81. NORMATIVE MODELLING WITH POLYGENIC SCORES REVEALS DISTRIBUTIONAL ASSOCIATIONS WITH DEPRESSION AND ENVIRONMENT

Background Phenotypic variance explained by polygenic scores (PGS) is of major interest in genetic studies, but individuals’ deviations from PGS-predicted outcomes have received less attention. Reframing residuals as deviation phenotypes shifts focus from mean effects to structured mental health/environmental correlates. This aligns with normative modeling, which estimates population heterogeneity and links deviation scores to mental health severity. Here, we applied this framework to predict cognitive performance from educational attainment (EA) PGS. Deviation scores then indexed how strongly an individual’s cognitive performance diverged from the genetic expectation. We assessed whether these deviations were associated with depressive symptoms and environmental factors, extending normative modeling to PGS. Methods We used ABCD data of N=10,831 participants at baseline and calculated an EA-PGS using PRS-cs auto and GWAS sumstats of ∼1.1 million participants. Participants were then split into three 70/30 train-test folds, balanced for depressive symptom scores. We fit hierarchical Bayesian models predicting cognitive scores from EA-PGS (+age, sex, site, ancestry PCs), yielding out-of-sample Z-score residuals. To the same train-test fold pairs we fit an ordinary least-squares (OLS) model for comparison. We tested 3 depressive symptoms and 16 environmental metrics against deviation scores across models and data folds, using linear regression for continuous variables and Spearman correlation for ordinal variables. We tested extreme deviators (|Z| > 1.645) for patterns in depressive symptoms and environmental distributions. Results Across folds, Anhedonia (Cohen’s d ∼ -0.29, p-FDR ∼ 0.0009) and excessive worrying (Cohen’s d ∼ -0.37, p-FDR ∼ 0.009) predicted downward deviation shifts, indicating lower-than-expected cognitive performance scores relative to EA-PGS. Several environmental measures were associated with deviation scores across folds, the strongest of which were Childhood Opportunity Index (COI; R2 ∼ 0.02, p-FDR ∼ 8.3*10-16) and Area Deprivation Index (ADI; R2 ∼ 0.016, p-FDR ∼ 1.3*10-11). Multiple environmental metrics showed distributional differences when comparing positive and negative extreme deviators, the strongest among which were ADI (Cohen’s d ∼ 0.66; pFDR ∼ 4*10-7) and COI (Cohen’s d ∼ -0.63; pFDR ∼ 8*10-12). Proximity to Roads, Social Vulnerability Index, Family Conflict, and Neighborhood Safety shifted in positive extreme deviators (> 95th percentile) relative to the middle distribution, effects missed by OLS. Normative modeling identified stronger tail-to-tail differences than OLS in 8/12 environments. Discussion EA-PGS-cognitive deviations showed downward shifts among participants with anhedonia and excessive worrying, as well as distributional differences in environmental metrics among extreme deviators. Extreme deviators may index vulnerability/resilience, as positive and negative deviations from EA-PGS-predicted cognitive performance tracked patterns of environmental deprivation and opportunity. Residuals capture measurement error and unmodeled genetic effects but also contain structured signal. Normative modeling detected distributional, tail-specific effects not captured by OLS across folds. Extending this approach to multivariate PGS may further improve sensitivity to shared genetic liability across correlated traits and reveal more structured patterns of environmental association.