Himito: a graph-based toolkit for mitochondrial genome analysis using long reads

Abstract

Understanding the genetic and epigenetic regulation of mitochondrial DNA (mtDNA) is essential for elucidating mechanisms of aging and disease. Long-read sequencing can span the mitochondrial genome and capture base-modification signals, yet analytical tools remain limited. We developed Himito, a graph-based toolkit for analyzing mitochondrial genome using long reads. Himito filters reads originating from nuclear mitochondrial insertions (NUMTs), constructs a sequence graph to represent mtDNA diversity, assembles primary haplotypes, calls variants, and analyzes 5-methylcytosine (5mC) modifications within a unified framework. Benchmarking on reference datasets shows Himito achieves superior performance in assembly and variant calling compared with existing tools. Applied to the All of Us (AoU) v8 dataset, Himito identified pathogenic mtDNA variants, revealed population-scale haplogroup diversity, and uncovered age-related genetic and epigenetic patterns. These results demonstrate that long-read sequencing, combined with graph-based analysis, enables integrated characterization of mitochondrial genomic and epigenomic variation. Himito is available at https://github.com/broadinstitute/Himito.

Himito is a graph-based toolkit for long-read mitochondrial genome analysis. It integrates NUMT filtering, haplotype assembly, variant calling, and methylation profiling, providing a comprehensive framework for studying mtDNA in aging and human disease.

Introduction

Mitochondria are essential organelles responsible for cellular energy production, metabolic regulation, and aging^1,2. Human mitochondrial DNA (mtDNA) is a 16 kb circular, double-stranded DNA molecule distinguished by its high copy number, elevated mutation rate relative to nuclear genomes, maternal inheritance, and tissue-specificity^3,4. Mitochondrial variants contribute to diverse disorders, from metabolic and neurodegenerative diseases to cancer⁵. These variants occur as homoplasmies, shared by all the mtDNA copies within a cell, or heteroplasmies, representing mixtures of mtDNA types. Characterizing the genetic variations in mtDNA is essential for decoding the mitochondrial dysfunction and related mitochondrial diseases. These variants include single-nucleotide variants (SNVs), small insertions or deletions (indels), and large structural variants (SVs). To date, over a hundred high-confidence pathogenic variants have been reported⁶, and the majority of these variants cause disease in the heteroplasmic state, when the level of mtDNA with these variants exceeds a disease-specific threshold. Thus, sensitive and accurate detection of heteroplasmy is crucial for understanding, accurate diagnosis and early intervention of mitochondrial disease⁷.

Accurate mitochondrial genome analysis is complicated by several biological and technical factors: the circular nature, the presence of heteroplasmy, and the confounding factor of nuclear-mitochondrial segments (NUMTs), fragments of mtDNA inserted into the nuclear genome⁸. The nuclear genome harbors hundreds to thousands of ancient NUMTs⁹, introducing complexity to variant calling and heteroplasmy analysis. NUMTs insertion is also an ongoing process¹⁰, occurring approximately once in every 10,000 births⁸. Short reads struggle to distinguish NUMTs-derived from true mitochondrial reads, generating false-positive variant calls and spurious heteroplasmy. Moreover, mtDNA harbors large structural variants¹¹ and variable homopolymer tracts that are poorly resolved by short reads¹². Current strategies for removing NUMTs-derived reads include the frequency-based filtering (filtered out variants with heteroplasmy <10%)³, or annotate variants using the known NUMTs sequences^12,13, yet these methods struggle for identifying genuine low-frequency mtDNA heteroplasmies or distinguishing reads from NUMTs sequences that are almost identical to the mitochondrial genome. Another persistent challenge is the controversial existence of 5-methylcytosine (5mC) modifications in mtDNA. Although several studies suggest negligible or artifactual mtDNA methylation^14–16, others report the low-level 5mC^17–19 and mitochondrial localization of DNA methyltransferases (DNMT1, DNMT3A, DNMT3B)^20–22, suggesting potential regulatory roles^23–25. These conflicting observations may stem from methodological artifacts in bisulfite sequencing²⁶, including incomplete conversion of super-coiled mtDNA and mis-mapping of methylated NUMT reads^15,27. Long read sequencing circumvents these limitations by generating native reads that preserve methylation information and span the entire mitochondrial genome without PCR amplification. Emerging literature and growing research interest suggest that the use of methylation patterns offers a promising strategy for distinguishing NUMTs-derived reads from authentic mtDNA reads^28,29.

Despite extensive development of long-read bioinformatics tools (https://long-read-tools.org/), most are primarily designed for diploid nuclear genomes rather than organelle-specific analysis. Common variant calling tools, such as DeepVariant³⁰, are not specifically designed for mtDNA variant calling and have limited power detecting heteroplasmic variants. Tools such as GATK-Mutect2 mitochondrial-mode³¹, mtDNA-server2³², MToolBox³³, are designed for short-read sequence analysis, but are not optimized for long reads. MitoHiFi³⁴ has been developed for mitochondrial genome assembly using HiFi long reads, but does not perform variant calling. Mitorsaw (unpublished) primarily focuses on mitochondrial variant calling using PacBio HiFi long reads, but has not been tested thoroughly in literature. Consequently, no existing methods integrate NUMT filtering, haplotype-resolved assembly, variant calling, and methylation profiling of mtDNA from long reads within a unified framework.

To address this gap, we present Himito, a graph-based toolkit designed for mitochondrial genome analysis using long-read sequencing. Himito offers compatibility with data produced by both Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) sequencing platforms. Himito classifies and removes NUMT-derived reads based on alignment signatures and methylation patterns, constructs sequence graphs capturing the circular topology of mtDNA, assembles haplotypes, detects both homoplasmic and heteroplasmic variants, and annotates methylation events on the resulting graph. Benchmarking across multiple reference datasets demonstrates that Himito improves mitochondrial assembly accuracy and heteroplasmic variant detection compared with existing approaches. Application to the All of Us (AoU) v8 long-read cohort further reveals population-scale mtDNA diversity, pathogenic variants, and age-related genetic and epigenetic changes, underscoring the power of integrating graph-based and methylation-aware analysis for mitochondrial genomics.

Results

Himito algorithm

Himito incorporates five key modules for long-read-based mitochondrial genome analysis (Fig. 1): Filter, Build, Asm, Call, and Methyl. Himito Filter module excludes the NUMTs contamination by identifying NUMTs-derived reads based on the whole-genome supplementary alignment signature (SA tags) and read-level methylation patterns (Fig. 1a). Himito Build module constructs a sequence graph by mapping unique k-mers in the standard human reference mitochondrial genome (rCRS), deemed as anchors, to the mtDNA-derived long reads (Fig. 1b). The Himito graph captures the circular nature of mtDNAs. The local topology of the graph reveals the sequence diversity within a genomic region. The graphical genomes are stored in the standard GFA file format with annotations in the JSON format. The Himito Asm module extracts the primary haplotype by traversing the graph following the most-supported paths (Fig. 1c). The primary haplotype of each sample is stored in the standard FASTA files. Himito Call module aims to call both homoplasmic and heteroplasmic variants from the graph. Pair-wise alignment is performed between reference paths and alternative paths separated by anchor pairs (Fig. 1d). Genetic variants are recorded on the graph in the form of Compact Idiosyncratic Gapped Alignment Report (CIGAR) strings and are mapped back to the reference coordinates. To distinguish true variants from sequence artifacts, a permutation test was employed, leveraging the variants’ phasing information to exclude false positive calls (Methods), an idea illuminated from the development of the low-frequency somatic variant calling tools, Mutect2³¹ and Permutect (unpublished). The variants are reported in a standard biallelic VCF file with annotations. Himito Methyl module annotates the methylation signals from the graph and aggregates the methylation signals along the major haplotype. The methylation modification scores on the major haplotype are recorded in a standard BED file (Fig. 1e).

Fig. 1. An overview of Himito. Himito comprises five modules: Filter, Build, Asm, Call and Methyl.

a Himito classifies NUMTs-derived reads and mtDNA-derived reads based on alignment signature and methylation patterns. b The long reads identified as originating from mtDNAs are utilized to construct a sequence graph. Unique k-mers (red) in the reference genome are selected as anchors and compressed the reads into a graphical genome. The graph automatically forms a circular genome, which reveals the circular nature of reads derived from mtDNAs. Within a specific genomic region, the local graph topology illustrates sequence diversity. Standard reference genomes are represented as blue paths between anchor pairs, while alternative paths, shown in gray, incorporate all sequence variations. c Traversing the graph by following the most supported path (gray path, wide), the major haplotype of each sample is constructed. RC: read counts. d Sequence variants are annotated to the graph in forms of CIGAR strings and are further reported in standard VCF files. e 5-methylcytosine (5mC) modification sites with their methylation likelihood are annotated to the graph and further aggregated in a standard BED file.

The innovation of Himito centered on representing both genetic and epigenetic variations in a single unified sequence graph with haplotype information, and simultaneously addressing the key challenges of mitochondrial genome analysis, including the circular genome structure, NUMTs interference and low heteroplasmic variants detection. Himito leverages both long-read sequence alignment and methylation features to filter out NUMTs-derived reads and applies a novel permutation test to exclude false positive variant calls. The graph topology captures the circular nature of mtDNAs and delineates conserved versus highly variable regions. Himito also ensures compatibility with other bioinformatic pipelines by using common file formats. Himito is written in Rust, and serves as a fast and comprehensive toolkit for mitochondrial genetic and epigenetic analysis using long reads.

Benchmark

We benchmarked Himito’s performance using PacBio long-read whole-genome sequencing (lrWGS) data from the Human Pangenome Reference Consortium (HPRC)³⁵ and the HapMap mixture benchmarking dataset released in Somatic Mosaicism across Human Tissues consortium (SMaHT)³⁶.

We first compared Himito’s major haplotype assemblies to those generated by MitoHifi³⁴ using the HPRC-Y1 PacBio samples, with the high-quality HPRC assemblies serving as the truth set. MitoHiFi frequently failed when the sequence depth was low (10×). For successful runs, we computed the Jaccard similarity between the assembled mitochondrial genomes and truth assemblies by breaking the sequences into 100-, 300-, and 500-mer windows (Supplementary Methods). Himito achieves higher assembly accuracy than MitoHiFi across all depths, especially excelling in low-coverage scenarios, such as 10× (Fig. 2a). We next applied Himito to ONT R9 and R10 long-read datasets from the HPRC Y2 release, benchmarking assembly performance against the released reference assemblies. Assemblies derived from ONT R10 data generally have better accuracy than R9 across all depths, while R9 assemblies showed lower accuracy in 10x and 20x settings due to the higher error rate in the reads. We further benchmarked the assembly accuracy using different anchor length (Fig. 2a, Supplementary Fig. 10). The anchor length critically impacts graph topology, downstream assembly and variant calling performance. Our results suggest that the optimal anchor size for robust assembly and variant calling performance appears to be within the range of 11 to 101 bp.

Fig. 2. Benchmarking of Himito assemblies and variant calls.

a Assembly evaluation. Himito and MitoHiFi PacBio HiFi assemblies are compared with the high-quality HPRC y1 reference assemblies (sample size n = 29, mean depth >600×). Himito can be also applied to ONT R9 and R10 data and the assemblies’ accuracy are benchmarked using HPRC Y2 ONT data and their high-quality reference assemblies as truth (n = 10, details in Supplementary Methods). b Callset evaluation at predefined heteroplasmic frequencies (HF = 10%) using in silico mixture experiment (n = 25). Himito and Mitorsaw callsets were evaluated and compared in terms of Precision, Sensitivity and F1 score (full analysis on Supplementary Figs. 6, 7). c Long read callset evaluation using HapMap mixture data (n = 6). d Callset evaluation between short reads and long reads. HapMap mixture short read and long read BAMs are downsampled to 600x respectively (Illumina short read sample size n = 11, long read PacBio sample size n = 6). MToolBox (mtoolbox), GATK Mutect2-mitochondrial-mode (mutect2-mt-mode), and mtDNA-Server2 (fusion mode) were applied to both short-read and long-read BAMs for cross-modality benchmarking. Himito and Mitorsaw were applied to long read BAMs. e Taking the HPRC full-coverage short read callset generated by GATK Mutect2-mitochondrial-mode as the truth set, the Himito call performance is evaluated at different depths (n = 26). f The R² values of allele frequency for true positive (TP) calls from srWGS and lrWGS were calculated (n = 26). All box plots show the median (center line), interquartile range for bounds of box (25th to 75th percentiles), and whiskers extending to data extremes within 1.5 * interquartile range from the lower and upper bounds of the box; individual points beyond the whiskers represent outliers. Source data are provided as a Source Data file.

We next evaluated the Call module’s performance. Himito Call outputs both homoplasmic and heteroplasmic variants in the standard VCF file with their heteroplasmic frequency (HF) recorded. Using the HPRC minigraph-cactus assembly-based VCF as the truth set, we assessed homoplasmic variants (HF > 95%) called by Himito using the standard VCF evaluation tool RTG vcfeval³⁷. Himito achieved high precision and sensitivity (F1 ≈ 0.95; Supplementary Fig. 4).

To assess heteroplasmic variant detection, we conducted an in silico mixture experiment simulating predefined heteroplasmic frequencies (Methods; Supplementary Fig. 2). The assembly-based variants identified from the two donor samples were merged to create the truth set. We contrasted Himito performance with Mitorsaw (v0.2.0), a recently developed long-read mitochondrial variant caller. Himito outperformed Mitorsaw in variant calling sensitivity and overall F1 scores at lower heteroplasmic frequencies (1–10% using default parameters) (Supplementary Fig. 6). In addition, we found that Mitorsaw’s sensitivity to detect low-level heteroplasmic variants can be improved by adjusting the minimum-maf parameter, with the cost of decreased precision (Supplementary Fig. 7). For a balanced comparison, we set the minimum-maf of Mitorsaw and the vaf-threshold of Himito at the same value 0.01. As shown in Fig. 2b, with a predefined frequency of 10%, both Himito and Mitorsaw show low F1 scores in low depth scenarios (10x), yet Himito outperforms Mitorsaw when the depth is over 50x (complete results in Supplementary Fig. 7). We also optimized the parameters within the Himito call function specifically for ONT-R9 and ONT-R10 data. Distinct default parameter configurations were implemented for the ONT R9 and R10 datasets, determined via a grid search strategy to achieve optimal performance. To assess Himito’s variant calling performance, we used HPRC-Y2 ONT-R9 and ONT-R10 samples to perform the in silico mixture experiment. While good precision was achieved across both R9 and R10 datasets, the sensitivity for R10 consistently outperformed R9 within the 5–20% predefined VAF range. R9 exhibited superior sensitivity only at a VAF of 1%, which is a consequence of the parameter selections (Supplementary Fig. 8).

We extended benchmarking of Mitorsaw and Himito to the HapMap mixture lrWGS dataset and its corresponding truth callset. This resource, developed by the SMaHT consortium³⁶, was generated by experimentally mixing six HapMap cell lines³⁸ to produce variants that span low (0.5–16.5% HF) and high (83.5–100% HF) heteroplasmy levels. The minigraph-cactus graph constructed from HPRC assemblies of these donors was used to derive the truth set³⁹. As shown in Fig. 2c, Himito outperforms Mitorsaw in both precision and sensitivity using PacBio long-read dataset, as well as in overall F1 score, especially at high depth settings. We further applied Himito to 5 ONT long-read datasets in the HapMap mixture. The results showed comparable performance with PacBio lrWGS data (Supplementary Fig. 9).

To compare the variant calling performance across different data modalities, we downsampled the HapMap mixture short-read and long-read data to 600× and called variants using different short-read whole genome sequencing (srWGS) and lrWGS callers. Himito shows higher precision and sensitivity among long-read callers (Fig. 2d). Among short-read tools, mtDNA-server2³² demonstrated the highest precision, while GATK-Mutect2 mitochondrial-mode³¹ achieved the greatest sensitivity. We further conducted cross-modality benchmarking by applying short read callers to lrWGS. MToolBox converts the BAM to FASTQ expecting paired-end Illumina reads, thus cannot process lrWGS in the program. mtDNA-server2 and Mutect2 mitochondrial-mode demonstrated low precision as they are not specifically designed to handle the NUMTs contamination and the short indel errors associated with long-read sequencing. Consequently, these findings underscore the need for developing a mitochondrial variant caller specifically tailored for long-read data. Overall, long-read callers (Himito, Mitorsaw) excelled in precision, and in the case of Himito, matched or exceeded the sensitivity of the best short-read callers. The improved performance stems from a combination of advancements in both sequencing technology and algorithm design. Specifically, long reads can span NUMTs and the flanking sequences. Long reads capture native methylation modification signals with sequence data. These features allow better differentiation between sequences originating from mtDNA and those derived from NUMTs. In the meantime, Himito employs a multi-layered filtering strategy that utilizes alignment signatures, methylation patterns, and a permutation test to effectively exclude false positive calls.

In addition, we compared Himito calls to the paired HPRC short-read callset generated by GATK-Mutect2 mitochondrial-mode. Himito achieved high F1 scores compared with short-read data as truth set (Fig. 2e), and high concordance of heteroplasmic frequencies at depth exceeding 50x (Fig. 2f). These results confirm good agreements between long-read and short-read results. Furthermore, Himito’s methylation analysis aggregated 5mC signals along each major mitochondrial haplotype, enabling downstream methylation analyses. Himito’s methylation calling achieves high concordance with PbCpG-count mode (R² > 0.99, Supplementary Fig. 14). To assess the efficacy of our NUMTs filtering strategies, we evaluated Himito’s variant calling performance across a range of methylation filtering thresholds (--fraction-max-methylation) from 0.01 to 1 (Supplementary Fig. 11). Our findings indicate that incorporating methylation-based filtering reduces false positive calls. However, applying a methylation threshold below 0.1 leads to overly aggressive filtering, which significantly compromises sensitivity.

Himito is developed in Rust in order to achieve fast and scalable performance for biobank-scale mitochondrial genome analysis. Himito is publicly available under an open-source MIT license. Users can easily run Himito using the provided static binary files or build it from the source. Its design allows for straightforward integration into existing workflows and pipelines. To test Himito’s computational performance, such as run-time and memory usage, we downsampled HapMap mixture PacBio data to 10×, 20×, 50×, 100×, 200×, 300×, 400×, 500× and tested the runtime performance of Himito (overall performance and by each module respectively), Mitorsaw, MitoHiFi, and mtDNA-server2 (Methods). Himito demonstrated the fastest compared to MitoHiFi, Mitorsaw, and mtDNA-server2, as measured by elapsed time (Supplementary Figs. 12, 13). MitoHiFi and mtDNA-server2 require more processing time when analyzing samples. It is worth noting that Mitorsaw’s runtime significantly increased for samples with coverage exceeding 100× (Supplementary Fig. 13). Himito demonstrates relatively low RAM consumption, with peak usage remaining consistently below 200 Mb and only slightly increasing with higher sample coverage. In contrast, mtDNA-server2 consumes significantly more RAM. Himito’s highest RAM requirement occurs during the Build module, as this is when read sequences are loaded into memory to construct the sequence graph.

All of Us v8 PacBio long read callset

We applied Himito to analyze the mitochondrial genetic variation in All of Us (AoU) v8 long-read dataset. The AoU v8 release incorporates 2517 PacBio samples and 324 ONT samples. We assessed the sequencing depth across the mitochondrial chromosome for all PacBio and ONT samples and excluded those with an average depth below 10x. There are 95 PacBio samples and 62 ONT samples are removed due to low chrM coverage (mean depth less than 10x). After filtering, 2422 PacBio samples and 262 ONT samples remained for downstream variant calling analysis. The average chrM depth is ~80x (PacBio) and ~150x (ONT) (Fig. 3a).

Fig. 3. Mitochondrial genetic variation across 2422 PacBio and 262 ONT samples from the AoU v8 cohort.

a Mitochondrial genome (chrM) coverage depth for the PacBio (blue) and ONT (orange) samples in AoU v8. Solid lines represent the mean base depth, with shaded regions denoting the 95% confidence intervals. b Principal Component Analysis (PCA) of mitochondrial genotypes across the AoU v8 PacBio and ONT cohorts (color-coded by major haplogroups). c Distribution of homoplasmic variant counts stratified by mitochondrial haplogroups. Shaded areas represent the kernel density estimation (KDE) of the underlying probability distribution. d Distribution of heteroplasmic variant counts across the AoU v8 PacBio and ONT cohorts. e Distribution of variant density across the mitochondrial genome, stratified by SNPs and indels for both sequencing platforms. f Functional annotation and genomic distribution of the identified mitochondrial variants. Source data is provided as a Source Data file in AoU Verily Workbench.

Mitochondrial variants were identified for each sample using Himito and subsequently merged with bcftools merge⁴⁰. Haplogroups for each sample were determined using haplogrep2⁴¹. Principal component analysis (PCA) was then conducted based on genotypes extracted from the Himito PacBio and ONT callset (Fig. 3b). PCA revealed clear clustering of samples by their haplogroups, reflecting global population diversity and different ancestral compositions in PacBio and ONT callsets. We then quantified the number of homoplasmic variants (HF > 95%) and heteroplasmic variants (HF between 1–95%) in each sample. The distribution of homoplasmic variant counts was multimodal, attributable to haplogroup diversity (Fig. 3c), whereas heteroplasmic variant counts followed a unimodal distribution centered around three variants per sample (Fig. 3d), consistent with the previously identified widespread presence of low-level heteroplasmies in human mtDNA¹².

We next examined the variant density (classified into SNVs and indels across different sequencing platforms) across the mitochondrial genome (Fig. 3e). We observed the high concentration of genetic variations in the D-loop region in AoU callset, which agrees with the previous study of the presence of hypervariable regions in the mitochondrial genome⁴². To characterize length heteroplasmies commonly arising from variable-length homopolymer tracts¹², we analyzed three known length heteroplasmic loci chrM:302-316 (ACCCCCCCTCCCCCG), chrM:567-574 (ACCCCCCA), and chrM:16174-16195 (CACATCAAAACCCCCTCCCCAT), and quantified the number of distinct haplotypes per sample. Joint haplotypes between each pair of these length heteroplasmies were analyzed (Supplementary Fig. 21), and co-occurrence was further assessed at the read level using Fisher’s exact test (Supplementary Figs. 22, 23). These analyses revealed that length-heteroplasmic variants often co-occur across distant genomic regions, suggesting shared mutational processes or linked replication dynamics.

We performed functional annotation of Himito-detected variants using the Variant Effect Predictor (VEP)⁴³, classifying each variant by predicted consequence: synonymous, missense, frame shift, and stop-gain. Following annotation, we stratified variants based on their predicted consequences and analyzed their distribution in the mitochondrial gene region (Fig. 3f). We next intersected the Himito callset with the pathogenic variants reported in the MitoMap database⁴⁴. Three pathogenic variants (m.1555A>G, m.11778G>A, and m.14484T>C) were identified among 2422 individuals in the AoU v8 long-read cohort at heteroplasmic frequencies ranging from 10–100% (8 carriers total). These variants are associated with deafness, autism spectrum, atherosclerosis, Leber’s hereditary optic neuropathy (LHON), and progressive dystonia⁴⁴. Examination of All of Us survey data confirmed that several high-frequency carriers reported related disabilities. Although low-frequency mitochondrial pathogenic variants may not cause overt disease, their identification in population-scale datasets may provide valuable opportunities for early detection and prevention.

Pangenome of AoU v8 PacBio assemblies

The development of third-generation sequencing technologies has enabled personalized genome construction and driven a paradigm shift from the linear reference genome to the pangenome models that integrate multiple genomes in a unified representation. Linear reference genomes have long served as the foundation of many bioinformatics pipelines, providing coordinate systems for read mapping, variant calling, and genome annotation. However, these linear reference genomes cannot capture the full breadth of human genetic diversity. In contrast, graphical pangenomes provide a comprehensive framework for representing sequence variation across populations.

Using Himito, we assembled individual mitochondrial genomes for 2422 PacBio long-read samples from the AoU v8 long-read cohort and combined them to build a population-scale mitochondrial pangenome graph (Fig. 4a). We showed the length difference between assembled and standard reference mitochondrial genomes (rCRS) in Fig. 4b. Overall, the assembly length differences are less than 50 bp, which indicates the primary haplotype length is relatively conserved in the human mitochondrial genome. The global topology of the population-scale pangenome graph is shown in Fig. 4c. We constructed a binary matrix representing the presence or absence of each graph vertex within each individual genome. Hierarchical clustering of this matrix grouped samples by haplogroup (Supplementary Fig. 15), confirming that the graph structure encodes haplogroup-specific mitochondrial diversity.

Fig. 4. Pangenome for 2422 mitochondrial genome assemblies constructed from AoU v8 PacBio samples.

a Pangenome graph construction workflow. b Length difference between Himito assemblies and the standard human mitochondrial reference genome (rCRS). c Global structure of the mitochondrial pangenome constructed from individual assemblies. d The number of distinct haplotypes within each 1 kb window in the mitochondrial pangenome. e Sequence diversity scores for each mitochondrial gene, measured as the number of unique haplotypes normalized by the gene length. f The local graph structure in the TRNN gene. Boxes indicate the anchor nodes (annotated by the rCRS coordinates) and curves represent the graph connectivity. Multiple sequence alignment was performed among all the unique haplotypes in the gene.

Local subgraphs illustrate regional sequence variation within specific genomic intervals. Leveraging the anchor-based sequence coordinate system, we identified subgraphs and haplotypes within predefined regions (Methods). For each non-overlapping 1 kb window, we enumerated all unique haplotypes across the mitochondrial genome. The D-loop region exhibited a significantly higher number of haplotypes than any other region in the mtDNA. In contrast, MT-RNR1 and MT-RNR2 gene regions, which encode the ribosomal RNAs, are relatively conserved across the 2,422 individuals. For each gene region annotated on the standard reference genome, we computed a sequence diversity score, defined as the number of distinct haplotypes for each gene normalized by the gene length (Fig. 4e).

As an example, we visualized the local graph structure for the TRNN gene, encoding mitochondrial transfer RNA for asparagine (Asn) (Fig. 4f). Multiple sequence alignment among haplotypes of this gene revealed base-level polymorphisms within this region, illustrating how the pangenome graph provides an intuitive and comprehensive view of conserved and variable features across the population. Together, these analyses demonstrate that the mitochondrial pangenome graph captures both global and local sequence diversity across individuals.

Methylation signals in the mitochondrial genome

Among the All of Us v8 PacBio long-read samples, 1490 samples contained methylation data. We investigated the methylation patterns across 1490 PacBio long-read samples and 324 ONT samples from the AoU v8 cohort. Reads aligned to chrM were stringently filtered by their alignment signatures and methylation levels. Reads with over 20% of their CpG sites with a methylation likelihood greater than 0.5 were stringently filtered out as potentially derived from NUMTs (Fig. 1a, Supplementary Methods).

After filtering, we calculated the modification rate of each CpG site using Himito methyl. Despite stringent filtering, methylated CpGs were consistently observed across the mitochondrial genome in both PacBio and ONT samples. However, notable technical variation existed between PacBio and ONT data (Supplementary Figs. 16, 41 paired PacBio and ONT samples in v8). To account for this, we applied distinct modification rate thresholds (0.417 for PacBio and 0.125 for ONT) determined from the 99.5th percentile of the modification rate distributions in each data modality (Supplementary Fig. 17).

We next examined the distribution of hypermethylated CpG sites across individuals (Supplementary Fig. 18). Several CpG sites appeared recurrently hypermethylated in multiple samples. Overall, these findings provide evidence supporting the presence of 5mC methylation modification in the mitochondrial genome, while also highlighting technical variability across sequencing technologies.

Age-related observations

The AoU v8 long-read cohort represents a large and diverse population with rich phenotypic data, including participant age, which ranges from 20 to 90 (Supplementary Fig. 19). A recent study, using srWGS data from AoU, showed that an increase with age in low heteroplasmy mtDNA SNVs was due to a reduction in blood cell clonal diversity after middle age that in turn increases the sensitivity of bulk sequencing data for detecting the mtDNA variants of the dominant clones⁴⁵. We found a similar age-associated accumulation of low-frequency SNVs (1-10% heteroplasmic frequency) in the lrWGS data (Fig. 5a). Our results also validate that the increase in low heteroplasmy SNVs with age is dominated by G>A and T>C transitions, consistent with these variants being introduced by a replication-linked mtDNA mutational process⁴⁵ (Fig. 5b; Supplementary Fig. 20). We also observed a decrease in the number of distinct haplotypes with age in the chrM:302-316 repeat expansion region (Supplementary Fig. 24).

Fig. 5. Age-related observations in 1490 PacBio samples with methylation data and 324 ONT samples.

a The number of low-frequency heteroplasmic SNVs (1–10%) accumulated with age (sliding window size of 10). The number of individuals in each bin is shown in the bottom. b A subset of low-frequency heteroplasmic SNVs, G to A and T to C shows the age-accumulating patterns. The full analysis shows in Supplementary Fig. 20. c The number of hyper-methylated CpG sites (mod rate > 0.417, 99.5th percentile) in the 1490 PacBio samples and in the 324 ONT samples (mod rate > 0.125, 99.5th percentile) stratified by age groups. Data are presented as mean values with 95% confidence intervals. d Negative correlation between the number of low-frequency heteroplasmic SNVs (1–10%) and the number of methylated CpG sites (Mod score >0.417) in PacBio samples (p-value = 3.3e-15). Ordinary Least Squares (OLS) multiple linear regression model (Methylated Sites ~ SNP + age; N = 1490 samples). Methylation level demonstrated a statistically significant negative association with low-frequency SNVs (beta = −0.35, two-sided t-test). No adjustments for multiple comparisons were performed as this was a single hypothesis test. The solid line represents the linear regression fit, and the surrounding shaded region denotes the 95% confidence interval. Source data is provided as a Source Data file in AoU Verily Workbench.

Intriguingly, analysis of both PacBio and ONT methylation data revealed an increase in the number of hypermethylated CpG sites with age (Fig. 5c). Moreover, a significant negative correlation exists between the counts of low-frequency SNVs and methylated sites per sample, including when correcting for age (Fig. 5d, p value < 1e-14 after age correction). A prior study showed that de novo mtDNA methylation prevents stress-induced mtDNA damage during the critical peri-implantation period of embryogenesis⁴⁶, which provides a possible explanation of the antagonistic relation between methylation and low-frequency SNVs in our data. Himito’s ability to jointly analyze variation and methylation levels could enable future explanation of the role of mtDNA methylation in mitochondrial biology.

Discussion

In this study, we present Himito, a graph-based toolkit designed for mitochondrial genomic and epigenomic analysis using PacBio or ONT long reads. Himito addresses several key challenges in mitochondrial genomics, including accurately handling the circular structure of mtDNA, removing the NUMTs-derived reads, identifying low-level heteroplasmies, and integrating 5mC methylation signals. We applied Himito to analyze the whole-genome aligned BAM files released in the AoU and SMaHT consortium (without realignment or other reprocessing). Himito integrates both genomic and epigenomic features in a single graph representation, and excludes reads derived from NUMTs that have been misaligned to improve the accuracy of the downstream genome assembly, variant calling, and methylation analysis. The Himito graph reveals the circular nature of mtDNA, which facilitates visualization of sequence diversities. The haplotypes, sequence variants, and methylation signals can be extracted from the graph in standard file formats, which are compatible with standard bioinformatics tools. Himito is written in Rust, which is easy to install and integrate into existing bioinformatics pipelines.

Compared with existing tools, Himito delivers more accurate mitochondrial genome assemblies, improved heteroplasmic variant detection, and integrated methylation analysis. Benchmarking against the state-of-the-art methods using publicly available HPRC and HapMap mixture lrWGS dataset demonstrated that Himito yields higher assembly accuracy than MitoHiFi³⁴ and higher sensitivity for low-frequency heteroplasmic variants than Mitorsaw. Together, these results establish Himito as a powerful and versatile toolkit for comprehensive long-read mitochondrial genome analysis.

Application of Himito to the All of Us long-read dataset enabled population-scale investigation of mtDNA structure, variation, and methylation. Analysis of the AoU mitochondrial PacBio long-read callset elucidates the haplogroup composition within the AoU cohort and identifies pathogenic variants in the AoU v8 samples. We then assembled the major haplotype for each sample and constructed a population-level pangenome graph to represent 2422 genomes in the AoU v8 cohort. We investigated the conserved and highly diverse regions in the mitochondrial genome. The sequence graph provides an intuitive visualization of sequence diversity in a specific genomic region and enables a better understanding of the genomic structure within the mitochondrial genomes. Our analysis of mtDNA methylation further suggests that, despite its low level, 5mC methylation exists in the mitochondrial genome. We investigated the age-related genomic and epigenomic patterns of the mitochondrial genome and validated an increase in low-heteroplasmy SNVs with age present in srWGS data that was shown to be the result of reduced blood cell clonal diversity with age⁴⁵, and further identified a similar trend in the number of hypermethylated CpG sites. We observed a negative correlation between the number of low-frequency SNVs and methylated CpGs, suggesting a potential interplay between these processes. These findings require further investigation into the underpinning mechanisms and their functional roles. Himito facilitates the integrative analysis of mitochondrial genomes, demonstrating the effectiveness of combining variation and methylation analyses through lrWGS.

Despite these strengths, there are several limitations to consider. First, Himito is capable of detecting low-frequency heteroplasmies with sufficient coverage; however, caution should be taken when distinguishing low-level heteroplasmic variants from sequencing errors. Orthogonal validation using multiple technologies, such as integrating short-read and long-read data, is helpful for confirming the presence of low-heteroplasmic variants within a single sample. For improved mitochondrial genome analysis with Himito, the raw variant callset from AoU v8 requires the implementation of more advanced filtering techniques before downstream association analysis. In addition, Himito’s variant calling performance on lower coverage or noisy datasets may be reduced. Himito only incorporates 5mC modification in the analysis. The accuracy of epigenetic modification detection is dependent on the sequencing platform, the methylation caller, and the library preparation. More comprehensive studies on the technical variations in methylation analysis are needed to be performed. Future work will aim to address these challenges by incorporating more sophisticated variant filtering strategies, supporting hybrid data integration, and expanding compatibility with additional epigenetic modification types.

As long-read sequencing becomes increasingly cost-effective, population-scale datasets will continue to expand. Himito provides a scalable foundation for exploring mitochondrial disease mechanisms, age-related processes, and tissue-specific variation across large cohorts such as the All of Us and SMaHT⁴⁷. Potential future works will include the development of a long-read NUMT caller that utilizes the reads classified as NUMT-derived by Himito. This new tool will be benchmarked against existing short-read based NUMT detection methods, such as recently published work that involves genotyping NUMTs in the pangenome graph using short-read samples⁴⁸. In addition, optimizing the graph-based algorithm to generalize Himito’s application to large, complex circular genomes, such as the plant mitochondrial genomes, is a key focus for future development. Taking Lycopodium japonicum mitochondrial genome as an example, we collected the PacBio long read sequence data from previous research (SRA accession: SRX20558281)⁴⁹, and then applied Himito (k = 801), MitoHiFi, HiMT⁴⁹ to assemble the plant mitochondrial genome. As shown in Supplementary Fig. 25, the k-mer composition of the three assemblies are highly concordant, where 99.8% k-mers are shared by all three assemblies. However, careful selection of the anchor k-mer size is necessary for the Himito graph, especially for plant mitochondrial genomes, due to their large size, structural complexity, and repetitive sequences. Future algorithmic modifications are planned to automate the anchor k-mer selection based on reference genome size and sequence complexity, thus better accommodating the complexity of other circular genomes. With future improvements, Himito can build a cross-species pangenome model designed to enhance the understanding of mitochondrial genome evolution across species, utilizing data obtained from consortium efforts such as the Darwin Tree of Life Project⁵⁰.

In summary, Himito expands the toolkit available for mitochondrial genome analysis by introducing a fast, scalable, and versatile toolkit for long read mitochondrial analysis. Its central innovation lies in providing a unified framework that integrates both genetic and epigenetic variations in a single graph. Himito simultaneously addresses several key challenges in mitochondrial genome analysis, including effectively removing NUMTs-derived reads, accurately handling the circular structure of mtDNA, detecting low-frequency heteroplasmic variants, and integrating the analysis of 5mC modifications. This unified approach enables integrated investigation of mitochondrial genetic and epigenetic diversity and offers a foundation for future studies of disease, aging, and evolution.

Methods

Algorithms

Himito includes five key modules for comprehensive mitochondrial genome analysis:

Filter (Fig. 1a): The Filter module aims to remove reads from NUMTs to ensure the accuracy of downstream analysis. It classifies reads using both alignment signatures (SA tags in BAM files) and read-level methylation patterns, identifying and separating NUMT-derived reads from mtDNA-derived reads.

Build (Fig. 1b): The Build module focuses on constructing a graphical genome model for mtDNA. The Himito graph construction begins by selecting unique k-mers in the reference genome as anchors, which are mapped to long reads deemed as mtDNA-derived. Sequences between anchor pairs are extracted and consolidated into vertices deemed as edges. The resulting Himito graph is automatically circular, representing the circular topology of the mtDNA. The connectivity of the graph vertices always follows the anchor-edge-anchor pattern, and thus formulate a sequence-based graph coordinate system to map graph vertices back to the standard reference coordinate. Edge vertices in the graph incorporate all the sequence variations represented in the sample, including SNVs, indels and SVs. The number of reads supporting each edge vertex are reported in the GFA files, e.g., RC:i:4. The annotations such as anchor reference coordinates, read name supporting each edge sequences are reported in the PG:J field in json format.

Asm (Fig. 1c): The Asm module is designed to assemble the primary mitochondrial haplotype from the constructed graph. Although mutations gradually accumulate in the population of mtDNAs over time due to errors in mtDNA replication⁵¹, most of the molecules in the mtDNA pool are nearly identical, with most bases in a state of homoplasmy and a minority of bases showing heteroplasmic variation, and usually at low levels of heteroplasmy⁵². Asm reconstructs the primary haplotype sequence by traversing the graph following outgoing vertices that have the highest count of supporting reads, then outputs the concatenated sequences in FASTA format. The number of reads supporting each edge vertex is recorded in the GFA file in JSON format. When traversing the graph, the primary path was identified by starting from the first anchor of the graph, walking to the outcoming edge vertices with the largest number of supporting reads until looping back to the starting vertex or any node is revisited during graph traversal. When the number of supporting reads are identical for two outcoming vertices, we apply a heuristic algorithm to determine the primary path by lexicographically sorting the path sequences and choosing the one appearing first. This enables us to get a deterministic primary assemblies from the sequenced sample.

Call (Fig. 1d): The Call module detects both homoplasmic and heteroplasmic variants from the graph. It enumerates paths between anchor pairs, aligns these paths to the reference segments, records alignment differences as CIGAR strings, maps variants to reference coordinates, and outputs them in VCF format, which includes read counts, depths, and heteroplasmic frequencies information, while also employing a permutation test to filter false positives.

Methyl (Fig. 1e): This module analyzes DNA methylation at both read and haplotype levels. It identifies 5mC modifications at CpG sites using MM and ML tags, annotates methylation likelihoods to graph vertices, aggregates methylation signals, estimates methylation rates for CpGs on the major haplotype, and outputs these rates in BED format.

Please see the Supplementary Methods for detailed description.

Permutation test for variant filtering

One major challenge for heteroplasmic variant calling is to distinguish the low level heteroplasmic variants with sequencing artifacts. We adopted a novel strategy to exclude false positive calls based on the co-occurrence patterns of variants on long reads. The sequencing artifact, especially the extensively observed short indels in both the PacBio and ONT long read sequences, randomly occurs among all of the reads. With the assumption that the true variants co-occur with other genuine variants in the same haplotype, we performed the permutation test to filter out variants that are likely to be sequence artifacts. A presence-absence matrix was initially built to represent each variant’s occurrence across all reads. We calculated the sum of each pairwise Jaccard similarity of variants as the statistics. By randomizing the variants occurrence in the reads, we calculated the null distribution by shuffling the occurrence of the variants in all reads and computing the sum of all pairwise Jaccard similarity. We then computed the empirical p-values for each observation to exclude the variants that agree with the null hypothesis. This approach effectively excluded many false positive calls in the raw graph and improved the precision of the Himito heteroplasmic variant calling.

In silico mixture experiment to evaluate heteroplasmic calling performance

To evaluate the precision and recall of variant calling at a defined heteroplasmic frequency, an in silico mixture experiment was conducted (Supplementary Fig. 2). This experiment utilized the sequencing reads aligned to the mitochondrial chromosome (chrM) from two distinct donor samples, which involves distinct sets of true variants. For each donor, the aligned BAM file was downsampled to a specified coverage depth. Subsequently, the downsampled BAM files from the two donors were merged at a predefined mixing ratio, such as 0.1 and 0.9, to simulate a sample with a known heteroplasmic variant frequency. Variant calling was then performed on the resulting mixed BAM file. The truth set for variant evaluation was established as the union of all assembly-based variants identified in the two donor samples. Finally, the performance of the variant caller on the mixed sample was assessed by comparing the called variants against the truth set using the RTG vcfeval tool, allowing for the quantification of precision and recall at the simulated heteroplasmic frequency.

The workflow is in https://github.com/broadinstitute/Himito/blob/main/wdl/MixSampleExperiment.wdl.

HapMap mixture experiment and mitochondrial truth callset generation

The SMaHT network³⁶ created an in vitro-constructed cell line mixture using six well-characterized cell lines from the International HapMap Project³⁸, composed of 83.5% HG005, 10% HG02622, 2% each of HG002, HG02257, and HG02486, and 0.5% HG00438. Genome Characterization Centers at Baylor College of Medicine, the Broad Institute, the University of Washington and Seattle Children’s Research Institute, and Washington University in St. Louis generated high-depth PacBio long-read whole-genome sequencing (WGS) data for this mixture. The HapMap mixture somatic variant truth set³⁹, including mitochondrial variants, was constructed using a minigraph–cactus pangenome graph built from de novo assemblies produced by the Human Pangenome Reference Consortium³⁵. The aligned BAM files and the corresponding truth set were obtained from the SMaHT data portal.

For each PacBio long read whole genome sequencing sample in the HapMap dataset, we first subset the whole genome BAM file to reads aligned to chrM. We downsampled the BAMs to 10×, 20×, 50×, 80×, 100×, 200×, 300×, 400×, 500×, 600×, then called Himito and Mitorsaw at each coverage level, respectively to generate the mitochondrial callset. We used RTG vcfeval to evaluate the performance of Himito and Mitorsaw, respectively. We further benchmark ONT long-read whole-genome sequencing samples in the HapMap dataset using the same strategy. The analysis workflow is in https://github.com/broadinstitute/Himito/blob/main/wdl/HapMap_downsample_experiment.wdl.

Run-time experiments

To measure the performance of each tool consistently, we conducted runtime experiments using 6 PacBio long read datasets from Hapmap mixture samples. The aligned BAM files from the six samples were first subset to chrM (samtools view to subset the BAM) and then downsampled to 50×, 80×, 100×, 200×, 300×, respectively (GATK downsampleSAM). To benchmark the runtime performance of different toolkits, each downsampled BAM file was further converted into FASTQ before experiments (samtools FASTQ). We measured the runtime performance for each module of Himito (Filter, Build, Call, Asm), respectively, and the total Himito Filter-Build-Call-Asm workflow. We also compared the performance with other long-read methods (Mitorsaw and MitoHiFi), as well as a short-read analysis tool (mtDNA-server2). We measured wall clock time (elapsed real time) and maximum resident set size (maximum rss, kb) with /usr/bin/time. The scripts conducting runtime experiments were available in https://github.com/broadinstitute/Himito/tree/main/runtime_experiments.

Computational performance experiments were run on a server (zep-gpu.broadinstitute.org) hosted by the Data Science Platform at Broad Institute. The server is Ubuntu 20.04.4 LTS server (Linux 5.4.0-122-generic) with AMD EPYC 7763 64-Core Processor (64 cores), NUMA CPU (128) 500 GB of RAM, and an NVMe SSD disk. No other user or major task was active.

Population-level pangenome construction

Mitochondrial genomes were assembled by Himito for each sample in AoU v8 release individually using PacBio long reads. These assemblies were then combined and used as input for the Himito Build function to construct a population-level genome graph. This graph was further processed and circularized using customized scripts. For specific genomic regions of interest, the number of distinct haplotypes represented in the graph was extracted and analyzed to assess regional diversity. Anchors served as a sequence-based coordinate system. We associated each graph vertex to a reference interval utilizing the bounding anchor sequences and their reference coordinates. For each genomic region, such as genes, we found all the graph vertices overlapping with the interval and reconstructed the subgraph in that genomic region. We then traversed the subgraph and extracted all the haplotypes in that region, without any coverage-based trimming.

The distinct haplotypes within each genomic region were recorded in the standard FASTA files. Mitochondrial haplotype differences were analyzed using ClustalW⁵³ for multiple-sequence alignment (MSA) and MSAplot (https://github.com/mourisl/MSAplot) for visualization. Haplotype diversity was quantified by counting distinct haplotypes in 1 kb windows across the genome and visualized with pyCirclize⁵⁴. Gene intervals from the rCRS GFF3 file (NCBI) were used to extract subgraphs from the AoU v8 pangenome. Distinct haplotype sequences were identified by graph traversal within these subgraphs. Genetic diversity per gene was calculated as the number of distinct haplotypes normalized by gene length, providing a measure of gene-specific diversity.

Author contributions

H.S., Y.H., and K.V.G. conceived the project and designed the algorithm. H.S. developed the toolkit and performed the benchmarking and AoU lrWGS analysis. T.J.D. contributed to the interpretation of AoU lrWGS callset discovery and age-related pattern. N.K., E.C., and S.C.J. contributed to HapMap mixture dataset development and validation. N.K., E.C. contributed to the srWGS benchmarking using HapMap dataset. D.B. contributed to the variant filtering algorithm. K.V.G. and Y.H. supervised the project. The All of Us Biobank collected, stored, and plated participant biospecimens. HudsonAlpha and Discovery Life Sciences generated the PacBio whole-genome sequencing data, and Baylor College of Medicine generated the ONT whole-genome sequencing data (N.G., N.L., S.G., S.L., C.B., J.G., D.M.M., R.A.G., G.A.M., F.J.S., J.D.S., E.E.E. and K.F.D.).

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

The All of Us Research Program’s Long-Read Project is supported by the U.S. National Institutes of Health (NIH), Office of the Director: (Broad Institute: OT2OD002750, OT2OD038121; Baylor College of Medicine/Johns Hopkins University: OT2OD002751, OT2OD038122; University of Washington: OT2OD002748, OT2OD038111; HudsonAlpha Institute for Biotechnology: OT2OD027070). Research reported in this publication was supported, in part, by the SMaHT Pilot and Feasibility Award (2025). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Data availability

Source data is provided as a Source Data file. The benchmarking statistics generated in this study are provided in the Supplementary Information/Source Data file. The raw data for the benchmarking analysis are listed as follows: the HPRC Y1 data are available in https://anvil.terra.bio/#workspaces/anvil-datastorage/AnVIL_HPRC; the HPRC Y2 ONT R9 and R10 data are downloaded from https://data.humanpangenome.org/raw-sequencing-data. The HapMap mixture dataset can be found in the SMaHT data portal at https://data.smaht.org/ with log in credentials. The plant genome assembly analysis is performed using public PacBio HiFi sequencing data (SRA accession: SRX20558281). Due to the data access restrictions of the All of Us Research Program, individual-level data are not permitted to be extracted from the Researcher Workbench. Control Tier access to AoU workbench can be obtained via Verily workbench with proper training. We have provided Source data (gs://rw-migration-aou-rw-2cf44b8f/SOURCE_data.xlsx) for AoU analysis accessible in the Research Workbench and provided instructions for accessing the AoU data through the authorized All of Us Research Workbench environment. The AoU v8 release can be accessed via the research workbench https://www.researchallofus.org/data-tools/workbench/. The AoU v8 lrWGS analysis is performed in Research Workbench Workspace: https://workbench.verily.com/workspaces/aou-rw-2cf44b8f. The generated mitochondrial callset and assemblies are available on the Verily workbench with proper credential. The whole-genome BAMs from both AoU and SMaHT HapMap mixture dataset were used as it is without realignment or other preprocessing. Source data are provided with this paper.

Code availability

Himito is publicly available at https://github.com/broadinstitute/Himito (DOI: 10.5281/zenodo.21522817)⁵⁵. The general user guide can be found here: https://github.com/broadinstitute/Himito/blob/main/docs/User_guide.md. The WDL workflow provided in the Himito repository can be run directly in the Cromwell server in the AoU Research Workbench, where input files are provided by a JSON file. The runtime parameters and docker images are specified in each WDL task. An example for job submission in AoU Research Workbench is provided here: https://github.com/broadinstitute/Himito/blob/main/docs/cromwell-Himito-quickstart.ipynb. The cromwell server setup instructions can be found here: https://support.researchallofus.org/hc/en-us/articles/14428263737620-How-to-use-Cromwell-in-the-All-of-Us-Researcher-Workbench.

Competing interests

K.V.G. is a co-inventor on a pending international patent application related to long-read RNA isoform sequencing, licensed to Pacific Biosciences, but not used in this study. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77418-y.