Associations of 2923 Olink proteins with demographic, lifestyle, environmental and health characteristics in middle-aged Chinese adults

Deciphering the human proteome could enhance our understanding of health and disease aetiology [1]. Plasma protein levels, secreted or leaked from cells or organs, may be affected by genetic and various non-genetic factors [2], and systematic investigation of plasma proteins-exposure relationships could improve our understanding of human biology and inform disease prevention and research strategies.

Traditionally, mass spectrometry has been used to measure plasma protein levels [3‐5], but studies using this method are often limited by small sample sizes and low protein coverage [2]. In contrast, recent advances in affinity-based technologies (e.g., Olink and SomaScan) have enabled proteomics in large-scale population and clinical studies, allowing comprehensive investigations of relationships of plasma proteins with different health outcomes and associated traits [2, 6, 7]. In particular, the Olink platform, which utilises antibodies as reagents to bind target proteins, has been widely used in epidemiological research due to its high sample throughput, assay specificity and cost-effectiveness [6]. Recently, the Olink Explore 3072 platform was used to measure 2,923 plasma proteins in 54,219 participants in the UK Biobank [8], leading to many novel associations between proteins, demographic and clinical exposures and genetic factors [8]. It also replicated many of well-established associations for specific proteins, such as the associations between sex and LEP, and age and GDF15 [6‐11]. Another affinity-based platform, SomaScan [7], has also been used in population-based and clinical studies, revealing proteomic associations with both genetic and non-genetic factors, such as age, sex, and adiposity [12‐14].

Nevertheless, most previous proteomics studies investigated a relatively small number of proteins or only a few pre-selected exposures [15], without simultaneously considering a broader range of factors. Moreover, few have investigated associations of plasma proteins with composite indices (e.g., frailty) reflecting general lifestyle and health [16‐19], which could be useful in population-level screening and disease prevention. Furthermore, previous studies typically focused on the discovery of protein quantitative trait loci (pQTLs) [8, 12, 13] for Mendelian Randomisation analyses and drug target discovery and validation [2], offering limited insights into the non-genetic factors influencing circulating proteins. Finally, evidence suggests protein concentrations vary across populations [20], driven not only by ancestry differences in the genetic architecture of the proteome, but also by differences in non-genetic factors, including environment. However, most large-scale proteomics studies have been conducted in European populations, highlighting the lack of diversity in proteomics research.

Exposome-wide association studies investigate the influence of a variety of life-course exposures from external and internal sources on phenotypic traits [21, 22]. When combined with proteomics, this method can offer important insights into the complex relationship between the exposome and proteome, as well as the biological mechanisms of how non-genetic factors impact human health [23]. Therefore, to fill the evidence gap, the present study aims to use the exposome-wide approach to [1] comprehensively explore the exposure profiles of ~ 3,000 Olink proteins in ~ 2,000 Chinese adults in the CKB; [2] assess the consistency of proteomic associations between the Chinese and European populations; and [3] identify priorities for future research.

Among the 2,006 participants, the mean baseline age was 50.8 (SD 10.5) years, 62% were female and the mean BMI was 23.9 (3.4) kg/m² (Table 1). Overall, 15% of participants were regular alcohol drinkers (men: 37%; women 3%) and 25% (men: 63%; women: 2%) were current smokers. The prevalence of prior diseases was similar in males and females, with 8% of participants having respiratory disease, 2% having kidney/liver disease or tested sero-positive for HBsAg, and 6% having diabetes (self-reported or screen-detected). Similar patterns with baseline characteristics were observed in the full CKB cohort (eTable 4).

Among 37 baseline characteristics examined, 31 were associated with at least one protein at the Bonferroni-adjusted threshold in the main model (Table 1; Fig. 1 and eTable 5). The four baseline characteristics that showed the largest number of associations with proteins were age (n = 1154), sex (n = 827) and BMI (n = 869), and frailty index (n = 597). Likewise, of the 2,923 proteins examined, 1900 (65%) were associated with at least one exposure, with three proteins (CDHR2, CK-BB, and PLAT) showing the most associations with baseline characteristics (n = 14), primarily involving demographic factors and clinical measurements (Fig. 2, eFig. 3). In the mutually adjusted model, the patterns of associations were similar, although there was a reduced number of significant associations for age (n = 766), sex (n = 467), BMI (n = 675), frailty index (n = 573) and other exposures (eTables 5, 6, 7 and 8). After mutual adjustments, five proteins (CDH2, ADGRE2, ADGRD1, ACY1, MEGF9) remained to be associated with > 10 exposures.

Of the 827 sex-related proteins (higher levels in females for 259 and in males for 568 proteins), the strongest associations were with LEP, XG, FSHB, CGA, and PZP in females, and with ACRV1, EDDM3B, INSL3, SPINT3, and KLK3 in males (Fig. 3I.a). Among the top 50 sex-related proteins, most were also associated with other exposures, chiefly age (e.g., FSHB, CGA, XG, LEFTY2, RELT) and BMI (e.g., LEP, FABP4, CDHR2, APCS, CA14; Fig. 3I.b). Additionally, 77 sex-related proteins were not associated with other exposures examined, of which 27 were uniquely associated with female sex (e.g., CSF3, MELTF, ITIH4, CTRC, FCGR3B) and 50 with male sex (e.g., EDDM3B, TEX101, CRISP2, HS6ST2, SPESP1; eTable 8). Among women, some proteins were also associated with several reproductive factors (e.g. menopause status), the number of which changed little in mutually adjusted models (eTable 6).

Of the 1,154 age-related proteins, the strongest positive associations were with EDA2R, ELN, LTBP2, CSCL17, and GDF15, while the strongest negative associations were with PAEP, CTSV, SDK2, IGDCC4, and TNFSF11 (Fig. 3 II.a). Among the top 50 age-related proteins, most were also associated with other exposures, chiefly sex (e.g., FSHB, CGA, PAEP, CST3, SUSD5) and BMI (e.g., DPT, HSPB6, FGF5, MSR1, HAVCR1; Fig. 3 II.b). Additionally, 168 age-related proteins (e.g., ITGB5, IL17D, TIMP4, SORCS2, TNFSF13) were not associated with other exposures (eTable 8). In sex-specific analyses, age was associated with 1,133 proteins in males and 612 in females (Table 1; eFig. 4 I.a & II.b). Of the 524 overlapping proteins in both sexes, nearly all (> 95%) associations were directionally consistent, but stronger in females (r = 0.62; eFig. 5). Among females, FSHB, ELN, EDA2R, CXCL17, and LTBP2 showed the strongest positive associations with age, while PAEP, SDK2, CTSV, IGDCC4 and STC2 showed the strongest negative associations (eFig. 4 I.a). Among males, EDA2R, ELN, LTBP2, KLK4, and GDF15 showed the strongest positive associations, and EGFR, INSL3, IGFBP3, ERBB3, and TNFSF11 showed the strongest negative associations (eFig. 4 II.a). Among females the top age-related proteins were predominantly associated with menopause, while among males they were mainly associated with clinical measurements (i.e., BMI, SBP, and RPG), exhaled CO and current smoking (eFig. 4 I.b & 4 II.b). Additionally, 394 and 352 age-related proteins in females and males, respectively, were not associated with other exposures (eTable 8). In overall analyses, regular alcohol consumption and current smoking were associated with 75 and 53 proteins (4 overlapping), respectively, with MAMDC4, CHI3L1, VWA1, VCAN, and FGF21 most strongly associated with alcohol drinking and CXCL17, LAMP3, ALPP, MSLN and SCGB3A1 with smoking (Table 1, eFigure 6). Among men, most of these protein associations with alcohol (n = 53) and smoking (n = 47) were significant (eFigure 7). Overall, outdoor temperature was associated with 292 proteins (e.g., SNED1, SPINK6, LRP1, STX7, WFDC12), reducing to 174 in female- and 57 in male-specific analyses (eFigs. 6 and 7).

Among clinical measurements, BMI was associated with the largest number of proteins (869) followed by SBP (n = 479) and RPG (n = 387; 274 overlapping with prevalent diabetes; Table 1). In mutually adjusted models, the number of significant associations was reduced (eTable 5, eTable 6), especially for SBP (n = 157) and RPG (n = 54). For BMI, the strongest associations were with LEP, FABPA, IGFBP2, CK-BB, and GHR (eFigure 8a). In sex-specific analyses, BMI was associated with 576 proteins in females and with 353 in males, with similar strength of associations among 300 overlapping proteins (r = 0.97; Table 1 and eFig. 4b). The leading BMI-related proteins demonstrated similar profiles, with > 90% of them also associated with age, sex, prevalent diabetes, and other clinical measurements in overall and sex-specific analyses (eFigs. 8b, 9 I.b & II.b).

Overall, prevalent diabetes and HBsAg positivity were associated with 340 and 282 proteins, respectively (Table 1), with more associations in females than males (217 vs. 39 and 198 vs. 45, respectively; Table 1 and eFig. 7 h). These were reduced to 212 proteins for prevalent diabetes but increased to 292 for HbsAg positivity in the mutually adjusted models (eTables 5, 6). Other health-related measures, including self-rated health and prior cancer, were associated with fewer than three proteins. For females, post-menopause was associated with 181 proteins, with the strongest positive associations being with FSHB, CGA, DPP4, CHAD, and COL1A1 and the strongest inverse associations with PAEP, CHRDL2, SDK2 and C1QA, and RSPO1 (eFig. 10a). Additionally, the top menopause-related proteins were predominantly associated with age (eFig. 10b).

Lifestyle index was associated with 342 proteins (e.g., IGSF9, VWA1, LEP, GGT1, CES1) in overall analyses, and with 126 and 99 proteins in male- and female-specific analyses (Table 1; Fig. 4I.a, eFig. 11 I.a & II.a), which was reduced to 188 proteins overall in the mutually adjusted model. Most of these proteins were also associated with the individual components of the index, particularly age, sex and clinical measurements, with alcohol and smoking being particularly notable among males (Fig. 4I.b, eFig. 11 I.b & II.b).

Overall, the frailty index was associated with 597 proteins (e.g., IGSF9, CDHR2, IL1RN, CES1, FABP4), with 300 (50%) proteins overlapping with the lifestyle index, albeit in opposite directions (r=-0.91; Table 1; Fig. 4c II.a). In the mutually adjusted model, there was only a small reduction in the number of significant proteins (n = 573). In sex-specific analyses, 102 and 279 proteins were associated with frailty index in females and males, respectively (eFig. 13 I.a & II.a). Many of the frailty index-related proteins were also associated with individual components of the index, including age, sex, clinical measurements, and prior diseases, with similar patterns in men and women (Fig. 4 II.b, eFig. 12 I.b & 12 II.b).

In UKB participants (mean age 57 [SD 8.1], 55% female), 1,871 proteins were significantly associated with sex, 1,659 with age, 2,144 with BMI, 1,658 with SBP, 702 with RPG, and 1,302 with prevalent diabetes (Fig. 5). Over 90% of the significant associations in CKB were replicated in UKB, except of RPG, which had a lower (~ 80%) replication rate. However, > 95% of RPG associations in CKB were replicated in UKB using HbA1c (eFig. 13). Moreover, overlapping significant proteins in CKB and UKB showed high correlations (r > 0.80) in effect sizes, with > 95% being directionally consistent. In sex-specific analyses, there were similar patterns and replication rates between the two populations (eFig. 14).

In this exposome-wide analysis of Olink proteins in Chinese adults, we identified a large number of associations between various exposures and levels of ~ 3,000 proteins. In particular, age, sex, BMI and frailty index each showed significant and apparently independent associations with plasma levels of 450–750 proteins. A range of other exposures including: socio-demographic, environmental factors, clinical measurements, health-related traits, and composite lifestyle and health indices, were also associated with levels of modest numbers of proteins. Many proteins were associated with multiple exposures, including 25 that showed associations with > 10 exposures. We also observed differences in proteomic-exposure associations between females and males, and replicated > 90% of proteomic associations with key exposures in the European populations.

A large number of proteins in our study showed associations with several different exposures, including some that remained associated with > 10 exposures even after adjustments for relevant covariates. These proteins may reflect the general and pervasive impact of the exposome on human health, as reflected by their biological functions. For example, CDHR2 is involved in the surface structure of epithelial cells, which are present in various human organs and related to cancer aetiology [32, 33]. Both ADGRE2 and ADGRD1 are adhesion G protein-coupled receptors involved in cell adhesion and signalling, with ADGRE2 particularly implicated in inflammatory responses of immune system cells [34, 35]. ACY1 is an enzyme that plays a key role in protein breakdown and amino acid salvage, and its deficiency has been linked to metabolic imbalance and developmental delays [36, 37]. Finally, MEGF9 is a transmembrane protein involved in cell adhesion and signalling, with evidence suggesting its role in neurodevelopment [38].

Among all exposures investigated, age yielded the most significant associations with plasma protein levels. The protein most strongly and positively associated with age was EDA2R, a member of the tumour necrosis factor receptor superfamily [39]. EDA2R is involved in cell signalling and tissue development, and its gene expression has been previously linked to ageing in plasma, muscle, and lung tissues [40‐42]. Older age was also associated with higher levels of ELN, a protein making up elastic fibres in various human organs, including the skin, heart, and blood vessels [43‐46]. Consistent with this finding, we also found a positive association between older age and higher levels of LTBP2, a component of micro-fibrils that interact with ELN [47]. Many age-related proteins were also significantly associated with other exposures. For example, higher levels of CXCL17, a protein involved in homeostasis at mucosal barriers and inflammatory response in respiratory diseases [48, 49], were associated with older age, smoking, and amount of exhaled CO in our study.

Our analyses in the main and mutually adjusted models demonstrated marked sex differences in plasma levels of proteins, including many involved in human reproductive processes, such as FSHB, which regulates follicular growth in females [50, 51], and ACRV1, EDDM3B, and TEX101, which are involved in spermatogenesis in males [52‐54]. We replicated the well-known finding of elevated LEP levels in females, a protein released by adipocytes regulating appetite and metabolism, which has been related to sex differences in body fat percentage/distribution [8, 10, 11]. Moreover, we found elevated levels of the XG protein in females, an antigen that defines the Xg blood group [55, 56]. Historically, the Xg blood group has received less research attention, but its strong associations with sex, age, and BMI in our study suggest a need for further investigation into its role in health and disease aetiology. Many of the sex-associated proteins were also independently associated with other exposures, including FSHB, which was also associated with age. Additional analysis revealed this association was primarily driven by female sex, since FSHB regulates the growth of ovarian follicles and elevated FSHB levels are an indicator of menopause [57, 58].

Our analyses also replicated several exposure-protein associations previously reported in Europeans. For example, BMI was associated with 675 proteins in the mutually adjusted model, including known associations of higher BMI with higher levels of LEP (regulating energy balance) [59] and FABP4 (lipid transporter in adipocytes) [60]. Other proteomic associations with clinical measurements included higher SBP and lower REN levels (part of the renin-angiotensin system that regulates blood pressure and fluid balance) [61] and higher RPG and lower LPL levels (involved in lipid metabolism by breaking down triglycerides) [62]. Consistent with previous study findings, we found certain behavioural factors may also be associated with plasma protein levels, such as smoking, associated with higher CXCL17 levels (involved in inflammatory response in respiratory diseases) [48, 49] and alcohol drinking, associated with higher CHI3L1 levels (involved in inflammatory response and liver fibrosis) [63]. However, we found little association of diet and physical activity with plasma levels of proteins. Sample collection factors, including outdoor temperature and fasting time, were also significantly and independently associated with 253 and 83 proteins, respectively. For example, higher outdoor temperature was associated with higher SPINK6 levels (maintaining skin homeostasis and restricting influenza activation) [64, 65], and shorter fasting time was associated with higher GIP levels (stimulating insulin secretion) [66]. These factors may confound exposure-protein associations and should be considered in analyses. Future blood-based studies should also collect relevant information to enhance the robustness and reliability of analyses.

In addition to individual exposures, we identified novel associations between many plasma proteins and two composite measures reflecting general lifestyle and health (i.e., lifestyle index and frailty index). The indices showed opposing associations with some proteins, including IGSF9, IGFBP2, and SHBG. IGSF9 and IGFBP2 are implicated in multiple cancers and are considered potential diagnostic or prognostic markers and treatment targets [67‐72]. SHBG, a liver-produced protein that regulates sex hormones [73] is associated with metabolic and reproductive system disorders [74‐77]. As expected, many proteins associated with the two indices also showed associations with individual exposures, especially BMI, SBP, and RPG. Sex differences were also observed in proteomic associations with the two indices. For example, lifestyle index showed stronger inverse associations with GGT1 (a liver function marker) [78, 79] and CXCL17 (involved in lung function) [48, 49] in males than females, which may be due to the higher prevalence of alcohol drinking and smoking in males than females (37.2% vs. 2.6% and 63.3% vs. 2.3%, respectively) in CKB and the general Chinese population [80‐82].

Apart from leveraging an East Asian population, the main strengths of this study include the large number of proteins assayed, the wide range of exposures considered simultaneously, and the exclusion of participants with prior cardiovascular disease to reduce bias from pre-existing conditions. Consequently, the findings are most directly generalizable to adults free of cardiovascular disease. Additionally, we examined potential sex differences in protein-exposure associations, revealing potential novel findings to inform future research. Nevertheless, the study also had limitations. First, although it is the largest proteomics study in East Asians to date, power was limited for rare binary exposures (e.g., liver disease, cancer, and mental disorder), but was adequate (> 80%; Cohen’s d = 0.3) for binary exposures with prevalence above 25% and high (> 95%; Cohen’s f²=0.023) for continuous exposures. Nonetheless, the study’s findings provide valuable insights in a hypothesis-generating context, although limited power precluded more detailed subgroup analyses beyond sex-specific comparisons. As such, the smaller number of males compared to females might explain fewer significant associations in males observed in our analyses. Second, the cross-sectional study design and lack of repeated measures prevent confirmation of the direction of observed associations. Third, despite extensive adjustment for key covariates to minimise confounding, residual confounding may still persist. Fourth, recall bias in self-reported variables, such as prior diseases and diet, may have affected some associations. However, recall bias is likely to be limited, as three resurveys (~ 5 years apart), each involving ~ 25,000 randomly selected participants from the full CKB cohort, included ~ 15,000 individuals who participated in all three, enabling longitudinal comparisons that suggest good consistency in self-reported data [24, 25]. Therefore, findings should be interpreted with caution, and future studies incorporating longitudinal data and using genetic approaches (e.g., Mendelian Randomisation) are required to clarify causality. Finally, the lack of similar datasets with proteomics data prevented the replication of our findings in independent East Asian cohorts. However, the findings in the current study are also broadly consistent with a study using ~ 7,000 proteins measured by the SomaScan platform in the same sample [83]. A number of top associations were found in both platforms, such as the association between sex and LEP, as well as the association between frailty index and SHBG [83]. The SomaScan study also offers complementary information to the current study, as it covered proteins unique to the SomaScan platform and assessed the influence of other platform-specific factors, such as data normalisation procedures and dilution factors [83]. In addition, over 90% of associations with sex, age, BMI, SBP and diabetes-related exposures were replicated in UK Biobank, supporting generalisability of our findings using the same Olink platform.

Overall, the present study in Chinese adults demonstrated a large number of proteomic associations across a diverse range of exposures, particularly sex, age, adiposity, and frailty indices. We also identified sex differences in proteomic associations with various exposures, mainly reflecting differences in reproductive processes and lifestyle habits between females and males. Future studies from diverse cohorts are still required to replicate our findings, which may guide biomarker discovery when combined with clinical studies. Future analyses using penalisation methods, such as LASSO and elastic net, may be beneficial for building prediction models and reducing the high dimensionality of the exposome and proteomic data [83]. The integration of genetic data with proteomics, such as Mendelian Randomisation and colocalization analyses, can also confirm the causal relevance of these associations and explore biological mechanisms linking exposures to specific proteins.