Hi-MC: a novel method for high-throughput mitochondrial haplogroup classification

Introduction

Human mitochondrial DNA (mtDNA) consists of a double-stranded, circular chromosome that spans 16,529 base pairs and encodes 22 transfer RNAs, two ribosomal RNAs, and 13 proteins that are part of the oxidative phosphorylation enzyme complexes. Compared with nuclear DNA, unique characteristics of mtDNA include uniparental (i.e., matrilineal) inheritance, lack of recombination, high copy number, and a high mutation rate. These characteristics make mtDNA a powerful tool for investigations in multiple disciplines including population and medical genetics, molecular anthropology, and forensics (Chaitanya et al., 2014). Strong evidence exists supporting the involvement of mtDNA variation in human disease phenotypes, underscoring the importance of integrating the mitochondrial genome into genetic association studies. Evidence includes the association of mtDNA single nucleotide polymorphisms (SNPs) and mitochondrial haplogroups with a number of phenotypes encompassing cancer, neurologic, ocular, cardiovascular, and metabolic traits (Fetterman et al., 2013; Hudson et al., 2014; Shamnamole et al., 2013; Mitchell et al., 2014a; Van der Walt et al., 2004; Wallace, 2013).

Mitochondrial haplogroups are collections of similar combinations of mtDNA SNPs inherited from a common ancestor. These haplogroups are formed via the sequential accumulation of mutations through the maternal lineage. As a result of population migration, distinct mitochondrial haplogroups are associated with different continental ancestries including African, European, Native American, Asian, and Oceanic (Forster, 2004; Maca-Meyer et al., 2001; Wallace, 2013), allowing for accurate classification of maternal genetic ancestry in large datasets using a small subset of mitochondrial markers.

Currently, several methods are available for mitochondrial haplogroup classification including Haplogrep2, HaploFind, MitoTool, HmtDB, MToolBox, and Phy-mer (Calabrese et al., 2014; Fan & Yao, 2011, 2013; Kloss-Brandstätter et al., 2011; Navarro-Gomez et al., 2015; Rubino et al., 2012; Vianello et al., 2013; Weissensteiner et al., 2016). While these methods are powerful tools for mtDNA sequence analysis, including classification of mitochondrial haplogroups, most require full or partial mtDNA sequence, and some are limited in the number of samples that can be processed at once. To address limitations of existing methods we developed a high-throughput method for automated mitochondrial haplogroup classification that can accommodate large sample sizes with SNP data recorded in the widely used pedigree (PED/MAP) file format.

Using a custom panel of mitochondrial SNPs, we constructed a reduced mitochondrial phylogenetic tree, and developed an algorithm (Hi-MC) for broad classification of European, African, and Native American mitochondrial haplogroups. We then genotyped and assigned mitochondrial haplogroups to DNA samples used by the International HapMap Project (CEU, YRI, CHB/JPT, and MXL) (International HapMap Consortium, 2003; International HapMap Consortium, 2005, International HapMap 3 Consortium, 2010). To evaluate the performance of the algorithm, we compared these Hi-MC mitochondrial haplogroup classifications in DNA samples used by the HapMap Project with haplogroup classifications previously reported by HapMap. We also compared these haplogroup classifications with classifications assigned for the same DNA samples with genotypes from the custom SNP panel using Haplogrep2, the most widely used web-based application for mitochondrial haplogroup classification. As a final comparison, we assigned haplogroups via Haplogrep2 to the subset of overlapping DNA samples sequenced by the 1,000 Genomes Project (The 1,000 Genomes Project Consortium, 2010, 2012). In general, Hi-MC assignments were highly concordant with both HapMap-reported and Haplogrep2 assignments for CEU. Hi-MC haplogroup assignments were also highly concordant with HapMap-reported and Haplogrep2 YRI assignments with the latter using 1,000 Genomes data. When compared with HapMap-reported assignments, neither Hi-MC nor Haplogrep2 were highly concordant for CHB/JPT samples. However, Hi-MC and Haplogrep2 classifications were highly concordant with each other suggesting HapMap-reported assignments for this population are incorrect. Overall, these data suggest that Hi-MC and its associated custom SNP panel perform well compared with a popular classifier. Although it may not yet offer the same resolution as full sequence data, Hi-MC and its associated custom SNP list provide a user-friendly method for high-throughput classification provided in an R software package that can be easily expanded in future revisions to capture additional mitochondrial haplogroups.

Materials and Methods

Algorithm

The algorithm input is a list of mitochondrial SNP genotypes for each individual DNA sample, and the output is haplogroup classification. The revised Cambridge reference sequence is used to specify SNP positions. Phylotree, a comprehensive phylogenetic tree of human mtDNA variation displaying relationships between mitochondrial haplogroups (Van Oven & Kayser, 2009), was used as a reference to create a reduced tree of 46 common haplogroups as presented in Mitchell et al. (2014b). This reduced classification tree was converted into a node-based tree structure. Each haplogroup node has a list of associated SNPs, a parent node, and zero or more child nodes. The SNPs associated with a node define which SNP genotypes a subject must possess to belong to the corresponding haplogroup. Classification into a haplogroup also requires a subject to recursively meet the definition for the parent haplogroup. Haplogroups that require the reversion to the ancestral genotype (e.g., 10398A–10398G) are accommodated by adding a second hierarchy of required SNP genotypes.

The algorithm determines the appropriate haplogroup in a two-step process (Fig. 1). In the first step, the algorithm passes mitochondrial SNP genotype data for each subject into the root node of the tree. The algorithm checks the list of SNP genotypes against those required by the root node. If the genotypes meet the criteria for the parent node, this haplogroup is added to an accumulator. The algorithm then passes the list of SNP genotypes to each of the child nodes connected to that parent node until the tree is exhausted. Next, the algorithm ranks the list of haplogroups in the accumulator according to their distance from the root node. Any haplogroup with a path length less than that of the haplogroup with the longest path length is dropped. The remaining haplogroups, along with their path from the root node to the end node, are returned as a result.

Results

Here we compared HapMap-reported haplogroups (from ftp://ftp.ncbi.nlm.nih.gov/hapmap/mtDNA_and_chrY_haplogroups/) to haplogroups assigned by Hi-MC and Haplogrep2 (Weissensteiner et al., 2016) using a custom SNP panel we genotyped in HapMap DNA samples as part of the PAGE I Study (Crawford et al., 2015; Matise et al., 2011). We also performed similar comparisons for Haplogrep2-assigned haplogroups using 1,000 Genomes Project (The 1,000 Genomes Project Consortium, 2010, 2012) mitochondrial SNP data for overlapping HapMap DNA samples. Overall, concordance between Hi-MC with both HapMap-reported (94.4%) and Haplogrep2-assigned (93.3%) haplogroups was high for CEU (Table 1A). Concordance in CEU was highest in the comparison between Haplogrep2 assignments and HapMap-reported haplogroups (97.8%; Table 1A). All five discordant haplogroup calls between Hi-MC and Haplogrep2 assignments were a result of missing genotypes at key SNPs for the Hi-MC classification tree (Table S4; Fig. S1). We observed two discordant haplogroups when comparing Haplogrep2 assignments with HapMap-reported haplogroups. One was observed in CEU sample NA12146, which has the MT11251 and MT15452 genotypes that support the JT Haplogrep2 haplogroup assignment (Table S4). Another was for CEU sample NA12875 where Haplogrep2 assigned JT while both HapMap reported and Hi-MC assigned H despite the sample having the JT-defining MT11251 genotype (Table S4). In general, the comparisons made with Haplogrep2 using 1,000 Genomes Project data supported high concordance for all CEU comparisons (Table 1B). Neither NA12146 nor NA12875 was included in the 1,000 Genomes Project.

Among YRI, 14 HapMap samples were assigned “unclassified” by Hi-MC due to missing SNPs or missing key SNPs at the beginning of the classification tree (Table S5). Of the remaining 76 YRI samples for this comparison, 69 Hi-MC haplogroup assignments were concordant with HapMap-reported haplogroups (90.8%; Table 1A). In comparison, concordance was lower for Haplogrep2-assigned haplogroups when compared either with HapMap-reported (60.5%) or Hi-MC-assigned (59.2%) haplogroups (Table 1A). In contrast, the concordance was uniformly high across YRI comparisons when haplogroup assignments were made with Haplogrep2 using 1,000 Genomes Project data (Table 1A).

Examination of the YRI discordant and concordant haplogroup comparisons yielded several observations (Table S5). First, the discordance observed between HapMap-reported and both Hi-MC and Haplogrep2-assigned haplogroups was not supported by the genotype data. That is, HapMap-reported L1 haplogroups had genotypes supportive of L0 in the custom SNP dataset (such as MT9042T, MT9347G, and MT5442C; Table S5) as well as in the 1,000 Genomes Project dataset suggesting L0 is the correct haplogroup. Second, HapMap-reported and Hi-MC-assigned L3 haplogroups, including L3b, L3d, and L3e, were assigned N or Y haplogroups by Haplogrep2 when applied to the custom SNP genotype data despite the lack of N haplogroup-supporting genotypes (MT10398A in the Hi-MC classification tree) in these samples. The discordance was resolved when Haplogrep2 was applied to 1,000 Genomes Project data suggesting that L3 is the correct assignment for these samples (Table 1B; Table S5). The remaining discordant assignments were the result of a key missing genotype (MT10115) defining L2, resulting in misclassifications by Hi-MC compared with Haplogrep2 (Table S5; Fig. S2).

Compared with the CEU and YRI populations, we observed less concordance among the CHB/JPT samples particularly for the Hi-MC- and Haplogrep2-assigned comparisons with HapMap-reported haplogroups (Table 1A). One HapMap-reported classification was listed as “unknown” (for NA19012), and four CHB/JPT samples were unclassified due to missing SNPs at key genotypes for Hi-MC. Of the remaining 85 CHB/JPT samples, there was high agreement (84.9%) between Hi-MC- and Haplogrep2-assigned haplogroups (Table 1A). The low concordance with HapMap-reported haplogroups was mainly the result of haplogroups reported in HapMap as B, C, D, or E despite the lack of genotype support for these haplogroups (Table S6; Fig. S3). This high concordance between Hi-MC and Haplogrep2 assigned haplogroups using both the custom SNP panel data (Table 1A) and the 1,000 Genomes Project data (Table 1B) suggest that the majority of haplogroups reported by HapMap (in ftp://ftp.ncbi.nlm.nih.gov/hapmap/mtDNA_and_chrY_haplogroups/) for CHB/JPT are incorrect.

The mitochondrial haplogroups for the samples of Mexican ancestry from HapMap Phase III (MXL) have not been previously reported. Samples in this dataset include 30 trios of Mexican ancestry from Los Angeles, CA. We applied Hi-MC to determine mitochondrial haplogroups in these samples and characterized the distribution of mitochondrial haplogroups among the MXL. Due to matrilineal inheritance of mtDNA, offspring have the same mitochondrial haplogroup as their mother; therefore, offspring were excluded when calculating the frequency distribution of mitochondrial haplogroups. We also excluded three samples with unclassified haplogroups and two samples with likely incorrectly assigned haplogroups due to missing genotypes, for a total of 55 unrelated MXL samples. Overall in the MXL samples, 89.1% of mitochondrial haplogroups identified were of Native American ancestry and 10.9% were of European ancestry (Fig. 2). The distribution of haplogroups in the HapMap MXL samples is similar to the distribution of haplogroups observed in Mexican Americans ascertained for the NHANES (Mitchell et al., 2014b).

To further evaluate the performance of Hi-MC, we compared the Hi-MC mitochondrial haplogroup classifications of MXL samples to Haplogrep2-assigned classifications applied to either the custom SNP panel or the 1,000 Genomes Project data. Among the 90 MXL samples, six were unclassified by Hi-MC due to missing data at multiple SNPs. Of the 84 classified MXL samples, there was high agreement between Hi-MC-assigned haplogroups and haplogroups assigned by Haplogrep2 based on the custom SNP panel (97.6%) and the 1,000 Genomes Project data (96.6%). Two MXL samples were misclassified by Hi-MC as haplogroup L3, and both of these samples were missing 11 key SNPs for the Hi-MC classification tree.

Discussion

Using a custom panel of mitochondrial SNPs that we previously applied to participants in the NHANES datasets (Mitchell et al., 2014b), we developed Hi-MC, a method for high-throughput classification of European, African, and Native American mitochondrial haplogroup lineages. We evaluated the performance of Hi-MC, and with genotype data from the custom SNP panel, demonstrate that Hi-MC performs comparably to the widely-used tool Haplogrep2. While Haplogrep2 is an excellent tool for mitochondrial haplogroup classification that accepts either sequence or SNP genotype data, it was developed primarily for sequence level data. The ability to alternatively genotype a relatively small number of SNPs allows for rapid haplogroup classification in a large number of genetic samples.

Mitochondrial SNPs captured by standard genome-wide genotyping arrays vary widely, and often the SNPs on these arrays are not informative for haplogroup determination. Hi-MC uses a defined panel of mitochondrial SNPs for classification of mitochondrial haplogroups eliminating the need for time-consuming SNP selection. Additionally, the relatively small number of SNPs in the custom panel makes Hi-MC particularly useful for large data sets where full mitochondrial genome sequencing is cost-prohibitive. As examples, approaches like Hi-MC promise to be of use to large biobank and cohort efforts such as the Million Veteran Program (Gaziano et al., 2016) and the UK Biobank (Sudlow et al., 2015), both of which continue to rely on cost-effective array-based assays rather than cost-prohibitive sequencing to generate genome-wide and mitochondrial data on hundreds of thousands to a million participants.

Hi-MC employs the commonly used PED/MAP file format as the input. There are a number of software programs that make use of the PED/MAP format, including PLINK (Purcell et al., 2007) which is widely used for analyzing genotypic data. Thus, in contrast to Haplogrep, many Hi-MC users will not have to reformat data prior to use. Additionally, Hi-MC is an R-based software package that can be downloaded and run locally allowing for memory limits that are dependent on the machine where R is being run, which grants greater flexibility in the number of samples that can be processed at once. Once samples have been classified using Hi-MC, figures or tables displaying haplogroup frequencies can be easily generated via other R packages such as ggplot2 (Wickham, 2009).

We determined that Hi-MC performs well with samples of European, African, and Native American descent. Even though the Hi-MC custom SNP panel was not designed to specifically resolve Asian-specific haplogroups, we present evidence that Hi-MC is comparable to Haplogrep2 in haplogroup assignments for samples of Asian maternal lineage. Compared with other maternal lineages, the Asian branches of the mitochondrial phylogenetic tree are not as well-defined as other parts of the tree despite recent progress made in characterizing the phylogeny of Asian mtDNA (Kivisild et al., 2002; Kong et al., 2006). Another known challenge is that a major Asian haplogroup (B) is defined primarily by a deletion (8,281–8,289), a mutation not targeted by genotyping arrays. As more mtDNA sequences are obtained from individuals of Asian descent the phylogeny of mitochondrial genetic variation will be better understood. Future versions of Hi-MC will be updated to incorporate additional knowledge regarding subjects of Asian descent.

We applied Hi-MC to the HapMap Phase III MXL samples as the mitochondrial haplogroups for these participants have not been previously reported. The haplogroup distribution observed in the HapMap Phase III MXL samples is somewhat similar to the recently reported Haplogrep2-generated distribution for the MXL samples sequenced as part of the 1,000 Genomes Project (Rishishwar & Jordan, 2017). In this newer reference dataset, the most common reported haplogroup is A (25%) followed by B (15%) and C (9%) (Rishishwar & Jordan, 2017) compared with a higher A (A2) frequency in the present study (40%; Fig. 2). Overall, the distribution of Native American and European haplogroups in the MXL samples from HapMap Phase III is similar to the distribution observed in the NHANES Mexican American samples (Mitchell et al., 2014b). No African lineage mitochondrial haplogroups were identified among the HapMap MXL samples. This differs from the NHANES Mexican Americans in which 4.4% had mitochondrial haplogroups of African ancestry (Mitchell et al., 2014b). The lack of African haplogroups in the HapMap MXL samples is likely due to the small sample size and the regional ascertainment of these samples. While the NHANES samples were collected from across the United States, the HapMap Phase III MXL samples were ascertained solely from Los Angeles, CA, therefore are likely to be more homogeneous.

While there are several benefits to Hi-MC, there are some limitations. Currently, Hi-MC employs a reduced mitochondrial phylogenetic tree for classification. As a result, it is currently limited to classification of the major haplogroups of European, African, and Native American lineages, and requires that SNPs from the described custom panel be genotyped. While this panel was customized for populations expected for the PAGE I study, it is notable that several SNPs in this panel (MT1736, MT2092, MT3552, MT4883, MT10400, MT11177, MT11251, MT11719, MT12007, MT12308, MT12705, MT13368, MT14766) overlap with previously published panels (Chaitanya et al., 2014; Van Oven, Vermeulen & Kayser, 2011), suggesting the potential for both greater resolution and generalizability in future extensions of Hi-MC. Additionally, because the method relies on a limited number of SNPs, it is not very robust to missing genotype data. And, while Hi-MC has the ability to classify mitochondrial haplogroups at a broad level, it currently cannot capture sub-haplogroups at finer resolution. As such, in instances where sequence level data are available another method for mitochondrial haplogroup classification, such as Haplogrep2, would be more appropriate.

Despite these limitations, Hi-MC offers several advantages including a defined panel of mitochondrial SNPs that is used in conjunction with the software for mitochondrial haplogroup classification. Hi-MC utilizes PED/MAP files for a user-friendly input file format, saving time and reducing opportunities for errors to be incorporated into the data. Also, Hi-MC is implemented in the commonly used statistical software environment R allowing for classification of relatively large sample sizes, as well as the ability to easily utilize other available R packages for visualization of results.

Conclusions

We have developed a custom SNP panel and algorithm for mitochondrial haplogroup classification. The algorithm, Hi-MC, is implemented in R and makes use of PED/MAP file format for data input. We evaluated the performance of Hi-MC and demonstrate that classifications are comparable to the widely-used tool Haplogrep2. Hi-MC offers an algorithm that leverages a validated mtDNA SNP panel for mitochondrial haplogroup classification and is particularly valuable for studies in which sequencing is not feasible.

Supplemental Information