The interplay between inflammatory cytokines and cardiometabolic disease: bi-directional mendelian randomisation study

Study design

Figure 1 is a schematic presentation of the study design. Firstly, we conducted genome-wide association studies on 47 cytokines, in up to 13 365 individuals, available in at least one of three Finnish cohorts: Northern Finland Birth Cohort 1966,21–23 Young Finns Study,24 or FINRISK study25 (online supplemental table 1). We used these estimates of genetic associations in a mendelian randomisation to investigate the effects of 15 cardiovascular, anthropometric, lipid, or glycaemic phenotypes (online supplemental table 2) on circulating cytokine concentrations. To generate instrumental variables for the cytokines, we integrated publicly available summary statistics from genome-wide association studies of 21 735 individuals and integrated data for expression quantitative trait loci (eQTL) from 15 201 samples across 49 tissues in 838 individuals. These variables were of biological relevance to the cytokines under consideration through their presence at a relevant gene locus and an association with the corresponding circulating protein, or with the gene expression level. These instruments were used in a two sample mendelian randomisation to investigate the effects of the circulating cytokines on each other. Additionally, the genetic associations were used to investigate the effects of the circulating cytokines on the cardiovascular phenotypes in mendelian randomisation and colocalisation analyses. Further details of the genome-wide association studies, mendelian randomisation, and colocalisation are given in the online supplemental methods.

Request permission

Request permission

Genome-wide association studies on circulating cytokines

To obtain genetic association estimates for the cytokines considered, genome-wide association studies were conducted separately in each of the three Finnish studies (Northern Finland Birth Cohort 1966, Young Finns Study, FINRISK), assuming an additive genetic model. The results were pooled via an inverse variance weighted fixed effects meta-analysis.

Instrument selection

The genetic associations for the cardiometabolic traits were obtained from the summary statistics derived from the large scale genome-wide association studies (online supplemental table 2). Instruments for each cardiometabolic trait were selected as single nucleotide polymorphisms (SNPs) that associated with that trait at P<5×10^-8 (P<1×10^-6 for fasting insulin) and were uncorrelated (r ²<0.001). The inverse variance weighted method was applied as the main mendelian randomisation analysis. We also conducted analyses using weighted median, MR-Egger, and MR-PRESSO methods that are more robust to violations of mendelian randomisation assumptions due to horizontal pleiotropy.19

To define the instrumental variables for circulating cytokine levels modelled as exposures in a mendelian randomisation, we enhanced our genome-wide association studies data with additional publicly available summary statistics from the INTERVAL study16 and SCALLOP consortium,17 as well as cross-tissue gene expression associations from the Genotype-Tissue Expression project (version 8).26 We generated the genetic instruments using two different criteria (online supplemental table 3). In the first approach, we selected variants within a deviation of plus or minus 500 kilobases of their corresponding gene locus that associated with corresponding circulating cytokine levels at P<1×10^-4, which we term cis-protein QTL (cis-pQTL) instruments. In the second approach, we chose variants within plus or minus 500 kilobases of the corresponding gene locus that associated with both gene expression aggregated across tissues at P<1×10^-4, and circulating cytokine levels at P<0.05, which we term cis-expression QTL (cis-eQTL) instruments. The cross-tissue gene expression associations were further compared with blood expression associations.

Mendelian randomisation and colocalisation

Mendelian randomisation analyses investigated the effects of cardiometabolic traits on circulating cytokine concentrations, the effects of circulating cytokines to other cytokine concentrations, and the effects of circulating cytokine concentrations on cardiometabolic phenotypes.

The mendelian randomisation analyses that had circulating cytokine concentrations as exposures were performed by use of the two sets of instruments (cis-pQTL and cis-eQTL). The ratio method (if one instrument available) or inverse variance weighted method (if two or more instruments available) was applied as the main mendelian randomisation analysis, complemented with weighted median, MR-Egger, and MR-PRESSO methods to investigate whether the results were driven by pleiotropic effects.19 In our secondary analyses, we also performed the same mendelian randomisation analysis by selecting instruments as uncorrelated (r²<0.001) variants from across the genome, associating with the cytokine concentrations at P<5×10^-8. The Pearson correlation coefficient estimated the similarity between mendelian randomisation estimates obtained for the same exposure and outcome associations as with the two main different instrument selection criteria (cis-pQTL and cis-eQTL). To account for multiple testing, we applied a Bonferroni correction for the number of outcomes, yielding a P value threshold of 0.05/47=0.0011 when cytokines were outcomes, and 0.05/15=0.0033 when the cardiometabolic phenotypes were outcomes.

For cytokines that generated mendelian randomisation evidence of having causal effects on concentrations of other cytokines, we conducted colocalisation. Specifically, we investigated the posterior probability (PP) of a shared causal variant (PP_shared) with the exposure and outcome cytokine concentrations at the exposure gene locus, and thus, any evidence for colocalisation would further support a causal association.27 28 By contrast, a high PP for distinct causal variants (PP_distinct) would suggest genetic confounding. PP_shared+PP_distinct >0.5 and PP_shared /(PP_shared+PP_distinct) >0.5 was considered as evidence for colocalising signals. Similar colocalisation analysis was done for the cytokine and cardiometabolic phenotype pairs with nominal evidence for association in mendelian randomisation (P<0.05).

Genome-wide association studies on inflammatory cytokines

The genome-wide association studies results for circulating levels of the 47 cytokines are presented in online supplemental figure 1. The obtained genetic association estimates were used for the subsequent mendelian randomisation analyses.

Effect of cardiometabolic traits on inflammatory cytokines

By use of mendelian randomisation, we found positive associations (P<0.0011) for genetically proxied body mass index, waist circumference, systolic blood pressure, high and low density lipoprotein cholesterol concentrations, total cholesterol levels, triglycerides, and smoking liability with circulating concentrations of at least one cytokine (figure 2, online supplemental figure 2, online supplemental tables 4 and 5). The absolute values of these effect sizes (|β|, per 1 standard deviation increase in the genetically proxied exposure) varied from 0.08 (lowest absolute value of the effect sizes) to 0.48 (highest absolute value of the effect sizes). In particular, genetically predicted body mass index was associated with circulating levels of 10 biomarkers, namely, active plasminogen activator inhibitor-1 (PAI1), beta nerve growth factor (βNGF), C reactive protein (CRP), cutaneous T cell attracting chemokine (CTACK), hepatocyte growth factor (HGF), interleukin (IL) 16, monocyte chemoattractant protein-1 (MCP1), soluble E-selectin, soluble intercellular cell adhesion molecule-1 (ICAM1), and tumour necrosis factor related apoptosis-inducing ligand (TRAIL), with |β| for these associations between 0.14 and 0.43. Genetically proxied concentrations of low density lipoprotein cholesterol were associated with circulating levels of IL6 (β=0.12, 95% confidence interval 0.05 to 0.19) (figure 2, online supplemental figure 2, online supplemental table 5).

The effect of circulating cytokines on other circulating cytokines

The cis-pQTL instruments were available for 30 cytokines and cis-eQTL instruments for 20 cytokines, with both types of instruments available for 19 cytokines (online supplemental tables 3, 6, and 7). For the cis-eQTL instruments, the Pearson correlation between cross-tissue expression z scores and blood expression z scores was 0.38 (95% confidence interval 0.01 to 0.66) (online supplemental table 7). Mendelian randomisation evidence (P<0.0011) showed associations between 34 different cytokine-cytokine pairs when using the cis-pQTL or cis-eQTL instrument selection criteria, with |β| ranging from 0.07 to 1.71 (figure 3, online supplemental figure 3, online supplemental table 8). The most associations were noted for genetically predicted vascular endothelial growth factor (VEGF) with circulating levels of ten (|β| from 0.14 to 1.71) other cytokines and IL18 with six (|β| from 0.14 to 0.26) other cytokines (figure 3). When comparing the results for the two types of instrumental variables, we found a strong positive correlation between the cis-pQTL main mendelian randomisation estimates and the cis-eQTL main mendelian randomisation estimates (Pearson correlation r=0.73, online supplemental figure 4).

Request permission

Colocalisation provided further evidence supporting causality (PP_shared /(PP_shared+PP_distinct) >0.5) for 20 exposure-outcome pairs of cytokines (figure 3, online supplemental table 9; online supplemental figures 5a-t), notably circulating VEGF concentrations (at the VEGF gene) with levels of seven other cytokines: interferon gamma (IFNγ), IL5, IL7, IL10, IL12p70, IL13, and macrophage inflammatory protein-1-alpha (MIP1α).

The effect of circulating cytokines on cardiometabolic traits

By use of the cis-pQTL instrument selection criteria, the evidence of association for 24 cytokine and outcome pairs was strong (P<0.0033; figure 4, online supplemental figure 6, online supplemental table 10). The odds ratio for coronary artery disease risk per 1 standard deviation increase in genetically proxied MCSF was 1.13 (95% confidence interval 1.06 to 1.20). Using the cis-eQTL instruments, mendelian randomisation showed evidence of an association for 16 cytokine and outcome pairs (figure 4; odds ratio per 1 standard deviation increase in genetically proxied soluble ICAM1 on type two diabetes mellitus risk was 0.79 (95% confidence interval 0.67 to 0.92).

Request permission

The cis-pQTL and cis-eQTL methods provided similar mendelian randomisation estimates (Pearson’s correlation coefficient r=0.66, P=4.8×10^-35, online supplemental figure 7). Analyses of both cis-pQTL and cis-eQTL instruments showed evidence for association (P<0.0033) for genetically increased IL1RA concentrations with decreased CRP concentrations (cis-pQTL β=−0.15 (95% confidence interval −0.18 to −0.13); cis-eQTL β=−0.13 (−0.17 to −0.09)) and increased total cholesterol concentrations (cis-pQTL β=0.05 (0.04 to 0.06); cis-eQTL β=0.05 (0.02 to 0.08)); additionally, for genetically increased macrophage migration inhibitory factor (MIF) concentrations with decreased CRP levels (cis-pQTL β=−0.08 (−0.12 to −0.04); cis-eQTL β=−0.07 (−0.10 to −0.04)), increased high density lipoprotein cholesterol (cis-pQTL β=0.03 (0.01 to 0.04); cis-eQTL β=0.03 (0.01 to 0.05)), and increased triglycerides concentrations (cis-pQTL β=0.06 (0.04 to 0.08); cis-eQTL β=0.07 (0.05 to 0.08)). For the 15 cytokines where three or more genetic instrumental variables were available (online supplemental table 3), weighted median, MR-Egger, and MR-PRESSO sensitivity analyses produced consistent mendelian randomisation estimates to the main inverse-variance weighted analysis (online supplemental methods, online supplemental figure 8, online supplemental table 10). This suggests that pleiotropic associations of the genetic variants are unlikely to be significantly biasing the mendelian randomisation estimates. The mendelian randomisation results that used genome-wide selection of the instruments for the cytokines (online supplemental table 4) provided evidence of association (P<0.0033) for 16 cytokine-outcome pairs. Most notably, genetically proxied circulating concentrations of soluble E-selectin were associated with eight cardiometabolic traits (|β| from 0.01 to 0.07; online supplemental table 11 and online supplemental figure 9).

In colocalisation analyses, evidence suggests a shared causal variant (PP_shared+PP_distinct >0.5 and PP_shared /(PP_shared+PP_distinct) >0.5) for 19 cytokine and outcome pairs. Notably, IL1RA colocalised with five cardiometabolic traits (CRP, low density lipoprotein cholesterol, total cholesterol, triglycerides, and glucose), and coronary artery disease risk colocalised with MCSF and TRAIL (figure 4, online supplemental table 12; online supplemental figures 10a-s). Thus, MCSF had both mendelian randomisation and colocalisation evidence to support effects on increasing coronary artery disease risk. For the other cytokines with nominal mendelian randomisation evidence (<0.05) for coronary artery disease risk, TRAIL also had evidence for colocalisation with coronary artery disease risk (PP_shared /(PP_shared+PP_distinct) >0.99). Although colocalisation of IL1RA and coronary artery disease risk showed stronger support for shared causal variant than distinct variants (PP_shared /(PP_shared+PP_distinct) >0.91), the statistical power was insufficient to show conclusive evidence against genetic confounding (PP_shared+PP_distinct=0.42; online supplemental table 12; online supplemental figure 10t).

Principal findings

Using novel data and approaches, we offer comprehensive genetic insight into the determinants, cascades, and effects of circulating cytokines in relation to cardiometabolic traits. Our approach generated genetic evidence for effects of obesity measures, hypertension, lipid concentrations, and smoking, on levels of circulating inflammatory cytokines, cytokine cascades underlying systemic responses, and a further three putative therapeutic targets for coronary artery disease (TRAIL, IL1RA, and MCSF).

Obesity, hypertension, and smoking are leading preventable threats to global health.29 Our results suggest that higher body mass index increases multiple mediators of inflammation that affect various processes, including thrombosis (via plasminogen activation inhibitor-1), metabolism (HGF), and endothelial dysfunction (MCP1, TRAIL, ICAM1, and soluble E-selectin). These results for body mass index were corroborated by similar associations for waist circumference and waist to hip ratio. Our analyses also suggested that cigarette smoking contributes to elevating CRP and MCP1 levels, higher systolic blood pressure elevates CRP levels, and low density lipoprotein cholesterol elevates circulating IL6 levels. Taken together, these risk factors seem to increase cardiovascular disease pathogenesis at least partly through inflammatory mediators.30

Comparison with other studies

To better understand cytokine regulatory networks, we also examined the associations of genetically predicted circulating cytokines with levels of other cytokines. Our approach suggests complex relations between circulating cytokines, with VEGF appearing as a master regulator. Indeed, VEGF signalling is already targeted clinically in the treatment of certain cancers and ophthalmic conditions.31 32 Our findings are consistent with an earlier report identifying VEGF as an upstream controller of IL12p70, IL7, IL10, and IL13.14 We replicated and extend these findings to show a wider range of cytokines within this cascade, including drivers of type two immune responses (IL5 and IL13), Jak-STAT cytokine signalling (IFNγ and IL12), and immune modulation (IL10). Although several of these inflammatory cytokines are directly targeted by biological drugs used in routine clinical practice or late stage clinical trials, further research is required to ascertain their potential benefit in the setting of cardiovascular disease. As such, therapeutic targeting of the pathways responsible for the expression or biological signalling of these cytokines might also be possible.

Considering the cardiometabolic outcomes under study, we identified consistent mendelian randomisation and colocalisation evidence for circulating TRAIL increasing coronary artery disease risk; IL1RA and MIF decreasing CRP levels; and IL1RA increasing circulating low density lipoprotein cholesterol, total cholesterol, and glucose concentrations. TRAIL is a ligand involved in initiating apoptosis that has previously been implicated in atherosclerosis.33–35 Our current work extends on this previous work by use of genetic data that supports that higher circulating TRAIL concentrations causally increase coronary artery disease risk, further implicating TRAIL as a therapeutic target. We also found evidence that higher body mass index causally increases circulating TRAIL levels, thus implicating TRAIL as a mediator in the effect of body mass index on coronary artery disease risk. Aligning with earlier work, we showed that variants within the IL1RN gene that increase IL1RA are positively associated with total cholesterol, low density lipoprotein cholesterol, and glucose, but negatively with CRP concentrations,36 which adds support for IL1RA as a therapeutic target in coronary artery disease. Our mendelian randomisation analysis similarly implicated MCSF as a drug target for reducing coronary artery disease risk, triangulating with evidence obtained in studies of mice.37 Although the association between IL1RA and coronary artery disease risk was not robustly supported in our colocalisation analysis, the absence of colocalising signals probably reflects insufficient power rather than genetic confounding. Of relevance, anakinra, which is a recombinant and modified human IL1RA protein, is already used in the treatment of rheumatoid arthritis.38

The mendelian randomisation results using genome-wide selection for instruments provided distinct results to the main results. In particular, genetically proxied circulating soluble E-selectin levels were associated with eight cardiometabolic traits, mostly driven by strong associations in the ABO locus.15 Selection of instruments across the full genome allow for detection of trans-QTL proxies for cytokines. However, this strategy is also likely to include variants that are not specific only to the relevant cytokine, therefore, potentially mis-specifying the exposure and introducing pleiotropic effects violating the mendelian randomisation assumptions.

Strengths and limitations of this study

In our analyses, we aimed to maximise the validity of instruments selected to proxy levels of circulating cytokines by considering only variants that were located at the corresponding gene locus, an established strategy for selecting instruments when investigating drug effects.20 39 Our novel use of both pQTL and eQTL instrument selection criteria offered complementary evidence, and we further used colocalisation analyses to strengthen the evidence for causal effects. However, our work also has limitations. We could not identify suitable genetic proxies for all considered cytokines, and further we may have missed some associations due to insufficient statistical power and the correction imposed for multiple testing. Of note, we could not identify the support for targeting IL6 and MCP1 to reduce cardiovascular disease risk that has been demonstrated in previous mendelian randomisation studies.40 41 The discrepancy for IL6 might be attributable to our approach identifying genetic proxies at the gene for the ligand, rather than its receptor. As a further limitation, mendelian randomisation analyses should not be directly extrapolated to infer the effect of a clinical intervention, as the instruments employed represent the cumulative effect of lifelong genetic predisposition, while a clinical intervention typically represents a discrete event at a particular time point.42 Finally, our genetic data were predominantly obtained from individuals of European ancestry and it is unclear whether these findings will extend to individuals of other ancestries.

Conclusions

By leveraging large scale genetic data, this work offers deeper insight into causal relations between cardiometabolic traits, circulating cytokines, and cardiovascular disease. Our findings replicate previously described therapeutic targets and further identify potential pharmacological opportunities, including perturbation of circulating TRAIL, IL1RA, and MCSF for reducing coronary artery disease risk.