The multi-tissue landscape of somatic mtDNA mutations indicates tissue-specific accumulation and removal in aging
Curation statements for this article:-
Curated by eLife
eLife assessment
The analysis of 89,000 independent somatic mtDNA mutations provides compelling evidence that allows the authors to refute the idea that reactive oxygen species (ROS) are a main driver of mtDNA mutagenesis, although ROS effects may still be tissue-dependent. These are fundamental results with convincing evidence, and they should appeal to a broad audience. The discovery of transversion mutations (C>A/G>T and C>G/G>C), which previously were assumed to be almost nonexistent, will nevertheless require additional validation.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Accumulation of somatic mutations in the mitochondrial genome (mtDNA) has long been proposed as a possible mechanism of mitochondrial and tissue dysfunction that occurs during aging. A thorough characterization of age-associated mtDNA somatic mutations has been hampered by the limited ability to detect low-frequency mutations. Here, we used Duplex Sequencing on eight tissues of an aged mouse cohort to detect >89,000 independent somatic mtDNA mutations and show significant tissue-specific increases during aging across all tissues examined which did not correlate with mitochondrial content and tissue function. G→A/C→T substitutions, indicative of replication errors and/or cytidine deamination, were the predominant mutation type across all tissues and increased with age, whereas G→T/C→A substitutions, indicative of oxidative damage, were the second most common mutation type, but did not increase with age regardless of tissue. We also show that clonal expansions of mtDNA mutations with age is tissue- and mutation type-dependent. Unexpectedly, mutations associated with oxidative damage rarely formed clones in any tissue and were significantly reduced in the hearts and kidneys of aged mice treated at late age with elamipretide or nicotinamide mononucleotide. Thus, the lack of accumulation of oxidative damage-linked mutations with age suggests a life-long dynamic clearance of either the oxidative lesions or mtDNA genomes harboring oxidative damage.
Article activity feed
-
-
Author Response
Reviewer #1 (Public Review):
The paper states that they observed a combined total of 77,017 single-nucleotide variants (SNVs) and 12,031 insertion/deletions (In/Dels) across all tissue, age, and intervention groups. Collectively, these data represent the largest collection of somatic mtDNA mutations obtained in a single study to date. However, A study with more somatic mtDNA mutations by the LostArc method (PMID 32943091) revealed 35 million deletions (~ 470,000 unique spans) in skeletal muscle from 22 individuals with and 19 individuals without pathogenic variants in POLG. Thus, the authors should reword this part to say that this study represents the largest collections of mouse mtDNA point mutations detected, but not the largest amount of mutations (deletions exceed this number).
Thank you for pointing this out. …
Author Response
Reviewer #1 (Public Review):
The paper states that they observed a combined total of 77,017 single-nucleotide variants (SNVs) and 12,031 insertion/deletions (In/Dels) across all tissue, age, and intervention groups. Collectively, these data represent the largest collection of somatic mtDNA mutations obtained in a single study to date. However, A study with more somatic mtDNA mutations by the LostArc method (PMID 32943091) revealed 35 million deletions (~ 470,000 unique spans) in skeletal muscle from 22 individuals with and 19 individuals without pathogenic variants in POLG. Thus, the authors should reword this part to say that this study represents the largest collections of mouse mtDNA point mutations detected, but not the largest amount of mutations (deletions exceed this number).
Thank you for pointing this out. When we wrote that sentence, we were more referring to small polymerase-based errors, as opposed to larger structural variants that likely arise from a different mechanism. However, the distinction between these two event classes is poorly defined. We have amended our statement and have added a citation to Lujan et al. Our statement now reads “We observed a combined total of 77,017 single-nucleotide variants (SNVs) and 12,031 small insertion/deletions (In/Dels) (≲15bp in size) across all tissue, age, and intervention groups. Collectively, these data represent the largest collection of somatic mtDNA point mutations obtained in a single study to date and is second only to Lujan et al. in terms overall In/Del counts (Lujan et al., 2012).” (Lines 252-256)
What is the theoretical limit of pt mutations in the mitochondrial genome, assuming only one pt mutation per genome? Doesn't 77000 detected independent pt mutations approach that limit? Can the authors estimate how many molecules contained two or more pt mutations? Did the analysis reveal any un-mutated regions implying an essential function? For example, on p.9 can the authors provide an explanation of why OriL and other G/C-rich regions were not uniformly covered as compared to the rest of the genome?
This is an interesting question and one we’ve given some thought to. In fact, this basic question was the inspiration for our recent Nucleic Acids Research paper (PMC8565317) where we asked how mutations were distributed in the genome. The short answer is that we likely exceed the limit for only dG site mutations (and only for G>A mutations, at that), but not the other reference sites. The reason is that there are only 2013 dG sites and the mutation spectrum is heavily skewed toward G>X (there are 47,680 dG site mutations, 42,924 of which are G>A). In comparison, we observe only 4,421 A>X, 9,277 T>X, and 15,632 C>X mutations, but with 5,629, 4,681, and 3,976 dA, dT, and dC genomic sites, respectively. Assuming the mutations are uniformly distributed along the genome (which they are not; see our NAR paper), then random binomial sampling would require a fair amount more mutations in order to reach saturation for the other genomic sites. The uneven distribution increases this number further.
With regard to the second question, we can’t actually do this estimation with this data set. The reason is because the ~77,000 mutations aren’t found in a single sample, but are distributed across may independent or semi-independent (i.e. different organs within a mouse), which means that most, if not all, of the mutations are necessarily on different mtDNA molecules.
With regard to the OriL and G/C rich regions, these presumably have some sort of secondary structure that prevents the sequencer from obtaining any useful information. However, this is all speculative and we don’t know why. Interestingly, human mtDNA doesn’t show this dip at the OriL, despite a similar function and location in the mtDNA.
Given that mitochondrial disease usually doesn't present until >60% of the genomes are affected, the very low level of detected pt mutations observed in the mouse (and presumably similar to human) would mean that they are well below a physiological level. Thus, these low-level pt mutations are well tolerated. Can the authors estimate a theoretical age of the mouse (well beyond their life span) where over 50% of the genomes carry at least one pt mutation?
The reviewer brings up a frequent noted point in mitochondrial biology that is very much worth addressing in this manuscript. The often-cited statistic that mitochondrial disease doesn’t present until ~60% of genomes are affected is, while true, only pertinent to overt mitochondrial diseases, such as LHON, MERRF, etc, where all or nearly all cells in an individual are affected by the mutation. However, the impact of mtDNA mutations is not only contingent on how many cells have the mutation, but also the fraction of mtDNA molecules within a cell that harbor the variant. Because the deleterious effects of a mtDNA mutation act at the level of individual cells, it is important to know both how many cells harbor a mutation as well as what the heteroplasmic level is within the cell before making claims on their pathological impact.
To date, nearly all studies on mtDNA mutations rely on bulk DNA analysis from thousands to millions of cells, which necessarily decouples variant phasing information between any two reads, resulting in a loss of important biological information such as the heteroplasmic level within any given cell. As such, with bulk sequencing it is impossible to tell the difference between a homoplasmic mutation in a small subset of cells and heteroplasmic mutation in all cells. In the first case, the cells harboring this mutation would be negatively impacted, whereas in the second example, it is unlikely. One can imagine a scenario where every cell contains a different homoplasmic pathogenic mutation which would negatively affect cellular function for every cell. In this case, mutations would be highly prevalent (100% of cells), yet individually rare. However, bulk sequencing would give the appearance that no mutation comes close to exceeding the phenotypic threshold. We highlight this issue in a recent review (Sanchez-Contreras and Kennedy, 2022; PMC8896747).
The point that the review brings up is extremely important, so we have added a section in the discussion related to heteroplasmy versus clones.
Also, the problem with this low level of pt mutations is that they are not physiological, the effect of the drug treatment causing a reduction in ROS-mediated transversions would not be expected to have a detectable effect on mitochondria. The improvement on mitochondrial seen by others is most likely independent of the mutations in the genome. There needs to be a cause and effect here and I don't see one.
It is important to note that we do not make the claim (no do we want to imply) that the reduction of mutations is the reason behind the improvements in mitochondrial function by these interventions. Instead, we believe that loss of ROS-linked mutations is a consequence of the mechanism by which these interventions work. We do hypothesize that the reduction in ROS-linked mutations suggests that “there is tissue specificity in how cells repair and/or destroy oxidatively damaged mitochondria and/or mtDNA resulting in a steady-state of ROS-linked mutations.” (Lines 551-553) and that “We propose that rather than the incidence and impact of ROS damage on mtDNA being minimal, recognition and removal of ROS-linked mutations are maintained at a steady state during aging.” (Lines 572-574).
In addition, as noted above, how “low level” these mutations are and their impact on cellular function is not easily determined in bulk sequencing studies, so a strong link between cause and effect is not an answerable relationship with this data set.
There's no mention in this paper and methodology about how point mutations in nuclear-encoded mtDNA (NUMTs) are excluded from the reads and I'm worried that these errors are being read as rare errors in the mtDNA genome. While NUMTs have been documented for decades, a recent report in Science (PMID: 36198798) documents how frequently and fluidly NUMTs occur. Can the authors provide a clear explanation of how mutations in NUMTs are excluded?
The reviewer is absolutely correct to call attention to this important aspect of mitochondrial biology. We don’t believe NUMTs are an important confounder in our data set for several reasons.
We used isogenic inbred C57Blk6/J which, frequently, were litter mates (siblings). Therefore, any mutations from NUMTS that are there would be expected to be uniform across samples, especially between tissues from a single sample animal. Unknown and variations of NUMTS would certainly be a potentially strong confounder in an outbred population, but the use of one isogenic inbred line for this study likely eliminates this confounder.
We used the mm10 reference genome which is based on the C57Blk6/J strain so any NUMTS derived variants present in our mtDNA data should preferentially align against the NUMT. Therefore, we perform a BLAST step of all reads containing at least one variant against the mm10. BLAST is much more sensitive to sequence variation compared to bwa but is far slower, so it is impractical to run as the initial aligner. We then reassign the read based to whatever genomic location has the lower e-score. The result is typically around a dozen reads are removed, demonstrating that NUMTS are not likely a major source of false mutations.
Because NUMTS are inherited, then any variants would be found across all the tissues and animals we used in this study. As part of our processing, we mark and remove variants shared between multiple individual samples.
We have made edits to the Methods section (Lines 198-206) to more explicitly highlight the filtering steps and the logic behind them. In addition, we have added a paragraph in the discussion that addresses NUMTs (Starting on line 642).
Reviewer #2 (Public Review):
A common problem in mutation analysis is that DNA damage (present in one strand) is difficult to separate from real mutations (present in both strands). One of the approaches to solve this problem based on independent tagging of the two strands by different unique molecular identifiers was developed by the authors about 10 years ago. This study summarizes the application of this method to a wide range of mouse tissues, ages, and drug treatment regimes. Much of the results confirm previous conclusions from this laboratory. This involves overall mutational levels of somatic mtDNA mutations (~10-6-10-5), their accumulation with age, the prevalence of GA/CT transitions, and their clonality. Although these results were not new, it is important that these were confirmed in a single study with high confidence in a huge number of independent mutations.
We thank the reviewer for the comment and really hope this data set will be of significant use to other researchers given its breadth of sample types and large number of mutations.
What really sets this study apart from other studies is the detection of a large proportion of transversion mutations, primarily of the C>A/G>T and C>G/G>C types. Transversions are traditionally considered 'persona non grata' in mtDNA mutational spectra and are typically associated with errors of mutational analysis (which they in fact are). The presence of these mutations in both strands of the duplex makes a good case that these mutations are real, rather than converted damage. However, because this is such a novel discovery and because regular controls do not work (I mean, for example, that these mutations never clonally expand. If there is a clonal expansion, then the mutation is real, only real mutation can expand. But in the case of non-expandable C>A/G>T and C>G/G>C this control does not help to validate these mutations), it would be nice to provide extra assurances that this is not some kind of artifact that somehow slipped through the ds sequencing procedure. I would recommend including in the supplement the data on the abundance of single-stranded base changes as detected by ds sequencing (i.e., changes confirmed in one and not in the other strand of a given molecule). An unusually high presence of such single-stranded changes of the C>A/G>T and C>G/G>C type would be a red flag for me. If ratios of single and double-stranded mutations were similar for transitions and transversions - that would reassure me and hopefully the reader.
Furthermore, a similar excess of C>A/G>T and C>G/G>C has been observed in a recent paper by Abascal 2021 (cited in the manuscript). In that paper, a UMI- free, but otherwise very similar ds sequencing approach in nuclear DNA (BotSeqS) was demonstrated to suffer from an artifact causing (among other effects) an excess of C>A/G>T and C>G/G>C transversions. This artifact is related to end repair and nick-translation of DNA fragments during library preparation. Because BotSeqS is very similar to ds sequencing, we expect that same artifact may be taking place in the study under review. We recommend running checks similar to those undertaken by Abascal et al (which include, at the very minimum, checking the distribution of the C>A/G>T and C>G/G>C transversions within the reads (artifacts tend to be concentrated towards the ends of the reads).
The reviewer is absolutely correct to bring up this extremely important point. We have addressed these concerns in two ways that are addressed on Lines 332-361. 1) by performing an analysis of the single-stranded consensus data, which is a measure of PCR artifacts that frequently arise as a function of DNA damage, across all the tissues of the aged cohort. We noted no differences between tissues, which indicates that the amount of ROS-induced PCR artifacts is no different between the tissues. Thus, it would require a different rate at which ROS artifacts lead to false “Duplex consensus” variants that is tissue specific. The analysis is presented in Figure 3-figure supplement 2. 2) we have included an experiment in which we show that treatment of post-fragmented DNA with FPG, a glycosylase that targets Fapy-dG and 8-oxo-dG, does not differ from untreated control DNA. Because Duplex-Seq requires that both strands of a parent DNA molecule be present to form a final Duplex Consensus Sequence, the scission of one strand by the lyase activity of FPG would prevent the formation of this final consensus and prevent this sort of error from “bleeding through”. This analyses can now be found in a Figure 3-figure supplement 3.
Of note, even if transversions detected in this study prove to be artifacts of the Abascal type (likely) they still may reflect real ss damage in mtDNA (not instrumental artifacts, like sequencing errors or in vitro DNA damage). This is supported by the strong variation in the levels of transversions across tissues and as a result of the ameliorating drug intervention. Artifacts, in contrast, would be expected to be at a constant level. This logic, however, does not differentiate between real ds mutations and ss damage. So UMI-based ds sequencing evidence remains the only (though very strong) independent proof. So, in my view, whereas the jury may be still out on whether the observed transversions are true ds mutations or some kind of single-stranded damage, this is a critically important observation. The evidence of ss damage greatly varied between tissues and detected with such precision on a single molecule level is a very important finding as well.
Out of caution, I would recommend mentioning the above-stated uncertainty and noting that more research is needed to fully confirm that C>A/G>T and C>G/G>C changes detected in this study are indeed double-stranded mutations.
We agree. Together with comments from Reviewer #1 regarding NUMTs (Comment #5), we have added a paragraph in the Discussion about potential alternative explanations for our observations.
-
eLife assessment
The analysis of 89,000 independent somatic mtDNA mutations provides compelling evidence that allows the authors to refute the idea that reactive oxygen species (ROS) are a main driver of mtDNA mutagenesis, although ROS effects may still be tissue-dependent. These are fundamental results with convincing evidence, and they should appeal to a broad audience. The discovery of transversion mutations (C>A/G>T and C>G/G>C), which previously were assumed to be almost nonexistent, will nevertheless require additional validation.
-
Reviewer #1 (Public Review):
This work applies duplex sequencing to study point mutations in mice across tissues in young (4.5 months) and old mice (26 months). In this study, they identified 89,000 independent somatic mtDNA mutations representing the largest collection of somatic 'point' mtDNA mutation (not considering mtDNA deletions). They find that mtDNA mutations accumulate linearly with age in a clock-like manner but are not uniformly represented in all tissues. This indicates a likely constant 'clock-like' accumulation analogous to what is seen in the nuclear genome. This part of the paper is a comprehensive extension of work done by Arbeithuber et al., 2020. They also find variability between tissues of the ROS-linked (transversions) mutations. Similar to prior work by Kennedy and Loeb (2013 Plos Genetics) they conclude that …
Reviewer #1 (Public Review):
This work applies duplex sequencing to study point mutations in mice across tissues in young (4.5 months) and old mice (26 months). In this study, they identified 89,000 independent somatic mtDNA mutations representing the largest collection of somatic 'point' mtDNA mutation (not considering mtDNA deletions). They find that mtDNA mutations accumulate linearly with age in a clock-like manner but are not uniformly represented in all tissues. This indicates a likely constant 'clock-like' accumulation analogous to what is seen in the nuclear genome. This part of the paper is a comprehensive extension of work done by Arbeithuber et al., 2020. They also find variability between tissues of the ROS-linked (transversions) mutations. Similar to prior work by Kennedy and Loeb (2013 Plos Genetics) they conclude that ROS-linked mutations do not accumulate significantly with age. Lastly, the authors apply this knowledge and technique to interrogate whether mtDNA mutations are affected by two known treatments, elimipretide and nicotinamide mononucleotide, that have been shown to improve mitochondrial function and reverse apparent aging phenotypes. Here they demonstrate that these treatments reduced the low level of ROS accumulated mtDNA mutations seen in untreated tissues.
Comments:
The paper states that they observed a combined total of 77,017 single-nucleotide variants (SNVs) and 12,031 insertion/deletions (In/Dels) across all tissue, age, and intervention groups. Collectively, these data represent the largest collection of somatic mtDNA mutations obtained in a single study to date. However, A study with more somatic mtDNA mutations by the LostArc method (PMID 32943091) revealed 35 million deletions (~ 470,000 unique spans) in skeletal muscle from 22 individuals with and 19 individuals without pathogenic variants in POLG. Thus, the authors should reword this part to say that this study represents the largest collections of mouse mtDNA point mutations detected, but not the largest amount of mutations (deletions exceed this number).What is the theoretical limit of pt mutations in the mitochondrial genome, assuming only one pt mutation per genome? Doesn't 77000 detected independent pt mutations approach that limit? Can the authors estimate how many molecules contained two or more pt mutations? Did the analysis reveal any un-mutated regions implying an essential function? For example, on p.9 can the authors provide an explanation of why OriL and other G/C-rich regions were not uniformly covered as compared to the rest of the genome?
Given that mitochondrial disease usually doesn't present until >60% of the genomes are affected, the very low level of detected pt mutations observed in the mouse (and presumably similar to human) would mean that they are well below a physiological level. Thus, these low-level pt mutations are well tolerated. Can the authors estimate a theoretical age of the mouse (well beyond their life span) where over 50% of the genomes carry at least one pt mutation?
Also, the problem with this low level of pt mutations is that they are not physiological, the effect of the drug treatment causing a reduction in ROS-mediated transversions would not be expected to have a detectable effect on mitochondria. The improvement on mitochondrial seen by others is most likely independent of the mutations in the genome. There needs to be a cause and effect here and I don't see one.
There's no mention in this paper and methodology about how point mutations in nuclear-encoded mtDNA (NUMTs) are excluded from the reads and I'm worried that these errors are being read as rare errors in the mtDNA genome. While NUMTs have been documented for decades, a recent report in Science (PMID: 36198798) documents how frequently and fluidly NUMTs occur. Can the authors provide a clear explanation of how mutations in NUMTs are excluded?
-
Reviewer #2 (Public Review):
A common problem in mutation analysis is that DNA damage (present in one strand) is difficult to separate from real mutations (present in both strands). One of the approaches to solve this problem based on independent tagging of the two strands by different unique molecular identifiers was developed by the authors about 10 years ago. This study summarizes the application of this method to a wide range of mouse tissues, ages, and drug treatment regimes. Much of the results confirm previous conclusions from this laboratory. This involves overall mutational levels of somatic mtDNA mutations (~10-6-10-5), their accumulation with age, the prevalence of GA/CT transitions, and their clonality. Although these results were not new, it is important that these were confirmed in a single study with high confidence in a …
Reviewer #2 (Public Review):
A common problem in mutation analysis is that DNA damage (present in one strand) is difficult to separate from real mutations (present in both strands). One of the approaches to solve this problem based on independent tagging of the two strands by different unique molecular identifiers was developed by the authors about 10 years ago. This study summarizes the application of this method to a wide range of mouse tissues, ages, and drug treatment regimes. Much of the results confirm previous conclusions from this laboratory. This involves overall mutational levels of somatic mtDNA mutations (~10-6-10-5), their accumulation with age, the prevalence of GA/CT transitions, and their clonality. Although these results were not new, it is important that these were confirmed in a single study with high confidence in a huge number of independent mutations.
What really sets this study apart from other studies is the detection of a large proportion of transversion mutations, primarily of the C>A/G>T and C>G/G>C types. Transversions are traditionally considered 'persona non grata' in mtDNA mutational spectra and are typically associated with errors of mutational analysis (which they in fact are). The presence of these mutations in both strands of the duplex makes a good case that these mutations are real, rather than converted damage. However, because this is such a novel discovery and because regular controls do not work (I mean, for example, that these mutations never clonally expand. If there is a clonal expansion, then the mutation is real, only real mutation can expand. But in the case of non-expandable C>A/G>T and C>G/G>C this control does not help to validate these mutations), it would be nice to provide extra assurances that this is not some kind of artifact that somehow slipped through the ds sequencing procedure. I would recommend including in the supplement the data on the abundance of single-stranded base changes as detected by ds sequencing (i.e., changes confirmed in one and not in the other strand of a given molecule). An unusually high presence of such single-stranded changes of the C>A/G>T and C>G/G>C type would be a red flag for me. If ratios of single and double-stranded mutations were similar for transitions and transversions - that would reassure me and hopefully the reader.
Furthermore, a similar excess of C>A/G>T and C>G/G>C has been observed in a recent paper by Abascal 2021 (cited in the manuscript). In that paper, a UMI- free, but otherwise very similar ds sequencing approach in nuclear DNA (BotSeqS) was demonstrated to suffer from an artifact causing (among other effects) an excess of C>A/G>T and C>G/G>C transversions. This artifact is related to end repair and nick-translation of DNA fragments during library preparation. Because BotSeqS is very similar to ds sequencing, we expect that same artifact may be taking place in the study under review. We recommend running checks similar to those undertaken by Abascal et al (which include, at the very minimum, checking the distribution of the C>A/G>T and C>G/G>C transversions within the reads (artifacts tend to be concentrated towards the ends of the reads).
Of note, even if transversions detected in this study prove to be artifacts of the Abascal type (likely) they still may reflect real ss damage in mtDNA (not instrumental artifacts, like sequencing errors or in vitro DNA damage). This is supported by the strong variation in the levels of transversions across tissues and as a result of the ameliorating drug intervention. Artifacts, in contrast, would be expected to be at a constant level. This logic, however, does not differentiate between real ds mutations and ss damage. So UMI-based ds sequencing evidence remains the only (though very strong) independent proof. So, in my view, whereas the jury may be still out on whether the observed transversions are true ds mutations or some kind of single-stranded damage, this is a critically important observation. The evidence of ss damage greatly varied between tissues and detected with such precision on a single molecule level is a very important finding as well.
Out of caution, I would recommend mentioning the above-stated uncertainty and noting that more research is needed to fully confirm that C>A/G>T and C>G/G>C changes detected in this study are indeed double-stranded mutations.
-
