Selection of genetic variants related to cardiometabolic traits and schizophrenia

For fasting insulin, triglycerides, and HDL, we used a set of 53 single nucleotide polymorphisms (SNPs) reported to be associated with all 3 traits, representative of an insulin resistance phenotype, from a recent meta genome-wide association study (GWAS) of 188,577 European adults, which adjusted for BMI [11]. In our study, we included SNPs reaching genome-wide significance for the corresponding trait. Summary statistics for genome-wide significant SNPs were also obtained for 6 related continuous (FPG, HbA1C, LDL, BMI, leptin, and glucose tolerance) and 1 binary (T2DM) cardiometabolic traits from recent large GWAS (S2–S10 Methods). We obtained summary statistics for schizophrenia from a recent GWAS from the Psychiatric Genomics Consortium (PGC) [24] based on 40,675 cases and 64,643 European controls. The degree of sample overlap between exposure and outcome samples was likely to be low since exposure and outcome data were obtained from different consortia [25].

Ethics statement

Our study was a secondary analysis of the above publicly available data. Informed consent was sought for all participants per the original GWAS protocols, and all ethical approvals for the GWAS were obtained by original GWAS authors.

Statistical analysis

The analysis plan was prospectively conceived by the authors in 2019 but was not formally deposited in a repository or database. All described analyses were planned a priori except for the following: a) the analysis of inflammation-related SNPs at a less-stringent significance threshold (see the ‘Analysis using inflammation-related SNPs’ section below); b) the MVMR analysis including CRP (see the ‘Adjustment for Inflammation’ section below). These analyses were conceived and conducted in light of findings from the primary analysis, to further probe whether inflammation could explain the results. We obtained summary-level data (SNP rs number, β-coefficient or log odds ratio (OR), standard errors or 95% confidence intervals (CIs), effect allele, other allele, p-value, effect allele frequency, sample size, and number of cases/controls) from each GWAS. Where a specific instrument SNP was not available in the outcome dataset, we located proxy SNPs using linkage disequilibrium (LD) tagging (r² > 0.8) via LDlink [26]. Alleles were harmonised based on matching alleles, and the resulting instruments were clumped for LD to ensure independence (10,000 kb pairs apart, r² < 0.001). In the event of palindromic SNPs, the forward strand was inferred where possible using allele frequency information. We performed bidirectional analysis (i.e., with schizophrenia as exposure and cardiometabolic traits as outcomes) to examine direction of association. Statistical analysis was conducted using the TwoSampleMR package (v0.5.4) [27] for R (The R Foundation for Statistical Computing, Vienna, Austria) [28]. Our primary MR analysis method was inverse variance weighted (IVW) regression where at least two exposure SNPs were available for analysis. Where one exposure SNP was available for analysis, we used the Wald ratio method. We also conducted weighted median and MR–Egger regression analysis (S11 Methods). For the binary outcome of schizophrenia, the estimates for continuous exposures (fasting insulin, HDL, triglycerides, LDL, FPG, BMI, HbA1C, glucose tolerance, and leptin) represent log-odds ratios converted into ORs, representing the increase in risk of schizophrenia per standard deviation (SD) of exposure, and 95% CIs. For binary exposures (T2DM), the estimates represent the OR for schizophrenia per unit increase in the log-odds of T2DM. For continuous cardiometabolic outcomes, β-coefficients represent the SD increase in exposure per unit increase in the log-odds of schizophrenia, with standard errors (SEs).

We performed several sensitivity analyses to check the validity of our results. Heterogeneity among SNPs included in each analysis was examined using the Cochran Q test. We checked for horizontal pleiotropy using the MR–Egger regression intercept alongside a more recent and robust method to detect horizontal pleiotropy and outliers, “MR pleiotropy residual sum and outlier” (MR-PRESSO) [29]. Using MR-PRESSO, we used the global test to examine for horizontal pleiotropy, and where evident, used the method to correct the IVW-estimate via outlier removal (S11 Methods). We examined for measurement error in SNP-exposure associations using the I²_GX statistic [30]. This study is reported as per the Strengthening the Reporting of MR studies (STROBE-MR) guideline [31] (S1 Checklist) and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement [32] (S2 Checklist).

Analysis using inflammation-related SNPs

Next, we repeated MR analysis using only inflammation-related SNPs for each cardiometabolic risk factor as an instrumental variable for the outcome of schizophrenia. We did this to test the hypothesis that these SNPs may represent a mechanism involving inflammation. This could be via, for example, a common causal basis (panel A in S1 Methods) or via vertical (mediating) pleiotropy [27] (panel B in S1 Methods). We used Phenoscanner v2 (University of Cambridge, United Kingdom) [33] to examine each SNP associated with each cardiometabolic risk factor, to identify SNPs that were also associated with a measure of inflammation, defined as blood concentration/count of cytokines (such as chemokines, interferons, interleukins, lymphokines, or tumour necrosis factors), acute phase proteins (e.g., CRP), or immune cells (e.g., neutrophils and lymphocytes). Primarily, we considered inflammation-related SNPs at genome-wide significance (p<5×10^-8) to maximise specificity. We also performed a sensitivity analysis by including inflammation-related SNPs at a less-stringent nominal significance threshold (p<1x10^-4) to increase sensitivity to inflammation-related SNPs [34] (S12–S17 Methods).

Using the same method, we identified inflammation-related schizophrenia SNPs (S18 Methods) and used them as instrumental variables in MR analysis examining cardiometabolic traits as outcomes.

Adjustment for inflammation

As a post hoc sensitivity analysis to estimate whether any associations evident above may be explained by inflammation, we carried out MVMR analysis [34,35] using the 53 SNPs for fasting insulin, triglycerides, and HDL, representative of an insulin resistance phenotype, as exposures with schizophrenia as the outcome, after conditioning on the associations of these 53 SNPs with CRP. We chose CRP because it is a widely used downstream measure of systemic inflammation, and publicly available data from large-scale GWAS for CRP are available. Summary statistics for CRP were obtained from a recent large GWAS based on 204,402 participants [36]. For CRP as an exposure in MVMR, we used all independent (10,000 kb pairs apart, r² < 0.001) SNPs reported to be conditionally associated with CRP and located within the CRP coding region (S19 Methods).

Correction for multiple testing

Statistical significance was estimated using the Holm–Bonferroni correction method [37], correcting for the number of exposures tested at each stage of analysis.

MR analyses using all genetic variants associated with insulin resistance and other cardiometabolic traits

We did not find significant evidence for associations between genetically-predicted levels of cardiometabolic traits and schizophrenia, using the primary IVW analysis method. Evidence using the weighted median method for associations between genetically-predicted levels of triglycerides (weighted median OR = 1.26; 95% C.I., 1.06–1.50; corrected p = 0.090) and HDL (weighted median OR = 0.79; 95% C.I., 0.65–0.95; corrected p = 0.126) with schizophrenia did not survive correction for multiple testing (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. MR analyses of cardiometabolic traits and schizophrenia using all SNPs.

https://doi.org/10.1371/journal.pmed.1003455.t001

MR analyses using inflammation-related genetic variants for insulin resistance and other cardiometabolic traits

After testing only genome-wide significant inflammation-related SNPs for cardiometabolic traits, we found evidence for associations of inflammation-related genetically-predicted fasting insulin (Wald Ratio OR = 2.95; 95% C.I., 1.38–6.34; corrected p = 0.035) and HDL (Wald Ratio OR = 0.55; 95% CI, 0.36–0.84; corrected p = 0.035) with schizophrenia. We could not include any genome-wide significant inflammation-related variants for triglycerides, leptin or glucose tolerance. In our sensitivity analysis featuring inflammation-related cardiometabolic variants at a less stringent significance threshold, evidence persisted for associations of inflammation-related genetically-predicted fasting insulin (IVW OR = 1.74; 95% C.I., 1.08–2.98; corrected p = 0.030) and HDL (IVW OR = 0.78; 95% C.I., 0.62–0.92; corrected p = 0.036) with schizophrenia. In addition, we found evidence for an association of genetically-predicted inflammation-related triglycerides (IVW OR = 1.24; 95% C.I., 1.07–1.55; corrected p = 0.036) with schizophrenia (Table 2; Fig 1 & Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

MR Analyses Testing Associations between Insulin Resistance Phenotypes (Fasting Insulin (A), Triglycerides (B) and HDL (C)) and Schizophrenia, Highlighting Inflammation-Related SNPs. Points in plots represent the association of the genome-wide significant insulin-resistance single nucleotide polymorphisms (SNPs) and their association with schizophrenia (Y axis) and the exposure (X axis). SNPs are denoted by green points in the plot. Inflammation-related SNPs at genome-wide significance are denoted by a purple border. Inflammation-related SNPs at nominal significance are denoted by a red border. Whiskers represent standard errors. Lines on the plot represent inverse-variance weighted (>1 SNP) or linear regression (1 SNP) of all-SNPs (green line), inflammation-related SNPs at genome-wide significance (purple line) and inflammation-related SNPs at nominal significance (red line).

https://doi.org/10.1371/journal.pmed.1003455.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. MR analyses testing associations between cardiometabolic traits and schizophrenia.

Forest plot presents ORs and 95% CIs for associations between cardiometabolic traits and schizophrenia using IVW / Wald Ratio MR analyses based on all single nucleotide polymorphisms (SNPs) associated with each risk factor (green), inflammation-related SNPs at genome-wide significance (purple), and inflammation-related SNPs at nominal significance (red). See Tables 1 and 2 for the number of SNPs used in each analysis. HDL = High Density Lipoprotein; T2DM = Type 2 Diabetes Mellitus; BMI = Body Mass Index; FPG = Fasting Plasma Glucose; LDL = Low-Density Lipoprotein; HbA1C = Glycated Haemoglobin; Glucose Tol = Glucose Tolerance.

https://doi.org/10.1371/journal.pmed.1003455.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. MR analyses of inflammatory-related cardiometabolic SNPs and schizophrenia.

https://doi.org/10.1371/journal.pmed.1003455.t002

Adjustment for inflammation

MVMR analysis for inflammation-related SNPs of fasting insulin, triglycerides, and HDL with schizophrenia showed that the univariable associations fully attenuated after controlling for the genetic associations of these variants with CRP, in analyses involving both inflammation-related SNPs at genome-wide and nominal significance levels. Controlling for CRP had negligible effect on MR estimates based on all genetic variants (Fig 3, S1 and S2 Results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Multivariable MR analysis testing associations between insulin resistance phenotypes and schizophrenia after controlling for genetic associations for CRP.

Forest plot presents ORs and 95% CIs for inverse-variance weighted regression (IVW) MR associations between insulin resistance phenotypes and schizophrenia using all single nucleotide polymorphisms (SNPs) (dark green), and after controlling for association of these SNPs with C-reactive protein (CRP) using multivariable MR (MVMR) (light green). The forest plot also presents ORs and 95% CIs for IVW / Wald Ratio MR associations between insulin resistance phenotypes and schizophrenia using inflammation-related SNPs (genome-wide significance = dark purple; nominal significance = red), and after controlling for association of these SNPs with CRP using MVMR (genome-wide significance = light purple; nominal significance = light red). HDL = high-density lipoprotein.

https://doi.org/10.1371/journal.pmed.1003455.g003

Test for bidirectionality using schizophrenia as exposure

We did not find statistically significant MR associations between schizophrenia and any cardiometabolic trait after correction for multiple testing (S3 Results, S1 Fig). Similarly, we did not find statistically significant MR associations of inflammation-related schizophrenia variants with cardiometabolic traits after correction for multiple testing (S4 Results, S1 Fig).

Test for horizontal pleiotropy

Using the MR-Egger regression intercept test, we found evidence of potential horizontal pleiotropy for BMI and HDL in the all-SNP analysis, but no evidence for horizontal pleiotropy for any cardiometabolic exposure in the inflammation-related SNP analysis. Using MR-PRESSO however, we found evidence that horizontal pleiotropy was likely to have affected estimates for all cardiometabolic exposures in the all-SNP analysis (p value for global test all ≤0.020), and both LDL and T2DM in the inflammation-related SNP analysis. Following MR-PRESSO outlier correction, evidence strengthened for the association of triglycerides with schizophrenia in the all-SNP analysis (MR-PRESSO IVW β = 0.23, S.E. 0.06, p = 0.008), but outlier-corrected IVW estimates for other exposures were not significantly altered.

In the bidirectional analyses, both MR-PRESSO and the MR-Egger regression intercept suggested horizontal pleiotropy affecting the outcomes of HDL, BMI and LDL (all p<0.05). Following outlier correction, there was evidence for a weak protective effect of schizophrenia on BMI (β = -0.04, S.E. 0.02, p = 0.014). MR-PRESSO additionally revealed possible horizontal pleiotropy affecting the outcomes of fasting insulin, triglycerides and T2DM (p for MR-PRESSO global test all <0.05) (S5–S12 Results), but outlier-corrected IVW estimates were not significantly altered.

Test for heterogeneity of instruments

In the analyses based on all SNPs, the majority of cardiometabolic traits demonstrated evidence of heterogeneity, which was reduced in the inflammation-related SNP analysis (S5–S8 Results). There was limited evidence of heterogeneity in the sensitivity analyses based on inflammation-related SNPs for T2DM, BMI, and HbA1C only.

Test for measurement error

Results for the I²_GX tests for SNP-exposure associations revealed some evidence for potential measurement error, which may have biased MR–Egger analyses in the analyses with leptin, glucose tolerance, T2DM, and schizophrenia as exposures (S13 Results).

Main findings

We conducted bidirectional uni- and multivariable two-sample MR analyses using large publicly available genomic datasets to first examine for associations that support a causal relationship between insulin resistance and related cardiometabolic traits and schizophrenia, and second, to examine whether there is evidence to support that inflammation may be a common causal mechanism for insulin resistance and schizophrenia. Using our primary IVW analysis method, we did not find evidence in support of a causal association between genetically-predicted cardiometabolic traits and schizophrenia. However, we found weak evidence using the weighted median method in support of a causal association of genetically-predicted levels of triglycerides and HDL with schizophrenia, but this association did not survive correction for multiple testing and the estimate may have been affected by horizontal pleiotropy. We found more consistent evidence for an association of the insulin resistance phenotype of fasting insulin, triglycerides, and HDL [11] with schizophrenia when we examined only genetic variants also associated with inflammation. Using two p-value cut-offs for inflammation-related SNPs, we found that the strength of association with schizophrenia increased as the specificity toward inflammation-related SNPs increased. In MVMR analyses adjusting for CRP, those estimates attenuated fully to the null. We found no evidence in bidirectional analyses in support of a causal relationship of schizophrenia with insulin resistance (panels C and D in S1 Methods). Together, our results are therefore most consistent with inflammation as a common cause for insulin resistance and schizophrenia (panel A in S1 Methods).

Inflammation as a common cause for schizophrenia and insulin resistance

Three aspects of our results point towards inflammation as a common cause for insulin resistance and schizophrenia (panel A in S1 Methods). First, we did not find convincing overall evidence for a causal relationship between insulin resistance and schizophrenia (likely ruling out panel B in S1 Methods). Second, in our analyses of inflammation-related variants for the cardiometabolic traits, we found strong and consistent evidence in support of a potential causal relationship of fasting insulin, HDL and triglycerides with schizophrenia, and the strength of association with schizophrenia increased as the specificity toward inflammation-related SNPs increased. Third, we used MVMR to evidence that after controlling for CRP, an archetypal general inflammatory marker, the associations between inflammation-related genetic variants for insulin resistance and schizophrenia completely attenuated. This result suggests that the observed associations for the inflammation-related variants are at least in part explained by inflammation. Together, the results are consistent with the idea that inflammation may be a common causal mechanism for insulin resistance and schizophrenia.

Evidence for a common causal mechanism between insulin resistance and schizophrenia may help to explain why schizophrenia is associated with higher rates of insulin resistance even in early stages of illness, when the cumulative effects of medication and lifestyle factors are relatively small [12,38]. Anti-inflammatory agents, of which several have shown promise in treating the symptoms of schizophrenia [39], should therefore be considered as a putative therapeutic target for prevention and treatment of cardiometabolic disorders in schizophrenia.

We used CRP, an archetypal downstream inflammatory marker, as a means of gauging the effect of systemic inflammation in MVMR analysis, rather than hypothesising a specific role for CRP in the relationship between insulin resistance and schizophrenia. Nevertheless, CRP has observationally shown in both cross-sectional [40] and longitudinal [41] research to be associated with schizophrenia, although such findings are limited by the potential of residual confounding and reverse causality. Interestingly, however, MR findings have reported that genetically predicted CRP may have a protective effect on schizophrenia [21], with authors positing that a genetically attenuated ability to produce CRP may predispose to more insidious and chronic infections. In our MVMR analysis, attenuation of insulin resistance schizophrenia associations after controlling for CRP is consistent with inflammation being associated with both exposure and outcome, albeit “negatively” with the latter. Further research is needed to explore potential mechanisms of association between CRP and schizophrenia.

Many of the SNPs included in the inflammation-related analysis were associated with neutrophils and/or lymphocytes. A raised neutrophil to lymphocyte ratio (NLR) is a marker of systemic inflammation and is known to be associated with schizophrenia [42] and insulin resistance [43]. We were unable to find large GWAS studies conducted in European populations for NLR or for other inflammatory markers, which we might have used in MVMR analyses in place of CRP.

Based on our findings, we are unable to completely rule out the possibility that insulin resistance may mediate an inflammation–schizophrenia association (panel B in S1 Methods), since there was weak evidence that did not survive correction for multiple testing for an association of triglycerides and HDL with schizophrenia using the weighted median method, and in our MR-PRESSO sensitivity analysis, evidence from the outlier-corrected IVW analysis suggested a possible association between triglycerides and schizophrenia. These findings are broadly similar to 1 previous MR study [17], which reported only weak evidence of an association between the homeostasis model assessment (HOMA), a measure of insulin resistance, on schizophrenia. Another MR study [16] reported a genetic association between fasting insulin and schizophrenia, although the evidence attenuated after adjustment for BMI. To account for BMI, we obtained summary statistics for genetic variants related to insulin resistance after controlling for BMI [11]. The previous MR study included an ethnically heterogeneous sample, increasing the potential for population stratification bias. We used genetic data from a more ethnically homogenous GWAS of schizophrenia [24]. Nevertheless, while our results in the all-SNP analysis suggested weak evidence for triglycerides and HDL, which may reflect an insulin resistance phenotype, the evidence did not survive correction for multiple testing and requires replication in future when larger GWAS samples are available.

The implications of our findings with regard to shared causal mechanisms should not distract clinicians from focusing on the assessment and management of malleable lifestyle factors related to cardiometabolic disorders in people with schizophrenia. Factors such as poorer diet, reduced exercise and smoking, which are associated with schizophrenia [7,44,45], may predispose to an inflammatory state [46]. Therefore, it is possible that lifestyle factors exacerbate a feedback loop between inflammation, insulin resistance, and schizophrenia by increasing both inflammation and insulin resistance, eventually leading to T2DM and other cardiometabolic disorders such as obesity and cardiovascular disease (CVD). In addition to the potential therapeutic potential of anti-inflammatory medications, malleable lifestyle factors must continue to remain crucial targets [47,48] for the prevention of cardiometabolic morbidity in people with schizophrenia.

Additional findings

We report that after outlier correction, schizophrenia had a weak protective effect on BMI. This finding complements estimates from previous research [53] using LD score regression, though we are able to advance previous findings since genetic correlation analyses are unable to test direction of association. This finding suggests that weight gain associated with schizophrenia is unlikely to be a feature of the illness itself but could be attributed to iatrogenic or lifestyle effects. Moreover, the “lean insulin-resistance” phenotype may be associated with higher levels of inflammation [54] and warrants further research in the context of schizophrenia, particularly since in younger patients, the “lean” nature of the phenotype may mean that important cardiometabolic investigations may be overlooked in the clinic.

Strengths and limitations

Strengths of this study include the use of a large set of cardiometabolic traits and large GWAS datasets, through which we were able to test specific biological mechanisms. We chose SNPs reaching genome-wide significance from large GWAS and meta-GWAS for insulin resistance and related cardiometabolic traits. We performed a comprehensive set of sensitivity analyses to check MR assumptions. Furthermore, while weak instrument bias may be a factor in MR analysis, in two-sample MR, this bias tends towards the null [55], thus would not explain the positive associations we describe. We corrected for multiple testing to minimise potential type I error.

Our study has some limitations. We did not select SNPs in known coding regions for the exposures, for example, the IRS-1 gene for insulin resistance [56]. We took this step on the assumption that many mechanisms at play may not yet be fully understood. For example, while the heritability of cardiometabolic traits such as obesity is as high as 70%, the variance currently explained by known genetic variants is but a small fraction of this [57]. In addition, selecting SNPs from many different GWAS studies featuring large sample sizes may increase the risk of sample overlap between exposure and outcome variables and can bias the results in either direction, depending on the proportion of overlap [27]. Also, for our primary inflammation-related SNP analysis, we chose a stringent p-value threshold to define inflammation-related SNPs. In doing so, we may have overlooked some SNPs with true inflammatory associations. As a result, only one genome-wide significant inflammation-related genetic variant was included in the analysis of fasting insulin and HDL, and none were included for triglycerides. Therefore, these results be considered with caution. However, we attempted to address this limitation by relaxing the p-value threshold for inflammation-related SNPs, thereby allowing a greater number of SNPs to be included, and the results for fasting insulin, HDL and triglycerides were consistent. Yet, the inclusion of inflammation-related genetic variants at a relaxed significance threshold may have increased the risk of weak instrument bias for those analyses. In the future, larger and better-powered GWAS may identify more SNPs for analysis and at greater resolution, potentially unearthing a larger number of inflammation-related SNPs, which would be helpful to confirm our findings. Additionally, the full range of gene products from the genetic variants we used as proxies for the cardiometabolic traits is unknown, and so we are unable to comment on potential biological mechanisms of association other than inflammation, which may also be relevant. Finally, our analyses were based on data from mostly European participants, so it is unclear whether our results apply to other populations. Large-scale GWAS and replication of our analyses in other populations are required to answer this question.