Participants
Data analyzed in this study were collected as part of IMAGEN, a longitudinal genetic × neuroimaging cohort study of adolescents recruited from eight study centers in England, Ireland, France and Germany50. The IMAGEN study was approved by local research ethics committees at each study site, and written informed consent was obtained from participants and their parents or guardians. A detailed description of the study protocol and data acquisition can be found in ref. 50. Information on specific ethnic categories was not collected but the study, aimed at identifying the genetic and neurobiological basis of individual variability in behavior, was designed to include predominantly participants of European (white) ancestry, based on their self-reports. To further account for population stratification, statistical approaches were applied to identify and exclude genetic ancestries other than European when analyzing the genetic data. Specifically, the SDQ data used in this study were acquired at ages 14, 16, 19 and 23 years; neuroimaging data (N = 949) were acquired at ages 14 and 23 years and the TFEQ data were obtained at age 23 (N = 996).
Neuropsychological assessments
Eating behaviors
The short version (18 items) of the TFEQ was used to assess eating behaviors. The TFEQ contains 3 subscales: CR, which measures the tendency to restrict one’s food intake constantly and consciously instead of using physiological cues, hunger and satiety as regulators of food intake (6 items); EE, which reflects the tendency to eat in response to negative emotions (3 items); and UE, which characterizes the tendency to overeat with the feeling of being out of control (9 items). It has good structural validity and has been used and validated in different European populations7,51 and was found to distinguish different eating patterns in the general population52.
ED symptoms
Dieting, binge eating and purging symptoms were assessed using the self-reports from the ED section (section P) of the Development and Well-being Assessment21,53. Dieting symptoms were evaluated based on responses to questions P18a, P18b and P18c, which asked about eating less at meals, skipping meals and fasting, respectively. Binge-eating symptoms were assessed using the question P15, which inquired about eating a large amount of food and losing control overeating. Purging symptoms were measured using the questions P1c, P18f and P18g, which asked about self-induced vomiting or taking pills or medicines to lose weight.
Emotional and behavioral problems
The SDQ was used to assess emotional and behavioral problems in adolescents. It has five hypothesized subscales, including emotional symptoms, conduct problems, hyperactivity/inattention, peer relationship problems and prosocial behaviors54. In low-risk and general population samples, the emotional and peer subscales can be combined into an ‘internalizing’ subscale (10 items) and the behavioral and hyperactivity subscales into an ‘externalizing’ subscale (10 items), respectively55. We used self-reported scores at ages 14, 16, 19 and 23 years for IPs and EPs in further analyses.
Structural magnetic resonance imaging acquisition and processing
Magnetic resonance imaging (MRI) scans were acquired with 3T MRI scanners from different manufacturers (Siemens, Philips, General Electrics and Bruker) from 8 IMAGEN recruitment sites. The high-resolution anatomical MRI images acquired included a 3-dimensional T1-weighted magnetization prepared gradient echo sequence based on the Alzheimer’s Disease Neuroimaging Protocol ( T2-weighted fast-spin echo and fluid-attenuated inversion recovery scans for visual assessment.
All raw images were visually inspected to exclude images with movement artifacts, brace artifacts or field inhomogeneities before preprocessing. The preprocessing procedures were then conducted using the Computational Anatomy Toolbox (CAT 12.8 (r1907); in SPM 12 (Wellcome Department of Cognitive Neurology). We used the ‘segment longitudinal data’ procedure with default settings. Intrasubject coregistration was performed on the baseline (at age 14) and follow-up (at age 23) images. The coregistered images were then realigned across participants and bias corrected with reference to the mean images computed from each subject’s baseline and follow-up images. Next, the baseline and follow-up images and their mean images were segmented into gray matter, white matter and cerebrospinal fluid based on the default tissue classification map. Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) normalization was subsequently performed on the segmented mean images using the default DARTEL template. The derived spatial normalization parameters were then applied to transform the segmented subject baseline and follow-up gray matter images into the standard Montreal Neurological Institute space. All normalized gray matter images were finally smoothed with an isotropic Gaussian kernel of 6 mm full width at half maximum. The quality measures created during preprocessing for each participant at each time point were examined, and images with sufficient quality (corresponding to grade D or above) were included in further analyses. Changes in GMV were analyzed using whole-brain VBM. Measures of CT and square root-transformed SD, which were then resampled to 12 mm in line with the recommendation for surface measures, were also derived. Longitudinal changes in mean CT and SD were extracted for different ROIs using the Desikan–Killiany atlas (N of ROIs = 68).
BMI PGSs
A total of 2,087 participants were genotyped with the Illumina Human610-Quad BeadChip and Illumina Human660-Quad BeadChip during the baseline assessments. Stringent quality control (QC) procedures were performed before imputation (Supplementary Information). In brief, multidimensional scaling analysis and principal component analysis were conducted to identify genetic ancestry. Participants who were outliers from the European superpopulation were excluded (Supplementary Figs. 4 and 5) due to the limited portability across ancestries for PGSs. Consequently, 1,899 participants (49.66% male participants) who passed genotyping QC and were identified to be of European ancestry were selected for generating the BMI PGSs. IMAGEN genotype data were integrated into the European ethnicity 1KGP (phase 3 release v.5) reference panel56 for imputation. Summary statistics of genome-wide association study data of BMI from ~681,275 individuals of European ancestry57 were used to calculate BMI PGSs. This was achieved using PRS-CS58, which uses high-dimensional Bayesian regression and a continuous shrinkage before single-nucleotide polymorphism effect sizes. The global shrinkage parameter was set to 0.01 as is recommended for highly polygenic traits. A total of 905,362 single-nucleotide polymorphisms were used to predict BMI PGSs. Participants with available TFEQ scores were included in the subsequent analyses (for RE, N = 255; for E/UE, N = 194; for HE, N = 347). The BMI PGS was residualized for the first 10 principal components and batch effects before being Z-scored for subsequent analyses.
Statistical analyses
Identification of groups with distinct eating behaviors by K-means clustering
K-means clustering using the TFEQ subscale scores (that is, CR, EE and UE) at age 23 was performed to identify clusters showing different eating behaviors. All continuous variables were transformed into Z-scores. We used the NbClust package to identify the optimal cluster number and validity of the cluster solution, and the fpc package to examine the clustering stability with the Jaccard coefficient and a bootstrap technique (N = 1000) in R.
Group differences in trajectories of ED symptoms across adolescence
Linear mixed models were used to examine group differences in the trajectories of dieting, binge eating and purging from ages 14 to 23. Age (that is, 14, 16, 19 and 23) was treated as a categorical variable. The models included age, group and age-by-group interactions as fixed effects and adjusted for sex. Random intercepts for participants nested within recruitment sites accounted for the dependence of repeated measures. Group-by-age interactions were investigated using HEs and age 14 as reference. A Bonferroni correction accounting for 18 tests (3 ED symptoms × 2 group comparisons × 3 age comparisons) was applied (that is, PBonferroni = 2.78 × 10−3).
Group differences in trajectories of IPs and EPs
The LGCMs were conducted in these analyses using the lavaan package in R.
Univariate LGCM analyses
Latent factors of intercept and slope were estimated for repeated measures (at ages 14, 16, 19 and 23 years) of IP and EP scores separately. Sex, groups and recruitment sites were considered time-invariant covariates. For these analyses, we included only those participants who had TFEQ scores at age 23 and who had at least 1 measure of IP or EP at ages 14, 16, 19 and 23. For both IP and EP, we attempted to fit a quadratic term; however, this specification resulted in a non-positive definite covariance matrix, driven by a correlation greater than or equal to one between the linear and quadratic terms. Hence, we decided not to include a quadratic term as the information contained within it was not adding any extra information over the linear term. The full information maximum likelihood estimator was used to account for data missing at random. We investigated group differences in intercepts and slopes of IP and EP trajectories, taking HEs as a reference. A Bonferroni-corrected P-value threshold of 0.05/(2 behaviors × 2 measures × 2 groups) = 6.25 × 10−3 was considered statistically significant.
Multivariate LGCM analyses within each group were also run to estimate models for IP and EP trajectories simultaneously and to investigate covariances between latent factors (that is, IP intercept, IP slope, EP intercept and EP slope). Sex and recruitment sites were included as covariates.
Longitudinal MRI analyses for group differences in brain maturation
Participants were excluded from the analysis if they had missing MRI data or failed to meet QC criteria (N = 47; see ‘Structural magnetic resonance imaging acquisition and processing’ for image preprocessing and QC). Consequently, a total of 949 participants (306 REs, 236 E/UEs and 407 HEs) were included in the whole-brain VBM analysis and linear mixed models for CT and SD.
VBM analysis
Longitudinal whole-brain VBM analyses were performed using the CAT 12.8 (r1932) toolbox. To identify brain regions reflecting significant changes in GMVs between ages 14 and 23 among the groups identified above, we performed a 2 × 2 mixed analysis of variance on the smoothed images using the ‘flexible factorial’ model. The two factors were age (age 14 or age 23; within subject) and group (that is, comparison of each of 2 groups, namely REs versus HEs, E/UEs versus HEs or REs versus E/UEs; between subject). Intracranial volumes (TIVs) were estimated by CAT 12.8 as the sum of the gray matter, white matter and cerebrospinal fluid volume. Analyses were controlled for the effects of participants’ sex, the scanning site and TIV at each time point (at ages 14 or 23). An absolute threshold masking of 0.1 was applied. The gray matter morphological differences showing significant age-by-group interactions were reported after a cluster-level, family-wise error correction with P < 0.05 and a cluster-forming threshold of P < 0.001 without correction.
Linear mixed models
For group differences in changes in CT and SD, we performed ROI-based linear mixed models, investigating interactions between age and groups. Models included age, groups and their interactions as fixed effects, with the participant nested within recruitment sites as a random effect and adjusted for sex. For both measures, the Bonferroni correction was applied to adjust for multiple testing (P = 0.05/68 ROIs × 3 group comparisons = 2.45 × 10−4).
Mediation analyses
Simple mediation models were performed using the PROCESS v.4.0 macro for R to test whether the between-group differences in brain changes mediated the relationships between differences in IP or EP trajectories and eating behaviors. We refer to this model as the psychopathology–brain maturation–eating behaviors model. Brain clusters that significantly differentiated REs from HEs, E/UEs from HEs or REs from E/UEs were considered ROIs. For group comparisons involving several brain clusters, these clusters were combined into a single ROI for each structural measure (GMV, CT or SD). For comparisons between REs and HEs, 1 mediation model was tested, relating GMV differences in the left cerebellum to differences in IP slope; therefore, a P-value threshold of 0.05 was considered significant. For comparisons between E/UEs and HEs, nine mediation models were tested because these two groups differed behaviorally in IP intercept, IP slope, and EP intercept, and in their changes of GMVs, CTs and SDs. The Bonferroni-corrected significance threshold of 0.05/(3 trajectory measures × 3 structural brain measures) = 5.56 × 10−3 was applied.
Subsequent analyses investigated the potential contributions of the BMI PGS on the brain mediation models identified above, referred to as the genetics–brain maturation–eating behaviors models. The same brain ROIs were considered as mediators in these models. For models comparing REs with HEs, the significance was set at P = 0.05/(1 structural measure) = 0.05. For models comparing E/UEs to HEs, the significance was set at P = 0.05/3 structural measures = 1.67 × 10−2.
Multivariate mediation models were conducted using AMOS 29 to explore the unique contributions of brain ROIs, psychopathology (IP and EP trajectories) and BMI PGS to mediation models identified in simple mediation analyses, referred to as the ‘genetics–psychopathology–brain maturation–eating behaviors’ model. Continuous variables were transformed into Z-scores for these analyses. CIs for the mediation effect were estimated from 5,000 bootstrap samples.
Covariates
Covariates for all analyses included sex and recruitment sites. For analyses involving GMV, CT and SD, TIV at the corresponding age was additionally included as a covariate. As there were no significant group differences in age at each data collection, age was considered a categorial variable in the linear mixed models and as time points in repeated measures in the LGCM analysis. For the longitudinal MRI analysis (VBM analysis and linear mixed models), participants nested within recruitment were modeled as a random effect and sex was considered a fixed effect in the model.
Other covariates
To examine the robustness of findings from our primary analyses, sensitivity analyses were conducted by including pubertal status, IQ, EA, and age- and sex-adjusted BMI as additional covariates. Pubertal status was assessed using the Puberty Development Scale59, an eight-item self-report measure of physical development based on Tanner stages, separately for male and female participants. IQ was calculated as the average of the Perceptual Reasoning Index (PRI) and Verbal Comprehension Index (VCI) scores based on age norms using the Wechsler Intelligence Scale for Children, Fourth Edition (WISC-IV; Pearson Clinical Assessment UK). We administered the block design, matrix reasoning, similarities and vocabulary subtests. Raw scores from each subtest were converted into scaled scores based on age norms. For both the PRI and VCI, we calculated prorated sums of scaled scores and then converted these sums into index scores according to the WISC-IV manual. EA was assessed by self-report of the ‘average grade at the end of the last term completed’. The age- and sex-adjusted BMI Z-score at age 14 was calculated using the jBmi R package based on the Centers for Disease Control and Prevention recommendations.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
link
More Stories
Veterinary DNA Testing Market Size
KFSHRC Achieves 39% Rise in
Genomic screening reveals hidden health risks of patients