Piousbox

Social relationships are adaptive and critical for wellbeing and survival in social species¹. Social isolation and loneliness, characterized as reflections of objective and subjective manifestations of impoverished social relationships, are increasingly recognized as important global public concerns². Cumulative evidence demonstrates that both social isolation and loneliness are linked to morbidity and mortality, with effects comparable to traditional risk factors such as smoking and obesity^3,4,5,6. Despite these empirical associations, the underlying mechanisms through which social relationships impact health remain elusive.

Experimental studies show that social interactions can causally alter animal physiology, such as sympathetic nervous system (SNS) and hypothalamic–pituitary–adrenal (HPA) activity, inflammation and antiviral responses, and directly influence disease risk^7,8,9. These patterns parallel observations in human correlational studies^10,11. Moreover, adverse social relationships have been associated with unhealthy lifestyles¹², potentially impacting these physiological pathways and subsequently affecting health. In comparison with behavioural moderators that indirectly influence health, there is a growing focus on understanding biological processes mediating the link between social relationships and health, given their relevance to improving disease prediction, prevention and intervention. It is noteworthy that proteins, as the final products of gene expression, serve as the main functional components of biological processes and represent a major source of drug targets¹³. Therefore, understanding the proteomic associations of social isolation and loneliness becomes imperative for unravelling the biology underpinning the effects of social relationships on health. One study reported that circulating brain-derived neurotrophic factor levels were associated with social relationships and partly mediated the effect between social support and dementia risk¹⁴. However, no comprehensive proteome-wide association study (PWAS) of social isolation and loneliness has been performed so far.

Here, leveraging high-throughput, population-scale proteomics alongside deep phenotypic data from the UK Biobank, we aimed, in this novel study, to answer two key questions: (1) What are the proteomic profiles associated with social isolation and loneliness? (2) How do proteomic alterations contribute to the relationship of social isolation, loneliness and health? To address the first question, we initially conducted PWASs and protein co-expression network analysis for social isolation and loneliness, respectively. The identified proteins and protein modules were subsequently examined potential causal relationships with social isolation and loneliness using bidirectional Mendelian randomization (MR) and colocalization analyses. To explore the relationship between the MR-identified proteins and broad physical functions, we investigated their associations with other blood biomarkers. On the basis of the social brain hypothesis¹⁵ and increasing research on the neurobiology of social isolation and loneliness¹⁶, we further related these proteins to brain volumes. For the second question, we delved into the prospective associations between proteins linked to social isolation and loneliness and the incidence of morbidity and mortality. Specifically, we focused on five diseases with well-documented associations with social relationships: cardiovascular diseases (CVDs)^17,18, dementia^19,20, type 2 diabetes (T2D)²¹, depression²² and stroke^18,23. Finally, we evaluated the role of plasma proteins in the connection between social relationships and the risk of morbidity and mortality in mediation analyses for time-to-event outcomes.

Cohort characteristics

Our primary study population included 42,062 participants (aged 56.4 ± 8.2 years, 52.3% female) from the UK Biobank who had quality-controlled proteomic data and complete behavioural data including social isolation, loneliness and all covariates. A flow chart of the participant selection is shown in Supplementary Fig. 1. Among these, 3,905 (9.3%) reported being socially isolated and 2,689 (6.4%) felt lonely. Detailed demographic characteristics, stratified by social isolation and loneliness, are presented in Table 1. During a median (s.d.) follow-up of 13.7 (2.1) years (ending on 30 November 2022 or death), 2,695 participants developed CVD, 892 developed all-cause dementia, 1,703 developed T2D, 1,521 developed depression, 983 developed stroke and 4,255 passed away.

Proteins associated with social isolation and loneliness

We conducted logistic regression for PWASs involving 2,920 plasma proteins, using social isolation or loneliness as the outcome. In simple models incorporating age, sex, site, technical factors and the first 20 genetic principal components (PCs) as covariates, we found 776 proteins significantly associated with social isolation and 519 proteins associated with loneliness (P < 0.05/(2,920 × 2) = 8.6 × 10⁻⁶) (Supplementary Fig. 2). After additional adjustments for ethnicity, education level, household income, smoking, alcohol consumption and body mass index (BMI), 175 proteins associated with social isolation (Fig. 1a and Supplementary Table 1) and 26 proteins associated with loneliness (Fig. 1b and Supplementary Table 2) maintained significance at the Bonferroni-corrected threshold.

Notably, growth differentiation factor 15 (GDF15), a protein belonging to the transforming growth factor-β superfamily that acts as an inflammatory marker²⁴, demonstrated the strongest association with social isolation (odds ratio (OR) of 1.22, 95% confidence interval (CI) 1.17 to 1.27, P = 1.2 × 10⁻¹⁹, N = 41,396), while proprotein convertase subtilisin/kexin type 9 (PCSK9), a key protein in the regulation of cholesterol metabolism²⁵, showed the strongest association with loneliness (OR of 1.15, 95% CI 1.10 to 1.20, P = 4.2 × 10⁻¹¹, N = 41,396). The majority of identified proteins exhibited a positive association, indicating that higher protein abundance was linked to an elevated risk of social isolation or loneliness. However, only four proteins emerged as protective factors against social isolation, and one against loneliness. Particularly notable among these was C-X-C motif chemokine ligand-14 (CXCL14), an immune and inflammatory modulator²⁶, emerging as the second most significant protein associated with social isolation (OR of 0.84, 95% CI 0.81 to 0.88, P = 2.4 × 10⁻¹⁷, N = 40,090).

Furthermore, 22 proteins (12.3% of the overall identified proteins) were common (Fig. 1c), and the proteome-wide associative patterns of social isolation and loneliness were moderately related (r = 0.54, 95% CI 0.52 to 0.57, P < 2.2 × 10⁻¹⁶, N = 2,920; Supplementary Fig. 3), implicating shared and distinct proteomic signatures between social isolation and loneliness. Restricted cubic splines²⁷ revealed that four proteins displayed a significant nonlinear association with social isolation (P < 0.05/(175 + 26) = 2.5 × 10⁻⁴; Supplementary Fig. 4), while no nonlinear proteomic association with loneliness was detected.

Sensitivity analyses

First, ordered logistic regressions using raw scores of social isolation and loneliness were performed to test the robustness and potential dose-dependent associations with plasma protein levels. This approach yielded results consistent with those from dichotomous variables and logistic regression but with greater statistical power (580 and 125 proteins associated with social isolation and loneliness, respectively; Supplementary Fig. 5).

Given the reported sampling bias in the UK Biobank Pharma Proteomics Project (UKB-PPP) subcohort composition, we replicated the primary analyses specifically within the randomly selected subset (N = 36,250), which is highly representative of the overall UK Biobank population. The proteins identified in this subset were consistent with those in all available samples and the proteomic associative patterns in these two populations were highly correlated (both r > 0.9, P < 2.2 × 10⁻¹⁶; Supplementary Fig. 6). Additionally, to mitigate potential population stratification, we conducted PWAS specifically in Caucasians (N = 35,697) and observed that the proteomic associative patterns were highly correlated with those found in the full sample (both r > 0.9, P < 2.2 × 10⁻¹⁶; Supplementary Fig. 7).

Next, interaction terms between sex or age and protein levels were included in logistic models to assess potential sex and age differences. No significant interaction effects between sex or age and proteins associated with social isolation or loneliness were observed (P > 0.05/(175 + 26) = 2.5 × 10⁻⁴). The PWAS results for males (N = 20,076) and females (N = 21,986) are shown in Supplementary Fig. 8. Additionally, the PWAS results for younger (<60 years, N = 24,171) and older (≥60 years, N = 17,891) groups are shown in Supplementary Fig. 9.

Considering the established link between loneliness and depression²², we explored the potential influence of depressive symptoms on the proteomic association of social isolation and loneliness (N = 38,778). The adjustment for depressive symptoms modestly impacted the association between social isolation and proteins (Supplementary Fig. 10a,b), whereas notably weakening the association between loneliness and proteins (for example, the OR of GDF15 decreased by 9.5% after adjustment; Supplementary Fig. 10c–e). Additionally, we examined the potential influence of physical activity (N = 34,548) and found that the associations between social isolation, loneliness and plasma proteins were largely independent of physical activity levels (Supplementary Fig. 11).

Incorporating both social isolation and loneliness into the model had minimal impact on their respective associations with proteins (Supplementary Fig. 12). To further investigate the underlying construct of impoverished social relationships, multinomial logistic regressions were conducted to test the association between the four-group classification and plasma protein levels. Compared with participants who were neither isolated nor lonely (N = 36,100), 116 proteins differed significantly in the socially isolated (SI) but not lonely (LO) group (SI+LO−; N = 3,273), 8 in the not isolated but lonely group (SI−LO+; N = 2,057) and 22 in the socially isolated and lonely group (SI+LO+; N = 632) (Supplementary Fig. 13). Interestingly, GDF15 was identified as the top differentiated protein in SI+LO− and SI+LO+, while PCSK9 was the top differentiated protein in SI−LO+.

Finally, we performed cross-validation by randomly splitting the UK Biobank samples 100 times. Our results showed that most of the significant proteins identified using the full sample retained significance in at least one of the two split samples, and proteomic associative patterns for social isolation (mean r = 0.56, s.d. 0.05) and loneliness (mean r = 0.35, s.d. 0.08) between the two split samples exhibited medium to large correlations (Supplementary Fig. 14).

PPI and pathway enrichment

To explore potential interactions among proteins associated with social isolation or loneliness, we analysed the protein–protein interaction (PPI) networks within the identified pool of 179 proteins using the STRING database²⁸. A prominent interconnected network featuring 150 proteins and 690 PPIs was found (Fig. 1d). The maximal clique centrality (MCC) method²⁹ revealed the top three hub proteins of interleukin 6 (IL6), intercellular adhesion molecule-1 (ICAM1) and cluster of differentiation 4 (CD4) all exhibiting associations with social isolation.

Furthermore, we investigated functional pathways for the proteins associated with social isolation and loneliness at a false discovery rate (FDR) q < 0.05. We observed several shared pathways between social isolation and loneliness. Both sets of proteins exhibited significant enrichment in immune pathways such as cytokine–cytokine receptor interaction and antiviral processes (Fig. 1e,f and Supplementary Tables 3 and 4). Additionally, proteins associated with social isolation specifically showed enrichment in complement system and mitogen-activated protein kinase (MAPK)/extracellular signal-regulated kinase (ERK) signalling.

Protein networks linked to social isolation and loneliness

Given that complex diseases are not caused by individual proteins, but rather result from highly interactive protein networks³⁰, we applied a data-driven approach to classify plasma proteins into clusters or modules based on protein co-expression patterns. Thirteen non-overlapping protein modules were identified, ranging in size between 30 and 1,051 proteins (Supplementary Fig. 15). We then correlated the module eigengene with social isolation and loneliness, revealing that M4, M8 and M12 showed associations with both conditions after Bonferroni correction (P < 0.05/(13 × 2) = 0.002) (Fig. 2a,b). Additionally, M3 was found to be associated with social isolation. Enrichment analyses suggested distinct functions associated with these modules. For example, protein module M4 exhibited enrichment in immune-related pathways (Fig. 2c) and included the top protein related to social isolation, GDF15 (Supplementary Table 5). M8 demonstrated enrichment in metabolic processes (Fig. 2d and Supplementary Table 6). M12 showed enrichment in neutrophil degranulation (Fig. 2e and Supplementary Table 7). M3 was enriched in complement systems (Fig. 2f) and included the top protein related to loneliness, PCSK9 (Supplementary Table 8). Actually, M3 was also associated with loneliness if relaxing the significance threshold to FDR correction. Sensitivity analysis demonstrated that the biological relevance of the identified modules associated with social isolation and loneliness remained robust across various co-expression network construction parameters (Supplementary Figs. 16–18).

Moreover, we examined the relationship between significant modules with the proteins identified by PWAS using two-sided Fisher’s exact tests. After Bonferroni correction, both protein sets associated with social isolation and loneliness were significantly enriched in M4, and the results were consistent across all proteins, the top 20% and the top 10% of proteins in each module (Supplementary Table 9). It is noteworthy that both approaches—the PWAS and the protein co-expression networks—were consistent and complementary, together providing more comprehensive information.

MR between social isolation, loneliness and proteins

To infer causality between social isolation, loneliness and the identified proteins and protein modules, we implemented a bidirectional two-sample MR. Genome-wide association study (GWAS) summary statistics were sourced from non-overlapping samples from the UK Biobank. In the forward direction, no protein or protein network was found to be associated with social isolation or loneliness at an FDR significance threshold, using either cis-protein quantitative trait loci (pQTLs) alone or a combination of cis- and trans-pQTLs (Supplementary Table 10). However, we uncovered significant associations between loneliness and five proteins in the backwards direction using the inverse-variance weighting (IVW) method (FDR q < 0.025) using both cis- and trans-pQTLs (Fig. 3 and Supplementary Table 11). The results of the sensitivity analyses were consistent with IVW estimates in direction and magnitude. No evidence of heterogeneity (Q statistic, all P > 0.3) and horizontal pleiotropy (MR–Egger intercept, all P > 0.3; heterogeneity in independent instrument-outlier test, P > 0.01) among instrument variables (IVs) was detected. A leave-one-out analyses demonstrated no potentially influential single-nucleotide polymorphisms (SNPs) driving the causal link (Supplementary Fig. 19). However, the causal analysis using summary effect (CAUSE) method could not distinguish a model of causality from correlated pleiotropy for the five proteins that showed significant causal links with loneliness using IVW (Supplementary Table 12).

For the MR-identified proteins in relation to loneliness, we implemented colocalization analysis to ensure that the results were not confounded by linkage disequilibrium. Two proteins, ADM and asialoglycoprotein receptor 1 (ASGR1), exhibited strong evidence of colocalization (posterior probability of hypothesis 4 (PPH4) > 0.8 for one common causal variant) for at least one of the instruments, while GDNF family receptor alpha 1 (GFRA1) showed medium support for colocalization between loneliness and pQTL signals (PPH4 of 0.741) (Fig. 3).

...