57. MULTI-PGS MODELS IMPROVE PREDICTION OF ADHD IN ADULTS AND CHILDREN

Background Attention deficit/hyperactivity disorder (ADHD) is an early-onset neurodevelopmental disorder characterized by impairing levels of hyperactivity, inattention or impulsivity. Polygenic scores (PGSs) indexing genetic susceptibility for this disorder explain only a small proportion of the phenotypic variance. New approaches such as the multi-PGS, which combines multiple single-trait PGSs using penalized regression models, have shown to improve the prediction accuracy for different traits, ADHD among them. Here we aim to provide new insights into the genetic architecture of ADHD by: (i) constructing a hypothesis-driven categorized multi-PGS framework in an adult clinical cohort and (ii) evaluating the generalizability of these models in a population-based child cohort. Methods ADHD diagnosis and symptoms were assessed using structured interviews in a clinical sample of adults (1,789 cases and 4,729 controls) and a population-based sample of children and adolescents (791 cases and 5,354 controls). Polygenic scores were generated using PRScs and PLINK. First, in the adult sample we compared the prediction accuracy of a single-PGS model including the PGS for ADHD (PGSADHD) against categorized multi-PGS models combining the PGSADHD with multiple PGS for risk factors grouped in six categories: environment, mental-health, cognition, personality, health-risk behavior and non-mental diseases. Second, the PGSs showing significant weights were subsequently tested in the child population-based sample. Results Multi-PGS models for environment, mental health, health-risk behaviors and personality substantially increased the explained variance by 11.2%, 87.2%, 16.8%, and 45.6%, respectively, compared with the single PGSADHD model. Significant contributors included PGSs associated with psychiatric disorders (e.g., bipolar disorder and cannabis use disorder), risk-taking behaviors, smoking-related traits, and personality dimensions such as openness and neuroticism. Genetic liability related to loud music exposure and mobile phone use also significantly contributed to the predictive accuracy of ADHD. Furthermore, these results were generalized to the independent sample of children where multi-PGS also increased the explained variance compared to the single PGSADHD (FDR-adjusted p < 0.05). However, no significant improvement in model fit was observed when ADHD symptoms were considered. Discussion These results further support the utility of multi-PGS methods to increase explained variance and the potential of classifying PGS into different domains to provide meaningful and more interpretable insights into the etiology of ADHD. The model fit improvement of the multi-PGS models relative to the single PGSADHD for ADHD diagnosis in the child sample suggests that part of the genetic liability captured by the multi-PGS framework is shared across developmental stages. However, the lack of improvement for symptom scores may indicate that part of the genetic signal captured by the multi-PGS framework is more strongly related to clinically significant manifestations of ADHD rather than to continuous variation in symptom presentation within the general population.