Introduction

Bipolar disorder (BD) is a major psychiatric disorder with manic and depressive episodes. BD affects 1% of the population and causes significant social losses [1]. While several medications are effective in alleviating BD symptoms, many challenges remain to be solved, such as treatment resistance and the necessity for long-term treatment. In the current clinical practice, diagnosis and treatment rely on symptomatology that may not align with the underlying biological backgrounds. We need to understand the biological basis of BD for better diagnosis and treatment beyond the current symptomatic nosology.

Based on the high heritability [2, 3], various genomic studies have explored BD’s genetic architecture [4,5,6,7,8,9,10,11]. Large-scale genome-wide association studies have identified associations between BD and genes involved in synaptic function and fatty acid metabolism [5,6,7]. Several copy number variations have been reported to increase the risk of BD [12, 13]. Whole-exome sequencing has identified a potential link between BD and AKAP11 [4]. Trio-based studies have uncovered de novo mutations, particularly in constrained genes, that may be associated with BD [9,10,11]. Despite these advances, the genetic architecture of BD has not been completely understood.

In addition to the traditional germline variants, somatic mosaic variants have emerged as promising contributors to BD [14, 15]. Specifically, mosaicism in mitochondrial DNA (mtDNA), known as heteroplasmy, has been identified as a potential factor for BD [14, 15]. We have reported the enrichment of mitochondrial heteroplasmic variants in tRNA genes, particularly the m.3243A>G variant, in peripheral samples from BD patients [14]. The m.3243A>G variant, a well-established causative variant for MELAS (mitochondrial myopathy encephalopathy lactic acidosis stroke-like episodes), has also been recurrently detected in postmortem brain samples of BD patients [15]. Conversely, mitochondrial diseases caused by specific mtDNA variants, including m.3243A>G variants, are frequently associated with BD [16,17,18].

Mitochondrial dysfunction is theoretically linked to neuropsychiatric abnormalities for their essential roles in ATP production and calcium ion regulation, both of which are critical for the nervous system [19,20,21]. Indeed, BD patients exhibit mitochondrial abnormalities [20], including mitochondrial protein damage by oxidative stress [22]. Hyper-excitability and mitochondrial abnormalities were reported as hallmarks of iPS-derived neurons induced from BD patients [23]. Brain-specific Polg mutant transgenic mice with accumulation of mtDNA mutations exhibit repeated depressive episodes [24, 25], and the human POLG variants could be associated with BD [26]. Besides rare variants, some common variants in mtDNA are reported to be potentially associated with BD [27, 28]. Building on these empirical findings and theoretical considerations, we hypothesize that brain mitochondrial DNA variants contribute to BD’s pathophysiology.

Here, we investigate the associations between BD and mitochondrial heteroplasmic variants in the brain. While previous comprehensive analyses of mitochondrial heteroplasmic variants have been conducted for the brains derived from BD, schizophrenia (SZ), and major depressive disorder [29, 30], extensive landscapes of low-level heteroplasmic variants in BD brains have been elusive [29]. We set two models for the mode of association between BD and mitochondrial heteroplasmic variants in the brain:

  1. (1).

    A limited number of specific pathogenic variants, with variant allele fractions (VAF) above a certain threshold (e.g., VAF > 1%), which arise early in development or are maternally inherited, impair neural functions. These pathogenic variants would be analogous to those implicated in mitochondrial diseases.

  2. (2).

    Accumulation of numerous non-specific low-level heteroplasmic variants (e.g., VAF < 1%), which emerge later in development or during aging, progressively degrade mitochondrial functions and impair neural activity.

Guided by these two models, we explored mitochondrial heteroplasmic variants in 163 brain DNA from 54 BD patients, 54 controls (CTRL), and 55 SZ patients, using duplex molecular barcode sequencing for single-molecule resolution. Our findings indicate potentially pathogenic heteroplasmic variants with VAF > 1%, including loss-of-function (LoF), rRNA, and m.3243A>G variants, as relevant for BD. In contrast, our single-molecule analysis revealed no increase in low-level heteroplasmic variants in BD. These findings provide a foundation for the development of future therapeutic strategies for BD, informed by genomic stratification by mitochondrial variant profiles.

Materials and methods

Detailed methods are provided in Supplementary Methods.

Postmortem brain DNA samples

We obtained 163 postmortem brain DNA samples from the Stanley Medical Research Institute, including 54 samples from BD patients, 54 from CTRL, and 55 from SZ patients. Fifty-eight samples were derived from the frontal cortex, and 105 from the superior temporal gyrus. SZ samples were included as disease control to identify BD-specific characteristics. No formal power calculation was performed; instead, we utilized the maximum number of available samples. No significant age differences were found between BD and CTRL groups (two-sided t-test, P = 0.64); however, a significant difference in sex distribution was observed (two-sided Fisher exact test [FET], P = 0.995 × 10−4) (Table S1). The average ages of the BD, CTRL, and SZ groups were 45.8, 46.7, and 44.4 years, respectively. The male-to-female ratios in BD, CTRL, and SZ were (26:28), (45:9), and (36:19), respectively. The observed differences in sex distribution were primarily influenced by the availability of postmortem brain tissue and the need to adjust for the average ages across the three groups. Brain tissues were predominantly collected from relatively young donors under the age of 65, with the exception of two samples. The lower proportion of female controls can be attributed to the demographic trend that females under 65 generally have lower mortality rates from accidents, cardiovascular events, and other causes within this age range compared to males. The investigators were blinded to group allocation during sequencing and the application of both bulk and single-molecule bioinformatic pipelines (described below). Following the identification of heteroplasmic variants, diagnostic information was provided by the Stanley Medical Research Institute. Case-control comparisons were subsequently conducted.

Ethics approval and consent to participate

This study was conducted in accordance with the Helsinki Declaration and was approved by the Research Ethics Committee of the Faculty of Medicine, Juntendo University (approval ID: M19-0278). All procedures followed the Ethical Guidelines for Medical and Biological Research Involving Human Subjects issued by the Government of Japan. Informed consent was obtained from the next of kin for the use of all specimens.

Ultra-deep mtDNA sequencing

The brain DNA was subjected to ultra-deep mtDNA sequencing at a target depth of 400,000× using duplex molecular barcodes. DNA fragments underwent library preparation with dual molecular barcoding via SureSelect XT HS2 and custom mtDNA probes (Agilent Technologies). The libraries were sequenced on NextSeq2000 or NovaSeq6000 (Illumina) using 151 cycles in paired-end mode. Heteroplasmic variants were classified based on their variant allele fraction (VAF): those with 50% <VAF < 90% were categorized as high-VAF, those with 1% <VAF < 50% as moderate-VAF, and those with VAF < 1% as low-VAF. We developed a “bulk pipeline” to detect high/moderate-VAF heteroplasmic variants, and a “single-molecule pipeline” to identify low-VAF variants.

Bulk pipeline

The bulk pipeline incorporated alignment and quality control using AGeNT-3.0.6 (Agilent), BWA-0.7.17 [31], and GATK-4.2.5.0 [32] with the human reference genome hg38, followed by PCR/optical-duplication removal using GATK MarkDuplicates (Table S2). Heteroplasmic variant candidates were called by bam-readcount-1.0.1 [33] for single nucleotide variants (SNVs) and GATK Mutect2 [34] for insertions/deletions (INDELs). Candidates in low-mappability mtDNA regions were filtered out.

Potential sequencing artifacts and commonly observed heteroplasmic variants were filtered out using a Panel of Normals (PoN). To detect heteroplasmic variants in D-loop region, we used a modified mtDNA reference sequence shifted by 8000 bp, following the gnomAD’s approach for mitochondrial genome [35].

Single-molecule pipeline

We employed duplex molecular barcode sequencing to achieve single-molecule resolution [36]. In this approach, each strand of a double-stranded DNA molecule is uniquely tagged with a duplex unique molecular identifier before PCR amplification during library preparation. This tagging enables the distinction between the two complementary strands of the original DNA molecule. Following PCR amplification and conventional sequencing, true variants are identified only if they are observed at the same genomic position on both strands, whereas PCR or sequencing errors typically appear on only one strand. This strategy allows for the differentiation between true variants and technical artifacts. Duplex molecular barcode sequencing can achieve an error rate below 1.0 × 10⁻7 per base, thereby enabling the detection of mutations at the single-molecule level [36, 37]. This level of sensitivity is appropriate for mtDNA, in which per-base mutation frequencies are estimated at approximately 1.0 × 10−5 [38].

Our “single-molecule pipeline” utilized Fgbio-2.2.1 to generate consensus sequences from molecule-tagged sequence read data (Table S2). Variants were retained if they were supported by three or more independent consensus reads of double-stranded DNA and had consensus base quality scores exceeding 70, corresponding to an estimated error rate of less than 1.0 × 10⁻7. To ensure high mapping reliability, candidate variants were restricted to genomic regions with a mapping quality (MAPQ) score of 60. In this study, we refer to the “single-molecule pipeline/analysis” as our analytical pipeline/method, and duplex molecular barcode sequencing as the underlying molecular sequencing technique.

Variant annotation

The detected variants were annotated with gnomAD-v3.1.2 [39], ToMMo54KJPN [40], and HelixMTdb [41], SnpEff-4.3 [42], ClinVar (http://www.ncbi.nlm.nih.gov/clinvar/), Mitocalss.1 [43], PolyPhen-2 [44], MitoTIP [45], PON-mt-tRNA [46], and mitochondrial local constraint (MLC) score [47].

Mitochondrial variants were categorized into six classes: D-loop, synonymous, missense, loss-of-function (LoF), tRNA, and rRNA. Deleterious missense variants were defined as PolyPhen2 prediction = “probably_pathogenic” and Mitoclass.1 prediction = “damaging”. Deleterious coding variants were defined as LoF and deleterious missense variants. Deleterious tRNA variants were defined as MitoTIP score > 12.66 and PON-mt-tRNA prediction = “Pathogenic/Likely_pathogenic”. Ultra-rare variants were defined as those absent as homoplasmic variants in neither gnomAD v3.1-2 nor ToMMo54KJPN databases. Common variants were those present as homoplasmic variants in over 1% of the population in either database, while medium-rare variants were those that fell between these two classifications. Haplogroups were annotated by Haplogrep 3 [48], using a quality threshold of 0.75. Mitochondrial haplotypes for all samples are described in Table S1. The most prevalent haplogroup was the H group. The ratios of individuals with haplogroup H to those with non-H haplogroups were (23:31) in BD, (21:33) in CTRL, and (20:35) in SZ.

Validation by target amplicon sequencing

We prioritized candidates in tRNA, rRNA, and nonsynonymous variants for target amplicon sequencing (TAS) validation using iSeq100 (Illumina) with 151 cycles pair-end mode. TAS libraries were prepared according to our previous analyses [49, 50].

The sequence reads were aligned to hg38 by BWA-0.7.17, and the alternate base calls with base quality ≥30 were counted to calculate the VAF of the candidate variants. The candidates with VAF ≥ 0.3% and MAPQ = 60 were considered validated. The candidates with VAF ≥ 0.3% and MAPQ < 60 underwent manual inspection of whether the PCR products’ sequences were derived from the target sequences in the mitochondrial genome or homologous sequences in the nuclear genome using IGV [51]. INDELs were checked and counted by IGV. We excluded candidates failing in TAS validation from subsequent analysis.

Bulk analysis for high/moderate-VAF variants

We compared the counts of heteroplasmic variants with VAFs > 1% in BD/SZ and CTRL. This threshold was set based on the observed VAF distribution and the following assumption. Variants with VAF < 1% (low-VAF) represent variants arising later in development [38], while those with VAF > 1% (high/moderate-VAF) represent variants arising early in development or maternally transmitted. We assessed the enrichment of high/moderate-VAF variants in BD or SZ by FET. No multiple testing correction was applied due to the limited number of detected variants. We evaluated the enrichment of these variants using linear models (lm in R-4.3.1) that incorporated potential covariates, including disease status, age, sex, tissue type, postmortem interval (PMI), brain pH, haplogroup, and sequencing depth (bulk deduplicated depth). In this analysis, haplogroup was simplified as H or non-H groups. Given the challenges in fitting parametric distributions to the detected variant data—attributable to the small number of high/moderate-VAF heteroplasmic variants per individual—this analysis should be interpreted as exploratory.

Single-molecular analysis for low-VAF variants

We counted the number of point mutation events arising later in development based on the low-VAF heteroplasmic variants to calculate per-base mutation frequencies. Following a previous study of single-molecule analysis using duplex barcoding [38], the per-base mutation frequency was defined as the number of variants with VAF < 1% relative to the total base molecules. The variants with VAF > 1% were excluded as maternally inherited variants or those arising during early embryonic development (clonality cut off [38]). We assessed the difference in per-base mutation frequencies between BD/SZ and CTRL by t-test. No multiple correction was applied in this analysis. We also employed in-depth regression analysis to accommodate potential covariates—disease status, age, sex, tissue, PMI, brain pH, haplogroup, and sequencing depth (single-molecule depth)—using lm in R-4.3.1. To evaluate proliferation/selection bias of mitochondrial variants, we assessed the difference of VAFs in BD/SZ and CTRL using lmer (linear mixed model) in lme4 with the same sample-level covariates described above as fixed effects and sample ID as a random effect.

Mutation pattern analysis

We analyzed tri-nucleotide-based mutation patterns specific to certain variant categories, such as disease status and functional regions. The differences in mutation patterns were assessed based on 96 tri-nucleotide sequence contexts, using cosine similarity to quantify similarities of the mutation pattern distributions.

Results

We obtained an ultra-deep sequencing dataset at an average depth of 6.85 × 104 for the mitochondrial genome after deduplication via bulk analysis pipeline. The average depth via the single-molecule pipeline was 4.76 × 103 (Table S3). One sample (CTRL-51) failed in library preparation and was excluded in the subsequent analysis. As a result, 162 samples from 54 BD patients, 53 controls, and 55 SZ patients underwent the subsequent analysis. Contamination levels assessed by Mutect/ContEst, revealed virtually no significant contamination across all samples.

Bulk analysis for high/moderate-VAF variants

We detected 15 high-VAF and 101 moderate-VAF heteroplasmic variants (VAF > 1%), including 36 ultra-rare variants, by the bulk analysis after TAS validation (Table S4 and summarized in Table S5). We validated 71 out of 76 TAS candidates (93.4%), excluding two questionable variants with low VAF (Table S6). While common and medium-rare variants were detected at high VAF values across the dataset, ultra-rare variants with VAF > 10% were only detected in the D-loop region. INDELs were significantly more frequent in the D-loop region (two-sided FET, P = 0.0036). These observations indicated the selection bias to the ultra-rare variants and INDELs at the mitochondria level. A total of 116 high/moderate-VAF heteroplasmic variants (Fig. 1A) in 162 cortical DNA underwent the subsequent analyses.

Fig. 1: High/moderate-VAF heteroplasmic variants detected by bulk analysis.
figure 1

A Circos plot of all the high/moderate-VAF variants detected in this study. The heteroplasmic variants of interest are indicated. B Per-individual counts of high/moderate-VAF variants in BD, CTRL, and SZ. C High/moderate-VAF variant counts per individual in BD and CTRL, stratified into D-loop, synonymous, missense, LoF, tRNA, and tRNA variants. D Per-individual counts of ultra-rare high/moderate-VAF variants in BD, CTRL, and SZ. E Ultra-rare variant high/moderate-VAF counts per individual in BD and CTRL, as in panel C. For B–E, P-values calculated from two-sided Fisher exact tests are described on the right, and error bars indicate the standard errors.

Full size image

BD exhibited a higher number of total high/moderate variants, compared to CTRL (two-sided FET, P = 0.000525). We also observed more variants in SZ than in CTRL (P = 1.98 × 10−5) (Fig. 1B). While no region-specific enrichment was observed in BD (Fig. 1C), D-loop variants were enriched in SZ (Fig. S1A). When limited the analysis to the ultra-rare variants that potentially include pathogenic variants, BD still exhibited a higher number of variants compared to CTRL (P = 0.0252), while SZ did not (P = 1.00) (Fig. 1D). Notably, ultra-rare rRNA variants were enriched in BD (FET, P = 0.0270, Fig. 1E), while this enrichment did not survive multiple testing (adjusted P = 0.162, n of tests = 6) due to our limited sample size.

When applying linear models (LM) incorporating age, sex, pH, PMI, tissue, haplogroup, and depth to the above analysis, the enrichment of ultra-rare rRNA variants in BD remained significant (P = 0.00155). The enrichment of total ultra-rare variants in BD was suggestive by LM, though not statistically significant at the conventional threshold (P = 0.0508). No significant enrichment of total variants was observed in BD or SZ by LM (P = 0.554 and 0.306, respectively). The enrichment of D-loop variants in SZ was not significantly supported (LM P = 0.108). While fitting parametric distributions using linear models for these variant data presented certain challenges, the observed enrichment of ultra-rare variants—particularly rRNA variants—in BD appears to be relatively robust.

Four of six ultra-rare rRNA variants showed the MLC scores above 0.5, demonstrating high pathogenicity [47]. Particularly, one variant had an MLC score of 0.896, surpassing the m.3243A>G score of 0.851. Among all the ultra-rare variants, 11 had MLC scores of >0.5, including eight in BD, two in SZ, and one in CTRL, indicating enrichment in BD (two-sided FET, P = 0.0315 vs. CTRL). Six rRNA/tRNA variants were exclusively detected in BD (two-sided FET, P = 0.0270, vs. CTRL), while four missense variants were present across all three groups, and one nonsense variant was identified in SZ. The observed enrichment of ultra-rare variants with MLC > 0.5 in BD was primarily driven by rRNA/tRNA variants.

In the tRNA region, the m.3243A>G variant, a major causative variant for MELAS, was recurrently detected in two BD patients as ultra-rare tRNA variants. The VAFs were 1.1 and 8.6%, which are lower than those reported in the postmortem brains of MELAS patients, where VAFs ranged from 39.9–85.2% in Miyahara et al. [52]. No other ultra-rare tRNA variants were observed in CTRL or SZ, demonstrating the specificity of ultra-rare tRNA variants to BD in this study. While the difference in ultra-rare tRNA variant count was not statistically significant due to our limited sample size (two-sided FET, P = 0.495 for BD vs CTRL), the finding of recurrent m.3243A>G variants is consistent with Munakata et al. [15], which reported two recurrent m.3243A>G variants in the brain of 15 BD patients but none in 14 controls. A meta-analysis incorporating the current findings and those from Munakata et al. [15] suggests a trend of association between m.3243A>G and BD (BD vs. CTRL = 4/69 vs. 0/67, two-sided FET P = 0.120).

We noticed four ultra-rare LoF variants in BD and SZ, two in BD and two in SZ with independent validation, while none in CTRL (Fig. 1E, Fig. S1B). The homoplasmic variants of The LoF variants are virtually none as homoplasmy in the gnomAD database [35], indicating its strong pathogenicity. Despite the non-significance in statistical assessment (BD vs CTRL: P = 0.495, SZ vs CTRL: P = 0.496), our findings of LoF variants only in BD/SZ suggest promising roles of LoF variants for psychotic disorders. Notably, three of the four LoF variants were m.6289A>AC variants in MT-CO1, indicating the vulnerability of this site. The mtDNA of two patients (BD-37, SZ-28) was classified within the H3ag1 haplotype, a haplotype unique to these two samples. The mtDNA of another patient (BD-31) was classified as J1c2a3. BD-37 had 13 of 16 homoplasmic variants present in SZ-28 (Fig. S1C), suggesting a possible relationship between these two patients; however, it is premature to draw definitive conclusions in the absence of pedigree information.

Heteroplasmic variants with potential pathogenicity for BD are listed in Table 1. Given the high proportion of the H haplogroup among the donors, no significant enrichment of this haplogroup was observed in samples containing these variants (two-sided FET, P = 0.169 in BD). The potential relationship between the m.6289A>AC variants in MT-CO1 and the H3ag1 haplotype warrants further investigation. In contrast, no significant differences were observed in synonymous or missense variants in BD or SZ, and there was no enrichment of specific classes of high/moderate VAF variants in SZ, except for D-loop and LoF variants (Fig. S1).

Table 1 Potentially pathogenic variants in BD and SZ.
Full size table

Single-molecule analysis for low-VAF variants

Second, we detected low-VAF variants in BD, SZ, and CTRL by single-molecule analysis. The average base quality in single-molecule analysis was 83.3, with a standard deviation of 1.67 (Fig. 2A). Our theoretical sequencing accuracy far exceeds the estimated per-base mutation frequency of approximately 1.0 × 10−5 by Kennedy et al. [38]. The VAF distribution exhibited a trend of bimodal pattern, with a trough around VAF = 1% (Fig. 2B). This observation indicated a qualitative difference between low-VAF (VAF > 1%) and high/moderate-VAF (VAF > 1%) variants, supporting the validity of the clonality cutoff of Kennedy et al. [38] and our distinction of low-VAF and high/moderate-VAF variants. A total of 52,312 low-VAF heteroplasmic variants (VAF < 1%), including 37,953 ultra-rare variants, were identified by our single-molecule pipeline (Fig. 2C, Fig. S2).

Fig. 2: Low-VAF mitochondrial variants detected by single-molecule analysis.
figure 2

A Base quality distribution of all the variants detected by single-molecule analysis. The estimated per-base mutation frequencies in mtDNA (1.0 × 10−5) and our cut-off threshold (1.0 × 10−7) are indicated by red dotted lines. B VAF distribution of all the variants detected by single-molecule analysis. The distribution around VAF = 1% is magnified on the right. C Circos plot displaying the distribution of low-VAF variants in BD and CTRL. The inner layer plots all the detected variants in single-molecule analysis with their VAFs.

Full size image

When calculating the per-base alternative allele count of low-VAF variants, the bulk pipeline yielded a value of 1.64 × 10⁻4 (mapping quality set to 60 for comparison), whereas the single-molecule pipeline produced a value of 1.64 × 10⁻5. This tenfold reduction observed in the single-molecule pipeline suggests improved accuracy, likely attributable to the removal of low-level technical artifacts that persisted in the bulk analysis. Notably, the error rate obtained from the bulk pipeline was lower than anticipated based on the average base quality score of 35.4 (corresponding to an expected error rate of 2.88 × 10⁻4), supporting the validity and potential conservativeness of the base quality estimation. This finding implies that the single-molecule pipeline should achieve even greater accuracy, with its theoretical base quality exceeding 70 (corresponding to an error rate of 1.0 × 10⁻7). The alignment between the theoretical base quality, the observed performance of the single-molecule pipeline, and the estimated mtDNA mutation rate (on the order of 10⁻5) supports the pipeline’s accuracy. Although orthogonal validation of single-molecule variants is inherently challenging, these findings suggest that the low-VAF variants detected by the single-molecule pipeline likely represent true mutations.

By leveraging this high level of precision, we identified a total of 52,312 low-variant allele frequency (low-VAF) heteroplasmic variants through single-molecule analysis, with an average sequencing depth of 4.76 × 103. This corresponded to a mean of 323 low-VAF variants per individual (52,312 variants across 162 individuals), which was considered sufficient to support subsequent analyses. At these 52,312 sites, the mean surrounding base quality was 85.8 ± 1.31 for alternative variants and 84.1 ± 1.35 for reference variants (mean ± standard deviation). Although the surrounding base quality was significantly higher for alternative calls (Mann-Whitney U-test, P = 0), both groups exhibited quality scores sufficient for the reliable detection of true variants. Furthermore, these scores were comparable to the overall mean base quality of all base calls (83.3 ± 1.67; see Fig. 2A). While we did not set minimum quality standards for surrounding sequence bases, these findings suggest that the low-VAF variants were unlikely to have resulted from poor sequencing contexts.

Per-base mutation frequency analysis

The per-base mutation frequencies averaged 1.19 × 10−5 in the D-loop region and 6.84 × 10−6 in the other regions in control samples, aligned with Kennedy et al.’s data [38] and estimates in the large-scale mitochondrial genome study [53]. The per-base mutation frequencies between BD and CTRL showed no significant differences by two-sided Welch’s t-tests (P = 0.137) (Fig. S3). Welch’s t-test was employed due to unequal variances between groups, as indicated by a two-sided F-test (P = 0.0165). We further analyzed the per-base mutation frequencies using a linear model with log-transformation. Disease status, age, sex, brain pH, PMI, and tissue type were included as covariates to accommodate the effect of potential confounding factors (Table S7). The distribution of per-base mutation frequencies approximately fit log-normal distribution (Fig. S4). Consistent with t-tests, we observed no significant differences in per-base mutation frequencies between BD and CTRL (P = 0.187) (Fig. 3A). No significant differences remained, even if stratifying the variants by genomic regions (Fig. 3A), functional effects (Fig. 3B), variant rarity (Fig. 3C), deleteriousness, or mitochondrial genes (Fig. S5). However, significant effects of age (P = 0.00564) and tissue type (P = 5.40 × 10−6) were observed (Table S7). The per-base mutation frequencies tended to increase with aging (Fig. 3D). This observation reinforced prior evidence of age-associated systemic mutation rate increases [38], and suggested potential brain-region-specific mutational events. The latter possibility warrants further investigation to clarify the role of regional variation in mutation events. When separating the analysis by the brain region, no effect of BD status for the per-base mutation frequencies was observed (Table S7). In this analysis, sequencing depths did not have a large effect on per-base mutation frequencies.

Fig. 3: Per-base mutation frequencies in BD and CTRL.
figure 3

A Per-base mutation frequencies of low-VAF variants on a log scale in BD and CTRL, stratified by all, D-loop, protein-coding, tRNA, and rRNA regions. B Relative per-base mutation frequencies on a log scale, with the mean of the CTRL group set to 1, stratified by synonymous, missense, LoF variants. C Relative per-base mutation frequencies on a log scale, stratified by common, medium-rare, and ultra-rare variants. For A–C, P-values calculated by log-transformed linear models are described at the top. D Age and per-base mutation frequencies. Star indicates the mean VAFs or relative value of mean VAFs.

Full size image

Reassuringly, no association was observed between the per-base mutation frequencies and Brain pH (P = 0.589) or PMI (P = 0.331), ruling out sample degradation as the main factor driving mutagenic events. The male-to-female ratios were different among BD, CTRL, and SZ, but no substantial effect of sex to the per-base mutation frequencies was observed (P = 0.220). When limiting the analysis to ultra-rare variants to exclude the possibility of sample cross-contamination that should be mostly included in the common or medium-rare variants, we observed consistent results (Fig. S6), indicating that our results were unlikely to be derived from sample cross-contamination. Similar to BD, no apparent increases in per-base mutation frequencies were observed in SZ (Fig. S7).

VAFs of low-level heteroplasmic variants

We subsequently analyzed the selection bias of low-VAF variants in BD/SZ vs CTRL by log-transformed linear mixed model (LMM). The distribution of VAF approximately fits the log-normal distribution, probably reflecting the exponential nature of mtDNA proliferation (Fig. S8). The log-transformed distribution of VAFs is slightly biased to the left, probably because mtDNA proliferation decelerates above certain VAF levels in a brain cell as a general tendency of an organism’s ecological proliferation.

BD and SZ showed no clear VAF differences compared to CTRL (P = 0.347 and 0.179, respectively). Consistent with this, no differences in VAFs were observed between BD/SZ and CTRL across variant regions and variant rarity (Fig. 4, Table S7). Sequencing depth was identified as the primary factor influencing VAFs in this analysis (Table S7). The single-molecule depths were highly variable among the 162 samples (Table S3). Samples sequenced at higher depths tended to exhibit lower VAFs, likely due to the increased sensitivity for detecting variants with low VAFs at greater sequencing depths.

Fig. 4: VAF analysis of low-level heteroplasmic variants in BD and SZ.
figure 4

A VAF changes of low-level heteroplasmic variants in BD and SZ compared to CTRL, as estimated by linear mixed models. The x-axis indicates log(Fold-change), and the y-axis indicates -log10(P-value). Variants are categorized into all, common, medium-rare, and ultra-rare variants. All the comparisons exhibited small FC changes with P > 0.05 (-log10(p-value) <1.30), suggesting that the hypothesis that BD patients have more VAFs than CTRL is rejected. B Similar to panel A, with variants categorized into D-loop, protein-coding, tRNA, and rRNA variants.

Full size image

The temporal gyrus exhibited slightly higher VAFs than the frontal cortex (FC = 1.035, P = 0.00527). However, this subtle difference should be interpreted cautiously, given the large effect of sequencing depths. If we performed the VAF analysis separately in each brain region, BD and SZ still showed no VAF differences (Table S8). Age, sex, pH, PMI, and haplogroup did not significantly affect the VAFs in this analysis (Table S8). Collectively, these findings suggest an absence of biased proliferation or elimination of mutated mtDNA among the BD, SZ, and CTRL.

To exclude the possibility of sample cross-contamination, we analyzed the VAFs of ultra-rare variants and observed consistent results (Fig. S9). Thus, our results were unlikely to be derived from sample cross-contamination. Expectedly, the VAFs of common and medium-rare were higher than those of ultra-rare variants by LMM with variant rarity and the same covariates described above (FC = 1.131 and 1.027, P = 5.83 × 10−33 and 1.38 × 10−13, respectively). This observation was consistent with the assumed weaker selection pressure on common variants, supporting the validity of our selection bias analysis for mitochondrial variants.

Mutation pattern analysis

We analyzed mutation patterns of mitochondrial SNVs based on trinucleotide contexts. C>A mutations were more prevalent in the low-VAF variants (VAF < 1%) than in the high/moderate-VAF variants (VAF > 1%), showing a cosine similarity of 0.636 (Fig. 5A). This is consistent with previous findings from bulk analyses, where C>A mutations were infrequent [54, 55]. Cosine similarity analysis revealed high similarity in mutation patterns between BD/SZ and CTRL (cosine similarity = 0.959 and 0.972 for BD vs CTRL and SZ vs. CTRL) (Fig. 5B). These similarities were also observed when stratifying the variants by tRNA, rRNA, protein coding, and control regions (Fig. 5C). These results indicated comparable mutational profiles across disease statuses and variant regions. The uniformity in mutational patterns across the mitochondrial genome suggests a consistent mutagenic mechanism, aligning with observations by Kennedy et al. [38].

Fig. 5: Mutation spectrum of mitochondrial DNA.
figure 5

Bar charts illustrating mutation count (y-axis) for 96 mutation patterns (x-axis). A Mutation spectrums in low-VAF and high/moderate-VAF variants. B Mutation spectrums in BD, CTRL, and SZ. C Mutation spectrums in all, D-loop, synonymous, nonsynonymous, tRNA, and rRNA variants.

Full size image

Discussion

In this study, we observed an enrichment of heteroplasmic rRNA variants and promising pathogenic variants, including recurrent m.3243A>G variants and four LoF variants in BD/SZ patients. This observation is consistent with our hypothesized model that a limited number of specific pathogenic variants with VAF > 1%, which arise early in development or are maternally inherited, impair neural functions. We observed no significant increase in the per-base mutation frequency and VAFs of low-level heteroplasmic variants in either BD or SZ. These findings do not support our second model, which hypothesizes that the accumulation of non-specific low-level heteroplasmic variants—arising later in development or during aging—contributes to impaired neural function.

Our findings of the recurrent moderate-VAF m.3243A>G variants, an established pathogenic variant linked to mitochondrial disease MELAS, along with the previous finding of recurrent m.3243A>G variants with VAF > 1% in BD brain [15] indicate that such pathogenic variants maternally transmitted or arising in early developmental stages contribute to psychiatric symptoms. This data also aligns with our previous findings of m.3243A>G variants in BD peripheral tissues [14]. Looking from the other side, patients with mitochondrial diseases are more likely to have BD [16, 17]. Especially, patients carrying the m.3243A>G variants are at particularly high risk for BD with an odds ratio of 12.8 [17]. Collectively, these data indicate a promising association between BD and the m.3243A>G variants. If the m.3243A>G variants were assumed to be observed in 1% of healthy individuals’ brains, the odds ratio of this variant for BD would be 6.09 ([4/65]/ [1/99]). To achieve a power of 80% with a significance level of P < 0.05, an expanded analysis involving 151 samples in each group is required for conclusive statistical assessment. Besides BD, the m.3243A>G variant has also been identified in postmortem brain tissue and peripheral samples from individuals with schizophrenia [14, 56]. The potential association between these conditions and mitochondrial variants warrants further investigation.

Our study identified six ultra-rare heteroplasmic rRNA variants, indicating a potential association between rRNA heteroplasmic variants and BD. Notably, the variants with MLC scores of >0.5 were enriched in BD, and four of six ultra-rare rRNA variants had MLC scores of >0.5, supporting their potential pathogenicity for BD. Some studies have suggested associations between rRNA variants and mitochondrial diseases [57] or Alzheimer’s disease [58]. However, few investigations have explored connections between rRNA variants and psychiatric disorders. We hypothesize that certain ultra-rare rRNA variants might impair protein translation in mitochondria, similar to tRNA disturbances. This could cause mitochondrial dysfunction and neuronal activity impairment, potentially contributing to psychiatric symptoms. Further studies are warranted to explore these rRNA variants’ molecular mechanisms and phenotypic effects in psychiatric disorders.

Besides m.3243A>G and rRNA variants, we found four ultra-rare LoF variants in BD/SZ and none in CTRL. The LoF variants in mitochondrial genome are highly constrained with virtually none observed as homoplasmy in gnomAD [35], reflecting the strong negative selection pressure and phenotypic disadvantage from LoF variants. The brain LoF variants in this study could affect the nervous system functions, possibly contributing to psychiatric symptoms. Despite insufficient statistical power for definitive conclusion, the contrast of individuals with LoF variants (four in 109 BD/SZ vs. none in 53 CTRL) indicates the promise of heteroplasmic LoF variants for understanding psychotic disorders.

Mitochondrial variant accumulation at single-molecule resolution has been examined in neurodegenerative diseases [29, 59]. However, no studies have examined mood disorders at single-molecule resolution, underscoring the novelty of our investigation. In this study, we found no increases in per-base mutation frequencies and VAFs of low-level heteroplasmic variants in BD or SZ. Besides, there were no apparent differences in trinucleotide-based mutation patterns among BD, SZ, and CTRL. The absence of significant differences in per-base mutation frequency and the VAFs of low-level heteroplasmic variants in BD does not support our model of nonspecific low-level heteroplasmic variant accumulation. Accordingly, our data does not support the hypothesis that low-VAF variants, which arise later in development or during aging, progressively degrade mitochondrial function and impair neural activity in BD. In contrast, we observed an enrichment of ultra-rare variants with VAFs > 1% in BD. These potentially pathogenic variants are likely to have originated independently of the accumulation of low-level heteroplasmic variants. Their probable sources include maternal transmission or de novo mutations arising early in development.

The oxidative stress is reported to increase in BD, probably damaging mitochondrial proteins [22, 60]. Oxidative stress induces the formation of 8-oxo-7,8-dihydroguanine (8-oxoG) from guanine residues. 8-oxoG has a propensity to mispair with adenine rather than cytosine during DNA replication, resulting in C>A transversions; however, such lesions are typically corrected by the DNA repair machinery. Although increased oxidative stress and DNA/RNA damage have been proposed in BD [38, 61], the frequency of C>A mutations did not differ significantly between BD and CTRL in this study. Therefore, increased DNA oxidation in BD was not evident in our dataset. This observation aligns with the findings of Kennedy et al., which argued that oxidative stress is not a major contributor to the accumulation of age-related mitochondrial mutations [38]. If oxidative stress is indeed relevant to BD, as suggested by previous studies, its primary impact may be on proteins or RNA, potentially resulting in neuronal dysfunction or mitochondrial abnormalities through alterations in protein function.

Our findings indicate that heteroplasmic variants above and below 1% VAF represent distinct variant categories, as illustrated by the bimodal distribution in Fig. 2. The variants with VAF > 1% are likely maternally transmitted or early-onset variants, which can reach higher VAFs through genetic bottlenecks during development. The variants with VAF < 1% should primarily represent variants arising later in development or aging. The clinical impact of heteroplasmic variants depends on the VAF threshold, which varies by variant types and tissues. Energy-demanding tissues like the central nervous system and cardiac tissue are hypothesized to have lower thresholds for clinical impact [62]. Besides, heteroplasmic variants with VAF near 1% could be localized within specific brain regions or cell types, where VAF may locally exceed 1%. The locally concentrated variants might surpass the threshold required for functional disruption, resulting in neural circuit dysfunction. The moderate-VAF heteroplasmic variants in this study, detected from bulks of cortex samples, might be localized to more focused areas in specific sub-regions.

Previous studies on heteroplasmic fractions were mainly discussed for rare mitochondrial diseases, and there was less argument on heteroplasmic fractions with regard to relatively milder phenotypes such as psychiatric disorders. Our study provides a new insight into heteroplasmic fractions and common diseases, consistent to the Stewart et al.’s model wherein highly pathogenic variants, when reaching a sufficient VAF, cause severe conditions such as mitochondrial diseases and, at lower VAFs, may contribute to relatively common disorders [63, 64]. Moderate-VAF variants, most of which probably arise earlier in development or are maternally transmitted, could be present across different tissues, including peripheral tissues, within the same individual. The enrichment of m.3243A>G in peripheral samples of BD patients [14] underscores the validity of examining peripheral samples in BD research.

This study is limited by a moderate sample size with insufficient power for rigorous multiple testing and assessment of postulated tRNA variant enrichment in BD. The sex disparity among the three groups was introduced during the age adjustment process. However, the sex disparity did not significantly affect the results of single-molecule analysis. The lack of maternal data precludes the determination of whether detected variants were maternally transmitted or de novo mutations. The absence of comprehensive clinical data, particularly metabolic markers associated with mitochondrial function and lithium response, constrains the interpretation of the clinical manifestations linked to mitochondrial variants. If such data are obtained in the future, they would substantially inform and advance future investigations in this area. This study included samples from both the frontal cortex and superior temporal gyrus tissue, which affect the mutation rates as indicated in Table S7. The differences derived from brain regions, including other brain regions such as hippocampus and amygdala, should be addressed in future research. The current short-read technology cannot confirm the heteroplasmic variants in the low-mappability genomic regions. This problem will be solved with long-read sequencing technology. High variability of the single-molecule depths for single-molecule analysis among the 162 samples (Table S3) was derived from our inability to control the PCR amplification bias in library preparation using custom mtDNA probes. Future studies employing methods to control PCR amplification bias would facilitate more precise comparisons among different groups.

Conclusion and future directions

This study finds the enrichment of potentially pathogenic mitochondrial variants with moderate VAFs in BD. Especially, m.3243A>G variants are the most promising single variant to explain BD. The LoF and rRNA variants are promising candidates for further investigations. Our single-molecule analysis did not reveal an increase in low-level heteroplasmic variants in BD. The potentially pathogenic mitochondrial variants with moderate VAFs observed in BD are likely attributable to maternal transmission or to de novo mutations arising early in development. These findings offer new insights that may inform the development of mitochondria-targeted therapies for BD, guided by genomic stratification based on mitochondrial variant profiles. Several compounds for mitochondrial diseases [65,66,67] could be repositioned for BD with mitochondrial etiology. Indeed, we are exploring potential compounds modulating mitochondria functions for BD treatment [68]. Investigations of brain mitochondria hold promise for advancing our understanding of pathophysiology and improving treatment strategies for BD.