Genetic regulation of cell type–specific chromatin accessibility contributes to migraine etiology

Abstract

Background

Migraine is a common and disabling neurological disorder, yet most causal regulatory variants and their target cell types remain unresolved. Disease-associated variants are enriched in non-coding cis-regulatory elements, indicating that they act by modulating gene regulation. Chromatin accessibility quantitative trait loci (caQTLs) capture genetic effects on these regulatory elements with cell-type resolution and can pinpoint the regulatory regions and cellular contexts mediating disease risk. However, brain cell type–specific caQTLs have not previously been integrated with migraine genetics.

Methods

We performed two-sample Mendelian randomization (MR) using brain cell type–specific cis-caQTLs from sorted neuronal (NeuN⁺) and non-neuronal (NeuN⁻) nuclei of 616 postmortem brain donors as exposures, with FinnGen Release 12 migraine GWAS (26,894 cases / 374,605 controls) as the discovery outcome. Open chromatin regions (OCRs) reaching Bonferroni-corrected significance were replicated in the UK Biobank (25,393 cases / 433,047 controls) and the Million Veteran Program (31,836 cases / 405,831 controls), and further evaluated by Bayesian colocalization and instrument-level phenome-wide screening. Multi-modal target gene annotation integrated Hi-C, cis-eQTL, ABC, and ENCODE-rE2G models; nominated genes were further evaluated for drug-target safety using gene-level rare-variant collapsing analyses and characterized for regional brain expression using the Human Protein Atlas.

Results

Of 15,334 OCR–migraine causal estimates, 19 reached Bonferroni-corrected significance in FinnGen R12. Seven were Bonferroni-replicated in both replication cohorts with concordant directions, and three showed strong colocalization (PP.H4 > 0.80). After excluding one OCR with extensive cardiovascular pleiotropy, two non-neuronal OCRs were retained: Peak_7234 (chr1:115,282,554–115,283,246) and Peak_124742 (chr4:35,524,861–35,526,079). Multi-modal annotation of Peak_7234 convergently nominated NGF as the highest-priority candidate target gene, with TSPAN2 as a secondary candidate; Peak_124742 lies in a gene-poor region with insufficient regulatory annotation to confidently nominate a target gene. Neither NGF nor TSPAN2 showed associations reaching phenome-wide significance.

Conclusions

Two non-neuronal brain open chromatin regions are prioritized as likely causal contributors to migraine risk. At the chr1:115.3 Mb region, NGF is nominated as the highest-priority candidate downstream target, with TSPAN2 as a complementary candidate. These findings extend cell type–specific causal mapping of migraine to the regulatory layer and identify candidate regulatory elements and target genes for further mechanistic and translational investigation.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s10194-026-02450-8.

Introduction

Migraine is a highly prevalent and disabling neurological disorder, affecting over one billion people worldwide and ranking among the leading causes of years lived with disability [1, 2]. Genome-wide association studies (GWASs) have identified more than 100 genetic loci associated with migraine risk [3, 4], yet translating these statistical associations into mechanistic understanding remains challenging. Owing to linkage disequilibrium (LD) between genetic variants in the associated regions, pinpointing the causal SNPs that drive disease risk is difficult [5]. Furthermore, most identified migraine risk variants reside in non-coding regions of the genome and are enriched within regulatory elements, indicating that their effects are mediated through modulation of gene regulation rather than alterations in protein structure or function [4, 6].

Expression quantitative trait loci (eQTLs) have been widely used to bridge GWAS findings with putative causal genes, including in migraine, where bulk-tissue and single-cell brain eQTL Mendelian randomization (MR) studies have nominated candidate causal genes [7–9]. However, eQTL–GWAS integration captures only the relationship between gene expression and disease, leaving the upstream regulatory mechanisms unresolved. Moreover, cis-eQTLs have been estimated to mediate only ~ 11% of complex trait heritability on average [10], a fraction substantially smaller than the heritability enrichment observed for variants within regulatory chromatin regions more broadly [11]. Consistent with this, recent work has shown that chromatin accessibility QTLs explain approximately 50% more GWAS loci than eQTLs in matched cellular contexts [12], suggesting that a substantial fraction of disease-associated regulatory variation acts at the chromatin level without producing detectable steady-state changes in mRNA levels.

Open chromatin regions (OCRs) encompass the regulatory sequences—including promoters, enhancers, insulators, and silencers—through which non-coding variants exert their effects on gene expression [13, 14]. Genetic variants influencing chromatin accessibility (caQTLs) can therefore directly identify the cis-regulatory elements through which disease-associated variants act, providing a layer of mechanistic resolution that is inaccessible to eQTL analysis alone [14–16]. Importantly, caQTLs in the human brain display a high degree of cell-type specificity—only ~ 10% are shared between neurons and non-neurons—highlighting the need for cell type–resolved caQTL data to dissect the regulatory architecture of brain disorders [15]. Despite the strong rationale for integrating caQTLs with disease genetics, no study to date has applied brain cell type–specific caQTLs to causal inference in migraine.

To address this gap, we extend the previous single-cell causal mapping of migraine from the transcriptional layer to the chromatin accessibility layer. Leveraging large-scale neuronal and non-neuronal cis-caQTL summary statistics from 616 postmortem brain donors generated by the PsychENCODE2 consortium, we performed two-sample MR using FinnGen R12 as the discovery cohort and the UK Biobank and Million Veteran Program migraine GWASs as two independent replication cohorts. Findings were further evaluated by Bayesian colocalization, instrument- and gene-level pleiotropy assessment, multi-modal target gene annotation, and regional brain expression characterization, to identify causal regulatory elements and candidate target genes for migraine.

Methods

Study design

We performed a two-sample Mendelian randomization (MR) study to evaluate the causal effects of brain cell type–specific chromatin accessibility on migraine risk. Genetic instruments were derived from cis-acting chromatin accessibility quantitative trait loci (caQTLs) of sorted neuronal and non-neuronal nuclei from human postmortem brain. Migraine genome-wide association study (GWAS) summary statistics from FinnGen Release 12 served as the discovery cohort, with two independent European-ancestry cohorts used for replication. Causal inference relied on the three core MR assumptions, the plausibility of which was evaluated through instrument strength (F-statistic), Bayesian colocalization, Steiger directionality testing, phenome-wide pleiotropy screening, and multi-modal target gene annotation. The study followed the STROBE-MR reporting guidelines [17]. As only publicly available summary-level data were used, no separate ethics approval was required.

Exposure data: brain cell type–specific cis-caQTLs

Cell type–specific cis-caQTL summary statistics were obtained from the PsychENCODE2 consortium, which generated ATAC-seq data from 1,932 fluorescence-activated nuclear sorting (FANS)–purified neuronal (NeuN⁺) and non-neuronal (NeuN⁻) samples derived from 616 postmortem brain donors across four cortical regions (dorsolateral prefrontal cortex, anterior cingulate cortex, superior temporal gyrus, and parahippocampal gyrus) [15]. Open chromatin regions (OCRs) and their associated cis-caQTLs were defined within ± 50 kb of each OCR midpoint, with multivariate Bayesian meta-analysis (mashR) performed within each cell type. For the present study, we used cell type–specific summary statistics for both neuronal and non-neuronal lineages.

Outcome data: migraine GWAS

Discovery analyses used migraine summary statistics from FinnGen Release 12 (endpoint G6_MIGRAINE; 26,894 cases / 374,605 controls; Finnish-ancestry) [18]. Replication was performed in (i) the UK Biobank whole-genome sequencing migraine GWAS (GCST90473326; 25,393 cases / 433,047 controls; non-Finnish European ancestry) [19] and (ii) the Million Veteran Program migraine GWAS (GCST90475837; 31,836 cases / 405,831 controls; European-ancestry US veterans) [20]. For an exploratory subtype-specific analysis, we additionally used the migraine-with-aura GWAS from the Million Veteran Program (GCST90477543; 4,445 cases / 441,492 controls; European ancestry); a corresponding migraine-without-aura GWAS was not available in this resource.

Instrumental variable selection

For each OCR, SNPs reaching genome-wide significance (P < 5 × 10⁻⁸) were selected and pruned using linkage disequilibrium clumping (r² < 0.001 within a 10,000 kb window; 1000 Genomes European reference panel). Instrument strength was assessed by the F-statistic, with all variants showing F > 10 retained. Exposure and outcome data were harmonized using the TwoSampleMR R package, removing palindromic SNPs with intermediate allele frequencies.

Mendelian randomization analysis

Causal effects were estimated using the Wald ratio for OCRs instrumented by a single SNP and the inverse-variance weighted (IVW) method for those instrumented by two SNPs. Cochran’s Q test was used to assess heterogeneity in IVW analyses. Steiger directionality testing was performed for all OCR–migraine pairs to confirm that the exposure explained more variance than the outcome. Statistical significance in the discovery cohort was determined using Bonferroni correction (P < 0.05/15,334 = 3.26 × 10⁻⁶), and OCRs reaching this threshold were carried forward to replication. Replication was considered successful if the same OCR achieved Bonferroni-corrected significance in each replication cohort independently (UKB: P < 0.05/19 = 2.63 × 10⁻³; MVP: P < 0.05/17 = 2.94 × 10⁻³), with concordant effect direction. All analyses were performed in R (version 4.4.3, RRID: SCR_001905); two-sample Mendelian randomization was conducted using the TwoSampleMR package (version 0.6.14, RRID: SCR_019010). Because the caQTL exposures are measured on a standardized scale of normalized chromatin accessibility, the resulting MR effect estimates (odds ratios per unit increase in normalized accessibility) are best interpreted as indicating the direction and approximate magnitude of effect rather than change per defined biological unit.

Bayesian colocalization

For OCRs achieving Bonferroni-corrected significance in both replication cohorts, Bayesian colocalization between the caQTL and migraine GWAS signals was performed using the coloc.abf function from the R package coloc [21] (version 5.2.3, RRID: SCR_026041). Variants located within ± 100 kb of the lead caQTL SNP were extracted from both summary statistics for each tested locus. Default prior probabilities were used: p₁ = 1 × 10⁻⁴ (probability of association with the exposure only), p₂ = 1 × 10⁻⁴ (probability of association with the outcome only), and p₁₂ = 1 × 10⁻⁵ (probability of joint association with both traits). Posterior probabilities for five hypotheses were estimated: PP.H0 (no association with either trait), PP.H1 (association with caQTL only), PP.H2 (association with migraine only), PP.H3 (both traits associated, distinct causal variants), and PP.H4 (both traits associated, shared causal variant). A PP.H4 > 0.8 was used as the threshold for strong evidence of colocalization, indicating a shared causal variant underlying both chromatin accessibility and migraine risk.

Pleiotropy assessment

Because all OCRs were instrumented by one or two SNPs, conventional MR-based pleiotropy tests such as MR-Egger and MR-PRESSO (which require ≥ 3 instruments per exposure) could not be applied. Instead, Horizontal pleiotropy of the IV SNPs underlying OCRs was evaluated using two phenome-wide association study (PheWAS) resources: the GeneATLAS database (UK Biobank, ~ 452,000 individuals, RRID: SCR_017577) [22] and FinnGen Release 12 (~ 500,000 individuals; >2,500 disease endpoints) [18]. Associations reaching genome-wide significance (P < 5 × 10⁻⁸) for phenotypes outside the migraine causal chain were flagged as potential violations of the exclusion-restriction assumption. OCRs whose instruments showed extensive cross-trait associations indicative of alternative causal pathways were excluded from downstream causal interpretation. Specifically, an OCR was excluded from downstream causal interpretation if its lead caQTL SNP showed genome-wide significant associations (P < 5 × 10⁻⁸) with phenotypes outside the migraine causal chain that formed a coherent alternative disease domain—multiple related endpoints converging on a single physiological system—indicating a systematic alternative pathway; OCRs whose only flagged associations were isolated, of small effect, or attributable to migraine itself (e.g., anti-migraine medication use) were retained, as the latter lie on the assumed causal pathway.

Target gene annotation

Candidate target genes of colocalized OCRs were nominated by integrating multiple resources from the SCREEN/ENCODE registry (RRID: SCR_015482; https://screen.wenglab.org/) [23, 24], including Hi-C contacts, cis-eQTL associations, and the Activity-by-Contact (ABC) and ENCODE-rE2G (regulatory Element-to-Gene) models. Brain biosamples were prioritized in all annotations.

Drug-target safety evaluation

To evaluate phenome-wide safety profiles of nominated target genes, we queried the AstraZeneca PheWAS Portal (https://azphewas.com) [25], which provides gene-level rare variant collapsing analyses across ~ 18,000 phenotypes in ~ 450,000 UK Biobank exome-sequenced participants. Phenome-wide significance was set at − log₁₀(P) > 8 and study-wide suggestive significance at − log₁₀(P) > 6, following the Portal’s default thresholds. Genes showing no associations reaching either threshold were considered to have no major adverse phenotypic associations indicative of pleiotropic safety concerns at the population level.

Regional gene expression

Bulk-tissue mRNA expression of nominated target genes across 13 anatomical regions of the adult human brain was retrieved from the Human Protein Atlas (HPA, RRID: SCR_006710, https://www.proteinatlas.org) [26], reported as normalized transcripts per million (nTPM) from RNA sequencing data.

Results

Genome-wide MR identifies 19 brain caQTL–migraine causal associations in the FinnGen R12 discovery cohort

Among 501,548 OCRs with available cis-caQTL data in the PsychENCODE2 dataset, 16,811 contained at least one cis-caQTL reaching genome-wide significance (P < 5 × 10⁻⁸); 15,334 were retained after LD clumping (r² < 0.001 within a 10,000 kb window), comprising 15,111 single-instrument OCRs and 223 OCRs instrumented by two SNPs, all with F-statistic > 10 (Figure S1). After harmonization with the FinnGen R12 migraine GWAS, 12,744 SNP–OCR pairs covering 12,607 unique OCRs were retained for downstream MR analyses (Table S1). Causal effects were estimated using the Wald ratio for single-instrument OCRs and the inverse-variance weighted (IVW) method for two-instrument OCRs. Nineteen OCRs reached Bonferroni-corrected significance (P < 0.05/15,334 = 3.26 × 10⁻⁶) in the discovery cohort, comprising 14 non-neuronal and 5 neuronal OCRs, with 12 risk-increasing and 7 risk-decreasing effects (Fig. 1A–B; Table S2). Cochran’s Q test for two-instrument OCRs is provided in Table S3, and Steiger directionality testing supported correct exposure-to-outcome directionality for all 19 Bonferroni-significant hits (Table S4).

Fig. 1.

Two-sample Mendelian randomization of brain cell type–specific cis-caQTLs identifies open chromatin regions prioritized as likely causal for migraine. (A) Manhattan plot of MR association results for neuronal (NeuN⁺) cis-caQTLs against migraine in the FinnGen Release 12 discovery cohort (26,894 cases / 374,605 controls). Each point represents one open chromatin region (OCR) tested as exposure, with the − log₁₀(P) of the Wald-ratio MR estimate plotted by chromosomal position. The dashed horizontal line indicates the Bonferroni-corrected significance threshold (P < 0.05/15,334 = 3.26 × 10⁻⁶). The five most significant OCRs are labelled with their peak identifier and genomic coordinates. (B) Manhattan plot of MR association results for non-neuronal (NeuN⁻) cis-caQTLs against migraine. (C) Forest plot of all 19 OCRs reaching Bonferroni-corrected significance in the FinnGen R12 discovery cohort. For each OCR, the peak identifier and genomic coordinates, MR method (Wald ratio for single-instrument OCRs), P value, and odds ratio (OR) for migraine per unit increase in normalized chromatin accessibility with 95% confidence interval are shown. Diamonds and horizontal bars represent point estimates and 95% confidence intervals, respectively. The vertical dashed line indicates OR = 1 (null effect)

Two independent migraine GWAS cohorts replicate seven OCR–migraine associations

The 19 Bonferroni-significant OCRs were carried forward for replication in two independent European-ancestry migraine GWASs. In the UK Biobank cohort, all 19 OCRs showed concordant effect directions with the discovery cohort, and 11 reached Bonferroni-corrected replication significance (P < 0.05/19 = 2.63 × 10⁻³) (Fig. 2A; Table S5); Steiger directionality testing supported correct exposure-to-outcome directionality across all tested OCRs (Table S6). In the Million Veteran Program cohort, 17 of 19 OCRs were testable, all with concordant effect directions, and 8 reached Bonferroni-corrected replication significance (P < 0.05/17 = 2.94 × 10⁻³) (Fig. 2B; Table S7); Steiger directionality testing again supported correct directionality across all tested OCRs (Table S8). Seven OCRs—six non-neuronal (Peak_7234 [chr1:115,282,554–115,283,246], Peak_8343 [chr1:156,480,770–156,481,078], Peak_124742 [chr4:35,524,861–35,526,079], Peak_163823 [chr6:96,449,901–96,451,022], Peak_202875 [chr9:116,482,445–116,483,426], and Peak_65776 [chr16:75,477,925–75,478,485]) and one neuronal (Peak_261762 [chr6:12,845,086–12,845,560])—reached Bonferroni-corrected replication significance in both UK Biobank and Million Veteran Program cohorts with concordant effect directions across all three cohorts (Fig. 2; highlighted in red), and were prioritized for downstream colocalization analysis.

Fig. 2.

Replication of 19 Bonferroni-significant OCR–migraine causal estimates in two independent migraine GWAS cohorts. Forest plots showing replication of the 19 open chromatin regions (OCRs) reaching Bonferroni-corrected significance in the FinnGen Release 12 discovery cohort in two independent European-ancestry migraine GWAS cohorts. (A) Replication in the UK Biobank migraine GWAS. (B) Replication in the Million Veteran Program migraine GWAS. For each OCR, the peak identifier and genomic coordinates, MR method (Wald ratio for single-instrument OCRs), P value, and odds ratio (OR) for migraine per unit increase in normalized chromatin accessibility with 95% confidence interval are shown. Diamonds and horizontal bars represent point estimates and 95% confidence intervals, respectively. The vertical dashed line indicates OR = 1 (null effect). OCRs highlighted in red reached Bonferroni-corrected replication significance in both the UK Biobank and the Million Veteran Program cohorts (n = 7), with effect directions concordant with the FinnGen R12 discovery cohort

Bayesian colocalization identifies three OCRs with shared causal variants

Bayesian colocalization between the cis-caQTL signals of the seven Bonferroni-replicated OCRs and the FinnGen R12 migraine GWAS revealed three OCRs with strong evidence of a shared causal variant (PP.H4 > 0.80): Peak_65776 (chr16:75,477,925–75,478,485; PP.H4 = 0.953), Peak_7234 (chr1:115,282,554–115,283,246; PP.H4 = 0.898), and Peak_124742 (chr4:35,524,861–35,526,079; PP.H4 = 0.811) (Table S9, Figure S2). The remaining four OCRs showed dominant PP.H3 signals (≈ 1.0), indicating distinct causal variants underlying the caQTL and GWAS associations at those loci, likely reflecting linkage disequilibrium between independent regulatory and disease-associated SNPs.

Instrument-level pleiotropy assessment excludes Peak_65776

Because all 19 Bonferroni-significant OCRs in our discovery cohort were instrumented by one or two SNPs, conventional MR-based horizontal pleiotropy tests such as MR-Egger and MR-PRESSO, which require at least three independent instruments per exposure, could not be applied. We therefore screened for potential violations of the exclusion-restriction assumption by performing instrument-level phenome-wide association screening of the lead caQTL SNPs underlying the three colocalized OCRs across the GeneATLAS (UK Biobank) and FinnGen Release 12 databases (Table S10–S11; Figure S3).

Applying this predefined rule, the lead caQTL SNP for Peak_65776 (rs4888422) met the exclusion criterion: it showed genome-wide significant associations (P < 5 × 10⁻⁸) forming a coherent, predominantly cardiovascular disease domain—coronary artery disease, myocardial infarction, hypertension, ischaemic heart disease, and angina pectoris—amounting to 17 distinct disease endpoints across the two databases. Rather than an isolated or borderline signal, this represents a mechanistically coherent cluster of associations converging on a single physiological system, each constituting a plausible independent pathway to the outcome. Whether these reflect direct pleiotropic effects or linkage disequilibrium with cardiovascular risk variants, both scenarios compromise the exclusion-restriction assumption of MR. Peak_65776 was therefore excluded from downstream causal interpretation despite its strong colocalization signal (PP.H4 = 0.953); because this locus (chr16) is independent of the two retained OCRs (chr1 and chr4), its exclusion does not affect the target-gene prioritization that constitutes our principal finding. Conversely, we note explicitly that retaining rather than excluding Peak_65776 would leave the NGF and TSPAN2 prioritization unchanged, as this peak resides on a separate chromosome (chr16) and is statistically independent of the two retained OCRs on chr1 and chr4; its removal is therefore a conservative safeguard that does not drive our principal conclusions.

In contrast, the lead caQTL SNPs for Peak_7234 and Peak_124742 met the retention criterion, showing no coherent alternative disease domain and phenotypic profiles consistent with the MR exclusion-restriction assumption. The lead caQTL SNP for Peak_7234 showed no associations with non-headache phenotypes outside the migraine causal chain; its only flagged associations were with migraine-related triptan medication use, a downstream consequence of migraine itself rather than an independent causal pathway, consistent with the directionality assumed by MR. The lead caQTL SNP for Peak_124742 showed only a single isolated genome-wide-significant association with a smoking-related phenotype of small effect (β = −0.009 on the standardized phenotype scale), which does not constitute a coherent alternative disease domain and is unlikely to introduce substantial bias to the MR estimate. Peak_7234 and Peak_124742 were therefore retained as final candidate regulatory elements for downstream target gene annotation.

Multi-modal regulatory annotation nominates candidate target genes for the two retained OCRs

To nominate downstream target genes of the two retained OCRs (Peak_7234 and Peak_124742), we integrated multi-modal regulatory annotations from the SCREEN/ENCODE registry, including candidate cis-regulatory element (cCRE) classification, Hi-C chromatin contacts, cis-eQTL associations, the Activity-by-Contact (ABC) model, and the ENCODE regulatory Element-to-Gene (rE2G) model.

For Peak_7234, the underlying regulatory element was annotated as a promoter-like cis-regulatory element (PLS; cCRE EH38E2832709), indicating proximity to a transcription start site and high promoter-like signatures across reference biosamples (Table S12). Multiple orthogonal lines of evidence converged on NGF as the highest-priority candidate target gene: (i) Hi-C chromatin contacts linked Peak_7234 to the NGF locus; (ii) cis-eQTL associations within the SCREEN registry identified NGF and the antisense long non-coding RNA NGF-AS1 as expression targets; (iii) the Activity-by-Contact (ABC) model further nominated NGF across multiple brain biosamples (posterior cingulate cortex, dorsolateral prefrontal cortex, and frontal cortex) (Table S13, maximum ABC score 0.08); and (iv) the ENCODE-rE2G model, which integrates DNase signal and 3D regulatory connectivity, prioritized NGF as the top-ranked candidate target (rE2G score 0.88, including the brain pericyte and dorsolateral prefrontal cortex biosample), with TSPAN2 nominated as a secondary candidate (rE2G score 0.38) (Tables S14). The convergence of cCRE, Hi-C, eQTL, ABC, and rE2G evidence across multiple brain biosamples therefore supports NGF as the highest-priority candidate downstream target of Peak_7234, with TSPAN2 retained as a complementary candidate.

For Peak_124742, although SCREEN cCRE annotation supported a distal enhancer-like signature (dELS), the underlying regulatory element lies within a gene-poor region of chromosome 4q13.2, with the nearest protein-coding gene located more than 200 kb away. Despite searching the same multi-modal regulatory layers, no Hi-C contact, cis-eQTL association, ABC link, or rE2G connection consistently nominated a specific target gene (Table S15). Peak_124742 was therefore retained as a candidate regulatory element with unresolved downstream target, representing a locus that may benefit from higher-resolution fine-mapping and functional validation in future studies.

Drug-target safety evaluation and regional brain expression of NGF and TSPAN2

Gene-level rare-variant collapsing analyses in the AstraZeneca PheWAS Portal showed no associations reaching either phenome-wide significance (− log₁₀(P) > 8) or study-wide suggestive significance (− log₁₀(P) > 6) for either NGF or TSPAN2 across ~ 18,000 phenotypes in ~ 450,000 UK Biobank exome-sequenced participants (Figure S4). Regional brain expression profiling using the Human Protein Atlas revealed that NGF was expressed at low but detectable levels across all 13 brain regions surveyed. In contrast, TSPAN2 showed marked enrichment in brain regions central to migraine pathophysiology, including the thalamus, pons, midbrain, hypothalamus, medulla, and amygdala (Figure S5).

Exploratory analysis in a migraine-with-aura subtype shows concordant effect directions

As an exploratory extension addressing the clinical heterogeneity of migraine, we applied the two retained OCRs to the migraine-with-aura GWAS using the same Wald-ratio and harmonization procedure as in the primary analysis. Both OCRs showed effect directions concordant with the overall migraine analyses. Peak_124742 (chr4:35,524,861–35,526,079) remained nominally associated with migraine with aura (OR = 1.14, 95% CI 1.00–1.29, P = 0.045), whereas Peak_7234 (chr1:115,282,554–115,283,246) showed a concordant but non-significant effect (OR = 0.92, 95% CI 0.83–1.01, P = 0.086); Steiger directionality testing supported the correct exposure-to-outcome direction for both OCRs (Tables S16 and S17). Given the markedly smaller case count for this subtype, these results should be interpreted with caution and regarded as hypothesis-generating.

Discussion

By integrating PsychENCODE2 brain cell type–specific cis-caQTLs with three independent migraine GWAS cohorts, our analyses extend cell type–specific causal mapping of migraine from the transcriptional to the chromatin accessibility layer. Conceptually, this shifts the unit of causal inference one step upstream of the standard GWAS-plus-eQTL-MR paradigm: instead of linking a risk locus to a gene through steady-state expression, we first identify the specific cis-regulatory element and the cell-type context through which non-coding variants act, and only then nominate downstream target genes—resolving a regulatory layer that eQTL-based mediation can miss when chromatin-level effects do not produce detectable changes in mRNA. Across 15,334 OCRs tested, we prioritize two non-neuronal cis-regulatory elements (Peak_7234 and Peak_124742) as likely causal contributors to migraine risk, with multi-modal regulatory annotation convergently nominating NGF and TSPAN2 as candidate downstream target genes at the chr1p13.2 locus. At the outset, we emphasize that our MR analyses primarily identify cis-regulatory elements as the unit of causal inference; the downstream target genes—NGF and TSPAN2—are computationally nominated through convergent multi-modal annotation rather than experimentally confirmed, and we therefore frame them throughout as prioritized, testable candidates rather than validated causal mediators.

At the chr1p13.2 locus, the underlying regulatory element Peak_7234 is classified by SCREEN as a promoter-like cCRE, and multi-modal evidence (Hi-C, cis-eQTL, ABC, and ENCODE-rE2G) converges on NGF as the highest-priority candidate downstream target gene. NGF encodes nerve growth factor, the founding neurotrophin and a central regulator of nociception, acting through the high-affinity TrkA receptor on peripheral sensory neurons—including the trigeminal ganglion neurons that mediate craniofacial pain in migraine [27–29]. Although NGF showed low cortical expression in the Human Protein Atlas, this expression pattern is consistent with the cell type origin of our caQTL signal: NGF is predominantly produced by non-neuronal cells in the brain—including pericytes, glia, and meningeal cells—and acts in a paracrine or systemic manner on TrkA-expressing peripheral neurons [28]. The ENCODE-rE2G nomination of NGF in a brain pericyte biosample provides further support for this non-neuronal regulatory origin. Notably, earlier GWAS have reported associations with migraine in the broader chr1p13.2 region [4], although the lead variants in those studies are distinct from the lead caQTL SNP underlying Peak_7234. Our analyses refine this statistical association to a specific regulatory element, identify its non-neuronal cell type origin, and prioritize NGF over TSPAN2 as the highest-priority candidate downstream target—dimensions that GWAS alone could not establish. Consistent with this mechanistic framework, elevated NGF levels in cerebrospinal fluid have been documented in patients with chronic daily headache, with positive correlations to substance P and CGRP [30]—biomarker-level evidence aligned with NGF involvement in trigeminal sensitization. Pharmacologically, the NGF–TrkA axis has been an actively investigated therapeutic target, with humanized anti-NGF monoclonal antibodies reaching late-stage clinical trials for chronic pain conditions [31]; our analyses provide independent human genetic support consistent with NGF as a candidate target relevant to migraine pathophysiology.

We acknowledge that the nomination of NGF and TSPAN2 represents a computational prioritization of candidate target genes based on the convergence of orthogonal regulatory annotations (cCRE classification, Hi-C, cis-eQTL, ABC, and ENCODE-rE2G), rather than an experimentally validated regulatory mechanism. These annotations infer, but do not directly measure, the regulatory connection between Peak_7234 and its target gene; establishing that the open chromatin region functionally controls NGF (or TSPAN2) expression and that this gene mediates migraine risk would require direct experimental perturbation in disease-relevant cell types, followed by assessment of target-gene expression and downstream phenotypes. As a concrete example, CRISPR interference (CRISPRi) targeting Peak_7234 in a disease-relevant non-neuronal cell model, followed by quantification of NGF and TSPAN2 expression and of nociceptive cellular phenotypes, would directly test whether this regulatory element controls the nominated target genes—the experiment that most clearly distinguishes our computational prioritization from mechanistic proof. The present findings should therefore be interpreted as prioritized, testable hypotheses for such functional studies rather than as established causal mechanisms.

The two prioritized OCRs showed modest but directionally consistent effect sizes across the three cohorts, with odds ratios ranging from approximately 1.09 to 1.14 for Peak_124742 (increased migraine risk) and from approximately 0.90 to 0.93 for Peak_7234 (decreased migraine risk). Effect sizes of this magnitude are characteristic of common-variant regulatory associations in migraine and other polygenic neurological disorders, in which each locus contributes only a small increment to overall liability; such modest per-locus effects are expected under a polygenic architecture and do not diminish the biological relevance of the implicated element. Indeed, a small genetically determined shift in the regulatory activity governing NGF expression is mechanistically consistent with the substantially larger therapeutic effect of direct pharmacological NGF–TrkA blockade, illustrating that a modest population-level effect size at a regulatory element can nonetheless mark a pathway of considerable biological and translational importance. Given that most OCRs were instrumented by one or two SNPs, however, the corresponding confidence intervals are wide, and these estimates (particularly the single-instrument Wald-ratio odds ratios, which carry no internal consistency check) should be interpreted as indicating the direction and approximate magnitude of effect rather than precise quantifications.

Although ENCODE-rE2G prioritized NGF over TSPAN2 (rE2G score 0.88 versus 0.38), TSPAN2 deserves substantial consideration as a complementary candidate at the same locus. TSPAN2 encodes a tetraspanin family member functioning as a marker of myelinating oligodendrocytes [32], a cellular identity directly aligned with the non-neuronal (NeuN⁻) origin of our caQTL signal. TSPAN2 further showed marked enrichment in brain regions central to migraine pathophysiology, including the thalamus, pons, midbrain, hypothalamus, medulla, and amygdala—an anatomical pattern that overlaps the trigeminothalamic and brainstem nociceptive circuitry implicated in migraine pathogenesis [29]. The lower rE2G score for TSPAN2 may partially reflect model-specific limitations, as ENCODE-rE2G is trained primarily on protein-coding gene CRISPRi-FlowFISH datasets and may underweight oligodendrocyte-specific regulatory connections. Together with the genetic-based prioritization of TSPAN2 [33, 34] and additional case-control validation in Han Chinese populations [35], our analyses do not exclude TSPAN2 as a causal mediator at this locus; rather, the convergence of two non-neuronal candidate genes—oligodendrocyte-marker TSPAN2 and pericyte/glia-derived NGF—at the same regulatory element points toward a unified non-neuronal regulatory architecture in which both genes may contribute to migraine pathophysiology in parallel or in concert.

Beyond chr1p13.2, Peak_124742 located in a gene-poor region of chromosome 4 (4q13.2) demonstrated a robust MR-estimated effect on migraine across all three cohorts and strong colocalization with the migraine GWAS signal (PP.H4 = 0.811). SCREEN cCRE annotation revealed that this 1.2-kb regulatory window contains two adjacent distal enhancer-like signature (dELS) elements, suggesting the presence of a clustered regulatory configuration rather than an isolated single enhancer. However, the surrounding genomic environment is gene-poor, with no protein-coding genes within 200 kb, and none of the multi-modal annotations (Hi-C, cis-eQTL, ABC, or ENCODE-rE2G) consistently nominated a specific target gene (Table S15). This combination of clustered enhancer-like signatures and the absence of nearby coding genes suggests that Peak_124742 may act over substantial genomic distance—a phenomenon increasingly recognized for distal enhancer clusters at complex disease loci [36]. The lack of an assigned target gene currently precludes mechanistic interpretation. Nonetheless, the robust MR-estimated effect across three independent cohorts, the clear colocalization signal, and the presence of dual dELS elements together nominate Peak_124742 as a candidate locus, warranting future investigation through high-resolution chromatin conformation capture, CRISPR-based perturbation, and finer-grained chromatin annotation in disease-relevant cell types.

Together, these findings extend cell type–specific causal mapping of migraine from the transcriptional to the chromatin accessibility layer, complementing prior brain cis-eQTL MR work in migraine—including our previous identification of inhibitory-neuronal BTBD16 and astrocytic RIMS1 as candidate causal mediators [7]. The convergence of multiple non-neuronal cell types as regulatory contributors—pericytes, glia, and oligodendrocytes—suggests that migraine risk architecture may be more substantially distributed across non-neuronal compartments than previously appreciated.

A further consideration concerns the use of chromatin accessibility as the MR exposure, a more recent extension of the molecular MR framework than eQTL- or pQTL-based designs. caQTL-based inference shares the three core instrumental-variable assumptions of established xQTL MR and has precedent within multi-omic summary-data-based MR pipelines that treat chromatin accessibility as one exposure layer alongside gene expression and protein abundance [37]. Because chromatin accessibility lies most proximal to the causal non-coding variant, caQTLs are strong cis-instruments; however, the exclusion-restriction assumption is more demanding than for a single-gene or single-protein exposure, as one regulatory element may influence multiple genes and one variant may be a caQTL for multiple peaks [14]. We addressed this through Bayesian colocalization and instrument-level phenome-wide screening, but note that with one or two instruments per peak, colocalization can establish a shared causal variant yet cannot, by itself, fully separate vertical from horizontal pleiotropy.

Interpretationally, a caQTL identifies a cis-regulatory element rather than a gene or protein: chromatin accessibility and steady-state expression are not perfectly coupled, and a fraction of caQTLs may not produce a detectable expression consequence. We therefore interpret our findings as the prioritization of cell type–specific regulatory elements and candidate downstream target genes—assigned through multimodal annotation rather than measured directly—rather than as definitive gene-level causal effects. These interpretational boundaries, together with the practical constraints arising from the sparse instruments available per open chromatin region, are considered in detail below.

Several limitations should be acknowledged. First, the great majority of OCRs were instrumented by only one or two SNPs, which constrains causal inference in three related respects. For pleiotropy assessment, regression- and outlier-based tests of horizontal pleiotropy (MR-Egger, MR-PRESSO) require multiple independent instruments per exposure and could not be applied. For sensitivity analyses, multi-instrument estimators (weighted-median, weighted-mode) and Cochran’s Q heterogeneity testing were likewise inapplicable, so the consistency of the causal estimate across instruments could not be interrogated. And for the robustness of causal inference more broadly, a single-instrument Wald-ratio estimate offers no internal means of separating a true causal effect from an undetected pleiotropic one. To mitigate these constraints, we shifted the burden of inference from within-MR tests to instrument-level and cross-cohort evidence: instrument-level phenome-wide screening of the lead caQTL SNPs (which led us to exclude Peak_65776), Bayesian colocalization to exclude confounding by distinct variants in linkage disequilibrium, Steiger directionality testing, and independent Bonferroni-corrected replication across two external cohorts with concordant effect directions. We therefore interpret our findings as the prioritization of candidate regulatory elements and target genes supported by this triangulation, rather than as definitive causal effects, with experimental perturbation as the necessary next step. Second, the brain cis-caQTLs were derived from cortical tissue and may not fully capture cell type–specific regulatory variation in trigeminal ganglion or meningeal vasculature, which are also implicated in migraine pathophysiology; complementary caQTL profiling in these tissues will further refine the regulatory architecture of migraine in future studies. Third, although multi-modal annotation prioritizes NGF as the most likely candidate target of Peak_7234, target gene nomination relies on biosample-level resources rather than direct functional perturbation; experimental validation through CRISPRi or related approaches in disease-relevant cell types will be required to confirm the causal regulatory link. Fourth, formal eQTL/pQTL-based triangulation and mediation analyses matched to the NeuN⁺/NeuN⁻ cellular context were not feasible with currently available resources, as existing brain eQTL data are either bulk-tissue or partitioned into transcriptomically defined cell types that do not correspond to the neuronal/non-neuronal lineage definition underlying our caQTL exposures; this constraint is further compounded for lowly expressed targets such as NGF, whose regulatory effects may act predominantly at the chromatin-accessibility level. As matched cell-type-resolved brain eQTL and brain pQTL resources become available, formal triangulation and mediation of the prioritized OCR→gene→migraine relationships—and assessment of their downstream therapeutic relevance—will be a valuable extension of the present work. Fifth, our analyses are based on European-ancestry data, and generalizability to other ancestries requires future caQTL and GWAS resources in diverse populations. Sixth, although we extended our framework to migraine subtypes, publicly available subtype-resolved GWAS resources remain limited and no matched migraine-without-aura GWAS was accessible, and the available subtype sample was substantially smaller than the overall migraine datasets, limiting statistical power; this analysis was therefore exploratory, and adequately powered subtype-specific resources will be needed to resolve subtype-specific regulatory mechanisms. Finally, although NGF and TSPAN2 showed no major safety concerns in rare-variant collapsing analyses at the population level, common-variant pleiotropic effects and tissue-specific consequences of pharmacological modulation require further evaluation before translational implications can be established.

Conclusion

Through two-sample Mendelian randomization integrating brain cell type–specific cis-caQTLs from PsychENCODE2 with three independent migraine GWAS cohorts, complemented by Bayesian colocalization, multi-layered pleiotropy assessment, and multi-modal target gene annotation, we prioritize two non-neuronal open chromatin regions as likely causal contributors to migraine risk: Peak_7234 (chr1:115,282,554–115,283,246), which prioritizes NGF as the most likely candidate downstream target and TSPAN2 as a complementary candidate, and Peak_124742 (chr4:35,524,861–35,526,079), a regulatory element in a gene-poor region whose downstream target awaits future fine-mapping. These findings extend cell type–specific causal mapping of migraine from the transcriptional to the chromatin accessibility layer, prioritize specific cis-regulatory elements and cell types underlying migraine pathophysiology, and provide a regulatory framework for further mechanistic dissection and the development of precision therapeutic strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

QX designed the study, performed the analyses, and drafted the manuscript. ZC contributed to data acquisition and replication analyses. YZ contributed to target gene annotation and biological interpretation. JS supervised the study and revised the manuscript. All authors read and approved the final manuscript.

Funding

This study was financially supported by grants from the National Natural Science Foundation of China (82071285), the Medical Science and Technology Project of Zhejiang Province (2022505533), and the Science and Technology Plan Project of Haixi Mongol and Tibetan Autonomous Prefecture, Qinghai Province (2026-JC-Q02).

Data availability

All datasets analysed during the current study are publicly available. Brain cell type–specific cis-caQTL summary statistics from the PsychENCODE2 consortium are available through the Synapse repository (https://www.synapse.org). Migraine GWAS summary statistics are available from FinnGen Release 12 (https://www.finngen.fi/en/access_results), the GWAS Catalog for the UK Biobank whole-genome sequencing migraine GWAS (accession GCST90473326; https://www.ebi.ac.uk/gwas/), and the GWAS Catalog for the Million Veteran Program migraine GWAS (accession GCST90475837). Phenome-wide association data were retrieved from the GeneATLAS database (http://geneatlas.roslin.ed.ac.uk), FinnGen Release 12 (https://r12.finngen.fi), and the AstraZeneca PheWAS Portal (https://azphewas.com). Multi-modal regulatory annotations were obtained from the SCREEN/ENCODE registry (https://screen.wenglab.org/). Regional brain expression data were retrieved from the Human Protein Atlas (https://www.proteinatlas.org). Additional supporting data are provided in the Supplementary Tables and Figures. The analysis code supporting this study is openly available at https://github.com/xuqiuhan/caqtl-migraine.

Declarations

Ethics approval and consent to participate

This study used only publicly available, de-identified, summary-level data from previously published genome-wide association studies and quantitative trait loci datasets. Original ethical approval was obtained by each contributing study from their respective institutional review boards or ethics committees, with informed consent obtained from all participants. No additional ethics approval was required for the present analyses.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.