We employed Two-sample MR analysis to investigate the genetic causal effects of exposures on outcomes based on GWAS summary statistics. The study design is shown in Fig. 1. In step 1, we assessed the causal effects of migraine on four common types of dementia, including Alzheimer’s disease (AD), vascular dementia (VaD), frontotemporal dementia (FTD), and Lewy body dementia (LBD). In step 2, we examined the causal effects of migraine on longitudinal changes in four global and three local brain measures. In step 3, we conducted a subtype analysis to replicate the findings of steps 1–2. In step 4, we investigated the mediation effects of longitudinal brain measures between migraine and dementia by conducting a two-step MR mediation analysis.

Study design for identification of the causal relationships between migraine, dementia, and longitudinal brain measures. Abbreviations AD, Alzheimer’s disease; VaD, vascular dementia; FTD, frontotemporal dementia; LBD, Lewy body dementia; MA, migraine with aura; MO, migraine without aura

To avoid the bias resulting from sample overlaps between exposures and outcomes (i.e., samples from the UK biobank) in Two-sample MR analysis, we did not use the latest GWAS summary statistics of migraine [29]. Instead, we included an earlier version of GWAS summary statistics for migraine, which has no sample overlaps with the outcomes used in this study [30]. This GWAS yielded also the genetic variants related to two subtypes of migraine, MA and MO. To ensure privacy protection for participants in the 23andMe cohort, we excluded their samples from the GWAS summary statistics. Consequently, the GWAS summary statistics used in the present study comprised data from 202,140 (Migraine: Ncase = 29,209, Ncontrol = 172,931), 151,215 (MA: Ncase = 6,332, Ncontrol = 144,883), and 147,970 (MO: Ncase = 8,348, Ncontrol = 139,622) individuals, respectively. These participants were recruited from six tertiary headache clinics (N = 20,395) and fifteen population-based cohorts (N = 181,745) through various methods, such as advertisements, the project’s website, national media campaigns, and referrals from headache centers. Detailed recruitment information is available in the respective cohort descriptions. Migraine cases, including those with current episodes and a history of migraine, were identified using self-reports, diagnostic questionnaires meeting full or modified ICHD-II criteria, or diagnoses by trained physicians or senior medical students. This enabled the inclusion of a large number of cases, thereby enhancing the statistical power. Only individuals who met the strict classification criteria standardized by the International Headache Society were included as migraine subtype cases because the migraine aura can be difficult to distinguish in a questionnaire-based setting.

We collected the latest GWAS summary statistics of four common types of dementia, including AD (Ncase = 111,326, Ncontrol = 677,663) [31], VaD (obtained from the FinnGen database R9; Ncase = 2,335, Ncontrol = 360,778), FTD (Ncase = 2,154, Ncontrol = 4,308) [32], and LBD (Ncase = 2,591, Ncontrol = 4,027) [33]. In the AD GWAS, cases were identified using multiple approaches, including clinical diagnosis by experts, self-report of the diagnosis, or self-report of a family history of AD. The inclusion of some proxy cases in the AD GWAS may lower the specificity of the findings. We thus conducted a replication analysis using the AD GWAS in the FinnGen database (R9; Ncase = 9,301, Ncontrol = 367,976) that has an overlap with the AD GWAS used in the primary analysis but only included clinically diagnosed cases. The VaD cases were collected from nationwide electronic health registers in Finland using the International Classification of Diseases, 10th Revision (ICD-10) codes, as defined by FinnGen clinical expert groups. Diagnoses were based on hospital billing codes rather than specific viral assays. The diagnosis of FTD was made by a neurologist using the Neary criteria (97% of the total sample) or in a minority (3% of the total sample), by pathological diagnosis (e.g., TDP-43 and FUS). LBD cases were confirmed based on either pathologically definite criteria (69% of the total sample) or clinically probable criteria (31% of the total sample), as recommended by the Dementia with Lewy Bodies Consortium. The diagnostic process integrated clinical features and biomarkers obtained from imaging and cerebrospinal fluid analyses.

We collected the GWAS of longitudinal brain measures from a study conducted on 15,100 participants of European ancestry where each participant underwent both baseline and follow-up MRI scans [28]. Participants in this study were recruited from various population-based, case-control, and family-based cohorts through multiple methods, including invitation letters, citizen registries, and the project’s website. Detailed recruitment information can be found in the descriptions of the respective cohorts. Considering that genetic risk for disease may be associated with genetic influences on brain changes, both healthy participants (89% of the total sample) and patients with neuropathic or psychiatric disorders (11% of the total sample) were included in the analysis, enhancing the applicability for inferring pathology and the adverse consequences of diseases. This study processed the MRI data using the FreeSurfer, a widely used tool for automated brain morphometry analysis. The phenotypes in this study are annual rates of change of 15 brain imaging-derived measures that are calculated by subtracting baseline brain measures from follow-up measures and dividing by the number of years of follow-up duration. Of the 15 longitudinal brain measures, we included four global (total brain volume, total cortical volume, total cortical surface area, and mean cortical thickness) and three local (volumes of the hippocampus, thalamus, and caudate) brain measures that were strongly associated with aging and presented an almost linear trajectory of change over time. It is noted that all GWAS included in the present study exclusively comprise participants of European ancestry.

Before conducting Two-sample MR analyses, we first selected and preprocessed the genetic IVs. The genetic variants with minor allele frequency (MAF) < 1% were removed from GWAS summary statistics. To meet strong associations between IVs and exposures, we selected the genetic IVs with a genome-wide significance threshold of p < 5 × 10− 8 and F value (β2 / se2) > 10. When the number of genetic variants reaching the genome-wide significance threshold was no more than 3, we relaxed the significance threshold to p < 5 × 10− 6 [34, 35]. The resulting genetic IVs were pruned to high independence with a r2 threshold of 0.001 and a window size of 1 Mb. To ensure that the genetic IVs are associated with outcomes only through exposures, we removed the genetic IVs that were strongly associated with the outcome (p < 5 × 10− 8). We also removed the Single Nucleotide Polymorphisms (SNPs) located in long LD regions due to their high potential for pleiotropy [36]. After converting all odds ratio (OR) values in case/control GWAS to log odds, the effects of genetic IVs on exposure and outcome were harmonized to the same alleles. Because palindromic SNPs are sensitive to strand-flipping issues that impede the harmonization of effect alleles, we removed the palindromic SNPs (i.e., A/T or G/C alleles) with MAF close to 50%. The genetic IVs that were not available in outcome GWAS were replaced with proxy SNPs (r2 > 0.8) using a web-based tool “LDlinkR” [37]. The outliers were detected and excluded using the “ivw_radial” (alpha = 0.05, weights = 1, tol = 0.0001) and “egger_radial” (alpha = 0.05, weights = 1) of the “RadialMR” package [38].

We used the TwoSampleMR R package (https://mrcieu.github.io/TwoSampleMR) to perform Two-sample MR analyses with the multiplicative random-effects inverse-variance weighted (IVW) estimate as the primary analysis method to evaluate the causal effects of exposures on outcomes. To examine the robustness of the IVW estimate, we employed three supplementary MR methods (weighted median, weighted mode, and MR-Egger method) to conduct MR analysis. The significant threshold was set as two-tailed p < 0.05 and corrected for multiple testing with Bonferroni within each step (step 1: 0.05/4 = 0.0125; step 2:0.05/7 = 0.0071; step 3: 0.05/6 = 0.0083).

To exclude the potential influence of pleiotropy, we validated our findings by conducting a succession of sensitivity analyses as follows: (1) MR-Egger regression and MR-PRESSO Global test; (2) Cochran’s Q heterogeneity test; (3) leave-one-out (LOO) analysis. We also conducted a replication analysis by excluding potential pleiotropic IVs that were strongly associated with some confounders in European ancestry. Briefly, we searched for SNPs exhibiting significant associations with smoking status, alcohol consumption, major depressive disorder, coronary artery disease, stroke, and hypertension in Phenoscanner (http://www.phenoscanner.medschl.cam.ac.uk) and removed them from the IVs. For the dementia diseases that were significantly affected by migraine, we additionally conducted reversed MR analyses to assess their causal effects on migraine.

We conducted a two-step MR mediation analysis to investigate the mediating pathway from migraine to dementia via longitudinal brain measures. In the first step, we estimated the causal effect of migraine on longitudinal brain measures. In the second step, we assessed the causal effect of longitudinal brain measures on dementia. Finally, we quantified the indirect effect of migraine on dementia via longitudinal brain measures. The “product of coefficients” and “delta” methods were used to assess the indirect effects and their standard errors, respectively [39]. A sample overlap was found between the GWAS of longitudinal brain measures and AD (i.e., samples from the UK biobank). To avoid the bias caused by the sample overlaps in Two-sample MR analysis, we used the GWAS of AD in the FinnGen database (R9; Ncase = 9,301, Ncontrol = 367,976) to complete the mediation analysis.