Shared biological pathways linking ADHD and cortisol variability are related to externalizing behaviors

Abstract

Altered cortisol regulation is implicated in Attention-Deficit/Hyperactivity Disorder (ADHD), but causality remains debated. While meta-analyses suggest that lower basal cortisol levels, especially in the morning, correlate with ADHD, study heterogeneity prompts further inquiry. Leveraging post-genome-wide association approaches, we examined morning cortisol levels (n = 25,314) and ADHD (n = 225,543). We employed seven Mendelian Randomization methods for causal inference, assessed global and local genetic correlations, and applied the conjFDR method. Additionally, we conducted polygenic score (PGS) analysis in an independent ADHD sample (n = 1,660). We found evidence of regional pleiotropy rather than causality, indicating shared genomic architecture. Morning cortisol levels and ADHD showed local genetic correlations in two regions—on chromosomes 5 and 22—one with variants correlating positively and the other negatively. These regions harbor genes associated with psychiatric disorders, including two previously linked to ADHD (RASGRF2 and TRIOBP). ConjFDR revealed one independent hit (rs28406364) jointly associated with ADHD and cortisol levels. Our PGS analysis linked cortisol PGS to externalizing behavior comorbidity only in the European ancestry group. Adjusting for psychiatric disorders, we found an association with ADHD, with cases exhibiting lower cortisol PGS. Our findings suggest ADHD and cortisol levels reflect a complex interplay involving arousal regulation rather than a simple stress-response mechanism. The data align with broader physiological models, supporting a U-shaped relationship between cortisol levels and ADHD traits. This conceptual shift has significant implications for understanding ADHD as systemic dysregulation rather than solely a cognitive or attentional disorder.

Graphical Abstract

INTRODUCTION

Attention-Deficit/Hyperactivity Disorder (ADHD) is a common and complex neurodevelopmental condition (Breda et al., 2021; Vitola et al., 2017). It arises from a multifactorial and highly polygenic etiology, resulting from a combination of alterations across multiple physiological systems (da Silva et al., 2023; Faraone et al., 2024; Koirala et al., 2024). The hypothalamic-pituitary-adrenal (HPA) axis and its primary effector, cortisol, are fundamental to the regulation of stress responses, arousal, and cognitive control, all of which are disrupted in ADHD (Saccaro et al., 2021). Arousal regulation is particularly relevant, as ADHD has been hypothesized to involve difficulties in maintaining optimal arousal levels, contributing to deficits in attention, impulsivity, and emotional instability (Isaac et al., 2024). Cortisol, as a key modulator of the HPA axis, follows a distinct diurnal rhythm, with peak levels occurring in the morning and a gradual decline throughout the day (O’Byrne et al., 2021). This rhythm is closely linked to arousal states and cognitive performance, with deviations potentially indicating altered stress reactivity or maladaptive physiological regulation in ADHD (Ramos-Quiroga et al., 2016).

In this context, a meta-analysis of 19 studies investigated the link between cortisol levels and ADHD (Chang et al., 2021). It found that youths with the disorder consistently displayed lower basal cortisol levels regardless of the time-point, as well as reduced cumulative cortisol levels throughout the day compared to neurotypical individuals. Moreover, morning cortisol was lower in cases with ADHD (Chang et al., 2021).

Although meta-analysis results place cortisol variability as a potential factor involved with ADHD, there is significant heterogeneity among the individual studies comprising these findings, highlighting the necessity for further investigations in this domain (Chang et al., 2021; Kamradt et al., 2018). Genomic studies offer a promising avenue to elucidate whether the associations between an outcome and an exposure factor are pleiotropic, causal, or driven by reverse causality (Ciochetti et al., 2023). In the case of ADHD and morning cortisol levels, a single-site study identified a shared genetic component for cortisol variability and ADHD in the SERPINA6/SERPINA1 region (Carpena et al., 2022). This locus contains the gene for corticosteroid-binding globulin (CBG), a protein responsible for transporting glucocorticoids in the blood and regulating their bioavailability in target tissues. Genetic variants in SERPINA6/SERPINA1 locus were found to be significant in two genome-wide association studies (GWAS) of morning plasma cortisol levels (Bolton et al., 2014; Crawford et al., 2021). Additionally, a Mendelian Randomization (MR) study further evaluated these genetic variants as instrumental variables, revealing no causal association of genetically predicted plasma morning cortisol levels on ADHD; however, a significant effect of ADHD on morning cortisol variability was observed (Jue et al., 2023).

However, with substantial increases in sample sizes of GWAS on morning plasma cortisol levels (Bolton et al., 2014; Crawford et al., 2021) and ADHD (Demontis et al., 2023, 2019), deeper examination of the topic through advanced post-GWAS approaches became possible (Ciochetti et al., 2023; Uffelmann et al., 2021). This exploration is a necessary step toward understanding the role of stress in ADHD and related phenotypes. Investigating diurnal cortisol levels under resting conditions allows for the characterization of baseline HPA activity without the confounding effects of acute stressors, offering insight into whether ADHD is associated with hypo- or hyperarousal patterns. Therefore, in this study, we dissect the biological connection between ADHD and morning plasma cortisol variability using various genomic techniques to scrutinize causality and pleiotropy at a local level.

METHODS

Figure 1 summarizes the study design and analytical workflow.

Figure 1 –

Study Design. Overview of the analytical framework used to investigate the genetic relationship between morning cortisol levels and ADHD. The study integrates two complementary data sources: (i) GWAS summary statistics and (ii) individual-level genotype and phenotype data from an independent cohort. The analysis begins with Mendelian Randomization (MR) to test for causal effects between cortisol and ADHD, under assumptions that minimize confounding. To explore the possibility of horizontal pleiotropy, Local Analysis of [co]Variant Association (LAVA) estimates regional genetic correlations across the genome. Next, Conjunctional False Discovery Rate (conjFDR) is used to identify shared loci jointly associated with both traits, enhancing discovery power by integrating both GWAS. Finally, Polygenic Score (PGS) analyses assess whether genome-wide genetic liability for cortisol levels is associated with individual differences in ADHD and related behavioral traits. This multi-layered approach allows complementary evaluation of causal inference, genetic correlation, locus-level overlap, and individual-level prediction.

Summary statistics

We utilized the most extensive GWAS meta-analysis to date on plasma cortisol levels. We selected morning cortisol as our primary metric based on both its biological and clinical relevance, as well as the availability of GWAS data with adequate statistical power. To date, there are no GWAS available for cortisol awakening response (CAR), and the only GWAS on diurnal area under the curve (AUC) (Valders et al., 2011) is not a meta-analysis, limiting its utility for robust post-GWAS analyses. In contrast, GWAS of morning cortisol variability have been conducted in larger cohorts, offering more reliable genetic signals. Morning cortisol also captures key features of HPA axis regulation and has been consistently linked to neuropsychiatric outcomes, supporting its relevance for our study.

The CORtisol NETwork (CORNET) consortium significantly expanded its GWAS on morning plasma cortisol from 12,597 (Bolton et al., 2014) to 25,314 (Crawford et al., 2021) participants. The raw cortisol levels reported in Crawford et al., 2021 are available in their Supplementary Table 1, with plasma cortisol ranging from 7 to 3641 nmol/L, depending on the cohort and sampling time. Cortisol levels were measured using immunoassay in blood samples collected between 07:00 and 11:00 hours in all cohorts, except for one cohort (TwinsUK), which measured cortisol using liquid chromatography-mass spectrometry. This meta-analysis included 17 European population-based cohorts: CROATIA-Vis (n = 886), CROATIA-Korcula (n = 897), CROATIA-Split (n = 493), ORCADES (n = 1974), Rotterdam Study (n = 2870), NFBC1966 (n = 1324), Helsinki Birth Cohort Study 1934–44 (n = 399), ALSPAC (n = 1487), PREVEND (n = 1151), PIVUS (n = 919), Raine Study (n = 860), ET2DS (n = 847), MrOS-Sweden (n = 969), KORA (n = 1651), TwinsUK (n = 5654), SHIP (n = 910), and VIKING (n = 2073). Exclusion criteria in each individual cohort included current glucocorticoid use, pregnancy, and breastfeeding. For twin participants, one twin from each pair was excluded. A primary GWAS analysis was conducted in each cohort using linear regression on z-scores of log-transformed morning plasma cortisol, adjusted for sex, age, and cohort-specific genetic ancestry, as well as smoking and body mass index. The results were then meta-analyzed, and the summary statistics were made publicly available.

The GWAS meta-analysis of ADHD combining European samples from the Danish iPSYCH, Iceland deCODE, and the Psychiatric Genomics Consortium (PGC ADHD) expanded from 20,183 individuals diagnosed with ADHD and 35,191 controls (Demontis et al., 2019) to 38,691 individuals with ADHD and 186,843 controls (Demontis et al., 2023). The summary statistics were subjected to quality control procedures, excluding single nucleotide variants (SNVs) with a minor allele frequency (MAF) ≤ 1%, duplicated or ambiguous SNVs, and variants located on sex chromosomes.

Causality approach

Two Sample MR

We analyzed the causal relationship between morning plasma cortisol and ADHD diagnosis using seven different MR approaches [1. MR-Egger, 2. Weighted Median, 3. Inverse Variance Weighted (IVW), 4. Simple Mode, 5. Weighted Mode, 6. MR-PRESSO - Mendelian Randomization Pleiotropy RESidual Sum and Outlier, and 7. MR-RAPS - Robust adjust profile score]. To select the instrumental variables (IV) we clumped the morning plasma cortisol summary statistics selecting SNPs reaching a less conservative genomic significance (P < 5e-06) with a 1000 kb window and a r² = 0.01, allowing for the inclusion of more IVs. Then, we harmonized the exposure beta and standard error coefficients with those calculated for the SNPs associated with the outcome (ADHD) to perform the MR. Ambiguous and palindromic SNPs were excluded. The same steps were also performed considering ADHD as the exposure. In this case, we were able to select SNPs reaching the traditional genome-wide significance (P < 5e-08). We utilized default parameters in the TwoSampleMR (Hemani et al., 2018), MR-RAPS (Zhao et al., 2020), and MR-PRESSO (Verbanck et al., 2018) R packages to conduct all analyses. The IVs selected are described in Supplementary Tables 1 and 2.

Sensitivity analyses

We evaluated the validity of the IVs utilized in the TwoSampleMR analysis by examining F-statistics and I² to scrutinize the potential violation of the assumption of no measurement error (NOME) in the SNP-exposure association. Subsequently, we conducted various sensitivity analyses to detect any breaches of MR assumptions, including Cochran’s Q Statistics, MR Egger regression, Steiger directionality test, Leave-one-out analysis, Funnel Plots, and MR-PRESSO. Our methodology adhered to the guidelines outlined in the STROBE-MR (Skrivankova et al., 2021) and recently updated recommendations (Burgess et al., 2023).

Pleiotropy approach

Global and local genetic correlations

The global genetic correlation between morning plasma cortisol and ADHD was calculated using Linkage Disequilibrium Score Regression (LDSC) (Bulik-Sullivan et al., 2015b, 2015a). Firstly, summary statistics were harmonized to contain approximately one million HapMap SNPs, excluding the major histocompatibility complex. LDSC is based on calculating a LD score that estimates the degree to which variants are associated with each other. Then, it performs a regression of the effect of SNPs on both traits in their LD score. After normalization by trait heritabilities, it generates a genetic correlation coefficient (rg), indicating the degree of shared variation in the genome underlying the pair of analyzed phenotypes. In other words, it indicates genetic sharing, measured from the degree of covariance between two phenotypes.

We used the Local Analysis of [co]Variant Association (LAVA) method to estimate local genetic correlations (local rgs) (Werme et al., 2022). This approach seeks to identify specific genomic regions involved in the global genetic correlation between morning plasma cortisol and ADHD. LAVA achieves local results by partitioning the genome into semi-independent LD blocks, called loci. These loci are determined using an algorithm that imposes a minimum size requirement of 1000 SNPs (after filtering for a minor allele frequency > 0.01 among all SNPs contained in the summary statistics), resulting in the creation of 2495 distinct loci (Werme et al., 2022). To quantify local rgs and eliminate unassociated loci, LAVA initiates the analysis by testing the local univariate association signal for each pair of phenotypes. Loci showing univariate associations with P < 0.00002 (0.05/2495) for both phenotypes are then subjected to bivariate local analysis. We deemed a local rg significant when the False Discovery Rate (FDR) P-value (Benjamini-Hochberg method) was less than 0.05.

Conjunctional False Discovery Rate

The Conjunctional False Discovery Rate (conjFDR) is an extension of the Conditional False Discovery Rate (condFDR) method (Smeland et al., 2020). The condFDR is a hypothesis-free strategy to discover more variants associated with a trait after initial GWAS. Under a Bayesian framework, condFDR proposes that SNPs showing effects on pleiotropic phenotypes are more likely to be truly associated with the target phenotypes. The method assesses the cumulative distribution function of nominal P-values and readjusts them for each SNP. The result is the posterior probability of the SNP’s association being null, given the observed P-value. Therefore, conjFDR is an extension of these premises, applied to two phenotypes simultaneously, used to enrich SNPs jointly associated between two traits. The conjFDR value is defined as the maximum between the two condFDR values; in this case, between plasma morning cortisol and ADHD. For each SNP, this value represents the probability that it is null in the first analyzed phenotype, given that the P-values for its association with both the first and second phenotype are as small (or smaller) than observed (Smeland et al., 2020).

Gene annotation

We used Genome Data Viewer (https://www.ncbi.nlm.nih.gov/genome/gdv/) to annotate the genes encompassing pertinent loci identified through LAVA. Subsequently, we assessed gene-phenotype associations using GWAS Catalog (https://www.ebi.ac.uk/gwas/) and the PheWAS tool within GWAS Atlas (https://atlas.ctglab.nl/PheWAS).

Analysis in an independent single-site case-control sample

Sample description

The sample comprised 665 adults diagnosed with ADHD from the adult division of the ADHD Outpatient Program (ProDAH-A) at Hospital de Clínicas de Porto Alegre (HCPA), alongside 995 blood donor controls from the same institution. Following the study’s public announcement in local media, patients self-referred to ProDAH-A, where an initial screening interview confirmed ADHD diagnosis before inclusion in the study. All participants were 18 years or older. Exclusion criteria encompassed evidence of clinically significant neurological diseases potentially impacting cognition (e.g., history of head trauma, epilepsy, or dementia) and an estimated intelligence quotient (IQ) below 70.

ADHD diagnosis adhered to the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) criteria from 2001 to 2012 (American Psychiatric Association, 1994) and DSM-5 criteria from 2013 onwards (American Psychiatric Association, 2013; Matte et al., 2015). Lifetime diagnoses of ADHD and oppositional defiant disorder (ODD) were conducted using the Portuguese version of the Kiddie Schedule for Affective Disorders and Schizophrenia (K-SADS-E) adapted for adults (Grevet et al., 2005). Antisocial personality disorder (ASPD) diagnosis was determined through the Brazilian version of the Mini-International Neuropsychiatric Interview (Amorim, 2000; Sheehan et al., 1998).

Other lifetime psychiatric comorbidities were assessed by trained psychiatrists using the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) from 2001 to 2012 (First, 1998), an adapted version of SCID from 2012 to 2015, and SCID-5 from 2015 onwards (First et al., 2015). The severity of ADHD symptoms was gauged using the 18 items Adult ADHD Self-Report Scale (ASRS-18) adapted to Portuguese (Mattos et al., 2006). Participants with ADHD were not taking psychostimulant medication at the time of the first assessment, when symptoms were evaluated and the diagnosis was established.

The control sample screened negative for ADHD, evaluated using the 6-item ASRS (Kessler et al., 2005). All participants were fully briefed on study protocol and procedures and provided signed informed consent, approved by the institutional review board of HCPA (IRB 0000921), in accordance with the Declaration of Helsinki and Brazilian regulations.

Genotyping, quality control, and imputation

In the Adult ADHD Porto Alegre cohort, the available GWAS data originate from 3 different waves of genotyping. The first sample batch was genotyped using PsychChip (Infinium PsychArray BeadChip; Illumina) and the remaining with Infinium Global Screening Array (GSA) BeadChip versions 1.0+MD and 3.0+MD. Pre imputation quality control (QC) and principal component analysis (PCA) were implemented in the PsychChip dataset separately and in the merged GSA batches using the RICOPILI pipeline (Lam et al., 2020) (//sites.google.com/a/broadinstitute.org/ricopili/home) following default parameters for inclusion of variants or individuals. TOPMed Imputation Server (Taliun et al., 2021) (https://imputation.biodatacatalyst.nhlbi.nih.gov/#!) was used for imputation, using Minimac4 and the TOPMed reference panel. The resulting imputed datasets underwent additional QC procedures using the following cut-offs to exclude SNPs or individuals: info < 0.8, individual and SNP call rate < 95%, MAF < 1%, HWE < 1e-06. The resulting dataset contains 419 ADHD cases, 463 controls, and 6,599,235 SNPs for PsychChip and 256 cases, 571 controls, and 7,213,318 SNPs for the merged GSA batches. These two datasets were combined using Plink2 software, ensuring to retain the only overlapping SNPs, with the same chromosomal location, and the same alleles referencing the same DNA strand. Additional QC and PCA steps were applied to the complete dataset, resulting in 665 ADHD cases, 995 controls, and 5,751,409 SNPs. Additional quality verification was conducted using the GWASinspector R package (Ani et al., 2021) (Supplementary Figure 1A and 1B).

Ancestry analysis

The ancestry distribution of the Adult ADHD Porto Alegre Cohort against 1000 genomes populations is presented in Supplementary Figure 2A. We further assessed the degree of admixture in this sample using Multidimensional Scaling (MDS) (Marees et al., 2018) and maximum likelihood estimation (ADMIXTURE) (Alexander et al., 2009). Thus, we used the values of MDS components 1 and 2 from the 1000 genomes European reference groups to classify the European individuals in our sample, while the remaining individuals were classified as admixed. As expected for a population from southern Brazil, the European component is quite prominent (Supplementary Figure 2B). On average, 84.5% of the ancestry is European, followed by 7.9% African and 7.6% Native American. The average European ancestry in the Euro-Brazilian group is 91%, and in the admixed group, 77%.

Polygenic scores calculation and analysis

A polygenic score (PGS) for morning plasma cortisol was calculated using the Crawford et al., (2021) GWAS as the reference sample. The PGS was computed in the Porto Alegre target sample using the continuous shrinkage method (Ge et al., 2019), which involves employing a high-dimensional Bayesian regression framework. This method takes into account the LD pattern of SNPs to perform “continuous pruning” of genetic variants, proving to be a robust methodology for various genetic architectures. Moreover, it offers computational advantages and allows for multivariate modeling of local LD patterns (Ni et al., 2021).

We investigated the correlation between the cortisol levels PGS (standardized into z-scores) and different outcomes in the ADHD sample. These included case-control status, comorbid profile, and measures of severity of inattention symptoms, hyperactivity/impulsivity, and total ADHD symptoms. These correlations were tested using logistic regression models for categorical outcomes (e.g., case-control status) and linear regression for continuous outcomes (e.g., severity of ADHD). The models included at least sex, age, and the first 5 principal components as covariates.

RESULTS

MR and sensitivity analyses

After conducting extensive MR methods and subsequent sensitivity analyses, we did not uncover any indication of a causal relationship between morning plasma cortisol levels and ADHD diagnosis, nor did we observe evidence supporting the reverse causality (Tables 1 and 2; Supplementary Figures 3 and 4). In the first Direction MR (from cortisol to ADHD), despite relaxing the GWAS threshold to 5e-06, the limited number of IVs analyzed (n SNPs = 7) poses a challenge for robust GWAS comparison. Moreover, the high level of heterogeneity revealed by Cochran’s Q test (Supplementary Table 3) suggests potential sources of bias, such as pleiotropy or alternative biological pathways. Additionally, while the reverse Direction MR did not exhibit clear evidence of pleiotropy or violations of other MR assumptions, it similarly failed to indicate a causal effect of genetically predicted ADHD on cortisol levels.

Table 1.

Two sample mendelian randomization results for model-based methods

Exposure	Outcome	Method	n SNP	b	SE	P
Plasma cortisol	ADHD	MR-Egger	7	−0.0013	0.1303	0.9926
		Weighted Median	7	0.0069	0.0072	0.3397
		IVW	7	−0.0697	0.0855	0.4148
		Simple Mode	7	−0.0260	0.0304	0.4266
		Weighted mode	7	0.0077	0.0067	0.2948
ADHD	Plasma cortisol
		MR-Egger	26	0.0823	0.1650	0.6226
		Weighted Median	26	−0.0632	0.0471	0.1798
		IVW	26	0.0010	0.0342	0.9776
		Simple Mode	26	−0.0783	0.0960	0.4226
		Weighted mode	26	−0.0806	0.0911	0.3844

IVW = Inverse Variance Weighted;

Table 2.

Two sample mendelian randomization results for outlier-corrected methods

Exposure	Outcome	Method	n SNPs	beta	SE	P
Plasma cortisol	ADHD	MR-RAPS	NA*	NA*	NA*	NA*
ADHD	Plasma cortisol		27	−0.0040	0.0333	0.9041
Exposure	Outcome	Method	n SNPs	beta corrected	SE corrected	P corrected
Plasma cortisol	ADHD	MR-PRESSO	4 (3 outliers)	−0.0991	0.0444	0.1118
ADHD	Plasma cortisol		NA**	NA**	NA**	NA**

RAPS = Robust adjust profile score;

PRESSO = Pleiotropy RESidual Sum and Outlier;

^*Not calculated due to small number of SNPs;

^**No outliers were detected.

Findings on global and local genetic correlations

There was no significant global genetic correlation between morning plasma cortisol and ADHD (rg = −0.13; P = 0.1946). However, in the local analysis, 36 loci achieved significance and proceeded to the final step, the bivariate local analysis (Table 3). Among these, ten exhibited P-values lower than 0.05. Notably, 40% of these displayed a direct relationship, while 60% showed an inverse relationship. This pattern possibly impacted the global correlation, contributing to a non-significant rg. Two loci survived FDR correction (loci 842 and 2482) and were selected for further examination in annotation analysis.

Table 3.

Summary of local correlations analysis results

Locus	Chromosome	Start	Stop	n SNPs	rho	rho lower	rho upper	P-value	FDR
84	1	96147577	97721186	3086	−0.36	−0.73	−0.03	0.0316	NS
344	2	157556097	159040979	2051	0.16	−0.21	0.57	0.3748	NS
409	2	233352723	234115092	1700	0.11	−0.32	0.58	0.5984	NS
441	3	20754843	21790867	3045	−0.08	−0.41	0.23	0.5947	NS
537	3	132124019	132925967	1786	0.26	−0.17	0.71	0.2157	NS
694	4	105319196	106479155	2242	0.39	0.02	0.84	0.0446	NS
782	5	5828695	6936733	3470	−0.18	−0.61	0.21	0.3343	NS
828	5	60930754	62180368	2111	−0.37	−0.83	0.02	0.0647	NS
842	5	80448226	81710194	2094	−0.56	−0.96	−0.22	0.0014	0.0342
920	5	173606996	174662885	2589	−0.48	−0.98	−0.09	0.0178	NS
931	6	5378214	6621937	3285	−0.35	−0.71	−0.03	0.0299	NS
1037	6	106053916	107309327	3139	0.11	−0.20	0.44	0.4695	NS
1088	6	164687463	165835953	3311	0.27	−0.04	0.60	0.0897	NS
1136	7	38966484	40261240	2429	0.32	−0.03	0.71	0.0742	NS
1351	8	125453323	126766827	2843	0.05	−0.37	0.49	0.8073	NS
1430	9	85135752	86769886	3398	0.23	−0.12	0.62	0.1892	NS
1592	10	118416050	119528730	1907	−0.24	−0.57	0.06	0.1246	NS
1597	10	123856185	124894142	2571	0.20	−0.10	0.51	0.1804	NS
1631	11	19780785	20686336	2338	−0.01	−0.35	0.33	0.9698	NS
1729	11	124403478	125266684	1868	−0.09	−0.46	0.28	0.6202	NS
1731	11	126304241	127306871	2530	−0.44	−0.80	−0.14	0.0068	NS
1744	12	3871115	4967643	2146	−0.19	−0.61	0.21	0.3319	NS
1940	13	101573737	103050417	3507	−0.10	−0.48	0.26	0.5838	NS
1955	14	20876081	21982654	2353	0.20	−0.11	0.54	0.2039	NS
2020	14	93386329	94892240	3712	0.41	0.12	0.77	0.0079	NS
2092	15	92675406	93707010	2760	−0.09	−0.51	0.32	0.6295	NS
2093	15	93707011	94565521	2459	−0.41	−0.83	−0.05	0.0235	NS
2198	17	31318754	32356061	2646	−0.04	−0.42	0.34	0.8420	NS
2251	18	12735559	13641479	1958	0.22	−0.18	0.67	0.2656	NS
2301	18	73568261	74471745	2478	−0.21	−0.64	0.18	0.2683	NS
2405	20	41351515	42185671	1690	−0.28	−0.61	0.03	0.0719	NS
2449	21	34383795	35729261	2561	0.34	0.01	0.71	0.0416	NS
2452	21	38046450	39252059	2714	−0.20	−0.58	0.15	0.2518	NS
2455	21	41344743	42187152	2319	−0.23	−0.58	0.09	0.1573	NS
2472	22	27192924	27952441	2355	−0.29	−0.67	0.05	0.0905	NS
2482	22	37364005	38718589	2608	0.47	0.19	0.79	0.0019	0.0342

NS = Not significant

Gene annotation

We detected a negative correlation between morning plasma cortisol levels and ADHD at locus 842 (rho = −0.56, CI95% = −0.22 – −0.96; P = 0.0014; P corrected by FDR = 0.0342). This locus is situated on chromosome 5 and spans twelve genes. Notably, nine out of the twelve genes have significant associations with a myriad of psychiatric-related phenotypes, documented in the GWAS Catalog and GWAS atlas (Supplementary Table 4). For instance, RASGRF2 has demonstrated involvement in various psychiatric traits and other phenotypes that exhibit correlations with ADHD, as well as hypothalamus-pituitary-adrenal (HPA) axis dysregulation, including substance use disorders, risk-taking behaviors, externalizing disorders, and response to methylphenidate treatment in ADHD patients. Another noteworthy gene, ATP6AP1L, boasts multiple associations, primarily with cognitive function, educational attainment, and neuroimaging measures such as cortical thickness and hippocampal volume.

At locus 2482, we detected a positive correlation (rho = 0.47, CI95% = 0.19 – 0.79; P = 0.0019; P corrected by FDR = 0.0342). This locus is on chromosome 22 and spans 52 genes, of which 15 are relevant for psychiatric conditions and related phenotypes according to the GWAS Catalog and GWAS atlas (Supplementary Table 5). TRIOBP is the most implicated gene for a wide range of phenotypes, including cognitive function, educational attainment-related phenotypes, brain measures, and ADHD.

ConjFDR results

Only one independent SNP (rs28406364) was found to be jointly associated with ADHD and morning plasma cortisol levels at conjFDR < 0.05 (Figure 2). This SNP is an intronic variant located within the ZNF652-AS1 gene (ZNF652 antisense RNA 1) on chromosome 17.

Figure 2 -.

Manhattan plot showing common genetic variants jointly associated with ADHD and morning plasma cortisol levels at conjunctional false discovery rate (conjFDR) < 0.05. The y-axis shows the −log10 transformed conjFDR. Chromosomal position is presented along the x-axis. The threshold for significant shared associations (conjFDR < 0.05) is represented by the horizontal dotted line.

PGS analysis in an independent sample

The PGS analysis results are consistent with pleiotropic effects. Initially, no significant association between the cortisol PGS and ADHD was observed (Table 4). However, nominal findings emerged where higher cortisol PGS was associated with increased odds for substance use disorders (SUD) (OR = 1.398; P = 0.021), oppositional defiant disorder (ODD) (OR = 1.308; P = 0.017), and conduct/antisocial personality disorder (TC/TPAS) (OR = 1.332; P = 0.05). Subsequently, we investigated whether the number of externalizing comorbidities is sensitive to increases in the cortisol PGS. Figure 3 depicts a significant (F = 5.299; df = 2; P = 0.005) linear association, as observed through post hoc Bonferroni analysis. Finally, we re-evaluated the regression model with ADHD as the outcome, incorporating psychiatric comorbidities as covariates, and observed a nominal and inverse association between cortisol PGS and ADHD (OR = 0.825; P = 0.043) consistent with previous meta-analyses on biomarker levels. As anticipated, the plasma cortisol PGS proved to be a predictor of behavioral outcomes only within the European ancestry subgroup. There was no significant association between the cortisol PGS and severity of ADHD (Supplementary Table 6).

Table 4.

Effects of the morning plasma cortisol levels polygenic score on Attention-Deficit Hyperactivity Disorder and its main comorbidities

	All								European
Outcome	N yes	N no	B	SE	Wald	P-value	OR (CI 95%)		N yes	N no	B	SE	Wald	P-value	OR (CI 95%)
Attention-deficit hyperactivity disorder	658	995	−0.016	0.052	0.093	0.760	0.984 (0.888–1.091)		358	580	−0.062	0.069	0.807	0.369*	0.940 (0.822–1.076)
Nicotine use disorder (ADHD case-only)	272	365	0.087	0.085	1.091	0.307	1.091 (0.923–1.289)		130	216	0.180	0.116	2.426	0.119	1.197 (0.955–1.501)
Substance use disorder (ADHD case-only)	129	509	0.183	0.107	2.916	0.088	1.201 (0.973–1.481)		68	276	0.335	0.145	5.332	0.021	1.398 (1.052–1.858)
Oppositional defiant disorder (ADHD case-only)	392	250	0.030	0.084	0.130	0.719	1.031 (0.875–1.214)		215	134	0.268	0.112	5.718	0.017	1.308 (1.050–1.629)
Major depressive disorder (ADHD case-only)	259	383	−0.088	0.085	1.057	0.304	0.916 (0.775–1.083)		144	202	−0.089	0.112	0.613	0.427	0.915 (0.735–1.139)
Generalized anxiety disorder (ADHD case-only)	155	487	0.005	0.097	0.062	0.962	1.005 (0.830–1.215)		82	265	0.252	0.131	3.693	0.055	1.286 (0.995–1.662)
Bipolar disorder (ADHD case-only)	105	529	0.099	0.113	0.763	0.382	1.104 (0.884–1.379)		52	291	0.280	0.154	3.308	0.069	1.323 (0.979–1.788)
Conduct/Antisocial personality disorder	129	486	0.181	0.108	2.78	0.095	1.198 (0.969–1.482)		69	265	0.287	0.146	3.847	0.050	1.332 (1.000–1.775)

Analyses were adjusted for age, sex, and first five principal components.

^*When adjusted for age, sex, first five principal components, substance use, oppositional defiant, generalized anxiety, bipolar, and conduct/antisocial personality disorders: P = 0.043, OR 0.825 (0.684–0.994), n yes = 330, n no = 576.

Figure 3 -.

The relationship between morning plasma cortisol levels polygenic score (PGS) and the occurrence of externalizing comorbidities, including substance use disorders, oppositional defiant disorder, conduct disorder, and antisocial personality disorder. The univariate generalized linear model shows significance for the PGS (F = 5.299; df = 2; P = 0.005). Asterisks denote Bonferroni post hoc comparisons (*P = 0.049; **P = 0.004). Sample sizes are as follows: None = 167; Only 1 = 141; 2 or 3 = 91. Sex, age, bipolar disorder, internalizing comorbidities, and the first five principal components were covariates. Only individuals with European ancestry were included in the analysis.

DISCUSSION

Our findings suggest that the relationship between ADHD and morning cortisol variability is better explained by localized pleiotropy rather than direct causality. We identified genomic regions exhibiting both positive and negative correlations between these traits, reinforcing the notion of a complex, multifaceted association. Furthermore, PGS analyses in an independent sample revealed associations between the genetic predisposition for high morning cortisol levels and comorbid externalizing behaviors, supporting previous findings that link cortisol and ADHD primarily through related behavioral pathways (Bernhard et al., 2021). Collectively, these results contribute to a refined understanding of the genetic influences on neuroendocrine regulation in ADHD, moving beyond a simple stress response model to one that implicates broader arousal mechanisms and behavioral regulation.

The MR results challenge previous reports suggesting a causal effect of ADHD on morning cortisol variability (Jue et al., 2023). Importantly, the latest ADHD GWAS meta-analysis has substantially increased the number of IVs, effectively doubling the available loci. In the context of MR methodology, a null result obtained with an expanded IV set strengthens the evidence against a causal relationship, even when pleiotropic effects remain a possibility (Burgess et al., 2023). Moreover, our findings align with those of Jue et al., (2023) who also failed to detect a causal effect of morning cortisol on ADHD. However, several factors warrant caution in the interpretation of these results: (1) the limited number of IVs, despite the expansion of the morning cortisol GWAS dataset, and (2) heterogeneity and sensitivity analyses, which suggest the presence of alternative biological pathways.

To further investigate the biological connection between ADHD and morning cortisol variability, we conducted global and local genetic correlation analyses. A key finding was the absence of a significant global correlation, despite the negative genetic correlation coefficient suggesting an inverse relationship between ADHD and cortisol levels—an association that would align with previous meta-analytic evidence (Chang et al., 2021). The lack of statistical significance likely reflects heterogeneous contributions from distinct genomic regions. Specifically, while some loci followed the expected inverse relationship (rho: −0.56 to −0.35), others exhibited a positive correlation (rho: 0.34 to 0.47), potentially diminishing the overall genome-wide genetic correlation signal. Notably, such bidirectional regional correlations can arise even when the global genetic correlation remains null (van Rheenen et al., 2019).

Given these findings, the localized pleiotropy observed in the LAVA analyses provides new hypotheses regarding the relationship between ADHD and cortisol regulation. An alternative interpretation of these findings is that the association between ADHD and HPA axis functioning may follow an inverse U-shaped curve, where both low and high cortisol levels are linked to ADHD symptomatology. This hypothesis aligns with physiological models of homeostasis, where optimal function occurs at an intermediate point along the inverted U-shaped curve. Deviations in either direction—whether hypo- or hyper-arousal—can lead to maladaptive physiological states, a phenomenon referred to as allostasis or, in its more dysregulated form, cacostasis (Chrousos, 2009; Stanojlović et al., 2022).

The implication of this model extends beyond ADHD to psychiatric conditions more broadly, reinforcing a conceptualization of these disorders as systemic, rather than purely neuropsychiatric, in nature (Finlay et al., 2022). Empirical data support this curvilinear association, with several studies demonstrating that moderate cortisol levels optimize cognitive performance, whereas both elevated and reduced levels impair memory function (Moriarty et al., 2014; Salehi et al., 2010; Schilling et al., 2013). Similar findings have been reported in the context of depressive symptoms, where hair cortisol levels exhibit an inverse U-shaped relationship with symptom severity (Ford et al., 2019).

We acknowledge that the inverse U-shaped hypothesis could be misinterpreted as suggesting a causal link between cortisol regulation and ADHD risk. Although this conceptual model is appealing, our MR findings do not support a causal relationship—that is, we found no evidence that genetically predicted levels of morning cortisol levels directly influence ADHD risk. In this context, a recent study demonstrated that the pleiotropic relationships between ADHD and pain-related traits involve both vertical pleiotropy (indicating potential causal pathways) and horizontal pleiotropy (reflecting correlated but non-causal associations) (Ciochetti et al., 2025). Nevertheless, it remains possible that causal mechanisms will emerge as additional loci associated with morning cortisol variability and ADHD are identified. As it stands, our findings are more consistent with horizontal pleiotropy, in which shared genetic variants influence ADHD and cortisol variability in non-causal pathways. Importantly, this model can still accommodate the curvilinear effects observed in phenotypic studies—particularly if the direction or magnitude of pleiotropic effects varies across loci, contributing to both ends of the cortisol spectrum. Given the current findings, we believe both interpretations remain plausible and not mutually exclusive, as the inverse U-shaped hypothesis, if confirmed, could represent a causal mechanism for other phenotypes related to ADHD—thus manifesting as horizontal pleiotropy when examining ADHD directly. Future studies should aim to disentangle these mechanisms more precisely.

Further supporting the pleiotropic nature of the ADHD-cortisol relationship, the two significant loci identified in LAVA contain genes implicated in a wide range of psychiatric and behavioral phenotypes. One notable example is RASGRF2, which encodes a calcium-regulated nucleotide exchange factor involved in neurogenesis in the dentate gyrus, a key hippocampal region (Amaral et al., 2007; Gómez et al., 2017). This gene has been associated with externalizing behaviors and risk-taking propensity (Baselmans et al., 2022; Linnér et al., 2021, 2019), substance use disorders (Brazel et al., 2019; Liu et al., 2019; Saunders et al., 2022), insomnia (Jansen et al., 2019), and cardiovascular responses to methylphenidate in ADHD (Mick et al., 2011). Similarly, TRIOBP, which regulates neural tissue development, has been linked to cognitive abilities, ADHD risk in cross-trait analyses, and substance use disorders (Davies et al., 2018; Demange et al., 2021; Rao et al., 2022). The involvement of ZNF652-AS1, identified through ConjFDR analysis, further reinforces the externalizing pathway hypothesis, as the SNP significant in our study (rs28406364) is a genome-wide hit for age of first sexual intercourse, a trait closely tied to externalizing behaviors from genome-wide analyses (Mills et al., 2021). Alternatively, the strongest association described for this SNP is with whole-body impedance (P = 1.4 × 10⁻²⁹; Watanabe et al., 2019). This may represent a pleiotropy effect, as both cortisol levels and ADHD have been linked to body fat and obesity (Chen et al., 2018; Ling et al., 2020; van der Valk et al., 2022).

The PGS analysis further supports this pleiotropic framework, with genetic predisposition for high cortisol levels being associated with CD, ODD, SUD, and ASPD. These findings are counterintuitive when we consider prior meta-analyses suggesting that ADHD is primarily linked to reduced cortisol levels through comorbid externalizing behaviors (Chang et al., 2021). Importantly, lifetime CD and ODD were associated with an increased PGS load for morning cortisol levels in our study, a pattern consistent with findings in adolescent samples. For instance, a study of 138 adolescents found that after adjusting for CD, anxiety, and depression, ADHD symptoms correlated with increased cumulative diurnal cortisol, morning cortisol, and afternoon cortisol, while ODD symptoms were linked to reduced cortisol levels (Berens et al., 2023). These complex findings underscore the need to account for both age and neuropsychiatric comorbidities when interpreting cortisol levels variability in ADHD. Moreover, sex differences and epigenetic influences may further contribute to the heterogeneity in observed HPA axis responses (Müller et al., 2021).

Several limitations of this study must be acknowledged. Although GWAS sample sizes have significantly expanded, the number of identified loci remains a small fraction of the SNP heritability for each trait, potentially limiting the power to detect global genetic correlations. Additionally, the power of LAVA analyses is dependent on sample size, meaning other genomic regions that exhibit ADHD-cortisol correlations may not have met the criteria for bivariate testing. The bidirectional nature of these regional correlations also presents challenges for PGS analyses, as it may obscure true associations.

Our clinical sample, while well-characterized, was modest in size, limiting the statistical power to examine psychiatric comorbidities. The absence of circadian hormonal measures in our dataset precludes direct validation of cortisol secretion patterns. Moreover, we recognize that sex may influence the relationship between variability in morning cortisol levels and ADHD. However, the GWAS meta-analyses for both phenotypes were not stratified by sex; instead, sex was included as a covariate in each cohort’s individual analysis prior to meta-analysis. Consequently, we were unable to conduct MR, LDSC, LAVA, or ConjFDR analyses separately by sex. In the independent cohort with individual-level data, sex was likewise included as a covariate in all analyses, as stratifying by sex would have reduced the sample sizes to levels insufficient for reliable statistical inference.

An additional limitation is the Eurocentric bias of the discovery GWAS used to construct the morning plasma cortisol PGS. The summary statistics were derived from 17 cohorts of exclusively European ancestry, which restricts the transferability of the PGS to more diverse populations. Given that PGS performance decreases with increasing genetic distance from the discovery population (Bruxel et al., 2025), it is likely that the cortisol PGS performs more accurately among individuals with higher European ancestry within our admixed Brazilian sample. This may explain why the observed association between cortisol PGS and externalizing behavior comorbidity was only detected in the European ancestry subgroup. These findings should not be interpreted as indicating a weaker cortisol–ADHD relationship in non-European populations, but rather as evidence of reduced predictive power in underrepresented groups due to current ancestry-biased GWAS resources.

Furthermore, heterogeneity in the mean and range of cortisol levels across individual cohorts included in the reference sample presents an additional consideration. While this heterogeneity might impact results, it would only pose a methodological concern if the genetic factors influencing morning cortisol variability differed significantly among cohorts. However, this was not the case. For instance, the SERPINA6/SERPINA1 locus was consistently associated with cortisol levels across all cohorts, with the direction and magnitude of genetic associations remaining consistent. This suggests that genetic variants at the associated loci influence cortisol levels similarly across populations.

Finally, the study relies heavily on genomic data, which, while informative, does not account for potential environmental and sociodemographic confounders that may also influence the regulation of morning cortisol levels and ADHD-related outcomes. This limitation is inherent to the study design and highlights the importance of integrating genetic, environmental, and multidimensional clinical data in subsequent research.

In conclusion, our findings challenge the proposed causal relationship between ADHD and morning plasma cortisol levels, instead supporting a model of localized horizontal pleiotropy. The genetic architecture of ADHD and morning cortisol regulation suggests a complex interplay involving arousal regulation, rather than a simple stress-response mechanism. Moreover, our findings align with broader physiological models, suggesting an inverse U-shaped relationship between cortisol levels and ADHD traits. This conceptual shift has important implications for understanding ADHD as part of a broader systemic dysregulation rather than a disorder solely of cognitive or attentional deficits.

Future research should aim to explore relationships between ADHD and morning cortisol in larger, diverse cohorts while integrating genetic, epigenetic, and hormonal data to refine our understanding of the neuroendocrine mechanisms underlying ADHD. In particular, cohorts that have both cortisol measurements and genome-wide genotype data offer valuable opportunities to examine whether ADHD PGS is associated with cortisol variability, potentially modulating individual differences in HPA axis functioning. We suggest that approaches modeling nonlinear genetic associations (e.g., PGSs interacting with quadratic cortisol terms), in combination with longitudinal phenotyping of HPA axis dynamics and ADHD symptom trajectories, could be particularly informative. Furthermore, ADHD-specific and transdiagnostic cohorts with multidimensional phenotyping—encompassing behavioral, neurocognitive, and psychiatric dimensions—are needed to disentangle indirect pathways linking cortisol regulation, ADHD, and co-occurring neuropsychiatric traits. MR analyses stratified by sets of IVs associated with baseline cortisol levels or stress reactivity profiles may also help distinguish causal from correlative mechanisms across the cortisol secretion spectrum. Longitudinal studies incorporating these factors could provide critical insights into how HPA axis dynamics contribute to ADHD symptom trajectories throughout the lifespan. While the current literature has often addressed these components in isolation, future efforts should prioritize integrative designs that combine biomarker profiling with fine-grained clinical and genetic data to reconstruct the complex neuroendocrine influences on ADHD.

Highlights.

• We used genomic approaches to dissect the link between morning cortisol variability and ADHD.

• The findings suggest localized pleiotropy rather than direct causality between ADHD and morning cortisol levels.

• Local genetic correlation analysis indicates both positive and negative correlations contribute to the ADHD-cortisol relationship.

• Polygenic score (PGS) analysis revealed associations between cortisol PGS and externalizing behaviors.

• A potential U-shaped physiological relationship between HPA axis functioning and ADHD is proposed.

Data availability

Due to IRB constraints, the data from the ADHD PoA cohort is not publicly available. However, the raw genotyped data can be accessed through requests to the ADHD working group of the Psychiatric Genomics Consortium or by contacting the senior authors of this study [DLR, CHDB, and EHG]. All summary statistics used in this study are publicly available at https://pgc.unc.edu/for-researchers/download-results/ and https://datashare.ed.ac.uk/handle/10283/3836.