Mitochondrial DNA methylation signals in WGBS are mostly technical artifacts

For years, scientists have debated whether mitochondrial DNA carries its own layer of epigenetic regulation in the form of cytosine methylation. A new analysis of hundreds of whole-genome bisulfite sequencing datasets now suggests that most of the methylation signals reported in mitochondrial DNA are not biology at all, but the product of technical artifacts, above all the failure to fragment the DNA before bisulfite conversion. The study, published in Epigenetics & Chromatin by Jing-Jing Yin, Zhi-Chao Gong and Shi-Qing Mao, systematically dissected the factors that distort mitochondrial methylation measurements and offers a framework for separating technical noise from potential genuine signals.

Whole-genome bisulfite sequencing, or WGBS, is widely regarded as the gold standard for mapping DNA methylation at single-base resolution. The method exploits a simple chemical principle: sodium bisulfite converts unmethylated cytosines into uracils, which are then read as thymines during sequencing, while methylated cytosines remain unchanged. The fraction of unconverted cytosines at each position is therefore interpreted as methylation. To monitor conversion efficiency, laboratories typically spike in unmethylated lambda phage DNA as an internal control, since any residual methylation detected in lambda DNA indicates incomplete conversion.

The problem, the authors argue, is that lambda DNA is a poor proxy for the mitochondrial genome. Mitochondrial DNA is a small, circular, double-stranded molecule that lacks nucleosome packaging and adopts distinctive secondary structures, particularly a supercoiled conformation in the D-loop region. These topological features can prevent complete denaturation and impede bisulfite penetration of the DNA strands. When intact or near-intact mitochondrial DNA enters the conversion reaction, incompletely converted cytosines are erroneously read as methylated, systematically inflating both CpG and non-CpG methylation estimates. Linear lambda DNA simply does not reproduce these structural barriers, so a clean lambda control does not guarantee accurate methylation calls in the circular mitochondrial genome.

To quantify these effects, the team assembled 421 publicly available zebrafish WGBS datasets spanning a broad range of tissues, developmental stages and experimental conditions, and processed them with a consistent bioinformatic pipeline. They complemented this with 40 datasets from human and mouse BioProjects. Methylation was quantified separately in CpG, CHG and CHH sequence contexts, and only cytosines covered by at least ten reads were retained. A threshold of 2 percent CpG methylation, calibrated against the observed lambda DNA distribution, was used to distinguish background-level from elevated signal.

The results revealed striking heterogeneity. Among the 304 zebrafish datasets with lambda spike-in data, lambda methylation was generally close to background, yet a subset of samples showed conspicuously elevated mitochondrial methylation across all three sequence contexts. This simultaneous enrichment of CpG and non-CpG signals contrasts sharply with the CpG-dominated methylation architecture of animal genomes and exceeds what the limited capacity of DNA methyltransferase enzymes would predict for non-CpG cytosines. Of the datasets with lambda data, 191 showed low methylation in both mitochondrial DNA and lambda DNA, but 91 displayed elevated mitochondrial CpG methylation despite low lambda methylation, and 20 showed elevated methylation in both. Among the 117 datasets without lambda spike-in data, 106 exceeded the 2 percent mitochondrial methylation threshold.

The single most influential factor was DNA fragmentation. The researchers classified the datasets into a Fragmented group, in which sonication or enzymatic shearing of genomic DNA was explicitly reported before library construction, and an Unmentioned group, in which fragmentation was not documented. Samples with documented fragmentation showed mitochondrial CpG methylation levels that clustered near background, whereas the undocumented group displayed a much broader distribution including a large subset with markedly elevated levels. A two-sided Mann-Whitney U test found a highly significant difference between the groups, with P less than 0.0001. The authors caution that the Unmentioned category reflects incomplete reporting rather than a confirmed absence of fragmentation; if some of those samples were in fact fragmented, the true effect would be even stronger than measured.

Kit annotations also tracked with the artifact. Elevated mitochondrial methylation was concentrated in datasets prepared with the EpiGnome Methyl-Seq Kit, where 46 of 54 datasets exceeded the threshold, and the TruSeq DNA Methylation Kit, where 47 of 53 did. Notably, the protocols associated with these kits do not specify a separate fragmentation step before bisulfite conversion, relying instead on fragmentation occurring during the bisulfite treatment itself. In a linear mixed-effects model with laboratory as a random intercept, library preparation kit was significantly associated with mitochondrial CpG methylation variation, whereas adding bisulfite conversion kit did not significantly improve model fit. Because kits were unevenly distributed across laboratories, the authors interpret this as a dataset-level technical effect rather than evidence that any single kit independently caused the elevated signals.

One long-standing hypothesis held that strand-specific sequencing bias might explain false-positive mitochondrial methylation, partly because an apparent enrichment of non-CpG methylation on the light strand had previously been interpreted as biologically meaningful. The new analysis argues against this. In datasets with documented fragmentation, mitochondrial CpG methylation remained low across the full range of strand coverage imbalance. Counterintuitively, high methylation in undocumented samples was preferentially observed where strand coverage was balanced rather than imbalanced. A linear model including the interaction between fragmentation status and absolute strand coverage bias found the interaction not statistically significant, while fragmentation status overall showed a strong association with methylation levels. Elevated methylation above 2 percent was far more frequent in datasets without reported fragmentation, with an odds ratio of 17.64. The authors conclude that strand bias is unlikely to be a major independent driver of the artifactual signals.

The pattern was not confined to zebrafish. In the human and mouse datasets, elevated mitochondrial CpG methylation above 2 percent appeared in 6 of 20 human and 5 of 20 mouse datasets, with several samples exceeding 30 to 60 percent. These elevated estimates occurred across datasets with and without documented fragmentation and were not restricted to a single conversion or library preparation kit, and some appeared even in samples with low lambda methylation. The authors note that the available metadata were insufficient to assign the signal to a specific protocol component, but the cross-species consistency suggests that the same technical factors inflate mitochondrial methylation estimates in vertebrate WGBS data generally.

The findings do not entirely close the door on mitochondrial methylation. The authors explicitly state that the possibility of low-abundance, functionally relevant methylation at specific sites cannot be categorically excluded by their study design, and they acknowledge that some studies reporting methylation signals used orthogonal methods such as HPLC-MS, which directly measures 5-methylcytosine abundance and is not subject to bisulfite conversion artifacts. Their analysis cannot directly address the validity of those mass spectrometry measurements, which might reflect nuclear DNA contamination, incomplete mitochondrial purification, or signals from nuclear mitochondrial DNA segments, or alternatively a genuine residual pool of methylation below the WGBS detection threshold. Resolving the question will require rigorous mitochondrial isolation, depletion of nuclear contaminants, and bisulfite-free single-molecule approaches such as Oxford Nanopore sequencing with appropriate controls. In the meantime, the authors urge caution: elevated mitochondrial methylation in WGBS data, especially when accompanied by substantial non-CpG signal or when lambda spike-in controls are unavailable, should be treated as presumptive evidence of technical artifact rather than authentic epigenetic regulation.

Subject of Research: Technical artifacts dominating mitochondrial DNA methylation measurements in whole-genome bisulfite sequencing

Article Title: Apparent methylation of mitochondrial DNA in WGBS is dominated by technical artifacts

Article References: Yin, J.-J., Gong, Z.-C., & Mao, S.-Q. (2026). Apparent methylation of mitochondrial DNA in WGBS is dominated by technical artifacts. Epigenetics & Chromatin, 19(1), Article 44. https://doi.org/10.1186/s13072-026-00692-z

Image Credits: AI Generated

DOI: 10.1186/s13072-026-00692-z

Keywords: mitochondrial DNA, DNA methylation, WGBS, bisulfite sequencing, technical artifacts, epigenetics, DNA fragmentation, zebrafish, lambda DNA spike-in, library preparation kits, strand bias, cross-species analysis