To address these questions, we employed a two-sample MR design in our study, in which two types of data are obtained from two non-overlapping datasets for MR analysis (22). This approach was chosen due to the inherent difficulty in obtaining comprehensive data encompassing genetic variation, exposure, and the relationship between genetic variation and outcomes within the same sample. Adopting a two-sample MR design gives us several advantages over a single-sample design. Firstly, weak instrument bias in a two-sample design tends to favor the null hypothesis, making it more conservative and reducing the likelihood of false positive results compared to bias in the direction of correlation. Secondly, in scenarios where it is challenging to simultaneously measure exposure and outcome data within the same group of individuals, the two-sample MR approach significantly expands the scope of MR studies (23). According to previous research, the heritability of morning plasma cortisol ranges from 30 to 60% (24, 25). Identifying genetic variations that contribute to changes in morning cortisol values could provide critical insight into the mechanisms of HPA axis activation (26). The CORNET consortium aims to identify genetic determinants of inter-individual variation in HPA axis function. For example, the consortium has discovered two genes, SERPINA6 and SERPINA1, highly expressed in tissues involved in cortisol physiology. SERPINA6 encodes corticosteroid-binding globulin, a protein that binds cortisol in the bloodstream, while SERPINA1 encodes α1-antitrypsin, a protein that regulates the release of cortisol from corticosteroid-binding globulin by inhibiting the cleavage of the reactive center loop (27). In the forward direction, the CORNET consortium provided the GWAS associated with the levels of serum morning plasma cortisol as the sample exposure, which was obtained from a sample of 12,597 individuals of European ancestry (on average 53.5 years old, 59.2% female) and replicated in an independent sample of 2,795 participants. Morning plasma cortisol levels were standardized using the plasma cortisol log-scale SD score corrected for genetic control genomic control and adjusted for age and sex (27). SNPs were identified in or in proximity to specific genes. The GWAS summary statistics from the Psychiatric Genomics Consortium (PGC) ADHD database, including 35,191 controls from 12 cohorts (19) and 20,183 patients with ADHD, were used as sample outcomes. Throughout our study, we considered effect sizes based on the European participants (19,099 cases and 34,194 controls) to establish the relationship between morning plasma cortisol levels and ADHD. In the reverse direction, we used the GWAS summary statistics from the ADHD Working Group of the PGC database (n = 55,347) as the discovery sample exposure and cortisol obtained from the CORNET consortium (n = 12,597) as the sample outcome to investigate the causal relationship of ADHD on morning plasma cortisol levels.

The CORNET consortium identified SNPs that were highly (P < 5 × 10^–6) and independently (r² < 0.01, window size = 10000 kb) associated with morning plasma cortisol levels. Typically, SNPs with P < 5 × 10^—8 are considered significant across the entire genome. However, only one SNP was screened at this threshold value, so the threshold was appropriately adjusted downward. Therefore, the selected instrumental variables meet the correlation assumption, i.e., there is a strong association between the exposure factors and instrumental variables. The Two-Sample MR R software package was used to exclude SNPs with disequilibrium in the links through the clump step. The parameters used were r² = 0.01 and kb = 10000, indicating that SNPs with r² greater than 0.01 within a 10 MB range, including the most important SNPs, were removed to eliminate the effect of linkage disequilibrium (LD). To verify whether the selected instrumental variables met the independence assumption, Pheno Scanner¹ was used to examine the association of the remaining SNPs with other phenotypes. The results showed that the remaining five SNPs were not associated with prenatal exposure to alcohol, tobacco, cocaine, prematurity, or other factors that may affect ADHD, confirming the independence hypothesis. Finally, the five extracted SNPs were matched with SNPs from the PGC ADHD database, and the corresponding data were collated and merged. We extracted the cortisol-associated SNPs for ADHD from a GWAS analysis in the PGC ADHD database, and we considered impact estimates from European participants, including 19,099 cases and 34,194 controls. SNPs that demonstrated a robust association with the statistical threshold for the GWAS (P < 5 × 10^–8, linkage disequilibrium r² = 0.01, window size = 10000 kb) were selected as instrumental genetic variables. The CORNET consortium (27) then collected genetic correlations among those with cortisol with European ancestry. In the end, Pheno Scanner was used to verify that the selected instrumental variables met the independence assumption (28), and seven SNPs were retrieved and matched with SNPs from the CORNET consortium.

In this study, we assessed the influence of a specific genetic variation of SNPs on exposure by determining its effect on an outcome. We then performed MR analyses to combine the estimates obtained using IVW, MR-Egger regression, Weighted median, Simple mode, and Weighted mode. The IVW method was predominantly used for analysis. The analysis outcome was presented as odds ratios (ORs) and a 95% confidence interval (CI).

We utilized various reliable methods to verify and correct the estimates obtained, including weighted median, MR-Egger regression, and MR-PRESSO (29–31). It has been demonstrated that combining the weighted median method with the MR-Egger regression method can help identify horizontal pleiotropy and provide an unbiased estimate of the causal influence. We used the weighted median approach to compare the two methods, allowing up to 50% of incorrect or pleiotropic instrumental factors in the analysis. MR studies (32, 33) require that instrumental variables can only influence outcomes through the studied exposure factor and are not directly related to outcomes. Testing the exclusivity hypothesis is challenging as genetic variants are all pleiotropic. The intercept term of MR–Egger regression is now commonly used to test for the presence of pleiotropy. When the linear regression intercept Egger–intercept of the MR–Egger model is close to 0 (P > 0.05), it indicates no pleiotropy in the instrumental variables, and the exclusion hypothesis can be considered valid; conversely, it indicates the presence of genetic pleiotropy and the exclusion hypothesis is invalid. In addition, the MR-PRESSO method was used to identify and correct outliers in the IVW linear regression. It consists of three components: (a) pleiotropy detection (MR-PRESSO global test), (b) pleiotropy correction by removing outliers (MR-PRESSO outlier test), and (c) study of statistically significant variations between the causal estimates before and after outlier removal (MR-PRESSO distortion test). We presented (31) the causative estimates for associations that have been corrected for outliers if the P-values of the global and distortion tests indicated a probability smaller than 0.05, suggesting the presence of horizontal pleiotropy. The Cochran Q statistic and P-value (34) were utilized to examine heterogeneity in the IVW estimations to corroborate the findings further. In addition, the results were analyzed separately using the leave-one-out method for sensitivity analysis. The leave-one-out method is a widely used technique for sensitivity analysis, whereby each SNP is removed one by one, and the results are observed to be statistically different before and after the removal. If P > 0.05 is obtained after excluding an SNP, the SNP does not have a non-specific effect on the effect estimation results (35). All analyses were conducted using R version 4.1.1 (Statistical Computing with R Foundation), and the R packages “Two Sample MR” (35) and “MR-PRESSO” (31) were used for MR analysis.

The statistical validity of SNPs and MR studies is crucial for the unbiased causal estimation of MR. The statistical validity of SNPs depends on the explanatory power of SNPs for the phenotype (R²), which can be assessed by the strength of the correlation between SNPs and the phenotype (F-statistic). The magnitude of R² ranges from 0 to 1, with a larger R² indicating a greater explanation of the phenotype by SNPs. The calculation of R² in this study was mainly based on the work of Papadimitriou et al. (36), published in Nature Communications. When the number of SNPs was less than or equal to 10, R² was calculated using the following formula: R² = 2 × EAF × (1-EAF) × beta². When the number of SNPs exceeds 10, the following equation was used to calculate it: R² = 2 × EAF × (1-EAF) × beta²[2 × EAF × (1-EAF) × beta² + 2 × EAF × (1-EAF) × N × se (beta) ²]. The F-statistic reflects the strength of the correlation between SNPs and phenotypes and is commonly used to determine whether SNPs are weak instrumental variables (37). A smaller F-statistic generally indicates a greater likelihood of bias in MR findings. In this study, the F-statistic was assessed using the following equation: F = (N-K-1)/K × [R^2/(1-R²)], where N is the sample size, K is the number of SNPs, and R² is the degree of explanation of exposure by SNPs. The size of the F-statistic decreases as the number of SNPs increases and increases as the sample size and the degree of explanation of exposure by SNPs increase. An F > 10 is considered a strong instrumental variable, while an F < 10 is weak. The F statistic corresponding to each SNP was calculated in this study (Tables 1, 3). Its distributions ranged from 22.57 to 52.72, with F-values > 10 indicating that the results are not weakly biased by IVs and are reliable.

No ethics approval was needed because all analyses in the present investigation used only publicly accessible data.

This study used MR analysis to evaluate the potential causal relationship between morning plasma cortisol levels on ADHD. Information on the five SNPs that achieved genome-wide significance (P < 5 × 10^–6) for morning plasma cortisol levels is presented in Table 1. The R² was 2.6%, and the distribution of the F-statistic corresponding to each SNP ranged from 22.57 to 48.81, while the F-statistic corresponding to all five SNPs was 67.28. These findings suggest that weak instrumental variable bias was less likely to impact the causal association. We observed no evidence of pleiotropy (MR-Egger Intercept: −0.008, SE = 0.0151; P = 0.653) or heterogeneity [Cochran’s Q (df = 3) = 0.879, P = 0.928], and MR-PRESSO did not detect any outliers. The aggregate estimations obtained by IVW or MR-Egger did not indicate any cortisol-related risk of ADHD (Table 2, Figures 2, 3 and Supplementary Figures 1, 2). Sensitivity analyses utilizing the weighted median, simple mode, and weighted mode did not reveal any significant associations (Table 2 and Figure 3), and MR-PRESSO failed to detect any significant link between cortisol and ADHD (β = 0.006, SD = 0.024, P = 0.815). Moreover, the leave-one-out sensitivity analysis of the IVW method demonstrated that even after gradually removing all five SNPs, the remaining four SNPs showed similar results to the combined effect OR of the IVW method, with P-values greater than 0.05. No SNPs with strong effects on the results were found in IV, indicating that the effect OR of the previous IVW method was robust (Supplementary Figure 2).

In this MR Analysis, we aimed to investigate the causal relationship between ADHD on morning plasma cortisol levels. Table 3 displays data on the associations of the seven selected SNPs with ADHD and morning plasma cortisol, with P-values and effect estimates calculated for Europeans. We calculated an R² of 2.2%, and the distribution of the F-statistic corresponding to a single SNP ranged from 31.76 to 52.72, while the F-statistic corresponding to seven SNPs was 173.137. Given that F > 10, the likelihood of weak instrumental variables was relatively small, indicating that the results of the MR analysis were unlikely to be affected by weak instrumental variable bias. The MR-Egger regression intercept term was 0.048, indicating no genetic pleiotropy between SNPs and morning plasma cortisol levels (P = 0.192). The MR-PRESSO test showed no genetic pleiotropy bias or outliers (β = −0.001, SD = 0.019, P = 0.892). Contrary to the previous MR analysis results, our findings revealed a negative causal association between ADHD and morning plasma cortisol levels (IVW method: b = −0.154, P = 0.018) (Table 4 and Figures 4, 5). Sensitivity analyses using IVW, MR-Egger regression, Weighted median, Simple mode, and Weighted mode were conducted, with results in Table 4 and Figure 5. Leave-one-out and funnel plot analyses indicated no effect (Supplementary Figures 3, 4). The IVW method showed that each 1 SD increase in clinical ADHD risk resulted in a 14.3% decrease in morning plasma cortisol levels (OR = 0.857, 95% CI: 0.755–0.974). Weighted median regression yielded similar results (OR = 0.842, 95% CI: 0.714–0.992). MR Egger’s results revealed no statistically significant association between clinical ADHD risk and morning plasma cortisol levels (OR = 0.509, 95% CI: 0.255–1.015). Simple mode and Weighted mode provided results similar to MR Egger’s results. The direction of the causal effect obtained by combining these five was consistent (Figure 5). Heterogeneity tests and sensitivity analyses of IVW (P = 0.594) and MR-Egger regression (P = 0.192) with Cochran’s Q-test [Cochran’s Q (df = 6) = 4.618, P = 0.594] indicated no heterogeneity in SNPs (Table 4). When using only one SNP as the independent variable (IV), the funnel plot (Supplementary Figure 3) reveals that the dots reflecting the effect of the causal relationship are symmetrically distributed, indicating that the causal association is less likely to be impacted by potential bias. The “Leave-one-out” sensitivity analysis results (Supplementary Figure 4) showed that the P-value range and IVW analysis results of the remaining six SNPs were similar to those of the included SNPs, and no SNPs were found to have a significant effect on the causal association estimates after excluding each SNP in turn.

Despite using MR analysis to reduce residual confounding bias, there are several limitations to consider in this study. First, one limitation is that a more liberal threshold (P < 5 × 10^–6, r² = 0.01) was used to select SNPs associated with morning plasma cortisol levels for analysis since only one SNP reached the genome-wide significance level (P < 5 × 10^–8, r² = 0.001) in the initial screening this may have introduced weak instrument bias. This study calculated the explanatory power of SNPs for phenotypes (R²) and the strength of association between SNPs and phenotypes to assess the statistical validity of SNPs in the MR study (F-statistic) (Tables 1, 3). The likelihood of weak instrument bias (60) may be mitigated by the fact that the F-statistic of each included SNP was greater than 10. However, even though the F-statistic was greater than 10 in all MR studies, the explanatory power of SNPs for cortisol (R² = 0.026) and ADHD (R² = 0.022) was low; therefore, bias in MR study results cannot be ruled out due to insufficient statistical power of SNPs. Second, it is widely believed that using different databases for different races may affect the results and cause population stratification bias. Population stratification refers to changes in the frequency of genetic variants across populations with various genetic origins, leading to false links between genetic variants and outcomes. Population stratification in MR investigations may lead to unjustified assumptions of independence or exclusivity and, therefore, incorrect causal conclusions. For example, Haworth et al. (61) found that genetic variation and primary health outcomes were associated with place of birth using UK Biobank data; failure to well-correct for population stratification can lead to spurious associations between genetic variation and primary health outcomes. To avoid population stratification bias, the most direct way is to include populations with the same genetic background for genetic association studies. However, due to traditional GWAS’s low statistical test efficacy, there has been a trend to expand the sample size for multicenter GWAS. In this study, the ADHD Working Group of the PGC database comprised a population-based cohort of 14,584 adults with ADHD and 22,492 control subjects. The GWAMA of morning plasma cortisol levels examined 12,597 people of European descent. Nonetheless, the analysis considered effect sizes based on the European members of the Working Group of the PGC database (19,099 cases and 34,194 controls). Only about 3.7% of the 55,347 participants were non-Europeans, so it is believed that such a small percentage of non-Europeans doesn’t affect the accuracy of the results. Third, our analysis was limited to morning plasma cortisol data; therefore, we could not examine the relationship between ADHD risk and cortisol levels throughout the day. It is worth noting that cortisol secretion exhibits a diurnal pattern, and incorporating data on cortisol fluctuations over a day could provide additional information. In addition, it is worth noting that after the submission of this manuscript, the Nature Genetics Journal updated the latest ADHD GWAS to include 27 markers (62). The results of this updated GWAS may refine and complement the conclusions of our study. We plan to conduct further discussions and studies in the future.