis eight of 116 patients (7%) did not have any appropriate mutation in the myeloid panel analysis and were excluded from further analysis. Of the remaining 108 patients the selected MRD mutation had a VAF >5% in the CR sample before alloHCT in 12 patients (10%), who were also excluded from primary analysis. The following 9 genes were found with high VAF in CR: ASXL2 (n=1), CBL (n=1), CUX1 (n=1), ETV6 (n=1), IDH1 (n=2), IDH2 (n=2), PPM1D (n=1), STAG2 (n=1), and TET2 (n=2). The VAF ranged from 6.5 to 53.4% and clinical characteristics of these patients are shown in Supplementary Table S3. These variants may indicate persistence of the mutation in differentiated cells, preexisting clonal hematopoiesis, or germline origin. Our main analysis cohort therefore consisted of 96 patients, who had an evaluable MRD marker with a VAF of 5% or less at the time of CR before alloHCT. Ninety-one patients (95%) had been transplanted in first CR and 5 patients (5%) in second CR. Genomic DNA from peripheral blood (n=56) or bone marrow (n=40) was used to analyse 1 (n=66), 2 (n=28) or 3 (n=2) mutations per patient for MRD in the pretransplant CR sample. Mutations in the following 24 genes were used for MRD analysis in the indicated number of patients: IDH2 (n=17), RUNX1 (n=17), NRAS (n=12), FLT3 (n=11), TP53 (n=8), IDH1 (n=7), KRAS, SF3B1, STAG2 (n=6 each), EZH2 (n=5), BCOR, BCORL1, PTPN11, TET2, WT1 (n=4 each), PHF6, (n=3), RAD21, ETV6 (n=2 each), CBL, DDX41, KDM6A, SETBP1, SMC3, STAG1 (n=1 each). One hundred mutations were single nucleotide variants and 28 mutations were indels. In the NGS analysis, the mean number of aligned reads was 414,603 (range 75,804 - 937,653) and the mean number of read families was 74,640 (range 8,964-305,795). Detailed results of MRD quantification are shown in Supplementary Table S4. Background error and sensitivity threshold (background error + 3 standard deviations) of the assay was calculated for each mutation and reached a median largest variant allel fraction (LVAF) of 0.0071% and 0.015%, respectively (Supplementary Table S4, median of read family and R1/R2 corrected analysis).

Incidence of MRD and associated patient characteristics

Forty three of 96 evaluable patients (45%) were MRD positive with a median VAF of 0.33% (range 0.016% to 4.91%, Supplementary Table S4), while 53 patients (55%) were MRD negative. MRD positive patients more often had an adverse cytogenetic profile according to MRC, more often had an adverse 2017 European LeukemiaNet (ELN) risk⁹ and complex karyotype, and by trend more often had a hematopoietic cell transplantation comorbidity index (HCT-CI)²⁵ higher than 2 (Table 1). Other clinical and transplantation-associated characteristics were similarly distributed between MRD positive and negative patients (Table 1). Importantly, the frequency of MRD positivity was similar whether MRD was determined in peripheral blood or bone marrow (43% and 48%, respectively). Molecular aberrations showed no significant differences between MRD positive and negative patients apart from more frequent TP53 and SF3B1 mutations and less frequent NPM1 mutations and a trend to less frequent U2AF1 mutations in MRD positive patients (Supplementary Table S5).

Table 1. Comparison of clinical characteristics between MRD positive (n=43) and MRD negative AML patients (n=53).

Characteristic	MRD positive n=43	MRD negative n=53	P
Age			0.89
median (years)	50.1	51.7
range (years)	18.8-67.7	21.1-69.9
Sex			0.23
male - no. (%)	15 (35)	25 (47)
female - no. (%)	28 (65)	28 (53)
ECOG performance status at diagnosis			0.13
0 - no. (%)	11 (26)	10 (19)
1 - no. (%)	31 (72)	36 (68)
2 - no. (%)	1 (2)	7 (13)
FAB-subtype			0.70
M0 - no. (%)	7 (16)	7 (13)
M1 - no. (%)	4 (9)	11 (21)
M2 - no. (%)	9 (21)	6 (11)
M4 - no. (%)	9 (21)	13 (25)
M5 - no. (%)	6 (14)	3 (6)
M6 - no. (%)	0 (0)	3 (6)
M7 - no. (%)	1 (2)	0 (0)
missing data - no. (%)	7 (16)	10 (19)
Type of AML			0.71
de novo - no. (%)	31 (72)	40 (75)
secondary* - no. (%)	12 (28)	13 (25)
Cytogenetic risk group§			0.02
Favorable – no. (%)	2 (5)	1 (2)
Intermediate – no. (%)	24 (56)	44 (83)
Adverse – no. (%)	17 (40)	8 (15)
2017 ELN risk group9			0.006
Favorable – no. (%)	5 (12)	15 (28)
Intermediate – no. (%)	11 (26)	19 (36)
Adverse – no. (%)	27 (63)	19 (36)
Complex karyotype			0.003
no – no. (%)	31 (72)	50 (94)
yes – no. (%)	12 (28)	3 (6)
WBC count			0.57
median - (x109/l)	8	19
range - (x109/l)	0.8-227	0.7-200.9
Hemoglobin			0.32
median – g/dL	9.5	9.7
range – g/dL	4.7-15	4.6-13.7
Platelet count			0.25
median - (x109/l)	52	63
range - (x109/l)	4-402	11-427
Number of chemotherapy cycles before alloHCT			0.19
one cycle – no. (%)	8 (19)	3 (6)
two cycles – no. (%)	26 (60)	37 (70)
three cycles – no. (%)	9 (21)	13 (25)
Type of CR sample for MRD	0.825
peripheral blood – no. (%)	25 (58)	32 (60)
bone marrow – no. (%)	18 (42)	21 (40)
Donor match			0.33
MRDonor – no. (%)	14 (33)	15 (28)
MUDonor – no. (%)	23 (53)	27 (51)
MMRDonor – no. (%)	1 (2)	1 (2)
MMUDonor – no. (%)	4 (9)	10 (19)
missing data - no. (%)	1 (2)	0 (0)
Type of conditioning regimen			0.84
Myeloablative – no. (%)	22 (51)	26 (49)
Reduced intensity – no. (%)	21 (49)	27 (51)
Stem cell source			0.58
peripheral blood stem cells – no. (%)	40 (93)	49 (92)
bone marrow – no. (%)	2 (5)	4 (8)
missing data - no. (%)	1 (2)	0 (0)
Remission status			0.83
CR1 – no. (%)	41 (95)	50 (94)
CR2 – no. (%)	2 (5)	3 (6)
HCT-CI score before transplantation25			0.07
0-2 – no. (%)	28 (65)	44 (83)
>2 – no. (%)	14 (33)	9 (17)
missing data -no. (%)	1 (2)	0 (0)
Donor sex			0.35
male - no. (%)	15 (35)	24 (45)
female - no. (%)	27 (63)	29 (55)
missing data - no. (%)	1 (2)	0 (0)
CMV status			0.24
donor neg/patient neg – no. (%)	15 (35)	13 (25)
any other combination – no. (%)	27 (63)	40 (75)
missing data - no. (%)	1 (2)	0 (0)

Abbreviations: ECOG, performance status of the Eastern Cooperative Oncology Group; FAB, French-American-British classification of acute myeloid leukemia; ELN, European LeukemiaNet; WBC, white blood cell count; CR, complete remission; MRD, measurable residual disease; MRDonor, matched related donor; MUDonor, matched unrelated donor; MMRDonor, mismatched related donor; MMUDonor, mismatched unrelated donor; HCT-CI, hematopoietic cell transplantation comorbidity index; P, P-value from two-sided chi-squared tests for categorical variables and from two-sided Student’s T or Kolmogorov-Smirnov tests for continuous variables.

^*secondary AML meaning leukemia secondary to chemo-or radiotherapy or an antecedent hematologic disease;

^§The cytogenetic risk group is defined according to Medical Research Council criteria⁴⁴

Prognostic effect of NGS-MRD

The median follow up of the 96 evaluable patients was 6.2 years. 27 of 43 MRD positive patients (63%) and 8 of 53 MRD negative patients (15%) relapsed after alloHCT. By competing risk analysis for cumulative incidence of relapse (CIR) and non-relapse mortality (NRM) patients with positive MRD had a significantly higher CIR than MRD negative patients (HR 5.58, 95%CI 2.47-12.60, P<0.001, 5-year CIR 66% vs 17%), while NRM was not different (HR 0.61, 95%CI 0.15-2.37, P=0.47, 5-year NRM 9% vs. 11%, Figure 3, Table 2 and Supplementary Table S6). RFS and OS were significantly shorter in MRD positive compared to MRD negative patients (RFS: HR 3.56, 95%CI 1.86-6.81, P<0.001, 5-year RFS 31 vs 74%; OS: HR 3.06, 95%CI 1.53-6.12; P=0.002; 5 year OS 41 vs 78%, Figure 3, Table 2). For multivariate analysis 60 variables were considered in univariate analysis for OS, RFS, CIR and NRM and were included in a multivariate model if P was ≤ 0.1 (Supplementary Table S6, FLT3-ITD and NPM1 were included despite P>0.1). In multivariate analysis MRD positivity was an independent predictor of CIR (HR 5.67, 95%CI 2.30-14.0, P<0.001) besides DNMT3A, NPM1 and FLT3-ITD mutation status, and of OS (HR 3.0, 95%CI 1.41-6.38, P=0.004) besides conditioning regimen, KRAS and TP53 mutation status (Table 2). MRD positivity was also an independent predictor of RFS (HR 3.41, 95%CI 1.72-6.75, P=0.001) besides patient age, TP53 and KRAS mutation status, while it had no effect on NRM (Table 2). KRAS mutated patients had a high risk of NRM and therefore were also predictive for OS and RFS, but not for CIR. We correlated the time to relapse with the VAF in MRD positive patients who relapsed after alloHCT. There was a negative correlation (Spearman's rank correlation coefficient -0.352, P=0.072), indeed suggesting that time to relapse is longer in patients with lower pre-transplant MRD load (see Supplementary Figure S3). Based on Cox regression analysis using VAF or log10 VAF as a continuous variable the hazard of relapse increased with a hazard ratio of 1.44 per 1% increase of VAF (95%CI 1.12-1.86, P=0.004); using log10 VAF, the hazard of relapse increased with a hazard ratio of 1.94 per 10-fold increase of VAF (95%CI 1.11-3.38, P=0.019).

Figure 3. CIR, NRM, RFS and OS for MRD positive and negative patients.

(A) CIR and NRM by competing risk analysis for MRD positive (n=43) and negative (n=53) patients. (B) RFS for MRD positive (n=43) and negative (n=53) patients. (C) OS for MRD positive n=43) and negative (n=53) patients.

Table 2. Univariate and multivariate analysis for CIR, NRM, RFS and OS in 96 AML patients.

Endpoint	Variables in the model	Univariate analysis				Multivariate analysis
		HR	95% CI	P		HR	95% CI	P
CIR	MRD positive vs negative	5.58	2.47-12.59	<0.001		5.67	2.30-14.0	<0.001
	DNMT3A mutant vs wildtype	0.41	0.17-0.97	0.042		0.33	0.12-0.89	0.029
	FLT3 ITD present vs absent	0.91	0.38-2.15	0.828		3.70	1.36-10.10	0.011
	NPM1 mutant vs wildtype	0.19	0.04-0.81	0.025		0.22	0.06-0.77	0.018
NRM	AML type secondary vs de novo AML	6.25	1.58-24.63	0.009		5.64	1.80-17.65	0.003
	Age above 60 years vs 18-60 years	3.93	1.07-14.43	0.039		7.13	1.75-29.16	0.006
	KRAS mutant vs wildtype	5.44	1.6-18.46	0.007		19.4	3.32-113	0.001
	NPM1 mutant vs wildtype	1.92	0.48-7.73	0.356		4.42	1.70-11.52	0.002
RFS	MRD positive vs negative	3.56	1.86-6.81	<0.001		3.41	1.72-6.75	0.001
	Age above 60 years vs 18-60 years	1.89	1.02-3.49	0.043		2.23	1.15-4.33	0.017
	KRAS mutant vs wildtype	2.13	1.03-4.39	0.041		3.51	1.64-7.50	0.001
	TP53 mutant vs wildtype	2.74	1.30-5.79	0.008		2.26	1.09-4.70	0.029
OS	MRD positive vs negative	3.06	1.54-6.12	0.002		3.00	1.41-6.38	0.004
	Conditioning MAC vs RIC	0.54	0.28-1.02	0.056		0.37	0.20-0.69	0.002
	KRAS mutant vs wildtype	2.84	1.33-6.05	0.007		5.58	2.68-11.62	<0.001
	TP53 mutant vs wildtype	3.57	1.67-7.61	0.001		3.97	1.90-8.29	<0.001

NOTE. Hazard ratios greater than or less than 1 indicate an increased or decreased risk, respectively.

The prognostic effect of MRD on CIR and OS after alloHCT was largely preserved when the analysis was restricted to patients in whom MRD was either measured in peripheral blood or in bone marrow (Figure 4). We also analysed paired peripheral blood and bone marrow samples from eight patients with available DNA and found very good concordance between the two tissues. Seven of eight patients were positive in both tissues, while one patient was positive in bone marrow but negative in peripheral blood (Supplementary Figure S4). CIR and OS remained separated by MRD into prognostic subgroups when the cutoff for MRD positivity was changed to 0.1% or 1.0% (Supplementary Figure S5). By changing the MRD cutoff to 0.1% and 1% the proportion of MRD positive patients decreased from 45% to 29% and 16%, respectively. We also found that the MRD status is prognostic for CIR, RFS and OS in the favorable and adverse ELN risk categories with a trend in the intermediate risk category (see Supplementary Figures S6-8) and when looking at SNVs and indels separately (see Supplementary Figures S9-10).

Figure 4. Prognostic effect of MRD on CIR, NRM, and OS if measured in peripheral blood or in bone marrow.

(A) CIR and NRM by competing risk analysis for MRD positive (n=24) and negative (n=32) patients, in whom MRD was quantified in peripheral blood (PB) (B) CIR and NRM by competing risk analysis for MRD positive (n=19) and negative (n=21) patients, in whom MRD was quantified in bone marrow (BM) (C) OS for MRD positive (n=24) and negative (n=32) patients, in whom MRD was quantified in peripheral blood (PB) (D) OS for MRD positive (n=19) and negative (n=21) patients, in whom MRD was quantified in bone marrow (BM)

In addition, we evaluated the clinical course of the 12 patients with a VAF >5% in the CR sample before alloHCT, who were initially excluded from analysis. Baseline characteristics and genetic profiles were similar among these patients and the remaining 96 patients and were similar when compared to the 43 MRD positive and the 53 MRD negative patients (Supplementary Tables S7 and S8). CIR and NRM of these 12 patients were similar to MRD positive patients (8 patients relapsed after alloHCT), while OS had an intermediate course between MRD positive and negative patients (Supplementary Figure S11). Combining the 43 MRD positive patients and the 12 patients with a VAF >5% in CR into one group and comparing these 55 patients to the 53 MRD negative patients confirmed the strong negative prognostic effect of persisting mutations in CR before alloHCT (Supplementary Figure S12, Supplementary Tables S9 and S10).

In a subset analysis of 19 patients we used DNMT3A as MRD marker. DNMT3A was positive in 15 patients and negative in 4 patients. All positive patients had VAFs above 5% (range 6-46.8%). Of the 15 DNMT3A positive patients, 9 were positive using the alternative MRD marker and only 5 of the 15 patients relapsed. CIR/NRM, RFS and OS were similar between patients with and without persistence of DNMT3A in CR before alloHCT (data not shown). In summary, NGS-MRD positivity is associated with a poor prognosis independently of other prognostic risk factors.

Accuracy of NGS-MRD and clonal evolution

The overall positive predictive value (PPV) and negative predictive value (NPV) of the MRD result to predict relapse were 62.8% and 84.9% in our cohort of 96 patients, respectively, with an overall accuracy of 75%. The lower PPV may reflect the fact that patients received an aggressive treatment after the MRD test, which may have eradicated the disease in some patients. Previous studies showed that mutations in some genes are more stable between diagnosis and relapse than in other genes.^26,27 To determine which genes might be preferred over others as MRD markers, we compared the PPV and NPV for genes that were studied in at least 6 patients. IDH1, IDH2, STAG1/STAG2 and KRAS were associated with a high NPV, suggesting that a negative MRD result for these genes reliably predicts a low risk of relapse. KRAS, NRAS, SF3B1 and TP53 mutations were associated with a moderate to high PPV, suggesting that a positive MRD result for these genes reliably predicts a high risk of relapse (Supplementary Table S11).

Sixteen patients were MRD positive in CR before alloHCT but did not relapse after a median follow-up of 3.43 years (range 0.33 – 8.9 years). In 9 of these patients MRD was measured in peripheral blood and in seven patients it was measured in bone marrow before alloHCT. AlloHCT may have eradicated residual disease in these patients, whose characteristics are listed in Supplementary Table S12. This explanation was supported by evaluating the MRD marker at day 90 after alloHCT: In thirteen relapse-free patients with an available sample post alloHCT the MRD marker was negative at day 90 after alloHCT (data not shown).

To identify potential markers of MRD positive patients which may predict sensitivity to alloHCT we compared clinical and molecular characteristics of the 16 MRD positive patients who never relapsed after alloHCT with all other patients. MRD positive patients who never relapsed after alloHCT more often had received only one cycle of chemotherapy before alloHCT, suggesting that one cycle of chemotherapy was not sufficient to eradicate MRD in these patients, as it would be expected, while these patients were chemosensitive and cleared their residual disease with the additional chemotherapy applied during conditioning treatment. MRD positive patients who never relapsed after alloHCT more often had an HCT-CI score >2, by trend a lower WBC count and a higher platelet count at diagnosis (Supplementary Table S13). They also more often had mutations in BCOR, RUNX1, SETBP1, SMC3 and by trend less frequently had mutations in FLT3-ITD compared to all other patients (Supplementary Table S14).

Eight patients were MRD negative in CR but relapsed after alloHCT (patients with false negative MRD result, patient characteristics are shown in Supplementary Table S15). For 6 of these patients the relapse sample was available. In five patients (83%) the reason for false negative MRD was clonal evolution leading to loss of the MRD marker at relapse (Supplementary Figure S13). For one patient the MRD marker (FLT3) was missed in CR, while it was present at relapse.

In summary, our MRD analysis had a higher negative than positive predictive value, suggesting that patients with a low risk of relapse can be more reliably detected by a negative MRD result. Patients with only one cycle of chemotherapy before alloHCT more often had positive MRD than patients with two or three cycles of chemotherapy. Residual leukemic clones with BCOR, RUNX1, SETBP1 and SMC3 mutations appeared sensitive to alloHCT.

Discussion

We provide evidence that NGS-based MRD monitoring can be applied to a large proportion of AML patients using almost any available molecular aberration and that it is highly predictive of relapse and survival when assessed in CR prior to alloHCT. This error-corrected sequencing approach allowed us to decrease the lower limit of detection to 0.005% (e.g. mutant IDH1) and showed a median background error of 0.0071% and a median sensitivity threshold of 0.015%. Prior published studies using NGS for MRD or mutational clearance monitoring have looked at a detection level that was significantly higher, e.g. VAF of 2-2.5%.^28–30 A few recent studies have also optimized the NGS detection level below 2%.^31–34 It was shown that ultra-sensitive NGS improves relapse prediction.³¹ In order to overcome the general NGS error rate of up to 1% and to reliably increase the sensitivity of NGS below that level, we had to adjust and refine the NGS sequencing approach. These refinements include the use of NGS-grade PCR primers with random barcodes for error correction,^19,21,35 a proof reading polymerase, avoidance of barcode contamination in consecutive sequencing runs, and standardized bioinformatics approaches for error correction that can be automated (Figure 1). Compared to MRD with multi-parametric flow cytometry (MFC), the NGS-based approach can be more easily standardized, which is important for routine clinical use.³⁶ Getta et al. showed that flow-MRD performed similar to NGS MRD using a 5% VAF cut-off.³⁰ However, a recent study in adult AML also proved that NGS was more sensitive compared to flow-MRD.³³

Our analysis included peripheral blood and bone marrow as specimens for MRD assessment. We received a similar percentage of MRD positive patients irrespective of the specimen type suggesting that peripheral blood could also be used for NGS based MRD monitoring as already shown for NPM1. ³ But further studies directly comparing both specimen types are needed to assess the level of achievable sensitivity.

A high sensitivity of error-corrected NGS-MRD assessment was confirmed by comparing NGS-MRD with the current gold standard of qRT-PCR in NPM1 mutated patients. NGS-MRD may become useful for rare NPM1 mutations, as only one primer pair is required to cover all potential NPM1 mutations. More importantly, this technique can be applied for the large group of AML patients that do not carry a NPM1 mutation. Ninety three percent of our patients had at least one mutation other than NPM1 or DNMT3A mutations using a custom myeloid sequencing panel with 46 genes. This is consistent with other reports,^7,37 but may be improved to nearly 100% by exome sequencing.³⁸ In 10% of our patients the MRD marker persisted in morphologic CR with a VAF >5%, which may indicate persistence of the mutation in differentiated leukemic cells,³⁹ pre-existing clonal hematopoiesis,^14–17 or germline origin.⁴⁰ We had excluded DNMT3A mutations upfront as they are often associated with persistence of an ancestral clone,^19,41 and are not predictive for MRD analysis.^12,13,33,42 Other genes, which had been associated with persistence of an ancestral clone with leukemic potential persisted with high VAF in our analysis (VAF >5% in CR before alloHCT: ASXL2, CBL, IDH1, IDH2, PPM1D, STAG2, TET2), while other persisting gene mutations previously had not been associated with clonal hematopoiesis (CUX1, ETV6).¹¹ We and others have proven that patients with persistence of an ancestral clone had clonal hematopoiesis several years before AML was diagnosed.^19,41 However, mutations in IDH1, IDH2 and TET2 were reliable MRD markers in other patients in our cohort, suggesting that mutations in genes associated with clonal hematopoiesis may provide important information about the disease course. This underscores that clonal hematopoiesis influences MRD detection in a complex manner.⁴² Further studies are needed in order to guide us on how to interpret persistence of an ancestral clone with leukemic potential for MRD detection. The introduction of treatments leading to maturation of leukemic blasts e.g. IDH inhibitors might further complicate the VAF cut-off. More data are needed to assess whether genes that indicate the persistence of an ancestral clone with leukemic potential may be used as MRD markers in cases where no other mutations are available. Our study is limited by the use of only 1-3 MRD markers per patient and not using the whole gene panel to monitor clonal evolution.⁴³ However, we focused on a high detection sensitivity aiming at a read depth of one million reads per marker, corresponding to 1/20^th of a high quality MiSeq run. Generally, NGS-MRD is advantageous because it is flexible, applicable to many patients and easy to standardize. Limitations of this technique are currently clonal evolution of the cells and the limitation to further increase the sensitivity of the assay. While the ideal turnaround time for this technique is 5 days, in real life several factors including number of samples can influence the time for analysis leading to a realistic turnaround time of 2-3 weeks. Germline mutations and clonal persistence can be clearly identified and do not pose a major hurdle.

In summary, we show that sensitive NGS-based MRD is widely applicable to AML patients, is highly predictive of relapse and survival when measured in CR before alloHCT, and may help refining transplant and posttransplant management in AML patients.

Key Points.

1. Error-corrected NGS-MRD can be applied to the majority of AML patients with high sensitivity.

2. NGS-MRD analysis in CR before alloHCT is highly predictive for outcome after alloHCT