Detection of mitochondrial DNA variants at low level heteroplasmy in pediatric CNS and extra-CNS solid tumors with three different enrichment methods

Abstract

The mitochondrial genome is small, 16.5kb, and yet complex to study due to an abundance of mitochondria in any given cell or tissue. Mitochondrial DNA (mtDNA) mutations have been previously described in cancer, many of which were detected at low heteroplasmy. In this study we enriched the mitochondrial genome in primary pediatric tumors for detection of mtDNA variants. We completed mtDNA enrichment using REPLI-g, Agilent SureSelect, and long-range polymerase chain reaction (LRPCR) followed by next generation sequencing (NGS) on Illumina platforms. Primary tumor and germline genomic DNA from a variety of pediatric central nervous system (CNS) and extra-CNS solid tumors were analyzed by the three different methods. Although all three methods performed equally well for detecting variants at high heteroplasmy or homoplasmy, only LRPCR and SureSelect-based enrichment methods provided consistent results for variants that were present at less than five percent heteroplasmy. We then applied both LRPCR and SureSelect to three successive samples from a patient with multiply-recurrent gliofibroma and detected a low-level novel mutation as well as a change in heteroplasmy levels of a synonymous variant that was correlated with progression of disease.

Implication: This study demonstrates that LRPCR and SureSelect enrichment, but not REPLI-g, followed by NGS are accurate methods for studying the mtDNA variations at low heteroplasmy, which may be applied to studying mtDNA mutations in cancer.

Introduction

The etiology of mitochondrial genomic disorders as well as the role of mitochondria in cancer involves both nuclear-encoded genes and maternally inherited mitochondrial DNA (mtDNA).¹ Unlike the nuclear genome, there are numerous mitochondria and hundreds to thousands of copies of the mtDNA genome per cell, which contributes to intra- and intercellular heterogeneity and heteroplasmy (the fraction of mutant mtDNA genomes compared to all mtDNA genomes in a tissue or cell).² Mitochondrial DNA (mtDNA) mutations have been found causal to a number of classic mitochondrial disorders^3,4 and also contribute to tumorigenesis^5,6. Mitochondrial metabolism has been linked to cancer since Oto Warburg described the Warburg effect in 1927.^7–9 Warburg observed that, even in the presence of oxygen, cancer cells prefer glycolysis over the more energy efficient oxidative phosphorylation in the mitochondria. He hypothesized that the glycolysis pathway generates precursors that are used by cancer cells to up-regulate biosynthesis and cellular proliferation, thus creating a link between dysfunctional mitochondrial metabolism and cancer. Additionally, there are other aspects of mitochondrial biology that are important in cancer including mitochondrial biogenesis and turnover, fission and fusion dynamics, oxidative stress, variation in mtDNA sequence and abundance, and a complex signaling mechanism through the mitochondrial unfolded protein response (UPR^mt) which activates transcription of nuclear encoded mitochondrial chaperones that rescue dysfunctional mitochondria and may provide a survival advantage to cancer cells.^10–15

Somatic mtDNA mutations are present in a variety of cancers, as best illustrated by several recent pan-cancer studies of adult and pediatric cancers.^16,17 Small-scale studies have demonstrated mtDNA mutations in a variety of pediatric malignancies including medulloblastoma^18,19, other pediatric brain tumors²⁰, pediatric acute myelogenous leukemia (AML)²¹, and neuroblastoma.²² Most recently, we described the landscape of germline and somatic mtDNA mutations in a diverse selection of pediatric cancers using whole genome sequencing (WGS) data from 616 primary tumor and matched normal samples.²³ We analyzed data from various hematologic malignancies, solid tumors (including 14 retinoblastomas and 13 rhabdomyosarcomas), and CNS tumors (142 total). We identified distinct somatic mtDNA profiles of different tumor types, the presence of multiple patients with germline pathogenic variants for classic mitochondrial diseases, four hotspot LoF somatic mtDNA mutations and one hotspot somatic mtDNA tRNA mutation. Our study established 391 mtDNA mutations in 284 tumors including a total of 45 LoF mutations which clustered in MT-COX3, MT-ND4, and MT-ND5, as well as a mutation hotspot in MT-tRNA-MET. We also observed distinct mtDNA mutation profiles in children compared to adults, with children having more variants at low heteroplasmy.

Prior studies have shed light on the high false positive rate of mtDNA mutations in existing publications.^24,25 In the present study, we explored three different methods for enriching and sequencing the mitochondrial genome that could be used to prospectively analyze primary pediatric tumors for mtDNA variants, which is very important given the ever-increasing recognition of the contributions of mtDNA mutations to tumorigenesis. The methods that we compared include REPLI-g²⁶, LRPCR²⁷, and a hybridization based capture method for the mitochondrial genome using custom Agilent SureSelect probes²⁸. REPLI-g requires a very small genomic DNA input of 5ng, LRPCR requires a 10-fold greater amount of genomic DNA with a mean quantity of 50ng used in the literature, while SureSelect requires 200ng of genomic DNA on average. Both LRPCR and REPLI-g work only with fresh frozen tissue samples and formalin-fixed paraffin-embedded (FFPE) is not recommended as the starting material. Although performance is suboptimal with FFPE, such samples can be used for SureSelect based mtDNA enrichment and sequencing. Our laboratory has a strong clinical focus that requires systemic validation with the intent of generating reproducible results. Thus, we compared these three methods to validate our work and to determine the best method to use for future studies.

In this manuscript, we describe the most accurate and efficient method for enriching and sequencing the mtDNA genome in tumor samples. The methods were compared using DNA isolated from blood, orbital fat pad-derived mesenchymal cells, buccal swab or primary tumor fresh frozen tissue from pediatric patients with CNS tumors, retinoblastoma, rhabdoid tumor, and sarcoma. We demonstrate the potential use of LRPCR as well as Agilent SureSelect methods followed by high depth NGS to characterize the mtDNA in primary tumors, as illustrated by results from the primary, first recurrence, and second recurrence tumors of a gliofibroma patient.

Materials and Methods

REPLI-g followed by NGS

The mtDNA genome was amplified using REPLI-g Mitochondrial DNA Kit (Qiagen, Germantown, MD).^26,29 REPLI-g provides enrichment of the mitochondrial DNA up to 40 million-fold through rolling circle amplification. One μl of template genomic DNA (20ng/μl) was adjusted to 20μl with RNAse-free water supplied from the kit. Of note, as little as 5ng of DNA can be used. The amplification mix was prepared using 27μl REPLI-g mt Reaction Buffer and 2μl REPLI-g Human mt Primer Mix. A total of 29μl of the amplification mix was then added to the DNA, incubated for 5 minutes at 75°C, and then cooled down to room temperature (25°C). One μl of REPLI-g Midi DNA Polymerase was added to the DNA and the sample was incubated at 33°C for 15 hours. The REPLI-g Midi DNA Polymerase was inactivated by heating the sample for 3 minutes at 65°C. The REPLI-g amplified mitochondrial DNA product was then sheared using Covaris M220 focused ultrasonicator. Library preparation was completed using KAPA Hyper Prep kit (KAPA Biosystems, Wilmington, Massachusetts). Sequencing was then performed using NextSeq™ 500 (Illumina, San Diego, CA) or MiSeqDx™ (Illumina).

Long Range PCR followed by NGS

The Long Range PCR (LRPCR) methods described by Zhang et al were utilized with slight modification as described below. ²⁷ The forward primer was mt16426F (5’-CCGCACAAGAGTGCTACTCTCCTC-3’) and the reverse primer was mt16425R (5’GATATTGATTTCACGGAGGATGGTG- 3’). PCR was performed using TaKaRa LA Taq Hot Start Polymerase kit (TaKaRa). We used 40–50ng of genomic DNA isolated from fresh frozen tissue and germline samples as template in a 50μL PCR reaction. However, up to 100ng of DNA may be needed depending on the type of tissue source. The PCR conditions included 94°C incubation for 2 minutes followed by 10 cycles of PCR with 10 seconds of denaturation at 94°C and 15 minutes of annealing at 65°C. This was followed by an additional 15 cycles of PCR under the same denaturation and annealing conditions with incremental increase of the annealing duration by 20 seconds per cycle. There was one final extension cycle at 72°C for 10 minutes followed by infinite hold at 4°C. The DNA product was sheared, and library preparation and sequencing were performed as detailed above for the REPLI-g enrichment protocol.

SureSelect hybridization based capture and sequencing

We previously designed SureSelect capture probes for the mtDNA genome in collaboration with Agilent (Santa Clara, California).²⁸ This bait and blend probe design involves sequences targeting the entire mtDNA genome created by standard 1X tiling across the entire mitochondrial DNA genome. The baits for the mtDNA genome were factory blended into nuclear baits at equimolar ratio or reduced concentration.²⁸ These probes have been validated in conjunction with the standard SureSelect exome capture probes for dual-genome sequencing for our clinical whole-exome sequencing (WES) test.³⁰ The standard laboratory methodology for WES library preparation, sequencing, and a modified bioinformatics pipeline were used in this study for sequencing the exome and the mtDNA genome of selected pediatric cancer samples.³⁰

For bioinformatic analysis of mtDNA mutations, please refer to methods used in the Triska et al. study.²³ We used a variant allele frequency (VAF) cutoff of 2.5%, all lower heteroplasmy variants were filtered out.

Results

Detection of known mtDNA mutations using LRPCR and SureSelect methods

To validate our methods, libraries were prepared from clinical samples with known pathogenic mtDNA variants of various heteroplasmy levels. Sequence data generated from these libraries were compared to clinical data previously generated by either Baylor College of Medicine (BCM) or GeneDx, Inc. DNA from the same four clinical samples with known mtDNA variants was used for all three methods. We were able to detect all the variants reported previously with a positive control method with both SureSelect exome sequencing and LRPCR/NGS at similar heteroplasmy levels between the different platforms (Table 1). The one outlier is sample 1 in which a mutation in MT-ND5 (m.13513G>A) was seen at 20% heteroplasmy using the reference method while we observed this mutation at 47% using SureSelect and 50% using LRPCR.

Table 1. Comparison of results from four samples with known mtDNA mutations using different methods.

The same four samples were sequenced in three clinical laboratories. Consistent heteroplasmy levels are demonstrated with reference techniques as well as SureSelect and LRPCR in our laboratory.

No.	Positive Control Method	Sample Type	Gene	DNA Change	Reference Heteroplasmy	SureSelect Heteroplasmy	LRPCR/NGS Heteroplasmy
1	mtDNA point mutation quantification	Blood DNA	MT-ND5	m.13513G>A	20%	47%	50%
2	GeneDx mitochondrial genome sequencing	Blood DNA	MT-ND5	m.13513G>A	61%	62%	69%
3	Targeted mtDNA analysis by massive parallel sequencing (MitoNGS)	Blood DNA	MT-ND3	m.10191T>C	76%	64%	75%
4	Dual genome panel by massive parallel sequencing (MitoNGS)	Blood DNA	MT-ND4	m.11777C>A	62%	59%	60%

Comparison of REPLI-g and LRPCR followed by NGS in retinoblastoma

In our initial comparison of the REPLI-g and LRPCR methods, we used retinoblastoma matched tumor (fresh frozen tissue) and normal (orbital fat pad-derived mesenchymal cells or buccal swab) samples from 8 cases (8 tumor and 8 matched normal). REPLI-g enrichment by NGS resulted in many variants compared to LRPCR (Fig. 1) (Supplemental Table 1–4). Using REPLI-g, 31% of all variants found in the 16 samples had a VAF less than 5%, while only 2.9% of mtDNA variants found in the same 16 samples had a VAF less than 5% with the LRPCR method (Supplemental Fig. 1). Additionally, the REPLI-g method resulted in a mean number of 7 frameshift variants per tumor sample while no frameshift variants were seen with LRPCR. Thus, we suspected that variants detected via REPLI-g at low VAF (< 5%) were potential artifacts. Overall, the landscape of mtDNA variants in retinoblastoma tumors was bland with few interesting (e.g. no LoF, few heteroplasmic) variants. With LRPCR we identified a total of three somatic mtDNA variants (D-loop, MT-ATP8, MT-TRNE/tRNA-Glu) in two of the eight sample pairs at 4%, 2.6%, and 3% VAF respectively. These results were consistent with the lack of LoF variants observed in the 14 retinoblastoma cases in our WGS data mining study which included many other pediatric cancer cases.²³

Figure 1. Distribution of mtDNA variants in retinoblastoma tumor samples and matched normal samples.

This is data from 8 tumor and 8 matched normal samples obtained with two different methods, LRPCR (Long Range PCR) and REPLI-g RCA (rolling circle amplification). High concordance is visually observed at high variant allele frequency (VAF) between the two methods, however numerous artifactual variants are observed with REPLI-g at low VAF. Outer scale represents heteroplasmy level. Each color-coded dot corresponds to a type of variant.

Comparison of LRPCR, SureSelect, and REPLI-g in a cohort of CNS and solid tumors

Next, we enriched the mtDNA of 13 fresh frozen primary tumor samples using SureSelect, REPLI-g, and LRPCR followed by NGS (Fig. 2). Those 13 samples included two glioblastomas, two anaplastic pleomorphic xanthoastrocytomas (PXA), one pilocytic astrocytomas, one diffuse anaplastic astrocytoma, one choroid plexus carcinoma (CPC), one embryonal tumor with multilayered rosettes (ETMR), one medulloblastoma, one atypical meningioma, one alveolar rhabdomyosarcoma of the pelvis, one embryonal rhabdomyosarcoma of the temporal skull base, and one soft tissue sarcoma NOS (not otherwise specified) of the lung. We observed very high concordance between the LRPCR/NGS and SureSelect exome-based results. However, we detected many variants with the REPLI-g enrichment method followed by NGS at low variant allele frequency (<5%) (Fig. 2, Supplemental Table 5). We interpreted the concordant calls between LRPCR and SureSelect to be valid calls (Supplemental Table 6, 7) and the discordant calls produced by REPLI-g to be artifacts. Interestingly, in one of these 13 samples (medulloblastoma) we initially detected 8 additional consecutive variants at the end of the D-loop region with LRPCR that were not identified by SureSelect. On closer examination, we determined that these variants correspond to an 8 base pair deletion (16,360–16,367) present in 6400 reads at 19% mean heteroplasmy level (Supplemental Table 8). It will be important in future studies to test the likely contribution of this small deletion to tumorigenesis. On the other hand, this case points to the advantage of the LRPCR/NGS method given the uniformity of the coverage profile and the higher likelihood of detecting structural variations.

Figure 2. Distribution of mtDNA variants in a cohort of primary CNS and solid tumor samples using three different methods.

Data is from a total of 12 primary tumor tissue fresh frozen samples. SureSelect, an exome-based method, is concordant with LRPCR (Long Range PCR) while with REPLI-g numerous artifacts are observed at low VAF.

Comparison of LRPCR, SureSelect, and REPLI-g in rhabdoid tumors

We analyzed six rhabdoid tumor samples (three fresh frozen tumor and three matched normal blood) with REPLI-g, SureSelect and LRPCR enrichment for mtDNA, followed by NGS (Fig. 3). Of those, one pair was from a paraspinal rhabdoid tumor while the other two pairs were from atypical teratoid/rhabdoid tumor (AT/RT) of the brain. We again observed a high level of concordance between LRPCR and SureSelect (Supplemental Table 10, 11) with numerous variants at VAF < 5% seen with REPLI-g/NGS (Supplemental Table 9). Overall, in the six rhabdoid samples, there was a lack of heteroplasmic mtDNA variants in the coding, tRNA, and rRNA regions. However, there were multiple heteroplasmic mtDNA calls in the D-loop region detected by both SureSelect and LRPCR methods (Fig. 3).

Figure 3. Distribution of mtDNA variants in 3 rhabdoid tumor samples and 3 matched normal blood samples by three different methods.

Each sub-panel demonstrates data from 6 rhabdoid tumor samples (3 tumor and 3 matched normal samples). SureSelect, an exome-based method, is concordant with LRPCR (Long Range PCR) while with REPLI-g numerous artifacts are observed at low VAF.

Clonal evolution of mtDNA mutations in tissue from a multiply-recurrent CNS gliofibroma

Given the knowledge of the robustness of the LRPCR or the SureSelect methods, we studied the mtDNA mutations in serial tissue biopsies from a single patient, a 20-month old male with pediatric CNS gliofibroma who unfortunately had multiple recurrences, first with the LRPCR method and then with the SureSelect method for validation purposes. This patient had a BRAF^V600E mutation and biallelic deletion of CDKN2A in the initial tissue sample³¹. The patient was initially treated with targeted therapy, using a BRAF inhibitor (vemurafenib) and MEK inhibitor (trametinib), with an excellent response, but had a recurrence after 15 months of therapy. At the first recurrence, the tumor had acquired two additional nuclear mutations, in PTPN11 and PDGFRA. The patient then had radiation therapy followed by a second recurrence outside of the radiation field which contained two additional deleterious nuclear mutations in TP53 and PIK3CA³¹. Considering the clonal evolution that we observed at the nuclear level, we wanted to see if we would detect a similar pattern in the mitochondrial DNA. We sequenced DNA from the initial fresh frozen tumor sample as well as both recurrent samples after LRPCR amplification and, although no new mutations were identified in the first recurrence, we observed a new loss of function stop-gain mutation in MT-ND2 (m.4788G>A) at 5% VAF with the second recurrence (Fig. 4). These results were replicated with our SureSelect capture followed by NGS sequencing platform. We therefore demonstrate progressive clonal evolution in the mitochondrial genome with progression of disease in this patient. All mtDNA variants detected using our SureSelect method were identical to the variants detected using the LRPCR/NGS method (Supplemental Table 12). With both methods we also observed a synonymous variant in MT-ND4L (m.10700A>G) that was present at 53% VAF in the initial tumor, at 79% VAF in the first recurrence, and 97% VAF (near homoplasmy) with the second recurrence (Fig. 4). This case clearly illustrates the potential clinical utility of mtDNA mutations as diagnostic markers.

Figure 4. Distribution of mtDNA mutations in 3 successive samples from a single patient with multiply-recurrent gliofibroma using a. LRPCR/NGS and b. SureSelect.

A novel mtDNA mutation was detected in the tissue sample from the second recurrence in MT-ND2 at 4% VAF with both methods. A synonymous variant in MT-ND4L was seen at 50% VAF in the initial biopsy, 75% VAF in the first recurrence, and near 100% VAF at the second recurrence.

We compared these results to the VAF for nuclear driver mutations and we observed BRAF^V600E (c.1799T>A) at 28% VAF in the original sample, at 40% VAF in the first recurrence, and at 46% VAF in the second recurrence. Although there was a small increase in the BRAF^V600E mutation VAF with each recurrence, this change was small compared to what we observed with the mitochondrial DNA MT-ND4L synonymous variant. While a potential difference in tumor purity or tumor fraction in each sample may have contributed to the incremental increase in mtDNA VAF, sampling variations alone cannot fully explain the observed phenomenon.

Discussion

The primary goal of this study was to identify a sensitive and accurate method for detecting low-level variants in mtDNA in primary tumor tissue. Initially we used REPLI-g because of the small amount of DNA required, ease of setting up the PCR reaction, as well as encouraging results in terms of sensitivity and accuracy based on published studies^26,29. However, comparing the data from three methods, we determined that REPLI-g/NGS introduces many artifactual variants at low VAF, i.e. < 5%. In contrast to REPLI-g/NGS, LRPCR/NGS and SureSelect based capture and sequencing demonstrate a high level of concordance across all heteroplasmy levels. Additionally, LRPCR provides more uniform coverage across the mitochondrial genome that allows for the detection of deletions that would otherwise be missed by SureSelect, as evidenced by the small 8bp deletion in the single medulloblastoma fresh frozen tumor sample. Thus, we have shown that LRPCR is an accurate method for detecting mtDNA mutations even at low VAF when compared to SureSelect. LRPCR is not only accurate, it also provides the added benefit of detecting deletions.

One reason for the high number of artifacts at low VAF generated by REPLI-g is potential contamination of non-mtDNA genome reads from nuclear mtDNA transcripts (“NuMTs”).²⁸ NUMTs can be unique to the individual which may explain the puzzling identical low heteroplasmy variant artifacts observed in tumor and matched normal samples as illustrated by the 8 retinoblastoma sample pairs. Additionally, errors were possibly introduced due to the REPLI-g rolling circle amplification using random hexamers. This method involves repeat single-strand extension of a relatively small amount of input DNA which is amplified up to 40 million-fold and can theoretically introduce amplification errors that wind up as low-level mtDNA heteroplasmy being detected in NGS assay.

Considering the high level of heterogeneity introduced by mtDNA mutations, it is crucial that the right method is used so one can detect true mutations across all heteroplasmy levels including very low VAF without the noise introduced by artifacts. A challenge with the LRPCR method, however, is the DNA quality. Although we did not perform such comparison, it is unlikely that LRPCR would work with FFPE samples in which the DNA is highly fragmented. With lower requirement for DNA fragment size, SureSelect would present an alternative given the demonstrated reliability of the mtDNA calls. While prior studies have indicated that REPLI-g is both a sensitive and accurate method for detecting mtDNA variants²⁶, we observed many artifactual variants in our data when compared to the other platforms. While cheaper and more efficient in terms of methodology and amount of genomic DNA required, we illustrate that REPLI-g is not an accurate method for enriching the mtDNA genome specifically, particularly when studying low VAF variants. However, the discordance we observed was mainly at low VAF so one may consider using REPLI-g if only interested in identifying homoplasmic or germline mtDNA variants.

Although the number of pediatric samples used in this study is not large, these results are concordant with our previously reported results where we interrogated a much larger WGS dataset ²³. We identified a total of 45 loss-of-function (LoF) mutations in the original study of 616 tumor-normal pairs, however, no LoF mutations were detected in the 14 retinoblastomas, four CPCs, 36 low-grade gliomas, or one T-cell acute lymphoblastic leukemia. Similarly, we did not find any LoF mutations in the current 8 retinoblastomas, one low-grade-glioma (pilocytic astrocytoma) or one CPC tumor samples.

Although functional studies are required to support the hypothesis that individual mtDNA mutations or mutation patterns are important for tumorigenesis, detection of these mutations may have clinical utility. For example, our results for the multiply recurrent gliofibroma case show that clonal evolution exists on the mtDNA level as demonstrated by the emergence of a new loss-of-function mutation in MT-ND2 (m.4788G>A) at the time of the second recurrence. Additionally, the same synonymous variant in MT-ND4L (m.10700A>G) was detected at increasing heteroplasmy (from ~50% to ~75% to ~100%) with each recurrence, which demonstrates the utility of using mtDNA variants as a potential marker for disease response or progression. Considering that the MT-ND2 (m.4788G>A) mutation was seen at 4.6–5% VAF with both LRPCR and SureSelect, we show that important deleterious mutations can be detected at VAF < 5% again highlighting the importance of using a method that allows for confident calling of mutations at very low VAF without concern for extraneous false variants.

In summary, given the recognized importance of somatic mtDNA mutations to tumorigenesis, selecting the right method for enriching and sequencing the mtDNA genome is of paramount importance to successfully identifying important mtDNA mutations in different samples including pediatric cancer samples as illustrated in this study. We have shown that both LRPCR/NGS and SureSelect-based methods are accurate methods for studying the mtDNA genome, while REPLI-g introduces many artifacts at low VAF. The limitations of this method need to be considered in further studies investigating mtDNA variants in clinical disease states.

Supplementary Material

Suppl Fig1

Supplemental Figure 1. Distribution of mtDNA variants in retinoblastoma tumor samples and matched normal samples. Data from 8 tumor and 8 matched normal samples obtained with two different methods, LRPCR (Long Range PCR) and REPLI-g RCA (rolling circle amplification). VAF=variant allele frequency. T = tumor, N = normal.