Dissecting the Association Between Inflammation, Metabolic Dysregulation, and Specific Depressive Symptoms

This genetic correlation and 2-sample mendelian randomization study uses large-scale genome-wide association data sources to explore the genetic overlap and associations between inflammatory activity, metabolic dysregulation, and individual depressive symptoms.


PGC MD
GWAS summary statistics for MD were retrieved from two original reports (Howard et al. 6

&
Wray et al. 7 ) to avoid sample overlap in LDSC regression and MR analyses. For the present investigation, we did not include data on participants from 23andMe.
Howard et al. 6  Covariates used by Wray et al. 7 were age, sex, and principal components as implemented in the RICOPILI pipeline. 11 Howard et al. 6 used age, sex, genotyping array, and the first eight principal components in the UK Biobank sample as covariates.
Importantly, definitions of MD from meta-analysed GWAS data differed, including a 'broad depression' definition, 8 self-reported diagnosed depression, 9 and meeting MD diagnostic criteria (Wray et al. 7 study). Recent work has emphasised that these phenotypic definitions are relevant and, in particular, that minimal phenotyping definitions such as 'broad depression' may prohibit finding specific signatures of MD. 10 While we acknowledge this limitation, we decided to use the term 'MD' to denote both the less specific depression phenotypes by Howard et al. 6 and the diagnosis-ascertained phenotype by Wray et al. 7 to make the present report more parsimonious as findings between depression phenotypes were similar.

Insomnia
GWAS summary statistics for insomnia were taken from Jansen et al. 12 , who meta-analysed data from UK Biobank (n=386,533) and 23andMe (n=944,477), resulting in a total sample size of 1,331,010 individuals. We included insomnia to provide a comparison to the PHQ-9 composite symptom of "sleep problems". Insomnia was defined in UK Biobank whenever participants answered "usually" (rather than "never/rarely", "sometimes", or "prefer not to say") to the question "Do you have trouble falling asleep at night or do you wake up in the middle of the night?".
independence between genetic variants. For the alternative IL-6 signalling instrument, SNPs with F-statistic greater than 15 (and not necessarily genome-wide significant) were clumped to the same threshold of R 2 <0.1.
Main genetic instruments were based on a recent report by Georgakis et al. 17 who investigated the association of CRP levels and IL-6 signalling on cardiovascular outcomes. Due to functional knowledge that IL-6 induces production of CRP from hepatocytes, 18 Georgakis et al. 17 used GWAS summary data for upregulated CRP levels from the CRP GWAS by Ligthart et al. 1 to define both the genetic instrument for CRP levels and for IL-6 signalling. This instrument selection strategy assessing different upstream effector molecules indexed using the same downstream readout has been extensively used in prior research. [19][20][21][22][23][24][25] Specifically, independent (R 2 <0.1), genome-wide significant SNPs within a 300kB region upstream or downstream of CRP and IL-6R genes, respectively, were selected that were associated with CRP levels. 1 We intentionally used the term "IL-6 signalling" in relation to the IL-6R genetic instrument, because the instrument was weighted based on GWAS summary data for CRP levels, which is a downstream substrate of IL-6 activity. Despite IL-6R SNP effect weighting being based on CRP levels, however, we are confident about the effects reflecting IL-6 signalling as IL-6 is an upstream inducer of CRP. When comparing genetic variants indexing increased IL-6 signalling to the Genotype-Tissue Expression (GTEx) platform 26 , we find that 3 of 6 SNPs (i.e., rs2228145, rs73026617 & rs11264224) are IL-6R-expression quantitative trait loci (eQTLs) in immune-/ vascular-relevant tissues (see eTable 3 & eFigure 1). Additionally, the strongest genetic variant (rs2228145), based on an F-statistic of 458.16, has been investigated in prior research and was shown to impair IL-6 classical signalling as the minor allele (C) reduces the expression of membrane bound IL-6R and decreases IL-6 production post-stimulation; 27 this aligns with the major allele (A) showing a positive, increasing effect on CRP levels. rs2228145 has also been associated with risk for severe depression and psychosis in a previous study. 28 Overall, these findings lend strong support for a functional role of our IL-6 signalling instrument on CRP levels via IL-6 signalling.
We compared these genetic instruments to genetic instruments used in a previous MR study by Khandaker et al. 29 in eTables 4-5 regarding LD and F-statistics, which shows that our genetic instruments include and extend information of these previously used instruments.
As alternative approaches, we used genetic variants throughout the genome that were associated with CRP levels and, based on a previous report, 15 variants within 250kB of the IL6R gene that were associated (F>15) with sIL-6R plasma levels as an indirect marker of IL-6 signalling. As sIL-6Rs are inversely associated with IL-6 signalling, we changed the effect valence of genetic variants by multiplying beta estimates by -1. We also compare main IL-6 signalling and alternative (indirect) IL-6 signalling instruments in terms of LD between included SNPs and Fstatistics (eTable 6). This shows that 3 of 6 SNPs from main IL-6 signalling instrument are in strong LD with the alternative IL-6 signalling instrument; rs11264224 with rs11264224 (R 2 =1); rs3766924 with rs12059682 (R 2 =0.994) and rs4129267 with rs2228145 (R 2 =1). Of note, this

eTable 4. Linkage Disequilibrium and F-statistics of IL-6R Gene SNPs between Studies
Kappelmann et al.

Assessment of Horizontal Pleiotropy
We performed various sets of additional analyses to assess the possibility of horizontal pleiotropy, which describes effects of genetic instruments on the outcome independent of the exposure.
For genetic instruments focussed on genetic loci (termed cis-MR in the literature 33 3.97 Note: a λGC refers to the genomic inflation factor, which is calculated as the median χ 2 statistic across SNPs divided by the median χ 2 statistic of the expected χ 2 distribution. If λGC>1, this indicates potential systematic biases in GWAS (e.g., population stratification).  7 so excludes UK Biobank participants. a P-values were FDR-controlled across depressive symptoms for each phenotype using the Benjamini-Hochberg method. 38 b Psychomotor changes, changes in appetite, and sleep problems reflect composite symptoms, which may obscure associations specific to one but not the other underlying symptom.  38 and additional Bonferroni correction was applied for analysed two main exposure phenotypes on CRP levels and IL-6 signalling. b Psychomotor changes, changes in appetite, and sleep problems reflect composite symptoms, which may obscure associations specific to one but not the other underlying symptom. *P<0.05, **P<0.01

eFigure 3. Significant Genetic Associations of BMI (Exposure) and Specific Depressive Symptoms (Outcome)
Note: Significant IVW MR associations of BMI with (A) anhedonia, (B) tiredness, (C) changes in appetite, and (D) feelings of inadequacy. Points represent GWAS-based effect sizes with standard errors. Orange, blue, and green lines show MR IVW, weighted median and MR Egger estimates, respectively.

eTable 13. MR Weighted Median Estimates of Alternative Genetic Instruments for CRP Levels (Genome-wide) and IL-6 Signalling (Indirect)
CRP levels (  38 b Psychomotor changes, changes in appetite, and sleep problems reflect composite symptoms, which may obscure associations specific to one but not the other underlying symptom. Significant results are highlighted in bold and marked with *P<0.05, **P<0.01.

Assessment of Horizontal Pleiotropy
We assessed horizontal pleiotropy for genetic instruments, which indicates if genetic variants are exerting an effect on the outcome variables independent of the exposure (i.e., violation of the exclusion restriction assumption).

Gene-Restricted IVW MR Analyses
Second, we repeated analyses for significant associations found with gene locus-based instruments by restricting the instruments to SNPs within CRP and IL-6R genes (eTable 16). With the restricted CRP levels instrument, as indexed by one SNP (rs1205), there was no evidence for association with any outcome. For IL-6 signalling, we replicated the IL-6 signalling-suicidality association for the main (direct) IL-6 signalling instrument (estimate=0.027, SE=0.011, P=0.013) and found an additional, significant association of this instrument with sleep problems (estimate=0.070, SE=0.034, P=0.037), both based on 3 SNPs within the IL-6R gene (cf. eTable 7). As the association with sleep problems was not found for the insomnia outcome (estimate=0.025, SE=0.107, P=0.814), it is likely that hypersomnia drives this association. We did not find significant associations of the gene-restricted alternative (indirect) IL-6 signalling instrument, based on 7 SNPs, with any symptom. However, the effect estimate with suicidality was similar in size (estimate=0.002, SE=0.002, P=0.176) and there was evidence for significant heterogeneity of indirect IL-6 signalling and suicidality MR analysis (Q=16.47, P=0.011). There was no evidence for heterogeneity in any other gene-restricted MR analysis (all P>0.05; eTable 15). In sum, we found evidence in favour of an association between IL-6 signalling and suicidality and some indication for an association between IL-6 signalling and sleep problems (likely driven by hypersomnia).

MR Egger Estimation of Genome-wide Instruments
Third, we conducted MR Egger estimation to evaluate directional horizontal pleiotropy for the CRP levels and BMI instruments based on genome-wide SNPs (eTable 17). MR Egger estimation allows estimation of an intercept (rather than fixing the intercept at zero as done in IVW and weighted median MR approaches), which describes directional effects of the instrument on the outcome not mediated via the exposure. 36 We did not find significant MR associations (i.e., slopes from MR Egger association) between CRP levels and any outcome variable. There was, however, evidence for significant heterogeneity based on a significant MR Egger intercept in the analysis of CRP levels and suicidality (intercept=0.001, SE=0.001, P=0.047).
For BMI, we found no significant MR Egger intercepts, so no evidence for directional horizontal pleiotropy. MR Egger slopes were only significant for the association with changes in appetite (estimate=0.183, SE=0.038, P<0.001). However, MR Egger slopes were similar in size between BMI and symptoms that were associated with BMI based on IVW and weighted median analyses (i.e., anhedonia, tiredness, changes in appetite, and feelings of inadequacy; cf. eFigure 3). As MR Egger estimation has reduced power compared to IVW and weighted median approaches and we do not find evidence for directional horizontal pleiotropy from MR Egger intercepts, 36 we deem the similarity in slopes as reflective of the robustness of estimates.

Leave-one-out (LOO) and Single-SNP MR Approaches
Lastly, we conducted leave-one-out (LOO) and single-SNP MR analyses. LOO and forest plots for all exposure-outcome MR combinations are available as additional files (https://osf.io/ub83a/). In this supplement, we include LOO and forest plots for our main finding of IL-6 signalling and suicidality (eFigure 4), which indicates that no single SNP is driving significant results. We also provide those LOO and forest plots in eTables 18 and 19, respectively, that arise from IVW MR analyses with evidence for significant heterogeneity (based on PQ<0.05). We further extracted all single SNP MR estimates from 'outlier SNPs', as defined by their highly significant (P<0.001) associations with outcome variables, that were included in instruments from these heterogeneous IVW analyses. We then manually assessed whether these SNPs were eQTLs in brain tissue by extracting the top brain eQTL information from GTEx. We further used the MR Base Phenome Wide Association Study (PheWAS) platform (http://phewas.mrbase.org/) to extract the top phenotype associations with these SNPs (eTable 20).
These sensitivity analyses indicated that the association between the alternative CRP levels instrument and changes in appetite and MD was unstable; that is 95% CIs included/ excluded zero depending on the left-out SNP. In general, however, results were stable and not dependent upon individual SNPs as indicated in LOO analyses. Forest plots showed that for both BMI and genome-wide CRP levels instruments, there were individual SNPs with strong effects (both protective and risk-increasing). Manual exploration of these SNPs (cf. eTable 20) showed that most BMI SNPs had strong associations with metabolic traits from PheWAS even though 'weight-increasing' alleles were not consistently associated with higher/ lower depression phenotypes. This could be an indication of horizontal pleiotropy and mechanisms of effect via other pathways.
For outlying SNPs associated with CRP levels, it was intriguing to see PheWAS traits were related to BMI, height, and other traits associated with metabolic dysregulation (e.g., broadband ultrasound attenuation). 39 This could potentially indicate that the heterogeneity in the genomewide CRP levels instrument was arising from a combination of SNPs associated with CRP levels and metabolic traits included in this instrument. This also emphasises further the value of choosing gene-based/ cis-instruments for MR analysis of CRP levels, which will be more specific to CRP activity. 33
These analyses indicated that main findings of IL-6 signalling and suicidality were stable and unlikely to be due to direct SNP-effects on suicidality. Significant IVW MR associations of higher BMI with anhedonia, tiredness, changes in appetite, and feelings of inadequacy only remained significant in MR Egger regression for changes in appetite but MR Egger slope effect sizes were similar in size and directionally consistent with IVW and weighted median MR estimates for all four symptoms. Accordingly, we report this as "directionally consistent" in main results as MR Egger regression has reduced statistical power as compared to IVW and weighted median approaches. 36

eTable 14. MR Heterogeneity Estimates (Cochrane's Q) of All Genetic Instruments
Main Note: Top Brain eQTL was obtained from GTEx, top Phenome-wide Association Study (PheWAS) hit in UK Biobank traits was obtained from the MR Base PheWAS platform (http://phewas.mrbase.org/), and MR estimates reflect single SNP MR analysis estimates using the Wald ratio. 26,37 a The trait-increasing allele was coded based on effect allele and valence of beta estimate (i.e., if beta estimate was negative, the reference allele was indicated here). b MR estimates were provided, so that A1 reflects the exposure-increasing allele (i.e., if A2 was the exposure-increasing allele, the MR estimate was inversed).
Our analyses have three major differences from the previous MR investigations, however. First, we included larger sample sizes for GWAS of both exposure and outcome phenotypes (note that Wium-Andersen et al. 40 conducted single-sample MR). Second, we used a larger number of instruments for MR analyses and our genetic instruments were characterised with large F-values highlighting their strength (cf. eTables [4][5]. As a result, our MR findings for the inflammation-MD association are likely more precise than that from the previous studies from a statistical perspective. Third, we defined IL-6 instruments based on downstream CRP levels and soluble IL-6Rs. Definition of IL-6 instruments based on IL-6 plasma levels can be argued to be preferable to our approach reflecting a more direct measurement approach. In contrast, IL-6 biology, 18,43-45 and in particular the dependency of downstream pro-inflammatory IL-6 effects on the amount of membrane-bound versus soluble IL-6Rs, could mean we selected a 'purer' index of IL-6associated inflammatory activity as genetic proxy. Both approaches are valuable and can potentially compliment each other to provide a comprehensive overview on IL-6 associations with MD.
In sum, our findings question the robustness of CRP and IL-6 associations with MD, so we hope that future MR investigations will add further to the currently mixed evidence base.