Figure 1: Experimental Workflow.

(A) Population: We included 5260 children (9-10 years old) from the Adolescent Brain Cognitive Development (ABCD) dataset, including 2689 males (51%) and 2571 females (49%). (B) Behavioral Data: We included 17 behavioral scores from the Child Behavior Checklist which includes syndrome scales and DSM-5 oriented scales. Syndrome scales included measures of composite and individual internalizing behaviors (shown in blue), composite and individual externalizing behaviors (shown in orange), other problems (shown in green), and a summary score of total problems (red). DSM-5 Oriented Scales included scores relating to affective, anxiety, somatic, oppositional defiant, conduct, and attention deficit/hyperactivity (ADHD) disorders. (C) Parcellation: We used the Schaefer cortical parcellation of 400 regions, and each region was assigned to one of 17 large-scale cortical networks. Image reproduced under a CC BY 4.0 license: https://doi.org/10.6084/m9.figshare.10062482.v1. We also included 19 subcortical regions in our analyses, which were assigned to a subcortical network. Image reproduced under a CC BY 4.0 license: https://doi.org/10.6084/m9.figshare.10063016.v1. (D) Neuroimaging Data: For each subject, we extracted their functional MRI time series data for the 400 cortical parcels and 19 subcortical parcels. Pairwise correlation was computed for all pairs of time series to obtain the estimated functional connectivity. (E) Predictive Models: Linear ridge regression models were trained to predict individual behavioral scores based on the upper triangular functional connectivity matrix in a sex-specific manner. Data were split into training and test sets. For each training set, a separate model was optimized and trained to predict each behavior. Once optimized and trained, models were evaluated for accuracy and generalizability. Prediction accuracy is defined as the correlation between the true and predicted behavioral scores in the test set for each split. We computed average accuracy by taking the mean across the distinct train-test splits. Once models were trained and tested within sexes and behaviors, we evaluated model generalizability across both sexes and all 17 behavioral scores. Model generalizability is defined as the accuracy obtained when a given model is evaluated on a population (i.e., sex) and/or behavioral score that is unique from the population/behavioral score that the model was trained on. This is distinct from model accuracy which is defined as the prediction accuracy obtained when evaluating the model on the same populations (i.e., sex) and behavioral score (using a hold-out test set) that it was trained on.