Shared genetic architecture between ADHD and intelligence varies across ADHD subtypes

Abstract

Background

Attention-deficit/hyperactivity disorder (ADHD) is a heterogeneous neurodevelopmental condition frequently accompanied by cognitive difficulties. Although previous genetic studies have demonstrated substantial overlap between ADHD and intelligence, most have treated ADHD as a single phenotype. However, whether this shared genetic architecture differs across ADHD subtypes remains unclear.

Methods

We conducted a genome-wide cross-trait analysis integrating large-scale genome-wide association study (GWAS) datasets of overall ADHD, its subtypes—childhood ADHD, persistent ADHD, and late-diagnosed ADHD—and intelligence (total N > 300,000). Genome-wide genetic correlations, polygenic overlap, local genetic correlations, and variant-level associations between ADHD phenotypes and intelligence were evaluated to characterize their shared genetic architecture. Shared variants were identified through cross-trait enrichment analyses and subsequently mapped to genes for functional annotation and gene-set enrichment. Bidirectional associations were evaluated using two-sample Mendelian randomization with sensitivity analyses. Additional GWAS datasets were used to validate the robustness of shared loci by assessing the consistency of effect directions.

Results

All ADHD phenotypes showed significant negative genetic correlations with intelligence (r_g ranging from -0.3442 to -0.4205). Despite these modest genome-wide correlations, cross-trait analyses revealed substantial genetic overlap, including polygenic overlap, local genetic correlations, and variant-level associations. We identified 184 loci jointly associated with ADHD traits and intelligence, including 64 novel loci, whereas no shared loci were detected for persistent ADHD under the current analysis. Functional annotation revealed biologically distinct enrichment patterns across subtypes: childhood ADHD loci were linked to early neurodevelopmental processes, while late-diagnosed ADHD loci were enriched in synapse-related and neuronal signaling pathways. Mendelian randomization analyses suggested bidirectional associations, with stronger evidence supporting a directional association from intelligence to ADHD risk. Furthermore, these shared loci showed largely consistent effect directions across additional GWAS datasets, providing support for the robustness of the findings.

Conclusions

The shared genetic architecture between ADHD and intelligence varies across ADHD subtypes, highlighting distinct biological pathways underlying cognitive heterogeneity in ADHD. These findings suggest that the relationship between ADHD liability and general cognitive ability is not uniform across ADHD subtypes and may inform future research on risk stratification and early identification in child and adolescent psychiatry.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12916-026-04998-3.

Background

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder characterized by persistent patterns of inattention, hyperactivity, and impulsivity [1, 2]. Affecting approximately 5–6% of children and 2.5% of adults worldwide, ADHD represents one of the most prevalent psychiatric conditions [3]. ADHD is clinically heterogeneous [4, 5] and comprises multiple subtypes based on developmental course: childhood ADHD diagnosed early in life, persistent ADHD continuing into adulthood, and late-diagnosed ADHD referring to individuals whose diagnosis occurs later in development [6]. These subtypes differ in developmental trajectories, symptom profiles, and genetic underpinnings [6–10]. Understanding these differences is essential for clarifying the broader impact of ADHD and identifying subtype-specific characteristics.

ADHD is frequently associated with cognitive difficulties across multiple domains [11], particularly in executive function, working memory, and attention control [12, 13]. These cognitive domains are crucial for adaptive functioning and are shaped by neurobiological processes including synaptic plasticity, neural connectivity, and cortical development [14, 15]. Among cognitive phenotypes, intelligence reflects general cognitive ability, a latent factor capturing variance shared across multiple neurocognitive tasks, and has been extensively studied and shown to be associated with ADHD [16–18]. Importantly, clinical presentations vary across ADHD subtypes—for example, adult ADHD is characterized by relatively reduced hyperactivity and increased inattentiveness compared to childhood ADHD [7–9], whereas persistent ADHD is associated with greater early-life symptom severity and higher comorbidity burden [10]. Such heterogeneity suggests that cognitive manifestations may also vary across subtypes, motivating a more detailed examination of ADHD in relation to intelligence as a general cognitive phenotype. Given the substantial heritability of both ADHD and intelligence, examining their relationship at the genetic level may provide important insights into the genetic basis of cognitive variability across ADHD subtypes.

Genome-wide association studies (GWAS) have advanced our understanding of the genetic basis of both ADHD and intelligence, identifying numerous associated variants [19–22]. Previous large-scale genetic studies have demonstrated shared genetic architecture between ADHD and intelligence, including substantial genetic overlap and pleiotropic loci identified through approaches such as bivariate causal mixture model (MiXeR) and conditional and conjunctional false discovery rate (cond/conjFDR) analyses [23, 24], as well as consistent negative genetic correlations reported across multiple studies [25–27]. Mendelian randomization (MR) analyses have also been conducted to explore potential causal relationships between ADHD and intelligence [28–30]. However, these previous studies have treated ADHD as a single entity, without systematically examining subtype-specific genetic influences. Whether the shared genetic basis between ADHD and intelligence differs across clinically and genetically heterogeneous ADHD subtypes—childhood, persistent, and late-diagnosed ADHD—remains unclear.

Building on these findings, this study systematically examined the shared genetic architecture between overall ADHD, its subtypes—childhood, late-diagnosed, and persistent ADHD—and intelligence, using the largest currently available ADHD GWAS dataset together with subtype-stratified data. Specifically, genome-wide genetic correlations were estimated to assess global genetic relationships, polygenic overlap was quantified to characterize the extent of shared genetic influences, and local genetic correlations were performed to localize shared effects to specific genomic regions. At the variant level, shared variants were identified using cross-trait enrichment strategies and subsequently functionally annotated to explore their biological relevance. Bidirectional MR analyses were further conducted to examine genetically informed directional associations between ADHD traits and intelligence. By integrating these complementary approaches and leveraging large-scale GWAS datasets, this study aims to clarify how ADHD heterogeneity shapes its genetic relationship with intelligence. An overview of the analytical pipeline is provided in Fig. 1.

Fig. 1.

Overview of the study workflow. Publicly available GWAS summary statistics for overall ADHD, childhood ADHD, late-diagnosed ADHD, persistent ADHD, and intelligence were harmonized and quality controlled. Global genetic correlation was estimated using LDSC, and polygenic overlap was quantified using MiXeR. Local genetic correlation was assessed using LAVA. Shared variants between ADHD traits and intelligence were identified using cond/conjFDR analyses. Genomic loci were then defined following the FUMA protocol and further merged across trait pairs to generate a nonredundant set of distinct loci. Functional annotation, gene mapping, and GO enrichment analyses were subsequently performed. Bidirectional MR analyses were further conducted to evaluate potential directional genetic associations. Abbreviations: ADHD, attention-deficit/hyperactivity disorder; condFDR/conjFDR, conditional and conjunctional false discovery rate; EUR, European; FDR, false discovery rate; FUMA, Functional Mapping and Annotation; GO, Gene Ontology; GWAS, genome-wide association study; IVW, inverse-variance weighted; LAVA, local analysis of [co]variant association; LD, linkage disequilibrium; LDSC, linkage disequilibrium score regression; MAF, minor allele frequency; MHC, major histocompatibility complex; MR, Mendelian randomization; SNP, single nucleotide polymorphism

Methods

Data sources

The dataset for overall ADHD was derived from the largest GWAS conducted to date, which included 38,691 individuals with ADHD and 186,843 controls, all of European ancestry [20]. Additionally, GWAS data stratified by ADHD subtypes—including childhood ADHD (N = 14,878), late-diagnosed ADHD (N = 6,961), and persistent ADHD (N = 1,473), along with 38,303 controls—were included to capture potential subtype-specific genetic architectures, all from individuals of European descent [6]. Subtype classification was based on age at first recorded ADHD diagnosis using 18 years as the cutoff: childhood ADHD was defined as a diagnosis before age 18, persistent ADHD as diagnoses both before and after age 18, and late-diagnosed ADHD as the first diagnosis recorded after age 18.

For intelligence, we used GWAS summary statistics focusing on general cognitive ability, with intelligence quotient serving as the primary outcome measure [21]. The intelligence dataset consisted of 269,867 individuals of European ancestry from 14 cohorts.

All data used in this study was obtained from publicly available GWAS summary statistics, with ethical approval and informed consent obtained from participants in the original studies. Details of datasets and methods are provided in the Additional file 1: Supplementary Methods.

Global genetic correlations between ADHD and intelligence

We estimated global genetic correlations using linkage disequilibrium score regression (LDSC) [31], which leverages the relationship between GWAS summary statistics and linkage disequilibrium (LD) patterns across the genome. Using pre-computed LD scores from the 1000 Genomes European reference panel [32] and restricting the analysis to HapMap3 single nucleotide polymorphisms (SNPs), we evaluated the genetic relationships between overall ADHD, its subtypes, and intelligence to obtain genetic correlation coefficients. To account for multiple testing, statistical significance was determined using the Benjamini-Hochberg false discovery rate (BH-FDR) procedure across all tested trait pairs, with q < 0.05 considered statistically significant.

Polygenic overlap revealed by MiXeR

We used the MiXeR [33, 34] tool to assess the polygenic overlap between ADHD and intelligence, estimating the number of causal SNPs based on GWAS summary statistics. First, we conducted univariate analyses to evaluate key parameters, including polygenicity, discoverability, and heritability [33]. Polygenicity refers to the number of genetic variants required to explain a significant portion of SNP heritability for a trait, with higher polygenicity reflecting a greater number of contributing variants. Discoverability measures the strength of additive genetic effects, with higher values suggesting that variants with larger effect sizes are more detectable. Heritability represents the proportion of phenotypic variance attributable to genetic factors, with higher heritability indicating a stronger genetic contribution.

Based on the results of these univariate analyses, we then performed bivariate MiXeR analyses to model the genetic effects of ADHD and intelligence as mixtures of distributions [34]. This analysis identified trait-specific variants, shared variants, and variants with no effect on either trait. The results were visualized in a Venn diagram, showing the shared and unique polygenic components across both traits, with the Dice coefficient (calculated from the bivariate output) quantifying the genetic overlap between ADHD and intelligence. The Akaike Information Criterion (AIC) was used to evaluate model fit in both univariate and bivariate MiXeR analyses. In univariate analyses, positive AIC values indicate that the GWAS summary statistics provide sufficient signal to support the MiXeR mixture model relative to alternative baseline models. In bivariate analyses, AIC values were used to assess the ability of the fitted model to discriminate between scenarios assuming minimal or maximal polygenic overlap, thereby evaluating the adequacy of the estimated shared genetic architecture between traits. To reduce potential biases from complex LD regions, we excluded the major histocompatibility complex (MHC) region (chr6:25,000,000–35,000,000) from both univariate and bivariate MiXeR analyses.

Local genetic correlation analysis

To further characterize the regional genetic architecture underlying the relationship between overall ADHD, its subtypes, and intelligence, we performed local genetic correlation analysis using local analysis of [co]variant association (LAVA) [35]. This framework partitions the genome into 2,495 approximately independent loci, each spanning ~ 1 Mb and defined according to local LD structure, thereby enabling the detection of locus-level genetic correlations that may not be evident in genome-wide analyses. For each ADHD-intelligence pair, univariate analyses were first performed across all loci to estimate local genetic signal (local h²_SNP) in both traits. Significance of local h²_SNP estimates was assessed using the BH-FDR procedure, and bivariate analyses were subsequently restricted to loci showing significant local genetic signal in both traits (BH-FDR q < 0.05). Within these loci, local genetic correlations (local r_g) were estimated using the bivariate method-of-moments framework implemented in LAVA, which models the local genetic covariance structure while accounting for local LD. Significance of local r_g estimates was determined using the test statistics implemented in LAVA and further adjusted using BH-FDR correction, with q < 0.05 considered statistically significant.

Identification of shared genetic variants

We utilized cond/conjFDR approach to identify genetic variants (i.e., SNPs) shared between ADHD and intelligence [36, 37]. CondFDR operates within an empirical Bayesian framework to enhance the detection of genetic associations for a primary phenotype by incorporating enrichment signals from a secondary phenotype. This is achieved by reordering the test statistics of the primary phenotype based on their association with the secondary phenotype, thereby recalibrating the significance of genetic variants. To explore cross-trait enrichment, conditional quantile-quantile (Q-Q) plots were generated to display the distribution of SNP associations for the primary phenotype stratified by varying levels of significance in the secondary phenotype (e.g., p-values less than 0.1, 0.01, and 0.001). A leftward deflection in the Q-Q plots at more stringent p-value thresholds indicated cross-trait enrichment, underscoring the presence of shared genetic signals.

Two condFDR analyses were subsequently conducted: one conditioning overall ADHD (or its subtypes) on intelligence, and the other conditioning intelligence on overall ADHD (or its subtypes). The conjFDR value for each SNP was determined as the maximum of the two mutual condFDR values, providing a conservative estimate of the false discovery rate for variants shared between the two traits. The significance thresholds were set at 0.01 for condFDR and 0.05 for conjFDR, consistent with previous studies [38–41]. To minimize potential biases, genomic regions with complex LD, including the MHC (chr6:25,000,000–35,000,000) and 8p23.1 (chr8:7,200,000–12,500,000), were excluded before fitting the FDR model [42–45].

Identification of independent genomic loci and novel loci definition

Independent genomic loci were defined through a stepwise procedure following the Functional Mapping and Annotation (FUMA) [46] protocol. Independently significant SNPs were identified as those meeting the predefined significance thresholds of condFDR < 0.01 or conjFDR < 0.05, and exhibiting low LD with one another (r² < 0.6). Lead SNPs were further selected from these independent significant SNPs, defined as those in minimal LD (r² < 0.1) with other independent significant SNPs. Candidate SNPs were then identified as variants with a cond/conjFDR < 0.1 and high LD (r² ≥ 0.6) with at least one independent significant SNP, and the genomic boundaries of each locus were determined by including all such candidate SNPs. Consistent with the default settings in FUMA [46] and commonly used thresholds in genetic association studies [39, 42, 47, 48], adjacent loci separated by less than 250 kb were merged into a single region, with the SNP showing the most significant FDR value selected as the representative lead SNP for the merged locus. To ensure a nonredundant locus set across ADHD traits, loci located within 250 kb across different trait pairs were further merged into distinct genomic regions. This approach reduces redundancy caused by nearby signals that may reflect the same LD structure. For each merged locus, the SNP with the most significant FDR value among the overlapping signals was selected as the representative lead SNP, while the original lead SNP from each trait pair was retained. After cross-trait merging, loci observed in only one trait pair were defined as trait-specific loci, whereas loci observed in two or more trait pairs were defined as trait-shared loci.

A locus was defined as novel if none of its candidate SNPs had been previously reported in the NHGRI-EBI GWAS Catalog [49] for associations with ADHD or intelligence, and if it was located more than 500 kb beyond the boundaries of any reported loci in the original GWAS summary (p < 1E-6) [6, 20, 21] or Kevin’s cond/conjFDR study [24]. Notably, novel loci for ADHD were determined by comparing each ADHD phenotype with all other ADHD phenotypes.

Functional annotation and enrichment analysis

We performed functional annotation using the FUMA [46] platform to explore the potential biological roles of the identified candidate SNPs. First, candidate SNPs were annotated with Combined Annotation Dependent Depletion (CADD) scores to predict their deleteriousness, with higher scores indicating a greater likelihood of functional impact. In this analysis, SNPs with CADD scores above 12.37, representing the deleterious variants in the genome, were considered potentially impactful [50]. RegulomeDB scores were employed to evaluate regulatory potential, incorporating data on transcription factor binding, chromatin accessibility, and histone modifications. Lower RegulomeDB scores (e.g., 1a, 1b) indicated strong regulatory evidence, whereas higher scores suggested limited or no regulatory relevance [51]. Chromatin states were also annotated using epigenomic datasets to determine whether candidate SNPs were located in active enhancers, promoters, or other regulatory regions, providing insights into their potential regulatory functions [52, 53].

Following annotation, the FUMA [46] platform was used to map candidate SNPs obtained from the conjFDR analysis to genes using three complementary approaches: positional mapping, expression quantitative trait loci (eQTL) mapping, and chromatin interaction mapping, with genes retained if mapped by any of these methods. Positional mapping assigned SNPs to the nearest genes within a 10 kb window upstream and downstream. eQTL mapping linked SNPs to genes whose expression levels were associated with genetic variation, based on multiple eQTL resources implemented in FUMA, including brain- and blood-related datasets such as BRAINEAC and GTEx v8. Chromatin interaction mapping associated SNPs in regulatory regions with genes through physical chromatin interactions, using tissue-specific Hi-C datasets from blood and brain to capture long-range regulatory effects. Finally, gene-set enrichment analysis was conducted using g: Profiler [54] to identify Gene Ontology (GO) biological processes associated with the mapped genes. The enrichment results were corrected for multiple testing using the BH-FDR method, with a significance threshold of q < 0.05.

Mendelian randomization analysis

We conducted MR [55, 56] analysis using a bidirectional two-sample framework to investigate potential directional genetic associations between ADHD and intelligence. The MR approach leverages genetic variants associated with the traits of interest as instrumental variables (IVs), allowing for the inference of directional effects. For each trait (overall ADHD, its subtypes, and intelligence) as an exposure, genetic variants were selected from GWAS summary data, and a genome-wide significance threshold of p < 5E-8 was applied to select robust IVs [57–60]. To ensure the independence of selected variants, LD clumping was performed with an r² < 0.001 and a 10,000 kb window. Proxy SNPs (r² > 0.8) were identified using the 1000 Genomes European reference panel for variants missing in the outcome dataset. Harmonization of alleles between exposure and outcome datasets was carried out to align the directions of effects, and palindromic variants with ambiguous strand alignment were excluded.

When multiple IVs were available, multiple MR estimators were applied, including the inverse-variance weighted (IVW) [61, 62], MR-RAPS [63], maximum likelihood [61], weighted median [64], and simple median [64] approaches. For analyses with a single instrument, the Wald ratio estimator was applied [65]. All MR results were adjusted for multiple testing using the BH-FDR correction. Associations were considered statistically significant at q < 0.05, with consistent effect directions across complementary MR estimators. Sensitivity analyses were performed to evaluate the reliability of the MR results. First, heterogeneity among the IVs was assessed using Cochran’s Q test, and horizontal pleiotropy was examined through the intercept term of the MR-Egger regression [66]. Second, MR-PRESSO was applied to identify and correct outlier IVs contributing to horizontal pleiotropy, with genetic estimates recalculated after excluding these outliers [67]. Finally, a leave-one-out analysis was performed to ensure that no single IV disproportionately influenced the observed genetic association.

Validation of shared genetic signals in additional GWAS datasets

To validate the robustness of shared loci identified by the conjFDR framework, we performed SNP sign tests [68, 69] to assess whether the directions of allelic effects were consistent across additional GWAS datasets, following approaches used in previous studies [70–72]. Specifically, we examined two additional GWAS summary-statistic datasets: an overall ADHD dataset [19] and an intelligence dataset [22].

For validation, lead SNPs from the shared loci identified in the discovery analyses were aligned to corresponding variants in the additional datasets. For each SNP, effect directions were compared between datasets to evaluate concordance. Validation was quantified as the proportion of lead SNPs showing concordant effect directions across datasets, and statistical significance was assessed using an exact binomial test under the null hypothesis of random effect-direction assignment. Multiple testing was controlled using the Benjamini-Hochberg procedure, with FDR q < 0.05 considered evidence of significant directional consistency.

Results

Global genetic correlation (LDSC)

We applied LDSC to estimate the global genetic correlation between ADHD and intelligence. The results showed a significant negative genetic correlation between overall ADHD and intelligence (r_g = -0.3828, se = 0.0232, BH-FDR q = 1.06E-60). For ADHD subtypes, late-diagnosed ADHD exhibited the strongest negative correlation (r_g = -0.4205, se = 0.034, BH-FDR q = 6.22E-35), followed by persistent ADHD (r_g = -0.3968, se = 0.0719, BH-FDR q = 3.36E-08) and childhood ADHD (r_g = -0.3442, se = 0.0303, BH-FDR q = 9.80E-30) (Fig. 2A, Additional file 2: Table S1).

Fig. 2.

Genome-wide genetic association between ADHD and intelligence. (A) LDSC results showing the genetic correlation (r_g) between overall ADHD, its subtypes, and intelligence. The color represents the strength of the correlation, with darker red indicating stronger negative correlations. The p values were adjusted for multiple testing using the BH-FDR procedure. (B) Estimates of discoverability, polygenicity, and heritability for ADHD traits and intelligence from univariate MiXeR analysis. (C) Venn diagrams derived from bivariate MiXeR analysis illustrating the genetic overlap between ADHD traits and intelligence. The values represent the number of genetic variants, shown in thousands, with standard errors provided in parentheses. Shared variants are represented in the overlapping regions, and trait-specific variants are depicted in the non-overlapping areas. (D) Bivariate MiXeR analysis showing the proportion of shared genetic variants between each ADHD trait and intelligence relative to the total variants of each ADHD (left), and the proportion relative to the total variants of intelligence (right). (E) Bivariate MiXeR-estimated Dice coefficient (left) and proportion of shared variants with concordant effect directions (right) for each ADHD-intelligence pair. Abbreviations: ADHD, attention-deficit/hyperactivity disorder; BH-FDR, Benjamini-Hochberg false discovery rate; LDSC, linkage disequilibrium score regression

Polygenic overlap (MiXeR)

The univariate MiXeR analysis provided estimates of discoverability, polygenicity, and heritability for ADHD and intelligence (Fig. 2B, Additional file 2: Table S2). Overall ADHD, childhood ADHD, late-diagnosed ADHD, and intelligence showed good model fit, as indicated by positive AIC values, whereas persistent ADHD showed a negative AIC value. Among these traits, persistent ADHD exhibited the highest discoverability (1.32E-4), with late-diagnosed ADHD (5.65E-5), childhood ADHD (4.81E-5), and overall ADHD (2.99E-5) showing progressively lower discoverability, and intelligence having the lowest discoverability (2.68E-5). Regarding polygenicity, persistent ADHD demonstrated the highest estimated number of trait-influencing variants (20,625), with intelligence having 10,135, late-diagnosed ADHD 8,590, childhood ADHD 7,825, and overall ADHD 7,774. For SNP heritability, late-diagnosed ADHD presented the highest estimate (0.308), followed by persistent ADHD (0.276), childhood ADHD (0.243), intelligence (0.176), and overall ADHD (0.151). Additionally, intelligence had the lowest discoverability, requiring the largest estimated sample size of approximately 19.5 million to capture 90% of its heritability. By contrast, the sample sizes needed for ADHD were relatively small, ranging from 6.7 million to 18.7 million (Additional file 2: Table S2, Additional file 3: Fig. S1).

The bivariate MiXeR analysis revealed substantial genetic overlap between ADHD and intelligence (Fig. 2C, Additional file 2: Table S3, Additional file 3: Figs. S2-5). Model fit evaluation using the AIC showed positive AIC_min and negative AIC_max values for overall ADHD, childhood ADHD, and late-diagnosed ADHD, whereas persistent ADHD showed negative AIC_min and AIC_max values (Additional file 2: Table S3). Overall ADHD shared the highest proportion of its trait-influencing variants with intelligence at 99.2% (7,708/7,774) compared to its subtypes, with childhood ADHD at 99.1% (7,758/7,825), late-diagnosed ADHD at 97.1% (8,342/8,590), and persistent ADHD at 21.3% (4,397/20,625) (Fig. 2D, Additional file 2: Table S3). Conversely, intelligence shared the highest proportion of its trait-influencing variants with late-diagnosed ADHD (82.3%, 8,342/10,135), followed by childhood ADHD (76.6%, 7,758/10,135) and overall ADHD (76.1%, 7,708/10,135), while persistent ADHD had the lowest overlap (43.4%, 4,397/10,135) (Fig. 2D, Additional file 2: Table S3). The Dice coefficient, which quantifies the proportion of shared variants relative to the total number of variants, was 86.1% for overall ADHD. Among the ADHD subtypes, late-diagnosed ADHD had the highest Dice coefficient of 88.7%, childhood ADHD followed with 86.3%, and persistent ADHD exhibited the lowest with 48.7% (Fig. 2E, Additional file 2: Table S3). The proportion of shared variants with concordant effects on ADHD and intelligence ranged from 14.7% for persistent ADHD to 35.3% for childhood ADHD (Fig. 2E, Additional file 2: Table S3).

Local genetic correlation (LAVA)

Using LAVA, we detected 75 genomic regions displaying significant local genetic correlations between ADHD and intelligence across all ADHD-intelligence pairs after BH-FDR correction. Of these, 72 regions exhibited negative correlations (r_g ranged from -0.383 to -1) and 3 showed positive correlations (r_g ranged from 0.609 to 1) (Additional file 2: Table S4, Additional file 3: Fig. S6). Specifically, 52 negative correlations were observed for overall ADHD, 10 for childhood ADHD, 6 for late-diagnosed ADHD, and 4 for persistent ADHD. Positive correlations were detected in 2 regions for overall ADHD and 1 region for childhood ADHD.

Shared genetic variants (cond/conjFDR)

To identify genetic variants associated with ADHD traits and intelligence, as well as their shared genetic architecture, we conducted cond/conjFDR analyses. The conditional Q-Q plots demonstrated substantial enrichment of ADHD-associated SNPs when conditioned on intelligence, and vice versa (Additional file 3: Fig. S7). At condFDR < 0.01, loci were identified separately for overall ADHD and each ADHD subtype conditioned on intelligence. In total, 133 loci were detected across all analyses, including 97 for overall ADHD, 23 for childhood ADHD, 12 for late-diagnosed ADHD, and 1 for persistent ADHD (Additional file 2: Table S5, Additional file 3: Fig. S8). To define independent genomic regions, overlapping loci across all ADHD traits were merged, resulting in 108 distinct loci, including 42 novel ones that had not been previously linked to ADHD traits (Additional file 2: Table S6). Subtype specificity was defined based on this merged set of loci to avoid redundancy arising from LD among lead SNPs. Among the 108 distinct loci, 77 were unique to overall ADHD (including 31 novel), 6 were unique to childhood ADHD (including 3 novel), 4 were unique to late-diagnosed ADHD (including 3 novel), and 1 was unique to persistent ADHD (Additional file 2: Table S6, Additional file 3: Fig. S9). Twelve loci were shared exclusively between overall ADHD and childhood ADHD, three between overall ADHD and late-diagnosed ADHD, and five were shared among overall ADHD, childhood ADHD, and late-diagnosed ADHD (Additional file 2: Table S6, Additional file 3: Fig. S9). No loci were shared across all four ADHD-intelligence analyses. Conversely, conditioning on all ADHD traits revealed 348 distinct loci associated with intelligence, including 49 novel loci (Additional file 2: Table S7, Additional file 3: Fig. S10).

At conjFDR < 0.05, we identified 242 loci jointly associated with ADHD traits and intelligence across all pairwise analyses (Fig. 3A-C, Additional file 2: Table S8). Of these, 158 were detected in the overall ADHD-intelligence comparison, 63 in childhood ADHD-intelligence, and 21 in late-diagnosed ADHD-intelligence, whereas no shared signals were identified for persistent ADHD (Fig. 3A, C). The most significant locus across all ADHD-intelligence comparisons was identified in the overall ADHD-intelligence analysis (chr5:87,712,913−88,253,916, top lead SNP: rs797419, conjFDR = 2.20E-6, Additional file 3: Figs. S11-14) and was also detected in both childhood and late-diagnosed ADHD analyses. Effect direction analyses revealed that most lead SNPs shared with intelligence had opposite effects, including 89.2% (141/158) for overall ADHD, 84.1% (53/63) for childhood ADHD, and 90.5% (19/21) for late-diagnosed ADHD (Additional file 2: Table S8).

Fig. 3.

Genomic loci shared between ADHD and intelligence identified by conjFDR analysis. (A) Network diagram illustrating the 242 top lead SNPs within the shared loci between overall ADHD, its subtypes, and intelligence. The size of the outermost points represents the -log₁₀-transformed conjFDR values. (B) Chromosomal distribution of 184 distinct loci shared between ADHD traits and intelligence, displaying loci associated with overall ADHD (red), childhood ADHD (yellow), late-diagnosed ADHD (blue), and overlapping loci shared across ADHD traits. Paired dots indicate loci shared between traits, while non-paired dots represent loci unique to each trait. Each colored circle corresponds to the top lead SNP within each locus, mapped across chromosomes 1–22. (C) Manhattan plot showing the variant-level genetic associations between ADHD and intelligence. Each point represents an SNP, with larger points outlined in black indicating top lead SNPs. The y-axis displays -log₁₀-transformed conjFDR values for each SNP, and the x-axis shows the chromosomal position. The dashed line represents the conjFDR threshold of 0.05, and the color of the dots indicates different ADHD traits. (D) LocusZoom plots of a genomic locus (chr14:47,249,450−47,314,938, top lead SNP: rs11844549), exclusively shared between overall ADHD and late-diagnosed ADHD, revealing divergence in the shared genetic architecture between ADHD subtypes and intelligence. These plots show the regional SNP associations (-log10(p)) for ADHD traits and intelligence from the original GWAS summary data. (E) LocusZoom plots of a genomic locus (chr2:233,557,043–233,727,071, top lead SNP: rs2197563), overlapping across all three ADHD-intelligence pairs, revealing convergence in the shared genetic architecture between ADHD subtypes and intelligence. These plots show the regional SNP associations (-log₁₀(p)) for ADHD traits and intelligence from the original GWAS summary data. Abbreviations: ADHD, attention-deficit/hyperactivity disorder; chr, chromosome; conjFDR, conjunctional false discovery rate; GWAS, genome-wide association study; SNP, single nucleotide polymorphism

After merging overlapping loci across ADHD traits, 184 distinct genomic loci were identified (Fig. 3B, Additional file 2: Table S9, Additional file 3: Fig. S15). Among the 184 distinct loci, 112 were unique to overall ADHD, 21 were unique to childhood ADHD, and 5 were unique to late-diagnosed ADHD (Fig. 3B, Additional file 2: Table S9, Additional file 3: Fig. S15). Thirty loci were shared exclusively between overall ADHD and childhood ADHD, 5 between overall ADHD and late-diagnosed ADHD (Fig. 3D, Additional file 2: Table S9, Additional file 3: Fig. S15), and 11 loci overlapped across all three ADHD-intelligence comparisons (Fig. 3E, Additional file 2: Table S9, Additional file 3: Fig. S15). No loci were exclusively shared between childhood ADHD and late-diagnosed ADHD. Additionally, across the 184 merged loci, 137 were novel for ADHD traits, 75 for intelligence, and 64 for both traits (Additional file 2: Table S9).

Functional annotation and enrichment analyses

A total of 8,579, 2,884, and 1,424 candidate SNPs were identified as shared with intelligence for overall ADHD, childhood ADHD, and late-diagnosed ADHD, respectively. The functional annotations of these candidate SNPs are presented in Fig. 4A and Additional file 2: Tables S10-12. The majority of these SNPs were located in intronic regions, with proportions of 52.5% (4,507/8,579) for overall ADHD, 66.7% (1,924/2,884) for childhood ADHD, and 60.3% (858/1,424) for late-diagnosed ADHD. Among the candidate SNPs, 4.2% (358/8,579) for overall ADHD, 5.0% (143/2,884) for childhood ADHD, and 3.9% (56/1,424) for late-diagnosed ADHD were identified as pathogenic based on CADD scores exceeding 12.37. Approximately 1.8% (153/8,579), 1.3% (37/2,884), and 2.5% (36/1,424) of the candidate SNPs for overall ADHD, childhood ADHD, and late-diagnosed ADHD, respectively, were predicted to have regulatory potential as suggested by a RegulomeDB score less than 2. Moreover, a chromatin state score less than 8 indicated that 82.5% (7,079/8,579) of candidate SNPs for overall ADHD, 86.2% (2,485/2,884) for childhood ADHD, and 81.5% (1,160/1,424) for late-diagnosed ADHD were located in open chromatin regions. Additionally, within the candidate SNPs, we identified three non-synonymous exonic variants (rs586339, rs1801251, and rs3816334) shared across overall ADHD, childhood ADHD, and late-diagnosed ADHD. Among the 184 lead SNPs in distinct shared loci, two—rs78648104 in TFAP2D and rs118134876 in AC004538.3:THSD7A—were non-synonymous mutations with CADD values of 24.3 and 23.8, respectively (Additional file 2: Table S9).

Fig. 4.

Functional annotation and enrichment analysis of shared loci between ADHD and intelligence. (A) Distribution of functional category, CADD scores, RegulomeDB scores, and minimum chromatin states for all candidate SNPs within the shared loci associated with ADHD and intelligence. The color gradient from dark to light represents overall ADHD, childhood ADHD, and late-diagnosed ADHD sequentially. The y-axis represents the number of candidate SNPs (log₁₀ scale) in each category, and the x-axis shows functional annotation categories used for SNP classification. (B) The bubble chart shows the representative subtype-specific enriched GO terms for biological processes of shared genes. The x-axis represents -log₁₀-transformed FDR q-values, and the y-axis displays the associated biological processes. The size of each bubble corresponds to the number of genes enriched in each biological process. The color intensity of the bubbles indicates the significance level. Abbreviations: ADHD, attention-deficit/hyperactivity disorder. CADD, Combined Annotation Dependent Depletion; FDR, false discovery rate; GO, gene ontology; SNP, single nucleotide polymorphism

Using the gene mapping strategy, candidate SNPs from loci shared between intelligence and each ADHD trait, including overall ADHD, childhood ADHD, and late-diagnosed ADHD, were mapped to 1,041, 397, and 192 protein-coding genes, respectively (Additional file 2: Tables S13-15, Additional file 3: Figs. S16-18). Overall ADHD-associated genes (N = 1,041) exhibited significant enrichment across 171 GO terms, covering a broad range of biological processes, predominantly related to cellular response to stimulus (BH-FDR q = 6.17E-5), cell population proliferation (BH-FDR q = 5.78E-4), B cell proliferation (BH-FDR q = 2.89E-3), regulation of protein metabolic process (BH-FDR q = 5.32E-3), and regulation of immune response (BH-FDR q = 1.25E-2) (Fig. 4B, Additional file 2: Table S16). Childhood ADHD-associated genes (N = 397) were significantly enriched for 36 GO biological processes, primarily involving early developmental processes such as regulation of neuron projection development (BH-FDR q = 3.69E-2), muscle structure development (BH-FDR q = 3.69E-2), synaptic vesicle localization (BH-FDR q = 3.81E-2), muscle cell differentiation (BH-FDR q = 3.81E-2), and regulation of neuron differentiation (BH-FDR q = 4.35E-2), highlighting their involvement in early neuronal and physical development (Fig. 4B, Additional file 2: Table S17). Late-diagnosed ADHD-associated genes (N = 192) displayed enrichment across 28 GO biological processes, particularly associated with synapse assembly (BH-FDR q = 1.10E-2), acyl carnitine transmembrane transport (BH-FDR q = 1.86E-2), dorsal/ventral axon guidance (BH-FDR q = 3.14E-2), modulation of chemical synaptic transmission (BH-FDR q = 4.09E-2), and regulation of trans-synaptic signaling (BH-FDR q = 4.09E-2), suggesting their role in the maturation and stability of synaptic functions and neuronal signaling (Fig. 4B, Additional file 2: Table S18).

Bidirectional genetic association (MR)

To assess the directional genetic associations between overall ADHD, its subtypes, and intelligence, we performed bidirectional two-sample MR analysis. For the genetic impact of ADHD traits on intelligence (Fig. 5, Additional file 2: Table S19, Additional file 3: Figs. S19-20), the overall ADHD model showed a negative association using the IVW estimator (beta = -0.104, 95% CI: -0.130 to -0.078, p = 2.70E-15, BH-FDR q = 1.35E-14), with broadly consistent estimates across complementary MR methods. For childhood ADHD, negative effects were observed using MR-RAPS (beta = -0.129, 95% CI: -0.184 to -0.075, p = 3.38E-06, BH-FDR q = 1.01E-05) and the maximum likelihood method (beta = -0.129, 95% CI: -0.186 to -0.073, p = 7.27E-06, BH-FDR q = 1.09E-05), whereas IVW estimates were not statistically significant. No significant effect was detected for late-diagnosed ADHD, and persistent ADHD could not be evaluated due to the lack of available IVs. For the reverse genetic effect of intelligence on ADHD traits (Fig. 5, Additional file 2: Table S20, Additional file 3: Figs. S21-25), genetically predicted intelligence showed inverse associations with ADHD-related outcomes across multiple MR estimators. For overall ADHD, effect estimates were consistently negative across methods, with the IVW model yielding an odds ratio (OR) of 0.695 (95% CI: 0.632 to 0.765, p = 1.13E-13, BH-FDR q = 2.82E-13). Similar inverse associations were observed for childhood ADHD (IVW OR = 0.587, 95% CI: 0.505 to 0.683, p = 4.16E-12, BH-FDR q = 7.75E-12) and late-diagnosed ADHD (IVW OR = 0.583, 95% CI: 0.480 to 0.707, p = 4.34E-8, BH-FDR q = 1.23E-7), with concordant effect directions across complementary MR approaches. For persistent ADHD, effect estimates were directionally consistent across methods; however, statistical support was weaker and not uniformly observed across all estimators, with the IVW estimate reaching statistical significance (IVW OR = 0.613, 95% CI: 0.416 to 0.902, p = 1.31E-02, BH-FDR q = 2.32E-02), whereas median-based approaches were not statistically significant.

Fig. 5.

Significant bidirectional MR associations between ADHD and intelligence. The effect estimates represent the beta coefficients for the directional effect of ADHD traits on intelligence and the OR for the directional effect of intelligence on ADHD traits. The error bars represent the 95% confidence intervals for the effect estimate. Abbreviations: ADHD, attention-deficit/hyperactivity disorder; CI, confidence interval; MR, Mendelian randomization; OR, odd ratio; SNP, single nucleotide polymorphism

Sensitivity analyses were conducted to evaluate the robustness of the MR findings (Additional file 2: Table S21). Evidence of heterogeneity was observed for the analysis of childhood ADHD on intelligence, as indicated by a significant Cochran’s Q test (p = 1.62E-03). For the remaining exposure-outcome pairs, Cochran’s Q tests did not indicate significant heterogeneity. MR-Egger intercept tests, performed for analyses with more than two IVs, yielded non-significant results, suggesting no evidence of directional horizontal pleiotropy. Leave-one-out analysis for these models further confirmed that no single SNP exerted disproportionate influence on the directional effects (Additional file 3: Figs. S20, 22–25).

Validation of shared genetic signals

The directional consistency of shared genomic loci identified by the conjFDR analysis was evaluated using additional GWAS datasets for ADHD and intelligence. For overall ADHD, 156 out of 158 top lead SNPs were available in the additional dataset, with all 156 showing consistent effect directions (BH-FDR q = 2.19E-47). For the ADHD subtypes, all top lead SNPs within the shared loci for childhood ADHD (63 SNPs) and late-diagnosed ADHD (21 SNPs) were present in the additional dataset. Among these, all SNPs for childhood ADHD (BH-FDR q = 1.45E-19) and 20 SNPs for late-diagnosed ADHD (BH-FDR q = 1.05E-5) exhibited consistent effect directions. For intelligence, 229 top lead SNPs were identified, all of which were present in the additional dataset, with 96.9% (222/229) showing consistent effect directions (BH-FDR q = 2.86E-56). These findings provide validation for the robustness of the identified shared loci.

Discussion

To our knowledge, this study provides the first comprehensive investigation of the shared genetic architecture between ADHD subtypes and intelligence using large-scale GWAS datasets. We observed consistent negative genetic correlations between ADHD phenotypes and intelligence, while the extent of shared genetic influences differed across ADHD subtypes. Childhood and late-diagnosed ADHD showed substantial polygenic overlap with intelligence, whereas persistent ADHD showed limited evidence of shared genetic architecture. In addition, we identified multiple pleiotropic loci and subtype-differentiated biological pathways linking ADHD liability with general cognitive ability. Together, these findings highlight the genetic heterogeneity of ADHD and suggest that the relationship between ADHD liability and general cognitive ability is not uniform across ADHD subtypes.

While previous studies have established a negative genetic correlation between overall ADHD and intelligence [25–27], our analysis expands on this finding by examining whether this relationship differs across ADHD subtypes. Using LDSC, we observed significant negative genetic correlations between overall ADHD, childhood ADHD, persistent ADHD, late-diagnosed ADHD, and intelligence. However, these findings should be interpreted cautiously. The intelligence phenotype used in this study reflects general cognitive ability [21], a latent factor capturing variance shared across multiple neurocognitive tasks. Within this framework, the observed negative genetic correlations reflect shared genetic influences between ADHD liability and general cognitive ability at the population level, rather than a deterministic relationship at the individual level. These correlations suggest that shared genetic factors may contribute to both traits, but in opposite directions. Accordingly, these findings should not be interpreted as evidence that ADHD universally implies lower intelligence. The strength of the correlation varied from -0.3442 for childhood ADHD to -0.4205 for late-diagnosed ADHD, suggesting that the relationship between ADHD liability and general cognitive ability may differ across subtypes. At the same time, these estimates may also be influenced by factors such as subgroup composition, diagnostic uncertainty, and the presence of common comorbidities, including learning difficulties and other neurodevelopmental or psychiatric conditions. In particular, inattentive symptoms may affect academic achievement and performance on cognitive tests without necessarily reflecting reduced underlying intellectual ability.

Although LDSC is effective for estimating genome-wide genetic correlation, it reflects the average magnitude of shared effects across the genome and therefore may not fully capture the extent of shared genetic architecture when signals are concentrated within specific loci [73–75]. In such cases, even when effect directions are largely consistent, the presence of strong effects in a subset of regions alongside weaker effects elsewhere may result in only modest genome-wide correlation estimates [76]. Consequently, a modest negative global genetic correlation does not preclude substantial overlap, including polygenic overlap and local genetic correlations. To further characterize this relationship, we applied MiXeR [33, 34], which estimates the number of shared variants irrespective of effect direction, and LAVA [35], which localizes shared genetic effects to specific genomic regions. MiXeR-based studies have previously estimated the genetic overlap between overall ADHD and intelligence at 66% [23], but by incorporating the updated largest GWAS summary data for overall ADHD, our study provides a refined estimate with a Dice coefficient of 86.1%. This enhanced overlap strengthens the evidence for shared genetic influences between ADHD and intelligence. Among ADHD subtypes, the late-diagnosed ADHD exhibited the highest Dice coefficient (88.7%), consistent with its strongest negative genetic correlation in our LDSC analysis, indicating a stronger genetic link between this subtype and cognitive ability. Model fit evaluation based on AIC supported the shared model for overall, childhood, and late-diagnosed ADHD, as indicated by positive AIC_min values, suggesting reliable estimation of polygenic overlap in these phenotypes. In contrast, persistent ADHD showed negative AIC_min and AIC_max values, suggesting limited statistical support for stable overlap estimation. Accordingly, the current data provide limited evidence for shared genetic architecture in persistent ADHD; however, this observation should be interpreted cautiously given weaker model support and the relatively small GWAS sample size. These findings underscore the importance of considering statistical power when interpreting subtype-specific overlap patterns. Importantly, the effect directions of shared variants were not uniform. Some shared variants showed opposing effects on ADHD liability and general cognitive ability, whereas others showed concordant effects, indicating a complex and multifaceted genetic relationship rather than a single, uniform pattern. This heterogeneity in effect direction may partly explain why substantial polygenic overlap can coexist with only modest genome-wide genetic correlations. LAVA provided a complement to these analyses by extending beyond genome-wide characterizations to identify specific genomic regions contributing to shared genetic architecture between traits. Unlike LDSC, which summarizes the overall correlation across the genome, LAVA localizes shared genetic effects to individual loci. The identification of 75 loci with significant local genetic correlations suggests that the shared genetic architecture between ADHD and intelligence is not uniformly distributed across the genome but is concentrated within specific genomic regions. This regional perspective further explains how substantial shared genetic architecture can be present despite only modest global genetic correlations.

We next applied the cond/conjFDR framework [36, 37], which extends beyond global and local correlation approaches by leveraging cross-trait enrichment to enhance variant discovery. In this framework, condFDR improves power to identify variant associated with one trait by conditioning on association with the other trait, whereas conjFDR more specifically identifies pleiotropic variant jointly associated with both traits. Therefore, unlike LDSC, MiXeR, and LAVA, which quantify the overall extent and distribution of shared genetic architecture, the cond/conjFDR framework enables variant-level discovery by identifying additional trait-associated loci through condFDR and shared pleiotropic loci through conjFDR. In this study, conditioning on intelligence allowed us to uncover multiple novel loci linked to overall ADHD and its subtypes, providing a more refined genetic analysis compared to conventional GWAS approaches. Previous GWAS study of ADHD subtypes reported 4, 1, and 0 genome-wide significant loci for childhood ADHD, late-diagnosed ADHD, and persistent ADHD, respectively [6]. By leveraging cross-trait enrichment through condFDR, we identified several subtype-specific loci that were not detectable in conventional single-trait GWAS analyses, thereby substantially expanding the genetic landscape of ADHD heterogeneity. Notably, we identified three childhood ADHD-specific loci (mapped to AC080125.1, MED10, and GID4) and three late-diagnosed ADHD-specific loci (mapped to DAZL, LRRCC1, and CTD-2171N6.1), none of which had been previously associated with ADHD traits (Additional file 2: Table S6). In addition, 31 novel loci were identified exclusively for overall ADHD. These subtype-restricted signals suggest that distinct genetic components may contribute to different developmental presentations of ADHD, supporting the view that ADHD subtypes are not merely clinical variants of a single disorder but may partially reflect divergent neurodevelopmental trajectories. Childhood ADHD-specific loci may preferentially relate to early neurodevelopmental mechanisms, whereas loci specific to late-diagnosed ADHD may reflect genetic influences emerging later in development or interacting with cognitive and environmental factors across the lifespan.

Comparison of merged loci derived from the conjFDR analysis further highlighted differences in subtype specificity across ADHD traits. Among the 184 distinct shared loci, 112 were unique to overall ADHD, compared with 21 for childhood ADHD and 5 for late-diagnosed ADHD, indicating variability in the extent of detectable genetic sharing across ADHD presentations. While differences in GWAS sample size and statistical power may partly contribute to these patterns, the observed distribution also supports partially distinct genetic architectures across ADHD subtypes. Notably, 30 loci were shared exclusively between overall ADHD and childhood ADHD, whereas only 5 loci were shared between overall ADHD and late-diagnosed ADHD, suggesting relatively greater detectable overlap between overall and early-onset ADHD presentations. In addition, 11 loci were shared across overall, childhood, and late-diagnosed ADHD, while none were uniquely shared between childhood and late-diagnosed ADHD. Although the two subtypes share genetic components through overall ADHD, the absence of direct overlap may suggest partially distinct neurodevelopmental processes or genetic risk factors emerging at different developmental stages. These findings may point to developmental timing as a critical factor in ADHD pathophysiology, with genetic influences on early-onset ADHD potentially distinct from those contributing to ADHD diagnosed later in life.

Among the 242 loci shared between ADHD traits and intelligence, the most significant shared locus, located on chromosome 5 (chr5:87,712,913−88,253,916) with the lead SNP rs797419 (conjFDR = 2.20E-6), was detected across overall, childhood, and late-diagnosed ADHD analyses. This locus is located within the intronic region of the MEF2C gene, a critical regulator of neural development, influencing synaptic plasticity and neuronal connectivity [77–79]. Extensive research has implicated MEF2C in several neurological disorders, including ADHD, and has highlighted its role in cognitive functions [80–83]. Our findings suggest that MEF2C may act as a genetic bridge between ADHD and intelligence, indicating that MEF2C may represent a shared genetic contributor to both traits. Among the 11 loci shared across all three ADHD-intelligence comparisons, one locus (chr2:233,557,043–233,727,071), a novel finding for both ADHD and intelligence, is located within the GIGYF2 gene, which encodes a protein critical for regulating cell growth, differentiation, and signaling pathways, playing a pivotal role in neuronal development [84]. Previous studies have suggested that GIGYF2 is involved in cognitive functions and is linked to several neurodevelopmental disorders, including ADHD [27, 85, 86]. Additionally, two non-synonymous mutations were detected among the distinct shared loci—rs78648104 in TFAP2D and rs118134876 in AC004538.3:THSD7A—which resulted in amino acid substitutions that may potentially disrupt the protein’s structure or function. With CADD scores of 24.3 and 23.8, respectively, these variants fall within the higher range of predicted deleteriousness. Nevertheless, as CADD scores reflect in silico functional predictions, they should be interpreted with caution in the context of complex polygenic traits. Furthermore, effect direction analysis revealed that the majority of shared loci exhibited opposing effects on ADHD and intelligence, implying that genetic susceptibility underlying ADHD risk might also contribute to cognitive difficulties [25–27]. Nonetheless, a few loci may influence ADHD and intelligence in a complementary manner, with concordant effects on both traits, highlighting the complex interplay between these traits at the genetic level [87, 88]. The SNP sign test showed a high level of agreement in effect directions between the discovery and additional datasets, validating the robustness of these findings.

Enrichment analyses of genes mapped from loci jointly associated with ADHD traits and intelligence indicate that the shared genetic architecture is biologically structured rather than confined to a single functional domain. For overall ADHD, enriched pathways converge on broad cellular regulatory processes, including stimulus response, proliferative mechanisms, and metabolic and immune-related pathways. This pattern suggests that the genetic overlap with intelligence may extend beyond canonical neurodevelopmental pathways to involve more systemic cellular and immune-regulatory mechanisms, aligning with accumulating evidence implicating immune and metabolic processes in ADHD pathophysiology [89, 90]. Subtype-level analyses further reveal biologically differentiated patterns consistent with developmental timing. Genes linked to childhood ADHD preferentially map to processes governing early neurodevelopment and structural growth, supporting the view that genetic influences contributing to cognitive variability in this subtype may operate during early developmental windows [91]. In comparison, late-diagnosed ADHD is characterized by enrichment in synapse-related and neuronal signaling processes, implying a greater contribution of mechanisms involved in synaptic organization and circuit-level functional modulation [92]. These findings suggest that while shared loci between ADHD and intelligence implicate common neurobiological themes, the underlying biological routes vary across subtypes. Rather than reflecting a uniform pleiotropic mechanism, the genetic overlap appears to manifest through developmentally stratified pathways, reinforcing the molecular heterogeneity underlying distinct ADHD trajectories.

The forward MR analysis provides evidence for a bidirectional genetic link between ADHD and intelligence. A significant negative genetic effect of overall ADHD on intelligence was observed, suggesting that greater genetic liability to ADHD may be associated with lower general cognitive ability. However, the relationship between childhood ADHD and intelligence showed significant heterogeneity, with some methods revealing a directional genetic effect, while others did not, suggesting that this relationship may be more variable and method-dependent. No significant association was found for late-diagnosed ADHD, possibly owing to the limited sample size. The reverse MR analysis suggested that lower general cognitive ability may be associated with greater genetic liability to ADHD across overall ADHD, childhood ADHD, and late-diagnosed ADHD [93]. For persistent ADHD, significant associations were observed using the IVW, MR-RAPS, and maximum likelihood methods, although the evidence was less consistent than for the other ADHD phenotypes. These findings support the presence of genetically informed bidirectional relationships between ADHD and intelligence, consistent with partially shared neurodevelopmental mechanisms, while emphasizing that these results reflect associations at the level of genetic liability rather than direct developmental or individual-level causal effects.

The complementary application of MiXeR and cond/conjFDR provides distinct but convergent perspectives on the genetic overlap between ADHD and intelligence. MiXeR estimates polygenic overlap by including all genetic variants contributing to both traits, even those with smaller effects, providing a broader view of the genetic relationship between ADHD and intelligence. In contrast, cond/conjFDR focuses on identifying loci with statistically significant associations, prioritizing variants with larger or more detectable effects, resulting in fewer shared loci. These methods complement each other: MiXeR offers a comprehensive, genome-wide perspective on polygenic overlap, while cond/conjFDR highlights the most statistically significant loci. An important consideration in interpreting these findings relates to differences in the age composition of the GWAS datasets analyzed across traits. The intelligence GWAS predominantly reflects genetic effects estimated in adult populations, whereas ADHD subtypes—particularly childhood ADHD—are defined by early-life onset. Because GWAS summary statistics capture genetic liability aggregated across the lifespan rather than age-specific phenotypic expression, cross-trait genetic associations identified in this study should be interpreted within a lifelong liability framework. Accordingly, the observed relationships are best understood as reflecting shared genetic influences operating across developmental stages rather than direct temporal relationships at the individual level. In this context, cross-trait genetic associations should not be interpreted in terms of chronological or developmental direction at the individual level, but instead point to partially overlapping neurodevelopmental mechanisms or pleiotropic genetic architectures contributing to both traits.

Beyond these interpretational considerations, several limitations should be acknowledged. First, all the data used in this study were derived exclusively from individuals of European ancestry, which limits the generalizability of the findings to populations with different genetic backgrounds. Replicating these results in cohorts from diverse ethnic and genetic backgrounds is essential to assess whether the identified genetic overlaps between ADHD and intelligence are consistent across different populations. Second, the study focused on common genetic variants that are well-represented in GWAS datasets, which means that rare variants or structural variations potentially influencing ADHD and intelligence were not captured. As a result, the findings may only explain part of the genetic architecture of these traits, and future studies should include rare variant analysis for a more comprehensive understanding. Third, the results for persistent ADHD were less consistent across analyses, including cond/conjFDR and MR. This could be due to the smaller sample size for persistent ADHD, which may have limited statistical power to detect shared genetic variants and bidirectional relationships. Future studies with larger sample sizes are necessary to clarify these associations. Finally, while functional annotation provided insights into the biological relevance of shared SNPs, the precise biological mechanisms underlying ADHD and intelligence remain unclear. Further studies are needed to explore how these genetic variants contribute to the neurobiological pathways that influence these traits.

Conclusions

This study shows that the shared genetic architecture between ADHD and intelligence varies across ADHD subtypes. Although all ADHD phenotypes exhibited negative genetic correlations with intelligence, the extent of genetic overlap differed across subtypes, with childhood and late-diagnosed ADHD showing greater shared genetic influences with intelligence than the persistent subtype. These findings highlight the genetic heterogeneity of ADHD and suggest that the relationship between ADHD liability and general cognitive ability is not uniform across subtypes. Understanding subtype-specific genetic architecture may help clarify ADHD heterogeneity and inform future research on risk stratification and early identification.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADHD

attention-deficit/hyperactivity disorder

AIC

Akaike Information Criterion

BH-FDR

Benjamini-Hochberg false discovery rate

CADD

Combined Annotation Dependent Depletion

chr

chromosome

condFDR/conjFDR

conditional and conjunctional false discovery rate

CTG

Complex Trait Genetics

eQTL

expression quantitative trait loci

EUR

European

FDR

false discovery rate

FUMA

Functional Mapping and Annotation

GO

Gene Ontology

GWAS

genome-wide association study

IVs

instrumental variables

IVW

inverse-variance weighted

LAVA

local analysis of [co]variant association

LD

linkage disequilibrium

LDSC

linkage disequilibrium score regression

MAF

minor allele frequency

MHC

major histocompatibility complex

MiXeR

bivariate causal mixture model

MR

Mendelian randomization

OR

odds ratio

PGC

Psychiatric Genomics Consortium

Q-Q

quantile-quantile

SNP

single nucleotide polymorphism

Author contributions

QZ1 (Qiyu Zhao) and FL conceived and designed the research. KY, ML1 (Mengge Liu), YG, QZ2 (Quan Zhang), YZ2 (Yang Zheng), and FL supervised research. QZ1 (Qiyu Zhao) acquired the data and performed the primary analyses. QZ1 (Qiyu Zhao), ZS, JZ, YW, QW, YZ1 (Ying Zhai), JX, ZZ, and ML2 (Minghuan Lei) interpreted the data. QZ1 (Qiyu Zhao) and FL drafted the article and made revisions to the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the China Postdoctoral Science Foundation (2025M782117), the Shandong Postdoctoral Science Foundation (SDZZ-ZR-202501075), the Tianjin Major Special Project on Public Health Science and Technology (24ZXGQSY00060), the Liaoning Province Science and Technology Plan Joint Program (2024-MSLH-565), and the Tianjin Key Medical Discipline Construction Project (TJYXZDXK-3–008 C).

Data availability

All datasets used in this study are publicly available. GWAS summary statistics for overall ADHD were obtained from the large-scale meta-analysis conducted by the Psychiatric Genomics Consortium (PGC) [20] and are available on the PGC website (https://www.med.unc.edu/pgc/download-results/; dataset https://doi.org/10.6084/m9.figshare.22564390). GWAS summary statistics for childhood ADHD, persistent ADHD, and late-diagnosed ADHD were obtained from the iPSYCH consortium [6] and can be accessed at the iPSYCH website (https://ipsych.dk/en/research/downloads/). GWAS summary statistics for intelligence were obtained from the Complex Trait Genetics (CTG) Lab [21] and are accessible via the CTG website (https://cncr.nl/research/summary_statistics/).

Code Availability

The code used in the present study is available in public repositories: LDSC (https://github.com/bulik/ldsc), MiXeR (https://github.com/precimed/mixer), LAVA (https://github.com/josefin-werme/LAVA), cond/conjFDR (https://github.com/precimed/pleiofdr), FUMA (https://fuma.ctglab.nl/), g: Profiler (https://biit.cs.ut.ee/gprofiler/gost), and Mendelian randomization (https://mrcieu.github.io/TwoSampleMR/). The custom code that supports the findings of this study is available at https://github.com/xiaoyu12-genetics/shared-genetics-ADHD-INT.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

Feng Liu is a member of the Editorial Board of BMC Medicine but was not involved in the editorial handling or decision-making process for this manuscript. The other authors declare that they have no competing interests.