Behavioral and brain signatures of substance use vulnerability in childhood

Abstract

The prevalence of risky behavior such as substance use increases during adolescence; however, the neurobiological precursors to adolescent substance use remain unclear. Predictive modeling may complement previous work observing associations with known risk factors or substance use outcomes by developing generalizable models that predict early susceptibility. The aims of the current study were to identify and characterize behavioral and brain models of vulnerability to future substance use. Principal components analysis (PCA) of behavioral risk factors were used together with connectome-based predictive modeling (CPM) during rest and task-based functional imaging to generate predictive models in a large cohort of nine- and ten-year-olds enrolled in the Adolescent Brain & Cognitive Development (ABCD) study (NDA release 2.0.1). Dimensionality reduction (n = 9,437) of behavioral measures associated with substance use identified two latent dimensions that explained the largest amount of variance: risk-seeking (PC1; e.g., curiosity to try substances) and familial factors (PC2; e.g., family history of substance use disorder). Using cross-validated regularized regression in a subset of data (Year 1 Fast Track data; n>1,500), functional connectivity during rest and task conditions (resting-state; monetary incentive delay task; stop signal task; emotional n-back task) significantly predicted individual differences in risk-seeking (PC1) in held-out participants (partial correlations between predicted and observed scores controlling for motion and number of frames [r_p]: 0.07-0.21). By contrast, functional connectivity was a weak predictor of familial risk factors associated with substance use (PC2) (r_p: 0.03-0.06). These results demonstrate a novel approach to understanding substance use vulnerability, which-together with mechanistic perspectives-may inform strategies aimed at early identification of risk for addiction.

Keywords: ABCD; Connectome-based predictive modeling; Substance use; Vulnerability.

Fig. 1

Analysis schematic. A) Behavioral measures associated with substance use were normalized and reduced using principal components analysis. B) Functional connectivity matrices were generated for each task and rest using a 268-node brain atlas (Shen et al., 2013). C) A 10-fold cross-validation procedure was used to train and test a ridge regression model to predict individual differences in PCA scores.

Fig. 2

Substance use measures. A) Correlation matrix of transformed behavioral measures related to substances. B) MDS plot visualizing distances among substance use measures. C) Scree plot depicting variance explained by individual components. D) PCA loadings for each behavioral measure (left). PCA loadings that demonstrated consistency across 10,000 bootstrapped samples (right).

Fig. 3

Predictive model performance. A) Cross validated predictions of substance-related risk components from functional connectivity. Model performance is defined as the partial correlation between predicted and observed values accounting for covariates. Error bars represent 95% confidence intervals. B) Pairwise distances (1 – Spearman correlation) between coefficients for all task- and rest-based models for PC1.

Fig. 4

Anatomical specificity of predictive models. Most informative connections (top 0.1%) in connectome-based models predicting PC1 (risk-seeking associated with substance use). Nodes are arranged in correspondence with networks functionally defined in an independent dataset. Edges represent model coefficient values. (DMN = Default mode network; MF = Medial frontal; CO = Cingulo-opercular network; FP = Frontoparietal network; CBL = Cerebellum).