Metformin use history and genome-wide DNA methylation profile: potential molecular mechanism for aging and longevity

Background: Metformin, a commonly prescribed anti-diabetic medication, has repeatedly been shown to hinder aging in pre-clinical models and to be associated with lower mortality for humans. It is, however, not well understood how metformin can potentially prolong lifespan from a biological standpoint. We hypothesized that metformin’s potential mechanism of action for longevity is through its epigenetic modifications. Methods: To test our hypothesis, we conducted a post-hoc analysis of available genome-wide DNA methylation (DNAm) data obtained from whole blood collected from inpatients with and without a history of metformin use. We assessed the methylation profile of 171 patients (first run) and only among 63 diabetic patients (second run) and compared the DNAm rates between metformin users and nonusers. Results: Enrichment analysis from the Kyoto Encyclopedia of Genes and Genome (KEGG) showed pathways relevant to metformin’s mechanism of action, such as longevity, AMPK, and inflammatory pathways. We also identified several pathways related to delirium whose risk factor is aging. Moreover, top hits from the Gene Ontology (GO) included HIF-1α pathways. However, no individual CpG site showed genome-wide statistical significance (p < 5E-08). Conclusion: This study may elucidate metformin’s potential role in longevity through epigenetic modifications and other possible mechanisms of action.


INTRODUCTION AGING
Considerable evidence from basic and pre-clinical models shows that several interventions, such as exercise, intermittent fasting, and even ingestion of certain compounds can prolong lifespan. These promising compounds include rapamycin [3,4], resveratrol [5][6][7], NAD [8], and metformin [9][10][11]. Our group also confirmed that inpatients using metformin had improved three-year survival rates compared to non-metformin users [12]. In addition, our data also showed that prevalence of delirium was lower among those who were on metformin compared to those without [12].
The mechanism (or mechanisms) of action that rationalizes how these interventions prolong lifespan, or potentially delay aging, has been investigated heavily. Nevertheless, no exact process is well understood, especially for metformin. It is believed that epigenetics is one of the most important molecular mechanisms of aging in animals and plants; thus, it is plausible that the "life-prolonging" effects of many interventions are through modification of epigenetic processes. For example, several reports show epigenetic changes from exercise [13], fasting [14], rapamycin [3], resveratrol [5], and NAD [8]. However, there are only a few studies investigating the direct influence of metformin on epigenetic changes [15][16][17], suggesting that information about the influence of metformin on the epigenetic profile in humans is currently limited.
To fill such gap of knowledge, we investigated the potential influence of metformin on the epigenetic profile by testing genome-wide DNA methylation (DNAm) in whole blood samples obtained from inpatients with and without a history of metformin use.

RESULTS
Demographics 173 subjects were enrolled in this study, but only 171 were included in downstream data analysis. The average patient age was 74.4 (SD = 9.8). 58 (33.9%) subjects were females while almost all the subjects were white per self-report (n = 167; 97.7%). 108 patients were nondiabetic (non-DM) while 63 were diabetic (DM). Among the DM group, 37 had diabetes with a history of metformin prescription DM(+)Met and 26 had diabetes without a history of metformin prescription DM(−)Met. Additionally, 43 (68.3%) diabetic subjects had a history of insulin use. Charlson Comorbidity Index (CCI) and body mass index (BMI) information are also included in Table 1. No variable revealed statistically significant differences between the DM(−)Met and DM(+)Met. However CCI, BMI, and insulin use were significantly higher among the DM group compared to the non-DM group, as expected. Table 2 shows the most significant genes that differed in methylation rates between patients with and without metformin use history regardless of diabetes status (171 subjects). None of the sites met the criteria for genomewide statistical significance (p < 5E-8).

Met vs. non-Met (including all patients regardless of diabetes status): top hits, KEGG, GO
Next, we conducted enrichment analysis using the top 330 CpG sites based on the absolute difference in methylation level (beta value) between metformin users (Met) and nonusers (non-Met) greater than 4% and the p-value less than 0.01. Enrichment analysis from the KEGG top signals showed relevant pathways to metformin's possible roles, such as "longevity regulating pathway", "longevity regulating pathwaymultiple species", and "AMPK signaling pathway" (Table 3). In addition, other pathways, such as "mTOR signaling pathway", "insulin secretion", "glutamatergic synapse", and "circadian entrainment" were discovered (Table 3). There were also relevant pathways revealed in the GO analysis, such as "regulation of hypoxiainducible factor-1alpha signaling pathway", "positive regulation of hypoxia-inducible factor-1alpha signaling pathway", and "canonical Wnt signal pathway" (Table 4), although none of the pathways in either KEGG or GO reached the False Discovery Rate (FDR) significance level (FDR <0.05) (Tables 3 and 4). Table 5 shows the most significant genes that differed in methylation rate between metformin users and nonusers among the diabetes group (63 subjects). Similar to the previous analysis, no gene reached genome-wide statistical significance (p < 5E-8).

Met vs. non-Met (including only patients with type 2 diabetes mellitus): top hits, KEGG, GO
The enrichment analysis was generated using consistent parameters in methylation level differences (beta >4%) and p-value (<0.01). This current analysis, however, included 1283 CpGs. KEGG showed many of the same signals discovered from the previous analysis, including "longevity regulating pathway", "glutamatergic synapse", "insulin secretion", "circadian entrainment", and "cholinergic synapse" (Table 6). GO also showed overlapping pathways compared to the first analysis, including "hypoxia-inducible factor-1alpha signaling pathway", but also new pathways, such as "interleukin-8-mediated signaling pathway", "negative regulation of leukocyte apoptotic process", "neutrophil homeostasis", and "neuron projection", although these pathways did       Table 7).

DNA methylation age acceleration
Among the diabetes group, metformin nonusers had a mean age acceleration of −8.07 compared to a mean age acceleration of −4.47 for metformin users (p = 0.11) (Figure 1). This difference was smaller among all the subjects included regardless of diabetes status (−5.92 for metformin nonusers vs. −4.47 for metformin users; p = 0.34) (Figure 2). Both analyses did not reach statistical significance.

DISCUSSION
In this study, we compared genome-wide DNA methylation rates among metformin users and nonusers to investigate the potential epigenetic effects of metformin exposure. Enrichment analysis was employed to elucidate the possible mechanisms of action induced by metformin. Our KEGG analysis revealed evidence of differences in epigenetic profiles involved in "longevity" such as "longevity regulating pathway" and "longevity regulating pathwaymultiple species" (Tables 3 and 6). Although it was not statistically significant, the appearance of these pathways among top signals in the KEGG analysis demonstrates the potential role of the epigenetic processes manifesting the effect of metformin on longevity. The same KEGG analysis also showed "AMPK signaling pathway" (Table 3). AMP-activated protein kinase (AMPK), an energy sensor that regulates metabolism, is commonly referred to as one of the targets of metformin's hypothetical mechanisms of action [18,19], although there is also evidence that metformin's effects are in part AMPK-independent [20]. Furthermore, AMPK activation is related to subsequent activation of hypoxia-inducible factors [21] which also appeared in our GO analyses as "regulation of hypoxia-inducible factor-1alpha signaling pathway" and "positive regulation of hypoxia-inducible factor-1alpha signaling pathway" (Table 4), as well as "hypoxia-inducible factor-1alpha signaling pathway" (Table 7). Hypoxia-inducible factor-1alpha (HIF-1α) is a transcription factor expressed in nucleated cells and mediated by oxygen levels. HIF-1α has been implicated in age-related diseases, endothelial senescence progression, AMPK, and many other pathways [22]. Beyond metformin's potential epigenetic medication AGING related to longevity, several pathways related to delirium, such as "circadian entrainment", "cholinergic synapse", and "glutamatergic synapse", were identified (Tables 3 and 6). These pathways are intriguing from metformin's possible "anti-aging" standpoint as age is a major risk factor of delirium.
The beneficial effects of metformin on lifespan have been widely studied. Previous studies reported that metformin increased median lifespan of C. elegans cocultured with E.coli by more than 35% [9,23], and prolonged the lifespan of mice [10]. Patients with agerelated diseases such as cardiovascular diseases and  cancer who take metformin also had lower rates of mortality [24,25]. Our recent study using a cohort of over 1,400 inpatients also revealed that diabetic patients with a history of metformin use have a significantly lower 3-year mortality than diabetic patients who have not taken metformin [12]. There are, however, conflicting reports as well. For example, the same effect was not observed in Drosophila [26]. Also, agedependent, dose-dependent, and gender-dependent variable effects on lifespan were reported in mice [27,28]. Although these previous studies' results are not consistent, our cohort mentioned above (from which the present data are an analysis of its subgroup) clearly showed a positive influence of metformin use on survival among diabetic inpatients [12].
Our epigenetics data presented herein support metformin's broad range of potential effects as indicated by the pathways identified through the enrichment analysis. The KEGG analysis (Table 7) showed several signals related to inflammation and the immune system, such as "interleukin-8 receptor activity" and "negative regulation of leukocyte apoptotic process." The appearance of inflammationrelated pathways is intriguing considering strong evidence showing that elderly people present with lowgrade, chronic inflammation [29]. These signals identified in our study may support our hypothesis that metformin can modify the inflammatory process through epigenetic modification and influence the likelihood of survival. Consistent with our data, Barath et al. also reported that metformin inhibited cytokine production from Th17 by correcting age-related changes in autophagy and mitochondrial bioenergetics, indicating its potential for the medication to promote healthy aging [30]. Among the literature supporting metformin's role in suppressing inflammation, clinical trials including the Diabetes Prevention Program (DPP) [31] and Bypass Angioplasty Revascularization Investigation 2 Diabetes (BARI 2D) [32] have provided further evidence of metformin's role in changing inflammatory biomarker levels among diabetic patients, while other clinical trials, such as the Lantus for Creactive Protein Reduction in Early Treatment of Type 2 Diabetes (LANCET) [33], have found opposing evidence. Although several studies mentioned here have investigated the relationship between metformin and its potential anti-inflammation, a clinical trial aimed to confirm metformin's role in aging is yet to be seen [2,34]. It is worth mentioning, nonetheless, a small clinical study that demonstrated the regression of epigenetic age of patients through the administration of recombinant human growth hormone (rhGH), dehydroepiandrosterone (DHEA), and metformin [15]. As the study team administered three medications to their subjects at the same time, it is impossible to distinguish epigenetic changes caused only by metformin. It is also worth mentioning the unexpected results from the Horvath epigenetic clock since subjects with history of metformin use had relatively higher age acceleration than subjects without history of metformin. Still, neither reached statistical significance (p < 0.05). Future prospective studies comparing epigenetics marks before and after metformin use would be needed to better understand the direct effect of the medication.
In DM-only subjects, A-kinase anchoring protein 13 (AKAP13) gene was found (Table 5). A recent study showed that AKAP13 inhibits mammalian target of rapamycin complex 1 (mTORC1), which was present in our enrichment analysis as "mTOR signaling pathway" (Table 3). Furthermore, the degree of AKAP13 expression in lung adenocarcinoma cell lines correlates with mTORC1 activity [35]. Metformin's antiinflammatory effect has been shown to occur through eventual AMPK activation, which also inhibits the mTOR signaling pathway [18]. Metformin's connection to AKAP13, which has yet been fully understood, deserves further investigation.
To the best of our knowledge, our study is the largest of its kind. A smaller, previous study also investigated metformin's effect on genome-wide DNA methylation in human peripheral blood, although their study power was limited to a sample size of 32 male subjects [36]. Enrichment analysis in the present study revealing the longevity pathway from a hypothesis-free approach further strengthens our hypothesis that metformin exhibits its potential benefit for longevity through epigenetic processes. We also identified other relevant pathways associated with metformin's mechanisms of action, such as the AMPK signaling pathway and HIF-1α signaling pathway [37].
Our study has several limitations. Although 171 subjects were analyzed retrospectively in this study, a controlled prospective study with a larger sample size would provide a better picture of the epigenetic mechanism of metformin on longevity. In addition, none of the individual CpG sites reached genome-wide significance (p < 5E-08). Thus, our findings should be interpreted as exploratory and hypothesis-generating. However, the fact that we found their biological relevance to metformin's roles is still worth noting. As diabetes and metformin use status of the subjects was determined based on a retrospective chart review of electronic medical records, there are possibilities for misclassification, although we were still able to find multiple relevant pathways and genes of interest related to metformin's action. Moreover, the duration of metformin use was not precisely assessed, making our definition of "metformin history use" broad since it AGING might have included patients who took metformin for only a few months and patients who took metformin for years, for instance. Also, other types of diabetic medications were not investigated, such as sulfonylureas and glinide drugs as we used an already completed study dataset from our previous work. The rationale for us not investigating the influence of other diabetic medications was based on past literature showing that those diabetic medications other than metformin did not show benefits for survival. In fact, sometimes they were associated with worse mortality [38][39][40].
In summary, the data presented here support our hypothesis that epigenetics, especially DNA methylation, may be altered by metformin use and that such epigenetic processes potentially contribute to molecular mechanisms leading to longevity. Further careful investigation with a larger sample size would be warranted.

Study participants and recruitment
We have previously recruited patients at the University of Iowa Hospital and Clinics (UIHC) for a separate study related to delirium from January 2016 to March 2020 [41][42][43][44]. Among them, we used data from a subgroup of patients recruited from November 2017 to March 2020 who had blood samples collected and processed for the epigenetics analysis [45][46][47]. Patients 18 years or older, who were admitted to the emergency department, orthopedics floor, general medicine floor, or intensive care unit were approached. Only those who consented, or whose legally authorized representative consented, were enlisted in the study. Written informed consent was obtained from all participants. Exclusion criteria included subjects whose goals of care were comfort measures only, those who were prisoners, or individuals with droplet/contact precautions. Further details of the study subjects and enrollment process are described previously [41][42][43][44].
We tested 173 subjects for genome-wide DNA methylation (DNAm) status, then conducted a post-hoc analysis of the available data to assess the influence of metformin. This study was approved by the University of Iowa Hospital and Clinics Institutional Review Board, and all procedures were compliant with the Declaration of Helsinki.

Clinical information
Clinical variables were gathered through electronic medical chart review, patient interviews, and collateral information from family members [41][42][43][44]. Metformin use, insulin use, and type 2 diabetes mellitus (DM) history were obtained by using the search terms "metformin", "insulin", and "DM" or "diabetes", respectively [12]. Only type 2 diabetes mellitus (DM) was included, excluding type 1 diabetes mellitus or gestational diabetes. If there was a history of metformin prescription before the study enrollment, patients were categorized as metformin users (Met). Those who were prescribed metformin after participation were not categorized as metformin users (non-Met) since the blood was obtained prior to such prescription.

Sample collection
Blood samples were collected in EDTA tubes during patients' hospital stay. Samples were shipped to the research laboratory and stored at −80°C until downstream analysis as a batch.

Statistics and bioinformatics analysis
All analyses were conducted using R. The R packages ChAMP [48] and minfi [49] were used to process the data. Data from a total of 175 samples from 173 subjects were included for the statistical and bioinformatic analysis. DNAm levels for each CpG site were first compared between those with and without a history of metformin prescription (first run; Supplementary Table 1). Then, comparison limited among only DM patients between those with and without a history of metformin prescription was conducted to avoid potential influence of DM on DNAm status (second run; Supplementary Table 2).
During quality control processes, 2 samples from the first run and no samples from the second run were excluded based on the density analysis plots as a part of our quality control pipeline. 2 samples were also excluded because two patients had their blood collected twice. The first collected samples were included for further analysis while the second samples were excluded to maintain AGING consistency between samples from all subjects. Therefore, 171 subjects from the first run and 63 subjects from the second run remained for the analysis. Furthermore, during the data loading process, probes were filtered out if they (i) had a detection p-value >0.01, (ii) had <3 beads in at least 5% of samples per probe, (iii) were non-CpG, SNP-related, or multi-hit probes, or (iv) were located on chromosome X or Y. Beta mixture quantile dilation [50] was used to normalize samples, while the combat normalization method was used to correct for batch effect in the first run [51,52]. The second run, which only included diabetic patients, was not corrected for batch effect because there were individual patients who were not part of any batches.
Top hits based on each CpG site difference were obtained through the RnBeads package using the limma method [53,54] and accounting for age, sex, insulin use, BMI and cell type proportions (CD8 T cells, CD4 T cells, natural killer cells, B cells, and monocytes) as covariates. DNAm Age Calculator available online [55] calculated the cell type proportions through the method reported previously [56].
After obtaining the top CpG sites, enrichment analysis followed using missMethyl [57] and unbalanced numbers of CpG sites on each gene were controlled using the EPIC array. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genome (KEGG) [58] analysis was conducted. The number of CpG sites included in the analysis was determined by the combination of p-value and beta value cutoffs of the methylation rates of each CpG site (p < .01 and beta >0.04). Genome-wide significance was set at a p-value of less than = 5.0E-08.
The chi-square test compared the categorical data (sex, race, and insulin use) between two groups, while the Welch's t-test compared the numerical data (age, BMI, and CCI) between two groups.

DNA methylation aging clock analysis
To investigate whether subjects with history of metformin use had slower "age acceleration" than subjects without history of metformin use, we submitted the raw DNA methylation beta values to a publicly available tool, which includes the Horvath [55] method. The calculated output was the difference between the DNA methylation age and the chronological age.

Availability of data materials
The datasets analyzed during the current study are available from the corresponding author upon reasonable request.