Introduction
Schizophrenia (SZ) is a complex disorder influenced by various factors, including genetic and environmental factors. Genetic factors explain a substantial portion of the risk, as indicated by twin studies demonstrating >80% heritability (Sullivan, Kendler, & Neale, Reference Sullivan, Kendler and Neale2003). However, environmental factors, particularly a history of significant childhood adversity (CA), also play key roles in the development and course of the disease (Murray et al., Reference Murray, Mondelli, Stilo, Trotta, Sideli and Ajnakina2020). The relationships among these factors have been extensively discussed, and hypotheses of gene–environment interactions (GEIs) for SZ have been proposed (Murray, Reveley, & McGuffin, Reference Murray, Reveley and McGuffin1986; Schulsinger, Parnas, Mednick, Teasdale, & Schulsinger, Reference Schulsinger, Parnas, Mednick, Teasdale and Schulsinger1987; Strahilevitz, Reference Strahilevitz1974). The identification of GEI effects in psychiatric disorders offers three key benefits: it allows new genetic and environmental main effects to be discovered (Kraft, Yen, Stram, Morrison, & Gauderman, Reference Kraft, Yen, Stram, Morrison and Gauderman2007), it enhances the understanding of the underlying biological pathways (Caspi & Moffitt, Reference Caspi and Moffitt2006), and it offers clinical insights for personalized medicine and lifestyle recommendations (Dempfle et al., Reference Dempfle, Scherag, Hein, Beckmann, Chang-Claude and Schäfer2008; Uher & McGuffin, Reference Uher and McGuffin2008). Although GEIs are important, their identification requires sample sizes ranging from thousands (for candidate genes) to tens of thousands (for genome-wide association studies); thus, the statistical power can be limited (Thomas, Reference Thomas2010).
When used as a single metric of molecular genetic risk, the polygenic risk score (PRS) has greatly enhanced the capacity for identifying associations with phenotypes and GEIs (Choi, Mak, & O’Reilly, Reference Choi, Mak and O’Reilly2020; Lin, Huang, Liu, Tsai, & Kuo, Reference Lin, Huang, Liu, Tsai and Kuo2019). There is evidence that substantially greater predictive power can be achieved using the PRS rather than a small number of genome-wide significant single nucleotide polymorphisms (SNPs) (Domingue, Trejo, Armstrong-Carter, & Tucker-Drob, Reference Domingue, Trejo, Armstrong-Carter and Tucker-Drob2020; Dudbridge, Reference Dudbridge2013). Many recent psychiatry studies have utilized PRS approaches to investigate GEIs (Soga, Teo, & Parhar, Reference Soga, Teo and Parhar2021; Woolway et al., Reference Woolway, Smart, Lynham, Lloyd, Owen and Jones2022; Yao, van der Veen, Thygesen, Bass, & McQuillin, Reference Yao, van der Veen, Thygesen, Bass and McQuillin2023). In SZ, prior studies have reported positive interactions between the PRS for SZ (PRS-SZ) and factors such as cannabis use (Gage et al., Reference Gage, Jones, Burgess, Bowden, Davey Smith, Zammit and Munafò2017; Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019; Pasman et al., Reference Pasman, Verweij, Gerring, Stringer, Sanchez-Roige and Treur2018), urbanicity (Colodro-Conde et al., Reference Colodro-Conde, Couvy-Duchesne, Whitfield, Streit, Gordon and Kemper2018; Maxwell, Coleman, Breen, & Vassos, Reference Maxwell, Coleman, Breen and Vassos2021; Paksarian et al., Reference Paksarian, Trabjerg, Merikangas, Mors, Børglum and Hougaard2018), and CA (Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019; Mas-Bermejo et al., Reference Mas-Bermejo, Papiol, Via, Rovira, Torrecilla and Kwapil2023; Pignon et al., Reference Pignon, Peyre, Ayrolles, Kirkbride, Jamain and Ferchiou2022; Pries et al., Reference Pries, Klingenberg, Menne-Lothmann, Decoster, van Winkel and Collip2020a; Pries et al., Reference Pries, Klingenberg, Menne-Lothmann, Decoster, van Winkel and Collip2020b; Saarinen et al., Reference Saarinen, Hietala, Lyytikäinen, Hamal Mishra, Sormunen and Lavonius2024; Sallis et al., Reference Sallis, Croft, Havdahl, Jones, Dunn and Davey Smith2021; Smigielski et al., Reference Smigielski, Papiol, Theodoridou, Heekeren, Gerstenberg and Wotruba2021; Trotta et al., Reference Trotta, Iyegbe, Di Forti, Sham, Campbell and Cherny2016). However, contrasting findings also exist; some authors have reported negative associations between PRS-SZ and cannabis use (Johnson et al., Reference Johnson, Colbert, Jeffries, Tillman, Bigdeli and Karcher2023) and CA (Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019; Mas-Bermejo et al., Reference Mas-Bermejo, Papiol, Via, Rovira, Torrecilla and Kwapil2023; Pignon et al., Reference Pignon, Peyre, Ayrolles, Kirkbride, Jamain and Ferchiou2022; Pries, Klingenberg, et al., Reference Pries, Klingenberg, Menne-Lothmann, Decoster, van Winkel and Collip2020a; Pries, Klingenberg, et al., Reference Pries, Klingenberg, Menne-Lothmann, Decoster, van Winkel and Collip2020b; Saarinen et al., Reference Saarinen, Hietala, Lyytikäinen, Hamal Mishra, Sormunen and Lavonius2024; Sallis et al., Reference Sallis, Croft, Havdahl, Jones, Dunn and Davey Smith2021; Smigielski et al., Reference Smigielski, Papiol, Theodoridou, Heekeren, Gerstenberg and Wotruba2021; Trotta et al., Reference Trotta, Iyegbe, Di Forti, Sham, Campbell and Cherny2016).
Although most studies have primarily focused on GEIs using single environmental factors, each factor represents a small portion of the dense network of possible environmental exposures. The introduction of a cumulative environmental score would provide a single measure, thereby enhancing risk prediction and advancing research to more fully comprehend the collective impact of the environment and its interactions with genetics in the context of psychosis. Several tools have been developed to measure cumulative environmental load in the form of a single aggregate score, similar to the PRS used in genetics. Examples include the exposome score for SZ (ES-SZ) (Cuesta et al., Reference Cuesta, Papiol, Ibañez, García de Jalón, Sánchez-Torres and Gil-Berrozpe2023; Pries et al., Reference Pries, Dal Ferro, van Os, Delespaul, Kenis and Lin2020), the Maudsley Environmental Risk Score (ERS) (Vassos et al., Reference Vassos, Sham, Kempton, Trotta, Stilo and Gayer-Anderson2020), the Polyenviromic Risk Score (PERS) (Padmanabhan, Shah, Tandon, & Keshavan, Reference Padmanabhan, Shah, Tandon and Keshavan2017), and the Psychosis Polyrisk Score (PPS) (Oliver et al., Reference Oliver, Spada, Englund, Chesney, Radua and Reichenberg2020; Oliver, Radua, Reichenberg, Uher, & Fusar-Poli, Reference Oliver, Radua, Reichenberg, Uher and Fusar-Poli2019). While these tools incorporate similar risk factors, they differ in how the aggregate score is calculated. ERS and PPS scores are estimated by scaling odds ratios (ORs) or relative risks (RRs) with population prevalence for each risk factor, and the PERS score is obtained by simply summing the ORs of the risk factors. ES-SZ employs a more advanced approach using weighted coefficients derived from a single predictive model, which accounts for the interdependency of exposures (Pries et al., Reference Pries, Lage-Castellanos, Delespaul, Kenis, Luykx and Lin2019). In contrast, the K-PERS-I score is calculated using ORs and RRs from Western studies, adjusted for the population proportions of risk factors identified from multiple resources of the Korean data. However, only a limited number of studies have investigated GEIs using these tools in SZ, and they have yielded mixed results (Cuesta et al., Reference Cuesta, Papiol, Ibañez, García de Jalón, Sánchez-Torres and Gil-Berrozpe2023; Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019; Mas et al., Reference Mas, Boloc, Rodríguez, Mezquida, Amoretti and Cuesta2020; Pries, Dal Ferro, et al., Reference Pries, Dal Ferro, van Os, Delespaul, Kenis and Lin2020). Notably, there has been a lack of studies examining GEIs in Asian SZ patients using these comprehensive measures of polyenvironmental factors.
We utilized the Korea-Polyenvironmental Risk Score-I (K-PERS-I) (Jeon et al., Reference Jeon, Kang, Piao, Kim, Kim and Lee2022), a comprehensive measure of multiple environmental factors associated with SZ, developed based on the proportions of risk factors (exposure) in the Korean population. Moreover, considering suggestions that GEIs can only involve monogenic factors (Caspi et al., Reference Caspi, Moffitt, Cannon, McClay, Murray and Harrington2005; Stefanis et al., Reference Stefanis, Henquet, Avramopoulos, Smyrnis, Evdokimidis and Myin-germeys2007) and comprehensive measures may not be appropriate tools for studying GEIs (Assary, Vincent, Keers, & Pluess, Reference Assary, Vincent, Keers and Pluess2018), we also adopted three specific environmental measures: subdomains of the K-PERS-I, as well as the total and subdomains of the Early Trauma Inventory-Self Report (ETI-SR) (Bremner, Bolus, & Mayer, Reference Bremner, Bolus and Mayer2007), which evaluates CA.
In this study, we aimed to calculate PRS-SZ using the Korean Genomics Center (KOGIC) dataset and East Asia (EAS) summary statistics from the Psychiatric Genomics Consortium (PGC). Subsequently, using the two aggregate scores (PRS-SZ and total K-PERS-I), we investigated the associations of two main factors and their interactions with case–control status, as well as with clinical phenotypes assessed using the Brief Core Schema Scales (BCSS) and the Positive and Negative Syndrome Scale (PANSS) in patients with schizophrenia spectrum disorders (SSDs). Additionally, we explored these associations using three specific environmental measures. The subsequent results were compared in order to see the possible advantages of using the comprehensive measure, K-PERS-I, over three specific environmental measures.
Methods
Participants
Patients (n = 818) were recruited from outpatients and inpatients treated at four hospitals (Jeonbuk [JB], Chonnam, Kyungpook, and Haeundae Paik National University Hospitals [NUH]) from December 2014 to February 2021. Inclusion criteria were as follows: (a) SSDs including SZ, schizoaffective disorder, schizophreniform disorder, and psychotic disorder not otherwise specified; (b) and age between 18 and 59 years. Diagnoses were established using the DSM-IV-TR criteria (American Psychiatric Association, 1994) and the Korean version of the Mini-International Neuropsychiatric Interview (Yoo et al., Reference Yoo, Kim, Noh, Oh, Kim and Namkoong2006). The exclusion criteria were as follows: (a) intelligence quotient (IQ) ≤ 70; (b) acute, unstable, or severe medical/neurological conditions; or (c) pregnancy or lactation. Healthy controls (HCs) were recruited through advertisements at JBNUH (n = 356) and underwent psychiatric interviews using the screening module of the Structured Clinical Interview for DSM-IV (First, Spitzer, Gibbon, & Williams, Reference First, Spitzer, Gibbon and Williams2002). HCs were excluded if they had a current or previous diagnosis of mental disorders, a clinically significant medical condition, or first-degree relatives with psychosis (to minimize the effects of genetic loading). The authors assert that all procedures contributing to this work complied with the ethical standards of relevant national and institutional Human Experimental Commissions and the 1975 Declaration of Helsinki, as amended in 2008. All procedures involving human subjects/patients were approved by the Ethics Committee of JBNUH (approval number CUH 2014-11-002).
Assessments
Psychopathology was evaluated using PANSS (Kay, Fiszbein, & Opler, Reference Kay, Fiszbein and Opler1987; Yi et al., Reference Yi, Ahn, Shin, An, Joo and Kim2001). To maximize inter-rater reliability across sites, psychiatrists with more than 3 years of experience in this field participated in the ratings process, and several workshops were held during the recruitment period (Cohen’s kappa ≥0.8). Self-rating scales, such as BCSS (Fowler et al., Reference Fowler, Freeman, Smith, Kuipers, Bebbington and Bashforth2006) and ETI-SR, were administered. For simplicity, we used only two subdomains of the ETI-SR, negative-self and negative-others, by reversing the scores of positive-self and positive-others. The reliability and validity of the Korean version of the ETI-SR were confirmed in Korean patients with depression (Cronbach’s alpha = 0.87) (Jeon et al., Reference Jeon, Lee, Lee, Jeong, Kim, Lee and Jeon2012). Environmental risk factors associated with SZ were measured using the K-PERS-I (Jeon et al., Reference Jeon, Kang, Piao, Kim, Kim and Lee2022), which consisted of six domains: paternal age at birth, parental socioeconomic status (pSES), obstetric complications, urbanicity, CA, and recent life events. For domain and scoring system definitions, see Supplementary Tables: Material 1.
Calculation of the PRS-SZ
The JBNUH genotype dataset (for details on the quality control of genetic data, see Supplementary Tables: Material 2 and Supplementary Figures 1 and 2), was converted from UCSC hg19 genomic coordinates to hg38 using CrossMap (v0.6.4). The KOGIC dataset of the Ulsan National Institute of Science and Technology consists of 28,692,913 SNPs in 1,047 Korean individuals, primarily from the Ulsan metropolitan region in the southern part of the Korean peninsula (Jeon et al., Reference Jeon, Bhak, Choi, Jeon, Kim and Jang2020). After the conversion, the dataset retained 437,188 SNPs. KOGIC genotype data were used as a reference panel for genotype imputation in the JBNUH dataset, conducted in two stages. First, pre-phasing of each chromosome was performed using Eagle (v2.4.1). Subsequently, Minimac3 (v2.0.1) was used to impute phased genotypes for approximately 28.7 million markers in the KOGIC genotype data, retaining variants with an imputation INFO score > 0.8. A global PRS-SZ was then generated using East Asian (EAS) summary statistics (Lam et al., Reference Lam, Chen, Li, Martin, Bryois and Ma2019) and KOGIC data as the linkage disequilibrium (LD) reference panel. For the PRS-SZ calculation, we utilized 2,525,966 SNPs that overlapped with the SNPs in the EAS summary statistics. Instead of limiting the analysis to significant SNPs, we included all available SNPs. The effect sizes were reweighted using the LDpred-inf algorithm to account for causal variants and non-infinitesimal genetic architecture. This approach incorporates a point-normal prior for posterior mean effect size estimation, implemented via the Markov Chain Monte Carlo (MCMC) method, which assumes a Gaussian mixture prior:
$ {\beta}_i\sim N\left(0,\frac{h_g^2}{M}\times p\right) $
with probability p, and
$ {\beta}_i\sim 0 $
with probability 1 − p, where p is the proportion of causal SNPs. The method is optimized by considering different values of p (10−3, 3 × 10−3, 0.01, 0.03, 0.1, 0.3, 1) (Privé, Arbel, & Vilhjálmsson, Reference Privé, Arbel and Vilhjálmsson2021; Vilhjálmsson et al., Reference Vilhjálmsson, Yang, Finucane, Gusev, Lindström and Ripke2015). The study population n = 1,073 participants (717 with SZ) was tested for the predictive performance of the global PRS-SZ. Personal scores were computed by coding each variant, weighting by its relative effect size on SZ, along with standardization. The PRS-SZ with the highest discriminative capability was chosen based on the maximum area under the curve (AUC). The estimated AUCs for different proportions of causal variants (10−3, 3 × 10−3, 0.01, 0.03, 0.1, 0.3, 1) were 0.529, 0.670, 0.695, 0.700, 0.699, 0.698, and 0.699, respectively. The LDpred model with a 3% proportion of causal variants demonstrated the highest predictive accuracy for SZ. Furthermore, age and sex adjustments were applied to the PRS-SZ. We calculated the heritability
$ \left({h}^2\right) $
for SZ with the Korean-specific LD scores and EAS summary statistics using LD score regression (Bulik-Sullivan et al., Reference Bulik-Sullivan, Loh, Finucane, Ripke, Yang and Patterson2015). The prediction accuracy for the PRS-SZ was assessed using AUC, which was subsequently converted to liability R 2, considering a population lifetime risk of SZ at 1% (Lee, Wray, Goddard, & Visscher, Reference Lee, Wray, Goddard and Visscher2011).
Statistical analyses
Gene–environment correlation between PRS-SZ and the subdomains of K-PERS-I in SSDs patients was assessed using a logistic regression model, adjusted for age, sex, and education. To examine the associations of PRS-SZ and/or total scores of K-PERS-I/ETI-SR with case–control status, we conducted four types of logistic regression models: (i) a genetic model with PRS-SZ only, (ii) an environmental model including either K-PERS-I or ETI-SR, (iii) an independent model with PRS-SZ and K-PERS-I/ETI-SR without an interaction term, and (iv) an interaction model with PRS-SZ and K-PERS-I/ETI-SR with an interaction term. For the total score of the K-PERS-I/ETI-SR, we performed analyses using a complete dataset and an imputed dataset. Interaction was assessed in two ways: first, through logistic regression to evaluate interaction as a departure from multiplicativity, and second, through the Relative Excess Risk due to Interaction (RERI) to assess interaction as a departure from additivity (Knol & VanderWeele, Reference Knol and VanderWeele2012). An RERI >0 indicated a positive deviation from additivity and was considered statistically significant if the 95% confidence interval (CI) did not contain zero. The RERIs were calculated using the delta method, which imposes an unnatural symmetry on the confidence limits of the underlying effect measure (Hosmer & Lemeshow, Reference Hosmer and Lemeshow1992), using the ORs derived from each model via the epiR package in R. In all analyses, we adjusted for covariates such as age, sex, and education. Multiple testing for subdomains of the K-PERS-I was corrected with false discovery rate (FDR). Nagelkerke’s R 2 value for each model was calculated with the DescTools package in R (version 0.99.52).
Additionally, logistic regression analyses were conducted with dichotomized PRS-SZ and K-PERS-I/ETI-SR scores, using a threshold of 75% of the HC group. This threshold was selected to more effectively capture the impact of high genetic risk while minimizing the potential masking of effects by intermediate or low scores (Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019; Mas et al., Reference Mas, Boloc, Rodríguez, Mezquida, Amoretti and Cuesta2020; Pries, Dal Ferro, et al., Reference Pries, Dal Ferro, van Os, Delespaul, Kenis and Lin2020; Segura et al., Reference Segura, Mezquida, Martínez-Pinteño, Gassó, Rodriguez and Moreno-Izco2023). The subgroup with low K-PERS-I/ETI-SR and low PRS-SZ with scores less than 75% was regarded as the reference group. Sensitivity analyses were performed using alternative thresholds of 50% and 25%. We also conducted the same analysis separately for males and females. Moreover, considering that simply adjusting for these covariates could be insufficient to eliminate potential bias in the Gene × Environment interaction term (Keller, Reference Keller2014), we performed further analyses with these covariates included as interaction terms and assessed the improvement of model fit with the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC).
To investigate the associations of PRS-SZ and K-PERS-I/ETI-SR (total and subdomains) and their interactions with clinical phenotypes, PANSS, and BCSS, linear regression with the interaction model was utilized. Multiple testing correction was applied to all analyses using the FDR method, with a significance threshold of FDR < 0.05. As we had smaller sample sizes for the total K-PERS-I/ETI-SR because of different missing data in the subdomains of the K-PERS-I/ETI-SR, we conducted the main analyses again after the imputation using the mice package in R (version 3.16.0). In all analyses, we used R software (ver. 4.1.1) and adjusted for age, sex, and education as covariates.
For power calculation, we performed post hoc power calculation (Rodriguez et al., Reference Rodriguez, Alameda, Aas, Gayer-Anderson, Trotta and Spinazzola2024) for the PRS-SZ by K-PERS-I interaction using a simulation method in R using the standardized coefficients for PRS-SZ, K-PERS-I, and PRS-SZ * K-PERS-I, from the comparison of case–control (For details on the code used, see Supplementary Tables: Materials 3 and 4).
Results
Demographic and clinical characteristics of the participants
The participants were divided into three groups: SSDs with K-PERS-I (n = 295–398), SSDs with ETI-SR (n = 474–476), and HCs (n = 320–356). Compared with HCs, the SSDs with K-PERS-I group were younger (p < 0.001); it had higher negative-self and negative-others scores (p < 0.001), and higher total and subdomains K-PERS-I scores (p < 0.001), with the exception of pSES. Similarly, compared with the HCs, the SSDs with ETI-SR group were younger (p < 0.001); it also had higher negative-self and negative-others scores (p < 0.001) and lower total and subdomain ETI-SR scores. Comparison of the SSDs with K-PERS-I and the SSDs with ETI-SR groups revealed no significant differences (Table 1).
Demographic and clinical characteristics of the participants
a SSDs with K-PERS-I (n = 280 ~ 398) versus HCs (n = 320 ~ 356); bSSDs with ETI-SR (n = 432 ~ 476) versus HCs (n = 320 ~ 356); cSSDs with K-PERS-I (n = 280 ~ 398) versus SZ with ETI-SR (n = 432 ~ 476); 269 SSDs patients with K-PERS-I group (n = 295) are overlapping with SSDs patients with ETI group (n = 474).
Note: BCSS, Brief Core Schema Scales; DI, duration of illness; DUP, duration of untreated psychosis; ETI-SR, Early Trauma Inventory Self Report; HCs, healthy controls; K-PERS-I, Korea Polyenvironmental Risk Score-I; NUH, National University Hospital; PANSS, Positive and Negative Syndrome Scale; PNOS, psychotic disorder not otherwise specified; SES, socioeconomic Status; SSDs, schizophrenia spectrum disorders; SSDs, schizophrenia spectrum disorders; All values were mean ± standard deviation (SD) unless otherwise specified.
AUC and liability R2 for PRS-SZ versus PRS-SZ and K-PERS-I/ETI-SR
The PRS-SZ alone showed an AUC of 0.724 and a liability R 2 of 8.7% (Supplementary Figure 3). However, when incorporating both PRS-SZ and K-PERS-I, we observed an AUC of 0.832 and a liability R 2 of 22.6%. Combining PRS-SZ and ETI-SR resulted in moderate values, with an AUC of 0.766 and a liability R 2 of 12.8% (Table 2). The SNP heritability for SZ calculated from the Korean-specific LD scores and EAS summary statistics was 0.20 ± 0.01.
AUC and liability R 2 of PRS-SZ versus PRS-SZ and K-PERS-I/ETI-SR on case–control status
All variables were adjusted for age, sex, and education.
Note: AUC, area under the receiver operating curve; ETI-SR, Early Trauma Inventory Self Report; HCs, healthy controls; K-PERS-I, Korea Polyenvironmental Risk Score-I; PRS-SZ, Polygenic Risk Score-Schizophrenia; SSDs, schizophrenia spectrum disorders.
Gene–environment correlation
We tested for correlations between PRS-SZ and K-PERS-I subdomains. No significant correlations were found in the patient group (Supplementary Table 1).
Main and interaction effects of PRS-SZ and K-PERS-I/ETI-SR on case–control status
When using the K-PERS-I as an environmental measure, the main effects in all four models were significantly associated with case–control status (PRS-SZ: adjusted OR = 2.20, 95% CI = 1.74–2.80, p = 1.68 × 10−11, K-PERS-I: adjusted OR = 3.26, 95% CI = 2.66–4.05, p = 7.41 × 10−28, PRS-SZ + K-PERS-I: adjusted OR = 3.39, 95% CI = 2.73–4.25, p = 7.01 × 10−27, PRS-SZ + K-PERS-I + PRS-SZ × K-PERS-I: adjusted OR = 3.37, 95% CI = 2.71–4.25, p = 3.74 × 10−26). No significant multiplicative interaction was observed, whereas a significant additive interaction was identified with a RERI of 3.20 (95% CI: 0.29–6.12, p = 0.015) (Figure 1a). Nagelkerke’s R 2 values were highest for the independent and interaction models. Simulation using the standardized regression coefficients for the PRS-SZ by K-PERS-I multiplicative interaction yielded a power estimate of 5.6%. It was estimated that a minimum sample size of 180,000 individuals would be required to achieve 80% power. In contrast, for the additive interaction term, the power was approximately 88%, and the minimum sample size required to achieve 80% power was 520 individuals. When using ETI-SR as the environmental measure, all four models showed significant results, but no significant interaction effects were identified. Notably, the ORs for ETI-SR associated with SZ risk were lower than those observed for K-PERS-I (Table 3).
(a) Additive effects of PRS-SZ and K-PERS-I, (b) additive effects of PRS-SZ and pSES, (c) additive effects of PRS-SZ and CA, and (d) additive effects of PRS-SZ and Recent life events. Note: CA, childhood adversity; K-PERS-I, Korea Polyenvironmental Risk Score-I; PRS-SZ, Polygenic Risk Score-Schizophrenia; pSES, parental socioeconomic status; RERI, relative excess risk due to interaction.
Main and interaction effects of PRS-SZ and K-PERS-I/ETI-SR on case–control status
All variables were adjusted for age, sex, and education.
Note: ETI-SR, Early Trauma Inventory Self Report; HCs, healthy controls; K-PERS-I, Korea Polyenvironmental Risk Score-I; OR, odd ratio; PRS-SZ, polygenic risk score-schizophrenia; RERI, relative excess risk due to interaction; SSDs, schizophrenia spectrum disorders.
In additional analyses using dichotomized PRS-SZ and K-PERS-I/ETI-SR with thresholds set at 75% of the HC group, the subgroup with high K-PERS-I and high PRS-SZ showed the highest OR (10.95, 95% CI = 5.89–20.35, p < 0.001) compared with the other three subgroups. For the subgroup with high ETI-SR and high PRS-SZ, the pattern was generally similar, but the OR was lower compared with the subgroup with a high K-PERS-I and high PRS-SZ (Figure 2). Sensitivity analyses using thresholds of 50% and 25% of the HC group showed similar results (Supplementary Figure 4). We found no different results between genders (Supplementary Table 2). Further analyses including three covariates (age, sex, and education) *GEss showed no significant results in all four models (Supplementary Table 3). AIC and BIC values of the new models assessing model fit were increased compared to the original models (Supplementary Table 4).
Regression results of subgroups of the K-PERS-I /ETI-SR and PRS-SZ were divided with a 75% cutoff of the control group. Note: ETI-SR, Early Trauma Inventory Self Report; K-PERS-I, Korea Polyenvironmental Risk Score-I; N, number; OR, odd ratio; PRS-SZ, Polygenic Risk Score-Schizophrenia.
When using the K-PERS-I subdomains in the interaction models, both PRS-SZ and each subdomain were significantly associated with SZ risk. However, a significant multiplicative interaction with PRS-SZ was observed for pSES only (adjusted OR = 1.39, 95% CI =1.17–1.66, p = 2.70 × 10−4, FDR = 0.002). Notably, its Nagelkerke’s R 2 value was 19.2%. Regarding additive interactions, PRS-SZ showed significant associations with pSES (RERI = 1.27, 95% CI = 0.32–2.23, p = 0.005, FDR = 0.013), childhood adversity (RERI = 3.69, 95% CI = 1.08–6.30, p = 0.003, FDR = 0.013), and recent life events (RERI = 1.50, 95% CI = 0.13–2.88, p = 0.016, FDR = 0.027) (Figure 1a–d, Supplementary Table 5). The additive interaction between PRS-SZ and pSES showed a power of 97.3%, with a minimum sample size of 450 individuals required to achieve 80% power. When using the ETI-SR subdomains, similar main effects were observed; however, no significant interactions remained after FDR correction (Supplementary Table 6).
Main and interaction effects of PRS-SZ and K-PERS-I/ETI-SR on the PANSS and BCSS in patients with SSDs
For the PANSS, an interaction model with the K-PERS-I showed neither main effects nor additive interaction. However, in the model with the ETI-SR, PRS-SZ was significantly associated with the total score (Beta = 2.66, SD = 1.26, p = 0.036, FDR = 0.036). In contrast, the total ETI-SR score was only associated with positive symptom score (Beta = 0.93, standard deviation (SD) = 0.28, p = 0.001, FDR = 0.003). No significant interaction was evident. For the BCSS, the interaction model with the K-PERS-I showed neither main effects nor interaction. However, in the model with the ETI-SR, the total ETI-SR score was significantly associated with negative-self (Beta = 0.99, SD = 0.19, p = 4.14 × 10−7, FDR = 4.14 × 10−7) and negative-others (Beta = 1.38, SD = 0.18, p = 1.34 × 10−13, FDR = 2.68 × 10−13) (Table 4).
Main and interaction effects of PRS-SZ and K-PERS-I/ETI-SR on the PANSS or BCSS in patients with SSDs
All variables were adjusted for age, sex, and education.
Note: BCSS, Brief Core Schema Scales; ETI-SR, Early Trauma Inventory Self Report; FDR, false discovery rate; K-PERS-I, Korea Polyenvironmental Risk Score-I; N, number of patients with SSDs; PANSS, Positive and Negative Syndrome Scale; PRS-SZ, polygenic risk score-schizophrenia; SSDs, schizophrenia spectrum disorders.
Additional analyses of the PANSS using K-PERS-I subdomains revealed no significant main effect of PRS-SZ, and no interaction effects remained significant after FDR correction. For the BCSS, no main effect of the PRS-SZ was evident, whereas CA was significantly associated with both negative-self (Beta = 0.72, SD = 0.27, p = 0.008, FDR = 0.008) and negative-others (Beta = 1.04, SD = 0.25, p < 0.001, FDR < 0.001). Intriguingly, a significant positive interaction between PRS-SZ and obstetric complications was identified in association with negative-others (Beta = 0.55, SD = 0.19, p = 0.004, FDR = 0.008) (Supplementary Table 7). In the model with ETI-SR subdomains, PRS-SZ was significantly associated with total, negative symptom, and general psychopathology scores; several subdomains of the ETI-SR were significantly associated with total, positive symptom, and general psychopathology scores. No significant interaction was noted. For the analyses with the BCSS, significant associations with negative-self and negative-others were only noted for the ETI-SR subdomains (Supplementary Table 8). After imputation, the results for Tables 3 and 4 were the same (Supplementary Tables 9–11).
Discussion
Studies of GEIs may partly explain why SZ only develops in some people who experience relevant environmental exposures. Here, we investigated associations of PRS-SZ and total K-PERS-I and their interactions with case–control status and clinical phenotypes. Additionally, we investigated these associations using the K-PERS-I subdomains, as well as the total and subdomains of the ETI-SR.
In the present study, the PRS-SZ explained approximately 9% of SZ risk, which is similar to the reported 7% of the variance on the liability scale (Ripke et al., Reference Ripke, Neale, Corvin, Walters, Farh and Holmans2014) but higher than other reported rates of ~3% (Lam et al., Reference Lam, Chen, Li, Martin, Bryois and Ma2019) or 3.4% (Agerbo et al., Reference Agerbo, Sullivan, Vilhjálmsson, Pedersen, Mors and Børglum2015). When the K-PERS-I and interaction term (PRS-SZ × K-PERS-I) were combined, the proportion of explained risk increased to 22.6%. This finding suggests that approaches considering both genetic and environmental factors and their interactions could enhance the accuracy of SZ risk prediction. Regarding the ORs for PRS-SZ and K-PERS-I in Table 3, the values of the K-PERS-I from the environmental model and two combined models (PRS-SZ + K-PERS-I and PRS-SZ + K-PERS-I + PRS-SZ × K-PERS-I) were greater than the values of PRS-SZ from the genetic model. These findings indicate that the contributions of environmental factors and the combined effects of genetic and environmental factors in association with SSD risk are greater than the contributions of genetic effects alone. However, this pattern was not observed when the ETI-SR was regarded as an environmental factor. The ORs of the ETI-SR from the environmental model and two combined models were smaller than the ORs of PRS-SZ from the genetic model. Moreover, the ORs of the ETI-SR were smaller than the ORs of the K-PERS-I. These findings suggest that using the K-PERS-I, a more comprehensive tool that measures multiple environmental factors, may be more effective than using the ETI-SR for predicting SSD risk. However, this interpretation should be approached with caution, as the sample sizes between the two datasets differ.
For the interaction with the K-PERS-I, we observed contrasting results between multiplicative versus additive models. Considering that an additive model can provide a superior representation of biological synergy and inform public health decisions within the sufficient cause framework (Kendler & Gardner Reference Kendler and Gardner2010; Rothman, Reference Rothman1976), the result of the additive model needs to be more highlighted. It suggests that the combined effect of genetic and environmental risk factors is greater than the sum of their individual effects, which is in align with the findings of GEI studies using the ES-SZ (Pries, Dal Ferro, et al., Reference Pries, Dal Ferro, van Os, Delespaul, Kenis and Lin2020) and ERS (Mas et al., Reference Mas, Boloc, Rodríguez, Mezquida, Amoretti and Cuesta2020). The implication of this synergistic interaction should be incorporated into psychoeducation and public health policy. The lack of significant interactions with the ETI-SR in both models may be due to different characteristics of the two measures: several subdomains of the K-PERS-I capture pre/perinatal or chronic exposures that are likely to affect or interact with genetic risk, while ETI-SR focuses on acute events, which may act independently of genetic predisposition. This emphasizes the need to consider the duration and nature of environmental exposures in GEI studies. One meta-analysis reported that the current four studies on the interaction between PRS and childhood adversity yielded inconsistent findings (Woolway et al., Reference Woolway, Smart, Lynham, Lloyd, Owen and Jones2022). Interestingly, when PRS-SZ and K-PERS-I/ETI-SR were dichotomized using a threshold of 75% of the control group value, the subgroups with high K-PERS-I/ETI-SR and high PRS-SZ showed the highest ORs, approximately 9–10, compared with the other three subgroups. If replicated in population-based cohort studies, these results could have important implications for the early identification of individuals at risk of developing SSDs and the provision of more intensive psychosocial interventions that target modifiable environmental factors in patients with SSDs.
When using the subdomains of the K-PERS-I, we found a significant interaction, departing from both multiplicativity and additivity, between PRS-SZ and pSES in association with SSD risk, indicating a synergistic effect. This suggests that the impact of pSES on SSD risk is moderated by PRS-SZ. To our knowledge, this is the first report demonstrating a significant interaction between pre-existing genetic liability and pSES in association with SSD risk, although a negative finding has been documented (Agerbo et al., Reference Agerbo, Sullivan, Vilhjálmsson, Pedersen, Mors and Børglum2015). Notably, Hatzimanolis et al., Reference Hatzimanolis, Stefanatou, Kattoulas, Ralli, Dimitrakopoulos and Foteli2020 found a significant interaction between familial risk of psychosis and pSES in terms of influencing social premorbid adjustment in childhood among patients with psychosis. There could be several explanations for this synergistic effect; poor pSES may lead to lower education, neighborhood deprivation, and health inequity relative to an affected individual’s siblings, all of which are factors related to mental health (González et al., Reference González, Estarlich, Murcia, Barreto-Zarza, Santa-Marina and Simó2023; Luo, van Grieken, Yang-Huang, van den Toren, & Raat, Reference Luo, van Grieken, Yang-Huang, van den Toren and Raat2022). Moreover, the cumulative effects of these interconnected factors may contribute to a synergistic effect, or the pSES may have exerted detrimental effects on SSD-associated mutants to produce more dysfunctional changes through epigenetic regulation. We identified significant additive interactions between PRS-SZ and both childhood adversity and recent life events. A synergistic interaction result between PRS-SZ and childhood adversity is in line with previous studies (Aas et al., Reference Aas, Alameda, Di Forti, Quattrone and Dazzan2023; Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx and Lin2019). Furthermore, given the interaction between PRS-major depressive disorder and childhood trauma (Peyrot et al., Reference Peyrot, Milaneschi, Abdellaoui, Sullivan, Hottenga and Peyrot2014), it reinforces the idea that genetic risk and early-life stressors can act together to increase susceptibility to mental health disorders. Similarly, recent life events, like stress or loss, can trigger or exacerbate genetic susceptibility to SZ. However, a study found that ES-SZ moderated the link between stressful life events and mental health, while PRS-SZ did not (Pries et al., Reference Pries, van Os, Ten Have, de Graaf, van Dorsselaer and Bak2020), highlighting the complexity of GESs. For the contribution of the K-PERS-I subdomains to SZ risk, all six factors were significant. Given that high paternal age is a risk factor for SZ and the average marriage age is increasing these days, this should be considered in psychoeducation and public health policy. Unexpectedly, urbanicity, which is strongly associated with SZ risk in other populations (Plana-Ripoll, Pedersen, & McGrath, Reference Plana-Ripoll, Pedersen and McGrath2018), was found to be a protective factor. This may be explained by easy access to mental healthcare services (Park, Park, Kwon, Kang, & Noh, Reference Park, Park, Kwon, Kang and Noh2016) and lower rates of drug abuse (Jang et al., Reference Jang, Lee, Park, Kang, Lee and Kim2023) in Korea compared to Western countries.
In the model with the K-PERS-I, neither PRS-SZ nor K-PERS-I were associated with the PANSS or BCSS, but in the model with the ETI-SR, PRS-SZ and ETI-SR were associated with several items. These findings suggest that the K-PERS-I is not an appropriate tool for investigating its association with clinical phenotypes. Concerning the results with the model that included the ETI-SR, PRS-SZ was significantly associated with the total score of the PANSS. Studies using PRS-SZ have yielded conflicting findings regarding associations with the symptom dimensions of SZ. Evidence supporting an association between PRS-SZ and positive symptoms is scarce, whereas studies of negative symptoms have led to mixed results (Mistry, Harrison, Smith, Escott-Price, & Zammit, Reference Mistry, Harrison, Smith, Escott-Price and Zammit2018; Ronald & Pain, Reference Ronald and Pain2018). Considering these inconclusive findings and the low prediction accuracy (~6%) of PRS-SZ for genetic variation in psychiatric phenotypes (Mistry et al., Reference Mistry, Harrison, Smith, Escott-Price and Zammit2018), the utility of PRS-SZ in clinical settings appears limited. In contrast, ETI-SR was associated with positive symptoms, negative-self, and negative-others. Because negative evaluations about the self and others or negative self-concepts are associated with the formation of delusions and hallucinations (Garety, Kuipers, Fowler, Freeman, & Bebbington, Reference Garety, Kuipers, Fowler, Freeman and Bebbington2001; Smith et al., Reference Smith, Fowler, Freeman, Bebbington, Bashforth and Garety2006), this finding may provide insights about how CA is involved in developing positive symptoms. Additionally, albeit at an uncorrected level, a synergistic effect between PRS-SZ and ETI-SR in predicting negative-self was observed. Taken together with the case–control status results, the model with the K-PERS-I may be better in terms of discriminating case–control status, whereas the model with the ETI-SR may be more suitable for predicting symptomatology.
In the subdomain models, CA of the K-PERS-I was associated with negative-self and negative-others. The positive interaction of PRS-SZ and obstetric complications in predicting negative-others is intriguing in that the association with negative-others only manifested when both genetic and environmental factors played a role; it was absent when each set of factors was present in isolation. For the synergistic effect, it is possible that individuals with genetic liability and obstetric complications have a higher likelihood of exposure to negative life events, which may lead to the formation of negative-others. Regarding the results with the model that included subdomains of the ETI-SR, the patterns were similar to those in the model with total ETI-SR.
Although this was the first study of GEIs using the K-PERS-I, several limitations should be considered. First, careful interpretation of the results is required, given that the sample size may have been insufficient for detecting multiplicative interaction potentially leading to type II errors. However, it is of note that in terms of additive interaction, the sample size was adequate. In addition, although we addressed missing data in the total K-PERS-I or ETI-SR by imputation, caution is needed when interpreting comparisons between total scores and subdomains due to differing sample sizes. Second, the study’s cross-sectional design did not allow for an investigation of the dynamic nature of the GEIs over time, hindering causal inference. Future research should focus more extensively on prospective cohort studies, including samples at different time points. Third, as we excluded foreigners to participate, the ethnic homogeneity of the sample, comprising original residents of South Korea, limits the generalizability of the findings to other populations. Fourth, although widely used, the RERI has limitations. Since it is based on ratios, it cannot be directly interpreted as the combined effects of exposures (Zhao & Thompson, Reference Zhao and Thompson2023). Fifth, K-PERS-I does not account for the interdependency of exposures like ES-SZ. Additionally, it excludes cannabis use and immigration, which may limit its ability to fully capture environmental influences on SZ risk and its applicability across diverse populations. Finally, the retrospective assessment of the K-PERS-I and ETI-SR may be associated with recall bias. Despite these limitations, this study has notable strengths, including being the first GEI study conducted in Korea and Asian countries and uniquely investigating and comparing the utilities of two environmental measures in the models.
In conclusion, using the aggregate scores of PRS-SZ and K-PERS-I, we did not observe any interactions in association with case–control status. However, when using the subdomains of the K-PERS-I or total and subdomains of the ETI-SR, several interactions were identified in association with case–control status and clinical phenotypes. These findings suggest that the aggregate scores of genetic and environmental measures, PRS-SZ and K-PERS-I, are unsuitable for investigating GEIs, which were instead detected with specific environmental measures. However, the contribution of the total K-PERS-I to case–control status was greater than the contribution of the total ETI-SR. There is a need for additional research concerning GEIs and modifiable environmental factors.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291725000753.
Acknowledgments
The corresponding and first authors would like to thank all participants in the study and family for their guidance and support (SDG). This study was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (HR18C0016) and by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00335201). Additionally, this work was supported by the National Research Foundation (NRF) grant, funded by the Korea government (2020R1C1C1A01012657 and 2021R1A6A1A10044154).
Competing interests
The authors declare no competing interests exist.
Open access
