Epigenetic remodeling of lipid metabolism in response to dietary restriction is conserved across mouse strains

Lifespan-extending interventions such as DR shape the DNA methylome by delaying age-related methylation changes across the genome [49–52]. Long-term DR has further been shown to alter DNA methylation specifically over lipid metabolism genes in C3B6F1 hybrid mice. This epigenetic remodeling was induced already in young DR animals, becoming more marked in older animals and differed from the pattern of age-related methylation changes [49]. As these DR-related DNA methylation changes have so far only been described for C3B6F1 hybrid mice, we asked whether they are a general feature of DR, independent of genetic background.

We therefore compared two recently published datasets comprising whole-genome BS-seq data of liver samples (S1 Table); one from DR fed female F1 hybrid mice (C3B6F1♀) and the other from genetically heterogeneous DR fed female UM-HET3 mice (UM-HET3♀), and their ad libitum (AL) fed controls [49,50]. UM-HET3 mice are employed by the National Institute of Aging intervention testing program, thus representing a robust context to investigate longevity-related phenotypes [53]. DR increases the median lifespan of C3B6F1♀ from 902 to 1087 days (~ +25%) [49] and the mean lifespan of UM-HET3♀ mice from app. 836 to 1179 days (~ +39%; values for median lifespan were unavailable) [54]. The two datasets were generated from cohorts housed at different locations, undergoing DR from early adulthood on (DR starting at 3 and 2 months, respectively). Mice from the two studies received food of differing composition and were sacrificed at slightly different late ages (26 and 22 months, respectively) [49,50]. Furthermore, sample extraction and sequencing were conducted at different facilities and with differing sequencing conditions. Hence, a mixture of biological, experimental and technical confounding factors differed between the two DR datasets. Given the relatively low hepatic levels of non-CpG methylation, we centered our analysis on CpG methylation only. The two datasets were pre-processed and analyzed in parallel, using an unbiased and conservative quantification approach as described in [49] (for details see “Methods”). In brief, we binned the methylome using flexible-size windows covering 50 CpGs each, and overlapping adjacent windows by 25 CpGs. Focusing on methylation profiles of both DR fed C3B6F1♀ and UM-HET3♀ mice together with their respective controls, the total of four treatment groups (Fig 1A) yielded a set of 931,702 bins with a median bin size of 4 Kbp, covering 23 million CpGs in total. For each bin we calculated a single methylation value averaged across CpGs, ranging from 0 to 100% (unmethylated to fully methylated).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Dietary restriction-related methylation changes are conserved across mouse strains.

(A) Schematic representation of the data sets analysed and comparisons made. We probed for Dietary Restricted (DR)-related DMRs at old age in B6C3F1 and UM-HET3♀ mice to test their conservation across mouse strains. Pre-processing was conducted in parallel for all experimental groups (B) Number of DMRs (p<0.05, ±10%< DNA methylation difference) between each DR cohort and its respective control. Proportion of hypermethylated DMRs are indicated in %. (C) Scatterplot comparison of bin-wise differences between DR-related DMRs in C3B6F1♀ mice versus changes in the same bin in dietary restricted UM-HET3♀. Bins overlapping the genes Elovl6, Acly, Scd1 and Irs2 are highlighted. Methylation differences in dietary restricted C3B6F1♀ mice were significantly associated with differences in dietary restricted UM-HET3♀ mice (One-sided Fisher’s exact test p <0.001). Number of bins in each quadrant is indicated in blue. (D, E) Venn diagrams depicting the overlap of significantly hypo- (D) and hypermethylated (E) DMRs in both DR cohorts. Proportions of overlap relative to the DR (C3B6F1♀) set are indicated in % (p-values; *** p<0.001, ** p<0.01, * p<0.05, Fisher’s exact test). (F) Functional enrichment of differentially methylated genes (with ≥ 2 common DMRs overlapping) detected in both DR cohorts (n = 51) (G, H) Venn diagrams depicting the overlap of significantly hypo- (G) and hypermethylated (H) genes (≥ 2 DMRs per cohort overlapping) present in both DR cohorts. Proportions of overlap relative to the DR (C3B6F1♀) set are indicated in % (p-values; *** p<0.001, ** p<0.01, * p<0.05, Fisher’s exact test). (I) Functional enrichment of differentially methylated genes (≥ 2 DMRs per cohort overlapping) detected in both DR cohorts (n = 71) (J-L) Differential methylation landscape of the Elovl6, Egf (J), Scd1 (K) and Fasn (L) gene loci in both DR cohorts. The methylation increase upstream of Scd1 did no pass statistical testing. Bins are represented as bars with color scale and height indicating methylation differences. Arrows indicate gene orientation; merged mRNA structure is depicted below.

https://doi.org/10.1371/journal.pgen.1007766.g001

In both strains, DR significantly altered methylation levels of specific bins in comparison to corresponding age-matched controls (adjusted p < 0.05, Chi-squared test; minimal required difference cutoff of 10%), thus overcoming the need for further SNP allele frequency correction. In total, DR induced 1,729 differentially methylated regions (DMRs) in C3B6F1♀ and 2,714 DMRs in UM-HET3♀ mice (Fig 1B; S2 Table). About 75% of DR-related DMRs were hypermethylated, indicating a similar net effect of DR in each strain. Approximately 16% of DR-related DMRs detected in C3B6F1♀ mice showed a significant overlap with DR-related DMRs in UM-HET3♀ (p < 0.001, one-sided Fisher’s exact test), thus confirming a similar epigenetic remodeling independent of genetic background or experimental conditions (Fig 1C–1E). Interestingly, the overlap between hypermethylated DMRs (19%) was clearly larger than between hypomethylated DMRs (6%), suggesting that DR robustly alters the epigenome by locally increasing DNA methylation (Fig 1D and 1E). These common DR-related DMRs affected particularly key genes involved in lipid metabolism (Fig 1C and 1F; S3 Table).

The overlap between strains further increased (30% overlap) when comparing genes instead of individual DMRs (for details see “Methods”), suggesting differential methylation of entire gene loci as a conserved feature of DR (Fig 1G–1I). Indeed, DR induced similar differential methylation profiles over fatty acid synthesis genes in both genetic backgrounds, including Elovl6, Scd1 and Fasn [55] (S4 Table), stretching over both introns and exons but not the promoter region or neighboring genes (Fig 1J–1L). Thus, targeted epigenetic reprogramming of lipogenic genes in the liver is a robust feature of DR independent of genetic background.

Epigenetic remodeling under dietary restriction is tissue-specific

In order to assess if DR would remodel the methylome similarly in other tissues, we profiled previously published reduced-representation bisulfite-seq (RRBS-seq) data from blood of DR fed male F1 hybrid mice (B6D2F1♂) and DR fed male inbred mice (C57BL/6♂) [48] (S1A Fig). B6D2F1 mice were generated by crossing C57BL/6 with DBA/2J animals [48], thus have considerable genetic similarity with C57BL/6 mice. RRBS-seq allows differential analysis of individual CpGs due to higher read coverage per cytosine and lessened filtering by multiple testing correction [56,57]. In contrast to the liver, the majority of differentially methylated CpGs (dCpGs) showed significant hypomethylation in response to DR (S1B Fig). The effects of DR on DNA methylation in blood were well-conserved across strains, with 19 (hypermethylation) to 40% (hypomethylation) common dCpGs in both strains (S1C Fig), which is in line with previous studies [58]. We finally compared differentially methylated genes (for details, see “Methods”) in both tissues and found substantial conservation across strains but no consistent overlap across tissues (S1D and S1E Fig). Even though the discrepancy between blood and liver may be in part explained by the different BS-seq techniques, our analysis indicates that DR induces robust and highly tissue-specific epigenetic programs.

Dietary restriction, reduced growth factor signaling and rapamycin treatment induced overlapping patterns of DNA methylation

Given the robust conservation of DR-related epigenetic reprogramming, we asked whether other longevity interventions also remodelled DNA methylation specifically (Fig 2A). We profiled the hepatic methylomes of female DR C3B6F1♀ mice, male Ames dwarf mice from a closed colony (strain referred to as Prop1♂) and rapamycin treated UM-HET3♀ mice, all relative to age-matched controls (Fig 2B) [49,50]. Compared to their control group, Ames dwarf mice show an almost 50% increase in mean lifespan (controls: 723 days; Ames dwarfs: 1076 days; values for median lifespan were unavailable) [25], while the median lifespan of UM-HET3♀ mice increases by ~26% in response to daily doses of 42 ppm rapamycin (controls: 896 days; rapamycin-treated: 1132 days) [59]. For DR, we included only C3B6F1♀ mice, given that WGBS-seq and RNA-seq data for both young and old mice were available only for this strain, and were required for downstream analyses (see below). With six treatment groups, our binning approach yielded a set of 922,788 bins covering 23 million CpGs with a median bin size of around 4 Kbp. Genome-wide levels of DNA methylation were largely unaffected by intervention or strain, with most bins showing levels in the range of 70–90% (Fig 2C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Cross-strain hepatic DNA methylation changes in response to longevity interventions.

(A) Schematic of treatment comparisons: Various longevity interventions act on DNA methylation through retardation of ageing. We explored direct longevity-related (green arrow) effects on DNA methylation. (B) Schematic representation of the data sets analyzed and comparisons made. We probed for intervention-related DMRs at old age by comparing long-lived mice to their respective controls. Pre-processing was conducted for all shown treatment groups (C) Boxplot representation of global DNA methylation levels in the liver of aged DR, Ames Dwarf and Rapamycin treated mice next to their respective controls. Strain and sex are indicated above. (D) Number of significantly differentially methylated regions (DMRs) (p<0.05, ±10%< DNA methylation difference) between each longevity intervention and it’s respective control. Proportion of hypermethylated DMRs are indicated in %. (E) Scatterplot representation of methylation values of each bin in control and long-lived animals. Intervention-related DNA methylation changes are highlighted. Bins that were not differentially regulated (background) are represented in grey. Data density is indicated by color code. (F, G) Venn diagram depicting the overlap of significantly hyper- (F) and hypomethylated (G) DMRs under DR, Dwarfism and rapamycin treatment relative to the control groups at old age (p-values; *** p<0.001, ** p<0.01, * p<0.05, Fisher’s exact test).

https://doi.org/10.1371/journal.pgen.1007766.g002

In total, 15,075 DMRs showed significant methylation change in at least one intervention (intervention-related DMRs), which represents less than 1.5% of the entire methylome. Of these, DR induced 1,619 DMRs, rapamycin 1,415 and genetic dwarfism 12,619 (S5 Table). Both non-genetic interventions thus had a weaker impact on the epigenome than genetically reduced somatotropic signaling (Fig 2D and 2E). Both DR and dwarfism caused predominantly gains in DNA methylation, whilst rapamycin treatment induced equal amounts of hyper- and hypomethylated DMRs (Fig 2D and 2E). Remarkably, 412 hypermethylated DR-related DMRs (33%), and 133 hypermethylated rapamycin-related DMRs (17%) were also detected in Ames dwarf mice (Fig 2F). In addition, 9 hypermethylated DMRs were common among all three interventions. Given that the global methylome remained unchanged, these overlaps were highly significant (each: p < e^-16, Fisher’s exact test), strongly indicating that their co-occurrence was not due to chance. In clear contrast, we did not detect any hypomethylated DMRs common to all three interventions and only few hypomethylated DMRs between any two interventions, for which we found no convincing functional association (Fig 2G).

The relatively small overlap of hypermethylated DMRs between DR and rapamycin-treated mice could partly result from low read coverage, which reduces the statistical power. To evaluate this possibility, we conducted a less constrained analysis and asked whether significant DR-related DMRs would associate with methylation changes in Ames dwarf or rapamycin treated mice, independent of their statistical significance (for details see “Methods”). DR-related DMRs associated with methylation changes in Ames dwarf (p < 0.001, one-sided Fisher’s exact test) but not in rapamycin-treated mice (Fig 3A and 3B), confirming our earlier result.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Conserved intervention-related remodeling of gene body methylation.

(A, B) Scatterplot comparison of bin-wise methylation differences between DR-related DMRs versus changes in the same bins in Ames dwarf (A) and rapamycin-treated (B) mice, respectively. Bins overlapping the Elovl6 gene are highlighted in red. Methylation differences under DR were significantly positively associated with differences in Dwarfs (One-sided Fisher’s exact test p <0.001) but not with rapamycin-treated mice. Number of bins in each quadrant is indicated in blue. (C) Enrichment analysis of common hypermethylated DMRs between DR and Dwarfs (n = 403) and rapamycin treatment and Ames dwarf mice (n = 124) over genomic and ENCODE chromatin segmentation. Bars indicate the ratio of the observed DMR frequency and the average frequency across the genome (log2-transformed; adjusted p-values; *** p<0.001, ** p<0.01, * p<0.05, Fisher’s exact test). (D) Venn diagram depicting the overlap of hypermethylated genes (≥ 2 DMRs overlapping) under DR, dwarfism and rapamycin treatment (p-values; *** p<0.001, ** p<0.01, * p<0.05, Fisher’s exact test). (E, F) Functional enrichment of common hypermethylated genes between Ames dwarf mice and DR (E; n = 101) or rapamycin-treatmed mice (F; n = 58). (G,H) Differential methylation landscape of the Elovl6 (G) and Ghr (H) gene loci. Bins are represented as bars with color scale and height indicating methylation differences. Arrows indicate gene orientation; merged mRNA structure is depicted below.

https://doi.org/10.1371/journal.pgen.1007766.g003

Finally, we considered that the relative large absolute bin size (~ 4 Kbp) might reduce the detectability of common methylation patterns and thus repeated the binning and differential methylation analysis with bins covering only 20 CpGs (2.3 mio. bins; median size: 1.49 Kbp) and 10 CpGs (4.6 mio. bins; median size: 649 bp) (S2A Fig). Consistent with the 50-CpG binning, the majority of DMRs under DR feeding and in Ames dwarf mice tended to be hypermethylated, and genetic dwarfism caused more DMRs than DR or rapamycin treatment (S2B Fig). Reducing the number of CpGs per bin increased the overall noise between long-lived and control animals (S2C Fig), yet the overlaps among the hypermethylated DMRs remained largely consistent (S3A Fig) and common 10-CpG-DMRs mapped to the same regions detected with 50-CpG bins (S3B Fig). Remarkably, we did not detect any hypomethylated DMRs common to all three interventions, thus validating our previous findings (S3C Fig).

Taken together, reduced growth factor signaling and reduced TOR signaling induce DNA methylation signatures in distinct regions of the genome, pointing to a similar epigenetic remodeling as identified before under DR. In addition, several hundred hyper- but not hypomethylated DMRs are common to at least two interventions, indicating a partly conserved epigenetic signature across interventions, mouse strains and sexes.

Ames Dwarf mice share functionally distinct gene body methylation patterns with rapamycin-treated and dietary restricted mice

Following our findings from DR-related hypermethylation of lipid metabolism genes, the overlapping hypermethylation signatures among interventions could suggest a partially shared epigenetic program, acting on specific, functionally connected regions of the genome. In line with our hypothesis, hypermethylated DMRs shared between Ames dwarf and either DR or rapamycin treatment were strongly enriched over gene bodies and open/active chromatin, confirming coordinated methylation increase over euchromatic, coding regions of the genome (Fig 3C; Chromatin segmentation data from the ENCODE project [60]). To assess if specific functional elements within genes were particularly associated with intervention-related methylation changes, we repeated the enrichment analysis using the 10-CpG-DMRs (S4A–S4C Fig) for improved resolution. We found equal enrichment of common intervention-related DMRs over exons, intron and exon/intron-boundaries, thus confirming that these hypermethylation patterns tend to equally affect each part of the gene body (S4D Fig). Using annotation from the ENCODE and Cistrome platforms [61], we also confirmed a strong enrichment of binding sites of several transcription factors and DNA-binding proteins (S4B and S4C Fig).

As described before [49], hypermethylated genes under DR were functionally enriched for fatty acid metabolism and regulation of lipogenesis (S5A Fig; S6 Table). In contrast, rapamycin treatment caused hypermethylation primarily over genes involved in carbohydrate and amino acid metabolism, and several cell proliferation and growth factor receptor genes (S5B Fig), including mediators of TGF-β signaling (Acvr2b [62] Ltbp1, Ltbp4 [63]) and the growth hormone receptor Ghr (S5B Fig; S7 Table). Differentially methylated genes in Ames dwarf mice exhibited the strongest functional association, with enrichment for lipid synthesis, amino acid metabolism, wound healing, cellular differentiation and growth factor binding (S5C Fig; S8 Table). Genetic dwarfism thus appeared to cause hypermethylation over similar pathways as observed under DR- and rapamycin-treatment separately.

We therefore analyzed the functional association of genes that were hypermethylated in at least two interventions (≥ 2 overlapping DMRs per intervention). Indeed, Ames dwarf mice shared hypermethylation patterns over lipid metabolism genes with DR fed mice (Fig 3D and 3E; S10 Table), while also showing hypermethylation of cell proliferation and growth factor signaling genes similar to rapamycin-treated mice (Fig 3D and 3F; S11 Table). As exemplified by Elovl6 and Ghr (Fig 3G and 3H, S4D Fig), DNA methylation was altered over larger regions of the gene body, similar to findings in both DR cohorts (Fig 1J–1K). In summary, the non-genetic interventions DR and rapamycin treatment thus resemble different components of the epigenetic program induced by genetic dwarfism. Interestingly, dwarfism reduces expression of genes involved in hepatic lipogenesis [64,65] and alters expression of genes involved in growth factor signaling and cell proliferation [64,66], thus paralleling the observed methylation signatures.

Hypermethylation patterns associated with longevity-extension are not explained by retardation of ageing alone

DR, rapamycin treatment and reduced somatotropic signaling slow down age-related changes of the methylome [47–50,52,67]. The common hypermethylation patterns described here were analyzed in old animals and could thus result from protection of specific genes against age-related methylation changes. Alternatively, common hypermethylation DMRs may have formed by an active reprogramming that is partly unrelated to ageing. To assess whether longevity intervention-related DMRs resulted solely from delayed ageing, we first profiled age-related changes in wild type, normally fed, non-drug-treated members of each of the three mouse strains (S6A Fig). Pre-processing and binning based on young and old control samples across strains yielded 859,750 bins, showing globally only mild age-related changes (S6B Fig). Ageing induced only 3,349 and 3,133 DMRs in C3B6F1♀ and UM-HET3♀ mice, respectively, but over 8,500 significant changes in the genetic background of the Prop1♂ strain, suggesting a strong interaction between genotype and age on DNA methylation (S6C Fig; S12 Table). Age-related DMRs in the genetic background of the Prop1♂ mice furthermore showed a bias towards age-related hypomethylation (S6C and S6D Fig), which was not observed in the other two wild type mouse strains. In order to minimize these strain-specific biases, we selected age-related DMRs present in at least two strains (S6E and S6F Fig) as reference for a comparison with intervention-related DMRs. Strikingly, 282 of 545 (53%) intervention-related, hypermethylation DMRs (found in at least two interventions) were not hypomethylated during ageing (S7A Fig; S5 Table). This result suggests that intervention-related methylation changes were not solely a consequence of general protection against ageing-related changes in the epigenome and instead reflect a specific feature of longevity interventions.

Intervention-related DNA methylation is rapidly induced at young age

We next determined whether intervention-related methylation changes were already detectable at young age. We analyzed DNA methylation of young DR-treated C3B6F1♀ (5 months) and Ames dwarf mice (2 months) in comparison to their respective controls (S3B; data for young rapamycin- or DR-treated UM-HET3♀ mice were unavailable). Our binning approach yielded 1,018,051 bins covering 25.5 million CpGs. Differential methylation analysis detected 1,944 DR- and 5,547 dwarfism-related DMRs at young age (S13 Table), with similar hyper-to-hypomethylation ratios as observed in old animals (compare Fig 2D, S7C Fig). DR- and dwarfism-induced methylation changes at young age were also significantly associated with each other, confirming common epigenetic patterns already at this age (S7D Fig). Strikingly, intervention-related, hypermethylated DMRs in young animals were significantly enriched over intervention-related DMRs that were also detected at old age. Reprogramming of the epigenome in response to DR and reduced somatotropic signaling is thus induced in young adult animals, prior to age-related methylation changes (S7E Fig).

We further asked if intervention-related DNA methylation patterns in young mice also affected the same genes as seen at old age and found a weak but significant overlap (12%) between young and old DR fed mice (S7F and S7G Fig), including lipid metabolism genes Elovl6, Scd1 (S7H and S7I Fig) and Pnpla3. The epigenetic reprogramming under DR is thus induced early in adulthood, but becomes more pronounced during chronic long-term treatment. In contrast, the set of common hypermethylated genes between young and old Ames dwarf mice covered over 650 genes (69% overlap), revealing that the extensive epigenetic reprogramming in response to reduced somatotropic signaling occurs early in life and is maintained life-long (S7G Fig). DR and reduced somatotropic signaling thus remodel the epigenome before age-related changes occur, additionally confirming that hypermethylated patterns detected in long-lived mice cannot solely result from deceleration of ageing.

Joint BS-seq profiling identifies candidate marker regions of longevity

While reliable DNA methylation markers (‘epigenetic clocks’) for murine ageing have been previously developed, including local markers like Elovl2, a specific epigenetic biomarker predicting interventions that will ameliorate ageing in mice does not exist [46–48,68]. Such a marker would probe immediately for responsiveness to interventions, instead of wait for the delay of age-related effects. The WGBS-seq profiles analyzed here comprised three different interventions in different genetic backgrounds, thus representing a suitable dataset to investigate the general feasibility of an intervention-related marker. We found nine hypermethylated DMRs shared among all three longevity interventions (see S7A Fig; S5 Table), as exemplified by the Tat gene locus (S8A and S8B Fig). Of these, four hypermethylated DMRs did not overlap with age-related hypomethylation (S7A Fig; S5 Table) making them suitable marker candidates to probe for longevity-intervention-specific effects. DNA methylation levels of individual cytosines present within the four DMRs increased in concert in all three longevity interventions (p<0.001, Paired Wilcoxon rank-sum test) (S8C Fig). We independently tested the putative marker regions in DR-treated UM-HET3 mice, which were not used to identify these regions previously, and found the expected hypermethylation pattern (p<0.001, Paired Wilcoxon rank-sum test) (S8C Fig).

Next, we reduced the DMRs to four representative 500 bp regions that can be investigated by PCR-based assays. We confirmed that hypermethylation of the 125 resulting cytosines was still well correlated for all three interventions (S9A Fig; S14 Table). We also detected significant hypermethylation of these cytosines in DR-treated UM-HET3 mice (S9A; p<0.001, Paired Wilcoxon rank-sum test). The four regions were selected to not overlap with age-related changes in DNA methylation (S7A Fig) and accordingly, we did not observe consistent methylation changes with age for the representative 500 bp regions (S9B Fig).

We asked if the putative marker regions would be able to predict extended lifespan when analysed in young DR or Ames dwarf mice. Indeed, young Ames dwarf mice showed significant hypermethylation over the marker cytosines, indicating that reduced somatotropic signaling during development primes for an extended lifespan (S9C Fig). Our marker readout is thus in agreement with analyses of the epigenetic clock in young Ames dwarf mice, which is already significantly decelerated at 2 months of age [47]. Surprisingly, however, young DR fed C3B6F1♀ mice (3 months ad libitum feeding followed by 2 months DR) showed no significant difference over the marker regions. The marker thus indicates that 2 months of DR feeding at young age do not suffice to extend lifespan later on (S9C Fig). Indeed, recent studies in rats confirmed that DR feeding limited to two to six months of age does not cause an extension of lifespan later on [12]. This result is also consistent with findings from DNA methylation clocks, where long-term, but not short-term DR, resulted in a significant reduction of DNA methylation age [48]. In addition, we tested whether the marker regions could also be used in other tissues and analysed the RRBS-seq data from blood of DR fed B6D2F1♂ mice (see S1 Fig). Only half of the 125 CpGs were covered by RRBS-seq and the corresponding methylation levels in blood of control animals did not match the observed levels in the liver (compare S9A and S9D Fig). However, despite the low coverage there was no evidence for a predictive function of this marker region in the blood.

In order to verify our marker experimentally, we measured DNA methylation at the marker regions in five DR and five AL fed C3B6F1♀ mice (24 months of age) from an independent batch of animals using targeted Bisulfite Amplicon sequencing (BS-AS; for details see “Methods”) [69]. To test if the marker would probe successfully for other lifespan-extending interventions, we also performed BS-AS in four female C3B6F1 mice genetically deficient for Irs1 (Irs1KO; 24 months of age) [70] and four controls. The BS-AS assay measured DNA methylation levels at 80 CpGs, thus covering most of the CpGs within the marker regions (S10A Fig). Unsupervised hierarchical clustering of methylation levels at these 80 CpGs separated most of the samples into controls and long-lived mice, indicating that DNA methylation at these sites can classify for extended lifespan (S10B Fig). Averaged methylation levels across all replicates clustered similarly, confirming that they are not biased by single outlier replicates (S10B Fig). Remarkably, we detected a significant concerted hypermethylation at these CpGs both in DR fed and Irs1KO mice (S10C; p<0.05, Paired Wilcoxon rank-sum test). Consistently, we found 18 hypermethylated dCpGs under DR feeding and 32 dCpGs in Irs1KO mice (4 and 12 hypomethylated dCpGs, respectively). One of the four marker regions, however, appeared to be false positive in the WGBS-seq data, as most long-lived animals showed no consistent change at these CpGs (S10B Fig). Excluding the corresponding CpGs improved the result of the detection of concerted hypermethylation (S10D; p<0.05, Paired Wilcoxon rank-sum test).

In summary, concerted hypermethylation of these marker cytosines in liver correlates specifically with longevity upon DR, reduced TOR and reduced growth factor signaling. By testing the marker cytosines in an independent model of longevity, we demonstrate the feasibility of developing novel epigenetic markers that can probe specifically for intervention outcome.

Epigenetic reprogramming parallels transcriptional repression of growth factor signaling networks in Ames dwarf mice

Gene body hypermethylation under DR has been associated with transcriptional repression of lipogenesis, pointing to a tight coupling between DR-related remodeling of the epigenome and gene expression [49]. We asked if expression and differential methylation would also be inversely associated in Ames dwarf mice, given their similar epigenetic patterns to DR-treated mice. Focusing on genes with at least 2 dwarfism-related DMRs mapping to the gene body, we plotted methylation differences against expression changes (Fig 4A; S15 Table). As previously described for DR [49], the expression of hypermethylated genes tended towards down-regulation in Ames dwarf mice (p < 0.001, Binomial test) and this was even more pronounced when considering only significantly differentially expressed genes (Fig 4A; S15 Table). We also found this inverse correlation to be present in young Ames dwarf mice (S10A Fig; S16 Table), suggesting this association to be early-onset and maintained life-long. Correspondingly, we found at young and old age similar functional enrichments of GO terms and Reactome pathways, including Unfolded Protein Response (UPR), lipid metabolism and wound healing (Fig 4B, S10B Fig; S16, S17 and S18 Tables). In addition, we found strong dwarfism-related hypermethylation and repression of several growth factor signaling and cell survival genes (Fig 4B–4D, S10B and S10C Fig; S16, S17 and S18 Tables). Notably, this included the growth hormone receptor (Ghr, Fig 3H) and several direct and indirect downstream targets, such as the systemic growth mediator Igf-1 [71] and the hepatoprotective genes Egfr, Lifr and Onecut1 (Hnf-6) [72–78]. These regulators of cell survival and growth carry out key functions in liver regeneration and also play major roles in promoting cancer onset and progression [79–87].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. DNA methylation differences in Ames dwarf mice predict expression changes.

(A) Scatterplot of expression differences versus methylation differences of dwarfism-related DMRs. Dashed lines indicate DNA methylation cutoff of > ±10%. DNA hypermethylation associated significantly with reduced gene expression (Binomial test p < 0.001). Number of differentially methylated genes in each quadrant is indicated in blue and red, for all genes and differentially expressed genes, respectively. (B) Gene ontology and Reactome enrichment of genes with a negative correlation of gene expression and methylation. Lengths of bars represent negative ln-transformed, adjusted pvalues using Fisher’s exact test. Cells indicate log2-foldchanges (log2FC) between Ames dwarf and control group per gene. (C) Differential methylation landscape of the Egfr gene locus at young and old age. Bins are represented as bars with color scale and height indicating methylation differences. Arrows indicate gene orientation; merged mRNA structure is depicted below. (D) Egfr mRNA expression by RNA-sequencing in control and Ames dwarf mice at young and old age, respectively (n = 4 vs 2 and 4 vs 4). (E) Schematic outline of differential methylation and gene expression along the Hypothalamic–pituitary–somatotropic axis in Ames dwarf mice: In Ames dwarf mice, the release of Growth Hormone (GH) is impaired. The corresponding Growth hormone receptor (Ghr) shows respective transcriptional downreguation in the liver. Direct (solid arrows) and indirect (dashed arrows) downstream targets of GHR signaling, show elevated methylation levels and corresponding transcriptional repression. Further hypermethylated/downregulated targets include the hepatoprotective factors Egfr, Lifr and Onecut1, involved liver growth regulation, cell survival and metabolism.

https://doi.org/10.1371/journal.pgen.1007766.g004

Whilst increased DNA methylation in gene bodies is typically correlated with increased expression [88], recent studies also demonstrated an inverse association [89–91]. For instance, the strongest inverse association between DNA methylation and expression in humans was previously found in so-called undefined intragenic DMRs (uiDMRs), which locate 0.3–8 Kbp downstream of the promoter. [89]. In line with these findings, we observed an enrichment of intervention-related methylation changes at uiDMRs (Fig 3C and S4A Fig). Moreover, the significant inverse association between dwarfism-related changes in methylation and expression remained consistent when restricting the analysis to uiDMRs (S10D and S10E Fig). Differential methylation over uiDMRs may thus be a putative mechanism mediating the transcriptional consequences of the epigenetic reprogramming.

In summary, our results thus reveal strong epigenetic and transcriptional silencing of the growth hormone signaling axis, which may contribute to delayed occurrence of neoplastic diseases in Ames dwarf and growth hormone signaling deficient mice [9,29,30].

Dwarfism and DR remodel methylation and expression of Fgfr3

Given the common epigenetic signatures occurring across longevity interventions (Fig 2E and 2F), we asked whether methylation and transcriptional changes would also be common between DR fed and Ames dwarf mice (RNA-seq data for rapamycin-treated mice were not available). In line with previous work [49], we found hypermethylation and transcriptional repression of lipogenesis genes to be significantly associated in DR fed mice (p < 0.05, Binomial test; S11A and S11B Fig; S19 and S20 Tables). DR and Ames dwarf mice showed similar transcriptional patterns, with an overlap of 421 commonly regulated genes (S11C and S11D Fig). Surprisingly, the overlap between commonly hypermethylated and significantly repressed genes comprised only a non-significant set of five genes, suggesting that DNA methylation and transcription are differently coupled under DR and reduced somatotropic signaling (S11E Fig). Interestingly, however, fibroblast growth factor receptor 3 (Fgfr3, S11F–S11H Fig), a key mediator of mitogenic signals and known activator of both the Ras-Mapk and PI3K-AKT pathways [92], was found to be hypermethylated and repressed in DR and Ames dwarf mice.

As a crucial regulator of cell growth and survival, Fgfr3 is frequently up-regulated in various cancers including hepatocellular carcinoma [93–96]. Hepatic Fgfr3 is also repressed in mouse models of genomic instability, which exhibit dwarfism and GH-IGF1-signaling deficiency, thus paralleling our findings in Ames dwarf mice [97,98]. Hence, the anti-tumorigenic effects of DR and dwarfism [9,29] may involve repression of Fgfr3. Furthermore, in-vitro studies with bladder cancer cell lines confirmed Fgfr3 as a direct activator of lipogenesis via Srebf1 [99–101], which itself becomes hypermethylated and repressed under DR [49,102,103]. Remarkably, gene expression changes induced by knockdown of Fgfr3 in RT112 cells [100] overlapped significantly with expression changes observed in livers of DR fed C3B6F1♀ mice (S12A–S12C Fig; S21 Table) including down-regulation of various hypermethylated lipogenesis genes (S22 Table). Our analysis thus point to a role of Fgfr3 in mediating the effects of DR and genetic dwarfism in the liver.

Conclusion

This study provides novel insights into the effects of longevity interventions on DNA methylation and how they may causally contribute to improved health at old age. Dietary restriction (DR), rapamycin treatment and reduced somatotropic signaling remodel DNA methylation partly independent of ageing, converging on defined marker regions predictive of intervention outcome. In addition, DR and rapamycin treatment resemble separate hypermethylation patterns of Ames dwarf mice, affecting lipid metabolism and growth factor signaling genes, respectively. For DR and Ames dwarf mice, we furthermore demonstrate significant association between DNA methylation and gene expression. Therefore, our work establishes hypermethylation signatures as a link between the long-term effects of several longevity models and their transcriptional networks controlling somaotropic signaling and metabolism.

Ethics statement

The study involving live mice was performed in strict accordance with the recommendations and guidelines of the Federation of the European Laboratory Animal Science Association (FELASA), with all protocols approved by the Landesamt für Natur, Umwelt und Verbraucherschutz, Nordrhein-Westfalen, Germany (reference number: 84–02.04.2015.A437). For dissections mice were killed by cervical dislocation.

Bisulfite Amplicon sequencing

Bisulfite Amplicon sequencing (BS-AS) libraries were prepared from liver tissue of 24 months old female C3B6F1♀ mice with the following treatment groups: 5 DR or 5 AL fed mice (for animals husbandry see [49]), 4 homozygous IRS1-mutant and 4 wild-type mice. DNA was isolated using the DNA/RNA Allprep kit (Qiagen) and bisulfit conversion was done using the Imprint DNA modification kit (Sigma) following the one-step modification procedure. The KAPA HiFi Uracil+ Master Mix kit (Kapa Biosystems) was used to amplify the 4 regions of interest using the primers indicated in S23 Table. A single round of PCR amplification was used per sample whereby the forward primer contained a 6bp sample barcode. The PCR program was: 95C 5min; 35 repeats of 98°C 20s, 60°C 15s, 72°C 60s; 72°C 10min. PCR products were purified using Ampure XP beads (Agencourt), analyzed on a 4200 Tape station system (Agilent) and subsequently pooled for library preparation. Two libraries were generated with 40 and 32 samples for the AL vs. DR and IRS-1 vs. wild-type comparison, respectively. Sequencing was performed using 250bp paired end reads on an Illumina HiSeq2500.

WGBS-seq quantification and differential methylation analysis

An overview of the analyzed datasets and treatment conditions is given in (S1 Table). Raw datasets were downloaded from the GEO repository under accession IDs GSE89275 and GSE92486. WGBS-seq data were preprocessed and analyzed according to [49]. Raw sequence reads were trimmed using Trim Galore! (v0.4.2). Trimmed sequences were aligned using Bismark [151] (v0.16.3). Methylation calls were extracted after duplicate sequences had been excluded. Data visualization and analysis were performed using SeqMonk and custom RStudio scripts. Each age and treatment group was present in biological triplicates or quadruplets. Data from replicates of the same treatment group were merged using SeqMonk’s data group option, in order to enhance coverage and detection of subtle differences. The high consistency between replicates across the used datasets has been previously verified in the original studies [49,50]. Regions with an unusually high number of observations were detected and filtered using non-overlapping 25 kb windows, followed by read count quantification and subsequent BoxWhisker filter implemented in SeqMonk with stringency >10. To achieve a fair and unbiased analysis of the methylome, we constructed windows containing 50 CpGs over the whole genome, spaced 25 CpGs apart. To demonstrate the validity of our findings independent of the window size, we partly repeated the analyses with windows containing 20 or 10 CpGs, spaced 10 and 5 CpGs apart, respectively. Each window therefore contained around the same amount of data and all windows had similar technical noise and statistical power. Furthermore, to reduce the effect of coverage differences between samples, only cytosines covered by at least three observations in all analyzed treatment groups were used for window construction, opposing the limitation for total number of bins and resulting bin size. Methylation for each window was calculated as the average of methylation for each covered CpG position. Windows that contained significantly different methylation levels (pairwise Chi-squared tests with subsequent multiple testing correction; adjusted p-value <0.05) and a minimal difference cutoff of 10% were defined as differentially methylated (DMR) [49].

Filtering of CpGs and binning was conducted for each of the following analyses separately: a) Old DR treated C3B6F1♀ or UM-HET3♀ mice including controls (Fig 1) b) old DR fed C3B6F1♀, Ames dwarf and rapamycin treated UM-HET3♀ mice, including controls (Fig 2) c) young and old C3B6F1♀, Prop1♂, UM-HET3♀ control mice (S2 Fig) d) young DR fed C3B6F1♀ and Ames dwarf mice, including controls (S3 Fig).

For the example in S2C, the methylation landscape over the Tat gene locus was visualized by calculating average CpG methylation levels of 500 bp 400 bp overlapping bins to improve resolution.

Comparative DMR analysis

Significance of DMR sets common between two interventions was tested using Fisher’s exact test with the entire methylome as background. Due to the observed gene body-wide methylation profiles, differentially methylated genes were defined as common when overlapped by at least two DMRs in each intervention (not necessarily the identical DMRs). For comparing longevity-related DMRs with age-effects (S3A Fig), we selected every bin overlapping age-related DMRs detected in at least two strains as reference. To test if methylation differences induced by two interventions would be associated in general, we asked if significant DMRs detected in one intervention would correlate with methylation changes in another intervention, independent of statistical significance. To this end, we selected DMRs detected in DR treated C3B6F1♀ mice and plotted the observed methylation differences for the same bins in other interventions (DR, rapamycin and dwarfism). The number of bins within the four resulting quadrants was counted and directionality of differences was tested using one-sided Fisher’s exact test.

Enrichment analyses of DMRs over genomic and epigenetic elements

For a given genomic element, we calculated the enrichment for a set of DMRs by counting the number of DMRs overlapping the element of interest and compared it to the set of background bins. One-sided Fisher’s exact test with subsequent multiple testing correction (if enrichment analysis was run over multiple elements) was performed to determine statistical significance of enrichment. Since most elements showed highly significant enrichment, we log-transformed the adjusted p-values. Ratio of the observed DMR frequency and the average frequency across the genome (‘observed/expected’) for individual elements were used to compare enrichments across elements and between DMR sets. This way, we were able compare the strengths of enrichments across elements in the same analysis.

RRBS-seq quantification and differential methylation analysis

Raw datasets were downloaded from the GEO repository under accession ID GSE80672. Raw sequence reads were trimmed using Trim Galore! (v0.4.2). Trimmed sequences were aligned using Bismark [151] (v0.16.3). Data visualization and analysis were performed using SeqMonk and custom RStudio scripts. Data from replicates of the same treatment group were merged using SeqMonk’s data group option, in order to enhance coverage and detection of subtle differences. To reduce the effect of coverage differences between samples, only cytosines covered by at least five observations were used for the differential methylation analysis. Differentially methylated cytosines (dCpGs) were identified by pairwise Chi-squared tests with subsequent multiple testing correction (adjusted p-value <0.05) and a minimal difference cutoff of 10%. Filtering of CpGs and binning was conducted for old DR treated C57BL/6♂ or B6D2F1♂ mice including controls (S1 Fig).

Functional enrichment analyses of differentially methylated genes

For the analyses employing WGBS-seq data, we performed functional enrichment of genes overlapped by at least two DMRs and only retained those showing at least ±10% average difference. We thereby avoided unclear cases with an equal extent of hyper- and hypomethylation occurring over the same gene. Enrichment analysis was carried out using topGO [152] and ReactomePA [153] with all genes in the genome as background.

For the analyses employing RRBS-seq data, we tested each gene for significant enrichment of overlapping dCpGs compared to all detected CpGs overlapping and only retained those with p-values < 0.05 from Fisher’s exact test (after multiple testing correction) and at least ±10% average difference.

Marker regions of longevity analysis

We quantified the methylation for each CpG position individually that mapped within the four hypermethylated DMRs shared between all three longevity interventions but not with age-related hypomethylation (S3A Fig). CpGs covered by at least three observations in at least one analyzed treatment group was included and the methylation change across CpGs tested using paired Wilcoxon rank-sum test. The analysis was repeated after manually selecting one representative 500 bp regions for each of the four DMRs.

BS-AS quantification and differential methylation analysis

Raw sequencing libraries were de-multiplexed before mapping and methylation extraction was conducted as described for WGBS-seq. De-duplication step was omitted. Data visualization and analysis were performed using SeqMonk and custom RStudio scripts. Data from replicates of the same treatment group were averaged using SeqMonk’s replicate set option. To reduce the effect of coverage differences between samples, only cytosines covered by at least three observations were used for the differential methylation analysis. Differentially methylated cytosines (dCpGs) were identified by pairwise, replicate-sensitive logistic regression tests with subsequent multiple testing correction (adjusted p-value <0.05). Methylation change across CpGs was tested using paired Wilcoxon rank-sum test.

RNA-seq analysis

For young and old Ames dwarf and DR-fed C3B6F1♀ mice, RNA-seq data derived from liver tissue of the same animals used for WGBS-seq was downloaded from the GEO repository under accession IDs GSE89275 and GSE92486 (S1 Table). Two replicate samples of young Ames dwarf mice showed very high amounts of low quality reads and insufficient coverage and were therefore removed from the analysis. Raw sequence reads were trimmed to remove adaptor contamination and poor-quality reads using Trim Galore! (v0.3.7, parameters:—paired—length 25). Trimmed sequences were aligned using Tophat2 [154] (v2.0.14, parameters:—no-mixed—library-type = fr-firststrand -g 2 -p 15 -r 500—mate-std-dev 525), supplying GENCODE annotation [155] (release M9, main annotation) for improved mapping. Multi-mapped reads were filtered using samtools [156] (v1.2, parameters: view -F 0x100 -b–h). Data visualization and analysis was performed using SeqMonk, custom RStudio scripts and the following Bioconductor packages: Deseq2 [157], topGO [152], ReactomePA [153] and org.Mm.eg.db. For Figs 4B, S5B, S6B and S7C, we further used the CellPlot package (https://github.com/dieterich-lab/CellPlot). To account for tissue-specific expression, we defined all genes passing the independent filtering of Deseq2 [157] as ‘expressed’ (14154 and 13894 genes for C3B6F1♀ and Prop1♂ mice, respectively). Differentially expressed genes were determined using Deseq2´s Wald test [157]. P-values were adjusted for multiple testing. Genes were considered to be significantly differentially expressed with a p adjust < 0.05 and no cut-off for fold change was used. Unless stated otherwise, the set of expressed genes was used as background for all functional enrichment analyses involving expression data.

Correlation analysis of methylation and transcription

To test for correlation between differential methylation and gene expression, we considered only genes overlapped by at least two DMRs. For associations between differential methylation over uiDMRs and expression, we considered only genes with their uiDMRs (0.3–8 Kbp downstream of gene start) overlapped by at least one DMR. We averaged methylation differences for multiple DMRs overlapping the same gene (uiDMR) before plotting, and only retained those showing at least ±10% average difference. We thereby avoided unclear cases with an equal extent of hyper- and hypomethylation occurring over the same gene (uiDMR). We plotted log2-fold expression changes versus methylation differences and the distribution of genes among the four resulting quadrants. Due to the very few hypomethylated genes common among longevity interventions, we focused on hypermethylated genes for statistical testing. For the group of differentially expressed and hypermethylated genes above the minimum methylation cut-off, we ran a binomial test to probe for a significant trend towards down- or up-regulation. Genes that showed a significant association of hypermethylation and transcriptional repression were analyzed for functional enrichment using topGO [152].

Comparative transcriptome analysis for DR liver tissue and Fgfr3 knockdown in RT112 cells

To compare the transcriptional changes induced by DR with in-vitro knockdown of Fgfr3, we obtained a list of Fgfr3-regulated genes from pre-preprocessed microarray results from [100]. The list comprises significantly differentially expressed genes in three independent doxacycline-treated RT112 cell lines expressing doxycycline-inducible shRNAs targeting Fgfr3 (each cell line expresses another Fgfr3-shRNA). Gene expression was compared to a line expressing a control shRNA [100]. For each gene showing differential expression in Fgfr3 knockdown cells and liver tissue of DR-treated mice, we plotted log2-fold expression changes under DR (compared to ad libitum feeding) versus log2-fold expression changes under Fgfr3 knockdown (compared to control). We plotted the distribution of genes among the four resulting quadrants and tested for directionality using Fisher’s exact test.

Definition of genomic and epigenetic elements

Gene annotation used in this study was obtained from the UCSC Genome Browser database [158]. We defined further genome annotation as follows:

1. Promoter region: Manually defined region stretching from 5 Kbp upstream and 100 bp downstream of the transcription start site.
2. uiDMR: According to [89], uiDMRs were defined as regions stretching from 300 to 8000 bp downstream of the transcription start site. Analysis was limited to genes with a total size of at least 16 Kbp.
3. CGI: Annotation based on CXXC affinity purification plus deep sequencing (CAP-seq) experiments [159] CGIs were further classified unambiguously into promoter, gene and intergenic CGIs. If a CGI overlapped both promoter and gene body, it was classified as a promoter CGI.
4. Chromatin states: Publicly available, pre-processed chromatin annotation in liver tissue based on four histone marks (ENCODE chromHMM segmentation) were obtained from the ENCODE table browser [60] and re-mapped to the GRCm38 genome using UCSC Genome Browser’s LiftOver [160]. Details about the corresponding methods can be found here: https://main.genome-browser.bx.psu.edu/cgi-bin/hgTrackUi?g=meryChromHmm7s&db=mm9).