Examining Gene–Environment Interactions Using Aggregate Scores in a First-Episode Psychosis Cohort

Abstract

Gene–environment (GxE) interactions have been related to psychosis spectrum disorders, involving multiple common genetic variants in multiple genes with very small effect sizes, and several environmental factors that constitute a dense network of exposures named the exposome. Here, we aimed to analyze GxE in a cohort of 310 first-episode psychotic (FEP) and 236 healthy controls, by using aggregate scores estimated in large populations such as the polygenic risk score for schizophrenia and (PRS-SCZ) and the Maudsley environmental risk score (ERS). In contrast to previous findings, in our study, the PRS-SCZ did not discriminate cases from controls, but the ERS score explained a similar percentage of the variance as in other studies using similar approaches. Our study supports a positive additive interaction, indicating synergy between genetic susceptibility to schizophrenia (PRS-SCZ dichotomized according to the highest quartile distribution of the control population) and the exposome (ERS > 75% of the controls). This additive interaction showed genetic and environmental dose dependence. Our study shows that the use of aggregate scores derived from large and powered studies instead of statistics derived from specific sample characteristics is a powerful tool for the study of the effects of GxE on the risk of psychotic spectrum disorders. In conclusion, by using a genetic risk score and an ERS we have provided further evidence for the role of GxE in psychosis.

Introduction

Psychosis spectrum disorders have been defined as corresponding to a complex phenotype due to gene–environment interactions (GxE).¹ Recent estimates establish heritability ranging from 73% for schizophrenia spectrum disorders to 79% for a narrow schizophrenia diagnosis,² thereby reinforcing the role of genetic background in the etiopathology of these disorders.^3,4 Meanwhile, several decades of epidemiological research has demonstrated the association of environmental variables with the risk of psychosis.⁵ Traditionally, research on GxE has been conducted by adopting a selective one-exposure × one candidate gene approach.

However, according to the polygenic theory of schizophrenia, a large fraction of the genetic risk is explained by many common genetic variants in multiple genes with very small effect sizes. In order to identify these variants using genome-wide association (GWA) technologies, the Psychiatric Genomic Consortium recruited a large sample to achieve the required statistical power.⁶ The results of that effort led to the identification of 108 genome-wide significant loci associated with schizophrenia risk,⁷ and thereby allowed the computation of a polygenic risk score for schizophrenia (PRS-SCZ) that explained 7% of the variation in the schizophrenia liability scale. A PRS is a single measure of molecular genetic risk estimated by summing the log of the ORs of individual single nucleotide polymorphisms (SNPs) multiplied by the number of alleles present at the corresponding loci.⁸ The PRS-SCZ uses the log of ORs from a GWA case-control study with high statistical power, thus increasing the potential to detect associations with phenotypes and GxE in other populations.

Moreover, exposure to several environmental factors has been associated with psychosis spectrum disorders, including an urban birth, cannabis use, season of birth, ethnic minority status, high paternal age, obstetric perinatal complications, and childhood adversity or abuse.⁵ However, each factor constitutes a small fraction of the dense network of exposures that makes up our environment: the exposome. An environmental risk score (ERS) would provide a single exposome measure that would improve risk prediction and facilitate research to improve our understanding of the overall impact of our environment and its interaction with our genes on the development of psychosis. Several approaches have been adopted to measure cumulative environmental load in the form of a single aggregate score, analogous to the PRS used in genetics.^9–11 The Maudsley ERS proposed by Vassos et al¹⁰ is based on the largest available meta-analyses of the most replicated and validated environmental risk factors for psychosis, and is the first systematic approach of this kind.

Here, we aimed to analyze GxE in a cohort of first-episode psychotic (FEP) patients from the PEPs study,^12,13 by using aggregate scores estimated in large populations such as the PRS-SCZ⁷ and the Maudsley ERS.¹⁰

Methods

Participants

This study forms part of the project “Phenotype–genotype interaction: application of a predictive model in first psychotic episodes, FIS PI080208” (known as the PEPs study, from the Spanish abbreviation for first psychotic episode). A complete description of the protocol for the PEPs study has been published previously.^14–17 The study was approved by the research ethics committees of all the centers involved. Informed consent was obtained from all participants or from parents or legal guardians of under-age subjects. Briefly, 335 FEP patients and 253 healthy controls were recruited between 2009 and 2011 at 16 Spanish hospitals that participated in the Biomedical Research Networking Center for Mental Health (CIBERSAM), a multi-institutional public research consortium.¹⁸ From this initial sample, 310 (101 females, 209 males; age mean = 24.7 ± 6.0) FEP patients (92.5%) and 236 (85 females, 151 males; age mean = 24.1 ± 6.4) healthy controls (96.0%) with genotype data available were included in the present analysis.

Environmental Exposure

For the present study, the 6 environmental factors included in the Maudsley ERS¹⁰ were selected: ethnic group, urbanicity, high paternal age, low birth weight, cannabis exposure, and childhood adversity. The assessment and recording procedures have been provided elsewhere.¹² Ethnicity was recorded using self-reported ancestries and was dichotomized as “native,” including all Caucasian subjects, and “any origin,” including subjects from other ethnic groups. Urbanicity, which according to the Maudsley ERS referred to the place of birth, was categorized as: “low” (includes rural population and towns <10 000 inhabitants), “medium” (includes towns and cities <100 000 inhabitants) and “high” (includes major cities >100 000 inhabitants). Paternal age was divided in 3 categories according to the Maudsley ERS: <40, between 40 and 50, and >50. Data on paternal age was missing for 47 cases (15.1%) and 12 controls (5.1%). Obstetric complications only considered low birth weight and the cut-point of 2.5 Kg was selected according to the Maudsley ERS. 90 cases (29.0%) and 63 controls (26.7%) did not provide this information. Cannabis use was evaluated using the European Adaptation of a Multidimensional Assessment Instrument for Drug and Alcohol Dependence (EuropAsi)¹⁹ considering the years of cannabis abuse or dependence following the DSM-IV criteria. Subjects who met criteria of abuse or dependence were categorized according to the time of exposure as “no exposure,” “little to moderate exposure” (≤1 y), and “high exposure” (>1 y). Seven cases (2.2%) and 3 controls (1.3%) were missing cannabis data. Information on childhood adversity was extracted from the events that appear in the Traumatic Experiences in Psychiatric Outpatients questionnaire (TQ),²⁰ a self–applied questionnaire of 18 items. Giving the high heterogeneity of traumatic experiences, we had only recorded the total number of experiences by age. This data was used to categorized subjects according to the Maudsley ERS as “No exposure” (no traumatic experience during childhood) and “Any exposure” (1 or more traumatic experiences during childhood). The questionnaire was not completed by 9 cases (2.9%) and 13 controls (5.5%).

These variables were categorized according to the sub-categories described by Vassos et al¹⁰. To each sub-category, we applied the numerical values derived from the log of the relative risk ratios of each factor, as computed by Vassos et al¹⁰. The corresponding numerical risk factor values are presented in table 1. The ERS was obtained by summing these numerical values, replacing missing values with 0.

Table 1.

Environmental Risk Factors Included in the Maudsley Environmental Risk Score (ERS)

Risk Factor	Sub-categories	RR	ERS	Cases	Controls	OR (95% CI)
Ethnicitya	Native (Caucasian)	1	−0.5	262	214	1
	Any origin	2.3	3	48	22	1.78 (1.04–8.05)
Urbanicityb	Low	1.16	−1.5	96	49	1.71 (1.15–2.54)
	Medium	1.55	0	51	17	2.57 (1.44–4.57)
	High	2.07	1	163	170	0.45 (0.31–0.65)
Paternal age	<40	1	0	228	209	1
	40–50	1.17	0.5	34	15	2.10 (1.11–3.98)
	>50	1.60	2	1	0	-
Obstetric complications	Birth weight >2.5 kg	1	0	205	159	1
	Birth weight <2.5 kg	1.67	2	15	14	0.83 (0.39–1.77)
Cannabisc	No exposure	1	−1	176	208	1
	Little to moderate	1.41	0.5	9	3	2.35 (0.63–8.77)
	High exposure	2.77	3	118	22	6.12 (3.72–10.05)
Childhood adversity	No exposure	1	−1.5	144	120	1
	Any exposure	2.78	2.5	157	103	1.27 (0.90–1.80)

Note: For each sub-category the relative risk (RR) and the numeric values (ERS) used to compute the ERS as calculated by Vassos et al¹⁰ are shown. The case-control distribution and ORs computed in the present study are also shown.

^aAny origin included Non-Caucasian participants from North-Africa (1.6%), South-Africa (0.7%), Asia (0.9%), South-America (8.6%), and Romani ethnic group (0.9).

^bUrbanicity: “low” includes rural population and towns <10 000 inhabitants; “medium” includes towns and cities <100 000 inhabitants; “high” includes major cities (>100 000 inhabitants).

^cCannabis: “little to moderate exposure” ≤1 y; “high” >1 y.

For the GxE analysis, ERS was dichotomized according to the highest quartile distribution of the control population (ERS > 75% of the controls).

Genetic Data Processing

Samples from all the individuals in the study were genotyped at the Centro Nacional de Genotipado (CeGen) using the Affymetrix Axiom Spain Biobank Array containing probes for 758 740 SNPs. Genotype data were then called using the Axiom Analysis Suite. Variants with a call rate <98% or with Hardy-Weinberg equilibrium P < 1 × 10⁻⁶ were excluded from the dataset.

After quality control, principal components analysis (PCA) was conducted using the SNPrelate R package,²¹ and the first 10 principal components scores were saved and were used for further analysis. Genotypes were imputed on the Michigan Imputation Server using the Haplotype Reference Consortium reference panel and the programs Eagle for haplotype phasing and Minimac3 for imputation.²²

The PRS-SCZ was calculated for each participant by multiplying the number of risk alleles possessed at each SNP by the natural logarithm of its OR and summing the products. Significant SNPs and its respective ORs were extracted from the clumped summary statistics derived from the Psychiatric Genomic Consortium GWAS results⁷ downloaded from the LD Hub (http://ldsc.broadinstitute.org/ldhub/). The PRS calculation was conducted by using PRSice software.²³ In accordance with previous research in the field, we constructed the PRS using 0.05 as the P-value threshold, given that it explained the most variation in the phenotype of the Psychiatric Genomic Consortium analysis. For the GxE analysis, the PRS-SCZ was dichotomized according to the quartile cutoff based on the control distribution of PRS-SCZ. The highest quartile (PRS-SCZ > 75% of the controls) was considered the genetic risk state for schizophrenia.

Statistical Analysis

Analyses were conducted on imputed data, resulting in missing data for some environmental exposure variables. Missing data patterns were analyzed and randomness was assumed. The multiple imputation chained equation was applied, as implemented in the mice R package,²⁴ with 20 imputations and restricting imputed values to observed ranges in the original variables. Imputed data were similar to observed values. All the regression analyses were run on multiple imputed data and pooled using Rubin’s rules.

Logistic regression analysis was used to test univariate associations with case status. To test the joint effect of environmental and genetic scores, binary states according to the highest quartile distribution were entered as independent variables in a multilevel logistic regression model. All the analysis was carried out using SPSS (version 22.0) for Windows. First 10 genomic principal components were used as covariates to control for population stratification in all statistical analysis that included genetic data.

In our study, we decide to apply additive models because they provide superior representation of biological synergy and inform public health decisions within the sufficient cause framework.^25,26 We tested for departure from additivity using the relative excess risk due to interaction (RERI) as the standard measure for interaction on the additive scale in case-control studies (RERI = OR_ERS75&PRS75 − OR_ERS75 − OR_PRS75 + 1).²⁷ A RERI greater than zero was defined as a positive deviation from additivity and considered significant when the 95% CI did not contain zero. The RERI was calculated using the delta method as implemented in the epiR R package.

Results

The PRS-SCZ did not discriminate cases and controls using logistic regression analysis adjusted by the first 10 principal components and gender (OR = 0.99; 95% CI: 0.99–1.00; P = .144: Nagelkerke’s R² = 0.034). Similar results were obtained when the PRS-SCZ₇₅ was considered the genetic risk state for schizophrenia (PRS-SCZ₇₅; OR = 0.94; 95% CI: 0.62–1.43; P = .797; Nagelkerke’s R² = 0.029).

Table 1 shows the distribution, in cases and controls, of the selected environmental risk factors and the corresponding sub-categories, in accordance to Vassos et al¹⁰. The computed ERS discriminates between cases and controls in the logistic regression analysis adjusted by age and gender (OR = 1.18, 95% CI: 1.11–1.21, P < .001, Nagelkerke’s R² = 0.084). Using the cutoff values proposed by Vassos et al⁷, ERS could be used to classify exposure as low-risk (ERS < 0, 42% of the control population vs 28.0% of FEP, OR = 0.48, 95% CI: 0.32–0.64, F = 19.14, P < .001), medium risk (0 < ERS > 6, 49% of the control population vs 56% of FEP, OR = 1.02, 95% CI: 0.73–1.52, F = 0.02 P > .05) and high risk (ERS > 6, 9.0% of the control population vs 16% of FEP, OR = 4.23, 95% CI: 2.13–7.81, F = 28.34 P < .001). The binary assessment of ERS accounted for a higher proportion of the variability (ERS₇₅, OR = 4.24, 95% CI: 2.57–6.99, P < .001, Nagelkerke’s R² = 0.113).

We tested for correlation between PRS-SCZ and ERS. Nonsignificant correlation was found when all samples were included (r = −0.061, P > .05), but when cases and controls were analyzed separately, a significant negative correlation was found in the control group (r = −0.141, P = .036).

The logistic regression model, adjusted by the first 10 principal components, reported the interactive effects (PRS-SCZ₇₅+ERS₇₅, OR = 6.38, 95% CI: 3.03–13.62, P < .001) of the PRS-SCZ₇₅ (PRS-SCZ₇₅, OR = 0.75, 95% CI: 0.43–1.39, P > .05) and ERS₇₅ (ERS₇₅, OR = 3.99, 95% CI: 2.20–7.23, P < .001) on case status. There was evidence of additive interaction (RERI = 2.64; 95% CI 0.65–10.87, P > .05). The sensitivity analysis replacing the a priori set PRS-SCZ₇₅ as the genetic risk in the model with alternative cutoff points of PRS-SCZ (50% and 25%) confirmed that the additive interaction was dose-dependent and evident across all PRS-SCZ cutoff points (RERI_PRS-SCZ50 = 2.03, 95% CI −2.03 to 6.08, P > .05; RERI_PRS-SCZ25 = 1.18, 95% CI −2.13 to 3.97, P > .05).

Discussion

The present study is, to the best of our knowledge, the first to examine the interaction between genetic liability and environmental exposure as a risk of psychosis spectrum disorders using aggregate metrics of both components computed using summary statistics of large populations. Our study supports a positive additive interaction, indicating synergy between genetic susceptibility to schizophrenia and the exposome. In summary, the combined influence of both genes and environment is larger than the sum of each individual effect.

In contrast to previous findings, in our study, the PRS-SCZ did not discriminate cases from controls.^28,29 This could be due to the diagnostic nature of our cohort, formed of FEP patients (including both affective and non-affective psychosis) instead of only schizophrenia patients. It may also be a result of limited sample size in comparison with previous studies; this limited the statistical power to detect discrete ORs, such as that detected for the PRS-SCZ (OR = 1.30).

Regarding the ERS score, it explained a similar percentage of the variance as in other studies using similar scores derived from meta-analysis.^9,11 Moreover, our results agreed with previous findings showing an additive effect of environmental factors,^30,31 as we observed a dose-dependent effect with increased ORs of psychotic risk as a function of the ERS. An individual in the top 25% of the control distribution was around 4 times more likely to develop a psychotic disorder than an individual below that cutoff point.

Although PRS approaches have gained attention in the study of GxE, to date few studies have been published using the PRS-SCZ for psychosis spectrum disorders. All such studies have used single environmental factors, such as childhood adversity and psychosis³²; intra-uterine environment and schizophrenia³³; or early life exposure to cannabis and cortical maturation.³⁴ In our study, we followed a similar approach to that of the largest case-control study of this kind conducted to date,²⁸ using the PRS-SC₇₅ as the binary genetic risk state and computing the RERI as a standard measure for interaction on an additive scale. However, Guloksuz et al²⁸ explored the interaction between the PRS-SCZ and several environmental factors, including cannabis use, bullying, emotional abuse, physical abuse, sexual abuse, emotional neglect, physical neglect, winter birth, and hearing impairment. Significant interactions were reported for cannabis use, emotional abuse, sexual abuse, emotional neglect, and bullying.²⁸

Although efforts have been made to combine environmental risk factors, there is no consensus on the optimal way of estimating cumulative environmental risk of psychosis.^{9,11,16,35,36} The studies differ not only in the method of calculating the aggregate score, but also in the choice and selection of the environmental factors to be included. The Maudsley ERS as described by Vassos et al¹⁰ summarizes the available evidence and critically appraises conceptual and methodological issues when combining different environmental factors into a single score. Moreover, unlike other approaches that assume equal contributions from each factor, the Maudsley ERS weights each factor by the best estimate of its effect from the latest meta-analysis of each risk factor.¹⁰

Our study shows that the use of aggregate scores derived from large and powered studies instead of statistics derived from specific sample characteristics is a powerful tool for the study of the effects of GxE on the risk of psychotic spectrum disorders. We show that those subjects with higher genomic load and higher exposure to environmental risk factors are more susceptible to develop a FEP than subjects that only showed high genomic load or only showed high environmental risk. In accordance, subjects with high genomic load that do not develop a FEP, showed lower levels of exposition to risk factors.

However, some limitations should be considered. First, given the sample size and the explorative nature of the present study, we focus on the main interactions between aggregate scores of both genetic liability and the exposome. We were unable to explore the complexity of specific genes or pathways and specific environmental factors that could explain causality and biological interactions. Second, due to the cross-sectional nature of the analysis presented, we could only report temporal associations, not causality. Finally, although we have shown evidence of additive interaction between genomic load and environmental risk exposure using aggregate scores, the observed RERI did not reach statistical significance. However, according to definitions, a RERI greater than zero was defined as a positive deviation from additivity, and considered relevant when the 95% CI did not contain zero. On this regard, we applied additive models instead of multiplicative models to test interactions. Although might be statistically convenient to use a multiplicative model, especially for dichotomous dependent variables, these models are more complex and error-prone in their estimations than additive models.

In conclusion, by using a genetic risk score and an ERS we have provided further evidence for the role of GxE in psychosis.

Funding

This study was supported by the Ministerio de Economía y Competitividad (Ref. ISCIII: PI 080208, PI0/00,283 and PI14/00,612)- Instituto de Salud Carlos III- Fondo Europeo de Desarrollo Regional (FEDER); the Centro de Investigación Biomédica en Red de Salud Mental (CIBERSAM); and by the Departament d’Innovació, Universitats i Empresa, Generalitat de Catalunya (2014SGR441).

Acknowledgments

E.V. has received grants and/or acted as consultant and/or speaker for the following companies: AB-Biotics, Abbott, Allergan, Angelini, Astra-Zeneca, Dainippon Sumitomo, Ferrer, Janssen, Lundbeck, Novartis, Otsuka, Pfizer, Richter, Sage, Sanofi, Servier, Sunovion, and Takeda. A.G-P. has received grants and served as consultant, advisor or CME speaker for the following entities: Janssen-Cilag, Lundbeck, Otsuka, Sanofi-Aventis, Exeltis, Angelini, the Spanish Ministry of Science and Innovation (CIBERSAM), and the Basque Government. M.B. has been a consultant for, received grant/research support and honoraria from, and been on the speakers/advisory board of ABBiotics, Adamed, Angelini, Casen Recordati, Eli Lilly, Janssen-Cilag, Lundbeck, Otsuka, Takeda, Somatics and has obtained research funding from the Ministry of Education, Culture and Sport, the Spanish Ministry of Economy, Industry and Competitiveness (CIBERSAM), by the Government of Catalonia, Secretaria d’Universitats i Recerca del Departament d’Economia i Coneixement (2017SGR1355), Foundation European Group for Research In Schizophrenia (EGRIS), and the 7th Framework Program of the European Union. M.B. has received honoraria from talks and consultancy of Adamed, has received honoraria from consultancy of Ferrer, has received research support and honoraria from talks and consultancy of Janssen-Cilag, has received honoraria from talks and consultancy of Lundbeck, has received honoraria from talks and consultancy of Otsuka, and a research prize from Pfizer. M.G. has been on the speakers/advisory board of Janssen-Cilag. M.P. has received educational honoraria from Otsuka, research grants from Instituto de Salud Carlos III (ISCIII), Ministry of Health, Madrid, Spain, has received grant support from ISCIII, Horizon2020 of the European Union, CIBERSAM, Fundación Alicia Koplowitz and Mutua Madrileña and travel grants from Otsuka, Exeltis and Janssen. She has served as a consultant for Servier, Fundación Alicia Koplowitz and ISCIII. R.R-J. has been a consultant for, spoken in activities of, or received grants from: Instituto de Salud Carlos III, Fondo de Investigación Sanitaria (FIS), Centro de Investigación Biomédica en Red de Salud Mental (CIBERSAM), Madrid Regional Government (S2010/ BMD-2422 AGES; S2017/BMD-3740), JanssenCilag, Lundbeck, Otsuka, Pfizer, Ferrer, Juste, Takeda, Exeltis. The other authors declare no conflict of interests.