Across methods of increasing genetic specificity, our results provide insights into the paradox that EA is positively genetically correlated with the use of alcohol or cannabis but negatively correlated with AUD and CUD. These directionally opposite r _g cannot be explained by differences in overall polygenic overlap, as EA has greater polygenic overlap with AUD and CUD (62.42% and 84.18%, respectively) than with AC and CanUse (57.57% and 48.07%, respectively). However, the effect direction of the shared variants highlighted a key distinction. Variants shared between EA and SUDs more often had opposing effects on the two traits, while those shared with substance use had greater concordance.

Disaggregating EA into its cognitive (CogEA) and noncognitive (NonCogEA) components clarified this distinction across levels of specificity, with important nuance across methods. At the global level, MiXeR estimated higher genome-wide concordance for CogEA with AC and CanUse than with AUD and CUD, consistent with the findings for EA overall. In contrast, conjFDR showed that the effect direction of jointly associated SNPs was more often concordant for CogEA and AUD/CUD than for CogEA and AC/CanUse. The discrepancy across the two methods is likely due to differences in scope. Whereas MiXeR estimates the polygenic overlap and concordance across all causal variants (including many with small effects), conjFDR identifies the SNPs that are most strongly jointly associated. Thus, although many variants positively associated with CogEA protect against developing AUD and CUD, a smaller set of variants with strong effects increases the risk for these disorders.

For NonCogEA, there were also differences across methods. MiXeR analyses showed higher concordance of NonCogEA with AC and CanUse than with AUD and CUD. For cannabis traits, however, conjFDR analysis showed the opposite pattern. SNPs jointly associated with NonCogEA and CUD were more directionally concordant than those jointly associated with NonCogEA and CanUse. Thus, although genetic variants associated with noncognitive traits promote cannabis use generally, the most strongly associated variants may disproportionately contribute to risk for developing CUD. Variants may influence both NonCogEA (increasing liability for EA through noncognitive traits) and CUD via risk-taking or socially adaptive forms of disinhibition (Moncel et al., 2025). In competitive academic environments, these tendencies facilitate goal pursuit and strategic risk-taking. For example, research shows that individuals in leadership positions often have elevated dominance (Lilienfeld et al., 2012, 2014) and that risk-taking is associated with better academic outcomes (Petzel & Casad, 2022). However, in other contexts, these same traits can confer vulnerability to CUD.

At the local level, results from LAVA showed that genomic regions where EA has local r _g with AC/CanUse are largely distinct from those where EA is correlated with AUD/CUD. This pattern was consistent regardless of whether EA was considered as a whole or partitioned into cognitive and noncognitive components. Previous research has also shown that experimentation with substances and the development of SUD are shaped by at least partially distinct etiological factors (Lynskey et al., 2003; Richmond-Rakerd et al., 2016). The minimal overlap identified in LAVA analyses suggests that transitions from use to disorder may involve a shift in biological pathways, such as from those associated with greater opportunity or openness to use substances to those linked with compulsivity and executive dysfunction, key elements of SUDs. Thus, local differences in pleiotropy may, in addition to differences in effect direction, contribute to the paradoxical associations of EA and its subcomponents with alcohol and cannabis use, and AUD and CUD.

Analyses performed at the most granular level provided insight into the specific biological pathways that differentiate CanUse from CUD. SNP-based tissue enrichment showed that whereas SNPs jointly associated with CUD and CogEA were enriched in the substantia nigra, those jointly associated with CanUse and NonCogEA were enriched in the cerebellum and its hemispheres. Elevated dopamine function in the substantia nigra has been implicated in increased psychosis risk among individuals with CUD (Ahrens et al., 2025), and psychosis is often preceded by cognitive deficits (Jonas et al., 2022). These dopaminergic pathways in the substantia nigra may help explain the overlap between CUD and CogEA. In contrast, chronic cannabis use has been associated with altered cerebellar structure and poorer performance on noncognitive behavioral tasks that involve the cerebellum (Blithikioti et al., 2019). Thus, the potentially distinct neurobiological pathways underlying substance use and SUDs may also differ in their associations with cognitive and noncognitive traits. Although we also observed enrichment in several non-brain tissues, the relevance of these findings to cognition or substance-related behaviors is unclear and warrants cautious interpretation.

In contrast to the divergent effects between CanUse and CUD that were revealed by enrichment analyses, we found evidence of shared effects between AUD and AC. SNPs jointly associated with AUD and EA, as well as those associated with AC and EA, were linked to enriched expression in brain tissues during prenatal development. Prenatal stress plays a causal role in the development of SUDs later in life (Pastor, Antonelli, & Pallares, 2017) and is related to lower school achievement during childhood (Li et al., 2013). Our findings point to a shared genetic etiology that may account for this co-occurrence.

When EA was broken down into its subcomponents, SNPs jointly associated with AUD and CogEA showed enriched expression in the hypothalamus and two components of the basal ganglia – the NAcc and caudate nucleus. The basal ganglia have already been a target for intervention in AUD, with deep brain stimulation of the NAcc producing significant reductions in alcohol-related craving and compulsive behavior (Davidson et al., 2022). Other evidence supports the importance of these regions for understanding alcohol-related risk and cognitive functioning. For example, lower white matter integrity between the NAcc and prefrontal cortex is prospectively associated with earlier initiation of binge drinking (Morales et al., 2020). Similarly, the caudate nucleus is sensitive to prenatal alcohol exposure, with lower volume in this region predicting poorer cognitive control in individuals exposed to alcohol in utero, even after controlling for total brain volume, intelligence, and age (Fryer et al., 2012). Taken together, these findings suggest that genetic variants jointly associated with CogEA and AUD may influence risk by shaping the striatal circuits integral to balancing reward motivation with executive control.

Our study has several limitations. First, analyses were conducted only among EUR individuals due to the limited availability of GWAS in other population groups, which potentially limits the generalizability of our findings to other genetically inferred ancestry groups. Additionally, educational systems, access to education, and the relationship between education and health differ across societies (Eikemo, Huisman, Bambra, & Kunst, 2008). Thus, the genetic architecture linking EA and substance-related traits may vary across groups as well. Partitioning EA into CogEA and NonCogEA offered a more nuanced view of this complexity, but studies are needed that examine whether the observed patterns are consistent in other settings and groups. Second, the AUD and AC GWAS comprised predominantly of the Million Veteran Program participants, of whom the vast majority are male (~90%). Although sex-stratified GWAS of substance use traits are limited, recent efforts have identified several sex-specific genetic loci for alcohol use behaviors (Vilar-Ribó et al., 2025), and these GWAS could prove useful for future studies investigating shared genetic architecture with educational traits. A third consideration is that AC and CanUse reflect different stages of substance involvement and differ in their associations with AUD and CUD, respectively. This likely contributed to the clearer separation between CanUse and CUD in our results relative to AC and AUD, which often showed more similar associations with educational traits. A recently published GWAS of cannabis use frequency in 23andMe participants provides an opportunity for future exploration of these associations (Thorpe et al., 2025). Fourth, differences in GWAS sample sizes (e.g. for CogEA [N ≈ 258k] and NonCogEA [N ≈ 511k]) influence the power to detect polygenic overlap across the methods we used, with greater effects on SNP-level discovery than genome-wide estimates. Finally, although our analytic approach captured the overlap and divergence in genetic architecture for EA and substance-related traits, it does not provide insights into the causal directions of these associations.

By integrating multiple analyses of increasing genetic specificity, we show that the paradoxical genetic correlations between EA and alcohol/cannabis use versus AUD/CUD reflect more than just differences in overall polygenic overlap. Instead, a combination of differences in effect direction, associated genomic regions, and neurobiological pathways better explains the paradoxical associations. Partitioning EA into its cognitive and noncognitive components highlights how traits generally viewed as promoting educational success – including curiosity, openness, and cognitive ability – may also increase vulnerability to the development of SUDs, particularly CUD. These findings reinforce our understanding of EA as a multidimensional trait with complex links to substance-related behaviors.