This study utilized a two-step MR analysis of genetic summary data to investigate the extent to which sleep traits, such as insomnia, sleep duration, daytime sleepiness, napping, and short sleep duration, explain the detrimental effect of depression on migraine risk.

The genome-wide association study summary data utilized herein are publicly available, and ethical approval and informed consent were obtained in each original study. Table 1 depicts the datasets that we included.

All selected single nucleotide polymorphisms and their associations with depression, mediators, and migraine were extracted from the genome-wide association studies depicted in Supplementary Table S1. Researchers screened out the eligible genetic variants that met the conditions based on strict quality control from the genome-wide association study summary statistics of depression and various sleep-related traits including insomnia, sleep duration, daytime sleepiness, napping, and short sleep duration. A schematic overview of the present study design is illustrated in Figure 1.

When performing MR analysis using genetic variants as instrumental variables, MR analysis should be based on three principal assumptions (Boef et al., 2015; Headache Classification Committee of the International Headache Society IHS, 2018) genetic variants should be associated with the exposure; (GBD, 2016 Neurology Collaborators, 2018); genetic variants should be associated with the outcome exclusively through the exposure; and (Murray et al., 2012) genetic variants should be independent of any measured and unmeasured confounders. The single nucleotide polymorphisms associated with depressive phenotype and five sleep-related traits with genome-wide significance (P < 5e-8) were extracted. Because the existence of linkage disequilibrium may lead to corresponding bias, controlling linkage disequilibrium before subsequent analysis was necessary. Herein, independent single nucleotide polymorphisms were selected by setting r2 < 0.001 and window size = 10,000 kb. To further explore whether genetic variants were interfered by other confounding factors, we searched all the selected instrumental variants in Phenoscanner database (http://www.phenoscanner.medschl.cam.ac.uk), which provides detailed information pertaining to human genotype-phenotypes (Supplementary Table S2).

Genetic associations for depression, such as depression (n = 500,199 individuals) and major depressive disorder (n = 480,359 individuals) (Wray et al., 2018), were obtained from genome-wide association study summary statistics in psychiatric genomics consortium and UK Biobank participants of European ancestry (Table 1).

We considered a genome-wide association study for five sleep-related traits ascertained in UK Biobank: insomnia (Lane et al., 2019) (n = 474,292 individuals), sleep duration (Dashti et al., 2019) (n = 446,118 individuals), short sleep duration (n = 411,934 individuals), daytime sleepiness (Wang et al., 2019) (n = 452,071 individuals), and napping (Dashti et al., 2021) (n = 452,633 individuals), as depicted in Table 1. We selected all available sleep traits to provide an unbiased investigation (Luo et al., 2022). It has been indicated that the questions utilized to define patient self-reported insomnia symptoms in UK Biobank exhibit sensitivity and specificity for clinically diagnosed insomnia disorders (Jansen et al., 2019). Although daytime sleepiness is usually studied as an exposure or an outcome, herein, it was included as a mediator because the genetic structure of daytime sleepiness indicates that this trait may partly reflect sleep fragmentation (Kim et al., 2016; Daghlas et al., 2020). Genetic variants associated with sleep traits in these genome-wide association studies were also strongly correlated with corresponding objective sleep indicators (Jones et al., 2019).

We acquisitioned genetic associations with migraine from the largest meta-analysis of genome-wide association migraine studies conducted by the International Headache Genetics Consortium (Hautakangas et al., 2022). This study comprised 873,341 individuals of European ancestry (102,084 cases and 771,257 controls) from five study collections, such as IHGC 2016 (Gormley et al., 2016), 23andMe (23andMe.com), UK Biobank (ukbiobank.ac.uk), GeneRISK (generisk.fi), and Nord-Trøndelag Health Study (ntnu.edu/hunt). The characteristics of each contribution cohort have been described in previous studies. Migraine cases were defined using a range of different methods, including self-reporting, questionnaires assessing diagnostic criteria, and diagnosis by trained clinicians. All participants exhibited genetically verified European ancestry.

With regard to depression and migraine, after harmonization of the effect alleles across the genome-wide association studies, we utilized MR analyses to determine depression-based MR estimates for migraine, which assumes the absence of invalid genetic instruments (Burgess et al., 2013), including inverse-variance weighted (IVW), MR-Egger, weighted median, maximum likelihood (ML), and penalized weighted median. The primary MR analyses were conducted using IVW regression analysis. When the MR assumptions are met, this odds ratio (OR) is an estimate of the causal effect of the exposure on outcome. All statistical analyses were performed using the R programming language (version 4.0.5), and MR analyses were conducted using the R-based package “TwoSampleMR” (version 0.5.6).

The IVW method provides the most precise estimates; however, it is sensitive to invalid instrumental variables and pleiotropy (Burgess et al., 2017). Therefore, we utilized the MR-Egger, weighted median, ML, and penalized weighted median as sensitivity analyses. The MR-Egger regression can detect possible pleiotropic effects and provide estimates after pleiotropy correction, albeit with low power (Bowden et al., 2015). The weighted median method can produce consistent causal estimates, assuming >50% of the instrumental variables from valid single nucleotide polymorphisms (Bowden et al., 2016). ML and penalized weighted median were mainly utilized to assess the robustness of MR results (Hartwig et al., 2017).

To estimate the indirect effect of depression on migraine through sleep traits, we performed a mediation analysis that included the causal estimates from two-step MR analyses and the total effect from univariable MR analyses of depression on migraine.

First, univariate MR analysis was performed to indicate the causal effect of depression on migraine, and it was considered to be the total effect of exposure (depression) on outcome (migraine), which was estimated to be β1; by contrast, the causal estimates of depression on risk factors related to sleep traits and risk factors on migraine were β2 and β3, respectively. The indirect effects of risk factors related to sleep traits were β2*β3, which was calculated using the coefficient product method. The standard error and the confidence interval of the indirect effect were calculated using the Delta method. Second, the direct effect was the causal estimates after excluding the indirect effect, which was β1-β2*β3 (Relton and Davey Smith, 2012). Finally, to verify the potential mechanism of risk factors, we established a causal association in each univariable MR analysis. It is important to note that the estimate of each causal effect should satisfy the following relationship: |β2*β3| < |β1|. If β2*β3 = β1, the exposure of interest completely affected the outcome through mediator, and it was considered to be a complete mediator. By contrast, |β2*β3| > β1 indicated that there may be a logical error.

IVW and MR-Egger regression were applied to test the heterogeneity, and Cochran’s Q statistics were calculated to quantitatively evaluate the heterogeneity. If the heterogeneity existed (p < 0.05), then the random effect IVW results were dominant; otherwise, it referred to the results of fixed effect IVW. MR-Egger intercept tests were performed to detect horizontal pleiotropy, and single nucleotide polymorphism and leave-one-out analyses were utilized to identify outlier single nucleotide polymorphisms driving relationships. Variance (R2) in the MR study refers to the proportion of total variation in the exposure that is explained by the genetic instruments. R2 for each trait were derived from the original study. To ensure that MR estimates minimize potential weak instrument bias, we considered a ≥ 10 F-statistic as sufficient for performing an MR analysis.

MR-Pleiotropy Residual Sum and Outlier methods (MR-PRESSO) analyses, as installed in the R-based package “MRPRESSO” (version 1.0), were also utilized to examine and correct potential horizontal pleiotropy through removing outliers (Ong and MacGregor, 2019), of which the distributions number was set to 1,000.

Summary information of instruments identified for depression and sleep traits are presented in Supplementary Tables S1, S2. Supplementary Tables S3, S4 depicts the harmonising results of instrumental variables filtering in detail. Ambiguous single nucleotide polymorphisms with incompatible alleles or palindromic single nucleotide polymorphisms with ambiguous strand were removed and illustrated in Supplementary Table S4.

Using the 46 depression-related single nucleotide polymorphisms, we observed the existence of a potential causal effect of depression on the risk of migraine (OR = 1.321, 95% CI: 1.184–1.473, p < 0.001), which was depicted in Table 2; Supplementary Table S7. The F-statistic of the depression-related single nucleotide polymorphisms utilized herein was approximately 14.92. In addition, we also observed that a genetic predisposition to depression leads to a higher risk of insomnia (OR = 1.088, 95% CI: 1.053–1.124, p < 0.001) with an F-statistic of 15.31. In the validation cohort containing 33 major depressive disorder-related single nucleotide polymorphisms, we also noted a potential causal effect of depression on the risk of migraine (OR = 1.262, 95% CI: 1.114–1.431, p < 0.001), and the F-statistic was 16.45. Moreover, we observed that depression was associated with an increased risk of insomnia (OR = 1.082, 95% CI: 1.046–1.120, p < 0.001), and the f statistics was 16.43. The output pertaining to the heterogeneity analysis of depression-migraine and depression-insomnia indicated possible heterogeneity, and the results of the heterogeneity test are illustrated in Supplementary Table S8. The MR-Egger intercept was utilized to evaluate the horizontal pleiotropy, and the results did not exhibit any evidence of horizontal pleiotropy, as illustrated in Supplementary Table S9. With outliers removed (major depressive disorder-migraine: rs2005864; depression-insomnia: rs12919291, rs12967143, rs28541419, rs2876520, rs354155, rs4730387, rs4936276, rs754287, rs76954012; major depressive disorder-insomnia: rs12552, rs12958048, rs17727765, rs2005864, rs7430565, rs76485002) in the MR-PRESSO analysis, the estimate did not change materially after correction (Supplementary Tables S5, S6).

Meanwhile, single nucleotide polymorphism did not affect the overall effect of depression-migraine and depression-insomnia in the leave-one-out sensitivity analysis (Figure 2B, D; Supplementary Figures S1, S2). The funnel plot was symmetrical, which indicated no pleiotropy (Supplementary Figures S3–S6). Furthermore, the scatter plot depicts the individual putative causal effect, and a significant positive correlation was observed between depression and migraine and between depression and insomnia (Figure 2A, C; Supplementary Figures S7, S8). The intercepts calculated in the MR analysis method were close to zero, which indicated that the probability of horizontal pleiotropy was low. Forest plots depicted the causal effect estimates between each single nucleotide polymorphism and the outcome, and they illustrated the combination of the effect estimates based on IVW and MR-Egger regression (Figures 3A, B; Supplementary Figures S9–S12).

In conformance with the observations on depression, a close association between genetically predicted insomnia and migraine risk was observed, as illustrated in Table 2; Supplementary Table S7 (IVW: OR = 1.766, 95% CI: 1.120–2.784, p = 0.014; MR-Egger: OR = 4.804, 95% CI: 0.641–36.016, p = 0.14; weighted median: OR = 2.253, 95% CI: 1.368–3.710, p = 0.001; maximum likelihood: OR = 1.827, 95% CI: 1.325–2.520, p < 0.001; penalized weighted median: OR = 2.409, 95% CI; 1.478–3.926, p < 0.001). In the primary analyses using IVW, other genetically determined sleep traits, such as sleep duration, daytime sleepiness, napping, and short sleep duration, were not associated with the risk of migraine (Table 2; Supplementary Table S7). It was observed that single nucleotide polymorphisms identified in insomnia and migraine were available instruments, with F-statistics = 14.06. In addition, the heterogeneity was detected using MR-Egger and IVW for insomnia (MR-Egger: Q statistics = 73.380, p < 0.001; IVW: Q statistics = 75.537, p < 0.001), and MR-Egger analysis exhibited no evidence of pleiotropy (Supplementary Tables S8, S9).

In the leave-one-out analysis, we observed that the risk estimates of genetically predicted insomnia and migraine did not change substantially after excluding single nucleotide polymorphism at each period (Figure 2F). The scatter plot, funnel plot, and forest plot also exhibit similar trends to depression-migraine analysis (Figure 2E; Supplementary Figure S13; Figures 3A, B).

Using mediation analyses, the effects of depression on migraine via insomnia were quantified. The mediation effect of depression on migraine mediated by insomnia were 0.048. Insomnia accounted for 17.3% of the total effect of depression on migraine (Supplementary Table S10). Meanwhile, as a validation cohort, the mediating effect of major depressive disorder on migraine mediated by insomnia was 0.045, whereas insomnia accounted for 19.3% of the total effect of major depressive disorder on migraine.