There is a gap in genetics that almost nobody outside the field talks about. A genome-wide association study finds statistical signals: stretches of chromosome where people with a disease carry one letter more often than people without it. What it does not tell you is which gene in that stretch matters, whether it is turned up or down, in which tissue, or what it does when it gets there. A GWAS hands you an address. It does not tell you who lives there. Caroline Ran and Andrea Carmine Belin's group at the Centre for Cluster Headache at Karolinska Institutet have just published the visit.
The paper is Genome-wide association susceptibility loci for cluster headache support a role for inflammation in the pathophysiology, published in The Journal of Headache and Pain in 2026. It is open access, and it is worth reading in full if you have the appetite. What follows is my account of what they did, what they found, what they are prepared to claim, and where I think it sits against everything else we have been collecting.
The map they were handed
Cluster headache runs in families more often than chance allows. First-degree relatives of a patient carry somewhere between five and thirty-nine times the population risk depending on which study you read, and second-degree relatives around eight times. But it has never behaved like a single-gene disease, and for two decades the candidate gene studies produced a graveyard of associations that failed to replicate.
Three genome-wide studies changed that between 2021 and 2023: two European studies published back to back in Annals of Neurology in 2021, and a trans-ancestry meta-analysis in 2023 pooling 4,777 cases from ten European cohorts and one East Asian cohort. That meta-analysis produced eight risk loci, which resolve to nine candidate genes because one locus on chromosome 6 covers two genes sitting inside the same block of inherited DNA.
Positions are schematic, not to scale. Eight risk loci from the 2023 trans-ancestry meta-GWAS, resolving to nine candidate genes. MERTK was the strongest signal in all three genome-wide studies and in the meta-analysis. CAPN2 was originally found only in the Taiwanese cohort and never reached genome-wide significance in Europeans alone. Source: Winsvold et al. 2023; Ran et al. 2026.
Look at that list as a headache researcher in 2023 and it is not obviously encouraging. A phosphatase. A protease subunit. A folate-binding enzyme nobody has characterised. A developmental signalling molecule best known from cancer biology. A lipoprotein receptor. Three of them (FHL5, PLCE1, LRP1) overlap migraine loci, which is interesting but not explanatory. There is no obvious story. There is certainly no obvious drug.
That is the problem the new paper sets out to solve, and the method is refreshingly unglamorous: go and measure the things.
What they actually did
Three separate pieces of work, on samples from the cluster headache biobank at Karolinska.
Genotyping. They re-genotyped 542 patients and 296 controls from the original GWAS (the markers of interest had previously only been inferred statistically, not directly read), and added a fresh replication cohort of 165 patients recruited between 2021 and 2025 against 506 new controls. They tested three specific variants that earlier analysis had flagged as likely to change protein function, in MERTK, FHL5 and PLCE1.
Methylation. They took whole blood from 62 patients and 26 controls and used pyrosequencing to read the methylation state of three CpG sites in MERTK, around a position that earlier computational work had predicted would be chemically modified in cluster headache.
Gene expression. This is the part that matters most. They measured how much messenger RNA each candidate gene was actually producing, in two human tissues: fibroblast cell lines grown from skin biopsies (12 patients, 9 controls) and white blood cells drawn fresh (20 patients, 22 controls).
Why fibroblasts? Because you cannot biopsy a living person's trigeminal ganglion. Skin fibroblasts are a standard workaround in chronobiology and headache research: they carry the patient's genome, they run their own molecular clock, and they can be serum-shocked in a dish to synchronise that clock. They are not neurons. The authors are explicit that this is a limitation, and I will come back to it.
The two variants that held up
Of the three functional variants tested, two survived meta-analysis across both Swedish cohorts and one did not.
Values exactly as published in Fig. 1 of Ran et al. 2026. Diamonds are the pooled estimate. A confidence interval that crosses 1.0 means the result is compatible with no effect. MERTK used a random effects model because the two cohorts disagreed substantially (I² = 65.7%). PLCE1 did not separate from the null.
Two things are worth pausing on here.
The FHL5 result is the cleaner one. Both cohorts point the same way, there is no heterogeneity at all between them, and the pooled odds ratio is 1.29 with the interval comfortably clear of 1.0. The minor allele raises risk by around a third. That is a modest effect, which is what you expect from a common variant in a polygenic condition, and it is a genuine independent replication in a cohort that did not exist when the GWAS was run.
The MERTK result is messier and the authors say so. The original GWAS cohort showed a strong protective effect of the A allele (0.60), the new replication cohort showed a much weaker one that did not reach significance on its own (0.81), and the pooled random effects estimate lands at 0.69. The two cohorts genuinely disagree about the size of the effect. That is not fatal, but it is a caveat, and it matters because MERTK has been the headline gene of cluster headache genetics since 2021.
PLCE1 did not replicate, and the gene was not detectable in either tissue they could sample. It stays unresolved rather than refuted.
The methylation result, in proportion
Earlier computational work predicted that a specific chemical tag would sit on MERTK in cluster headache. The team went and read three CpG positions in that region. Two showed no group difference. The third showed a difference that was statistically significant after correction: 89.2% methylated in patients against 87.9% in controls.
That is a real result and I am not going to inflate it. It is a 1.3 percentage point difference, at an intronic site, in whole blood, and the authors themselves write that predictions about what such a change does "cannot be made". What it establishes is narrow and still useful: the computational prediction was pointing at the right neighbourhood. Nothing more.
The expression data, which is the real payload
This is where the paper earns its title. Statistical association tells you a letter of DNA travels with a disease. Expression tells you whether the gene is doing more or less work in the people who have it.
Group medians read from the box plots in Fig. 3 of Ran et al. 2026, drawn as slopes. The published figure shows full distributions; this redraw keeps the direction and the significance and drops the spread, so read it as direction of travel rather than effect size. WNT2 is omitted: it was not detectable in blood and showed no difference in fibroblasts. PLCE1 was not expressed in either tissue. P-values are as reported after false discovery rate correction.
Look at the top panel. In white blood cells, four of the five measurable candidate genes are turned down in cluster headache, and three of them at P below 0.001, which for sample sizes of 20 and 22 means the separation is not subtle. CAPN2 is down in both tissues, the only gene consistent across blood and skin, and it shows the largest difference in the study.
There is one more finding in this section that gets less attention than it deserves. In blood, the expression levels of DUSP10, CAPN2, UFL1 and LRP1 were strongly correlated with one another across individuals. Whatever is turning them down appears to be turning them down together. That is the signature of a shared upstream regulator, not four independent coincidences.
The scoreboard
| Gene | Result in this study | Status |
|---|---|---|
| MERTK | Protective variant replicated in meta-analysis; methylation difference at one CpG site; previously shown to be raised in patient tissue | validated |
| FHL5 | Risk variant replicated cleanly in an independent cohort; risk allele lowers FHL5 expression in arteries | validated |
| UFL1 | Lower in blood, P<0.001; shares the chromosome 6 locus with FHL5 and matches the predicted direction | validated |
| DUSP10 | Lower in blood, P<0.001; matches the direction predicted from the GWAS variant | validated |
| CAPN2 | Lower in blood and in fibroblasts, both P<0.001; largest effect in the study | validated |
| LRP1 | Lower in blood, P=0.007; direction does not match the mixed eQTL data in public databases | validated, with a caveat |
| FTCDNL1 | Higher in fibroblasts, P=0.004; unchanged in blood; almost nothing is known about what this gene does | validated, unexplained |
| WNT2 | Not detectable in blood; no difference in fibroblasts, but the sample was small | not validated |
| PLCE1 | Variant did not replicate; gene not expressed in either available tissue | not validated |
Seven of nine, in human tissue, from a list that three years ago was nothing but coordinates.
Then they noticed what the survivors had in common
Here is the turn in the paper. The four genes that moved together in blood, DUSP10, CAPN2, UFL1 and LRP1, plus MERTK, are not a random assortment of immune genes. They all regulate the same thing.
The NLRP3 inflammasome is a protein complex that assembles inside immune cells and acts as an alarm. When it fires, it activates caspase-1, and caspase-1 does two things: it cuts interleukin-1β and interleukin-18 into their active, inflammatory forms, and it triggers pyroptosis, a violent form of cell death that ruptures the cell and spills its contents into the surrounding tissue. It is one of the most consequential switches in innate immunity, and it is heavily restrained, because a cell that fires it by accident causes real damage.
Five of the seven validated cluster headache genes sit on that restraint system.
Wiring redrawn from Fig. 4 of Ran et al. 2026, which they label a hypothesis. Press the button to overlay the direction of change measured in patient tissue. Four of the five restraints are transcribed at lower levels in cluster headache. MERTK is the exception and is discussed below.
Follow the individual arms and the logic holds up.
- CAPN2 is down. Calpain-2 suppresses transcription of calpain-1, and calpain-1 activates NLRP3. Remove calpain-2 and calpain-1 rises. This is not speculative wiring: it comes from work showing that deleting CAPN2 released calpain-1 driven cytokine production.
- DUSP10 is down. Its job is to dephosphorylate p38 and JNK, two pro-inflammatory kinases that feed NLRP3. Less phosphatase, less braking.
- LRP1 is down. LRP1 inhibits NF-κB, and NF-κB is the transcription factor that supplies the priming signal NLRP3 needs before it can fire at all.
- UFL1 is down. This one is genuinely ambiguous and the authors say so: in one cell type suppressing UFL1 makes NLRP3 activation worse, in another UFL1 enhances the priming signal. It moves, but the direction of consequence is not settled.
Four brakes, transcribed at lower levels, in the same people, correlated with each other.
These results confirm the validity of seven candidate genes for cluster headache identified through GWAS. Five of the candidate genes are abundantly expressed in both immune and blood cells, converge functionally on inflammatory signalling pathways and suggest a potential role for inflammasomes in cluster headache.
Ran et al., The Journal of Headache and Pain, 2026What does not fit, and what the authors cannot answer
I want to be careful here, because this is the part that usually gets dropped when a paper gets written up, and it is the part that decides whether the finding survives.
The MERTK direction is not reconciled
MERTK inhibits NF-κB, which makes it anti-inflammatory in this diagram. But the risk allele raises MERTK expression, and the group's earlier work found MERTK messenger RNA elevated in cluster headache tissue. More of an anti-inflammatory brake, in a disease the paper argues is under-braked. The authors offer a partial route out, that lower DUSP10 removes negative control from the MERTK/ERK arm, but they do not claim to have resolved it, and neither will I. It is an open inconsistency in the middle of the central argument.
There is a directly contradictory cluster headache study
In 2022, a group in Istanbul measured inflammasome components in peripheral blood in cluster headache and concluded the opposite: that the complex is suppressed and does not appear to play a role. That paper is cited in this one. It is worth knowing that it rested on four patients with chronic cluster headache. Four. That is a sample from which almost nothing can be concluded in either direction, and I would not use it to dismiss the present findings, but nor should anyone pretend the CH-specific inflammasome literature currently points one way.
The tissue is not the tissue
Everything measured here came from blood and skin. The disease happens in the trigeminal ganglion and the hypothalamus. The authors are direct about this: peripheral cells "do not directly translate into physiological events observed in headache pathophysiology and specifically neuroinflammation". They note that blood and skin already showed different expression profiles from each other, which is itself a warning that a third tissue could show a third pattern.
They do offer one anatomical argument in mitigation that I think is the strongest sentence in the discussion: the trigeminal ganglion sits outside the blood-brain barrier. Whatever is circulating in the periphery reaches it. A peripheral inflammatory signature is not automatically irrelevant to a structure that is bathed in peripheral blood.
The numbers are small
Twenty patients in the blood expression analysis. Twelve fibroblast lines. The authors list this first among their limitations, and note that small samples both miss real effects and exaggerate the ones they find. WNT2 in particular should not be written off on this evidence. Every participant in the expression work was diagnosed by a neurologist, which is a real strength, but it does not manufacture statistical power.
What this study does not show. It does not show that NLRP3 is activated in cluster headache. Nobody has measured that. It does not show that inflammation causes attacks. It does not show that anti-inflammatory treatment would help. It shows that seven genes flagged by association studies are measurably altered in patient tissue, and that most of them regulate the same switch. That is a finding about where to point the next experiment.
Where this lands against everything else we have
Everything from here on is my reading rather than the authors', and I will mark it clearly. It draws on the living analysis we prepared at the Cluster Headache Research Hub at the end of April, which synthesised roughly 3,950 primary papers across twelve mechanistic domains, alongside a community corpus of 2,637 posts and 31,116 comments.
That analysis listed ten open questions. Question three read: what is the full polygenic architecture of cluster headache? This paper does not answer it. But it does something the analysis did not anticipate, which is to take the architecture we already had and give it a single functional address.
Because here is what struck me when I finished reading. The NLRP3 inflammasome is not a new arrival in our model. It is the point where six of our layers were already converging, and none of them arrived there through genetics.
This figure is my synthesis, not the authors'. Each spoke is an established mechanism in its own literature, cited below. None of them is demonstrated in cluster headache tissue. What the diagram shows is convergence of independent lines of evidence on one node, which is a reason to investigate, not a finding.
My reading, not the authors'
Take the six spokes one at a time.
Vitamin D. The vitamin D receptor is a direct negative regulator of NLRP3. It physically binds NLRP3 and blocks BRCC3 from removing the ubiquitin tags that keep the complex switched off; without VDR, caspase-1 activation and IL-1β release rise. Set that beside what our corpus already holds: a Grade A meta-analysis of lower serum 25(OH)D in cluster headache patients, mechanistic coherence across five domains, and fifteen years and 85,000 downloads of a patient-developed D3 regimen with no controlled trial behind it. For the first time there is a specific molecular endpoint that a trial could measure rather than just counting attacks. That is not evidence the regimen works. It is a place to look for a reason if it does.
The clock. This is the connection I find hardest to dismiss. NLRP3 is transcriptionally gated by the core clock: deleting Bmal1 or Nr1d1 (REV-ERBα) drives enhanced NLRP3 activation through loss of transcriptional control of Nlrp3 and Il1b themselves. REV-ERBα binds the Nlrp3 promoter directly. And the same Karolinska group published, this year, a study finding genetic variation in core clock genes including BMAL1 in cluster headache. So we have one group reporting clock gene variants in cluster headache, and the same group reporting inflammasome regulator variants in cluster headache, and the independent immunology literature saying those clock genes control that inflammasome. Cluster headache's two most conspicuous features, its inflammatory character and its brutal timekeeping, may not be two findings that need reconciling. They may be one mechanism seen from two angles.
Hypoxia. Intermittent hypoxia activates NLRP3 through ROS and NF-κB signalling, and monocytes from patients with severe obstructive sleep apnoea show higher NLRP3 activity that tracks with the apnoea-hypopnoea index. Sleep apnoea is over-represented in cluster headache; attacks lock to sleep and to first REM; a 2024 review framed the disorder as an overzealous response to hypoxia. That is our hypoxia layer arriving at the same switch by a completely separate route.
The gut. Lipopolysaccharide is the canonical priming signal for NLRP3, and it is the same molecule our microbiome layer proposes leaks from a dysbiotic gut. Ran and colleagues note that the direct activating pathways of NLRP3 demonstrated so far run through LPS stimulation.
CGRP. This is the one that closes the loop back to the attack itself. In the trigeminal ganglion, spreading depolarisation raises CGRP expression, and that rise is suppressed by inhibiting NLRP3. In a nitroglycerin model, blocking NLRP3 or IL-1β reduced hyperalgesia and lowered CGRP in the trigeminal nucleus caudalis. Nitroglycerin, note, is also a reliable provoking agent for cluster headache attacks. CGRP is the single most pharmacologically validated target in the disease. If NLRP3 sits upstream of CGRP release in the trigeminal system, then a genetic predisposition toward an under-restrained inflammasome and the effector molecule we already drug are on the same wire.
I want to state the limits of that argument plainly, because it would be easy to read it as more than it is.
Every one of those six links is real in its own literature. Not one of them has been demonstrated in cluster headache tissue. The vitamin D work is in cell and mouse models of other conditions. The clock and inflammasome work is in macrophages and microglia. The CGRP work is in rodent migraine models. Convergence across independent literatures is a reason to take a hypothesis seriously; it is not evidence for it. A model that explains everything explains nothing until somebody tries to break it.
And there is a specific reason for caution that I cannot argue my way past. The one measurement anybody has made of inflammasome components in cluster headache patients found them suppressed. It was four patients and it should not be treated as decisive, but it is the only direct data point that exists, and it points the wrong way for the story I have just told.
What would settle it
Very little of this needs new technology. It needs somebody to measure the right things in the right tissue.
NLRP3, ASC and caspase-1 have never been properly quantified in a decent-sized cluster headache cohort, in and out of bout. The Danish and Swedish biobanks that produced the recent cytokine work already hold the samples. IL-1β has been measured and the results conflict: lower in episodic patients in bout in the Danish cohort, higher in chronic patients, higher during attacks in another study. Nobody has yet asked whether that scatter resolves once you stratify people by which of these seven variants they carry.
The trigeminal ganglion question is harder but not impossible. Post-mortem tissue exists. The group have already localised MERTK and its ligand galectin-3 in rat trigeminal ganglion, so the assays are built.
And the cleanest test of the whole convergence idea is the least glamorous one. If NLRP3 restraint really is the shared node, then the people who respond to vitamin D, the people who respond to treating their sleep apnoea, and the people whose attacks track hardest to the clock should not be three unrelated subgroups. They should overlap, and they should overlap in a way that maps onto genotype. That is answerable with a cohort, a questionnaire and a saliva kit.
None of that is a prediction about what the answers will be. I do not know, and anyone claiming to on the current evidence is guessing. It is a list of experiments that would tell us whether the pattern in this paper is a mechanism or a coincidence.
The part that is not about mechanism
There is a version of this write-up that stops at the science. I want to add one thing that is not scientific but is, I think, true.
For most of the last century, cluster headache patients were told a story about themselves. The vascular theory, then the personality theory, then the long stretch where the honest answer was that nobody knew. What has happened over the past five years is that a disease with no molecular account has acquired one, piece by piece, from people who went and measured things: eight loci, then the clock genes, then the cytokines in cerebrospinal fluid, and now seven genes read out of the blood and skin of people who volunteered for a biobank in Stockholm.
That biobank is the reason this paper exists. Five hundred and forty-two people gave blood for the original genotyping. Twelve people let somebody take a punch of skin so that their fibroblasts could be grown in a dish and shocked with serum at a fixed hour. Those twelve people are the entire fibroblast dataset in Figure 3. If you have ever wondered whether it is worth answering the recruitment email, that is the answer.
Sources
- Ran C, Deborgies Sanches C, Swedblom J, et al. Genome-wide association susceptibility loci for cluster headache support a role for inflammation in the pathophysiology. The Journal of Headache and Pain. 2026;27:198. doi:10.1186/s10194-026-02485-x (open access)
- Winsvold BS, Harder AVE, Ran C, et al. Cluster headache genomewide association study and meta-analysis identifies eight loci and implicates smoking as causal risk factor. Ann Neurol. 2023;94:713-726.
- O'Connor E, Fourier C, Ran C, et al. Genome-wide association study identifies risk loci for cluster headache. Ann Neurol. 2021;90:193-202.
- Harder AVE, Winsvold BS, Noordam R, et al. Genetic susceptibility loci in genomewide association study of cluster headache. Ann Neurol. 2021;90:203-216.
- Edvinsson JCA, Ran C, Olofsgård FJ, et al. MERTK in the rat trigeminal system: a potential novel target for cluster headache? J Headache Pain. 2024;25:85.
- Şahin E, Karaaslan Z, Şanlı E, et al. Reduced expression of inflammasome complex components in cluster headache. Headache. 2022;62:967-976.
- Lund NLT, Westgate CSJ, Søborg MLK, et al. Distinct alterations of inflammatory biomarkers in cluster headache: a case control study. Ann Neurol. 2025;98:4-18.
- Ran C, Olofsgård FJ, Wellfelt K, et al. Elevated cytokine levels in the central nervous system of cluster headache patients in bout and in remission. J Headache Pain. 2024;25.
- Sanches CD, Spulber S, Olofsgård FJ, et al. Genetic variability within molecular core clock genes in cluster headache. Cephalalgia. 2026;46.
- Rao Z, Chen X, Wu J, et al. Vitamin D receptor inhibits NLRP3 activation by impeding its BRCC3-mediated deubiquitination. Front Immunol. 2019;10:2783.
- Pourcet B, Zecchin M, Ferri L, et al. Nuclear receptor subfamily 1 group D member 1 regulates circadian activity of NLRP3 inflammasome to reduce the severity of fulminant hepatitis in mice. Gastroenterology. 2018;154:1449-1464.
- Kou L, Chi X, Sun Y, et al. The circadian clock protein Rev-erbα provides neuroprotection and attenuates neuroinflammation against Parkinson's disease via the microglial NLRP3 inflammasome. J Neuroinflammation. 2022;19.
- Díaz-García E, García-Tovar S, Alfaro E, et al. Inflammasome activation: a keystone of proinflammatory response in obstructive sleep apnea. Am J Respir Crit Care Med. 2022;205:1337-1348.
- Chen PY, Yen JC, Liu TT, Chen ST, Wang SJ, Chen SP. Neuronal NLRP3 inflammasome mediates spreading depolarization-evoked trigeminovascular activation. Brain. 2023;146:2989-3002.
- He W, Long T, Pan Q, et al. Microglial NLRP3 inflammasome activation mediates IL-1β release and contributes to central sensitization in a recurrent nitroglycerin-induced migraine model. J Neuroinflammation. 2019;16:78.
- Kursun O, Yemisci M, van den Maagdenberg AMJM, Karatas H. Migraine and neuroinflammation: the inflammasome perspective. J Headache Pain. 2021;22:55.
- Cluster Headache Research Hub. CH Pathogenesis Convergence Analysis, 30 April 2026. Internal living synthesis: ~3,950 primary papers, 12 domains, community corpus of 2,637 posts and 31,116 comments.