Discordant associations of educational attainment with ASD and ADHD implicate a polygenic form of pleiotropy

ASD and ADHD are highly polygenic neurodevelopmental disorders that frequently co-occur and share genetic architecture. Prior studies show substantial genetic correlations between ASD and ADHD, yet their polygenic associations with cognitive outcomes diverge: ADHD genetic risk correlates negatively with educational attainment (EA), while ASD genetic risk correlates positively. The mechanisms underlying these discordant associations are unclear and may involve various forms of pleiotropy, co-localization of risk variants in linkage disequilibrium (LD), or ascertainment biases. This study aims to investigate genetic mechanisms that manifest as ASD-related positive and ADHD-related negative polygenic associations with EA, and to identify and annotate underlying variants, using multivariable regression on GWAS summary statistics without making causal inferences.

Evidence from twin and GWAS studies supports shared genetic influences between ASD and ADHD across population and clinical levels. Reported SNP-based genetic correlations between ASD and ADHD range from about 0.36 in molecular studies to 0.87 in twin analyses. Despite overlap, polygenic scores reveal opposite associations with EA: increased ADHD risk relates to lower cognitive abilities and EA, whereas increased ASD risk relates to higher cognitive functionality and EA, particularly for years of schooling and college completion. Potential mechanisms for discordant associations include: (I) independent markers tagging independent risk alleles; (II) ascertainment bias related to socioeconomic status; (III) opposite alleles at the same marker tagging opposite or independent risk alleles; (IV) identical marker alleles tagging independent risk alleles via high LD (biological or spurious pleiotropy); and (V) identical risk alleles exerting different effects (biological pleiotropy). Prior cross-disorder GWAS identified pleiotropic loci across psychiatric disorders and suggested complex shared biology, but did not specifically resolve ASD-ADHD discordance with EA.

Design: A bidirectional multivariable regression (MVR) framework analogous to multivariable MR-Egger was applied to GWAS summary statistics to simultaneously estimate polygenic ASD- and ADHD-associated effects on EA while allowing for pleiotropy and minimizing collider bias by using genetically predicted phenotypes. Data: European-ancestry GWAS summary statistics included: ASD(iPSYCH, woADHD) N=32,985 (10,321 cases); ADHD(iPSYCH) N=37,076 (14,584 cases); ASD(PGC) N=10,610 (5,305 cases); EA(SSGAC, years of schooling, excl. 23andMe) N=766,345; intelligence(CTG); and for specificity, MDD, SCZ, BD from PGC. iPSYCH controls were shared across ASD and ADHD. Imputation predominantly to 1000 Genomes phase 3 (EA and iPSYCH) or HRC (intelligence). All samples were of European descent. Variant selection: Independent, common, well-imputed SNPs (minor allele frequency >0.01; LD-r2<0.25 within ±500 kb; INFO>0.7) were selected from ASD and ADHD GWAS across 11 P-value thresholds (5e-8 to 0.5), following polygenic scoring guidelines. Two primary thresholds were emphasized: Pthr<0.0015 and Pthr<0.05. Models: Two complementary MVRs were run across variant sets: (1) ASD-MVR using ASD-selected variants, estimating aggregate EA association effects for ASD (βASD) and ADHD (βADHD) simultaneously, with an intercept to capture residual pleiotropy; (2) ADHD-MVR analogously using ADHD-selected variants. SNP effect alleles were aligned to increase the disorder risk corresponding to the variant selection (ASD for ASD-MVR, ADHD for ADHD-MVR). Model fit was compared to univariable regressions; collinearity was assessed via VIF. Concordant alleles: To distinguish mechanisms, analyses were repeated restricting to variants carrying the same risk-increasing allele for both disorders (~80% overlap at Pthr<0.0015 and <0.05). Replication: ASD SNP estimates from ASD(PGC) replaced ASD(iPSYCH, woADHD) to replicate MVR findings at Pthr<0.0015 and <0.05. Regional analysis (gwas-pw): The genome was partitioned into approximately independent LD blocks. gwas-pw estimated posterior probabilities that blocks contain shared vs non-shared effects between ASD and ADHD, accounting for sample overlap. Blocks with posterior probability >0.9 for shared effects (biological pleiotropy or high-LD co-localization) were followed by MVR using independent variants from those blocks without P-value filtering. Single-variant identification (conditional thresholding): Starting from ASD and ADHD discovery variant sets at Pthr<0.0015, overlapping independent SNPs associated with both disorders were identified across conditional P-value thresholds for the other disorder (0.0015 to 0.5), yielding nested conditional subsets. The most stringent joint threshold (ASD and ADHD Pthr<0.0015) defined 83 loci. Permutation testing (10,000 permutations) evaluated chance findings. Functional enrichment: The 83 loci were mapped to 52 genes (RefSeq, build 37) and tested for gene-set enrichment (Molecular Signature Database v7.0; WikiPathways v20191010; GWAS Catalog e96_2019-09-24) using MAGMA within FUMA with FDR control. Genetic correlations and meta-analyses: SNP heritability and genetic correlations were estimated by LDSC. Cross-disorder meta-analyses (ASD+ADHD; and in specificity analyses other pairs) allowing for sample overlap were conducted with METACARPA to assess effects on genetic correlations with EA. Multiple testing: Discovery analyses corrected at P<0.0023 (22 tests); replication and specificity thresholds adjusted accordingly (e.g., P<0.0125 for four tests).

- Shared EA-related variation across ASD and ADHD: Using MVR, the same sets of variants captured discordant polygenic associations with EA for ASD (positive) and ADHD (negative), inconsistent with mechanisms requiring independent markers. - Discovery MVR (Pthr<0.0015): • ASD-MVR (NSNPs=1,973): EA increased by 0.009 years per log-odds ASD liability (βASD=0.009, SE=0.003, P=0.002), while the same alleles captured a decrease of 0.029 years per log-odds ADHD liability (βADHD=−0.029, SE=0.004, P<1×10−10). • ADHD-MVR (NSNPs=2,717): EA decreased by 0.012 years per log-odds ADHD liability (βADHD=−0.012, SE=0.003, P=4×10−5) and increased by 0.022 years per log-odds ASD liability (βASD=0.022, SE=0.003, P<1×10−10). - Including more variants (Pthr<0.05) increased power; multivariable models explained up to 3% more variance in genetically predictable EA than univariable models (VIF ≤1.2). - Concordant risk alleles: Restricting to variants with the same risk-increasing allele for both disorders retained discordant MVR effects, arguing against opposite alleles at the same marker (scenario III) and supporting identical marker alleles tagging either independent nearby risk variants (scenario IV) or the same risk variant with different effects (scenario V). - Replication with ASD(PGC): Discordant patterns were confirmed at Pthr<0.05 despite lower power (e.g., positive ASD–EA association using ASD(PGC) SNP estimates: βASD=0.005, SE=0.001, P<1×10−10). - gwas-pw regional signals: Three LD blocks showed high posterior probability (>0.9) for shared effects (pleiotropy/high-LD co-localization): 1p21.3 (PTBP2, DPYD), 5q14.3 (TMEM161B, LINC00461, MIR9-2, MEF2C), 20p11.22–p11.23 (PLK1S1/KIZ, XRN2, NKX2-2, PAX1). MVR across these regions confirmed discordant polygenic associations, albeit weaker than discovery analyses. - Conditional joint loci: At joint ASD and ADHD Pthr<0.0015, 83 loci (99% same risk-increasing allele) were identified genome-wide (30 identical SNPs + 53 LD proxies; LD-r2≥0.6, 500 kb). These loci produced markedly larger polygenic effects: e.g., ASD-MVR with 83 loci: βASD=0.15 (SE=0.025, P=1×10−7), βADHD=−0.15 (SE=0.025, P=1×10−7) years of schooling per log-odds liability. Permutations supported non-randomness (empirical P<2×10−4). - Functional enrichment: The 83-locus gene set was enriched for microRNA targets, notably MIR19A/MIR19B (FDR=7.7×10−4) and MIR9 (FDR=0.028), implicating regulatory mechanisms (e.g., CACNA1C, ERBB4). - Effect cancellation: Meta-analyzing ASD and ADHD summary statistics attenuated the genetic correlation with EA, consistent with cancellation of opposing EA-related effects across shared regions. - Specificity: Discordant associations generalized to intelligence; exploratory analyses indicated EA-related variation shared with adult-onset psychiatric disorders, sometimes with discordant patterns, and predicted attenuation of EA genetic correlations upon meta-analysis of discordant pairs.

The study demonstrates that ASD- and ADHD-associated polygenic variation shares EA-related genomic signals captured by identical marker alleles, yet these alleles encode opposite aggregate associations with EA for the two disorders. This rules out mechanisms relying solely on independent markers and suggests either biological pleiotropy at the same causal variants (different effects on ASD vs ADHD) or high-LD co-localization of distinct nearby causal variants within the same locus. At the polygenic level, these independent aggregate effects across shared markers constitute a polygenic form of pleiotropy. The observation that cross-disorder meta-analyses reduce genetic correlations with EA supports effect cancellation across EA-related regions, aligning with reports that accounting for socio-economic genetic variation can modify detectable ASD–ADHD correlations. Enrichment for MIR19A/19B and MIR9 target genes points to miRNA-mediated regulatory mechanisms in neurodevelopment as contributors to the observed pleiotropy. Multivariate modeling improved fit over univariable models, indicating that simultaneous modeling of multiple disorders can de-stratify overlapping polygenic signals and reveal discordant relationships that univariate approaches may obscure.

EA-related polygenic variation is shared between ASD and ADHD architectures and, when aggregated across identical marker alleles, yields ASD-positive and ADHD-negative associations with EA. These findings support a polygenic pleiotropy framework—via biological pleiotropy and/or high-LD co-localization—that shapes the detectable genome-wide overlap between ASD and ADHD and explains attenuation of EA correlations in cross-disorder meta-analyses. A focused set of 83 jointly associated loci with strong polygenic effects, enriched for microRNA target pathways (notably MIR19A/MIR19B and MIR9), highlights regulatory mechanisms in neurodevelopment. Future work should leverage larger ASD and ADHD cohorts, incorporate independent ADHD replication, dissect symptom dimensions, and integrate functional assays to validate miRNA-related mechanisms and clarify causal variant architecture.

- Power limitations: The ASD(PGC) sample is relatively small, reducing replication power and regional detection via gwas-pw. gwas-pw’s conservative overlap correction may underestimate shared polygenic signals. - Winner’s curse: ADHD-MVR may be affected due to lack of an independent ADHD discovery/replication set. - Ascertainment bias: Although considered unlikely given multivariate results, the study cannot fully exclude biases related to case ascertainment or socioeconomic status. - Phenotypic heterogeneity: Variation in ASD/ADHD symptom dimensions could influence genetic overlaps with EA and cognition but was not dissected here. - Sample overlap in specificity analyses: Some adult-onset disorder datasets share cohorts with EA, introducing potential confounding; these analyses are exploratory. - Ancestry: Analyses are limited to European ancestry, which may affect generalizability. - LD-based inference: Distinguishing true biological pleiotropy from high-LD co-localization at single variants remains challenging without fine-mapping and functional validation.