Polygenic and machine learning analysis of medication response in a prospective pediatric cohort initiating ADHD medication

While medication for attention deficit hyperactivity disorder (ADHD) is effective, predictors of treatment response remain poorly understood. Given the use of polygenic scores (PGS) in psychiatric conditions, we investigate their potential to predict ADHD medication response. In this work, we recruited 309 participants aged 6-17 years from the Swedish ADHD medication and predictors of treatment outcome (ADAPT) study (ClinicalTrials.gov: NCT02136147; June 2015), and calculated their PGS. Compared with a reference cohort (N=780), ADAPT showed higher PGS for ADHD (p<2.22e-16), and lower PGS for autism (p=0.0051), educational attainment (EA) (p=8e-10), and intelligence (p=7e-08). To assess the association between genetic effects and changes in Swanson, Nolan, and Pelham ADHD Rating Scale-version IV (SNAP-IV) scores, linear mixed-effects models tested the interaction between each PGS and time (baseline, 3-months). PGS for ADHD was associated with higher SNAP-IV scores overall ({beta}=0.1419, p=0.0078, FDR=0.0293) but not with a change at follow-up; whereas, PGS for EA was associated with lower overall scores ({beta}=-0.1161, p=0.0256, FDR=0.0679) and smaller reductions at follow-up ({beta}=0.0946, p=0.0453, FDR=0.1130). Subgroup analyses indicated that these effects were more prominent in adolescents ([&ge;]13 years) and individuals classified as responders to treatment. We further explored machine learning models combining clinical and genetic information across 14 models with 49 features. Although models could not empirically generalize, estimated information-theoretic upper bounds for classification accuracy were 72% with PGS, 84% with clinical and demographic variables, and 92% when combined. This study highlights the potential of using PGS and combining genetic and clinical data to improve treatment stratification in ADHD.