Abstract
BACKGROUND: Externalizing liability is a strong risk factor for early substance initiation, but the neurobiological pathways linking polygenic risk to initiation remain incompletely characterized. METHODS: Using the ABCD Study, we implemented a four-stage framework linking an externalizing polygenic risk score (extPRS) to baseline multimodal neuroimaging-derived phenotypes (IDPs) and longitudinal substance initiation (alcohol [primary], nicotine, cannabis, and any substance). First, we screened extPRS-IDP associations using covariate-adjusted linear models (age, sex, ancestry principal components, site/scanner variables; modality-specific covariates where applicable) and controlled multiple testing using false discovery rate (FDR) procedures. Second, we estimated direct extPRS associations with time-to-initiation using Cox proportional hazards models. Third, we fit joint Cox models including extPRS and each discovery-significant IDP, retaining outcome-IDP associations after within-outcome FDR correction. Fourth, we conducted mediation analyses for prioritized outcome-IDP pairs using an extPRS → IDP mediator model and an initiation model including both extPRS and IDP, estimating indirect (ACME) and direct (ADE) effects via bootstrap with multiple-testing control. RESULTS: Among 10,608 participants, higher extPRS was associated with earlier initiation across outcomes, with the largest effects observed for nicotine and cannabis and a modest but significant effect for alcohol. Stage 1 identified thousands of extPRS-associated IDPs that were highly concordant across robustness specifications. Stage 3 prioritized outcome-specific IDPs associated with initiation beyond extPRS, with the number of retained IDPs varying across sensitivity settings (site-clustered vs. HC3 standard errors; SES covariates on/off) but showing a replicated core set across models. In Stage 4, mediation analyses showed that indirect effects of extPRS through IDPs were small in magnitude (ACME ≈ 10(-4)) and accounted for less than 2% of the total effect, while direct effects (ADE ≈ 0.02-0.05) remained strong across outcomes. FDR-significant mediation signals were observed only for alcohol and any-substance initiation, whereas no mediation effects survived multiple testing correction for cannabis or nicotine. Across outcomes, direct genetic effects were substantially larger than mediated effects, indicating that genetic liability operates primarily through direct pathways rather than through baseline brain measures. CONCLUSIONS: Externalizing polygenic liability is broadly associated with substance initiation, with robust and consistent direct effects across substances. Although specific frontal structural and microstructural phenotypes show statistically significant mediation signals, their contribution is small, suggesting that baseline brain measures explain only a limited proportion of genetic risk. This framework provides a scalable approach to prioritize neurobiological pathways linking genetic liability to early substance initiation while highlighting the dominant role of direct genetic effects.