Adolescent binge drinking disrupts normal trajectories of brain functional organization and personality maturation

Adolescent binge drinking has been associated with higher risks for the development of many health problems throughout the lifespan. Adolescents undergo multiple changes that involve the co-development processes of brain, personality and behavior; therefore, certain behavior, such as alcohol consumption, can have disruptive effects on both brain development and personality maturation. However, these effects remain unclear due to the scarcity of longitudinal studies. In the current study, we used multivariate approaches to explore discriminative features in brain functional architecture, personality traits, and genetic variants in 19-year-old individuals (n = 212). Taking advantage of a longitudinal design, we selected features that were more drastically altered in drinkers with an earlier onset of binge drinking. With the selected features, we trained a hierarchical model of support vector machines using a training sample (n = 139). Using an independent sample (n = 73), we tested the model and achieved a classification accuracy of 71.2%. We demonstrated longitudinally that after the onset of binge drinking the developmental trajectory of improvement in impulsivity slowed down. This study identified the disrupting effects of adolescent binge drinking on the developmental trajectories of both brain and personality.


Introduction
Adolescence is characterized by significant developments (Caspi et al., 2005;Giedd and Rapoport, 2010), including the functional segregation and integration of different brain networks through a process of modular evolution (Gu et al., 2015) and developmental improvement of some personality traits, such as impulsivity (Harden and Tucker-Drob, 2011) and agreeableness (Klimstra et al., 2009). These developmental processes are influenced by genetic and environmental factors, such as alcohol misuse during adolescence (Brown et al., 2008;Squeglia and Gray, 2016;Squeglia et al., 2014). The incidence of alcohol misuse peaks between the ages of 18-25 (Blanco et al., 2008;Chen and Kandel, 1995) and appears to decline after 26 years of age (Substance Abuse and Mental Health Services Administration, 2018), which aligns to the brain development and personality maturation. Alcohol misuse can disrupt standard developmental trajectories and these effects can persist into adulthood and subsequently increase the risk of alcohol dependence, violence, drunk driving, and other adverse outcomes later in life (Miller et al., 2007). Therefore, delineating the consequences of adolescent alcohol drinking on brain and personality developmental trajectories may provide new insights into alcohol-affected neural and behavioral changes that are responsible for long-term adverse outcomes (McCambridge et al., 2011;Spear, 2018).
Significant processes in functional segregation, such as weakening connectivity between brain systems; and integration, such as strengthening connectivity within brain systems, have highlighted that brain functional connectivity network organization is particularly vulnerable to adolescent alcohol drinking (Gu et al., 2015). However, few longitudinal investigations have measured the functional consequences of alcohol drinking in the adolescent brain. In a previous study of the IMAGEN consortium (Whelan et al., 2014), the authors focused on establishing a profile at an early age to predict the onset of binge drinking in a 2-year follow-up; however, brain functional connectivity was not investigated. Based on previous epidemiological results (Dawson et al., 2008), we hypothesized that drinking-altered functional connectivity in the brain should satisfy the following three conditions: 1) classify binge drinkers from non-binge drinkers and be independent of other features, such as personality, cognition, and genetics; 2) earlier and heavier binge drinking induces more discriminative alterations; we named this the hypothesis of trend; and 3) not associated with other substance use (e.g. smoking, cannabis use) during adolescence.
In this study, we aimed to clarify the consequences of adolescent binge drinking by exploiting the longitudinal design of the IMAGEN study (G Schumann et al., 2010). This study collected neuroimaging, personality traits, drinking behavior, and genetic variant data from a large population. The IMAGEN study recruited healthy individuals; therefore, alcohol misuse was defined as episodes of lifetime drunkenness (i.e. binge drinking) (Whelan et al., 2014). Different from previous studies, the longitudinal design enabled us to stratify the drinkers according to their onset of binge drinking. By identifying groups of extreme drinkers with the longest and the shortest histories of binge drinking in our sample, we were able to select the features that were associated with binge drinking and that also satisfied the hypothesis of trend. With such a hypothesis-guided feature selection, we imposed more constraints on the developmental course of any selected feature. Other substance use (e.g., cannabis) may have a different onset age compared with binge drinking, and thereby require its associated feature to have a different trend. In other words, our procedure could increase the specificity of the feature selection associated with binge drinking. As the selected features satisfy the hypothesis of trend, we expected to see an intermediate disruptive effect of these features in a medium-term history of binge drinking group; therefore, we confirmed our findings using an independent sample (i.e. the medium-term drinkers). Finally, we applied a longitudinal analysis to assess if the developmental trajectories for the selected features were significantly affected by the age of binge drinking onset.

Participants
The IMAGEN study recruited a cohort of healthy adolescents at the age of 14. Informed consent was obtained from all subjects and their parents/guardians. The genetic data were collected at baseline, the neuroimaging data were collected at ages 14 and 19, and environmental and behavioral data were collected at ages 14, 16 and 19. The details of the study design and data quality are described in a previous report (G Schumann et al., 2010). In total, 212 participants were included in the current study; a flowchart of the inclusion criterion is presented in eFig. 1.

Resting-state functional imaging data
A proportion of the participants in the IMAGEN study underwent resting-state fMRI (rsfMRI). In the IMAGEN study, much less rsfMRI data (n = 389) were collected at the base line (i.e. age 14). As rsfMRI got more and more attention, more data (n = 1069) were collected at the follow-up stage (i.e. age 19). All fMRI data were preprocessed by the Data Processing Assistant for Resting State fMRI (rfmri.org/DPARSF). For each individual, we calculated the resting-state functional connectivity (rsFC) between each pair of brain regions [19,900 links for 200 atlas-defined brain regions, Craddock 2011 Atlas (Craddock et al., 2012) eTable 3].

Genome-wide genotype data
Details of genetic data have been reported previously (Desrivieres et al., 2015); we have included detailed descriptions in the eMethods. Briefly, we converted all single nucleotide polymorphisms (SNP) into binary variables; no mutations were assigned the number "zero" and mutations were assigned the number "one".

Substance use
We used the European School Survey Project on Alcohol and Drugs [ESPAD (G. Schumann et al., 2010;Whelan et al., 2014)] for ages 14, 16, and 19 years to assess the alcohol consumption, smoking, and cannabis use. The ESPAD category scores are as follows: 0(0), 1(1-2), 2(3-5), 3(6-9), 4(10-19), 5(20-39), 6(≥40). The primary questions of interest concerned lifetime alcohol use [for example, On how many occasions (if any) in your lifetime have you had any alcoholic beverage?]; lifetime drunken episodes [for example, On how many occasions (if any) in your lifetime have you been drunk from drinking alcoholic beverages?]; lifetime smoking (for example, On how many occasions during your lifetime have you smoked cigarettes?); and lifetime cannabis use (for example, On how many occasions in your whole lifetime have you used marijuana or hashish?). Variables associated with the occasions of alcohol use and the episodes of binge drinking were measured at three time points (ages 14, 16, and 19), and the data were organized according to the duration and amount of binge drinking during adolescence (Fig. 1, eFig. 1). More details are provided in the eMethods.

Cognition and personality
We used the Cambridge Neuropsychological Test Automated Battery [Cantab (Sahakian et al., 1988)], Monetary-Choice Questionnaire (KIRBY rate) (Kirby et al., 1999), personality characteristics tests, including the Revised NEO Personality Inventory [NEO-PI-R (Costa and Mccrae, 1992)], and the Substance Use Risk Profile Scale [SURPS (Woicik et al., 2009)] to assess cognition and personality factors. Further details are provided in the eMethods.

Univariate approach
We used two-sample Student's t-tests, controlling for covariates (sex and site) for univariate analyses, and employed a false discovery rate (FDR) correction for multiple comparisons.

Multivariate approach with hypothesis-guided feature selection
To identify discriminative features in high dimensional space and control for the over-fitting problem, we combined several dimension reduction approaches with the multivariate pattern analysis approach [Support-Vector Machine (SVM), (Suykens and Vandewalle, 1999)]. The main steps are described below and additional details are provided in the eMethods.
2.3.2.1. Candidate feature selection and dimension reduction. To reduce dimensionality, we first used the lasso-regularized logistic regression to select the most discriminative rsFCs between the binge drinkers and non-drinking controls. Chi-squared tests were used to select the SNPs with significantly different frequencies in subjects with a history of binge drinking compared to those without a binge drinking history. Next, instead of using these features as inputs to the SVM, we further reduced the dimensionality by calculating four summary scores, including increased/decreased rsFC (iFC/dFC) or risk/protective SNP (rSNP/pSNP) scores ( Fig. 2A). To prevent model over-fitting in the training sample, we applied the leave-one-out (LOO) cross-validation procedure to the SVM, which was built by the two steps described above. Only those features (rsFC and SNP) that were repeatedly selected in > 90% of iterations during the LOO were considered robustly discriminating features (namely, the candidate features) between the binge drinkers and the control subjects. We trained one SVM for the long-term drinkers (SVM-long classifier) and the other SVM for the short-term drinkers (SVM-short classifier). To assess the contribution of each domain (SNP, rsFC, and covariates) in classification, we compared the model performance before and after the removal of each domain from the model inputs. The candidate features in the domains significantly contributed to SVM-long and SVMshort were summarized into four new scores (iFC, dFC, rSNP, and pSNP) for next stage.
2.3.2.2. Hypothesis-guided feature selection. According to our hypothesis of trend for the consequences of adolescent binge drinking, we focused on the personality scores and the four new scores established by candidate features. We employed the Jonckheere-Terpstra trend test to identify the features that were consistent with the hypothesis of trend (i.e. the discriminative features having the largest, intermediate and smallest deviations from the control group in the binge drinkers with the longest, medium, and shortest history of binge drinking, respectively; Fig. 3A, B, and C).

Hierarchical classifier.
Finally, we built a hierarchical classifier, which consisted of two layers: the in-layer and the out-layer (Fig. 3D). To further reduce dimensionality, we constructed the in-layer with three SVMs taking the inputs from covariates (SVM1), rsFC and SNP (SVM2), and personality (SVM3). Each of these SVMs was trained using the LOO procedure on the training data, including the Control I group, Fig. 1. Stratified drinking and control groups. Drinking groups stratified by the onset age of binge drinking. "Binge (ESPAD)" corresponds to the score from the question "on how many occasions in your whole lifetime have you been drunk from drinking alcoholic beverages" while "Occasion (ESPAD)" corresponds to the score from the question "on how many occasions in your whole lifetime have you had any alcoholic beverage to drink." The long-and short-term drinkers and Control I group were used as the training sample, while the medium-term and Control II group were used as the test sample. the long-term binge drinking group, and the short-term binge drinking group. In the out-layer, one SVM (SVM4) received only 3 input values from the output values of SVM1-3 (a greater output value means more likely to be a binge drinker). With this hierarchical classifier, we significantly reduced the model complexity by taking the inputs from three kinds of features, so as to achieve a better generalizability. The model was then tested using an independent test sample (i.e. the medium-term drinkers and the Control II group).

Longitudinal analysis
We could not conduct a longitudinal analysis on the rsfMRI data because we only had 32 subjects (18 binge drinkers and 14 controls) with rsfMRI data available at the age of 14 after participant selection (eFig. 1). However, personalities were assessed at three time points (ages 14, 16, and 19); therefore, we estimated changes in personality between 14 and 16 years (difference  ) and between 16 and 19 years (difference  ). To compare the developmental trajectory of personality between the two different time periods (i.e. from 14 to 16 and from 16 to 19), difference 16-14 and difference  in binge drinkers must be established before and after binge drinking onset, respectively, thus only the short-term drinkers satisfied this condition. These differences were compared before and after the onset of binge drinking (significance was given by 1000 permutations of the paired median test). By conducting an Analysis of Variance (ANOVA) with repeated measures using SPSS (Released 2011. IBM SPSS Statistics for Windows, Version 20.0. Armonk, NY: IBM Corp), we were able to compare the longitudinal evolutions of both the impulsivity and the sensationseeking scores between the short-term drinkers and the control subjects. The same approach was also applied to difference 16-14 and difference 19-16 . The covariates (sex and data collection sites) were regressing out from the measurements before the group comparisons.

Stratification of binge drinkers by the onset age of binge drinking
All participants were stratified into 5 groups, including long-term (n = 54), medium-term (n = 52), and short-term (n = 41) drinkers; Control I (n = 44) and Control II (n = 21). These were further divided into the training and test samples (Fig. 1, eFig. 1 and Table 1). To summarize the analyses in this study, a diagram of the study design depicting the relationship between different analyses is presented in eFig. 2.
Among the established summary scores (Fig. 3), we found that iFC (z 212 = 8.54, p = 1.34 × 10 −17 ) and impulsivity (z 212 = 2.59, The black arrow indicates a significant monotonically decreasing trend confirmed by the Jonckheere-Terpstra trend test (****p < 10 −16 ; ***p < 10 −4 ; **p < 10 −2 ). Five bars (groups) from left (red) to right (blue) were long-term, medium-term, short-term drinkers, Control II and Control I, respectively. (D) Structure of the hierarchical model for adolescent binge drinking. (E) Comparison of the classification accuracies using the SVM4 on the subgroups of Control II (test sample). This was stratified by lifetime drinking occasions, where fewer drinking occasions achieved higher accuracy. (F) Comparison of classification accuracies using the SVM4 on the subgroups of medium-term drinkers. This was stratified according to lifetime binge drinking by age 19, where a higher number of episodes of binge drinking achieved a higher accuracy. (G) Contribution of iFC, pSNP, rSNP, and impulsivity scores. Contribution was characterized by the absolute value of the correlation between the scores and the model output. p = .0047) scores showed a trend that increased across controls, short-, medium-, and long-term binge drinkers (eFig. 5 and eTable 5). We also found higher rSNP scores (t 210 = 4.20, p = 3.88 × 10 −5 ) and lower pSNP scores (t 210 = −9.09, p = 7.40 × 10 −17 ) in binge drinkers than controls. Thus, the final classifier (SVM4) for binge drinkers was created from the iFC, impulsivity, and p/rSNP scores (Fig. 3D).

Validation of models using an independent sample
In the test sample, SVM4 achieved a classification accuracy of 71.2% for medium-term binge drinkers. The pSNP score contributed the most to this classifier ( Fig. 3G; r = 0.75, p = 6.69 × 10 −26 ), followed by the iFC score (r = 0.64, p = 2.88 × 10 −17 ). The rSNP and the impulsivity scores showed relatively small, but significant, contributions (rSNP: r = 0.36, p = 1.79 × 10 −5 ; impulsivity: r = 0.38, p = 6.94 × 10 −6 ). The covariates, including sex and site of data collection, contributed to the classifier significantly less than the rsFC and SNP (0.235-0.539, 95% confidence interval of contribution difference by 5000 bootstraps, eFig. 6). Without the hierarchical design (eTable 6) or hypothesis-guided feature selection (eFig. 7), the multivariate model tended to over-fit the training sample, which significantly reduced the classification accuracy of the test sample. As the onset time of binge drinking may be different from the onset time of other substance use, the hypothesis-guided feature selection may improve the specificity of the selected features to be associated with binge drinking instead of general substance use. Indeed, applying the established classifier to identify cannabis users and smokers in the test sample yielded accuracies of 49.3% and 54.8%, respectively, which is comparable to chance.

Onset of binge drinking slowed down the developmental trajectory of improvement in impulsivity
It was difficult to elucidate the cause and effect of binge drinking because the impulsivity score was higher in the binge drinkers (including the medium-term and the short-term groups) than controls (including Control I and Control II) at both ages (14: t 146 = 2.14, p = .034; 19: t 145 = 4.1, p = 6.93 × 10 −5 ). For each period (14-16, or 16-19), we computed the difference between two time points to represent the developmental improvements in impulsivity. In controls (including Control I and Control II), we observed that impulsivity improved at a steady speed from ages 14 to 16 and 16 to 19 (Fig. 4B). However, in short-term drinkers (n = 41), we found that the improvement in impulsivity between ages 16 and 19 was lower than between ages 14 and 16 (p = .003, Fig. 4A). This effect was not observed for sensation seeking (Fig. 4C-D). Using ANOVA with repeated measures, we identified an interaction effect between time (ages 14, 16 and 19) and group (short-term drinkers and controls) on impulsivity (F 97 = 3.745, p = .027) and the developmental change of impulsivity (i.e. difference  and difference 19-16 ) (F 97 = 5.516, p = .021). This interaction effect was not observed for sensation seeking.

Discussion
We conducted a systematical study to delineate the consequences of binge drinking in multiple domains, including the brain, personality, and cognition of adolescents, while considering multiple confounders, including genetics and other substance use. In the brain, we found that the frontal connectivity significantly contributed to the binge-drinking associated brain feature (11 out of 16 links selected for the iFC score). In the domain of personality, the developmental improvement of impulsivity was slowed down after the onset of binge drinking during adolescence. The current findings identified the disruptive effects of adolescent binge drinking on the developmental trajectories of both brain and personality.
The functional connectivity between the frontal cortex and precentral gyrus and occipital cortex are particularly vulnerable to adolescent binge drinking. The standard developmental trajectory of the functional architecture is to segregate the brain into different functional systems; for example, interactions among frontal and sensory-motor regions decrease with age (Gu et al., 2015;Stevens et al., 2009); and frontal areas undergo significant pruning processes during adolescence (Gogtay et al., 2004). It has been reported that the earlier age of the Fig. 4. Analysis of the relationship between personality scores and binge drinking. (A), (B) Boxplots of "impulsivity" scores in the controls (blue) and the short-term drinkers (green) at different ages. Controls included Control I and Control II groups, as they did not binge drink at all. The statistical significance displayed in (A) refers to the statistical significance of the difference between difference 16-14 of "impulsivity" scores and difference 19-16 of "impulsivity" scores in short-term drinkers. (C), (D) Boxplots of "sensation seeking" scores in the controls (blue) and short-term drinkers (green) at different ages. Boxplots showing the median, 25th, and 75th percentiles; whiskers show ± 2.7 standard deviation. (*p < 10 −1 , **p < 10 −2 , ***p < 10 −4 ).
first drink of alcohol predicted the poorer performances in both the psychomotor speed and the visual attention (Nguyen-Louie et al., 2017), and the earlier onset age of the weekly drinking was associated with higher frontoparietal context-dependent functional connectivity between the bilateral posterior cingulate and both cortical and subcortical areas implicated in the attentional processes in young adults (Nguyen-Louie et al., 2018). Our results indicated an accumulating effect of adolescent binge drinking on frontal connectivity because more episodes of binge drinking increased the accuracy of the classification between binge drinkers and non-binge controls. Therefore, if binge drinking is not totally avoidable in adolescents, efforts should be considered to control the number of incidences.
The maturation process of personality is characterized by a decline in negative emotionality (Blonigen et al., 2008;Donnellan et al., 2007); however, we found that the standard decline of impulsivity was slowed down by the onset of adolescent binge drinking. Compared with our observation of sensation seeking, it suggests that the adolescent binge drinking may be particularly disruptive to the developmental processes normally taking place at the time period that is later than the onset age of binge drinking. Interestingly, sensation seeking becomes relatively stable after mid-adolescence, while impulsivity shows an explicitly monotonic decline through adolescence (Harden and Tucker-Drob, 2011;Steinberg et al., 2008). Therefore, adolescent binge drinking had a greater impact on the maturation of impulsivity but not sensation seeking. Furthermore, we found higher impulsivity scores in the binge drinkers than in controls at baseline (before the onset of binge drinking), which is consistent with previous studies that used the impulsivity score as a predictor for later alcohol misuse Elkins et al., 2006;King and Chassin, 2004). Therefore, the current findings provide longitudinal evidence to suggest that there is a negative feedback loop between impulsivity and adolescent drinking behavior.
Increased functional connectivity and impulsivity scores both contributed to our classification model for binge drinkers independent of genetic features; therefore, these features are unlikely to be independent manifestations of common genetic predispositions for adolescent binge drinking. In fact, gene expression patterns in the brain can be altered by alcohol intake, especially in the frontal areas (Lewohl et al., 2000;Liu et al., 2006;Mayfield et al., 2002). This altered expression pattern may lead to dysfunctional brain connectivity; rsFC has been correlated with gene expression linked to ion channel activity and synaptic function (Richiardi et al., 2015). In addition, the brain serotonin system is responsible for impulsivity (Leyton et al., 2001;Sachs and Dodson, 2017) and this system can be disturbed by alcohol (Burnett et al., 2012;Sachs and Dodson, 2017;Shibasaki et al., 2010), which may account for the impact of binge drinking on the developmental trajectory of impulsivity in this study.
The advantages of this study include the longitudinal design, independent test sample, and disassociation with cannabis use. However, several limitations must be considered when interpreting our results. First, we used binge drinking as a behavioral indicator for alcohol-induced neurotoxicity, which is not necessarily comparable to clinically diagnosed alcohol dependency. We measured drinking behavior at ages 14, 16, and 19 only; therefore, more detailed measurements outlining the frequency and quantity of drinking would assist with the identification of heavy drinkers from occasional drinkers. Second, a limited number of individuals in the IMAGEN study participated in the rsfMRI experiments at age 14, and only 24 were defined as binge drinkers across the three binge drinking groups. Future longitudinal studies employing rsfMRI at baseline will enable estimations of developmental trajectories of rsFC before and after the onset of binge drinking and thereby, provide more insights into the disruptive effects of adolescent binge drinking on the developmental trajectory of brain functional architecture.

Conclusions
We found new evidence for disrupted brain functional organization in adolescents who participate in binge drinking behaviors and highlighted a negative feedback loop that interacted with impulsivity following binge drinking during early adolescence. Alcohol is the most abused substance in adolescents; therefore, its effect on the brain and personality must be considered in remedy programs to prevent further development of alcohol-related adverse outcomes. The identified disruptive effects of adolescent binge drinking provide potential targets for such interventions in adolescents with a history of binge drinking.

Data availability
IMAGEN data are available by application to consortium coordinator Dr. G. Schumann (http://imagen-europe.com) after evaluation according to an established procedure.

Code availability
Matlab codes for main algorithms used in this work are provided at the following website https://github.com/qluo2018/ RetraceHistoryOfBingeDrinking.

Funding information
Dr