Data sources for diagnoses and study sample

The sample and the phenotypes used in this study have been described in our previous publications [Reference Nudel, Wang, Appadurai, Schork, Buil, Agerbo, Bybjerg-Grauholm, Børglum, Daly, Mors, Hougaard, Mortensen, Werge, Nordentoft, Thompson and Benros7, Reference Nudel, Thompson, Borglum, Hougaard, Mortensen, Werge, Nordentoft and Benros11–Reference Nudel, Benros, Krebs, Allesøe, Lemvigh, Bybjerg-Grauholm, Børglum, Daly, Nordentoft, Mors, Hougaard, Mortensen, Buil, Werge, Rasmussen and Thompson15], and we repeat the details here, with changes relevant to the present study: the iPSYCH Consortium linked data from the Danish medical registers to biobank data via the unique civil registration number used in Denmark since 1968 [Reference Pedersen16]. Biological data from the Danish Neonatal Screening Biobank include dried blood spots taken 4–7 days after birth from nearly all infants born in Denmark after 1981 [Reference Pedersen16, Reference Norgaard-Pedersen and Hougaard17], which were used for downstream genotyping. Infection diagnoses were obtained from the Danish National Hospital Register, which, since 1977, has included records of all inpatients treated in Danish non-psychiatric hospitals, and, since 1995, has included information regarding outpatient and emergency room contacts [Reference Andersen, Madsen, Jorgensen, Mellemkjoer and Olsen18]. The Psychiatric Central Research Register includes data from all psychiatric inpatient facilities since 1969 and outpatient contacts since 1995 [Reference Mors, Perto and Mortensen19]. In Denmark, diagnoses were based on the 8th Revision of the ICD-8 [20] from 1977 to 1993, and, since 1994, they have been based on ICD-10 [21]. For the psychiatric phenotypes included in this study, all diagnosis types apart from henvisningsdiagnose (referral diagnosis) were included, and the types of contact included were inpatient, outpatient, and emergency care unit contact. Data for psychiatric phenotypes are from the Psychiatric Central Research Register. For infection phenotypes, the included types of diagnosis were the following: ICD-8: hoveddiagnose and bidiagnose (main and auxiliary diagnosis, respectively); ICD-10: aktionsdiagnose, grundmorbus, and bidiagnose (main, basic, and auxiliary diagnosis, respectively), and the included types of contact were inpatient and outpatient hospitalisations and emergency room contact, all from the Danish National Hospital Register. Tillægsdiagnoser (associated diagnoses) were not considered, and diagnoses of the following types were excluded: henvisningsdiagnose (referral) and komplikation (complication). ICD-8 diagnoses with the following modifications: ‘Obs. Pro’ and ‘Ej befundet’ (suspected and not found, respectively) were also excluded. All individuals in this study are part of the iPSYCH2012 sample [Reference Pedersen, Bybjerg-Grauholm, Pedersen, Grove, Agerbo, Bækvad-Hansen, Poulsen, Hansen, McGrath, Als, Goldstein, Neale, Daly, Hougaard, Mors, Nordentoft, Børglum, Werge and Mortensen22], selected from among all individuals born in Denmark between 1981 and 2005 (N = 1,472,762), and which included individuals diagnosed with at least one of schizophrenia, bipolar disorder or depression (affective disorder), ASD, ADHD, and anorexia, and individuals who were selected as part of a randomly selected sample from the Danish population. The iPSYCH sample used in this study has undergone several rounds of quality control (QC) as described in our previous studies [Reference Nudel, Wang, Appadurai, Schork, Buil, Agerbo, Bybjerg-Grauholm, Børglum, Daly, Mors, Hougaard, Mortensen, Werge, Nordentoft, Thompson and Benros7, Reference Nudel, Thompson, Borglum, Hougaard, Mortensen, Werge, Nordentoft and Benros11–Reference Nudel, Benros, Krebs, Allesøe, Lemvigh, Bybjerg-Grauholm, Børglum, Daly, Nordentoft, Mors, Hougaard, Mortensen, Buil, Werge, Rasmussen and Thompson15] using data from high-quality genetic markers prior to the imputation. The main sample QC steps included the removal of individuals based on ancestry (individuals who did not have Danish ancestry, as determined from register data of family history and genetic principal component analyses had been removed), as well as relatedness (if they were first- or second-degree relatives of other individuals in the sample; this step prioritised iPSYCH cases and then individuals with a higher genotype call rate). Other QC steps involved the removal of individuals based on missingness (>1%), abnormal heterozygosity, ambiguous sex (discrepancies between annotated sex and genetic data), or if they were duplicates of other individuals. The first studies employing this QC protocol have more information about the procedures [Reference Schork, Won, Appadurai, Nudel, Gandal, Delaneau, Revsbech Christiansen, Hougaard, Bækved-Hansen, Bybjerg-Grauholm, Giørtz Pedersen, Agerbo, Bøcker Pedersen, Neale, Daly, Wray, Nordentoft, Mors, Børglum, Bo Mortensen, Buil, Thompson, Geschwind and Werge23, Reference Erlangsen, Appadurai, Wang, Turecki, Mors, Werge, Mortensen, Starnawska, Børglum, Schork, Nudel, Bækvad-Hansen, Bybjerg-Grauholm, Hougaard, Thompson, Nordentoft and Agerbo24], including the supplementary methods of a preprint of the former [Reference Schork, Won, Appadurai, Nudel, Gandal, Delaneau, Revsbech Christiansen, Hougaard, Bækved-Hansen, Bybjerg-Grauholm, Giørtz Pedersen, Agerbo, Bøcker Pedersen, Neale, Daly, Wray, Nordentoft, Mors, Børglum, Bo Mortensen, Buil, Thompson, Geschwind and Werge25] and another study [Reference Aarøe, Appadurai, Hansen, Schork, Werge, Mors, Børglum, Hougaard, Nordentoft, Mortensen, Thompson, Buil, Agerbo and Petersen26]. Before QC, 78,050 individuals from 23 genotyping waves were included. Following QC, 65,534 unrelated Danish individuals were retained, of whom 34,705 were male and 30,829 were female. Data up to the end of 2012 were included for infections, and data up to the end of 2013 were included for psychiatric diagnoses. We had diagnoses for the following infection categories: bacterial, viral, central nervous system infection, gastrointestinal, genital, hepatitis, otitis, pregnancy infection, respiratory, sepsis, skin infection, HIV/AIDS, urological, and other infections (e.g., protozoan infections). ICD codes for these are provided in Supplementary Table S1. For psychiatric disorders, the following ICD codes were used: any/overall psychiatric diagnosis (ICD-8 code within the range 290–315 and/or ICD-10 code within the range F00–F99); ADHD (ICD-10: F90.0); anorexia (ICD-8: 306.50; ICD-10: F50.0); ASD (ICD-10: F84.0, F84.1, F84.5, F84.8, and F84.9); bipolar disorder (ICD-8: 296.19, 296.39, and 298.19; ICD-10: F30 and F31); depression (single and recurring) (ICD-8: 296.09, 296.29, 298.09, and 300.49; ICD-10: F32 and F33); schizophrenia (ICD-8: 295.x9 (excl. 295.79); ICD-10: F20).

Defining phenotypes for infection load and psychiatric diagnoses

For specific psychiatric disorders, case (control) status was determined based on having (not having) the relevant ICD diagnosis as per the above codes. For any/overall psychiatric disorder, cases had any ICD code from ICD-8: 290–315 and/or ICD-10: F00–F99, and controls did not have any of the codes from those ranges. For the quantitative phenotype of infection load, individuals who did not have any diagnosis from the above 14 infection categories received an infection load phenotype value of 0; for individuals with infection diagnoses, we counted the number of infection categories they had excluding the broad categories of bacterial, viral, and other infections, and the total count was used as their infection load value. This approach was chosen because these broad categories include codes that were also found in specific categories, and, moreover, they do not indicate specific infection sites. Individuals who were cases only for those three categories without any site-specific category (N = 3,074) received a missing infection load value and were not used in analyses for this phenotype. The final sample used in this study, therefore, included 62,460 individuals for the infection load phenotype (NB: when using the binary overall infection phenotype in some analyses, the full sample was used). The phenotype distribution figure was exported using Daniel’s XL Toolbox v7.3.4 [Reference Kraus27].

Statistical and epidemiological analyses

Statistical analyses were performed in R [28] v.3.5.1. Logistic regressions of psychiatric diagnosis on infection load were performed with the glm function (family = binomial(link = ‘logit’)) with covariates for age (in years), age squared, and sex, resulting in two-sided p-values from a Wald test for the infection load coefficient’s being different from zero. Confidence intervals were calculated with the confint function. The regressions were performed in the random population sample (N = 21,706 before exclusions; N = 20,822 after exclusions for the infection load phenotype) to avoid potential biases and obtain population estimates, and controls for psychiatric disorders were ‘normal controls’, that is, they did not have the diagnosis which was being tested but they could have other (psychiatric) diagnoses. Note that the age covariates in this study were censored for a minority of individuals who emigrated (0.43%), died (1.02%), or lost contact with the Danish authorities (0.02%) by July 2013 (i.e., we set their age to what it was when their status changed, whereas, for other individuals, the age at the end of 2013 was used). These figures apply to the full sample. Tetrachoric correlations for overall infection and overall psychiatric diagnosis in the full sample and in the random population sample were calculated with the polycor R package v.0.8-1 [Reference Fox and Dusa29].

Genetic dataset and genome-wide association study for infection load

We performed a discovery genome-wide association study (GWAS) for infection load, and GWASs for other phenotypes (overall psychiatric diagnosis; overall infection) for downstream heritability and genetic correlation analyses. The description of the genetic dataset used in this study has appeared in several of our previous publications [Reference Nudel, Thompson, Borglum, Hougaard, Mortensen, Werge, Nordentoft and Benros11, Reference Nudel, Allesoe, Werge, Thompson, Rasmussen and Benros12], but we repeat it briefly here, with further information relevant to the present study: the marker dataset used in this GWAS has undergone several rounds of QC. The starting point for the QC was a dataset which had 78,050 samples genotyped in 23 of the original 25 waves. This dataset is described in detail elsewhere [Reference Grove, Ripke, Als, Mattheisen, Walters, Won, Pallesen, Agerbo, Andreassen, Anney, Awashti, Belliveau, Bettella, Buxbaum, Bybjerg-Grauholm, Bækvad-Hansen, Cerrato, Chambert, Christensen, Churchhouse, Dellenvall, Demontis, de Rubeis, Devlin, Djurovic, Dumont, Goldstein, Hansen, Hauberg, Hollegaard, Hope, Howrigan, Huang, Hultman, Klei, Maller, Martin, Martin, Moran, Nyegaard, Nærland, Palmer, Palotie, Pedersen, Pedersen, dPoterba, Poulsen, Pourcain, Qvist, Rehnström, Reichenberg, Reichert, Robinson, Roeder, Roussos, Saemundsen, Sandin, Satterstrom, Davey Smith, Stefansson, Steinberg, Stevens, Sullivan, Turley, Walters, Xu, Stefansson, Geschwind, Nordentoft, Hougaard, Werge, Mors, Mortensen, Neale, Daly and Børglum30]. Further QC of this dataset including a description of the procedures for sample and marker QC is provided in the first study that utilised it [Reference Schork, Won, Appadurai, Nudel, Gandal, Delaneau, Revsbech Christiansen, Hougaard, Bækved-Hansen, Bybjerg-Grauholm, Giørtz Pedersen, Agerbo, Bøcker Pedersen, Neale, Daly, Wray, Nordentoft, Mors, Børglum, Bo Mortensen, Buil, Thompson, Geschwind and Werge23]. Note, however, that that study had a minor allele frequency (MAF) threshold (for dosage data) different from the one below. Before the imputation of more markers, markers with rare alleles and non-autosomal markers were removed. The pre-imputation QC is described in more detail in other studies which used this sample [Reference Schork, Won, Appadurai, Nudel, Gandal, Delaneau, Revsbech Christiansen, Hougaard, Bækved-Hansen, Bybjerg-Grauholm, Giørtz Pedersen, Agerbo, Bøcker Pedersen, Neale, Daly, Wray, Nordentoft, Mors, Børglum, Bo Mortensen, Buil, Thompson, Geschwind and Werge23, Reference Erlangsen, Appadurai, Wang, Turecki, Mors, Werge, Mortensen, Starnawska, Børglum, Schork, Nudel, Bækvad-Hansen, Bybjerg-Grauholm, Hougaard, Thompson, Nordentoft and Agerbo24, Reference Aarøe, Appadurai, Hansen, Schork, Werge, Mors, Børglum, Hougaard, Nordentoft, Mortensen, Thompson, Buil, Agerbo and Petersen26]. For the imputation, genotypes were phased with SHAPEIT3 [Reference O’Connell, Sharp, Shrine, Wain, Hall, Tobin, Zagury, Delaneau and Marchini31] and the imputation was performed with IMPUTE2 [Reference Howie, Donnelly and Marchini32]. Following the imputation, markers were excluded if they had an INFO score (calculated with QCTOOL) <0.2; MAF <0.001; best-guess genotypes missing in >10% of subjects, whereby the missingness in imputed genotypes was determined by treating individual genotypes with probability <0.9 as missing; Hardy–Weinberg equilibrium P < 1 × 10⁻⁶; and/or a genome-wide significant association with the genotyping wave or with the imputation batch itself (in controls in a homogeneous European subset of the sample). Lastly, markers with differential missingness between psychiatric cases and controls (P < 1 × 10⁻⁶) were also removed. In this study, we used only best-guess genotypes, or hard calls, and only genotypes with a probability of 0.9 or above (i.e., the imputation resulted in a best-guess genotype with a probability of at least 0.9 for a given marker and a given individual) were retained (others were set to missing). Furthermore, markers were retained only if they had an INFO score (following the imputation) of at least 0.8 and MAF of at least 0.01 in the QCed sample (these steps were additional steps not included in some of the previous studies mentioned [Reference Schork, Won, Appadurai, Nudel, Gandal, Delaneau, Revsbech Christiansen, Hougaard, Bækved-Hansen, Bybjerg-Grauholm, Giørtz Pedersen, Agerbo, Bøcker Pedersen, Neale, Daly, Wray, Nordentoft, Mors, Børglum, Bo Mortensen, Buil, Thompson, Geschwind and Werge23, Reference Erlangsen, Appadurai, Wang, Turecki, Mors, Werge, Mortensen, Starnawska, Børglum, Schork, Nudel, Bækvad-Hansen, Bybjerg-Grauholm, Hougaard, Thompson, Nordentoft and Agerbo24, Reference Aarøe, Appadurai, Hansen, Schork, Werge, Mors, Børglum, Hougaard, Nordentoft, Mortensen, Thompson, Buil, Agerbo and Petersen26]). Ultimately, 7,071,055 markers were used in the GWASs. The genome build for our dataset was hg19. The infection load discovery GWAS was performed with PLINK v.1.90b6.18 using a linear regression model (--linear) and included covariates for age, age squared, sex, the first 10 principal components, and overall psychiatric diagnosis status. The Manhattan plot and the QQ plot were generated with the ‘qqman’ R scripts by Stephen Turner and Daniel Capurso (with the (major update) version from 19 April 2011 for the former plot and the version from 10 June 2013 for the latter, available at https://github.com/stephenturner/qqman/blob/v0.0.0/qqman.r). Other GWASs (with logistic regression in the case of binary phenotypes, with --logistic) were performed for downstream analysis with LDSC, as described below. Those GWASs included covariates for age, age squared, sex, and the first 10 principal components. A covariate for overall psychiatric diagnosis status was included in some analyses for the heritability estimates and not in others (as indicated in the Results section), and it was not included in GWASs used in downstream genetic correlation analyses. For the top markers, a post hoc association test was run in R using a Poisson regression with the glm function (family = poisson(link = ‘log’)) and the same covariates as in the GWAS, as the infection load phenotype consisted of counts and was not normally distributed.

Heritability estimates and genetic correlations

Using the same dataset of markers and covariates as the ones used in the GWAS, we estimated the heritability of infection load in our full sample. This was achieved with GCTA [Reference Yang, Lee, Goddard and Visscher33] as well as LDSC [Reference Bulik-Sullivan, Loh, Finucane, Ripke, Yang, Patterson, Daly, Price and Neale34, Reference Bulik-Sullivan, Finucane, Anttila, Gusev, Day, Loh, Duncan, Perry, Patterson, Robinson, Daly, Price and Neale35]. For GCTA, the genetic relationship matrix (GRM) was calculated for each autosomal chromosome separately with --make-grm and merged with --mgrm with GCTA v1.91.1 beta as previously described [Reference Nudel, Appadurai, Schork, Buil, Bybjerg-Grauholm, Børglum, Daly, Mors, Hougaard, Mortensen, Werge, Nordentoft, Thompson and Benros14]. The heritability of infection load was estimated with --reml in GCTA v1.93.2 beta (covariates included age, age squared, sex, the first 10 principal components, and overall psychiatric diagnosis status). The number of markers in the final GRM for the full sample was 6,941,439. For LDSC, LD score files were generated using the QC-passing random population sample (from the marker dataset with best-guess genotype hard calls updated to have genetic positions using the genetic map from 1000 Genomes phase 3), with a 1 cm window (--l2 --ld-wind-cm 1), as described previously [Reference Nudel, Wang, Appadurai, Schork, Buil, Agerbo, Bybjerg-Grauholm, Børglum, Daly, Mors, Hougaard, Mortensen, Werge, Nordentoft, Thompson and Benros7]. The summary statistics (PLINK output) from the GWAS were processed with the munge_sumstats.py script with the default parameters (after adding A2 from the PLINK bim file and using the NMISS column as the N for LDSC), and LDSC (ldsc.py) v.1.0.1 was used (with the default parameters) in the estimation of the heritabilities (with --h2) and genetic correlations (with --rg). We used the same LD score dataset as both reference and regression (--ref-ld-chr and --w-ld-chr) datasets, as recommended in the LDSC tutorial for non-partitioned LD Score regression (https://github.com/bulik/ldsc/wiki/Heritability-and-Genetic-Correlation, version from 13 July 2017). In the iPSYCH datasets, 5,464,859 markers remained after processing with munge_sumstats.py. We tested the heritability (h ²) as being different from zero using a Wald test and the χ ² distribution with 1 degree of freedom (d.f.) in the following way: χ ² = (h ²/SE)², where h ² is the heritability on the observed scale from LDSC and SE is the standard error for the heritability, and p-values were calculated with 0.5*pchisq(χ², df = 1, lower.tail = F) in R (we multiply by 0.5 since h ² should be nonnegative). For GCTA, the likelihood ratio test statistic from the GCTA output was used to derive p-values using the χ ² distribution with 1d.f., as described above for the LDSC h ² estimates. Genetic correlations were estimated only with LDSC, as it is robust to sample overlap for the studied traits and allowed the use of only summary statistics-level data from the replication sample. Note that when testing for genetic correlations between infection phenotypes and psychiatric disorders in iPSYCH, the summary statistics for the infection phenotypes in iPSYCH were from a GWAS that did not include a covariate for overall psychiatric diagnosis status, as stated previously. For heritability estimation with LDSC, a covariate for overall psychiatric diagnosis status was included in the corresponding GWAS, or not, as indicated in the Results section. All GWASs used in downstream LDSC analyses included covariates for age, age squared, sex, and the first 10 principal components. The genetic correlation (r_g) was tested for being different from zero using a Wald test as well: χ ² = (r_g/SE)², where r_g is the genetic correlation from LDSC and SE is the standard error for the genetic correlation. P-values were calculated in R using pchisq(χ², df = 1, lower.tail = F). Note that in the case of the heritabilities and genetic correlation for or between any infection and any psychiatric diagnosis, which we reported previously [Reference Nudel, Wang, Appadurai, Schork, Buil, Agerbo, Bybjerg-Grauholm, Børglum, Daly, Mors, Hougaard, Mortensen, Werge, Nordentoft, Thompson and Benros7], we include in this study results that used the censored age covariates, so as to be in line with the present study’s methodology. For the binary traits, heritabilities were transformed to the liability scale [Reference Lee, Wray, Goddard and Visscher36] using the proportion of cases in each GWAS and a population prevalence estimate from the iPSYCH random population sample (0.1 for overall psychiatric diagnosis and 0.357 for overall infection).

Replication sample

To replicate the results from the present study and our previous study with a binary overall infection phenotype, we used summary statistics from FinnGen [Reference Kurki, Karjalainen, Palta, Sipilä, Kristiansson, Donner, Reeve, Laivuori, Aavikko, Kaunisto, Loukola, Lahtela, Mattsson, Laiho, PDB, Lehisto, Kanai, Mars, Rämö, Kiiskinen, Heyne, Veerapen, Rüeger, Lemmelä, Zhou, Ruotsalainen, Pärn, Hiekkalinna, Koskelainen, Paajanen, Llorens, Gracia-Tabuenca, Siirtola, Reis, Elnahas, Aalto-Setälä, Alasoo, Arvas, Auro, Biswas, Bizaki-Vallaskangas, Carpen, Chen, Dada, Ding, Ehm, Eklund, Färkkilä, Finucane, Ganna, Ghazal, Graham, Green, Hakanen, Hautalahti, Hedman, Hiltunen, Hinttala, Hovatta, Hu, Huertas-Vazquez, Huilaja, Hunkapiller, Jacob, Jensen, Joensuu, John, Julkunen, Jung, Junttila, Kaarniranta, Kähönen, Kajanne, Kallio, Kälviäinen, Kaprio, Kerimov, Kettunen, Kilpeläinen, Kilpi, Klinger, Kosma, Kuopio, Kurra, Laisk, Laukkanen, Lawless, Liu, Longerich, Mägi, Mäkelä, Mäkitie, Malarstig, Mannermaa, Maranville, Matakidou, Meretoja, Mozaffari, MVK, Niemi, Niiranen, O’Donnell, Obeidat, Okafo, Ollila, Palomäki, Palotie, Partanen, Paul, Pelkonen, Pendergrass, Petrovski, Pitkäranta, Platt, Pulford, Punkka, Pussinen, Raghavan, Rahimov, Rajpal, Renaud, Riley-Gillis, Rodosthenous, Saarentaus, Salminen, Salminen, Salomaa, Schleutker, Serpi, Shen, Siegel, Silander, Siltanen, Soini, Soininen, Sul, Tachmazidou, Tasanen, Tienari, Toppila-Salmi, Tukiainen, Tuomi, Turunen, Ulirsch, Vaura, Virolainen, Waring, Waterworth, Yang, Nelis, Reigo, Metspalu, Milani, Esko, Fox, Havulinna, Perola, Ripatti, Jalanko, Laitinen, Mäkelä, Plenge, McCarthy, Runz, Daly and Palotie37] Release 7 (309,154 individuals). We used two phenotypes which were the closest possible to our any/overall infection and any/overall psychiatric diagnosis phenotypes: certain infectious and parasitic diseases (AB1_INFECTIONS, based on ICD-10 codes starting with A or B) and any mental disorder (KRA_PSY_ANYMENTAL, based on ICD-10 codes F00–F03, F051, G30, and F1–F9, or ICD-8/9 equivalent codes) with the following numbers of cases and controls: 86,892 and 222,262, and 76,073 and 233,081, respectively. The FinnGen summary statistics were processed in the following way: first, they were cross-referenced with the HRC reference panel (ftp://ngs.sanger.ac.uk/production/hrc/HRC.r1-1/HRC.r1-1.GRCh37.wgs.mac5.sites.tab.gz) based on the marker ID. From that, the chromosome and position in genome build hg19 were added to the summary statistics. Based on the chromosome and position in hg19, marker IDs were changed to corresponding marker IDs in the iPSYCH dataset, and only overlapping markers were retained. This made it possible to use the same LD score files and to have greater marker overlap between the iPSYCH and FinnGen datasets for the genetic correlation analyses. Duplicate markers and markers with mismatched alleles (identified using PRSice [Reference Choi and O’Reilly38] v2.3.5) were removed from the summary statistics. The QCed summary statistics were then processed with munge_sumstats.py (as they did not contain sample sizes per marker, the total sample size as per the above numbers was used). After QC and processing, 4,887,803 markers were included in the FinnGen datasets. It should be noted that our LD score files were based on a homogenous Danish-European sample; while there exist LD differences between Finns and Danes, studies have used the same (European) LD weights when analysing both Finnish and Danish samples for genetic correlation estimation with LDSC [Reference Hautakangas, Winsvold, Ruotsalainen, Bjornsdottir, Harder, Kogelman, Thomas, Noordam, Benner, Gormley, Artto, Banasik, Bjornsdottir, Boomsma, Brumpton, Burgdorf, Buring, Chalmer, de Boer, Dichgans, Erikstrup, Färkkilä, Garbrielsen, Ghanbari, Hagen, Häppölä, Hottenga, Hrafnsdottir, Hveem, Johnsen, Kähönen, Kristoffersen, Kurth, Lehtimäki, Lighart, Magnusson, Malik, Pedersen, Pelzer, Penninx, Ran, Ridker, Rosendaal, Sigurdardottir, Skogholt, Sveinsson, Thorgeirsson, Ullum, Vijfhuizen, Widén, van Dijk, de Boer, van den Maagdenberg, Aromaa, Belin, Freilinger, Ikram, Järvelin, Raitakari, Terwindt, Kallela, Wessman, Olesen, Chasman, Nyholt, Stefánsson, Stefansson, van den Maagdenberg, Hansen, Ripatti, Zwart, Palotie and Pirinen39]. However, we also checked the heritabilities and genetic correlations within FinnGen using European LD scores (https://data.broadinstitute.org/alkesgroup/LDSCORE/eur_w_ld_chr.tar.bz2). In these analyses, we did not convert marker IDs, as both FinnGen datasets and the LD score files included SNP rsIDs. Preprocessing with munge_sumstats.py in this case included adding the total sample size as per the above; otherwise, the default thresholds we used. After processing, 12,809,821 markers in each dataset (AB1_INFECTIONS and KRA_PSY_ANYMENTAL) were retained (excluding markers with no rsID in the summary statistics files). The number of overlapping markers across the FinnGen datasets and the European LD score dataset was 1,169,856. Heritabilities were transformed to the liability scale using the proportion of cases for each trait as reported by FinnGen, and the same values were used for the population prevalence (since FinnGen is a full population cohort).

Post hoc analyses for genetic correlations

We performed several post hoc analyses and sensitivity analyses to evaluate whether various factors affected the genetic correlation analyses and/or provide further insight into the genetic relationship between infection and mental illness. The first of these was a test for cross-trait assortative mating, which could influence the r_g estimate. It was shown that, for a single trait, assortative mating would induce a correlation between a PRS generated using only odd- or only even-numbered chromosomes (in the target sample) at a time and that, under random mating, these scores should not be correlated [Reference Yengo, Robinson, Keller, Kemper, Yang, Trzaskowski, Gratten, Turley, Cesarini, Benjamin, Wray, Goddard, Yang and Visscher40]. This approach was later extended to cross-trait assortative mating [Reference Border, Athanasiadis, Buil, Schork, Cai, Young, Werge, Flint, Kendler, Sankararaman, Dahl and Zaitlen41], whereby the PRS for trait 1 generated using only odd-numbered chromosomes is tested for correlation with the PRS for trait 2 generated using only even-numbered chromosomes, and vice versa. We used a simplified version of this procedure, testing for both correlations at both an inclusive p-value threshold (pT = 1) and the genome-wide significance threshold (pT = 5 × 10⁻⁸), with FinnGen summary statistics (as processed above for genetic correlations with iPSYCH) as the training dataset and iPSYCH as the target dataset. The PRSs were generated with PRSice v.2.3.5 with an r ² threshold of 0.1 in a window of 250 kbp and using the --score sum method. Otherwise, the default parameters were used. The correlation was tested with the cor.test function in R, using the Pearson correlation as the method. The tests were two-sided.

To test for genetic causality between overall infection and overall psychiatric diagnosis in iPSYCH, we used the latent causal variable (LCV) model [Reference O’Connor and Price42]. This model tests whether trait 1 (in our case, overall infection) is causal to trait 2 (overall psychiatric diagnosis) using a latent causal variable which is assumed to mediate the genetic correlation between the two traits. If trait 1 is strongly genetically correlated with this variable, then it is partially genetically causal to trait 2. This relationship is quantified using the genetic causality proportion (GCP). GCP ranges from 0 (no genetic causality) to 1 (full genetic causality; if it is negative, it means that trait 2 is partially or fully genetically causal to trait 1). LCV uses LD scores and summary statistics (we used the same LD scores as used with LDSC, and the summary statistics processed with LDSC), as LDSC does, but it uses them in a modified procedure. We used both the full dataset of summary statistics and markers with MAF ≥ 0.05 as recommended on the LCV GitHub page. The scripts used were https://github.com/lukejoconnor/LCV/blob/master/R/ExampleRealdataScript.R, https://github.com/lukejoconnor/LCV/blob/master/R/MomentFunctions.R, and https://github.com/lukejoconnor/LCV/blob/master/R/RunLCV.R. The versions used were from 1 July 2020 for the first script and 14 March 2019 for the last two scripts.

Lastly, we performed sensitivity analyses for the LDSC results after removing the major histocompatibility complex (MHC) region on chromosome 6 from the dataset used to generate the LD scores: we removed markers on chromosome 6 between coordinates 25,392,021 and 33,392,022, based on the coordinates from a GWAS protocol [Reference Anderson, Pettersson, Clarke, Cardon, Morris and Zondervan43] converted to genome build hg19 using the UCSC Genome Browser LiftOver tool, and regenerated the LD scores with LDSC v1.0.1 using the dataset of the random population sample (as described earlier). The LDSC analyses were then repeated using these new LD scores. The numbers of markers included in the h ² and r_g analyses after the removal of the MHC region were 5,424,093 for the iPSYCH datasets and 4,857,454 for the FinnGen datasets (when used with iPSYCH LD scores). Note that the external European LD score dataset did not include the MHC region to begin with.