Reversible and causal epigenetic information loss in liver aging and disease

Abstract

The loss of epigenetic information has been proposed as a driver of aging and diseases, but the reversibility and causality of this process remain underexplored. Here, we analyze liver-unique methylation sites—genomic loci that show distinct methylation patterns in the liver compared to other tissues. Upon disease progression, these sites overwhelmingly regress toward the pan-tissue average. In addition, we demonstrate that this regression also occurs in a majority of these sites during normal aging. Using causal sites previously identified by Mendelian randomization analysis, we identify significant enrichment of liver-unique methylation sites in causal aging-associated loci, particularly sites that are highly methylated in healthy liver. Remarkably, repeated fasting, a metabolic intervention known to improve liver function, partially restores the DNA accessibility patterns at these sites. This restoration also occurs in isolated hepatocytes subjected to fasting-mimicking conditions, suggesting the effect is cell-autonomous rather than due to changes in tissue composition. The liver-unique methylation sites are enriched for binding sites of key metabolic transcription factors and show significant overlap with genetic variants associated with liver disease risk, suggesting a mechanistic link between epigenetic information loss and liver dysfunction. Our findings establish epigenetic information loss as both a marker and mediator of liver aging and disease, while demonstrating its potential reversibility through metabolic interventions.

Significance Statement.

The loss of epigenetic information, herein characterized by the regression of tissue-specific methylation patterns toward pan-tissue averages, is proposed as a driver of aging. We demonstrate that liver-unique methylation sites regress during liver disease and normal aging, and through Mendelian randomization, causally link these changes to aging. Remarkably, repeated fasting partially restores DNA accessibility at these loci, in a cell-autonomous manner. These sites are enriched for metabolic transcription factor binding regions and overlap with liver disease risk variants, suggesting a mechanistic basis. Our findings establish epigenetic information loss as both a marker and mediator of liver aging while demonstrating its potential reversibility through metabolic interventions.

Introduction

Methylation patterns

DNA methylation is a fundamental epigenetic mechanism that regulates gene expression and chromatin organization in mammals (1). It primarily occurs at CpG dinucleotides. While hypermethylation is generally associated with transcriptional repression and hypomethylation is associated with transcriptional activity, the effects on gene expression are dependent on genomic context. Promoter hypermethylation is generally associated with transcriptional repression, often through recruitment of methyl-binding proteins that condense chromatin into inactive heterochromatin (2). In contrast, hypermethylation within gene bodies has been linked to active transcription in proliferating cells (1). CpG islands, particularly those in housekeeping gene promoters, are typically unmethylated and support stable gene expression (2). Meanwhile, enhancers exhibit heterogeneous methylation sensitivity, with most remaining active regardless of methylation status, while a subset is epigenetically regulated through inhibition of transcription factor binding by DNA methylation (3). Such changes in methylation correlate with DNA accessibility and transcriptional activity (4, 5). Together, these methylation patterns help establish and maintain cell-type-specific identity, and their disruption has been implicated in aging and disease (2).

Liver diseases and aging

The liver is a vital organ with remarkable regenerative capacity, yet it remains susceptible to age-related deterioration and disease. With advancing age, liver function generally declines, and aging is associated with increased severity and poor prognosis of multiple liver diseases (6–8). This decline manifests through various molecular and cellular changes, including alterations in hepatic blood flow, immune responses, and regenerative potential (9). Recent evidence suggests that the loss of epigenetic information—measured here via the regression of tissue-specific methylation patterns toward the pan-tissue average—accompanies aging and some (10) diseases (10). However, whether this loss is reversible, and to what extent it causally contributes to liver aging and dysfunction, remains unknown. Understanding the reversibility and causality of epigenetic information loss could open new therapeutic avenues for age-related liver conditions.

Epigenetic clocks and liver aging

DNA methylation patterns at specific loci change predictably with age, enabling the development of epigenetic clocks that measure chronological and biological age. These clocks, originally developed using blood samples, typically combine methylation levels at multiple CpG sites using elastic net regression or similar machine learning approaches (11–13). Recent advances using Mendelian randomization have identified specific methylation sites that causally contribute to aging, distinguishing between detrimental (DamAge), and adaptive (AdaptAge) age-related changes (14). Epigenetic clocks spanning multiple tissues, and even pan-tissue and pan-species have been developed (15, 16). In mouse models, liver-specific methylation clocks have been constructed (17), though their relationship with liver function remains unclear. Understanding how these aging-associated methylation changes interact with tissue-specific epigenetic patterns could provide insights into the mechanisms of liver aging.

Tissue-specific methylation and epigenetic information loss

The epigenetic theory of aging suggests that the erosion of methylation patterns contributes to aging (18–20). Recent evidence indicates that this loss of epigenetic information affects tissue function during aging and pathological conditions (10, 21–23). Restoration of a youthful epigenetic state was demonstrated using exogenous factors (22, 24, 25). Multiple studies have identified tissue-specific and cell-type-specific methylation patterns using combinations of various experimental and computational approaches (26). Here, we focus on a specific subset of methylation sites identified through a comparative tissue analysis approach (10). Due to differences in selection criteria, statistical thresholds, and reference tissue sets, these sites differ substantially from previously reported tissue and cell specific methylation sites. To avoid confusion with existing nomenclature, we refer to these as tissue-unique methylation sites. These sites show two distinct patterns: unique-low (UL) sites have lower methylation values in a specific tissue compared to other tissues, while unique-high (UH) sites show higher tissue-specific methylation. We note these sites are defined relative to values of the same loci in other tissues. In the liver, these sites do not map well to hypo and hyper methylated sites, and overlap in ranges (10). During aging and disease, in multiple tissues, these sites regress toward the average methylation levels found across tissues (10). This phenomenon has been observed across multiple organs and pathologies, including kidney disease (23). In the liver, these changes correlate with decreased organ function, though the causal relationship remains unexplored (10). Curiously, in other tissues and cell types, the exact opposite phenomenon is observed, as these sites diverge from the mean upon aging and disease (10).

Dietary interventions and epigenetic changes

Several dietary interventions, such as dietary/caloric restriction (CR) and intermittent fasting (IF), have shown promise in protecting against and ameliorating liver diseases, including inflammation and fibrosis. In both humans and animal models, these interventions have demonstrated the ability to delay or even prevent liver deterioration (27). Notably, CR and IF increase lifespan in multiple animal models (28, 29). These dietary interventions have been shown to influence DNA methylation, histone acetylation, and chromatin remodeling, potentially delaying age-associated epigenetic changes (30–32). However, the exact relationship between these interventions and tissue-specific epigenetic modifications remains unclear (33).

Study aims and findings

To explore the reversibility and causality of epigenetic information loss, we investigated tissue-unique methylation sites, where information loss can be easily observed (Fig. 1). Our analysis revealed that liver-unique methylation sites are enriched for age-correlated sites and are found near disease-associated single nucleotide polymorphisms (SNPs). Furthermore, these sites show significant overlap with detrimental age-related methylation sites identified through Mendelian randomization. Importantly, we demonstrate that metabolic interventions in mice can partially reverse these changes, at least in terms of DNA accessibility, both in whole livers and isolated hepatocytes, suggesting epigenetic information loss is at least partially reversible in the liver.

Figure 1.

Reversible information loss at liver-unique methylation sites. Liver-unique sites, showing higher (UH) or lower (UL) methylation levels, regress to the pan-tissue average upon aging and disease. UH sites are enriched for methylation sites causal to the aging process, while UL sites are enriched for liver-specific enhancers and PPAR-α binding sites. Upon repeated fasting, both UL and UH diverge away from the pan-tissue average, partially restoring the more youthful and disease-free epigenetic state.

Results

Cirrhotic liver methylation patterns are enriched for liver-unique methylation sites

To validate the relationship between tissue-specific methylation patterns and liver disease, we analyzed methylation changes in a public, independent dataset comprised of cirrhotic (N = 15), hepatocellular carcinoma (HCC; N = 12), and control (N = 12) human livers (34), analyzed using the HumanMethylation27 BeadChip (>27 K, ∼90% covered by HumanMethylation450). Of 1,750 differentially methylated sites identified, 1,615 were specific to HCC samples. Of these HCC-specific sites, only a small fraction (42 sites; <3%) overlapped with liver-unique methylation sites. In contrast, among the 120 sites showing differential methylation in cirrhosis, in a unified analysis of UH and UL methylation sites, we observed a striking enrichment of liver-unique methylation sites (53 sites; 44%, P = 0.0001). To avoid misinterpreting a global shift in methylation as regression to the mean, we also analyzed UL and UH sites separately, requiring both UH and UL sites to regress to the mean. This enrichment was independent of HCC status and showed a clear directional pattern: hypomethylated sites in cirrhosis overlapped exclusively with UH sites, while hypermethylated sites overlapped exclusively with UL sites (Fig. 2A; Supplementary Datasets S1 and S2). These findings demonstrate that liver cirrhosis is associated with regression of liver-unique methylation patterns toward the mean methylation levels observed across tissues, consistent with disease-associated epigenetic information loss.

Figure 2.

Liver-unique methylation sites methylation patterns regress to the mean in A) cirrhosis and B) NAFLD. A) Overlap of methylation sites hypomethylated (left) and hypermethylated (right) in cirrhosis with UH, UL, and control sites reveals a significant overlap of hypomethylated with UH, and hypermethylated with UL, suggesting a regression to the mean. A single control site overlapped with the hypomethylated sites. Monte Carlo simulation demonstrates that this observed enrichment of liver-unique methylation sites is statistically significant (P = 0.0001). B) Epigenetic information loss (see Methods) increases progressively with fibrosis grade, consistent with a loss of liver-specific methylation identity. Each dot represents a patient. Pairwise t test comparisons demonstrate significant stepwise differences (0 vs. 3: P = 8.5 × 10⁻¹⁸, d = 1.90; 3 vs. 3.5: P = 0.0071, d = 0.56; 3.5 vs. 4: P = 0.0010, d = 0.89).

To explore the functional implications of this distinction, we analyzed genes linked to each CpG list using the STRING database (Supplementary Dataset S3). Genes near liver-unique CpGs (N = 1,611) were significantly enriched in metabolic pathways [false discovery rate (FDR) < 10⁻⁶] and complement and coagulation cascades (FDR < 10⁻¹⁰). In contrast, genes near liver age-associated CpGs (N = 5,708; randomly subsampled for gene number parity for comparison) were enriched in neuronal system pathways (FDR < 10⁻⁴) and potassium channel-related processes (FDR = 0.0014), suggesting distinct biological functions. These results support the idea that liver-unique methylation patterns define core liver identity, while age-associated sites may reflect broader systemic aging programs.

Erosion of liver methylation patterns correlates with NAFLD disease severity

To test the connection between tissue-specific methylation patterns and liver disease severity, we analyzed methylation profiles from a cohort of nonalcoholic fatty liver disease (NAFLD) patients (N = 341) in relation to clinical features including body mass index (BMI), age, type 2 diabetes, and fibrosis grade (35). We computed epigenetic information loss for each patient relative to controls (fibrosis = 0; see Materials and methods). This loss showed a strong and statistically significant correlation with fibrosis grade (Spearman's ρ = 0.70, ρ² = 0.49, P = 1.68 × 10⁻⁵¹; Pearson's r = 0.74, R² = 0.55, P = 4.4 × 10⁻⁶⁰). In contrast, other clinical traits such as BMI, age, and type 2 diabetes exhibited much weaker associations (Figs. 2B and S1; Table S1). These findings suggest that the erosion of liver-specific methylation identity is tightly linked to disease severity, as reflected by fibrosis progression.

Degree of erosion of liver methylation patterns

To evaluate the degree to which disease-associated epigenetic changes affect the distinctiveness of liver-specific methylation patterns, we computed a regression score for each liver-unique CpG site for a cohort of liver diseases with age and sex matched controls ((36); see Materials and methods). This score quantifies how far the methylation level in diseased liver shifts toward the distribution observed across other tissues, where 0 indicates no change from healthy liver and 1 indicates full regression to the 5-percentile of the pan-tissue reference. Consistent with our previous findings, 99.6% of liver-unique sites exhibited a shift toward the pan-tissue average in disease (10) (Figure S2). However, the mean regression score across all sites was modest (mean = 0.29, SD = 0.22), indicating that while regression is widespread, liver retains a distinct epigenetic identity at most sites. Only 0.5% of sites (18 CpGs) fully regressed to the adjacent 5-percentile of pan-tissue methylation levels. These results suggest that even in disease, these sites maintain a substantial degree of tissue-specific methylation.

Disease-associated genes are enriched for liver-unique methylation sites

To investigate potential causal relationships between tissue-specific epigenetic regulation and liver disease, we examined the spatial distribution of liver-unique methylation sites relative to genes containing established genetic risk variants for liver diseases (37). Among 17 genes previously identified as containing causal variants for various liver pathologies, we found significant enrichment of liver-unique methylation sites within 20 Kb of these loci. Specifically, 7 genes harbored a total of 13 liver-unique methylation sites (Fig. 3A). These genes encode proteins central to liver metabolism: PNPLA3 (hepatic stellate cell lipase), MTTP (microsomal triglyceride transfer protein), VLDL (triglyceride loading), ADH1B (alcohol metabolism), GCKR (glucokinase regulatory protein), MOSC1 (amidoxime reduction), Apo-E (fat metabolism), and GPAM (glycerolipid synthesis). The enrichment was highly significant compared to random control regions (P = 0.0001) and decreased with distance from the genetic variants (Supplementary Dataset S4).

Figure 3.

Liver-unique methylation sites are enriched in proximity to pathogenic SNPs. A) Distribution of liver-unique methylation sites (black circle) relative to risk factor SNPs for liver diseases (red line). B) Genome-browser view of the six UH methylation sites (red, indicated by an arrow), which overlap the TSS and first exon of GPAM, as well as H3K27ac in multiple cell lines. However, liver-derived HepG2 cells do not show H3K27ac at that location. C) Genome-browser view of the UL site (blue, indicated by an arrow) adjacent to the TSS of MTTP, as well as enhancer-associated chromatin marks including H3K27ac and H3K4me1. The other seven cell lines from ENCODE do not show significant H3K27ac levels at this location.

To understand the functional implications of these liver-unique methylation sites, we examined their relationship with gene expression and chromatin states. The presence of liver-unique methylation sites did not necessarily correspond to liver-specific expression. For instance, GPAM, despite having 6 UH sites near its transcription start site (TSS), is expressed in multiple tissues including the liver. Interestingly, while its TSS shows H3K27ac enrichment in several cell lines, this mark is notably absent in liver-derived HepG2 cells (Fig. 3B). Conversely, MTTP, which is expressed predominantly in liver and gastrointestinal tissue, contains a single UL site near its TSS that coincides with active chromatin marks (H3K27ac, H3K9ac, and H3K4me1) specifically in HepG2 cells but not in nonliver cell lines (Fig. 3C). These results suggest that causal genetic variants for liver disease often present near regions with liver-unique methylation patterns, though the mechanistic basis of this association requires further investigation.

Characterization of liver-unique methylation sites in mice

Mouse experiments offer better control over genetic and environmental effects, as well as sample access. To expand our analysis, we generated a dataset of liver-unique methylation sites in 92 mice, using methods similar to Refs. (10, 26), based on public Infinium Mouse Methylation BeadChip (MM285; >285k) data from GSE184410 (38, 39) (Figure S3). We identified 6,038 liver-unique methylation sites, of these 38% were UH and 62% were UL (Supplementary Dataset S5). Sequence analysis using Hypergeometric Optimization of Motif EnRichment (HOMER) revealed distinct transcription factor motif enrichment patterns. UL sites were strongly enriched for liver function regulators, including PPAR-α (P = 10⁻⁸³), a key metabolic regulator, and C/EBPβ (CCAAT/enhancer-binding protein beta) (P = 10⁻⁵⁴), which regulates liver regeneration (Supplementary Dataset S6). Analysis using published ChIP-seq data (40–47) validated these findings. The liver transcription factors (TFs) FoxA1, C/EBPβ, and HNF4α, as well as the fasting modulator PPAR-α, all showed enrichment in UL, but not in UH (Figs. 4A–D and S4A–D). These patterns parallel previous findings in human liver and hepatocytes (10, 26, 48, 49). UH sites showed enrichment for Krüppel-like factors (KLF) transcription factor family motifs (P = 10⁻¹⁷), proteins crucial for liver physiology and pathology (50–54) (Supplementary Dataset S7). Although some KLF family members regulate hepatic metabolic pathways in response to specific stimuli, many exhibit low basal expression in adult liver and are more highly expressed during fetal development or in other tissues. Consistent with this, the KLFs enriched at UH sites (KLF1, KLF3, KLF4, KLF5, KLF9, KLF14, and KLF16) predominantly function as transcriptional repressors and are primarily expressed in nonhepatic tissues (55, 56).

Figure 4.

UL sites are bound by liver TFs and enhancer histone modifications. ChIP-seq analysis comparing the reads per site of UL, UH, and controls for HNF4α (A), FoxA1 (B), C/EBPβ (C), and PPAR-α (D). ****P-value < 0.0001. H3K4me1 (E) and H3K27ac (F) ChIP-seq signal at a 5 kb window relative to UL, UH, and control site.

To enable a comprehensive chromatin-state analysis, we generated an additional list of mouse liver-unique methylation sites (Supplementary Dataset S8), using the smaller mammalian methylation array (15, 57). Chromatin-state analysis using eForge 2 (58) identified that like in humans (10), UL sites showed strong enrichment for liver-specific enhancers (Supplementary Dataset S9). UH sites were enriched for both enhancers and areas flanking active TSS, but with significantly lower P-values and with no tissue specificity (Supplementary Dataset S10). Indeed, ChIP-seq analysis of liver samples identified UL sites as enriched for the enhancer histone modifications H3K27ac and H3K4me1 (Fig. 4E and F). By contrast, UH sites in liver showed no enrichment in H3K27ac, and a weak enrichment over background for H3K4me1 (Fig. 4E and F). Importantly, analysis of ChIP-seq data from other tissues revealed the inverse pattern: in nonliver tissues, the regions corresponding to liver UH sites were enriched for both H3K27ac and H3K4me1, while the regions corresponding to liver UL sites showed no enrichment (Figure S4E–H). Thus we concluded that tissue-specific UL methylation sites significantly overlap with tissue-specific enhancers.

Liver-unique methylation sites are enriched for aging-causal sites

To investigate whether liver-unique sites may causally contribute to aging, we analyzed their overlap with CpG sites previously identified as causal for aging-related traits through epigenome-wide Mendelian randomization analysis (14). Specifically, we tested for enrichment of our sites among the full set of ∼20,000 causal CpGs reported in that study. UH sites were significantly enriched among damaging CpGs—those associated with accelerated biological aging (e.g. Aging-GIP1: P = 3 × 10⁻⁴; Longevity-90th: P = 7.9 × 10⁻⁴). In contrast, UL sites showed no enrichment and were slightly depleted among damaging CpGs (0 sites found; Longevity-90th: P = 0.044; Fig. 5A). Testing the directionality of the change, in the damaging category, in Aging-GIP1, 85% of sites have negative beta values (P = 0.006), while Longevity-90th had 70% negative beta values, but did not reach statistical significance (P = 0.17), possibly due to the low sample size (N = 10).

Figure 5.

UH sites are enriched for causal-detrimental methylation sites. A) Enrichment and depletion of UH and UL liver sites over expected values in Human causal CpGs sites. B) Mouse liver-unique Methylation site age correlation plotted against site “uniqueness” (see Methods) demonstrated that the majority of UH sites showed a negative correlation, while UL sites showed a positive one, indicating regression to the mean upon aging. C) Mouse liver-unique methylation sites are enriched in the mouse multitissue and liver clocks, but not in the blood clock.

We also examined overlap with the DamAge and AdaptAge clocks, which represent machine-learning-selected subsets of causal CpGs designed to predict detrimental and adaptive aging trajectories, respectively. Consistent with our findings from the full causal set, UH sites were strongly overrepresented in DamAge (P = 2 × 10⁻⁵), and all overlapping CpGs carried positive weights (P = 0.01), indicating their contribution to age acceleration. In contrast, AdaptAge showed only marginal enrichment (P = 0.046), and UL sites were not enriched in either clock and were depleted in DamAge (0 sites found; P = 0.009; Figure S5).

Genes associated with liver-unique methylation sites in mice and human show limited overlap

We note that UL and UH should not simply be interpreted as expressed and repressed genes, respectively. The role of methylation in gene regulation is context dependent (59, 60). For example, both low methylation levels adjacent to the TSS and high methylation levels along the gene body positively correlate with expression level. This pattern is observed in the human coagulation factor XII (F12), predominantly expressed in the liver, where UL sites are found in the TSS, while UH sites are scattered across the gene body and later exons (Figure S6).

We compared genes associated with liver-specific methylation sites in mice (N = 2177) and humans (N = 1611), observing a 17% match in mice and a 23% match in humans (Figure S7). While statistically significant (P = 10⁻⁵), this overlap appears modest. We thus tested if a significant overlap exists on the pathway level using the STRING database (61). This analysis identified several common pathways enriched in both species. Notably, both human and mouse analyses showed enrichment in metabolic pathways (human: FDR < 10⁻⁶, mouse: FDR < 10⁻⁷) and complement and coagulation (human: FDR < 10⁻¹⁰, mouse: FDR = 0.00026). These results can reflect on processes conserved across species particularly in the context of liver sole function in several metabolic processes and coagulation cascade regulation.

Liver-unique methylation sites regress to the mean upon aging in mice

To test for potential overlap between age-associated changes in methylation and liver-unique methylation sites, we plotted the site “uniqueness” (see Materials and methods) vs. correlation to aging using the mouse GSE184410 dataset. UH sites have an average negative correlation to aging, while UL sites have an average positive correlation (Fig. 5B). Among UH sites, 92.6% show negative correlation with age and only 7.4% positive. Conversely, 84.3% of UL sites show positive correlation with age, and only 15.7% negative (Fig. 5B). Overall, 87.4% of liver-unique sites regressed to the mean upon aging (P = 10⁻⁸²⁶). Interestingly, within each UH and UL group of CpGs, there was no obvious relationship between site uniqueness and age correlation. This is in contrast to a more quantitative relation observed for example in kidney-unique sites in chronic kidney disease (CKD) (62).

Liver-unique methylation sites are enriched in epigenetic clocks in mice

To explore the relationship between epigenetic information loss and aging, we examined whether liver-unique methylation sites are enriched in established mouse DNA methylation clocks (17), which accurately measure chronological age and may capture aspects of biological aging. Liver-unique methylation sites demonstrated no enrichment in a mouse blood clock ((63); P = 0.25; Fig. 5C), and only a modest, nonsignificant 1.7-fold enrichment in a multitissue clock ((17, 64); P = 0.1; Fig. 5C). However, a mouse liver-specific clock showed a 5.3-fold enrichment ((65); P = 0.0001; Fig. 5C), suggesting that liver-unique methylation sites are relevant to aging in a tissue-specific manner (Supplementary Dataset S11). Interestingly, both UH and UL sites were represented in these clocks at ratios proportional to their overall prevalence in the liver methylome, suggesting that the clock draws broadly from liver-unique features rather than preferentially selecting one subtype. These findings support the idea that epigenetic aging in the liver involves erosion of tissue-specific methylation patterns.

Repeated fasting partially restores chromatin state in mice livers

DNA methylation patterns can change due to environmental effects. Moreover, both beneficial behaviors and recovery from disease can result in temporal “reversion” of at least some epigenetic clocks (17, 66–69). However, while the reversibility of epigenetic information loss was established with exogenous interventions like treatment with “Yamanaka Factors,” its reversibility under natural settings has been underexplored. We note that this is distinguished from exploring the reversibility of various epigenetic clocks that have been studied more carefully (67). DNA accessibility often correlates to methylation levels in enhancers (70–72). This relationship was shown to be causal in some cases, but not in others (70). To investigate the potential reversibility of epigenetic information loss, we examined liver DNA accessibility in mice subjected to alternate-day fasting (ADF). ADF, a form of IF, reduces liver pathologies and enhances enhancer sensitivity (73–75). We analyzed liver ATAC-seq data (GSE212776) from 24 6-week-old mice (12 ADF, 12 control), generated following 30 days of ADF, using the Illumina NextSeq 500/550 platform with the High Output v2 kit (75 cycles) to assess DNA accessibility ((73); Supplementary Dataset S12, see Materials and methods). Enhancers showing increased accessibility were 8-fold enriched in liver-unique sites vs. control sites (Fig. 6A; Supplementary Dataset S12). This effect was exclusively driven by UL sites that showed a 12-fold enrichment vs. controls and a 44-fold enrichment vs. UH (Fig. 6B; Supplementary Dataset S12). Enhancers showing decreased accessibility were 4-fold enriched in liver-unique sites vs. controls (Fig. 6A), however, this effect was driven by UH sites that were 7-fold enriched vs. controls and 3-fold enriched vs. UL (Fig. 6C). This phenomenon was not observed when comparing livers of mice subjected to a single 24-h fast (Fig. 6A–C), nor was it observed in UL sites, which were slightly decreased in ADF as compared to single-fasted mice (Fig. 6C). We performed functional analysis of the unique (predominantly UL) methylation sites showing increased accessibility using Genomic Regions Enrichment of Annotations Tool (76). The two strongest associated mouse phenotypes were (i) abnormal liver physiology and (ii) abnormal hepatobiliary system physiology. Similarly, GO terms related to metabolic processes were highly enriched (Supplementary Dataset S13). These findings demonstrate that liver epigenetic information lost can be partially restored, at least in terms of DNA accessibility, through repeated fasting (Fig. 6D), which results in increased accessibility specifically in functionally relevant UL sites and decreased accessibility specifically in UH sites, suggesting a potential therapeutic approach for reversing tissue-specific epigenetic deterioration.

Figure 6.

Reversibility of epigenetic information loss in the liver of fasting mice. A) Overlap of liver-unique methylation sites and random controls with enhancers that show increased (left) or decreased (right) accessibility following ADF or a single fast, shows a greater overlap upon ADF. B) The overlap in increased accessibility enhancers is driven exclusively by UL sites in ADF. C) The overlap in decreased accessibility enhancers is mostly driven by UH sites in ADF, but some UL sites change following ADF and a single fast. D) Divergence from the mean and regression to the mean in liver unique sites following ADF or a single fast. E) Overlap of liver-unique methylation sites with corticosterone-activated and dual-activated enhancers.

Epigenetic information is restored in mouse hepatocytes upon fasting mimicry

While bulk-tissue analysis reveals epigenetic information loss and restoration, it does not inform us if these changes are driven by changes in cell composition, changes within cells, or both. To address this limitation, we leveraged an ex vivo system employed by Goldberg et al. (77) that mimics long-term fasting in isolated hepatocytes through agonist treatment—corticosterone to activate the glucocorticoid receptor and Wy-14643 to activate PPARα. Analysis of liver-unique methylation sites revealed a striking 10-fold enrichment in corticosterone-activated enhancers compared to controls (143 enhancers vs. 14 in random controls; Fig. 6E; Supplementary Dataset S14). Remarkably, all identified enhancers (143/143) mapped to UL sites (P < 10⁻²⁹). These findings were further validated using dual enhancers identified through combined activation with Wy-14643 and corticosterone, where an 8-fold enrichment was observed (211 vs. 26; Fig. 6E; Supplementary Dataset S14). Again, nearly all identified enhancers (210/211; P < 10⁻⁴⁰) mapped to UL sites. Together, these results demonstrate that DNA accessibility loss and restoration occurs at the cellular level within hepatocytes.

Discussion

Our findings establish the reversibility of epigenetic information loss in liver aging and disease, while demonstrating its causal role in these processes. The regression of liver-unique methylation patterns toward pan-tissue averages during cirrhosis validates and extends previous observations of epigenetic information loss in liver pathologies. Notably, this regression occurs independently of canonical age-associated drift, indicating that cirrhosis-associated methylation changes affect a distinct set of sites with tissue-specific regulatory functions. To extend this observation beyond cirrhosis, we analyzed methylation profiles from NAFLD patients and found that the degree of information loss at liver-unique sites varied across individuals and strongly correlated with fibrosis severity. These findings suggest that regression of liver-specific methylation identity is not merely a consequence of aging or metabolic status but is tightly linked to liver disease progression itself. Importantly, the strong enrichment of liver-unique methylation sites near disease-associated genetic variants suggests a mechanistic link between genetic predisposition to liver disease and tissue-specific epigenetic regulation.

The relationship between methylation changes and DNA accessibility in our study suggests a mechanism underlying epigenetic information loss. Our results demonstrate that liver-unique methylation sites exhibit directional changes during aging and disease: UH sites become hypomethylated while UL sites become hypermethylated, both regressing toward pan-tissue averages. This pattern aligns with the established inverse correlation between DNA methylation and chromatin accessibility at enhancers, where hypomethylation typically promotes increased accessibility and hypermethylation restricts it. However, our fasting intervention data, which measured DNA accessibility as a functional readout, reveal that this relationship operates bidirectionally in the restoration process. Upon repeated fasting, UL sites—which are hypermethylated in disease and aging—show increased DNA accessibility, suggesting potential demethylation and restoration of their enhancer function. Conversely, UH sites show decreased accessibility during fasting, consistent with potential remethylation toward their healthy hypermethylated state. While direct methylation measurements in the fasting studies would strengthen these conclusions and are currently being pursued, the observed accessibility changes provide strong functional evidence for epigenetic restoration. Another caveat is that information loss reversibility was demonstrated in young and healthy animals. The benefits of dietary interventions have been demonstrated on the liver across different ages and pathologies, in both mice and humans (74, 78–80). While focusing on young healthy mice, prior to the onset of fibrosis or metabolic drift, makes the findings more robust, future experiments will address this phenomena in old and obese animals, testing the efficacy of metabolic interventions to restore epigenetic information in more pathologically relevant models. This bidirectional restoration is particularly significant given that UL sites are enriched for liver-specific enhancers and PPAR-α binding sites, key regulators of metabolic function that are activated during fasting. The cell-autonomous nature of these changes, demonstrated in isolated hepatocytes under fasting-mimicking conditions, indicates that the methylation-accessibility relationship at liver-unique sites represents an intrinsic cellular response mechanism rather than tissue-level compositional changes. These findings extend previous observations that DNA accessibility correlates with methylation patterns at enhancers by demonstrating that this relationship can be therapeutically modulated through metabolic interventions, offering a potential pathway for reversing disease-associated epigenetic deterioration.

The distinct behavior of UH and UL sites in aging presents a nuanced picture of epigenetic information loss. Both site classes show age-correlated methylation changes and epigenetic information loss, revealed through correlation analysis of methylation patterns in mouse liver samples. This association was further strengthened by Mendelian randomization analysis in humans, which revealed significant enrichment of UH sites among CpGs identified as causal for damaging, age-accelerating traits. In contrast, UL sites were not enriched and were slightly depleted in the same set. These findings were further supported by analysis of the DamAge and AdaptAge clocks, which represent machine-learning-based aging models derived from causal CpGs. UH sites were overrepresented in DamAge with modest enrichment in AdaptAge, which tracks adaptive age-related changes. Conversely, UL sites are depleted in DamAge and showed no enrichment in AdaptAge, despite their clear susceptibility to disease-associated changes, as demonstrated in our cirrhosis analysis. Together, these findings suggest a model where UH sites actively participate in the aging process, possibly due to their importance in other tissues, while UL sites primarily respond to disease states.

The conservation of pathway-level associations between mouse and human liver-unique methylation sites, despite limited gene-level overlap, points to the fundamental role of these epigenetic patterns in liver function. The enrichment of metabolic and coagulation pathways in both species aligns with core liver functions, while species-specific enrichments may reflect a lower level of conservation, evolutionary adaptations, or experimental constraints.

Perhaps most significantly, our demonstration that repeated fasting can partially restore liver-unique methylation patterns provides evidence for the reversibility of epigenetic information loss through physiological interventions. The cell-autonomous nature of this restoration, as shown in isolated hepatocytes under fasting-mimicking conditions, indicates that these changes occur at the cellular level rather than through alterations in tissue composition. This finding has important implications for understanding the nature and reversibility of disease- and age-related epigenetic changes.

Interestingly, we did not observe an age-associated trend in the human liver data (Figure S1C), in contrast to the regression toward the pan-tissue average seen in mouse liver. This discrepancy may reflect the higher variability in the human NAFLD cohort, including differences in diet, medication, comorbidities, and environmental exposures, which could obscure age-related effects. In contrast, the mouse data derive from healthy animals of uniform genetic background and controlled conditions, enabling clearer detection of aging-associated changes. Alternatively, epigenetic information loss in the liver may proceed differently between species or on different timescales. Although we examined available healthy human liver data, limited sample size and age range precluded definitive conclusions. Future studies in larger, well-characterized human cohorts will be essential to resolve this question.

Several limitations of this study should be considered. While our findings demonstrate the reversibility of epigenetic information loss in mouse models, the extent to which these results translate to humans remains to be determined. Our analysis of human samples was limited to cross-sectional data, preventing direct assessment of temporal dynamics and individual variation in epigenetic changes. The cell-autonomous effects observed in mouse isolated hepatocytes, while promising, were studied under artificial fasting-mimicking conditions that may not fully recapitulate the complexity of in vivo fasting responses. Additionally, while we show enrichment of liver-unique methylation sites near disease-associated variants, the precise molecular mechanisms linking these epigenetic patterns to disease progression remain unclear. One significant limitation is our intermittent use of DNA methylation and accessibility data. Although differential DNA accessibility provides both a functional readout and a strong indicator of methylation profile changes, direct measurement of DNA methylation levels in the fasting intervention studies would strengthen our conclusions. Finally, while we demonstrate partial restoration of epigenetic information through fasting, the long-term stability of these changes and their impact on liver function and disease outcomes requires further investigation in longitudinal studies.

These results advance our understanding of epigenetic information loss as both a marker and mediator of aging and disease, while offering promising directions for therapeutic interventions. Future studies should investigate the molecular mechanisms linking metabolic interventions to epigenetic restoration and explore the potential of targeted approaches to reverse tissue-specific epigenetic deterioration in aging and disease.

Materials and methods

Identification of mouse liver-unique sites

Liver-unique methylation sites in mice were generated from healthy tissues using methods similar to Ref. (10), based on public data from GSE184410, generated by Infinium MouseMethylation BeadChips (MM285, >285k) (38, 39). 92 Mus musculus liver samples were selected, along with 404 M. musculus samples from 19 other tissues as the background (Figure S4A). A UH methylation site in the liver is defined as a site where the average methylation value is higher by 0.1 or more than the 0.95 quantile of all other tissues, and a UL methylation site is one where the average methylation value is lower by 0.1 or more than the 0.05 quantile of all other tissues. In addition to this binary classification, for each such site, we also calculate a “site uniqueness” for each site, defined as the absolute distance between the average liver methylation and the corresponding quantile threshold (0.95 or 0.05) across other tissues (10). Similarly, we extracted mouse liver unique sites from 1,195 M. musculus liver samples along with 2,816 samples from 34 other tissues found in GSE223748 (Figure S4B) (15), generated by the Mammalian Methylation Consortium using the Mammal40k array (57).

CpG sites correlation with age

Pearson correlation between age and methylation levels was plotted for all mouse liver-unique CpG sites. Data were sourced from the GSE184410 database, including all liver samples with no pathologies.

Statistics

Unless otherwise noted, P-values were calculated using log-scale binomial or hypergeometric distribution on R. In the cirrhosis overlaps (Fig. 2), pathogenic SNP analysis (Fig. 3), causal sites enrichment and mouse epigenetic clocks (Fig. 5), the P-values were determined using a Monte Carlo simulation with 10,000 iterations. Control groups of the same size as the liver-unique sites (N = 3,674 for human; N = 6,039 for mouse) were randomly sampled from tissue-unique sites excluding liver-unique ones [N = 84,249 for human; N = 25,128 for mouse (10)]. When the observed intersection of liver-unique sites exceeded all values from the simulated distribution, the P-value was calculated as 1/number of simulations, yielding a P-value of 0.0001.

Enrichment analysis

Enrichment analysis for cirrhosis, DamAge, AdaptAge, ADF and hepatocytes was done using randomly selected controls as described above, requiring either an identical CpG site (methylation arrays) or an overlap between the sites and range (DNA accessibility). For pathogenic SNPs and mouse tissue clocks, sites considered overlapping if found below a distance threshold, found in Supplementary Dataset S4 for the SNPs, or below 2.5 kb for the clocks. Reducing this threshold did not change the enrichment of the liver clock but reduced the number of overlapping sites.

Epigenetic information loss analysis

Epigenetic information loss for NAFLD (35) was calculated using liver-specific CpG sites classified as high or low methylated based on their levels in healthy liver tissue compared to other tissues. Control means were computed from samples with fibrosis grade 0. For each patient, loss was defined as the average deviation from the control mean across sites: directly for low-methylated sites, and inverted (multiplied by −1) for high-methylated sites to reflect loss of methylation. Scores for high and low groups were averaged separately and summed to obtain a total loss score per sample.

Regression score

Regression scores were calculated based on the methylation shift in liver diseases (36) relative to healthy liver and a nonliver tissue reference (95th or 5th percentile), where M_healthy is the methylation level in healthy liver, M_disease is the average level in diseased liver, and M_reference represents the 95th or 5th percentile methylation level across all nonliver tissues. A score of 0 indicates no change while a score of 1 indicates full regression to the 95th or 5th percentile of the pan-tissue average.

$$\mathrm{Scor}{\mathrm{e}}_{\mathrm{UH}}=\frac{M_{\mathrm{healthy}}-M_{\mathrm{disease}}}{M_{\mathrm{healthy}}-M_{\mathrm{reference}(0.95)}}\mathrm{Scor}{\mathrm{e}}_{\mathrm{UL}}=\frac{M_{\mathrm{disease}}-M_{\mathrm{healthy}}}{M_{\mathrm{reference}(0.05)}-M_{\mathrm{healthy}}}.$$

Epigenetic state analysis

Chromatin state analysis was performed on liver-unique methylation sites in mice using the eFORGE (81) web server (https://eforge40k.altiusinstitute.org/). It was set to eFORGE 40 K (mammalian array version), “Consolidated Roadmap Epigenomics—Chromatin—All 15 state marks” and default setting of 1 kb window, 1,000 background repetition, 0.01 strict, and 0.05 marginal.

Motif enrichment analysis

Known and de novo motifs were identified and enrichment was calculated using the HOMER (82). Analysis was done separately for mouse UL and UH sites, analyzing sequences extracted from the mm10 genome assembly and spanning ±500 bp around the methylation probe start sites (83). The Infinium mouse methylation manifest file served as the background for motif enrichment analysis.

ChIP-seq analysis

Tag density of ChIP signal around methylated sites (±100 bp) were analyzed using the HOMER suite. In aggregate plots, the tag count (averaged across all sites) per site per base pair was calculated using the HOMER suite (annotatePeaks, option -size 8000 -hist 10). In box plots, tag count ±200 bp around the site (averaged across all sites) was calculated using the HOMER suite (annotatePeaks, option -size 400 -noann). In box plots for H3K27ac, which is a broader signal, tag count ±500 bp around the site (averaged across all sites) was calculated using the HOMER suite (annotatePeaks, option -size 1000 -noann). In both aggregate plots and box plots, the data are an average of all available replicates. In all box plots, the 10–90 percentiles are plotted.

ADF regimen and DNA accessibility

Detailed description of the feeding regimen is described in Ref. (73). Briefly, female C57BL6J mice, 6 weeks old, were randomly divided into an unrestricted feeding (URF) group and an ADF group, with 12 mice in each. Following a 1-week acclimation period, the 30-day experiment began. The URF group received ad libitum access to Teklad TD2018 food and water for the entire duration. The ADF regimen consisted of 15 cycles over 30 days: 24 h of ad libitum food and water followed by 24 h of water-only access. Food was removed at the start of the inactive phase (zeitgeber time 1) and returned 24 h later. On day 31, all mice underwent a final 24-h fasting period. For the terminal procedure, half the mice from each group (n = 6 URF, n = 6 ADF) were euthanized immediately at the end of this final fast. The remaining mice (n = 6 URF, n = 6 ADF) were refed for 24 h before euthanasia. A control group that was continuously fed ad libitum and never fasted was also included. All mice were anesthetized and euthanized using ketamine:xylazine (30:6 mg/mL) to allow for liver excision and plasma collection.

ATAC-seq was performed as previously described (84). Briefly, nuclei were isolated using a hypotonic buffer and Dounce homogenizer. Nuclei were tagmented using Tn5 transposase loaded with Illumina adapters. Tagmented DNA was PCR-amplified with sample-specific indices. The resulting library was size-selected to DNA fragments of 150–800 nt. The multiplex sample pool (1.6 pM including PhiX 1%) was loaded on NextSeq 500/550 High Output v2 kit (75 cycles) cartridge with paired-read sequencing conditions. Each sample was sequenced at a depth of at least 5 × 10⁷ reads.

Supplementary Material

Supplementary material is available at PNAS Nexus online.

Competing Interests

The authors declare no competing interests.

Funding

This work was supported by the Israel Science Foundation (ISF) (grants 654/20 and 632/20 to D.Z.B.), the Center for Artificial Intelligence & Data Science in Tel Aviv University (TAD to D.Z.B.) and the European Research Council (grant 947907 to I.G.).

Author Contributions

Roni B. Shtark (Conceptualization, Investigation, Writing—review & editing), Naor Sagy (Conceptualization, Investigation, Writing—review & editing), Noga Korenfeld (Investigation, Methodology), Maayan Gal (Writing—review & editing), Ido Goldstein (Investigation, Methodology, Supervision, Writing—review & editing), and Daniel Zvi Bar (Conceptualization, Funding acquisition, Supervision, Writing—original draft, Writing—review & editing)

Preprints

This manuscript was posted on a preprint: https://www.biorxiv.org/content/10.1101/2025.02.24.639802v1.

Data Availability

All experimental data, including methylation data and ATAC-seq data used came from previously published sources, identified and references in the text and Supplementary files. Generated data are available as Supplementary files.

Ethics Statement

Ethical approval by an Institutional Animal Care and Use Committee (IACUC) was not required for this study, as it comprises solely the analysis of publicly available data. Animal ethics approvals and permit numbers for the generation of the primary data analyzed herein were obtained by the original authors and are detailed in their respective publications.