Polygenic risk for immuno-metabolic markers and specific depressive symptoms: A multi-sample network analysis study

Background About every fourth patient with major depressive disorder (MDD) shows evidence of systemic inflammation. Previous studies have shown inflammation-depression associations of multiple serum inflammatory markers and multiple specific depressive symptoms. It remains unclear, however, if these associations extend to genetic/lifetime predisposition to higher inflammatory marker levels and what role metabolic factors such as Body Mass Index (BMI) play. It is also unclear whether inflammation-symptom associations reflect direct or indirect associations, which can be disentangled using network analysis. Methods This study examined associations of polygenic risk scores (PRSs) for immuno-metabolic markers (C-reactive protein [CRP], interleukin [IL]-6, IL-10, tumour necrosis factor [TNF]-, BMI) with seven depressive symptoms in one general population sample, the UK Biobank study (n=110,010), and two patient samples, the Munich Antidepressant Response Signature (MARS, n=1,058) and Sequenced Treatment Alternatives to Relieve Depression (STAR*D, n=1,143) studies. Network analysis was applied jointly for these samples using fused graphical least absolute shrinkage and selection operator (FGL) estimation as primary analysis and, individually, using unregularized model search estimation. Stability of results was assessed using bootstrapping and three quality criteria were defined to appraise consistency of results across estimation methods, network bootstrapping, and samples. Results Network analysis results displayed to-be-expected PRS-PRS and symptom-symptom associations (termed edges), respectively, that were mostly positive. Using FGL estimation, results further suggested 28, 29, and six PRS-symptom edges in MARS, STAR*D, and UK Biobank samples, respectively. Unregularized model search estimation suggested three PRS-symptom edges in the UK Biobank sample. Applying our quality criteria to these associations indicated that only the association of higher CRP PRS with greater changes in appetite fulfilled all three criteria. Four additional associations fulfilled at least two quality criteria; specifically, higher CRP PRS was associated with greater fatigue and reduced anhedonia, higher TNF- PRS was associated with greater fatigue, and higher BMI PRS with greater changes in appetite and anhedonia. Associations of the BMI PRS with anhedonia, however, showed an inconsistent valence across estimation methods. Conclusions Our findings align with previous studies suggesting that systemic inflammatory markers are primarily associated with somatic/neurovegetative symptoms of depression such as changes in appetite and fatigue. We extend these findings by providing evidence that associations are direct (using network analysis) and extend to genetic predisposition to immuno-metabolic markers (using PRSs). Our findings can inform selection of patients with inflammation-related symptoms into clinical trials of immune-modulating drugs for MDD.

3 suggested 28, 29, and six PRS-symptom edges in MARS, STAR*D, and UK Biobank samples, 24 respectively. Unregularized model search estimation suggested three PRS-symptom edges in the 25 UK Biobank sample. Applying our quality criteria to these associations indicated that only the 26 association of higher CRP PRS with greater changes in appetite fulfilled all three criteria. Four 27 additional associations fulfilled at least two quality criteria; specifically, higher CRP PRS was 28 associated with greater fatigue and reduced anhedonia, higher TNF-α PRS was associated with 29 greater fatigue, and higher BMI PRS with greater changes in appetite and anhedonia. 30 Associations of the BMI PRS with anhedonia, however, showed an inconsistent valence across 31 estimation methods. 32 INTRODUCTION 5 studied less frequently and primarily for CRP (Badini et al., 2020;Kappelmann et al., 2020;66 Milaneschi et al., 2017b66 Milaneschi et al., , 2016. Moreover, inflammatory markers such as CRP are influenced by 67 metabolic factors (Timpson et al., 2011), which may causally underlie some inflammation-68 symptom associations (Kappelmann et al., 2020), so a combined investigation of immuno-69 metabolic factors is needed to disentangle their etiological roles. 70 Regarding clinical complexity, studies have shown associations of inflammatory markers with 71 specific depressive symptoms including fatigue, changes in appetite, anhedonia, and suicidality 72 (Badini et al., 2020;Fried et al., 2019;Jokela et al., 2016;Kappelmann et al., 2020;Köhler-73 Forsberg et al., 2017;Lamers et al., 2020Lamers et al., , 2018Milaneschi et al., 2017a;Moriarity et al., 2020a;74 Simmons et al., 2018;White et al., 2017). Most prior research, however, has restricted its 75 investigation to complexity on one side, that is focusing on multiple immune markers (e.g., cell 76 counts/ serum cytokine levels) while studying a composite depression phenotype or focusing on 77 multiple depressive symptoms in the context of a single inflammatory marker (mostly CRP). 78 Moreover, previous studies have usually considered associations of inflammatory markers with 79 each depressive symptom in isolation. Although these prior approaches have led to important 80 findings, they cannot address potential causal interactions between symptoms, thus conflate 81 evidence for indirect and direct associations. For example, analyses of isolated symptoms could 82 hypothetically provide evidence for associations of CRP with both fatigue and sleep problems 83 even if CRP was only indirectly associated with fatigue via its effect on sleep problems. A 84 network-based approach provides one means of disentangling such direct from indirect 85 inflammation-symptom associations. 86 Network theory and related analysis techniques have recently been put forward to accommodate 87 the symptomatic complexity of mental disorders (Borsboom, 2017). Network theory proposes 88 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;https://doi.org/10.1101https://doi.org/10. /2021 putative causal interactions between symptoms (e.g., fatigue causing concentration problems 89 causing low mood), which could result in self-reinforcing vicious symptom cycles triggering and 90 maintaining mental disorders. Such associations have been investigated in an increasing amount 91 of studies on psychological symptom networks (Contreras et al., 2019;Robinaugh et al., 2020). 92 To accommodate etiological factors beyond symptoms, however, recent work has proposed an 93 expansion of symptom networks to so-called 'multi-plane' networks, for instance also including 94 genetic, metabolic, immunological, or environmental variables (Guloksuz et al., 2017). To our 95 knowledge, so far, two studies have evaluated such multi-plane networks in the context of 96 inflammation and depression by jointly analysing serum CRP and cytokine concentrations with 97 individual depressive symptoms (Fried et al., 2019;Moriarity et al., 2020a). Findings suggested 98 unique associations of CRP with fatigue and changes in appetite. A third study has recently also 99 provided evidence that the symptom structure itself was a function of CRP levels; that is, 100 interconnections between symptoms were moderated by CRP (Moriarity et al., 2020b). All of 101 these previous studies were based on serum markers for inflammatory proteins, however, 102 reflective of acutely elevated inflammatory activity. Therefore, it remains unclear if 103 inflammation-symptom associations generalise to genetic/lifetime predisposition to higher 104 immuno-metabolic marker levels. 105 In the present study, we explored associations of polygenic risk scores (PRSs) for four major pro-106 and anti-inflammatory markers (i.e., CRP, IL-6, IL-10, & TNF-α) and Body Mass Index (BMI), 107 as a metabolic marker, with individual depressive symptoms using a multi-sample, multi-plane 108 network analysis approach. We evaluated associations in three large samples including the 109 inpatient Munich Antidepressant Response Signature (MARS) study (n=1,058), the outpatient 110 Sequenced Treatment Alternatives to Relieve Depression (STAR*D) study (n=1,143), and the 111 general population UK Biobank cohort (n=110,010) (Hennings et al., 2009;Rush et al., 2004;112 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted January 11, 2021. ; https://doi.org/10.1101/2021.01.07.20248981 doi: medRxiv preprint Sudlow et al., 2015). This investigation aimed to contribute to the study of inflammation and 113 depression by simultaneously addressing (i) combined immunological and symptom complexity 114 (using network analysis), (ii) unclarity regarding the influence of genetic/lifetime predisposition 115 to higher immuno-metabolic marker levels on depression (defining immuno-metabolic markers 116 using PRSs), and (iii) issues of reproducibility and generalisability (testing associations in one 117 large general population and two clinical samples). 118 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted January 11, 2021. ;https://doi.org/10.1101https://doi.org/10. /2021 National Coordinating Center, a Data Coordinating Center, and the Data Safety and Monitoring 141 Board at the National Institute of Mental Health (Rush et al., 2006(Rush et al., , 2004 1986) while the UK Biobank used the self-report Patient Heath Questionnaire (PHQ)-9 (Löwe et 150 al., 2004). From these questionnaires, we selected seven depressive symptoms for joint analyses 151 across samples. These symptoms included completely overlapping symptoms of depressed mood, 152 anhedonia, fatigue, and suicidality, but also partially overlapping symptoms of sleep problems, 153 changes in appetite, and psychomotor changes. Supplementary Table 1 provides an item-level  154   overview of depressive symptoms and Supplementary Table 2 displays symptom coding, where  155 this differed from original item Likert scale ratings. 156 Regarding partially overlapping symptoms of sleep problems, changes in appetite, and 157 psychomotor changes, the PHQ-9 only assesses information on conflated symptoms (e.g., 158 insomnia and hypersomnia are conflated to sleep problems) while the HAM-D incorporates 159 disaggregated symptoms. To harmonise these symptom data for retention in network analyses, 160 we conflated HAM-D symptoms of psychomotor retardation and agitation to "psychomotor 161 changes". For sleep problems and changes in appetite (available in the PHQ-9), only insomnia 162 and loss of appetite are available in the HAM-D, so we included both conflated and 163 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted January 11, 2021. ;https://doi.org/10.1101https://doi.org/10. /2021 unidirectional symptoms in network analyses as previous studies have specifically highlighted 164 associations of inflammation with these symptoms (Jokela et al., 2016;Milaneschi et al., 2017b). 165 We reasoned that comparative appraisal of associations, for example with changes in appetite and 166 loss of appetite, could give further indications on potential specificity of associations to symptom 167 directions, as observed in a previous report (Kappelmann et al., 2020). 168

Immuno-metabolic marker selection and GWAS data sources 183
PRSs for CRP, IL-6, IL-10, TNF-α, and BMI were computed based on available summary 184 statistics from genome-wide association studies (GWAS; Ahola-Olli et al., 2017;Ligthart et al., 185 2018;Locke et al., 2015). These inflammatory markers were selected, because (i) they showed 186 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. robust differences in case-control studies; (ii) CRP, IL-6, and TNF-α have been the most 187 frequently investigated inflammatory markers overall in the context of depression; and (iii) IL-10 188 was the most frequently studied anti-inflammatory cytokine, so could be informative on direction 189 of associations between depressive symptoms and innate immune activity (Köhler et al., 2017;190 Osimo et al., 2019). BMI was selected as the most frequently investigated metabolic marker. 191 GWAS data for CRP were obtained from a large European GWAS of 88 studies including 192 204,402 individuals (Ligthart et al., 2018). GWAS data for IL-6, IL-10, and TNF-α were obtained 193 from a GWAS of 8,293 Finns (Ahola-Olli et al., 2017). GWAS data for BMI were obtained from 194 the Genetic Investigation of Anthropometric Traits (GIANT) consortium that included up to 195 322,154 individuals (Locke et al., 2015). 196

PRS computation 197
PRSs can be computed by summing the GWAS association estimates of risk alleles for each 198 individual. Classically, this summation is done using an approach termed "clumping and 199 thresholding" (C+T), which first reduces summary statistics to independent SNPs and then 200 applies one or multiple thresholds (usually based on P-values) to restrict summation to SNPs with 201 high evidence for associations with phenotypes (Choi et al., 2020). As the optimal threshold for 202 the C+T approach is unknown and should ideally be estimated in a separate dataset with available 203 phenotype data, we computed PRSs using the Bayesian regression and continuous shrinkage 204 priors (PRS-CS) approach, which has been shown to perform similar to or outperform other PRS 205 computation approaches such as C+T (Ge et al., 2019;Ni et al., 2020). 206 PRS-CS takes a linkage disequilibrium (LD) reference panel into account (we used European 207 ancestry data from 1000 Genomes Project phase 3 samples) to update SNP effect sizes in a 208 blocked fashion, thus providing accurate LD adjustment. We pre-specified the global shrinkage 209 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; https://doi.org/10.1101/2021.01.07.20248981 doi: medRxiv preprint parameter using suggested defaults for less polygenic ( =1e -4 ) and more polygenic ( =1e -2 ) 210 phenotypes as =1e -4 for CRP, IL-6, IL-10, and TNF-α, and as =1e -2 for BMI; see details in 211 Supplementary Methods. Following PRS computation in individual samples, polygenic scores 212 were corrected for age, sex, and the first two genotyping principal or multidimensional scaling 213 (MDS) components using linear regression; genotyping MDS components were computed based 214 on raw Hamming-distances in MARS and STAR*D, and using principal component analysis on 215 high-quality, unrelated individuals in the UK Biobank sample (Bycroft et al., 2018). PRSs in 216 MARS were additionally corrected for the genotyping array. Following computation, higher 217 PRSs reflect higher genetic predisposition to respective immuno-metabolic phenotype levels. 218

PRS evaluation 219
In Supplementary Table 3, we provide the number of SNPs included in PRS computation in each 220 sample, which was approximately around one million SNPs for each phenotype-sample 221 combination. The proportion of SNP overlap between samples (for the same phenotype) was 222 >0.89 suggesting that mostly overlapping SNPs contributed to PRSs (Supplementary Table 4). 223 Taking these overlapping SNP sets, correlations between the posterior SNP effect sizes between 224 samples were large for CRP (Pearson's r range: 0.69-0.76) and BMI (Pearson's r range: 0.79-225 0.80) and relatively smaller for IL-6, IL-10, and TNF-α (Pearson's r range: 0.41-0.46; see 226 Supplementary Table 5). This suggests polygenic risk was quantified more similarly across 227 samples for CRP and BMI as compared to IL-6, IL-10, and TNF-α. 228 We quantified the impact that pre-specification of the hyperparameter had on resulting PRSs, 229 which was likely small (Supplementary Table 6). Specifically, PRSs with pre-specified 230 exhibited large correlations with PRSs based on automatic learning of from GWAS summary 231 data (termed PRS-CS-auto in the literature; Pearson's r range: 0.82-0.98). Furthermore, 232 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; moderate-to-large correlations remained to PRSs based on extreme grid search boundary values 233 of (Pearson's r range: 0.47-0.93). 234 Since MARS utilised three different genotyping arrays, we evaluated if our approach of 235 combining data from these arrays into a combined sample was justified. Based on highly similar 236 PRS distributions following adjustments for age, sex, genotyping principal/MDS components, 237 and genotyping array as well as absence of edges between genotyping array and PRSs in an λ 2 ) were selected using 10-fold 250 cross-validation to optimise the Bayesian Information Criterion (BIC). As recommended, we set 251 weights for the importance of each sample as 'equal' to ascertain that a single sample would not 252 dominate estimation (Danaher et al., 2014). 253 As secondary analysis, we also estimated networks for each sample individually using an 254 unregularized gaussian graphical stepwise model selection ("ggModSelect") algorithm 255 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; implemented in the qgraph package (version 1.6.5; Epskamp et al., 2012). The model search 256 algorithm used Spearman correlations and started from an empty model. Throughout results, we 257 will refer to this estimation strategy as "unregularized model search" or "model search" for 258 simplification. 259 We also estimated node predictability with the mgm package (version 1.2-10), which uses node-260 wise estimation to estimate networks for each sample (Haslbeck and Waldorp, 2020). Node 261 predictability describes the amount of variance in a node that is explained by all other nodes in 262 the network, so can be interpreted akin to R 2 (Haslbeck and Fried, 2017). Tuning parameter 263 selection for λ in mgm estimation was based on optimising the BIC based on 10-fold cross-264

validation. 265
Networks were visualised with the qgraph package using an average layout estimated with the 266 Fruchterman-Reingold algorithm for the FGL networks. This algorithm places nodes close to 267 each other that are connected by large edges (Epskamp et al., 2012). While this simplifies 268 network appraisal, it is important to note that nodes and edges should not be interpreted based on 269 their relative position within the network, which can be unstable. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;

Interpretation 278
We interpreted estimated networks based on the presence and reproducibility of edges as defined 279 using three quality criteria. First, we tested if edges were nonzero in FGL networks as well as 280 nonzero and directionally consistent in >50% of bootstrapped analyses (quality criterion 1) akin 281 to a previous PRS-symptom network study in psychosis by Isvoranu and colleagues (2020). 282 Second, we tested if edges between PRSs and symptoms were present (according to criterion 1) 283 across FGL networks of the three samples (quality criterion 2). Third, we tested if edges were 284 present in secondary analyses using unregularized model search estimation in individual samples, 285 again confirmed in >50% of bootstrapped estimations exhibiting directionally consistent 286 estimates (quality criterion 3). 287

Availability of data and materials 288
Data from original studies is not openly available, but can be requested; see details in 289 Supplementary Table 7. GWAS summary data for IL-6, IL-10, and TNF-α is openly available 290 from the original publication by Ahola-Olli and colleagues (2017), for BMI from the GIANT 291 consortium, and can be requested for CRP from the CHARGE inflammation working group. We 292 provide analysis scripts and estimated network matrices (including bootstrapped network 293 matrices) on the Open Science Platform (OSF) under https://osf.io/q4vw9/. 294 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;

295
Baseline characteristics of study populations are displayed in Table 1 analyses were conducted to assess stability of networks. We defined three quality criteria to 302 denote consistency of results across estimation techniques, bootstrapping, and samples. Focus of 303 this network investigation were unique associations (termed edges in network analysis) between 304 PRSs and symptoms, which are summarised in Table 2. 305

Fused Graphical LASSO (FGL) estimation suggests four consistent PRS-symptom edges 306
Using FGL estimation, we obtained networks that are visualised in Figure  As expected, nodes within the same plane displayed relatively stronger within-plane (i.e., 310 symptom-symptom & PRS-PRS) than between-plane (i.e., PRS-symptom) associations. Among 311 PRSs, CRP displayed associations with BMI (edge weight range across samples: 0.16-0.19) while 312 IL-6, IL-10, and TNF-α (based on the same GWAS) were associated with each other (edge 313 weight range across samples: 0.08-0.52). Associations of BMI and CRP with IL-6, IL-10, and 314 TNF-α were largely absent or very small (edge weight range across samples: -0.02-0.01). Among 315 symptoms, the largest associations were present between the core symptoms depressed mood and 316 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; anhedonia (edge weight range across samples: 0.14-0.55). Supplementary Figure 4  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; edges. Networks were comparable to FGL estimation, but generally sparser than those using FGL 340 estimation; see network graphs in Supplementary Figure 6 and bootstrapping results in 341 Supplementary Figures 7-9. 342 Regarding PRS-symptom edges, only three edges were estimated, which were all observed in the 343 UK Biobank sample and fulfilled quality criterion 3 (nonzero edges are also nonzero and 344 directionally consistent in >50% of bootstraps); these edges have been manually unfaded in 345 Supplementary Figure 6. The specific PRS-symptom edges were between the BMI PRS and 346 changes in appetite and anhedonia and between the CRP PRS and changes in appetite. 347 Comparing these edges to FGL estimation, the edge of the CRP PRS with changes in appetite 348 reproduced one of the edges fulfilling quality criteria 1 and 2 while the two edges observed for 349 the BMI PRS were only fulfilling quality criterion 1 (presence in FGL estimation and >50% of 350 bootstraps). Moreover, the BMI PRS association with anhedonia was negative using 351 unregularized model search estimation, but positive using FGL estimation. 352 - Table 2-353 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;

354
The present study investigated associations of PRSs for immuno-metabolic markers with 355 depressive symptoms using a multi-plane, multi-sample network analysis approach. Based on 356 three quality criteria emphasising consistency of network analysis results across statistical 357 bootstraps, samples, and estimation methods, we observed a unique association between the CRP 358 PRS and changes in appetite that met all three quality criteria. In addition to this association, we 359 observed five additional PRS-symptom associations that met two quality criteria. These included 360 edges of the CRP PRS with anhedonia (negative association) and fatigue, the TNF-α PRS with 361 fatigue, and the BMI PRS with anhedonia and changes in appetite. However, the BMI PRS-362 anhedonia association switched association direction depending on the estimation method, so 363 may not be fully consistent despite fulfilling our consistency criteria. Due to the novelty of our 364 analysis approach, we highlight several methodological considerations below, which we hope 365 provides a helpful framework to the discussion of our findings afterwards. 366 367

Methodological challenges and opportunities 368
Combining PRSs with psychological symptom networks is a relatively recent extension of 369 network analysis and, to our knowledge, has only been applied in one previous investigation 370 incorporating a schizophrenia PRS into a psychotic symptom network (Isvoranu et al., 2020). 371 Therefore, it is important to emphasise the unique challenges and opportunities of this approach. 372 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; network. Inclusion of PRSs into psychological symptom networks, and especially of potential 377 pathomechanistic (e.g., inflammatory) rather than main illness (e.g., depression) scores into these 378 networks, aggravates the sample size requirements for network analysis as PRSs only explain a 379 fraction of variance in the heritable component of their target phenotypes (Choi et al., 2020;380 Wray et al., 2020). 381 Second, and because PRSs only measure a fraction of variance in their target phenotype, unique 382 associations observed in network analyses are inevitably smaller than actual target phenotype-383 symptom associations. Taking this study as an example, absolute sizes of CRP-symptom 384 associations were 5-to 10-fold smaller than those from a prior network investigation using serum 385 CRP concentrations by Moriarity and colleagues (2020a). Therefore, PRS-symptom associations 386 are unlikely to give meaningful insights into size of association with the target phenotype, but 387 should, in our opinion, be interpreted based on robust presence/absence of specific associations. Despite these challenges, PRS-symptom networks also provide multiple opportunities. First, 397 PRSs reflect estimates of genetic liability to phenotype expression, so can give an indication on 398 the influence of lifelong predisposition to higher phenotype levels on the symptom level. In this 399 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; way, PRS-symptom associations also provide an indication regarding temporality of association, 400 which Bradford-Hill defined as one of the viewpoints for causality (Bradford Hill, 1965). It is 401 important to note, however, that evidence for a unidirectional temporal association does not 402 preclude bi-directionality. Moreover, PRSs combine information from a multitude of genetic 403 variants (in our case from ~1 million SNPs) that are not restricted to functional SNPs only, can 404 include false positive associations (i.e., noise), and can also tag information of pleiotropic 405 environmental confounding factors. Therefore, causal inferences should rely on separate evidence 406 from clinical trials and/or more focused genetic approaches such as Mendelian Randomisation 407 studies (Lawlor et al., 2008). 408 Second, the PRS-symptom network analysis approach allows the concurrent investigation of 409 multiple immuno-metabolic markers with multiple symptoms. Thereby, immunological and 410 clinical complexity is addressed concurrently, which is an advantage to previous investigations. 411 Furthermore, network analyses usually estimate partial/unique associations, so any emerging 412 associations could suggest direct causal paths from PRS phenotypes to individual symptoms, so 413 may pinpoint so-called 'bridge symptoms' that act as etiological docking sites of risk effects on 414 the symptom plane. 415 Third, large-scale population-based or patient cohort studies, commonly used in network 416 analysis, often do not have detailed immunophenotyping data available. If at all, studies mostly 417 have data available for serum CRP, but rarely for more specific cytokines. Conversely, the advent 418 of large GWAS investigations has produced a substantial amount of large cohort databases with 419 in-depth genotyping and phenotyping information. Combining such databases with GWAS 420 summary statistics from more focused investigations, such as on individual cytokines (Ahola-Olli 421 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021.  (Kappelmann et al., 2020;Milaneschi et al., 2017bMilaneschi et al., , 2016. Evidence from Mendelian 433 randomisation analyses further suggests that BMI could be a potential causal factor for changes 434 in appetite and specifically for increased appetite (Kappelmann et al., 2020;Milaneschi et al., 435 2020b). 436 In addition to these PRS associations with changes in appetite, we also observed associations of 437 higher CRP PRS with lower anhedonia and greater fatigue and of higher TNF-α PRS with greater 438 fatigue. Fatigue in particular has long been considered to have a neuroimmune basis (Dantzer et 439 al., 2014), is common across other medical illnesses characterised by chronic inflammation, and 440 has been reliably associated with inflammatory markers in previous studies including two 441 network investigations (Fried et al., 2019;Jokela et al., 2016;Lamers et al., 2020;Moriarity et 442 al., 2020a;van Eeden et al., 2020;White et al., 2017). While there have also been some studies 443 suggesting associations of inflammatory markers with anhedonia (Köhler-Forsberg et al., 2017;444 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; van Eeden et al., 2020), it is important to note that associations of the CRP PRS with anhedonia 445 observed in the present report were negative, so do not offer straightforward replication of these 446 findings. Nonetheless, we have recently shown in Mendelian randomisation analyses that BMI 447 could be a potential causal factor for both fatigue and anhedonia (Kappelmann et al., 2020), so 448 continued investigation of these symptoms is warranted. 449 Together, our findings add to the notion of an immuno-metabolic subtype of depression 450 characterised by neurovegetative symptoms of changes in appetite and fatigue (Dantzer et al., 451 2008;Milaneschi et al., 2020a). We also expand upon previous work by showing that 452 genetic/lifetime predisposition to higher inflammation and metabolic dysregulation increases risk 453 for depression and, based on network analysis results, these etiological factors may specifically 454 confer their risk on the broader depression syndrome through symptoms such as changes in 455 appetite and fatigue. These results can inform the design of clinical trials of anti-inflammatory 456 approaches and metabolic interventions by specifically selecting patients with an atypical, 457 neurovegetative symptom presentation. As clinical trials for immune-modulating drugs are 458 currently still characterised by relatively small sample sizes (Husain et al., 2020;Khandaker et 459 al., 2018;McIntyre et al., 2019;Raison et al., 2013), it may be worthwhile to pilot new 460 interventions with neurovegetative symptoms/phenotypes as outcome variables. This might 461 increase statistical power and sensitivity to detect effects for these proof-of-concept trials and 462 could then be followed up by larger trials testing broader clinical efficacy measures. 463

Strengths and limitations 464
Strength of this study include availability of large general population-based and patient samples 465 (maximising generalisability), polygenic definition of immuno-metabolic risk variables (indexing 466 lifetime predisposition to higher immuno-metabolic marker levels), and application of network 467 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;https://doi.org/10.1101https://doi.org/10. /2021 analysis (addressing immunological and clinical complexity concurrently). We have addressed 468 some of the more general limitations of combined PRS-symptom network analysis above, but 469 there are two more specific limitations that warrant mentioning. 470 First, data used in the current study included inpatients, outpatients, and individuals from the 471 general population and was based on different scales to measure depressive symptoms. 472 Depressive symptom structure varies between acutely ill patients versus those in remission (van 473 Borkulo et al., 2015), which may have influenced PRS-symptom associations. Moreover, two of 474 the seven symptoms used in the present report only overlap partially; the UK Biobank study 475 includes conflated items on sleep problems and changes in appetite while MARS and STAR*D 476 include items on insomnia and loss of appetite, respectively. This difference may explain some of 477 the inconsistencies observed in the current report such as the diverging valence of edge estimates 478 between CRP and changes in appetite. However, this may have also reduced statistical power to 479 detect associations. Future studies would benefit from inclusion of studies with the same 480 questionnaire and disaggregated symptom measures. 481 Second, PRSs are based on GWAS with highly diverging samples sizes as a large number of 482 individuals were included in the GWAS for BMI and CRP (>200 thousand individuals) and 483 smaller numbers of individuals (~8 thousand individuals) for IL-6, IL-10, and TNF-α. 484 Consistency of effect sizes following the PRS-CS approach was also larger for CRP and BMI as 485 compared to IL-6, IL-10, and TNF-α. This is likely to have shifted the balance of statistical 486 power towards detection of PRS-symptom associations to BMI and CRP rather than IL-6, IL-10, 487 and TNF-α. Therefore, our findings require replication once larger individual cytokine GWAS 488 become available. 489

Conclusion 490
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; https://doi.org/10. 1101/2021 The present investigation studied associations between four major pro-and anti-inflammatory 491 markers, BMI, and depressive symptoms by applying network analysis across one large general 492 population and two patient samples. Defining immuno-metabolic markers using polygenic risk 493 scores expanded previous reports by suggesting direct associations of genetic/lifetime 494 predisposition to immune-metabolic markers with depressive symptoms and provided evidence 495 for temporality of association. Despite methodological restrictions of the presented approach, we 496 observed associations of polygenic risk for CRP with changes in appetite and fatigue, for TNF-α 497 with fatigue, and similar associations for BMI. These findings align with recent 498 conceptualisations of an immuno-metabolic subgroup of depressed patients characterised by 499 atypical, neurovegetative symptom profiles. Results can inform future clinical trials of anti-500 inflammatory approaches by prioritising these patients for selection into clinical trials. 501 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

DECLARATION OF INTERESTS
The authors do not have any competing interests.

ACKNOWLEDGMENTS
We are grateful to all original authors, technical assistants and patients who contributed to the MARS study. We are grateful for the National Institute of Mental Health (NIMH) and the NIMH Repository and Genomics Resource (NRGR) for the possibility of analysing the STAR*D data.
We are also grateful to the original STAR*D authors, and particularly for the contributions of all patients and families who participated in the study. Data were obtained from the limited access datasets distributed from the NIH-supported 'Sequenced Treatment Alternatives to Relieve Depression' (STAR*D). The study was supported by NIMH Contract No. N01MH90003 to the University of Texas Southwestern Medical Center. The ClinicalTrials.gov identifier is NCT00021528. This research has been conducted using the UK Biobank Resource. We are grateful for all scientists and participants who made this large-scale effort and resource possible.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ;https://doi.org/10.1101/2021 CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; analysis of interleukins 6 and 1β, tumour necrosis factor α and C-reactive protein in patients with major depressive disorder. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. ; is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021. Note: Cell values reflect edge weights (i.e., partial correlation coefficients) and the percentage of 500 bootstrap estimations that edges were present. Estimates are restricted to those edges, for which >50% of bootstrapped samples were non-zero and directionally consistent (i.e., criteria 1 & 3). * Changes in appetite and sleep problems are measured as composite symptoms in UK Biobank, but as loss of appetite and insomnia in MARS and STAR*D samples. .

CC-BY 4.0 International license
It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted January 11, 2021.

Figure 2. Estimated FGL networks across samples
Legend: Networks are visualised with the qgraph package. Blue lines indicate positive and red lines negative associations, respectively, with larger associations displayed with thicker lines.
Circles around nodes display node predictability, which can be interpreted similar to explained variance. Maximum size of edge associations is 0.55. As the primary focus of this investigation was to identify consistent PRS-symptom associations, we manually unfaded edges between PRSs and symptoms if these edges met quality criteria 1 and 2 (see Table 2). Changes in appetite and sleep problems are measured as composite symptoms in UK Biobank, but as loss of appetite and insomnia in MARS and STAR*D samples.

Figure 3. Bootstrapped 95% quantile intervals of PRS-symptom edges using FGL estimation
Legend: Bootstrapped 95% quantile intervals (i.e., 95% of the distribution of raw bootstrapped edge estimates) are highlighted as shaded area for each edge. Black points indicate the raw FGL sample estimate while red points indicate the raw bootstrapped mean estimate. Edges are indicated on the y-axis and sorted by mean edge weight across samples in descending order.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted January 11, 2021.