CROSS-COHORT EVALUATION OF GENETIC CONTRIBUTIONS TO CHILDHOOD PSYCHOPATHOLOGY USING MACHINE LEARNING

Background At the individual level, genetic susceptibility to a psychiatric disorder can be conceptualised as comprising disorder-specific genetic risk and shared genetic risk with other disorders and traits. Knowledge on whether this shared genetic risk enhances the prediction of psychiatric outcomes is limited. Including multiple polygenic risk scores (PRSs) enhances our understanding of the genetic architecture underlying adolescent psychiatric disorders and improves prediction. Methods This study aimed to investigate the combined and distinct contributions of genetic predisposition using a PRSs approach to predict depression (DEP) and ADHD in adolescents at age 14 using data from two population-based longitudinal studies - UK Millennium Cohort Study (UK-MCS) and the Adolescent Brain Cognitive Development (ABCD). It proposes and internally validates across seven machine learning models, on UK-MCS, with early childhood predictors at age 11, in addition to polygenic scores. It externally validates these models on the independent ABCD cohort. It aims to predict DEP and ADHD caseness at age 14 and compare four predictor subsets: models including only environmental predictors, models including single disorder PRSs, models including multi-PRSs that include scores from 32 genetically correlated traits selected based on published genetic correlation scores, and combined models that include both environmental and multi-PRSs features. This approach enables quantification of the incremental predictive value of genetic information, examine whether correlated PRSs have predictive value in addition to the primary disorder score, whether correlated PRSs independently predict the risk, and examine the generalisability of predictive models in an independent external cohort. Results In the UK-MCS cohort, a baseline model using sex as the sole predictor achieved an AUC of 0.63 ± 0.01 for DEP at age 14. Incorporating multiple polygenic risk scores (PRSs) of genetically correlated traits alongside environmental variables — including parental education, income, and employment status, marital status, teachers’ assessments, and school and neighbourhood-related measures — improved model performance to an AUC of 0.69 ± 0.01. In the case of ADHD, the AUC values were 0.59 ± 0.012 for the sex-only model and 0.76 ± 0.02 for the full model. In the external validation for the ABCD cohort, a modest decline in discriminative ability was observed for DEP outcome: AUC for the sex-only model was 0.55 and for the full model was 0.61; for the ADHD outcome, the AUC for the sex-only model was 0.53 and for the full model was 0.60. SHAP-based feature importance analysis indicated that the two strongest predictors for the model were sex and PRS-MDD for the DEP outcome, while the top predictors for the ADHD outcome were the attention scores from the teachers' assessment and sex. Conclusion This study contributes to the understanding of the genetic risk factors for mental health outcomes in adolescence. The model combining various polygenic risk scores for genetically correlated traits and environmental/sociodemographic factors shows increased discriminative ability for the outcomes of DEP and ADHD in early adolescence.