A total of 255 Korean volunteers were included in this study. Twenty-four participants were excluded due to missing values for some psychological measures and brain images, which left a final sample of 231. Healthy participants numbered 137 (59.3%), recruited from community advertisements, and patients with PTSD numbered 94 (40.7%) and recruited from notices on the bulletin board in the hospital. The PTSD patients reported their traumatic events as follows: 63 (67.02%) severe motor vehicle accidents, 9 (9.57%) defections from North Korea, 8 (8.51%) physical or sexual violence, 8 (8.51%) relationship issues, 5 (5.32%) death of a family member, and 1 (1.06%) fire. We mixed PTSD patients and healthy participants with PTE experiences in the final sample to make data on PTEs and PTSD symptoms normally distributed and to investigate the phenomenon at the level of the general population.

The total participants consisted of 75 (32.5%) men and 156 (67.5%) women, with a mean age of 46.13 years (standard deviation, SD = 13.51). The mean years of education were 13.04 (SD = 3.27). Each participant signed a written informed consent form before participating in the study. All the experimental protocols were approved by the Institutional Review Board at Inje University Ilsan Paik Hospital (IRB no. 2015 - 07-025). All measurements and experiments were carried out in accordance with the guidelines and regulations of the board and the Declaration of Helsinki.

The Korean-validated version of the Life Events Checklist (LEC) was used to assess the experience of PTEs (44). The LEC comprises 17 items about PTEs, and the responses include experiencing, witnessing, and learning about a PTE. This study analyzed responses to experiencing PTE, since recalling witnessing or learning about it throughout their lifetime seemed difficult and confusing to some respondents. For each of the 17 items, responses were coded as 1 if the event was experienced and 0 if not, and the sum of these scores was used in subsequent analyses.

To examine the severity of PTSD symptoms, we administered the Clinician-Administered PTSD Scale for DSM - 5 (CAPS - 5) (45). It consists of 30 items and provides information about the frequency and severity of PTSD symptoms. The severity of PTSD symptoms is rated from 0 (“absent”) to 4 (“extreme/incapacitating”). The total severity score was obtained by adding up all the item scores for each individual. A licensed psychiatrist measured PTEs using LEC and PTSD symptom severity using CAPS through a 1:1 interview, and licensed nurses or a clinical psychologist observed the entire process to secure reliability.

To adjust for the possible confounding effect of alcohol problems from other variables, the Alcohol Use Disorders Identification Test (AUDIT) was used to assess alcohol consumption, drinking behaviors, and alcohol-related problems as a covariate. The AUDIT is a 10-item screening tool developed by the World Health Organization and is well-validated in the Korean population (46). The AUDIT is assessed on a 5-point Likert scale ranging from 0 (“never”) to 4 (“4 or more times a week”).

We used the same DNA genotyping process as in previous studies (12, 21). Participants had their blood sampled to extract DNA using a NanoDrop^® ND - 1000 UV-Vis Spectrophotometer. Then, genomic DNA was diluted to a 5 ng/µl concentration on 96-well polymerase chain reaction (PCR) plates. TaqMan SNP genotyping assays were obtained from Applied Biosystems. Afterward, the probes were labeled with either the FAM or the VIC dye at the 5’ end and with a minor-groove binder and a non-fluorescent quencher at the 3’ end. PCR was performed in 5 μl of the mixture containing 2 μl of a DNA sample, 0.125 μl of each TaqMan™ SNP Genotyping Assay (Thermo Fisher Scientific, USA), 2.5 μl of the TaqMan™ Genotyping Master Mix (Thermo Fisher Scientific, USA), and 0.375 μl of distilled water. Amplification and detection were carried out with a detection system (QuantStudio 12K Flex Real-Time PCR System, Thermo Fisher Scientific, USA), using the following profile: 50 °C for 2 min, 95 °C for 10 min, followed by 60 cycles of 95 °C for 15 sec, and 60 °C for 1 min. After the PCR amplification, allelic discrimination (an endpoint plate read) was performed on the same machines (the QuantStudio 12K Flex Real-Time PCR System). The QuantStudio 12K Flex Software calculated the fluorescence measured during the plate read and plotted Rn values based on the signals coming from each well. Subsequently, automatic or manual allele calls were performed on the analyzed plates. Three positive samples and one negative control sample were present on each plate. We confirmed the clustering image with positive controls. Intra-genomic DNA (gDNA) samples of known genotypes were used for the positive control. Guided by two recent comprehensive studies (22, 23) and other multiple empirical reports, we therefore selected the following nine genes (18 SNPs) for inclusion in our IG‐GSCA model: SLC6A4, FKBP5, ADCYAP1R1, BDNF, COMT, HTR3A, DRD2, NR3C1, and OXTR. A total of nine genes with 18 SNPs were analyzed and are described in Table 1 . For further analyses, the number of minor alleles was coded for each genotype: 0 for wild types with zero minor alleles, 1 for hetero types with one minor allele, and 2 for mutant types with two minor alleles.

MRI was performed using a 1.5 T scanner (Magneton Avanto, Siemens, Erlangen, Germany). Head motion was minimized with restraining foam pads provided by the manufacturer. High-resolution T1-weighted MRI images were acquired with the acquisition parameters of a 227 × 384 acquisition matrix, a 210 × 250 field-of-view, 0.9 × 0.7 × 1.2 voxel size, a total of 87,168 voxels, a TE of 3.42 ms, a TR of 1,900 ms, 1.2-mm slice thickness, and a flip angle of 15°.

We used the same MRI process as the one used in the previous study (47). Images were inspected visually for motion or other artifacts before and after preprocessing. The voxel-based volumetry (VBM) was conducted using CAT12 (http://dbm.neuro.uni-jena.de/cat/) implemented in SPM12 (Wellcome Department of Cognitive Neurology, London, UK). SPM12 tissue probability maps were used for the initial spatial registration. The structural T1 images were regularized with an ICBM East Asian template and normalized using the DARTEL algorithm (48). The images were then segmented into gray matter, white matter, and cerebrospinal fluid (49). Jacobian-transformed tissue probability maps were used to modulate images. The volume of the brain parts was extracted using the Neuromorphometric atlas (http://neuromorphometrics.com/) from 142 parts from 67 ROIs. Divided by a subject’s total brain volume, the extracted volumes of the brain parts were transformed into the relative sizes of volumes to adjust for individual differences in overall brain volume (50). Among the 67 ROIs in total, we excluded seven subcortical ROIs that were far from cortical function related to PTSD; cerebral exterior for it is considered residual non‐brain tissue from skull‐stripping (51), cerebellar vermis, brain stem, and optic chiasm since they have not shown consistent volumetric abnormalities in PTSD voxel‐based morphometry meta‐analyses (52), cerebellar and cerebral white matter for they are conceptually distinct from the gray‐matter volumetric framework (53), and cerebrospinal fluid since it is considered as an indirect marker of brain atrophy rather than as a mediator in structural models (54). A total of 60 remaining ROIs as potential mediators for the pathways from the nine genes to PTSD in the model. The list of the 60 ROIs and the means and SDs of their volumes are provided in Table 2 .

We used a component-based structural equation modeling technique specifically developed for imaging genetics research, called Imaging Genetics Generalized Structured Component Analysis (IG-GSCA) (12, 43). IG-GSCA enables researchers to statistically model and test path-analytic relationships among various constructs, representing those constructs as weighted composites of observed variables—termed components. For instance, a gene can be viewed a genetic construct composed of multiple SNPs located within the gene (55), whereas a brain ROI can be seen as an imaging construct comprising the voxels within the ROI (56). IG-GSCA generates a component for each gene and ROI from the relevant SNPs and brain voxels, respectively, and specifies biologically plausible pathways between genes, ROIs, and behavioral phenotypes within a single modeling framework. It also accommodates various well-documented interactions between constructs (57) in imaging genetics studies, such as gene-gene interactions (epistasis; G×G) (58) or gene-environment interaction (G×E) (59). The observed genetic or imaging variables used to define are referred to as indicators of their respective constructs.

Unlike univariate approaches—which estimate parameters for each equation separately and thus require post hoc corrections (e.g., Bonferroni, Šidák, Benjamini–Hochberg) to control the family‐wise error—IG-GSCA jointly estimates all parameters within a unified multivariate model. This simultaneous estimation accounts for shared variance and inherently mitigates Type I error inflation without the need for external corrections. Moreover, it enables IG-GSCA to statistically test mediation effects, such as whether ROIs mediate the relationship between genes and behavioral phenotypes. These features make IG-GSCA especially well-suited for imaging genetics studies that examine pathways from genetic variations to changes in brain structure to behavioral outcomes (60).

Lastly, IG-GSCA incorporates regularization techniques (61) to address multicollinearity, a common issue in imaging genetics (62) For example, neighboring genes often exhibit high correlations, which may cause multicollinearity when included simultaneously as predictors for the same outcome. In general, multicollinearity leads to unstable parameter estimates with inflated standard errors, especially in studies with small sample sizes. Regularization helps prevents overfitting to sample-specific noise, yielding more stable estimates and better generalizability. Given our hypothesized model includes a large number of genetic and brain-imaging variables as well as their interaction terms —relative to a modest sample size (N = 231), we employed IG-GSCA with regularization. A simulation study using a model of comparable complexity (e.g., nine-gene components, nine G×E interactions, 60 ROI components, and one outcome) demonstrates IGSCA performs as expected with a sample size of N = 250 (12).

Since recent studies suggested that the change of brain structures could mediate the pathways from genes and environmental factors to psychiatric disorders (12), we hypothesized a structural model where brain ROIs with cortical structural features mediate the effects of genes and PTEs on PTSD symptom severity. A total of nine genes with 18 SNPs were selected based on their relevance to PTSD (25, 27, 29, 31, 33, 36, 37, 39, 41). The number of minor alleles was coded for each genotype: 0 for wild types with zero minor alleles, 1 for hetero types with one minor allele, and 2 for mutant types with two minor alleles for further analyses. Also, the interaction effects of genes and PTEs were additionally considered based on the previous findings (63, 64). Among the 67 ROIs from the whole brain, we excluded seven subcortical ROIs that are far from cortical function related to PTSD (i.e., cerebral exterior, cerebellar vermis, brain stem, optic chiasm, cerebellar and cerebral white matter, and cerebrospinal fluid) from the model, leaving a total of 60 ROIs as mediating variables in the model. In addition, we set age, gender, and alcohol-related problems measured with AUDIT as covariates to control for possible contaminating effects (65–68). In all, we contemplated the structural model involving the pathways from nine genes and PTEs to 60 brain ROIs, the paths from 60 brain ROIs and PTEs to PTSD symptom severity, and the paths from three covariates (gender, age, and AUDIT score) to 60 brain ROIs and PTSD symptom severity. The total number of pathways specified in the structural model was 1,384. The structural model is displayed in Figure 1 , in which each individual pathway between two constructs, termed a path coefficient parameter, is symbolized by an arrow between the two constructs.

The demographic and psychological characteristics of healthy and PTSD participants are presented in Table 3 . The two groups did not show statistically significant gender differences (t = 1.642, p = 0.252), age difference (t = 1.309, p = 0.192), or AUDIT (t = -1.144, p = 0.254), whereas their PTE and PTSD scores were significantly different (t = -4.823, p < 0.001 for PTEs; t = -13.348, p < 0.001 for PTSD). Meanwhile, all variables, including both LEC scores and PTSD symptom severity scores, were normally distributed even for the total pool of participants since each skewness score was below 2.0 and each kurtosis score was below 7.0.

We tested an IG-GSCA model in which nine candidate genes (18 SNPs), their interactions with PTEs, and three covariates (age, gender, AUDIT score) were linked to PTSD symptom severity via 60 cortical ROIs. In total, the model specified 1,384 pathways, and its structure is shown in Figure 1 .

Each set of indicators for genes and ROIs is described in Tables 1 and 2 . The number of SNP indicators per gene ranged from 1 to 8. Each ROI construct was composed of the corresponding ROIs on both brain hemispheres except the inferior frontal gyrus (IFG). The three sub-parts (e.g., opercular, orbital, and triangular) of IFG on each side of the brain were used as indicators of IFG, as indicated in the Neuromorphometric Atlas. The constructs for G×E interaction did not have indicators since they were defined as products of a pair of genes and PTE.

We fitted our hypothesized IG-GSCA model to data via GSCA Pro (69). We adopted ridge-regularization for the structural part of our model and considered values between 0 and 1,000 as candidate values of the ridge tuning parameter λ. We selected λ = 62 based on 10 × 5 fold cross-validation. We employed 5,000 bootstrap samples to estimate standard errors and 95% confidence intervals for the parameter estimates.

We calculated the values of three goodness-of-fit measures: FIT, the goodness-of-fit index (GFI), and the standardized root mean square residual (SRMR). FIT is computed based on the residuals for components and indicators, whereas GFI and SRMR are calculated based on the discrepancy between the sample and reproduced covariance matrices of indicators. A higher value of FIT or GFI indicates a better fit, whereas a lower value of SRMR suggests a better fit. IG-GSCA provided the FIT value of 0.640, indicating that the model explained 64.0% of the total variance of all the components and indicators in the model. It also yielded a GFI value of 0.989, which is far greater than the recommended cutoff criterion (0.93) (70). Lastly, it produced the SRMR value of 0.04, which is much smaller than the recommended cutoff criterion (0.08) (70).

Table 4 presents standardized loading estimates, which can be interpreted as the correlations between constructs and their indicators. All the loading estimates were statistically significant, and their effect sizes were large (i.e., at least larger than 0.7), indicating that components were constructed to well explain the variations of their indicators. The average R² for all the indicators was 0.87.

Among 187 significant path coefficients, we paid most attention to reporting and interpreting the statistically significant path coefficient estimates from ROIs to the severity of PTSD symptoms and the ones from genes or PTEs to ROIs, since this study aimed to identify the mediating pathways from genes or PTEs to ROIs to the severity of PTSD symptoms. Figure 2 presents the mediating model, and Table 5 provides the standardized estimates for those path coefficients. The estimates of the entire path coefficients are presented in Supplementary Material .

For the relationships between brain ROIs and the severity of PTSD, the severity of PTSD turned out to be negatively associated with three brain ROIs—central operculum (b = -0.11, SE = 0.04, 95% CI = [-0.18, -0.00]), occipital fusiform gyrus (b = -0.13, SE = 0.05, 95% CI = [-0.21, -0.03]), and anterior cingulate gyrus (b = -0.10, SE = 0.05, 95% CI = [-0.19, -0.00]) ¹ , suggesting that people with smaller volume in those ROIs had more susceptible to the severity of PTSD symptoms. The PTEs also had a negative association with the severity of PTSD symptoms (b = 0.35, SE = 0.04, 95% CI = [0.26, 0.41]) even after the effects of ROIs and covariates were manipulated. It would imply that the traumatic experiences had a direct effect on the severity of PTSD symptoms even after adjusting the effects of covariates and decreasing the volume of ROIs. Figure 3 shows three brain areas significantly mediated between genes and PTSD symptom severity.

In the pathways from genes to the three brain ROIs above, HTR3A had a negative influence on the anterior cingulate gyrus (b = -0.11, SE = 0.04, 95% CI = [-0.18, -0.03]), indicating that people with a more minor allele in HTR3A were likely to have a smaller anterior cingulate gyrus. However, the indirect effect of HTR3A on the severity of PTSD symptoms through the anterior cingulate gyrus was not statistically significant (b = 0.01, SE = 0.01, 95% CI = [-0.00, 0.02]). On the other hand, the effect of NR3C1 on the central operculum was positive (b = 0.09, SE = 0.03, 95% CI = [0.02, 0.15]), implying that people with a mutant allele in NR3C1 had a larger central operculum. Moreover, NR3C1 had an indirect effect on the severity of PTSD symptoms mediated through the central operculum (b = -0.01, SE = 0.01, 95% CI = [-0.02, -0.00]), which means that the mutation of the NR3C1 led people to be less vulnerable to PTSD, though enlarging their central operculum. None of the interaction effects of genes and PTE on the three ROIs were statistically significant. The ratios of the amount of explained variance for each ROI or the severity of PTSD to their total variances are in Table 6 (average R² = 0.25).