Epigenome-wide association study of loneliness in a sample of U.S. middle-aged twins

ABSTRACT

Loneliness is a complex human trait that is highly polygenic and found to affect gene expression related to inflammatory and immunological functioning. To date, no epigenome-wide association studies of loneliness have tested whether differentially methylated sites are annotated to genes associated with inflammatory and immunological processes. Using 281 individual adult twins’ DNA methylation data from the Louisville Twin Study, we performed an epigenome-wide analysis of loneliness to address this gap in the literature. In the discovery analysis, 169 twins were used to prioritize probes and test associations with DNA methylation age acceleration, and 56 independent monozygotic (MZ) twin pairs (112 individuals) were used in a within-family replication analysis. Among the 837,274 sites analyzed, no probe sites were statistically significant at the genome-wide level (p < 5.97 × 10⁻⁸), but 25 suggestive sites (p < 5 × 10⁻⁵) were annotated to genes related to various biological processes, including inflammatory response and protein-binding functions that extend prior findings. The nominal associations at these suggestive probe sites were highly correlated (r = .72) between the discovery sample and the MZ pair replication sample. Finally, loneliness significantly correlated with the DunedinPACE DNA methylation measure, suggesting that higher levels of loneliness were associated with accelerated epigenetic age as quantified by a measure that indexes longitudinal changes across multiple organ systems.

In recent years, loneliness (sometimes referred to as perceived social isolation) has been declared a public health crisis in many countries [1,2], despite no clear diagnostic criteria for loneliness as there is for depression and anxiety disorders [3]. The prevalence of loneliness ranges from approximately 5.3% to 11.9%, depending on age, nationality, and operationalization of loneliness [4], with considerable variability in levels and trajectories of loneliness across the lifespan [5,6]. Loneliness has been deemed problematic in part because of its association with various disease risks (e.g., cardiovascular disease) [7]. The purpose of the current study is to investigate whether epigenetic mechanisms have a role in why loneliness is frequently observed to associate with human disease risks.

The genetic and epigenetic markers that contribute to the measureable phenotypic expression of loneliness likely are numerous but have not been studied adequately [8]. Although a few twin studies suggest heritability estimates of loneliness from 37% to 55% [8–10] and one genome-wide association study suggests estimates from 16% to 27% [11], we understand little about the epigenetics of loneliness. One gene expression study points to the possibility that loneliness may alter the epigenome, as environmental exposures among lonelier people altered gene expression [12], leading to up-regulated expression of pro-inflammatory gene sets and down-regulated expression of antiviral and antibody-related gene sets. As the epigenome serves as the interface between people’s environments and gene expression, epigenetic mechanisms (e.g., DNA methylation) can help to clarify how loneliness might correlate with physiological functioning and disease risk.

DNA methylation (DNAm) occurs at regions of DNA where cytosine precedes guanine from the 5’ to 3’ direction. Levels of DNAm at these cytosine–phosphate–guanine (CpG) sites may affect gene expression [13]. To date, two studies [14,15] have reported the association between loneliness and differential methylation in an a priori selected subset of CpG sites. In an unpublished dissertation thesis that used an older adult sample of Swedish twins [15], the association between loneliness and 2,324 CpG sites that included genes enriched in inflammatory and immunological response pathways were tested [12]. No CpG sites were significantly associated with loneliness scores after adjusting for multiple testing, but the five sites most strongly correlated with loneliness were annotated to genes associated with blood clot formation, inflammation, and immune function. In a second study [14] that followed a sample of Belgian pre-adolescents, loneliness correlated with DNAm levels in a selected subset of CpG sites that included a targeted set of genes – specifically, BDNF, SLC6A4, and NR3C1 – associated with the hypothalamic–pituitary–adrenal (HPA) axis. Both studies, however, tested a priori, theoretically motivated, probe sites, and the results were mixed. Moreover, loneliness may be associated with hypermethylation and hypomethylation in other regions of the genome that regulate biological processes beyond immunological and inflammatory pathways. The primary aim of the current study is to explore CpG sites across the entire human genome that are associated with differences in levels of loneliness [8].

As part of the current invesigation, we also tested the association between loneliness and six measures of DNAm age that index epigenetic aging. Significant associations support the relation between age-based measures of sets of CpG probe sites and loneliness [6,16]. Commonly referred to as ‘epigenetic clocks,’ DNAm age measures use a combination of sets of CpG sites, demographic variables, physiological biomarkers (e.g., albumin), and behavioral measures (e.g., smoking) to quantify people’s epigenetic age. Collectively, DNAm age measures outperform chronological age when predicting phenotypic outcomes including chronic health conditions, diseases (e.g., cancer), mortality, and cognitive impairment [17]. Initial DNAm age measures were developed using small sets of CpG sites (e.g., 71–513), chronological age [18], and sex [19]. Subsequent measures have incorporated inflammatory and immunological biomarkers [20], lifestyle factors and mortality risk [21], and longitudinal trajectories of multiple organ systems [22] in addition to sets of CpG sites for broader applications of predicting age-related outcomes. The utility of DNAm age measures, however, involves quantifying DNAm age acceleration in which DNAm age scores are regressed on chronological age to yield residual scores. Positive residual scores indicate accelerated biological aging, whereas negative values indicate slower aging.

In the current study, we conducted an epigenome-wide association study using genomic DNA extracted from whole blood tissue. Although we made no specific predictions about specific CpG sites that would associate with loneliness, we hypothesized that CpG sites most strongly correlated with loneliness would annotate to genes related to inflammatory and immune response. We then conducted a pathway enrichment analysis to explore whether CpG regions associated with loneliness were annotated to gene sets enriched in cellular, molecular, and biological processes, with a specific interest in testing whether any of these pathways were related to inflammatory and immunological functioning. Finally, we estimated correlations between loneliness, six DNAm age, and associated age acceleration measures. Here, we hypothesized that higher loneliness would be associated with higher age acceleration scores across all measures.

Method

Participants

The sample included 281 adult-aged twins (32–65 years) from the Louisville Twin Study (LTS) [23], which is a longitudinal twin study of physical, cognitive, and psychosocial development that began in 1957. Same-sex twins’ zygosity was determined by blood-typing on 22 antigens [24]. Data were drawn from twins who participated in a midlife data collection in which twins completed a self-reported psychosocial assessment and provided 50 cc of whole blood that was used for the DNAm assay. One hundred sixty-nine individual twins representing 39 complete dizygotic (DZ) pairs, 2 unknown zygosity pairs, and 87 incomplete (i.e., singleton) pairs (39 MZ, 42 DZ, and 6 unknown) were used in the discovery analysis. In the replication analysis, an independent sample of 56 complete MZ pairs (112 individuals) was used. The study was approved by the Institutional Review Board at the University of Louisville, the University of Southern California, and the University of Virginia. Participants provided demographic characteristics (chronological age, biological sex (male, female), and smoking status (current, never/former) by questionnaire report.

Loneliness assessment

Loneliness was measured with three items from the UCLA Loneliness Scale – Revised [25]. The brief 3-item version is a measure of social loneliness that assesses whether people feel disconnected from a broader social network or that one’s network feels inadequate. The items specifically measure whether people feel like they lack companionship, feel left out, and feel isolated from others (all rated on a scale of 1 (‘always’) to 4 (‘never’), have been used in prior research [26], and show substantial reliability in the current samples ($\mathit{\omega }$_Discovery = .91; $\mathit{\omega }$_{MZ Pairs} = .90). Responses were reverse-scored and summed so that higher scores indicate higher loneliness.

DNA extraction and preparation for epigenetic analysis

Genomic DNA was extracted from whole-blood samples that were collected at the University of Louisville and Norton Healthcare medical campuses. Venous blood was collected in ethylenediamine tetraacetic acid (EDTA) tubes and shipped to the Norris Comprehensive Cancer Institute at the University of Southern California (USC) Keck School of Medicine. DNA was extracted with Promega Maxwell 16 LEV blood DNA kits and then treated with bisulfite reagents at the USC Molecular Genomics Core following manufacturer’s protocol. With the bisulfite-converted DNA samples, greater than 865,859 DNAm loci were assessed using the Illumina Infinium Human MethylationEPIC BeadChip (Illumina, San Diego, CA, USA).

Microarray processing and quality control

All DNAm assays (N = 286) were performed at the USC Molecular Genomics Core at the Norris Comprehensive Cancer Institute at the USC Keck School of Medicine. Quality control was conducted in R 4.3.0 [27] using the minfi package [28] to identify aberrant samples and CpG sites, remove cross-reactive probes, conduct background correction, adjust for batch effects, and adjust for cell composition. Failed CpG sites were assessed across samples and probe sites. Samples with mean detection p-values across all CpG sites $\ge$.01 were not used in the current study (n = 1). CpG sites with detection p-values $\ge$.10 were removed from the final set of CpG sites (n = 1,160). Cross-reactive CpG sites were removed using a minimum number of 47 bases used to identify unintended targets (n = 29,233) [29,30]. Failed CpG sites, samples, and cross-reactive probes were removed following background correction. After quality control, 281 participants remained and 837,274 CpG sites (96.70% of raw CpG sites) were included in the analysis.

Next, we conducted background correction using the normal-exponential out-of-band (noob) method [31]. Samples were measured in four groups, and we adjusted for laboratory batch using the ComBat method [32]. We estimated blood cell composition using the Houseman method [33] and obtained percentages of CD8+ T cells, CD4+ T cells, natural killer cells (NK), B cells, monocytes, and granulocytes. Because cell proportions can be correlated with each other, we performed principal component analysis (PCA) on the estimated cell proportions. Principal components 1 and 2 explained 86.12% of the variance in the cell proportion data, so we used both in our analyses. Our primary analysis examined DNAm levels at each site individually.

DNA methylation age

To estimate biological age and age acceleration, six DNAm age measures were used in the current study. Three were designed to assess chronologic age [18,19,34], and three were designed to assess phenotypic age [20–22]. Horvath’s [18] DNAm age was defined by 353 CpG sites, whereas Horvath et al.’s [34] skin-based measure included 391 CpG sites based on associations in fibroblasts and other skin cell tissue types. Hannum et al. [19] included 71 CpG sites and biological sex. DNAm PhenoAge included nine physiological biomarkers (albumin, creatine, serum glucose, C-reactive protein (CRP), lymphocyte percent, mean red cell volume, red cell distribution width, alkaline phosphatase, and white blood cell count) and chronological age to estimate a measure of phenotypic age that, in turn, was associated with 513 CpG sites [20]. DNAm GrimAge was developed by regressing time-to-death on surrogate DNAm biomarkers of adrenomedullin, CRP, plasminogen activation inhibitor 1 (PAI-1), and growth differentiation factor 15 (GDF15), and smoking pack-years [21]. Both Horvath measures, Hannum’s measure, DNAm PhenoAge, and DNAm GrimAge were constructed using a principal component methodology to improve the reliability of DNAm age scores [35] using R scripts available on GitHub (https://github.com/MorganLevineLab/PC-Clocks). These five DNAm age measures were scaled to reflect years of biological age.

We also used the Dunedin Pace of Aging Calculated from the Epigenome (DunedinPACE), which is a ‘third-generation’ measure trained using longitudinal measures across 19 organ systems from the Dunedin Cohort Study from 26 years to 45 years [22,36]. DunedinPACE estimates interindividual differences in intraindividual change in multi-organ-system integrity conditioned on 173 CpGs measured at age 45. DunedinPACE was estimated using R scripts available on github (https://github.com/danbelsky/DunedinPACE), with scores scaled to represent the average rate of physiological change per calendar year.

DNAm age acceleration scores were estimated by regressing each DNAm age measure on chronological age in an ordinary-least squares bivariate regression model and saving the residuals for analysis.

Data analysis

Epigenome-wide association study (EWAS) analyses were conducted in the discovery sample of 169 individual twins using the lmFit function from the limma package in R [37] with separate regression models for each CpG probe site. The overall purpose of this analysis was to identify CpG probe sites that were statistically significantly associated with loneliness. The level of DNAm was regressed onto loneliness, age at midlife assessment, current smoking status, biological sex, and the first and second principal components of cell composition. Empirical Bayes estimates were used (also in the limma package) to compute t statistics and probability values for differential methylation at each CpG site. To adjust for any dependency between samples from twins in the same family, we used the duplicateCorrelation function in the limma package to adjust for sample dependency. The consensus correlation estimate, which is a robust average of the correlations between sites, was included in each CpG regression and is appropriate when observations (i.e., twins) are not independent [38]. The following linear function was fit to each CpG site:

$$\mathit{DNA}{m}_{\mathit{ijk}}=b_{00\mathit{i}}+u_{0\mathit{ik}}+b_{1\mathit{i}}\mathit{Lonely}+b_{2\mathit{i}}\mathit{Age}+b_{3\mathit{i}}\mathit{Sex}+b_{4\mathit{i}}\mathit{Smoke}+b_{5\mathit{i}}\mathit{PC}1+b_{6\mathit{i}}\mathit{PC}2+{\epsilon }_{\mathit{ijk}}$$

Bonferroni’s correction was applied to hold the Type I error rate at .05. The adjusted $\mathit{\alpha }$ cut-off value was 5.97 × 10^–8. For hypothesis generation, we additionally report CpG sites using a less stringent cut-off value of 5 × 10^–5 that takes into account the small sample size and the expected small regression coefficient values while using a quantitative loneliness predictor instead of using previously suggested cut-off values designed for larger samples and larger effect sizes [39]. We assessed the model fit using scatterplots of observed and expected p-values and we calculated lambda ($\mathit{\lambda }$) values. We used the Illumina manifest file (downloaded 9 November 2023) to annotate the top CpG sites to human genome assembly GRCh38 (hg38) (https://webdata.illumina.com/downloads/productfiles/methylationEPIC/infinium-methylationepic-v-1-0-b5-manifest-file-csv.zip).

We visualized adjusted associations between loneliness and site-specific DNA methylation using a volcano plot, showing the regression coefficient on the x-axis and the -log₁₀ p-value on the y-axis, highlighting sites reaching the suggestive p-value threshold. Among sites reaching the suggestive threshold, to examine patterns in the overall level of methylation (hypermethylation defined as >75% methylation in the sample versus hypomethylation defined as <25% methylation in the sample), we used Fisher’s exact test. To further annotate the suggestive sites, we referenced the probe mapping quality and genetic polymorphism information [40]. To facilitate future replication opportunities, we examined suggestive site probe availability on the most recently available EPIC v2.0 beadchip.

Pathway enrichment analyses were performed using the missMethyl package [41,42]. The gometh function included all suggestive CpG sites with a cut-off value of p < 5 × 10^–5. Paths from gene ontology (GO) and KEGG terms were used in the current analysis, and a prior probability was permitted for significant differential methylation due to numbers of probes per gene. Overrepresentation in GO and KEGG pathways was evaluated using a false discovery rate (FDR) of .05 to adjust for multiple testing.

Pearson product-moment correlations were estimated between chronological age and each DNAm age measure. Correlations were then estimated between loneliness and each DNAm age acceleration measure. Null hypothesis significance tests were performed in the discovery sample using an $\mathit{\alpha }$ cut-off value of .05.

Participants of this study did not agree for their data to be shared publicly, so supporting data are not currently available. Interested researchers, however, may consult with the corresponding author about potentially acquiring data used in the current study.

Replication analysis with complete MZ pairs

We tested whether statistically significant associations found in the discovery sample replicated in 112 individual MZ twins (56 complete pairs). As noted above, the reason for using a sample of independent MZ twin pairs was to control for the familial similiarity between MZ twins due to their shared genotype and any nongenetic (environmental) factors that contribute to similarity in their DNAm values. In the case where no CpG probe sites were genome-wide significant, we tested whether any suggestive sites observed in the discovery sample were significant in the MZ replication sample. Statistical significance of each CpG site was tested using an $\mathit{\alpha }$ cut-off value of .05. In this analysis, we examined the direction and magnitude of the correlation between the suggestive CpG regression coefficients in the discovery and replication samples and used a scatterplot for visualization.

Results

Table 1 provides the demographic characteristics of the discovery sample and the replication sample of MZ twin pairs. In our discovery sample, the average age was 51.54 and the average loneliness score was 5.22, which is indicative of mild loneliness (Table 1). Our discovery and replication samples were similar with respect to loneliness, age, biological sex, and smoking status, as no significant differences were observed between groups.

Table 1.

Characteristics of discovery and MZ twin pair replication samples in the Louisville Twin Study.

	Discovery (n = 169)		Complete MZ Pairs (n = 112)
Variable	M (%)	SD	M (%)	SD	t/$x^2$(df)	p
Loneliness	5.22	2.22	5.35	2.42	−0.44(279)	.661
Age in years	51.54	6.91	52.27	7.24	−0.84(279)	.400
Female	55.62%	-	64.29%	-	1.75(1)	.186
Current smoker	26.04%	-	16.96%	-	2.69(1)	.101
Horvath AA	0.09	4.02
Horvath-Skin AA	0.05	4.33
Hannum AA	0.22	4.18
DNAm PhenoAA	0.44	5.71
GrimAA	0.29	4.10
DunedinPACE	0.99	0.13

Note. Discovery sample consisted of 169 individuals. Complete MZ twin pair sample consisted of 112 individuals (56 complete pairs).

In the discovery sample, we tested for adjusted associations between loneliness and DNAm in an EWAS framework. The $\mathit{\lambda }$ value across these models was 1.71 with more significant CpG sites underrepresented (see Q–Q plot in Supplemental Figure S1). Figure S2 presents the Manhattan plot of the p-values in the EWAS. In this sample, no CpG site was genome-wide significant. Table S1 provides summary statistics of all associations. We prioritized 25 suggestive CpG sites with p-values <5 × 10^–5 for follow-up (Table 2). In most of these sites (68%), DNAm increased with higher loneliness scores, as indicated by the positive regression coefficient. Moreover, Figure 1 shows the −log₁₀ p-values plotted against the regression coefficients of all 837,274 CpG sites, suggesting higher DNAm with higher loneliness scores (59.72%). Fisher’s Exact Tests showed that the 25 suggestive probe sites were no more likely to be hypermethylated (p = .690) or hypomethylated (p = .825) than all other probe sites across the entire discovery sample.

Table 2.

EWAS summary statistics for the 25 suggestive CpG sites showing the most significant association found in the discovery sample and tested for statistical significance (p < .05) in an independent sample of complete MZ twin pairs.

	Discovery Sample						Replication in Complete MZ Pairs
CpG Site ID	Chr	Location	Genes	Est.	AMV	p	Est.	AMV	p
cg27552821$\mathit{†}$	22	22124530	MAPK1	0.005	0.91	2.56 x 10–6	−0.0002	.90	.859
cg11767117	1	76252650	RABGGTB, SNORD45A, SNORD45C	−0.001	0.04	5.56 x 10–6	0.0004	.04	.327
cg16624054	1	247171886	ZNF695, ZNF670-ZNF695	0.002	0.96	1.32 x 10–5	−0.0004	.96	.318
cg07956751	7	150211939	GIMAP7	−0.002	0.06	1.44 x 10–5	−0.0005	.06	.404
cg00490603a,$\mathit{†}$	22	46262041		−0.019	0.87	1.51 x 10–5	−0.0031	.87	.566
cg18361395	15	101100451		0.005	0.88	1.93 x 10–5	0.0010	.88	.329
cg02721852	19	3818830	ZFR2, MIR1268A	0.010	0.38	1.95 x 10–5	0.0038	.38	.155
cg18169886	2	25517869	DNMT3A	−0.001	0.03	2.73 x 10–5	0.00002	.03	.931
cg09618998	8	1950965	KBTBD11	0.007	0.48	2.86 x 10–5	0.0009	.47	.683
cg09697104	18	55212722	FECH	−0.006	0.09	2.88 x 10–5	−0.0011	.09	.470
cg27349120	17	17106053	PLD6	0.001	0.96	2.95 x 10–5	−0.0004	.96	.197
cg05957504	5	124014639	ZNF608	0.007	0.78	2.99 x 10–5	0.0013	.78	.461
cg03754700	12	132626213	DDX51	0.004	0.90	3.67 x 10–5	0.0002	.91	.822
cg03779490a,$\mathit{†}$	6	158069542	ZDHHC14	0.009	0.68	3.67 x 10–5	0.0015	.67	.601
cg15971114	15	42064995		0.010	0.69	3.85 x 10–5	−0.0006	.69	.832
cg01064150	3	119096032	ARHGAP31	0.002	0.92	3.86 x 10–5	−0.0003	.93	.567
cg18806912	15	42067504	MAPKBP1	0.006	0.45	4.04 x 10–5	−0.0008	.46	.657
cg22285168	21	35304407	LINC00649	−0.001	0.03	4.07 x 10–5	0.0001	.03	.562
cg18344365	2	30585859		0.002	0.93	4.21 x 10–5	−0.0001	.93	.807
cg07773800	20	43991521	SYS1, SYS1-DBNDD2	−0.001	0.03	4.48 x 10–5	0.0002	.03	.313
cg26046150b,$\mathit{†}$	16	33457500		0.005	0.88	4.51 x 10–5	−0.0003	.89	.779
cg19903702	8	76577776		0.001	0.98	4.61 x 10–5	−0.0004	.98	.355
cg22837499	11	18417924	LDHA	−0.001	0.07	4.66 x 10–5	0.0003	.07	.383
cg16268636	6	35016581	ANKS1A	0.007	0.47	4.73 x 10–5	0.0010	.46	.587
cg10074996$\mathit{†}$	15	72881989		0.002	0.93	4.80 x 10–5	0.0002	.94	.710

Note. Chr = chromosome. AMV = average methylation value.

^aMasked probe that includes SNP with minor allele frequency (MAF) > .01 [40].

^bMasked probe with low quality mapping [40].

$\mathit{†}$Excluded from Infinium MethylationEPIC v2.0 beadchip.

Figure 1.

Volcano plot of the regression coefficients (x-axis) versus the p-values for testing the adjusted association between loneliness and DNA methylation levels (y-axis). Each point represents one of the 837,274 CpG sites tested. Upside down triangles indicate negative effect sizes and squares indicate positive effect sizes. The 25 suggestive CpG sites are uniquely represented for clarity. The dotted vertical line indicates coefficients at exactly zero, and the dotted horizontal line indicates the suggestive -log₁₀ p-value of 5 x 10^–5.

Four CpG sites were located on chromosome 15, whereas chromosomes 1, 2, 6, 8, and 22 had two significant sites each. Only seven sites were annotated with locations relative to reference genes. On chromosome 1, a 1-unit increase in loneliness was associated with 0.15% increase in DNAm at cg16624054 and is downstream of the ZNF695 and ZNF670-ZNF695 genes. On chromosome 8, cg09618998 positively correlated with loneliness, such that for every unit increase in loneliness, there was a 0.72% increase in DNAm and is in the body of the KBTBD11 gene. On chromosome 11, cg22837499 was associated with a 0.12% decrease in DNAm with increasing loneliness and was in the body of the LDHA gene. On chromosome 12, cg03754700 was upstream of the DDX51 gene and was associated with a 0.36% increase in DNAm. On chromosome 15, cg18806912 was associated with a 0.64% increase in DNAm with a 1-unit increase in loneliness and is downstream of the MAPKBP1 gene. On chromosome 19, loneliness was positively correlated with the cg02721852 site such that there was 1.02% increase in methylation with increasing loneliness and is upstream of the ZFR2 and MIR1268A genes. Lastly, on chromosome 20, loneliness negatively correlated with the cg07773800 site with a 0.08% decrease in methylation with increasing loneliness and was upstream of the SYS1 and SYS1-DBNDD2 genes.

It is worth noting that one of the 25 prioritized CpG sites had lower than expected mapping quality (cg26046150) while two sites (cg00490603 and cg03779490) overlapped with a single nucleotide polymorphism with a minor allele frequency greater than 1% [40]. Future studies should take into account these technical and genetic details, particularly if conducting follow-up mQTL analyses. Lastly, future studies that include probes from the Infinium MethylationEPIC v2.0 beadchip should be aware that five prioritized probes (cg27552821, cg00490603, cg03779490, cg26046150, and cg10074996) were not included on the EPIC v2.0 beadchip.

Gene ontology analysis

Using the 25 suggestive CpG sites associated with loneliness presented in Table 2, we annotated them to genes and tested for enrichment of pathways. No GO or KEGG pathway was significant after adjusting for multiple testing (FDR = .05). Nevertheless, the general themes that emerged in our enrichment pathway analyses (a complete list of GO and KEGG terms and their associated p-values are in Supplemental Tables S2 and S3, respectively) suggest that these 25 CpG sites were associated with genes overexpressed in a mixture of biological processes and molecular functions. The three most significant KEGG pathways were also associated with an array of biological processes including regulation of gene expression, metabolic processes, and oxygen delivery.

MZ pair replication

Results of a paired sample t-test showed that lonelier MZ twins reported 1.73 units higher loneliness, on average, than their less lonely co-twins ($\mathit{t}$(55) = 7.11, p < .001, $\hat{d}$ = 0.95). In the replication analysis with the complete MZ twin pairs, we tested the associations between loneliness score and DNAm of the 25 CpG sites with p-values <5 × 10^–5 from the discovery analysis. The loneliness regression coefficients observed in the discovery analyses were highly correlated (r = .72) with the regression coefficients observed in the replication sample (Figure 2, as indicated in the 3 righthand columns of Table 2), indicating consistent direction and magnitudes of association. The average methylation values of these CpG sites were similar in both samples.

Figure 2.

Scatterplot of the regression coefficients observed in the discovery sample plotted against the regression coefficients from the MZ pair replication analysis. Twenty-five sites were selected based on their suggestive p-values in the discovery sample analyses and the Pearson correlation coefficient is shown.

DNA methylation age

Chronological age correlated strongly with all DNAm age variables: Horvath’s [18] measure (r = .77), Horvath et al.‘s [34] skin-based measure (r = .69), Hannum et al.‘s [19] measure (r = .74), DNAm PhenoAge (r = .65), and GrimAge (r = .79). We accounted for chronological age and estimated age acceleration measures for each clock, as described in the Methods. Correlations between loneliness and DNAm age acceleration measures were small: −.06 with Horvath; −.09 with Horvath’s skin-based measure; −.08 with Hannum; .02 with DNAmPhenoAge; .13 with GrimAge; and .18 with DunedinPACE. Only the association with DunedinPACE was statistically significant (t(167) = 2.36, p = .019), in which higher levels of loneliness significantly correlated with higher average rates of physiological change per year (Figure 3).

Figure 3.

Scatterplot of DunedinPACE values plotted against loneliness scores with overlayed line of best fit.

Discussion

Prior research has demonstrated that higher levels of loneliness correlate with differential expression of genes associated with inflammation and immunological responses [12,43]. Yet, to date, there are no published EWASs of loneliness [8] and only two studies have examined the associations between loneliness and a priori sets of CpG sites associated with genes found to be overexpressed in inflammatory and immunological pathways [14,15]. In 169 twins from the Louisville Twin Study, we showed that no CpG site correlated with loneliness at genome-wide significance. Twenty-five CpG sites had probability values less than 5 × 10^–5, but these were not replicated in the MZ twin pair sample. With the current results, we cannot conclude that there exists an association between DNAm and loneliness, although findings suggest promising avenues for further research.

CpG sites nominally associated with loneliness were spread across chromosomes, suggesting that loneliness possibly affects the regulation of numerous biological systems. In the current study, several CpG sites included genes that may be associated with basic RNA-binding functions, including ZNF608, DDX51, PLD6, FECH, and ZFR2. Broadly, these genes regulate transcription of DNA into RNA at multiple stages of protein transcription. Some sites included genes that regulate small nucleolar RNAs that guide modifications and productions of the other RNAs (e.g., SNORD45A and SNORD45C). Other sites nominally associated with loneliness included genes involved in cardiovascular diseases (ANKS1A), obesity risk (ZNF608), and protein coding required for healthy skeletal musculature (LDHA). Finally, several sites included genes that may be implicated in immunological functioning (e.g., DNMT3A, GIMAP7, MAPKBP1, ZNF608, and LINC00649), as hypothesized in the current study and elsewhwere [14,15]. These results may suggest a potential role of loneliness in human immunology. Our findings may also indicate that the effects of loneliness on human physiology are broader and potentially more fundamental than immunological and inflammatory processes alone. The disease consequences of feeling lonely, especially chronic loneliness, may not have direct, obvious pathways to human physiology but rather result from the accumulation of many small effects on the regulation of gene expression among the most basic cellular and molecular pathways as well as more complex biological pathways.

The exploratory EWAS results here diverge from the theoretically motivated analyses between loneliness and DNAm reported in prior studies [14,15] but also may extend their findings. Using DNAm levels estimated from saliva samples from 622 pre-adolescent children, Koopmans et al. [14] showed that loneliness correlated with CpG probe sites annotated to genes that promote activity along the HPA axis. Similarly, Phillips [15] observed in blood-based tissue of 356 older Swedish adults that the most significant CpG sites associated with loneliness regulated immune and inflammatory response systems. Among the 25 suggestive CpG sites that were nominally associated with loneliness in the Louisville twins, none of the reference genes included those hypothesized by Koopmans et al. [14] or Phillips [15]. This discrepancy could be attributed to sampling characteristics (e.g., children and older adults compared to middle-aged adults in the LTS) or sample size, as the current sample is smaller than both studies. Despite these differences, we found that the mitogen-activated protein kinase 1 (MAPK1) gene was located in the most significant CpG site (cg27552821), which is a protein-coding gene implicated in humans’ innate immune response [44]. The MAPK1 gene also is involved in cellular processes (i.e., cell differentiation, cell proliferation, and cell death) relevant for marshalling effective inflammatory responses [45]. That loneliness may be associated with differential methylation of the MAPK1 gene is bolstered by the observation that this gene has been found to be associated with other human complex traits and diseases, including various cancers [46], memory and learning [47], and depression [48].

Current findings, thus, provide preliminary support for the overall hypothesis that loneliness is associated with differential methylation of DNA regions related to immune and inflammatory systems [12,43].

Macrophage signaling processes regulated by the MAPK1 gene are also part of inflammatory response pathways that may consist of altered pro-inflammatory and anti-inflammatory responses in people with higher levels of loneliness. Similarly, genes involved in protein binding suggest that people with higher levels of loneliness may respond to drug performances (e.g., benzodiazepines, warfarin, and nonsteroidal anti-inflammatory drugs) differently than those with lower levels of loneliness [49].

The correlation between the regression coefficients of the 25 suggestive CpG sites observed in the discovery sample analyses and the replication analyses was large. Moreover, the direction of the association between loneliness and differential methylation among these sites was generally the same, although the effect sizes in the MZ twins were smaller. The diminished associations could be due to the greater genetic similarity in MZ twins than in other sibling types. Moreover, the holdout sample was smaller than the discovery sample, meaning that much larger sample sizes would have been required to observe statistical significance. Taken together, we interpret the large correlation coefficient as providing partial support for our findings in the discovery sample.

Lastly, the correlations between loneliness and the six DNAm age acceleration measures in the current study were small and mostly nonsignificant, a finding consistent with prior research [6]. Although loneliness has been found to correlate with measures of epigenetic age acceleration [6,16], all but one association in the current study did not statistically significantly differ from zero. Despite a significant positive association with DunedinPACE, loneliness accounted for less than 3.5% of the total variance. Of note, DunedinPACE is a multisystem measure of biological aging that incorporates information across 19 organ systems, the most of any of the measures used in the current study. Moreover, 165 of the 173 CpG sites do not overlap with the Horvath, Hannum, DNAm PhenoAge, or GrimAge measures [22], suggesting it might consist of additional CpG sites related to loneliness (none of the 25 suggestive CpG sites identified in the current study overlap with those used in DunedinPACE). Consistent with our finding that CpG sites significantly associated with loneliness were spread across the genome, DNAm age measures that incorporate information from sites associated with multiple systems may be more relevant for demonstrating whether and how loneliness affects epigenetic aging.

Limitations and conclusion

The obvious limitation of the present study is that the LTS sample size is small. Both the discovery and independent sample of MZ twin pairs were about 10g–20% the size of samples used to identify differential methylation of other psychological traits (e.g., PTSD [50]) and 45% the size of other epigenetic studies of loneliness [14,15]. Second, the measure of loneliness used in the current study was a measure of social loneliness. Although not an uncommon type of loneliness [26], the available items are an abbreviated version of the full scale. Moreover, emotional loneliness tends to be the most common form of loneliness and more variable than social loneliness across the lifespan [51,52] and, therefore, may correlate with different sets of CpG sites. Future studies should use measures of loneliness that better characterize the lack of intimacy and close friendships that characterize people’s typical experience of loneliness, particularly in middle and older adulthood. Third, the Louisville twins may not generalize to other populations in the United States, as twins tend to be wealthier, on average, and more of them are of European ancestry. The Louisville twins, however, are representative of the racial and ethnic composition of the State of Kentucky. Lastly, the causal association between loneliness and differential methylation remains unknown. Although mQTL analyses can be used in all studies, twin studies have a unique role in this regard; studying the association between loneliness and DNA methylation within MZ twin pairs controls for genetic and shared environmental variance to test directly whether within-family differences in environmental exposure account for the association between loneliness and DNAm [53].

Overall, the current study extends prior research suggesting that loneliness may correlate with DNAm of genes associated with immunological and inflammatory functions (e.g., MAPK1) [14,15]. Moreover, loneliness also was found to be correlated with methylation at numerous regions across the human genome consisting of genes that regulate general molecular and cellular functions necessary for a wide array of biological processes. Replication is needed, however, as the sample size in the current study is modest, none of the CpG probe sites reached genome-wide significance, and none replicated in the current study.

Funding Statement

Funding was provided by the National Institute on Aging NIH Grant No. [R01AG063949].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Author contributions

CRB, DWD, and ET designed the study. CRB, KMB, and TEA prepared and analyzed the data. CRB, DWD, EZ, ET, and JB acquired the data. CRB, KMB, and TEA discussed the results. CRB, ML, AIG, SAB, and ACK performed the literature review. CRB wrote the article. All authors read and approved the final version of the manuscript.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15592294.2024.2427999