Abstract
Patients with primary mitochondrial diseases present with fatigue and multi-system disease, are often lean, and die prematurely, but the mechanistic basis for this clinical picture remains unclear. Integrating data from 17 cohorts of patients with mitochondrial diseases (n=690), we find that clinical mitochondrial disorders increase resting energy expenditure, a state termed hypermetabolism. In a longitudinal cellular model of primary patient-derived fibroblasts from multiple donors, we show that genetic and pharmacological disruptions of oxidative phosphorylation (OxPhos) similarly trigger increased energy consumption in a cell-autonomous manner, despite near-normal OxPhos coupling efficiency. Hypermetabolism was associated with mtDNA instability, activation of the integrated stress response, increased extracellular secretion of age-related cytokines and metabokines including GDF15, as well as an accelerated rate of telomere erosion and epigenetic aging, and a reduced Hayflick limit. Finally, we generate a longitudinal RNASeq and DNA methylation resource dataset, which reveals conserved, energetically demanding, genome-wide recalibrations to OxPhos dysfunction. Hypermetabolism, or the increased energetic cost of living in mitochondrial diseases, has important biological and clinical implications.
Introduction
Mitochondrial diseases are caused by mutations in either the mitochondrial (mtDNA) or nuclear (nDNA) genomes, which impair oxidative phosphorylation (OxPhos) and the ability to convert food substrates into ATP 1. However, cellular dysfunction arises even when ATP levels are normal 2–4, suggesting that energy deficiency may not be the primary disease initiator. In animal models of mitochondrial disease, OxPhos defects trigger nuclear transcriptional responses, including the integrated stress response (ISR) 3,5–8, and downstream gene products such as growth differentiation factor 15 (GDF15) are secreted systemically where they impact metabolic functions 9,10. This implicates conserved systemic signaling pathways in the pathogenesis of mitochondrial disease 11. Considering that these stress pathways entail fundamentally energetically-demanding cellular processes, OxPhos defects could therefore increase energy consumption at the cellular and organismal levels. However, the metabolic costs of cellular and systemic recalibrations in mitochondrial disorders have not been defined. Here we examine this question in clinical datasets and in two in vitro models.
Clinically, OxPhos dysfunction causes a broad spectrum of multi-system disorders where symptoms include, among others, fatigue and exercise intolerance 12,13. As a result, most patients with mitochondrial diseases curtail physical activity and exercise 14,15. The reduced mitochondrial capacity to oxidize substrates16 coupled to minimal physical activity levels would be expected to promote positive energy balance and body fat accumulation, leading to obesity. However, patients with mitochondrial diseases are rarely obese. In fact, patients with moderate to severe disease on average classify as underweight 17. Although gastro-intestinal symptoms that limit food intake or absorption could contribute to this phenotype, the rarity of obesity in mitochondrial disease remains a clinical paradox. However, this may be resolved by the counterintuitive notion that mitochondrial OxPhos dysfunction may not decrease energy consumption and expenditure, but may rather increase the energetic cost required to sustain basic physiological functions.
Living organisms avoid thermodynamic decay to grow and survive by consuming energy. The amount of energy expended relative to the minimal metabolic rate required to sustain life is defined as metabolic efficiency. Strong evolutionary pressures have optimized metabolic efficiency in organisms, thereby minimizing the amount of ATP required to sustain life 18. One evolutionary strategy includes the choice of metabolic pathways to derive ATP (OxPhos vs glycolysis), which have different ATP yields and metabolic costs 19. Within cells, metabolic costs arise mainly from transcription/translation processes (∼60% of total energy demands), the maintenance of ionic balance, as well as organelle biogenesis and degradation 20,21, which includes mitochondrial turnover. Mitochondrial biogenesis comes at a substantial cost because of their extensive proteome 19. In mitochondrial diseases, the intracellular heterogeneous mixture of mitochondria with mutant and wild-type mtDNA (i.e., heteroplasmy) triggers exaggerated biogenesis 22, a phenomenon predicted to increase the basal metabolic cost of organelle maintenance and total energy expenditure 23. Accordingly, resting energy expenditure (REE) in animal models of mitochondrial OxPhos dysfunction is elevated by 15-85%, including in Crif1-/- mice with impaired mitochondrial translation 10, Clpp-/- mice with deficient proteostasis 24, Polg mutator mice 10, ANT1-/- mice with impaired ATP/ADP exchange 25, and ATP6-mutant flies 26. Therefore, we reasoned that similarly, patients with OxPhos defects would exhibit impaired metabolic efficiency and increased REE – a state known as hypermetabolism. Other causes of partial OxPhos dysfunction including mutations in nuclear genes encoding respiratory chain assembly factors like SURF1 27, which cause disease and decrease lifespan in humans 28, could also trigger hypermetabolism.
Shortened lifespan is a ubiquitous feature of mitochondrial diseases 29–31 and most animal models with OxPhos dysfunction die prematurely 32–35. But is there a causal link between hypermetabolism and lifespan in humans? Among healthy individuals, elevated REE or hypermetabolism measured by indirect calorimetry (oxygen consumption, VO2) predicts more rapid age-related physiological decline 36 and independently predicts 25-53% higher mortality over the following 20-40 years 37,38 – an effect double than that incurred by smoking cigarettes 38. In human stem cells, hypermetabolism was also correlated with senescence and other aging phenotypes 39. Mechanistically, multiple processes compete for limited energetic resources within cells 40,41, particularly under energy-restricted conditions. Because some cellular operations are prioritized over others 20, the energetic cost of stress responses and their associated increase in transcription/translation can inhibit growth and cell division, even triggering premature senescence 42,43. Recently, it was reported that excessive activation of the ISR itself inhibits cell population growth 8. Thus, OxPhos-induced ISR activation and the resulting hypermetabolism could curtail growth and/or cause premature death by forcing an energetic tradeoff between stress-related cellular recalibrations in competition with growth.
Taken together, the observations that: i) genetic mitochondrial OxPhos defects trigger integrated stress responses, ii) cells operate under energetic constraints where the prioritization of stress responses and transcription/translation costs can precipitate senescence, and iii) decreased metabolic efficiency predicts shorter lifespan in humans and other animals, lead to the following hypothesis for mitochondrial disease: genetic defects causing OxPhos dysfunction trigger hypermetabolism both physiologically and cell-autonomously, a phenotype associated with premature mortality.
Here we test this hypothesis by re-analyzing data from multiple clinical cohorts of primary mitochondrial diseases with direct and indirect assessments of energy expenditure and lifespan, and via longitudinal in vitro studies in patient-derived human fibroblasts. This cellular system provides high temporal resolution, repeated-measures of bioenergetic and multi-omic molecular recalibrations across the cellular lifespan. Using this model, we show that both genetic and pharmacological mitochondrial OxPhos defects trigger marked hypermetabolism in a cell-autonomous manner. We identify mtDNA instability, activation of the ISR, increased secretory activity, and transcriptional upregulation of transcriptional/translational stress pathways as potential contributors to hypermetabolism. Finally, we report that OxPhos defects and hypermetabolism are linked to accelerated telomere shortening and epigenetic aging in fibroblasts, and provide a publicly available longitudinal dataset to query epigenetic and transcriptional signatures conserved across both cellular models. Together, these translational data implicate hypermetabolism as a pathophysiological feature of mitochondrial diseases.
Results
Meta-analysis of metabolic rate and physiology in primary mitochondrial disease
To test the hypothesis that mitochondrial OxPhos defects are associated with increased energy expenditure and shortened lifespan (Figure 1A), we integrated and re-analyzed data from a total of 17 cohorts representing a total of 690 subjects with mitochondrial diseases and 225 healthy controls (provided by the authors or directly from publications) (Table 1). The heterogenous mixture of functional and dysfunctional mitochondria within single cells is well known to cause mitochondrial hyperproliferation and increase mtDNA copy number 44–47, as illustrated within a single patient skeletal muscle cell in Figure 1B. Increased biogenesis must naturally incur increased energy expenditure at the cellular level 23, which we reasoned may translate to elevated whole-body REE.
In subjects with mitochondrial diseases, resting heart rate, which correlates with whole-body REE 48, was on average 10.7% higher than healthy controls (p<0.01, Figure 1C). This tachycardia reached up to +46% when patients and controls performed mild exercise at the same absolute workload (data not shown). Both at rest and during mild physical activity, as initially reported in a small study 49, patients had on average 244% higher blood or urine catecholamine levels (p<0.05, Figure 1D), particularly norepinephrine (NE), a neurohormone sufficient to elevate REE when administered systemically to healthy individuals 50. To quantify REE in mitochondrial disease patients, resting whole body VO2 expressed relative to body weight was used as it is the most direct available measure of mass-specific energy expenditure. Strikingly, VO2 measured by indirect calorimetry across 6 cohorts of patients with mtDNA defects was on average 30% higher at rest (p<0.0001), a difference characterized by a very large effect size (Hedge’s g=2.4, Figure 1E). Notably, elevated VO2 was exaggerated by more than half (+51%) during mild physical activity in mitochondrial diseases (see Figure 1E). These substantially higher REE values reveal increased energy consumption (i.e., lower metabolic efficiency) in mitochondrial disease, at rest and particularly during mild physical challenges.
The increase in REE is particularly striking given that patients with mitochondrial diseases on average have lower muscle mass 17, which is the major site of activity-dependent energy consumption. Therefore, the lower muscle mass in patients would be expected to reduce energy expenditure, unless the tissues intrinsically exhibited impaired metabolic efficiency, and thus consumed more energy per unit time just to sustain homeostasis. Therefore, the meta-analysis of these clinical data from multiple cohorts combining hundreds of patients reveals an increased energetic cost of living per unit of body mass – or hypermetabolism – in mitochondrial diseases.
Physiologically, hypermetabolism is expected to produce a negative energy balance, expending more energy substrates than are ingested, generally preventing the accumulation of body fat. Accordingly, body mass index (BMI), a gross estimate of adiposity, was on average 9.8% lower (p<0.05) patients with mitochondrial diseases compared to controls (23% lower than national averages across 3 countries) (Figure 1F). In one study, fat mass index, a more precise indicator of body fat, was 21.9% lower in mitochondrial disease patients 17. Moreover, although not all patients are thin, patients with more severe disease manifestations tended to have lower BMI (r=-0.25, p=0.018) 17, suggesting that more severe mitochondrial OxPhos dysfunction in humans contribute to override the propensity to obesity. Again, this result is in line with those in animal models of OxPhos defects, which similarly show hypermetabolism and reduced adiposity 10,24–26.
This clinical picture of mitochondrial diseases marked by increased REE and reduced body fat was associated with a 3-4-decade reduction in lifespan among adults (Figure 1G) 31. In a 10-year longitudinal observational study from the UK, peak mortality in mixed genetic diagnoses of mitochondrial diseases occurs up to 3 decades earlier than the national reference (Figure 1H). In children with severe pediatric forms of mitochondrial diseases, including diseases caused by autosomal recessive respiratory chain defects (e.g., SURF1 mutations: median lifespan 5.4 years 28), lifespan can be reduced by >90%. Heterogeneity between genetic diagnoses also highlights possible mutation-specific effects on hypermetabolism (Extended Data Figure 1). Together, these multimodal physiological data establish hypermetabolism as a clinical feature of mitochondrial diseases, which could account for the rarity of obesity and also contribute to shortened lifespan in this population.
Longitudinal analysis of primary human fibroblasts with SURF1 mutations
To examine if mitochondrial OxPhos dysfunction alters the REE and lifespan in a cell-autonomous manner independent of clinical, medical, and socio-behavioral confounds, we next performed a longitudinal study of primary human fibroblasts with genetically-defined or pharmacologically-induced OxPhos dysfunction. We used cells with a stable nuclear mutation in SURF1 (Surfeit Locus Protein 1), which causes partial mis-assembly and dysfunction of respiratory chain complex IV (cytochrome c oxidase, COX) 51 , leading to Leigh syndrome and death in early childhood (see Figure 1G). Primary dermal fibroblasts were obtained from 3 patients with SURF1 mutations presenting with Leigh syndrome, and from 3 healthy donors with no known mitochondrial dysfunction (Control) (Extended Data Table 1-2). Each group included one female and two male donors. To capture both baseline as well as trajectories of metabolic parameters across the entire lifespan, we passaged each fibroblast line over multiple cellular generations until growth arrest, a model that recapitulates in vivo molecular features of human aging, including canonical age-related changes in telomere length 52 and DNA methylation 53. By sampling cells across the lifespan, longitudinal profiles of multiple cellular, bioenergetic, transcriptomic, epigenomic, and secreted molecular features can be modeled for each donor (Figure 2A). Although healthy cells survive for up to 250 days, here we limit our analyses to the maximal lifespan of SURF1-mutant cells, ∼150 days.
Beyond allowing longitudinal assessments of molecular and bioenergetic parameters as cells transition from early-, mid-, and late-life, one major advantage of time-resolved trajectories with repeated-measures is that this approach de-emphasizes potential bias of any single time-point and provides more accurate estimates of stable cellular phenotypes for each donor and treatment condition. The use of primary human cells obtained from multiple donors, compared to the same experiment repeated in immortalized cell line(s), also provides a more robust test of generalizability of the data. Throughout the text, we report standardized measures of effect sizes (Hedge’s g) where g>0.2 is considered a small, g>0.5 a medium, and g>0.8 represents a large effect size, which are considerably more informative to compare small groups (3 donors per group) 54.
SURF1 mutations cause hypermetabolism
We first examined the effect of SURF1 mutations using extracellular flux analysis (Seahorse XFe96) of oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) (Figure 2B-C). Using standard stoichiometric ratios for oxygen consumed, protons pumped, and linked ATP synthesis under standard conditions, OCR and ECAR can be transformed into interpretable ATP production rates using the methods described in 55. When added together, OxPhos-derived (JATP-OxPhos) and Glycolysis-derived ATP flux (JATP-Glyc) reflect the total energetic demand (JATP-Total) of each cell population (Extended Data Figure 2A). This approach is the cellular equivalent to REE measurements through indirect calorimetry in humans (Figure 1E) and mice 10,24,25.
Trajectories of JATP-OxPhos and JATP-Glyc across 150 days of lifespan are presented in Figure 2D. As expected from the SURF1 deficiency, SURF1-mutant cells (hereafter SURF1 cells) exhibited a 44% decreased JATP-OxPhos, but a 3-4-fold increased JATP-Glyc. Both parameters remained relatively stable across the lifespan (although a potential oscillatory behavior cannot be ruled out). Computing total energy expenditure showed that total ATP demand per unit of time was strikingly 91% higher in SURF1 cells relative to control cells (JATP-Total, p<0.001, g=2.4) (Figure 2E). These data demonstrate a robust SURF1-induced hypermetabolic state similar, albeit of greater magnitude, to that observed in patients with mitochondrial diseases.
To confirm this finding, potential confounds had to be ruled out. Non-glycolytic extracellular acidification rate (ECAR) could inflate estimates of JATP-Glyc56. However, measured resting non-glycolytic ECAR (in the absence of glucose or in the presence of the glycolysis inhibitor 2-deoxyglucose) was not elevated in SURF1 cells (it was in fact 31% lower), confirming the specificity of the ECAR signal in SURF1 cells to glycolysis (Extended Data Figure 3D). We also confirmed that non-OxPhos-related oxygen consumption by cytoplasmic and other oxidases did not differ between experimental groups (Extended Data Figure 2D). This parameter is not included in computing JATP-OxPhos, which formally excludes non-mitochondrial respiration as a potential contributor to the hypermetabolism measured in SURF1 cells.
Primary fibroblasts are continually dividing and a portion of total energy budget is expected to support cell division-related processes including DNA replication, transcription/translation, and other intracellular processes 20. Early in life (20-50 days), when division rates were mostly constant, SURF1 fibroblasts compared to control cells divided on average 31.8% slower (P<0.0001, g=-1.53 and 48.4% slower when quantified across 150 days). Therefore, hypermetabolism in SURF1 cells cannot be accounted for by an accelerated division rate. In fact, normalizing JATP-Total per rate of division further exaggerates apparent hypermetabolism, where SURF1 cells expend more than double the amount of energy than control to complete each cell cycle. Moreover, optically monitoring cell size at each passage showed that the SURF1 cell volume were moderately larger in early life and became smaller with increasing age, reaching similar volume as control cells by 150 days (Extended Data Figure 4A). Cell death was not significantly elevated (p=0.69, g=0.15, Extended Data Figure 4D-E). After accounting for cell volume, energy expenditure remained significantly elevated in SURF1 fibroblasts (p<0.0001, g=1.2, Figure 2F), demonstrating an increase in mass-specific REE. This increase is consistent in magnitude with that observed in humans (Figure 1) and animals 10,24–26 with OxPhos defects.
In control cells, the balance of estimated ATP derived from OxPhos and glycolysis was 64:36%, such that under our specific tissue culture conditions (5mM glucose, with glutamine, pyruvate and fatty acids), healthy fibroblasts derived the majority of ATP from OxPhos. In contrast, SURF1 deficiency robustly shifted the relative OxPhos:Glycolysis contribution to 23:77% (p=4.1e-6, g=-5.1), reflecting a significant shift in OxPhos-deficient cells towards an alternative, and less energy efficient, metabolic strategy (Figure 2G-H). As expected, removing glucose from the media did not substantially affect growth in control cells, but the absence of glucose was lethal to SURF1 cells within 5 days, confirming their dependency on glycolysis for survival (Extended Data Figure 3).
In response to this metabolic shift towards glycolysis, we expected SURF1 cells to naturally decrease maintenance-related energetic costs by decreasing mitochondrial mass and mtDNA copy number (mtDNAcn). However, in early life, SURF1 cells had the same mtDNAcn as control cells (5-40 days, p=0.99, g=0.04). And across the lifespan, SURF1 cells contained 32% more mtDNA copies, which manifested as an earlier age-related rise in mtDNAcn that reached maximal levels on average 30% higher than control cells (p=0.52, g=1.9, Figure 2I). Thus, although total mitochondrial mass was not directly assessed, elevated mtDNAcn similar to that observed in patient tissues (see Figure 1B) could contribute to increased maintenance cost and overall hypermetabolism in OxPhos-deficient cells, as suggested by mathematical modeling studies 23.
One potential mechanism for the lowered metabolic efficiency is a decrease in OxPhos coupling (i.e., uncoupling) at the inner mitochondrial membrane. However, both estimated proton leak (Figure 2J) and coupling efficiency (Figure 2K) measured by the proportion of OxPhos-dependent respiration not linked to ATP synthesis, were not different between control and SURF1 groups, and showed no measurable drift across the lifespan, thus ruling out mitochondrial uncoupling as a mechanism for hypermetabolism.
Inhibition of the mitochondrial ATP synthase triggers hypermetabolism
Next, to test if hypermetabolism is a specific manifestation in SURF1 cells or a more general feature of mitochondrial OxPhos dysfunction, we took an orthogonal pharmacological approach to chronically perturb OxPhos, and repeated the lifespan assessments of energy metabolism. Starting at day 20, fibroblasts from the same three healthy donors as above were treated chronically with a sublethal concentration of the mitochondrial ATP synthesis inhibitor oligomycin (Oligo, 1nM), which induces the ISR 3,57 (Figure 3A). Oligo reduced cellular oxygen consumption rate by ∼90% while largely maintaining viability, reflected in only a moderate elevation in cell death over time (2.7% in Oligo-treated cells vs 1.4% in control cells, 20-50 days, p=0.078, g=0.70) (Extended Data Figure 4D, 5B).
In relation to energy expenditure, Oligo doubled JATP-Total across the lifespan for each of the three healthy donors (+108%, p=5.9e-9, g=2.2), recapitulating the hypermetabolic state observed in SURF1 cells (Figure 3B). This robust elevation in cellular energy expenditure was already evident by 5 days of treatment and remained relatively stable across the lifespan, indicating the rapidity and stability of this adaptive hypermetabolic state. As in SURF1 cells, the hypermetabolic state in Oligo-treated cells was attributable to a markedly increased JATP-Glyc in excess of the decline in JATP-OxPhos, resulting in a shift outside of the optimal (i.e., normal) window of the OxPhos:Glycolysis ratio for these cells (Figure 3C-D).
Reductions in cell size and division rates are strategies to minimize energetic costs. Oligo caused a small but stable 4.8% decrease in cell size (p<0.001, g=-0.35), and decreased cell division rates by 39.1% (days 20-50, p=1.3e-5, g=-1.31; 49.6% slower across 150 days) (Extended Data Figure 4). Taking cell size into consideration showed that Oligo increased energy expenditure per unit of cell volume by 131% (p<0.001, g=0.97) (Figure 3E-F). Here also, hypermetabolism was not driven by significant increase in estimated proton leak (p=0.19, g=0.27) (Figure 3G) although we observed a 34.4% reduction in estimated coupling efficiency (p<0.05, g=-0.59) (Figure 3H), likely arising from the expected elevation in membrane potential from ATP synthase inhibition. Unlike SURF1 mutations, Oligo decreased mtDNAcn by 39.0% early in life (20-50 days, p=3.1e-5, g=-2.42), which subsequently normalized; peak levels were similar to control levels (Figure 3I-J).
Monitoring weekly the influence of Oligo on cell morphology also revealed an unexpected morphological phenotype. Oligo-treated cells developed into a reticular network, which involved contraction of the cell body and extension of multiple cellular appendages reminiscent of neuronal dendrites (Extended Data Figure 5). This reversible phenotype exhibited regular oscillatory behavior (one-week normal morphology, one-week reticular formation). We note that oscillatory behaviors are naturally energy-dependent 58, and that such dramatic and repeated changes in cell morphology must necessarily involve the remodeling of cell membranes and cytoskeleton through energy-dependent motor and cytoskeletal components. This morphological phenotype unique to the Oligo treatment could contribute to the higher energy expenditure in Oligo-treated cells (+131% ATP consumption per unit of cell volume) vs SURF1 (+91%).
OxPhos defects trigger the ISR and mtDNA instability
To understand the specific organelle-wide mitochondrial recalibrations in hypermetabolic SURF1 and Oligo-treated cells, we performed bulk RNA sequencing across the lifespan in each donor cell line (total 60 time points, average of ∼7 timepoints per cell line). We then systematically queried mitochondrial pathways from MitoCarta 3.0 59, in addition to all mtDNA-encoded transcripts (37 genes), and core ISR-related genes (ATF4, ATF5, CHOP/DDIT3, GDF15). Both SURF1 defects and Oligo treatment downregulated the majority of intrinsic mitochondrial pathways, including mtDNA stability and decay, which was downregulated in both SURF1 (−15%, p=1.7e-8, g=-1.65) and Oligo-treated cells (- 19%, p<0.001, g=-0.57) relative to control (Figure 4A). Although SURF1 and Oligo-treated cells exhibited similar overall mitochondrial transcriptional changes, some pathways showed opposite responses (e.g., expression of mtDNA-encoded genes, Pathway 3 in Figure 4B), suggesting the existence of partially specific mitochondrial recalibrations among SURF1 and Oligo models. In hierarchical clustering analysis across all pathways, the ISR pathway diverged most strongly from other pathways, and was upregulated +110% in SURF1 (p=6.5e-7, g=1.76) and +217% in Oligo-treated cells (p=1.2e-8, g=0.99), reaching up to a 16-fold elevation relative to the average of the young healthy donor cells (Figure 4B). Thus, both models of OxPhos dysfunction and hypermetabolism were associated with upregulation of the ISR, and downregulation of most mitochondrial pathways, notably mtDNA maintenance, suggesting a potential effect on mtDNA stability.
To our knowledge, neither SURF1 mutations nor Oligo treatment are established to cause mtDNA instability, but given the transcriptional changes described above and that heteroplasmy among mtDNA species is predicted to increase energetic maintenance costs 23, we directly examined mtDNA stability using two approaches. We first used long-range PCR at multiple time points across the lifespan of control, SURF1, and Oligo-treated cells, then validated the presence of mtDNA deletions across the lifespan by mtDNA sequencing, and quantified mtDNA deletion burden using eKLIPse 60 (Figure 4C-D). Circos plots in Figure 4D show the break points and heteroplasmy level for each mtDNA deletion, at early and late time points along the cellular lifespan. Circos plots for all timepoints investigated (4-14 timepoints per condition) are presented in Extended Data Figure 6.
Consistent with previous work, healthy fibroblasts do not accumulate appreciable heteroplasmy levels of mtDNA deletions in culture. However, SURF1 cells contained on average 17-fold more unique mtDNA deletions than control cells (p<0.01, g=1.38), reaching up to 126 unique deletions at a given time point (Figure 4E-F). The effect of Oligo treatment was more modest but reached levels 3-fold higher than untreated cells (p<0.01, g=0.79), and up to 20 unique deletions per time point. The majority of deletions eliminated segments of the minor arc and were on average 6.8-7.3kb in length; deletion size was similar among the three groups (Extended Data Figure 7A-C). Point mutations were not significantly elevated in SURF1 and Oligo-treated cells, suggesting specificity of mtDNA instability to deletions (Extended Data Figure 7D). Compared to controls where the maximal heteroplasmy levels was 0.13%, SURF1 and Oligo accumulated individual deletions reaching up to 0.40% and 0.19% heteroplasmy among the cell population (Extended Data Figure 7E-F), which remains low but similar to that observed with aging in human blood and brain tissues 61,62, and possibly noteworthy for replicating fibroblasts.
SURF1 mutations increases aging-related secretory activity
We next investigated outputs of the ISR, including the production of metabokines and cytokines. To broadly characterize changes in the cytokine stress response in patient-derived SURF1 cells across the lifespan, we designed a custom Luminex array targeting age-related proteins identified by plasma proteomics to be upregulated with human aging 63 (Figure 5). Compared to healthy donors, hypermetabolic SURF1 cells secreted higher levels of cytokines on a per-cell basis, including several pro-inflammatory cytokines, chemokines, and proteoglycans associated with the senescence-associated secretory phenotype (SASP) 64 (Figure 5A). Of the 27 cytokines detected in extracellular media, SURF1 cells achieved the highest cytokine concentration across the lifespan for 23 (85%) of the cytokines, reaching up to 10-fold higher concentration than control for one of the cytokines (insulin-like growth factor binding protein, IGFbp-rp1) (Figure 5B). Upregulated cytokines also included the canonical pro-inflammatory cytokines IL-6 and IL-8. The metabokine GDF15, which is elevated in both mitochondrial disease 65,66 and human aging 63,67, and which also appears sufficient to trigger hypermetabolism in mice 10, was also upregulated by 110% in SURF1 vs control cells (20-80 days, p=0.035, g=1.0, Figure 5C).
We attempted to validate IL-6 and GDF15 levels in both SURF1 and Oligo-treated cells by ELISAs. The ELISAs confirmed that IL-6 increased exponentially in aging fibroblasts, displaying altered onset and trajectories in both SURF1 (upregulated) and Oligo-treated cells (downregulated) (Figure 5D). Compared to control fibroblasts where GDF15 was undetectable in early passages, SURF1 mutant fibroblasts began to secrete GDF15 prematurely, and Oligo treatment acutely induced robust GDF15 secretion by 1-2 orders of magnitude over the first few weeks (Figure 5E), consistent with the rapid induction of the ISR particularly in Oligo-treated cells (see Figure 4B).
As cell-free mitochondrial DNA (cf-mtDNA) is associated with human aging 68 and was recently found to be elevated in the plasma of patients with mtDNA mutations/deletions 69, we quantified cf-mtDNA in the media along the lifespan. Both mtDNA and nDNA were detectable at appreciable levels (Extended Data Figure 8A-B). Compared to media of control cells, cf-mtDNA levels were 73% higher in SURF1 (g=0.5) and 100% higher (g=0.3) in the media of Oligo-treated cells (Figure 5F), although these differences did not reach statistical significance due to the high temporal variation of this phenotype. Parallel measurements of cell-free nuclear DNA (cf-nDNA) showed that the released mitochondrial-to-nuclear genome ratio was on average 117% higher in SURF1 than control cells (p<0.01, g=0.85, Extended Data Figure 8C-D), indicative of selective mtDNA release. Given the energetic cost associated with protein secretion 21,70, we suggest that the cytokine/metabokine and mtDNA hypersecretory phenotype in SURF1 and Oligo cells must contribute to hypermetabolism along with other cellular processes.
OxPhos defects upregulate energy-demanding cellular programs
From our longitudinal RNAseq dataset, we noted changes in the totality of genes related to the ribosomal machinery, which is produced in proportion with cellular biosynthetic demands (Extended Data Figure 9A). Despite their significantly reduced growth rate, both SURF1 (+19%, p=2.4e-10, g=2.59) and Oligo-treated cells (+50%, p<0.01, g=0.86) showed a marked time-dependent upregulation in the ribosomal machinery, consistent with the hypersecretory phenotype (secreted proteins must be transcribed and translated) and the elevated metabolic demands of translation that is in competition with cell growth 42,70 (Extended Data Figure 9B). To characterize the genome-wide gene regulatory changes associated with these cellular phenotypes, and to gain insights into the potential cause(s) of hypermetabolism in response to OxPhos dysfunction, we next deployed time-sensitive models of gene regulation and DNA methylation.
We first visualized the transcriptomic profiles of SURF1 and Oligo-treated cells using t-distributed stochastic neighbor embedding (t-SNE). Spatial embedding along the two major tSNE components captured three main features of the transcriptome: i) substantial interindividual differences separating each donor/cell line, ii) age-dependent shifts in transcriptional profiles, iii) clustering among both SURF1 and Oligo cells (Figure 6A). To harness the longitudinal nature of these data, we used a linear mixed effects model (LMER) to identify time-dependent differentially expressed genes (DEGs, FDR<0.05 threshold) between SURF1 and Oligo relative to control, across the cellular lifespan (Supplemental Files 1-2). Consistent with the similar degree of hypermetabolism and metabolic shift of both cellular models (see Figures 2D-F and 3B-F), there was a relatively high degree of overlap in DEGs between SURF1 and Oligo-treated cells (Supplemental Files 3-4). Genes with the largest effect sizes conserved across SURF1 and Oligo showed up to 2-4-fold upregulation (39%, n=1,503) or downregulation (35%, n=1,344) (Figure 6B-C). Differences were larger and more stable between SURF1 and control, compared to more progressive effects following the beginning of the Oligo treatment (Extended Data Figure 10), consistent with the constitutive genetic deficiency in SURF1 cells compared to the novel insult with Oligo treatment. The effects of OxPhos dysfunction on the expression of the 37 mtDNA genes across the cellular lifespan are shown in Extended Data Figure 11.
To gain insight into the molecular underpinnings of hypermetabolism, we analyzed gene expression changes using iPAGE, an information-theoretic computational framework that enables the systematic discovery of perturbed cellular pathways from gene expression data 71. Both SURF1 and Oligo-treated cells displayed a significant perturbation of transcription and translation processes (Figure 6D). Upregulated genes were enriched for pathways related to Golgi vesicle transport, fibroblast growth factor (FGF) binding, VEGF receptor signaling pathway, and the unfolded protein response, a signature consistent with increased secretion and inter-cellular signaling activity. Downregulated genes were over-represented for processes relating to cell division, consistent with the slower division rates (i.e., quiescence or senescence) of SURF1 and Oligo-treated cells.
Lifespan gene expression trajectories in this dataset showed some noteworthy features of OxPhos dysfunction and hypermetabolism at the single-gene level: i) ISR-related genes are robustly upregulated in a time-dependent manner by up to ∼16-fold for the transcription factor CHOP (DDIT3), and ∼60-fold for its downstream target GDF15; ii) the age-related upregulation of senescence-related genes (e.g., p21/CDKN1A) occurs prematurely in hypermetabolic SURF1 and Oligo cells; iii) key nucleotide metabolism enzymes such as thymidine kinase 1 (TK1) are robustly downregulated in SURF1 cells, possibly contributing to mtDNA instability 72; and iv) telomere and longevity-related genes such as the metabolic sensor SIRT1 and the telomere protection complex component CTC1 are upregulated 2-4-fold (Figure 6E). These broad changes in gene expression, largely consistent with previous in vitro work 3,73, prompted us to examine another major layer of gene regulation, DNA methylation.
DNA methylation recalibrations in OxPhos-induced hypermetabolism
To examine nuclear DNA methylation (DNAm) and create a resource dataset with broad utility, we measured DNA methylation levels at 865,817 CpG sites (Illumina EPIC array) in Control, SURF1, and Oligo-treated cells at multiple time points across their cellular lifespan (n=66). We then leveraged these high dimensional data by building mixed-effects models that consider the underlying data structure (donors, longitudinal observations) to identify robust conserved DNA methylation changes associated with OxPhos dysfunction and hypermetabolism. Visualizing the general data structure using t-SNE showed, that: i) the methylome signature of each donor was relatively distinct; ii) DNAm exhibited consistent age-related shifts, iii) SURF1 cells clustered separately from control, while iv) Oligo cells caused a modest time-dependent shift away from their respective controls (Figure 7A). These data therefore add to previous evidence in HEK293 cells74 and mice75, providing a robust platform for discovering conserved nuclear DNAm effects of hypermetabolism-causing OxPhos defects in primary human cells.
At the single CpG level, we asked which differentially methylated positions (DMPs) were stably and consistently either hypo- or hypermethylated in both SURF1 or Oligo-treated cells relative to control. Because transcriptionally relevant DNAm changes may operate across multiple CpGs, we complemented this approach by systematically examining differentially methylated regions (DMRs), which include multiple nearby CpGs exhibiting similar hypo- or hypermethylated changes in our statistical model 76 (see Methods for details). Figure 7B shows the overlap in significant DMPs and DMRs (threshold FDR<0.05). Of the overlapping DMPs between SURF1 and Oligo, 14.8% were hypermethylated, and 45.9% were hypomethylated. Global hypomethylation is a feature of human aging and replicative senescence 53. For DMRs, the corresponding proportions were 11.1% and 46.6%, showing high agreement in the methylome recalibrations between DMPs and DMRs approaches. A notable number of significant and highly differentially methylated changes in either SURF1 or Oligo-treated cells were specific to each condition (Supplemental Files 5-10), but here we focus exclusively on the changes conserved across two independent models, which therefore have the highest probability of being specifically caused by OxPhos dysfunction and related to hypermetabolism (Figure 7C-D).
The most robust changes in DNA methylation were targeted at CpG islands near or on gene bodies. Relative to control cells, as in the RNAseq results, the effect sizes were larger for SURF1 compared to Oligo, which induced directionally consistent but smaller effect size changes than SURF1 defects (Figure 7E). A stringent analysis of the most differentially methylated genes (based on both DMSs and DMRs) showed strong enrichment for processes involving: i) development and morphogenesis, ii) regulation of cell-cell signaling and organismal communication, iii) neural development, and iv) cell adhesion (Figure 7F). As highlighted above, increased regulation of signaling and communication, along with development and morphogenesis, must entail energetically dependent processes. These data, supported by the activation of corresponding downstream transcriptional programs (Figure 6) and the observed hypersecretory phenotype in OxPhos-deficient cells (Figure 5), document genome-wide epigenomic recalibrations consistent not with energy conservation, but with increased total energy expenditure.
RC dysfunction accelerates telomere shortening and decreases lifespan
Finally, given the deleterious effect of hypermetabolism-causing OxPhos defects on lifespan in patients with mitochondrial diseases and in animal models, these genome-wide data prompted us to examine how hypermetabolism relates to genomic markers of cellular aging and senescence. The complete population doubling curves of each donor (Figure 8A) provided initial evidence that cellular lifespan was reduced in SURF1 and Oligo-treated cells. The Hayflick limit (i.e., total number of cell divisions 52) was on average 53% lower in SURF1 cells (p=0.072, g=2.0), and Oligo decreased the Hayflick limit by 40% (p<0.066 g=2.0) relative to the untreated cells of the same donor (Figure 8A-B). Interestingly, the magnitude of these effects (40-53%) on total population doubling loosely corresponds to the 3-4-decade loss in human lifespan documented among adults with mitochondrial diseases (see Figure 1G-H), which would represent 38-50% for an average 80-year life expectancy.
To directly measure the pace of biological aging in response to OxPhos defects, we performed repeated-measures of telomere length across the cellular lifespan, which allowed us to compute the average rate of telomere shortening per population doubling or cell division (Figure 8C). Consistent with observations of dramatically shortened telomeres in skeletal muscle of patients with mtDNA mutations 77 and recent work causally linking mitochondrial dysfunction to telomere dysfunction 78, both SURF1 mutations and Oligo treatment strikingly increased the rate of telomere erosion per population doubling by 162% (SURF1: p=0.53, g=3.2) and 769% (Oligo: p=0.09, g=1.2) (Figure 8D). This means that for each cell division, OxPhos-deficient hypermetabolic fibroblasts lose 1.6-7.7 times more telomeric repeats than healthy fibroblasts. We note that these results rely on the estimated slope across the whole cellular lifespan (single value per donor, n=3 per group) so the p values are less meaningful than the effect sizes, which are large (>1). The terminal telomere length coinciding with growth arrest tended to be moderately higher in SURF1 and Oligo groups (Figure 8E). This could suggest that growth arrest is driven by factors other than absolute telomere length, such as the prioritization of transcription/translation over growth-related functions, sufficient to induce growth arrest and senescence in human fibroblasts 42,43.
Next, we leveraged our DNAm dataset to quantify biological age using validated multivariate algorithms or “clocks” (DNAmAge, or epigenetic clocks) trained, in human tissues, to predict chronological age and mortality 53,79. Five different validated clocks that rely on different CpG sets and include a modification that improves their accuracy 80 were applied directly to our fibroblast time series DNAm data. These results showed that relative to the rate of epigenetic aging in control cells with normal OxPhos function, the rates of biological aging per population doubling were accelerated by an average of 131% in SURF1 cells (p<0.05, g=1.5), and to a lesser extent in Oligo-treated cells (+54%, p<0.05, g=0.6, Figure 8F-G), independently supporting the findings of accelerated telomere shortening. Trajectories and DNAm aging rates for each donor using all five epigenetic clocks, including those computed relative to “time in culture” rather than to population doublings, produced variable results and are presented in Extended Data Figure 12.
Together, the decreased Hayflick limit, the accelerated telomere attrition rate, and increased rate of epigenetic aging converge with the secretome analysis and senescence-related gene expression to link mitochondrial hypermetabolism-causing OxPhos dysfunction to reduced cellular lifespan.
Discussion
Integrating available clinical and animal data together with our longitudinal fibroblast studies has revealed hypermetabolism as a conserved feature of mitochondrial OxPhos dysfunction. A major advantage of our cellular system is that it isolates the stable influence of genetic and pharmacological OxPhos perturbations on energy expenditure, independent of other factors that may operate in vivo. Thus, these data establish that OxPhos dysfunction induces hypermetabolism in a cell-autonomous manner. Moreover, despite the diverging mode of action of SURF1 and Oligo models, as well as some divergent molecular features, both models converge on the same hypermetabolic phenotype, adding confidence around the generalizability of this phenomenon. Our data also rule out mitochondrial uncoupling as a main driver of hypermetabolism in this system, and instead implicates the activation of energy-demanding gene regulatory programs, including but likely not limited to increased metabokine/cytokine secretion, that can compete with growth and longevity (Figure 9). Although phenomenological in nature, our cellular lifespan data provide several novel observations that agree with previous work 73, and that are relevant to understanding how primary mitochondrial OxPhos dysfunction triggers some of the physiological and phenotypic hallmarks of mitochondrial diseases.
First, we observed that the mitochondrial disease marker GDF15 was largely undetectable in the media of young, healthy fibroblasts, but increased progressively across the cellular lifespan. This finding recapitulates the age-related increase in GDF15 in humans 63,67, and adds to previous evidence of conserved age-related changes in DNA methylation in primary human fibroblasts cultured over several months 53. Consistent with the higher GDF15 levels in primary OxPhos disorders in humans 66 and mice 10, extracellular GDF15 tended to be elevated in both models of OxPhos dysfunction. Likewise, OxPhos dysfunction increased extracellular cf-mtDNA levels, in line with recent reports that cf-mtDNA is elevated in primary OxPhos disorders 69 and with aging in humans 68. The link between OxPhos dysfunction and cf-mtDNA release requires further investigation.
Second, we observed that OxPhos dysfunction from SURF1 mutations, and to a lesser extent Oligo treatment, both caused secondary mtDNA instability. mtDNA instability was associated with the variable accumulation of mtDNA deletions, but not point mutations, across the cellular lifespan. Our confidence in this result is reinforced by the longitudinal nature of the mtDNA sequencing data, from the same primary cell lines examined at multiple time points. Notably, the time course data also showed that cell populations can eliminate a large fraction of mtDNA deletions within 12-14 days (different mtDNA deletions appear at different passages). This is consistent with the fact that replicating fibroblasts eliminate some deleterious mtDNA deletions 81, and also that several de novo deletions removed the origin of replication of the light strand (OL), thereby preventing their replication. Whether the clonal amplification of some mtDNA deletions in SURF1 fibroblasts occurs through population selection at the cellular level, or through intracellular quality control mechanisms, or a combination of both, remains to be determined.
Third, mitochondrial OxPhos dysfunction dramatically increased the telomere erosion rate per cell division, despite the adaptive transcriptional upregulation of telomere protection complex components. This finding agrees with a role of mtDNA on telomere maintenance in conplastic mice 82, with the life-shortening effect of pathogenic mtDNA variants 32 and OxPhos dysfunction in mice 34, and with the reduced lifespan in patients with mtDNA disease shown in Figure 1G-H. A study in skeletal muscle of children with high heteroplasmic mtDNA mutations also reported excessively short telomeres, similar in length to the telomeres of healthy 80 year old controls 77. Because skeletal muscle is a post-mitotic tissue, this previous result also implies that OxPhos dysfunction could accelerate telomere attrition at a disproportionate rate, or perhaps independent from cell division, as suggested by the disconnect between the loss of telomeric repeats and genome replication/cell division observed in our hypermetabolic fibroblasts.
Why OxPhos-induced hypermetabolism is associated with both mtDNA instability and accelerated telomere erosion remains unclear. One potential principle at play may relate to the fact that DNA maintenance (mtDNA, and telomeres) relies on the accuracy of the molecular processes that ensure accurate mitochondrial and nuclear DNA replication. The energetic tradeoff between translation and growth 42 could explain why OxPhos-deficient cells, which expend a large fraction of their energy budget to upregulate transcription/translation and secretory, also grow more slowly. DNA replication is also constrained and sits at the bottom of a hierarchy of energy-consuming processes where vital processes, such as ionic balance and translation, are prioritized over division and DNA replication 20. Furthermore, cells under stress experience an energetic tradeoff between the accuracy of molecular operations and the speed of these processes, known as the energy-speed-accuracy tradeoff 83. Although it is currently not possible to mechanistically test this point, we speculate that the diversion of energetic resources, as well as substrates including nucleotides 72, contributes to reduced DNA replication fidelity, which in turn could contribute to both mtDNA instability and telomere attrition.
Fourth, our longitudinal RNASeq and DNAm datasets revealed conserved recalibrations implicating developmental and translation-related pathways, as well as cell-cell communication, with OxPhos dysfunction and hypermetabolism. The identified pathways overlap with previously identified multi-omic overrepresentation analysis performed on iPSC-derived neurons from SURF1 patients 84. In both this and our study, multi-omic analyses identified neural development, cell signaling, morphogenesis, cell cycle, and metabolism as the predominant processes altered in SURF1-related disease. The induction of these energetically-demanding pathways that constrain growth at the cellular and possibly at the organismal level 41, could help explain why a major feature of pediatric mitochondrial disorders (including our SURF1 donors) is developmental delay, and why short stature is common in adult patients 30. In relation to cell-cell communication, we note that the biomarker picture of adult patients with mitochondrial encephalopathy, lactic acidosis, and stroke-like episodes (MELAS) is dominated by elevated (not reduced) signaling and metabolic markers in blood 66. Thus, the organism under metabolic stress does not turn to an energy-saving hypometabolic state, but instead activates energivorous integrated stress responses that must divert and consume energetic resources, in a tradeoff with other processes such as growth and longevity pathways.
Finally, OxPhos dysfunction triggered a shift towards glycolytic ATP production. The glycolytic shift in fibroblasts is consistent with the physiological shift in substrate oxidation from lipids/amino acids to carbohydrates (whole body respiratory quotient) among patients 85 and mice 86 with OxPhos defects. The active shift towards glycolysis occurs even when OxPhos is not completely obliterated. For example, although basal respiration was markedly lower in SURF1 cells, the maximal FCCP-uncoupled respiration in SURF1 cells was relatively preserved (see Figure 2B & Extended Data Figure 2C). This cellular decision to route metabolic flux towards an energetically less efficient pathway (i.e., glycolysis) can be explained on the basis of energetic constraints and proteome efficiency, since the proteome cost of OxPhos is at least double that of glycolytic fermentation 19. Thus, cells can “choose” to divert metabolic flux towards glycolysis even when OxPhos is at least partially functional, as in cancer, because of rising intracellular energetic constraints driven by hypermetabolism. We note again that hypermetabolism is apparent across multiple animal models of primary OxPhos dysfunction, manifesting as an elevated cost of living, even during rest and sleep in mice 10,24–26. In particular, deep phenotyping of Ant1-/- mice across three studies 25,87,88 reveals a systemic physiological picture highly consistent with mitochondrial diseases, including excessive mitochondrial biogenesis, elevated circulating catecholamine levels, severe hypermetabolism (+82-85% REE), reduced adiposity, elevated mtDNAcn and mtDNA instability, and decreased median lifespan. These in vivo data provide direct converging evidence, beyond the clinical data (Figure 1), that mitochondrial OxPhos dysfunction impairs whole-body energetic efficiency and cause physiological hypermetabolism.
Identifying hypermetabolism as a feature of mitochondrial disease has potential clinical utility as it provides a potential basis to explain some of the major symptoms in patients. First, fatigue and exercise intolerance are evolutionary conserved and subjective experiences that arise when the organism consumes more energy than it would under optimal conditions (e.g., subjective fatigue during the oxygen debt after strenuous exercise, or during an infection). Thus, symptoms of fatigue could be direct consequences of impaired metabolic efficiency and hypermetabolism. Second, as noted above, severely affected patients with mitochondrial disease are usually thin, which may be attributable to not only reduced energy intake or to intestinal malabsorption, but to chronic hypermetabolism, effectively burning excess ingested calories, preventing the accumulation of excess adiposity. Third, alcohol appears to be poorly tolerated and associated with symptom onset in some patients with mtDNA defects 89–91, but the basis for alcohol intolerance remains unknown. Alcohol itself causes hypermetabolism in healthy individuals – increasing whole-body REE by as much as 16%, and inhibiting lipid oxidation by 31-36% 92,93. Alcohol may therefore aggravate pre-existing hypermetabolism, thus imposing further energetic constraints on vital cellular or physiological functions. Finally, chronic hypermetabolism could in part explain why infections can trigger symptoms exacerbations, representing the major cause of decompensation and death in this population 29. The metabolic cost of immune activation to viral and bacterial infection is high, and cytokine production in human leukocytes is under mitochondrial regulation 94. Thus, immunity must therefore compete with other host maintenance systems 95. We speculate that in mitochondrial diseases, because the more limited energetic resources are consumed at a higher rate than normal due to hypermetabolism, patients may particularly lack the necessary energetic reserve required to sustain vital organs while mounting adequate immune responses.
Together, the meta-analysis of clinical data from 17 cohorts and two cellular models of OxPhos dysfunction identify hypermetabolism as a feature of mitochondrial disease. Our longitudinal patient-derived fibroblasts data further delineate some of the cellular and molecular features of OxPhos-induced hypermetabolism, including sustained induction of the ISR, genome instability, hypersecretion of cyto/metabokines, and genome-wide DNA methylation and transcriptional recalibrations that emphasize the upregulation of processes related to signaling and communication (see Figure 9). A resource webtool with all data from this study, including the RNAseq and DNAm data, is available and can be explored for genes or processes of interest (see Data Availability Statement). Although further work is needed, these descriptive translational data provide a basis to rationalize some unexplained clinical features of mitochondrial diseases, providing a foundation to design well-controlled studies to understand to what extent hypermetabolism is a harbinger of morbidity and early mortality in patients with mitochondrial diseases.
Limitations
The major limitation of this work is the lack of mechanistic data about the origin of hypermetabolism, or about the relative contributions of specific signaling pathway(s) and cellular process(es) that drive elevated energy consumption in OxPhos deficient cells and humans. The extracellular flux analysis used to derive ATP consumption rates are also indirect 56, and metabolic tracing experiments are likely to yield useful information to better understand energy consumption and partitioning in hypermetabolic cells. Rather than pursuing a single potential mechanistic explanation for this phenomenon, motivated by the consistent evidence in clinical and animal studies, we attempted to deeply phenotype both cellular models of hypermetabolism and to produce an abnormally broad dataset (also an unusually long manuscript) covering key processes and pathways previously implicated in the pathogenesis of OxPhos defects. Deployed longitudinally across multiple donors, this publicly available dataset can now be used to rationally design testable hypotheses about the origin(s) and consequence(s) of hypermetabolism in OxPhos defects, to be tested in cellular and animal models, and eventually in humans. The small sample size and the in vitro nature of the cellular data can also be regarded as limitations of this work, potentially limiting generalizability. However, the stability of metabolic and molecular phenotypes in two distinct experimental models, across three unrelated donors (female and male) repeatedly monitored across the replicative lifespan – when cells undergo dynamic age-related changes – is a strong test of robustness for this finding. Finally, we note that the clinical phenotyping presented in Figure 1 is not exhaustive, focusing exclusively on clinical outcomes related to energy expenditure. Much work remains to fully define the clinical heterogeneity, as well as the neuroendocrine and metabolic manifestations of hypermetabolism. Therefore, our translational findings highlight more than ever the need for collaborative, multi-center partnerships that bridge the cellular, clinical, as well as the patient-reported aspects of mitochondrial diseases.
Methods
Human cohorts
Data were meta-analyzed from 17 mitochondrial disease cohorts listed in Table 1. Inclusion criteria included 1) cohorts with a genetic diagnosis for all participants and 2) including measures for at least one of the primary outcomes (resting heart rate, catecholamine levels, resting VO2 or VO2 relative to work rate, BMI, mortality). Eligible cohorts included participants from 5 countries including China, Denmark, England, Italy, and the USA. Studies were published between 2003 and 2019, covering a 16-year period. Each cohort with its sample size, female/male distribution, genetic diagnoses (nDNA vs mtDNA), and symptomatology is listed in Table 1, with additional information about data extraction provided here. Each cohort included their own control group, so group level averages (not patient-level data) were use to compute effect sizes as % difference between mitochondrial diseases and control, and standardized Hedges g for each outcome measure (e.g., resting heart rate, resting VO2). Cohorts with available source data to calculate intragroup variance include error bars denoting standard error of the mean in Figure 1.
Cohort 1 96 included data on resting HR and resting VO2 in patients with mixed genetic defects. Cohort 2 included four sub-studies: a) 97, b) 98, c) 99, and d) an unpublished cohort of patients with single large-scale mtDNA deletions with measures of resting HR, resting VO2, and BMI. Cohort 3 100 included data on resting HR, resting urinary catecholamines, and BMI. Cohort 4 101 included data on resting HR and BMI. Cohort 5 is an unpublished cohort (the Mitochondrial Stress, Brain Imaging, and Epigenetics Study – MiSBIE) of patients with m.3243A>G mutations, which included data on resting HR and BMI. Cohort 6 102 included data on circulating catecholamines at rest and during exercise. Cohort 7 85 included data on VO2 during fixed workload (65W) and BMI. Cohort 8 103 included data on resting VO2 and BMI. Cohort 9 104 included VO2 during constant work rate (40% of max), and VO2 values in ml/kg/min were adjusted to average workload achieved by each group to obtain comparable estimates of energetic demand relative to work performed. Cohort 10 14 included metabolic efficiency during constant-rate cycle ergometry (30 watts), including before and after a home-based exercise training protocol, and these values were compared to reference values in healthy individuals from 104. Cohort 11 105 overlaps with Cohorts 9 and 10 and included BMI data. Cohort 12 17 included BMI data averaged between both mutation groups. Cohort 13 30 is a natural history study of adult patients with mortality data. Cohort 14 29 is a retrospective study of the causes of death in adult patients with mortality data. Cohort 15 106 is a pediatric natural history study with mortality data. Cohort 16 28 is a multi-center pediatric natural history study with mortality data. Cohort 17 is an ongoing natural history study (McFarland et al., Newcastle Mitochondrial Disease Cohort) with mortality data.
Reference BMI for the USA (29.9 kg/m2) was obtained from the National Health and Nutrition Examination Survey (NHANES) for wave 2015-2016 (n=9,544) (e-link), for the UK (28.6 kg/m2) from the Health Survey for England 2018 (n=6,600) (link), and for Italy (25.8 kg/m2) from the NCD Risk factor collaboration (link), with the combined average presented in Figure 1F. Reference values for life expectancy were obtained from the World Bank (https://data.worldbank.org/) and the average value for the USA (78.6 yr), UK (81.2 yr), and Italy (82.9 yr) (representing most cohorts included) is reported in Figure 1G. Data presented in Figure 1H represent mortality rates in the UK (reference population) for 2018, and the mortality data for individuals for mitochondrial disease was collected between 2010-2020.
The clinical data demonstrating hypermetabolism are derived from more than a dozen laboratories over a >15-year period, illustrating the stability of this finding. The apparent cross-study stability of clinical hypermetabolism is also unlikely to be influenced by publication or reporting bias for three main reasons: i) most studies were exploratory (as opposed to confirmatory) in nature, such that the motivation for their publication depended neither on the significance nor direction of these results, ii) baseline group differences for most parameters (e.g., resting VO2) was not a primary outcome in any studies, and in several cases these data were not analyzed nor reported in the original reports, and iii) variables such as BMI are ubiquitously reported. Moreover, to further reduce the potential of bias, the overall sample includes new, previously unpublished cohorts of clinically and genetically well-defined patient populations (see Table 1). Together, these factors increase likelihood that the findings revealing the existence of a hypermetabolic state are robust and generalizable to mitochondrial diseases represented here, which includes a relatively broad diversity of mtDNA mutations. Further work directly quantifying hypermetabolism across the diurnal cycle, for example, are needed. Studies linking hypermetabolism to disease severity and progression are also warranted.
Skeletal muscle histology, mtDNA heteroplasmy, and mtDNA density
Human skeletal muscle from the diaphragm was subjected to sequential cytochrome c oxidase (COX, diaminobenzidine, brown) and succinate dehydrogenase (SDH, nitrobluetetrazolium, blue) staining as described previously 107. This technique reveals segments of myofibers deficient for mtDNA-encoded COX but positive for exclusively nDNA-encoded SDH 22. Sub-cellular segments of the same myofiber highlighted in Figure 1B were dissected from a 20um-thick cryosection by laser-capture microdissection (LCM) on a Leica AS LMD 6000 microscope, transferred and digested (Tween20, Proteinase K) overnight, and used as template DNA in a multiplex real-time PCR reaction that amplifies MT-ND4 and MT-ND1 amplicons within the minor and major arcs of the mtDNA, respectively, to calculate heteroplasmy levels for major arc mtDNA deletions 108. Total mtDNA density was quantified by deriving MT-ND1 copies from a standard curve, normalized per surface area (um2) of tissue used as input 107.
Tissue culture
Primary human dermal fibroblasts were obtained from distributor or in local clinic from 3 healthy and 3 SURF1-patient donors (IRB #AAAB0483, see below table for descriptive information and distributor). Fibroblasts were isolated from skin tissue biopsies using standard procedures. After isolation, fibroblasts were stored in 10% DMSO (Sigma-Aldrich #D4540), 90% fetal bovine serum (FBS, Life Technologies #10437036) in cryogenic tube under liquid nitrogen. To avoid freeze-shock necrosis cells were frozen gradually in an isopropanol container (Thermofisher #5100-0001) at -80°C overnight before storage in liquid nitrogen.
Genotypes were confirmed by whole genome sequencing. Paired-end (PE) reads were obtained from Illumina HiSeq and processed using SAMtools (v1.2) and BaseSpace workflow (v7.0). PE reads were aligned to hg19 genome reference (UCSC) using Isaac aligner (v04.17.06.15) and BAM files were generated. Small variants including single nucleotide variants (SNVs) and insertion/deletion (Indels) were called from the entire genome using Strelka germline variant caller (v2.8). Variants specific to SURF1 gene were obtained from the genome-wide annotated vcf files using SnpSift and annotated using web ANNOVAR.
To initiate cultures, cryopreserved fibroblasts were thawed at 37°C (<4min) and immediately transferred to 20ml of preheated DMEM (Invitrogen #10567022). Cells were cultured in T175 flasks (Eppendorf #0030712129) at standard 5% CO2 and atmospheric (∼21%) O2 at 37°C in DMEM (5.5 mM glucose) supplemented with 10% FBS, 50 μg/ml uridine (Sigma-Aldrich #U6381), 1% MEM non-essential amino acids (Life Technologies #11140050), 10 μM palmitate (Sigma-Aldrich #P9767) conjugated to 1.7 μM BSA (Sigma-Aldrich #A8806), and 0.001% DMSO (treatment-matched, Sigma-Aldrich #D4540). Cells were passaged approximately every 5 days (+/- 1 day). Oligo-treated healthy control cells were cultured in the same media as control cells supplemented with 1nM oligomycin (in 0.001% DMSO, Sigma-Aldrich #75351) starting on Day 15.
Brightfield microscopy images (10x, 20x magnification) were taken before each passaged using inverted phase-contrast microscope (Fisher Scientific #11350119). Cell counts, volume and death were determined at each passage using the Countess II Automated Cell Counter (ThermoFisher Scientific #A27977). Growth rates were used to determine replating density, by pre-calculating number of cells needed to reach ∼90% confluency (∼2.5 million cells) at time of next passage. Cells were never plated below 200,000 cells or above 2.5 million cells to avoid plating artifacts of isolation or contact inhibition, respectively. The timing and frequency of time points collected vary by assay, with an average sampling frequency of 15 days109. Cell media was collected at each passage. Individual cell lines were terminated after exhibiting less than one population doubling over a 30-day period. The Hayflick limit was determined as the total number of population doublings of a cell line at the point of termination.
Because oxygen tension can have a marked effect on the metabolism and replicative lifespan of cultured fibroblasts 110, and chronic hypoxia improves survival in fibroblasts with complex I defects and the Ndufs4 mouse model of Leigh syndrome 111, longitudinal experiments in SURF1 cells were repeated at low (3%) O2, in parallel with the same donor cell lines at atmospheric O2 (∼21%). Oxygen levels did not prevent or alter hypermetabolism, nor did they improve cell population growth rate (data not shown). Results of the low O2 “hypoxia” experiments as well as the full lifespan aging trajectory of control cells beyond 150 days are available in the resource dataset (see Data Availability Statement).
Mycoplasma testing
Mycoplasma testing was performed according to the manufacturer’s instructions (R&D Systems #CUL001B) at the end of lifespan for each treatment and cell line used. All tests were negative.
Calculations of energy expenditure and normalization to division rate and cell size
Bioenergetic parameters were measured using the XFe96 Seahorse extracellular flux analyzer (Agilent), oxygen consumption rate (OCR) and extracellular acidification rate (pH change) was measured over confluent cell monolayers. Cells were plated for Seahorse measurement every 3 passages (∼15 days) with 10-12 wells plated per a treatment group. Each well of a seahorse 96-well plate was plated with 20,000 cells and incubated overnight under standard growth conditions, following the manufacturer’s instructions, including a plate wash with complete Seahorse XF Assay media. The complete XF media contains no pH buffers and was supplemented with 5.5 mM glucose, 1 mM pyruvate, 1 mM glutamine, 50 μg/ml uridine, 10 μM palmitate conjugated to 1.7 μM BSA. After washing, the plate was incubated in a non-CO2 incubator for one hour to equilibrate temperature and atmospheric gases. The instrument was programmed to assess various respiratory states using the manufacturer’s MitoStress Test (Brand & Nicholls, 2011). Basal respiration, ATP turnover, proton leak, coupling efficiency, maximum respiration rate, respiratory control ratio, spare respiratory capacity and non-mitochondrial respiration were all determined by the sequential additions of the ATP synthase inhibitor oligomycin (final concentration: 1 μM), the protonophore uncoupler FCCP (4 μM), and the electron transport chain Complex I and III inhibitors, rotenone and antimycin A (1 μM). The optimal number of cells and concentration for the uncoupler FCCP yielding maximal uncoupled respiration was determined based on a titration performed on healthy fibroblasts (data not shown).
The final Seahorse injection included Hoechst nuclear fluorescent stain (ThermoFisher Scientific #62249) to allow for automatic cell counting. After each run, cell nuclei were counted automatically using the Cytation1 Cell Imager (BioTek) and raw bioenergetic measurements were normalized to relative cell counts on a per-well basis. ATP metrics were determined using the P/O ratios of OxPhos and glycolysis as previously described by Mookerjee et al. 55. These conversions assumed energy sourced was derived entirely by glucose. All JATP measurements take into account non-mitochondrial and proton leak derived oxygen consumption thereby reflecting the mitochondrial ATP-synthesis related flux (Extended Data Figure 2A). The code and raw data are available as detailed in the Data Availability statement.
To assess if increased ECAR in experimental conditions were due to non-glycolytic activity, a glucose-dependency test was performed using the Seahorse XF Glycolysis Stress Test Kit (Agilent, 103020-100). Prior to extracellular flux measurements, young healthy control (Donor2) and young SURF1 (Patient3) cells were grown overnight in differing nutrient conditions: physiological 5.5mM glucose, 0mM glucose, 25mM glucose. The glycolysis stress test kit was performed according to the manufacturer’s protocol. To monitor growth and cell death, cells were cultured for 7 days in each glucose condition and monitored daily (see Extended Data Figure 3).
mtDNA deletions
mtDNA deletions were initially detected by long-range PCR (LR-PCR) from DNA extracted from cultured fibroblasts using DNeasy blood and tissue kit (Qiagen #69504) following manufacturer’s instructions. Isolated DNA was amplified using 12 F (np 5855-5875) and D2 R (np 129-110) oligonucleotide primers to yield a 10-Kb product. PCR reactions were carried out using Hot Start TaKaRa LA Taq kit (Takara Biotechnology, #RR042A) with the following cycling conditions: 1 cycle of 94°C for 1min; 45 cycles of 94°C for 30sec, 58°C for 30sec, and 68°C for 11min with a final extension of 72°C for 12 min. Amplified PCR products were separated on 1% agarose gels in 1X TBE buffer, stained with GelGreen (Biotium #41005), imaged using a GelDoc Go Imager (Biorad). Primers (5’-3’) were: Forward (12F): AGATTTACAGTCCAATGCTTC (nucleotide position 5,855-5,875); Reverse (D2R): AGATACTGCGACATAGGGTG (129-110).
mtDNA next-generation sequencing and eKLIPse analysis
The entire mtDNA was amplified in two overlapping fragments using a combination of mtDNA primers. The primer pairs used for PCR amplicons were tested first on Rho zero cells devoid of mtDNA to remove nuclear-encoded mitochondrial pseudogene (NUMTS) amplification (PCR1: 5’-AACCAAACCCCAAAGACACC-3’ and 5’-GCCAATAATGACGTGAAGTCC-3’; PCR2: 5’-TCCCACTCCTAAACACATCC-3’ and 5’-TTTATGGGGTGATGTGAGCC-3’). Long-range PCR was performed with the Kapa Long Range DNA polymerase according to the manufacturer’s recommendations (Kapa Biosystems, Boston, MA, mtDNA next-generation sequencing and USA), with 0.5µM of each primer and 20ng of DNA. The PCR products were analyzed on a 1% agarose gel electrophoresis.
NGS Libraries were generated using an enzymatic DNA fragmentation approach using Ion Xpress Plus Fragment Library Kit. Libraries were diluted at 100 pM before sequencing and pooled by a maximum of 25 samples. Sequencing was performed using an Ion Torrent S5XL platform using Ion 540 chipTM. Signal processing and base calling were done by the pre-processing embedded pipeline. Demultiplexed reads were mapped according to the mtDNA reference sequence (NC_012920.1) before being analysed with a dedicated homemade pipeline including eKLIPse (Goudenège et al, GIM, 2019; https://github.com/dooguypapua/eKLIPse ) using the following settings:
Read threshold: min Quality=20 | min length =100bp
Soft-Clipping threshold: Read threshold: Min soft-clipped length =25pb | Min mapped Part=20 bp
BLAST thresholds: min=1 | id=80 | cov=70 | gapopen=0 | gapext=2
Downsampling: No
mtDNA copy number
Cellular mtDNA content was quantified by qPCR on the same genomic material used for other DNA-based measurements. Duplex qPCR reactions with Taqman chemistry were used to simultaneously quantify mitochondrial (mtDNA, ND1) and nuclear (nDNA, B2M) amplicons, as described previously 4. The reaction mixture included TaqMan Universal Master mix fast (life technologies #4444964), 300nM of custom design primers and 100nM probes: ND1-Fwd: GAGCGATGGTGAGAGCTAAGGT, ND1-Rev:CCCTAAAACCCGCCACATCT, ND1-Probe: HEX-CCATCACCCTCTACATCACCGCCC-3IABkFQ. B2M-Fwd: CCAGCAGAGAATGGAAAGTCAA, B2M-Rev: TCTCTCTCCATTCTTCAGTAAGTCAACT, B2M-Probe: FAM-ATGTGTCTGGGTTTCATCCATCCGACA-3IABkFQ). The samples were cycled in a QuantStudio 7 flex qPCR instrument (Applied Biosystems) at 50°C for 2 min, 95°C for 20 sec, 95°C for 1min, 60°C for 20 sec, for 40 cycles. qPCR reactions were setup in triplicates in 384 well qPCR plates using a liquid handling station (epMotion5073, Eppendorf), in volumes of 20ul (12ul mastermix, 8ul template). Triplicate values for each sample were averaged for mtDNA and nDNA. Ct values >33 were discarded. For triplicates with a C.V. > 0.02, the triplicates were individually examined and outlier values removed where appropriate (e.g., >2 standard deviations above the mean), with the remaining duplicates were used. The final cutoff for acceptable values was set at a C.V. = 0.1 (10%); samples with a C.V. > 0.1 were discarded. A standard curve along with positive and negative controls were included on each of the seven plates to assess plate-to-plate variability and ensure that values fell within instrument range. The final mtDNAcn was derived using the ΔCt method, calculated by subtracting the average mtDNA Ct from the average nDNA Ct. mtDNAcn was calculated as 2ΔCt x 2 (to account for the diploid nature of the reference nuclear genome), yielding the estimated number of mtDNA copies per cell.
Cytokines
Two multiplex fluorescence-based arrays were custom-designed with selected cytokines and chemokines most highly correlated with age in human plasma from 63, listed as available analytes on the R&D custom Luminex arrays (R&D, Luminex Human Discovery Assay (33-Plex) LXSAHM-33 and LXSAHM-15, http://biotechne.com/l/rl/YyZYM7n3). Media samples were collected at selected passages across cellular lifespan and frozen at -80°C. After thawing, samples were centrifuged at 500xg for 5min and supernatant moved to a new tube. Wells were loaded with media samples diluted 1:5 with assay diluent, incubated, washed, and read on a Luminex 200 (Luminex, USA) as per the manufacturer’s instructions. Positive (aged healthy fibroblast) and negative controls (fresh untreated media) samples were used in duplicates on each plate to quantify batch variations. Data were fitted and final values interpolated from a standard curve in xPONENT v 4.2, normalized to the cell number at the time of collection to produce estimates of cytokine production on a per-cell basis. IL-6 and GDF15 measures were repeated using enzyme-linked immunosorbent assays (ELISA), according to the manufacturer’s instructions (Abcam #ab229434 and R&D #DGD150).
Media cell-free DNA
Total cell-free DNA (cf-DNA) was isolated from cell culture media using a previously published automated, high throughput methodology (Ware et al. 2020). Quantitative polymerase chain reaction (qPCR): cf-mtDNA and cf-nDNA levels were measured simultaneously by qPCR. Taqman-based duplex qPCR reactions targeted mitochondrial-encoded ND1 and nuclear-encoded B2M sequences as described previously 112,113. Each gene assay contained two primers and a fluorescent probe and were assembled as a 20X working solution according to the manufacturer’s recommendations (Integrated DNA Technologies). The assay sequences are: ND1 forward 5’-GAGCGATGGTGAGAGCTAAGGT-3’, ND1 reverse 5’-CCCTAAAACCCGCCACATCT-3’, ND1 probe 5’-/5HEX/CCATCACCC/ZEN/TCTACATCACCGCCC/2IABkGQ/-3’, B2M forward 5’-TCTCTCTCCATTCTTCAGTAAGTCAACT-3’, B2M reverse 5’-CCAGCAGAGAATGGAAAGTCAA-3’, and B2M probe 5’-/56-FAM/ATGTGTCTG/ZEN/GGTTTCATCCATCCGACCA/3IABkFQ/-3’. Each reaction contained 4 µL of 2X Luna Universal qPCR Master Mix (New England Biolabs, cat#M3003E), 0.4 µL of each 20X primer assay, and 3.2 µL of template cf-DNA for a final volume of 8 µL. The qPCR reactions were performed in triplicates using a QuantStudio 5 Real-time PCR System (Thermo Fisher, cat#A34322) using the following thermocycling conditions: 95°C for 20 s followed by 40 cycles of 95°C for 1 s, 63°C for 20 s, and 60°C for 20 s. Serial dilutions of pooled human placenta DNA were used as a standard curve.
Digital PCR (dPCR): mtDNA and nDNA copy number (copies/µL) of the standard curve used in cf-mtDNA/cf-nDNA assessment were measured separately using singleplex ND1 and B2M assays using a QuantStudio 3D Digital PCR System and associated reagents (Thermo Fisher, cat#A29154) according to the manufacturer’s protocol. The values obtained for the standard curve were used to calculate the copy number for the experimental samples. All reactions were performed in duplicate (two chips). Because the same standard curve was used on all plates, its copy number was applied uniformly to all qPCR plates.
RNA sequencing and transcriptomic analyses
Total genomic RNA was isolated every ∼11days across cellular lifespan and stored in 1ml TRIzol (Invitrogen #15596026). RNA was extracted on-column using the RNeasy kit (Qiagen #74104), DNase treated according to the manufacturer’s instructions, and quantified using the QUBIT high sensitivity kit (Thermo Fisher Scientific #Q32852). RNA samples underwent QC on bioanalyzer and Nanodrop 2000, all samples had a RIN score >8.0 and no detectable levels of DNA. RNA (1500ng/sample, 50ng/μl) was then submitted for sequencing at Genewiz Inc. (Illumina HiSeq, single index, 10 samples/lane), and underwent RiboZero Gold purification. Sequenced reads yielding approximately 40 million paired-end 150bp single-end reads per sample. Sequenced reads were then aligned using the pseudoalignment tool, kallisto v0.44.0 114. These data were imported using txi import (‘tximport’, v1.18.0, length-scaled TPM), and vst normalized (‘DEseq2’, v1.30.1).
Dimensionality reduction was performed using ‘Rtsne’ v0.15 with perplexity value of 10 and initial dimensions of 30 on the log2 transformed normalized expression values after removing genes without any variation in expression across all samples. Linear mixed modeling was performed using the ‘lme4’ v1.1 r package with the fixed effects of time grown and clinical group for SURF1-differential expression and fixed effects of time grown and treatment with a mixed effects of the cell line for Oligo-differential expression. P values were obtained by running an Anova comparing the model for each gene to a null model that had a fixed effect of days grown (mixed effects of cell line for Oligo models) and then the value was corrected for multiple comparisons using FDR-adjustment (p<0.05). We used iPAGE to discover perturbed pathways in SURF1 and Oligo-treated cells (https://tavazoielab.c2b2.columbia.edu/iPAGE/) 71. iPAGE enables the systematic and comprehensive discovery of pathways that are significantly informative of gene expression measurements without any explicit thresholding requirements. Additionally, iPAGE is also able to detect pathways whose constituent genes are both up- and down-regulated in the treatments. Input to iPAGE included gene symbols and for each gene, a cluster identifier indicating if it was upregulated, downregulated or not differentially expressed in both (i.e. intersection) SURF1-mutant and Oligo-treated cells compared to controls. For discovering significantly over- and under-represented pathways using iPAGE, we used a stringent p value cutoff of 0.001 along with minr=1, ind=0 (to produce the most expansive set of pathway terms) and, catMin=30 (to exclude pathways with fewer than 30 genes). Timecourse and heatmaps show transcript levels relative to the median of the youngest control timepoints. Categorized genes were selected based on known mitochondrial and aging literature. Categorized pathways were categorized into meta-categories based on shared gene ontology parent processes.
DNA methylation and methylome analysis
Global DNA methylation was measured using the Illumina EPIC microarray ran at the UCLA Neuroscience Genomic Core (UNGC). DNA was extracted using the DNeasy kit (Qiagen cat#69506) according to the manufacturer’s protocol and quantified using QUBIT broad range kit (Thermo Fisher Scientific cat#Q32852). At least 375 ng of DNA was submitted in 30 µl of ddH2O to UNGC for bisulfite conversion and hybridization using the Infinium Methylation EPIC BeadChip kit. Sample positions across plates were randomized to avoid batch variation effects on group or time-based comparisons. All DNA methylation data were processed in R (v4.0.2), using the ‘minfi’ package (v1.36.0). Quality control preprocessing was applied by checking for correct sex prediction, probe quality, sample intensities, and excluding SNPs and non-CpG probes. Data were then normalized using Functional Normalization. Using the R package ‘sva’ (v3.12.0), both RCP and ComBat adjustments were applied to correct for probe-type and plate bias, respectively. After quality control, DNAm levels were quantified for 865,817 CpG Sites.
Dimensionality reduction was performed using the ‘Rtsne’ package (v0.15) with a perplexity value of 10 and initial dimensions of 30 on the normalized beta values. We ran linear mixed effects models (LMER) using ‘lme4’ v1.1. For our differential methylation analysis of SURF1, the fixed effects were assigned to ‘days_grown’ and ‘clinical_group’. For the Oligomycin treatment, the fixed effects ‘time_grown’ and ‘treatment’ and the mixed effect was assigned of the ‘cell_lines’ (i.e., donors). P values were obtained from an analysis of variance (ANOVA) comparing the model for each CpG to a null model with a fixed effect of days grown (mixed effects of cell line for Oligomycin models) and then corrected for multiple comparisons using FDR-adjustment (p < 0.05) to identify differentially methylated CpGs (DMPs). Differentially methylated regions (DMRs) were derived using the modified comb-p method in the ‘Enmix’ package (v1.26.8), with a maximum distance for DMR combination of 1000bp, a bin size for autocorrelation of 310, and FDR-adjustment cutoff of 0.01, and minimum of 3 CpGs per a DMR. Each DMP and DMR were assigned to the nearest annotated gene (IlluminaHumanMethylationEPICanno.ilm10b4.hg19 package, v0.6.0). Gene set enrichment analysis was then performed using ShinyGO 115 (v0.66, http://bioinformatics.sdstate.edu/go/) on the top 1,000 DMPs- or DMRs-associated genes based on the combined negative log p value across hyper- and hypo-methylated DMPs and DMRs.
Relative telomere length
Relative telomere length was measured by quantitative polymerase chain reaction (qPCR), expressed as the ratio of telomere to single-copy gene abundance (T/S ratio). The telomere length measurement assay was adapted from the published original method by Cawthon 116,117. The telomere thermal cycling profile consisted of: Cycling for T(celomic) PCR: Denature at 96°C for 1 minute, one cycle; denature at 96°C for 1 second, anneal/extend at 54°C for 60 seconds, with fluorescence data collection, 30 cycles. Cycling for S (single copy gene) PCR: Denature at 96°C for 1 minute, one cycle; denature at 95°C for 15 seconds, anneal at 58°C for 1 second, extend at 72°C for 20 seconds, 8 cycles; followed by denature at 96°C for 1 second, anneal at 58°C for 1 second, extend at 72°C for 20 seconds, hold at 83°C for 5 seconds with data collection, 35 cycles. The primers for the telomere PCR are tel1b [5’-CGGTTT(GTTTGG)5GTT-3’], used at a final concentration of 100 nM, and tel2b [5’-GGCTTG(CCTTAC)5CCT-3’], used at a final concentration of 900 nM. The primers for the single-copy gene (human beta-globin) PCR are hbg1 [5’ GCTTCTGACACAACTGTGTTCACTAGC-3’], used at a final concentration of 300 nM, and hbg2 [5’-CACCAACTTCATCCACGTTCACC-3’], used at a final concentration of 700 nM. The final reaction mix contained 20 mM Tris-HCl, pH 8.4; 50 mM KCl; 200 µM each dNTP; 1% DMSO; 0.4x SYBR Green I; 22 ng E. coli DNA; 0.4 Units of Platinum Taq DNA polymerase (Invitrogen Inc.); approximately 6.6 ng of genomic DNA per 11 microliter reaction. Tubes containing 26, 8.75, 2.9, 0.97, 0.324 and 0.108ng of a reference DNA (Human genomic DNA from buffy coat, Sigma cat# 11691112001) are included in each PCR run so that the quantity of targeted templates in each research sample can be determined relative to the reference DNA sample by the standard curve method. The same reference DNA was used for all PCR runs. Assays were run in triplicate wells on 384-well assay plates in a Roche LightCycler 480. The average concentrations of T and S from the triplicate wells were used to calculate the T/S ratios after a Dixon’s Q test to remove outlier wells from the triplicates. T/S ratio for each sample was measured twice. When the duplicate T/S value and the initial value varied by more than 7%, the sample was run the third time and the two closest values were reported. 26 out of the 512 samples (5%) has a CV greater than 10% after the third measurement. The inter-assay coefficient of variation (CV) for this study is 3.0%±4.3% (including the 26 samples) and 2.2%±2.0% (excluding the 26 samples). Telomere length assay for the entire study were performed using the same lots of reagents. Lab personnel lab who performed the assays were provided with de-identified samples and were blind to other data.
DNAmAge
DNAmAge was calculated using the online calculator (https://dnamage.genetics.ucla.edu/new) with normalization using the age of cell line donor as the input age. This outputted the Horvath1 (i.e. PanTissue clock), Horvath2 (Skin&Blood clock), PhenoAge, Hannum, and GrimAge estimated DNAmAges. PC-based DNAmAges were then obtained using the principal component method 80. The rates of epigenetic aging for each cell line were determined from the linear slope of timepoints between 25 to 75 days. This period ensures that Oligo treatment has taken effect, and avoids late-life changes in the behavior of DNAm clocks, providing the time window where the signal is most stable.
Data analysis and statistics
All statistical analyses were performed using GraphPad Prism (v9.0) and RStudio (v1.3.1056) using R (v4.0.2). Comparisons of groups between control, SURF1 and treatment groups were performed using mixed effects model, except for peak and rate measurements (unpaired T-test, assuming unequal variance or two-way ANOVA for concurrent measures). Interpolated curves for each experimental group is the best fit non-linear third order or fifth order polynomial functions depending on the kinetic complexity a given measurement. Data visualization and statistical analyses were generated in R (ggplot2) and Prism 8.
The time windows for specific statistical analyses were selected based on a combination of cellular growth behavior including: i) population doubling curves (e.g., stable division rates for all groups early in the cellular lifespan between days 20-50), ii) the availability of matching timepoints between treatment groups (at least 3 timepoints for all groups), and iii) potential delay to reach stable cellular phenotypes in Oligo-treated cells. To allow for adjustment to the in vitro environment, treatments began after 15 days of culture. Therefore, overall “lifespan effects” were determined between 20-150 days, which represents the maximal replicative lifespan of SURF1 cells. “Early life” effects that isolate most clearly the effects of OxPhos dysfunction, and avoid the potential accelerated aging phenotypes in SURF1 and Oligo cells, were examined using timepoints between 20-50 days. For analyses of differentially expressed genes (RNASeq, Figure 6) and differential methylation (DNAm, Figure 7) where a greater datapoint density was necessary to achieve robust mixed effects models, SURF1 cells were analyzed between 0 and 75 days (genetic defects in SURF1 are constitutive so do not require time in culture to manifest) whereas models for Oligo-treated cells used timepoints between days 35 to 110 (allowing 15 days for the effects of ATP synthase inhibition to manifest in the transcriptome, while avoiding late-life changes). All timepoints are shown in time series graphs.
Supplementary Material
Extended Data Figures 1-12, Extended Data Table 1-2, and Supplemental Files 1-10 are available in the online supplement to this article.
Author contributions
G.S. and M.P. designed experiments. M.H. provided cell lines. G.S. performed cellular studies and processed samples with assistance from A.S.M and A.T. G.S. analyzed data with assistance from A.S.M., B.S.S. and A.C. K.R.K. performed long-range PCR, cytokine arrays, and WGS analysis. S.A.D. and B.K. measured cf-mtDNA. J.L. and E.S.E. measured telomere length. A.H.C., M.L., and S.H. contributed epigenetic clocks. B.S.S. and S.T. performed the iPAGE analyses. C.B., V.P., G.L. performed mtDNA sequencing. T.T., S.R., R.S., V.K.M., B.G., R.K.M., R.H. provided data for the meta-analysis of clinical data. M.P., G.S., M.H., and E.L.S. drafted the manuscript. All authors reviewed the final version of the manuscript.
Data availability statement
The complete dataset for the present study is available without restriction and can be accessed, visualized, and downloaded using our webtool: https://columbia-picard.shinyapps.io/shinyapp-Lifespan_Study/. Data presented in this manuscript was generated as part of the Cellular Lifespan Study, which includes metabolic and endocrine experimental treatments across multiple donors described in detail in 109. The RNAseq and DNA methylation datasets for this project are available under the GEO SuperSeries GSE179849. All data preprocessing and analysis code is available on GitHub (https://github.com/gav-sturm/Cellular_Lifespan_Study).
Conflict of interests
The authors have no conflict of interest to declare.
SUPPLEMENTAL MATERIAL
Supplemental File 1. Table of differentially expressed genes for SURF1 cells over 0 to 75 days of growth.
Supplemental File 2. Table of differentially expressed genes for Oligo cells over 35 to 110 days of growth.
Supplemental File 3. Table of differentially expressed genes shared across SURF1 and Oligo cells.
Supplemental File 4. Gene expression heatmaps of select pathways in control, SURF1, and Oligo cells.
Supplemental File 5. Differentially methylated CpGs for SURF1 cells over 0 to 75 days of growth.
Supplemental File 6. Differentially methylated CpGs for Oligo cells over 35 to 110 days of growth.
Supplemental File 7. Differentially methylated CpGs shared across SURF1 and Oligo cells.
Supplemental File 8. Differentially methylated regions for SURF1 cells over 0 to 75 days of growth.
Supplemental File 9. Differentially methylated regions for Oligo cells over 35 to 110 days of growth.
Supplemental File 10. Differentially methylated regions shared in both SURF1 and Oligo cells.
Acknowledgements
We are grateful to Jane Newman, Renae Stefanetti, Robert McFarland, Robert Taylor, and Grainne S Gorman (Wellcome Center for Mitochondrial Research), Rohit Sharma and Vamsi Mootha (Massachusetts General Hospital), and other investigators whose work contributed to the meta-analysis in Figure 1; and to Marlon McGill for technical assistance with parts of this project. The cellular studies were supported by NIH grant AG066828 to M.P., the J. Willard and Alice S. Marriott Foundation, Muscular Dystrophy Association, Nicholas Nunno Foundation, JDF Fund for Mitochondrial Research, and Shuman Mitochondrial Disease Fund to MH. All research at Great Ormond Street Hospital NHS Foundation Trust and UCL Great Ormond Street Institute of Child Health is made possible by the NIHR Great Ormond Street Hospital Biomedical Research Centre. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health.
Footnotes
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.
- 7.
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.
- 34.↵
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.
- 46.
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.
- 91.↵
- 92.↵
- 93.↵
- 94.↵
- 95.↵
- 96.↵
- 97.↵
- 98.↵
- 99.↵
- 100.↵
- 101.↵
- 102.↵
- 103.↵
- 104.↵
- 105.↵
- 106.↵
- 107.↵
- 108.↵
- 109.↵
- 110.↵
- 111.↵
- 112.↵
- 113.↵
- 114.↵
- 115.↵
- 116.↵
- 117.↵