The AD GWAS summary statistics were obtained from a large-scale meta-analysis^[10]. In stage 1, 11,480,632 SNPs of 21,982 AD cases and 41,944 cognitively normal controls were genotyped and imputed first, and then meta-analysis was conducted to identify the 12 significant loci. In stage 2, the authors genotyped 11,632 variants on the I-select chip and tested the association in an independent sample set, which is composed of 8,362 AD cases and 10,483 normal controls. For the variations that the I-selected chip failed to accurately capture, stage 3A (N = 11,666) and 3B (N = 30,511) analyses were further performed. In this study, we adopted the stage 1 results of the AD meta-analysis. The International Genomics of Alzheimer’s Project (IGAP) was releasing this summary data with the accession number: NG00075.v1. We can download the P-values only summary statistics from https://dss.niagads.org/open-access-data-portal/#NG00075. Complete GWAS summary data were obtained upon application.

The migraine GWAS summary statistics were obtained from the FinnGen initiative, which is a large-scale genomic research project with 500,000 Finnish individuals^[30]. Here, we used the R12 version of the GWAS datasets released in 2024. The GWAS summary statistics of overall migraine, MA, and MO were conducted on genotyped 21,325,483 variants of 26,894 cases and 374,605 controls, 21,325,193 variants of 11,757 cases and 374,605 controls, and 21,325,144 variants of 9,690 cases and 374,605 controls, respectively. The above GWAS summary datasets are freely available at https://storage.googleapis.com/finngen-public-data-r12/summary_stats/release/finngen_R12_G6_MIGRAINE.gz, https://storage.googleapis.com/finngen-public-data-r12/summary_stats/release/finngen_R12_G6_MIGRAINE_WITH_AURA.gz, https://storage.googleapis.com/finngen-public-data-r12/summary_stats/release/finngen_R12_G6_MIGRAINE_NO_AURA.gz, respectively.

In this study, we utilized the GTEx V10 RNA-seq data to plot the gene expression in specific tissues. The GTEx project aims to investigate the relationship between genetic variation and gene expression in multiple healthy human tissues. The GTEx RNA-seq V10 dataset can be accessed from https://www.gtexportal.org/.

We carried out preprocessing on the downloaded GWAS datasets according to the following process. Firstly, we unified the reference genome of all GWAS data to genome reference consortium human build 37 (GRCh37)/hg19. For datasets utilizing other reference genomes, the liftOver tool was used to convert coordinates to GRCh37. Secondly, the non-standard SNPs lacking an “rs” prefix were excluded. Thirdly, we only kept the SNPs with single-nucleotide variants and filtered indels. Finally, the SNP with minor allele frequency (MAF) ≥ 0.01 in autosomes would be considered for subsequent analysis. These quality control procedures can ensure that all GWAS follow a unified standard and minimize bias to the greatest extent in the subsequent analysis process.

To ensure the robustness and reproducibility of our findings, several quality control measures were implemented. Firstly, potential bias arising from sample overlap between the AD (IGAP) and migraine (FinnGen) cohorts was assessed using the LDSC intercept. The intercept values were consistently close to 1.0, suggesting that the influence of sample overlap or residual population stratification was minimal. Secondly, linkage disequilibrium (LD) was estimated using the European ancestry reference panel from the 1000 Genomes Project Phase 3, ensuring consistency with the underlying GWAS datasets. Finally, multiple testing was rigorously controlled. A false discovery rate (FDR) correction was applied to the TWAS analyses, and a genome-wide significance threshold of P < 5 × 10^-8 was adopted for the cross-trait meta-analysis to limit false-positive findings.

We used the linkage disequilibrium score regression (LDSC) to estimate the genome-wide heritability for every trait using only GWAS summary statistics, without requiring individual-level genotype data^[31].

In order to obtain the shared genetic variants for multiple traits, we then conducted a cross-trait GWAS meta-analysis using the cross-phenotype association (CPASSOC)^[32]. CPASSOC is a cross-trait association analysis method based on summary statistics, which is used to simultaneously test whether a genetic variation is significantly correlated with multiple traits. The assumption is that if a certain SNP is associated with multiple traits, its cross-trait combined effect should be significantly stronger than that of any single trait.

CPASSOC provides two algorithms, Shom and Shet^[32]. Shom employs a fixed effect model, assuming that the effect size of the SNP is exactly the same for each trait. This model is suitable for traits with highly consistent effect sizes. Shet allows SNPs to have heterogeneity in their genetic effects for different traits (with different directions or magnitudes). Through random-effect meta-analysis, Shet gives higher weights to traits with large effects, reducing the interference from traits with small effects or heterogeneity. Thus, in the practical research where heterogeneity of effects is widespread, Shet becomes a more reliable choice for cross-trait meta-analysis. To formally justify this, Cochran’s Q test was applied to confirm significant heterogeneity across the top loci, supporting the use of the Shet random-effects model.

After cross-trait meta-analysis, a number of genetic variants shared by two traits were identified. To further ascertain the independent pleiotropic SNPs, we used the PLINK1.9 to clump the SNPs by using the parameter “-clump-p1 5e-8 -clump-p2 1e-5 -clump-r2 0.001 -clump-kb 500”. In this clumping process, the genetic variant with the lowest P value within each LD block was selected as the index SNP. SNPs that simultaneously meet the criteria of P value less than 5 × 10^-8 after meta-analysis and a P value less than 0.05 for single trait are considered as suggestive pleiotropic loci^[33]. To rigorously control for false positives and asymmetric chance associations, loci were declared as significant pleiotropic regions only if they simultaneously met a highly stringent cross-trait genome-wide significance threshold (P_CPASSOC < 5 × 10^-8) and a strict single-trait threshold of P < 5 × 10^-3 for both underlying phenotypes. In order to further obtain the function annotation of pleiotropic loci, we adopted the variant effect predictor (VEP) tool to obtain the nearest gene based on physical proximity^[34]. In addition, 3DSNP was used to obtain the regulated target gene based on chromatin three-dimensional interactions^[35].

Colocalization is a Bayesian statistical method to determine whether the association signals of two traits are driven by the same causal variant. For the independent pleiotropic loci, we then conducted colocalization analysis by using R package Coloc^[36]. This approach provides five mutually exclusive hypotheses (H0-H4), and calculates the posterior probability (PP) through Bayesian factor: H0 (no association), H1 (only associated with trait 1), H2 (only associated with trait 2), H3 (two distinct variants associated with two traits), and H4 (one shared variant associated with two traits). We extracted the summary statistics of independent pleiotropic SNPs ±500 kb from two GWAS as the input to calculate the PP. Loci with PPH3 > 0.5 or PPH4 > 0.5 can be considered colocalized^[37].

To capture the broad polygenic architecture of the shared genetic etiology, candidate genes for pathway enrichment analysis were mapped from variants meeting genome-wide significance (P_CPASSOC < 5 × 10^-8) using positional mapping and expression quantitative trait loci (eQTL) mapping in GTEx V8 brain tissues by functional mapping and annotation of GWAS (FUMA)^[38]. Pathway enrichment analysis was performed by web-based tool Metascape^[39]. Redundant enriched terms were clustered, and statistical significance was strictly defined at a Benjamini-Hochberg FDR-corrected P value < 0.05. The pathway enrichment results were plotted using R.

The shared SNPs between a pair of traits can be identified by cross-trait meta-analysis. Many genetic variation sites often participate in the progression of diseases and phenotypes by regulating the expression of genes. Therefore, we further identified the shared genes between the two traits by transcriptome-wide association studies (TWAS) using Fusion^[40]. GTEx V8 expression weights for TWAS are provided from Fusion. In addition to whole blood and brain tissues, other tissues and organs of the organism also play a very important role in the development of the disease. Firstly, we conducted the TWAS in 49 tissues to identify the significant gene-tissue pairs for single trait. Then, we intersected the significant gene-tissue pairs of different traits to obtain the shared genes. Considering the extensive multiple testing burden across 49 evaluated tissues, we employed a highly conservative global FDR correction (Benjamini-Hochberg) simultaneously across all tissue-gene pairs to strictly minimize false-positive discoveries.

To prioritize potential therapeutic drugs, we queried the Drug-Gene Interaction Database (DGIdb) and Drug Signatures Database (DSigDB) for Food and Drug Administration (FDA)-approved compounds whose targets overlap the shared genes, yielding a pharmacological landscape that could inform treatment strategies for both AD and migraine^[41,42].

PheWAS systematically relates a given variant or gene to the full spectrum of phenotypic outcomes recorded in large biobanks, thereby revealing pleiotropic effects. Using the AstraZeneca PheWAS portal, which is built on UK Biobank exome and phenotype data, we surveyed the disease landscape linked to each drug-target gene to flag potential on- or off-target liabilities of the corresponding compounds^[43].

All computational analyses were performed utilizing publicly available software and established bioinformatics pipelines. The primary tools employed in this study include LDSC (https://github.com/bulik/ldsc) for heritability estimation, CPASSOC (https://hal.case.edu/~xxz10/zhu-web/) for cross-trait meta-analysis, PLINK 1.9 (https://www.cog-genomics.org/plink/) for linkage disequilibrium clumping, FUSION (http://gusevlab.org/projects/fusion/) for TWAS, and the R package coloc (https://cran.r-project.org/web/packages/coloc/) for Bayesian colocalization. The positional and eQTL mapping of genetic variants were performed using FUMA (https://fuma.ctglab.nl/). Pathway enrichment was conducted via the Metascape platform (https://metascape.org). Downstream statistical analyses and data visualizations were executed in R software (v4.3.0) using standard packages including ggplot2 and pheatmap. All analyses strictly adhered to the standard protocols and official tutorials provided by the respective software developers.

We first calculated the SNP-based heritability for each trait by using univariate LDSC. The SNP-based heritability estimates for AD, overall migraine, MA, and MO are 0.0713, 0.0326, 0.0184, and 0.0177, respectively [Supplementary Table 1].

To explicitly address potential systematic biases arising from population heterogeneity between the IGAP and FinnGen cohorts, we assessed the global effect size (BETA) correlations using variants not expected to be associated with either trait. Specifically, we filtered for background null variants (single-trait P > 0.05 in both corresponding datasets) and randomly downsampled to 50,000 SNPs per comparison to plot their aligned effect sizes. We observed near-zero overall correlations across all comparisons (AD and overall migraine: R = -3.0 × 10^-5, P = 0.99; AD and MA: R = 0.0051, P = 0.26; AD and MO: R = 0.019, P = 2.3 × 10^-5; Figure 1). Although the correlation for the AD and MO pair reached nominal statistical significance due to the large number of evaluated variants, the absolute magnitude of the correlation coefficient remained biologically negligible. The absence of pervasive, directional genomic correlation among these null variants confirms that macroscopic methodological confounders, such as unadjusted population stratification or cryptic sample overlap, do not artificially inflate the genetic overlap. Collectively, these findings indicate that the cross-trait associations identified in the subsequent analyses are unlikely to be driven by systematic methodological artifacts, thereby increasing confidence in the robustness of the detected signals.

Several studies have shown that migraine is significantly associated with an increased risk of developing dementia in different ancestry cohorts^[22-27]. In addition, MR study demonstrated that migraine has a significant causal effect on AD^[29]. Therefore, we performed genome-wide cross-trait meta-analysis to identify the pleiotropic loci shared by two traits by CPASSOC.

As for the AD and overall migraine trait pair, we identified 4 suggestive independent shared loci and 1 significant shared locus after clumping by PLINK [Figure 2A and Supplementary Table 2]. The significant shared locus rs76968534 was identified under stringent dual-threshold criteria (P_CAPSSOC < 5 × 10^-8, P_single-trait < 5 × 10^-3). The significant locus rs76968534 is located in an intergenic region. Among the suggestive loci, rs117310449 (P_CPASSOC = 4.20 × 10^-46) is in close proximity to genes PVRL2, TOMM40, and CTB-129P6.4. PVRL2 plays a key role in immune response and is associated with the development of AD^[44-47]. TOMM40 is closely related to neuroinflammation and AD^[45,47,48]. CTB-129P6.4 is associated with the risk of intracerebral hemorrhage and AD^[49,50]. The other suggestive locus (index SNP: rs3851179, P_CPASSOC = 2.02 × 10^-15) is near the gene RNU6-560P, which is a non-coding gene associated with AD^[51,52]. The other suggestive locus (index SNP: rs10909892, P_CPASSOC = 9.81 × 10^-10) is close to the gene PRDM16. This gene is expressed in astrocytes and downregulated in AD patients^[53].

We identified 5 suggestive independent pleiotropic loci and 2 significant pleiotropic loci for AD and MA [Figure 2B and Supplementary Table 3]. The loci rs9271060 (near HLA-DRB1, CR753309.2, CR753835.1) and rs79254581 were shared by AD and MA. The suggestive shared locus (index SNP: rs8106813, P_CPASSOC = 9.06 × 10^-20) is close to the gene APOC1P1. It plays a crucial role in the development of AD risk^[54]. The next locus (index SNP: rs3740688, P_CPASSOC = 5.54 × 10^-11) is near the SPI1, which is a risk gene for AD. rs16035 (near CACNA1A) and rs11767557 (near EPHA1, EPHA1-AS1) were shared by AD and MA. CACNA1A is mainly expressed in the nervous system and associated with hemiplegic migraine^[55,56]. HLA-DRB1, EPHA1, and EPHA1-AS1 are risk factors for AD^[57-59].

As for the AD and MO trait pair, we identified 2 independent suggestive shared loci [Figure 2C and Supplementary Table 4]. The suggestive locus (index SNP: rs3851179, P_CPASSOC = 2.21 × 10^-16) is close to the gene RNU6-560P, which is closely linked with AD risk^[60]. Although two suggestive loci were initially identified, neither satisfied our more stringent locus-level filtering criteria, leaving no robust shared loci for downstream analyses.

The purpose of colocalization is to distinguish whether the associated signals of two traits are driven by the same causal variant or by different but linked variants. We found that 1 shared locus (index SNP: rs76968534) between AD and overall migraine demonstrated strong posterior probabilities for a single shared causal variant (PPH4 > 0.5), pointing to a shared causal variant [Figure 3A and Supplementary Table 5]. As for the AD and MA, the critical locus (rs9271060) exhibited a high PPH3 (> 0.5), reflecting distinct but tightly correlated regulatory variants driving the two traits independently within the same genetic region, while rs79254581 was colocalized at the same genetic variant (PPH4 > 0.5) [Figure 3B and C, Supplementary Table 6].

Based on the annotated genes from the genome-wide significant cross-trait loci, we performed pathway enrichment analysis. After rigorous Benjamini-Hochberg FDR correction (q-value < 0.05), the significantly enriched pathways for AD and overall migraine/MA were highly concentrated in neuro-immune regulation, lipid metabolism, and amyloid pathology. The most significantly enriched biological processes included the regulation of amyloid-beta clearance and cholesterol transport and efflux. Furthermore, robust enrichment was observed in immune-related cascades, notably the regulation of complement-dependent cytotoxicity [Figure 4A and B, Supplementary Tables 7 and 8]. Importantly, consistent with our stringent locus-level analyses, which identified no robust shared loci between AD and MO, downstream pathway enrichment analysis was inherently not applicable and thus intentionally not pursued for the MO subtype.

We performed TWAS to identify the shared genes between AD and migraine, including its subtypes, in 49 tissues. Firstly, the shared genes related to single trait were identified in different tissues. Then, we obtained the shared genes between two traits by intersecting the single-trait associated genes according to different tissues. For AD and MA, following a conservative global FDR correction across all 49 tissues, the complement component 2 (C2) gene emerged as a highly robust pleiotropic driver specifically within the central nervous system (Brain Cerebellar Hemisphere). TSBP1-AS1 was significant in thyroid tissue [Table 1].

We obtained the 2 shared genes between AD and MA. In order to further explore the expression in different tissues, we further quantitatively characterized gene expression using the GTEx V10 data.

According to the results, we found that the C2 gene is highly expressed in liver and lung tissues, and it is also expressed in the cerebellar hemisphere of the brain [Figure 5 and Supplementary Table 9]. TSBP1-AS1 is highly expressed in brain spinal cord (cervical c-1), with relatively lower expression across other tissues, including the thyroid [Figure 5 and Supplementary Table 9].

After identifying the shared genes between AD and migraine subtypes, we further identified potential target genes. Based on TWAS prioritization, C2 and TSBP1-AS1 were carried forward for downstream drug-target analyses.

Then, we made full use of DGIdb and DSigDB to obtain FDA-approved drugs for these target genes. According to DSigDB, there are two drugs (ESTRADIOL and PROGESTERONE) with reported gene associations. We did not obtain candidate small-molecule drugs or biologics for the target gene TSBP1-AS1 from DGIdb and DSigDB. This is expected, as TSBP1-AS1 is a long non-coding RNA (lncRNA), and current pharmacogenomic databases are heavily biased toward conventional protein-coding targets [Supplementary Table 10].

We have obtained the candidate therapeutic drugs for the potential gene targets. To evaluate the broader phenotypic associations of the prioritized target gene, we then conducted the PheWAS based on potential drug targets. According to the AstraZeneca PheWAS portal based on sequencing and phenotype data from the UK Biobank, the target gene C2 was not significantly associated with infectious, neoplasms, blood/immune, endocrine/metabolic, mental, nervous, eye, ear, cardiovascular, respiratory, digestive, skin, musculoskeletal, urinary/renal, pregnancy, congenital, lab findings, health services, and special phenotypic categories [Figure 6]. This PheWAS result suggests that C2 showed no significant associations across major disease categories. However, we caution that gene-level phenome associations act strictly as biological proxies and do not guarantee the clinical pharmacological safety of corresponding therapeutic compounds.

We performed genome-wide cross-trait meta-analysis to identify the pleiotropic loci for a pair of traits, followed by functional annotation. As for AD and overall migraine, 4 independent suggestive shared loci and 1 significant shared locus were identified by CPASSOC. In addition, we found 5 suggestive pleiotropic loci and 2 suggestive pleiotropic loci for AD and MA. Consistent with previous epidemiological observations suggesting a stronger association between MA and dementia than MO, we identified more shared loci between AD and MA than between AD and MO^[24-26]. Collectively, these findings support a shared genetic basis between AD and migraine at the locus level.

Furthermore, we identified the shared genes for AD and migraine, including their types, in 49 tissues by TWAS. At gene levels, we found the C2 and TSBP1-AS1 shared by AD and MA. C2 is a serum glycoprotein and is part of the classical pathway of the complement system. It is associated with certain autoimmune diseases^[61]. TSBP1-AS1 is a lncRNA and is highly expressed in immune system cells^[62]. Additionally, it was proven that TSBP1-AS1 is shared by immune and bone diseases^[63]. These genes were expressed in specific brain tissues, as well as peripheral tissues, including the liver and thyroid.

Regarding the associations identified in non-central nervous system (CNS) tissues, such as thyroid, we interpret these findings as a reflection of the systemic nature of pleiotropic genes. For example, C2 is a fundamental component of the systemic immune system. While its primary disease relevance in this context is likely within the neuro-immune environment of the brain, its genetic signals are detectable across multiple tissues due to shared regulatory elements. In this study, we have prioritized the interpretation of results from CNS-pertinent tissues, such as cerebellum and basal ganglia, as they align more closely with the known pathophysiology of neurodegeneration and headache disorders.

Rather than positing a singular shared causal etiology, our refined colocalization and systems biology analyses reveal a highly complex genetic intersection between AD and MA. The identification of C2 and TSBP1-AS1 implicates immune-related mechanisms. C2 participates in complement activation, whereas TSBP1-AS1 may contribute to immune regulation through mechanisms that remain to be fully characterized. These pathways are implicated in AD-associated neuroinflammation, a hallmark of migraine pathophysiology. The lack of existing drugs targeting TSBP1-AS1 likely reflects the historical focus on protein-coding genes. Nevertheless, the pleiotropic associations observed for TSBP1-AS1 support its prioritization for future functional investigation and evaluation as a potential target for RNA-based therapeutics, including antisense oligonucleotides.

Taken together, these findings support a neuro-immune-vascular axis as a unifying framework, providing mechanistic insight into the reported epidemiological association between migraine and AD.

Despite the integrative multi-omics framework employed in this study, several limitations should be acknowledged. First, to reduce potential bias arising from population stratification, our analyses were confined to GWAS summary statistics derived from European-ancestry cohorts (FinnGen and IGAP). Notably, the migraine data utilized the FinnGen cohort, which possesses a well-documented distinct population history and founder effect compared to the multi-cohort European IGAP dataset. While our LDSC intercept evaluations suggested minimal bias from cryptic sample overlap, this population heterogeneity represents a potential confounder. Future replication in independent, non-Finnish European cohorts is essential to independently validate these specific loci. Second, the shared risk genes identified in this study, including C2 and TSBP1-AS1, were inferred through statistical and computational approaches. Although these findings provide important hypotheses, experimental validation remains necessary. In particular, functional studies in relevant cellular and animal models, such as clustered regularly interspaced short palindromic repeats (CRISPR)-based perturbation in neuronal or microglial systems, will be essential to establish causal roles and underlying biological mechanisms. Third, gene expression patterns across 54 tissues were characterized using bulk tissue data from GTEx data. While informative, such data may obscure cell-type-specific regulatory effects, particularly within complex structures such as the neurovascular unit. Future studies leveraging single-cell RNA sequencing or spatial transcriptomics will be critical to resolve these signals at higher resolution. Finally, our drug target and PheWAS analyses are strictly hypothesis-generating. And predicting therapeutic efficacy or safety from gene-level associations is inherently limited by unknown pharmacokinetic and off-target pharmacological variables.

In conclusion, our work has characterized the distinct and shared genetic correlations between AD and migraine in multiple aspects. Additionally, this study provides important insights for understanding the genetic etiology between the two traits. These findings are crucial to interpreting the molecular mechanisms of the disease and identifying potential therapeutic targets.