Clinical impact of panel-based error-corrected next generation sequencing versus flow cytometry to detect measurable residual disease (MRD) in acute myeloid leukemia (AML)

We accrued 201 patients of adult AML treated with conventional therapy, in morphological remission, and evaluated MRD using sensitive error-corrected next generation sequencing (NGS-MRD) and multiparameter flow cytometry (FCM-MRD) at the end of induction (PI) and consolidation (PC). Nearly 71% of patients were PI NGS-MRD+ and 40.9% PC NGS-MRD+ (median VAF 0.76%). NGS-MRD+ patients had a significantly higher cumulative incidence of relapse (p = 0.003), inferior overall survival (p = 0.001) and relapse free survival (p < 0.001) as compared to NGS-MRD− patients. NGS-MRD was predictive of inferior outcome in intermediate cytogenetic risk and demonstrated potential in favorable cytogenetic risk AML. PI NGS-MRD− patients had a significantly improved survival as compared to patients who became NGS-MRD− subsequently indicating that kinetics of NGS-MRD clearance was of paramount importance. NGS-MRD identified over 80% of cases identified by flow cytometry at PI time point whereas FCM identified 49.3% identified by NGS. Only a fraction of cases were NGS-MRD− but FCM-MRD+. NGS-MRD provided additional information of the risk of relapse when compared to FCM-MRD. We demonstrate a widely applicable, scalable NGS-MRD approach that is clinically informative and synergistic to FCM-MRD in AML treated with conventional therapies. Maximum clinical utility may be leveraged by combining FCM and NGS-MRD modalities.


a. Design of smMIPS:
We reviewed somatic mutations in 393 patients of adult AML (Supplementary Figure 1) diagnosed at the Tata Memorial Centre as well as mutations in TCGA AML cohort (www.cbioportal.org). Based on this data we designed a 35 gene myeloid hotspot panel comprising of 302 single molecule molecular inversion probes (smMIPS). The genes covered by this "hot-spot" panel can be seen in Supplementary  Table 1. The panel comprised of a main panel and an add on module which covered uncommon mutations. The latter panel was added if mutations (covered by addon panel) were present at diagnosis. smMIPS were designed using MIPgen 1 software with the parameters -max_capture_size 162, -min_capture_size 152, -logistic_priority_score 0.5. Extenstion arm length parameters were set between 18-21 and ligation arm length between 21-24. Each smMIP was designed to include a four basepair (bp) unique molecular identifier (UMI) at each end (total of 8bp degenerate nucleotide sequence per smMIP).

a. Sequencing:
Initial standardization and balancing experiments using NA12878 control DNA were performed on a MiSeq (standard V2 flow cell 150PE chemistry). Once the assay was standardized, error modelling, limit of detection and MRD detection experiments were performed on multiple S4 flow cells of a NovaSeq 6000 using 150PE chemistry.

b. Data Analysis:
The bioinformatics approach was similar to that published by Waalkes et al 2 with a few modifications. Demultiplexing was performed using bcl2fastq-v2.17. Adapter sequences, smMIPs backbone and reads less than 53bp were trimmed using fastq-mcf tool of ea-utils (https://expressionanalysis.github.io/ea-utils/). Paired end assembly was carried out using PEAR (v0.9.8). 3 A custom script was used to trim, concatenate the 4bp UMI at the 5' and 3' end of assembled read and add it to the read header. Reads were mapped to the human genome (build hg19) using bwa-0.7.12 4 and pre-processed using SAMtools-1.1 5 . For computational efficiency, mapped reads were split by chromosome and into files sourced from reads mapping to individual smMIPS based on their genomic start and stop coordinates. Each of these files is processed individually for downstream variant calling. All reads originating from a single UMI were discarded (singleton reads). The rest of the reads (two or more UMI) were used to create a consensus .sam file using CallMolecularConsensusReads function of fgbio-0.4.0 with the following parameters --error-rate-post-umi=30 --min-reads=2 --min-input-base-quality 20 (https://github.com/fulcrumgenomics/fgbio). These reads were converted to fastq using picard-2.17.2 (http://broadinstitute.github.io/picard/) and mapped, sorted and indexed using bwa and samtools as described above. A .mpileup was created using samtools-1.1 and sequence variants detected using a bespoke variant caller. (type A)] were serially diluted in normal BM as seen in Supplementary Figure 5. The expected variant allele frequency (VAF) for that mutation was calculated from the original VAF found in undiluted sample. The range of the expected VAFs was from 1.25% to 0.02%. We could successfully detect VAFs in nearly all cases up to a lower limit of approximately 0.05%. The limit of detection of NGS-MRD assay was thus at 0.05% for all mutations and 0.03% for the NPM1 mutation. A lower LOD threshold was acceptable for NPM1 mutation because of the uniqueness of the indel mutation and a previous observation (based on a limit of blank study) that complex 4bp indels (seen in NPM1) are not observed for short read sequencing data. 7 6 Supplementary Figure 5: Serial dilutions of OCIAML3 and AML DNA in normal bone marrow. The expected VAF for a mutation was calculated from the original VAF found in undiluted samples.

d. Error Modelling:
A site and mutation specific error model was setup to ascertain the occurrence of variations observed in the smMIP MRD panel. We sequenced a NA12878 and four normal bone marrow controls to measure sequencing errors using smMIPS as described by Waalkes and colleagues. 2 As mentioned by Waalkes we discovered a reduction in error rates using UMI based sequencing over standard NGS based sequencing (data not shown). As described, we fitted a  distribution for each base position and probable base substitution error. We observed a higher frequency of C>T and G>A changes consistent with oxidative DNA damage (Supplementary Figure 6) occurring in template DNA before sequencing. Where no variation was detected a 1:15,000 error rate was presumed and a  distribution was modelled. For each variation observed, a site-specific p-value was annotated using these pre-calculated  distributions. Sites with p>0.005 were excluded as artefacts.
Supplementary Figure 6: Frequencies of errors observed using a site and mutation specific model.

e. Criteria for variant calling using smMIPS MRD assay:
i. Variants filtered by focussing on exonic regions (including splicing variants if any) followed by population frequency (<0.01) filtering. ii.
The variant must have been detected at baseline. iii.
Background error modelling at that site must have a P value <0.005. iv.
A minimum of 10 alternate variant reads must be present for an SNV. v.
A minimum of 3 alternate variant reads must be present for an indel. vi.
The highest VAF was taken as MRD. vii.
If the highest VAF was in a variant associated with DNMT3A, TET2 or ASXL1 (DTA mutation) it was ignored.

Detection of FLT3-MRD using a one-step PCR assay:
a. Assay Design: We observed that we could not monitor FLT3-ITD using smMIPs. To overcome this issue, we designed a one-step PCR assay that incorporated locus specific primers, dual indices and Illumina adapters in a single step (Supplementary Table 3). The primers were designed to amplify common internal tandem duplication site (chr13: 28608024 -28608353). 8 The assay was setup using 600ng of genomic DNA as template.

b. Data Analysis and Limit of Detection:
We adopted a recently described algorithm 9 for accurate detection of FLT3-ITD using next generation sequencing. We could demonstrate good correlation (Supplementary Figure 7, inset) between conventional and NGS testing for accurate detection of ITD length in 71 FLT3-ITD positive AML. We could validate this assay till a maximum ITD length of 100bp.

c. Limit of Detection of FLT3 NGS MRD assay:
In order to determine the limit of detection of this assay, we diluted a FLT3-ITD (30bp ITD) positive sample into normal BM. All dilutions were performed in triplicates (Supplementary Figure 7). We could successfully detect this mutation till a lower limit of 0.002% VAF, as shown in Supplementary Figure 7. Based on this data, we established the limit of detection of the FLT3-ITD NGS MRD assay at 0.002. All FLT3-ITD clones >1% VAF were tracked        These were detected using a previously published ultradeep sequencing based MRD assay for NPM1 mutated AML. 7 smMIPS MRD positive NPM1 mutated AML cases (45 out of 75; 60%) could be detected by the NPM1 NGS MRD assay. For the rest of the cases, NPM1 mutation was either negative or detected at a threshold below the smMIPS MRD assay limit. For smMIPS MRD positive cases the median MRD value was 0.86% as compared to 0.82% for NPM1 NGS MRD (at a LOD cut-off of 0.03%).

Outcome of patients in whom NGS MRD could not be performed (no mutations or not covered by panel)
A total of 65 patients (as detailed in supplementary figure 18) were excluded from NGS-MRD analysis as they lacked mutations or harboured mutations not covered by the NGS panel. Of these patients, 12 patients were refractory to chemotherapy. Of the rest (n=53), 11 patients have relapsed of which 7 have died. In addition, 10 have died of unrelated causes. The rest (n=36) are alive at last follow up.