Migraine: two attempts with public genetics

This is not medical advice and not a treatment. We tried twice to find something new about migraine with public data. The first attempt failed its own quality check. The second found one solid detail about a known gene.

Neither attempt is a breakthrough. The page shows both because the failure tells readers what these methods cannot do.

We combined 2,990 genes linked to migraine (Open Targets) with 18,854 drugs (DGIdb) and checked the ClinicalTrials.gov registry for novelty. The rules were written down and locked before any data was downloaded, with a check built in: the ranking had to bring back at least 2 of 3 known migraine drug classes in its top 30.

It brought back one. Topiramate ranked first. The CGRP drugs and the triptans, the modern core of migraine treatment, ranked far down. By the rule, we report no drug candidates.

Source: our attempt 1 report. The check needed at least 2 of 3 drug classes in the top 30 and found 1 of 3.

• Association scores are inflated by text mining, so broad GABA-A drugs (benzodiazepines, barbiturates, anaesthetics) filled the top of the table.
• The summary score rewards non-specific drugs: a drug with 40 weak targets collected more points than a triptan with 3 precise ones.
• The purely genetic signal is thin: with 509 genes of genetic evidence, 0 of 1,595 drugs passed correction for many tests (FDR below 0.05).

The general lesson holds beyond migraine: a naive gene-to-drug screen on public association scores cannot reproduce even known pharmacology, so it cannot be trusted to find new pharmacology.

Instead of drugs, genes were scored on genetic evidence, whether they can be targeted, and whether a drug exists. The map recovered real migraine biology: KCNK5 (TRESK), CACNA1A, ATP1A2, NOTCH3, TRPM8 and CALCA. Of 509 genes with genetic evidence, only 32 have an approved drug and 11 have already been tried in migraine. It listed 21 targets whose approved drugs did not appear in a registry search for migraine trials. That search was automated and is not a full literature review, so the list is a set of hypotheses, not confirmed gaps.

One example from that list is tanezumab (an anti-NGF antibody). Whether anti-NGF has been studied for migraine was not checked exhaustively against the primary literature, so this is a pointer for researchers and not a novelty claim. The drug was developed for osteoarthritis pain, and FDA advisory panels voted against its risk-mitigation plan in 2021, citing rapidly progressing osteoarthritis. Its developers stopped that program (Healio, Rheumatology Republic). The novelty check also caught one of our own hits: metformin already had a completed migraine prevention trial (NCT02593097, Mayo Clinic), so it is not new. The registry lists the trial as completed. Its result was not checked.

We tested whether migraine risk signals and gene activity signals in body tissues come from the same variant. This is colocalization, and we used our own implementation. We ran 5,834 tests (20 regions in 8 tissues), with migraine data from FinnGen (26,894 cases) and GTEx tissue data. The plan was locked before the analysis, with three validation checks.

PHACTR1

The strongest finding. In three arteries the migraine-protective variant lowers PHACTR1 activity in our analysis. It held in three independent FinnGen datasets, always in the same direction.

TSFM

A weaker lead in the aorta, repeated in one dataset but not in subtypes.

H2BC6-AS1

The weakest lead, repeated only for migraine with aura.

What was already known (checked in the full text of this 2017 paper): the PHACTR1 region is linked to migraine and to coronary artery disease. The variant rs9349379 allele G increases coronary disease risk and decreases migraine risk, so the two diseases go in opposite directions. That paper argues the effect may act through the nearby EDN1 gene, so PHACTR1 may not be the causal gene. Our contribution is the measured colocalization in arteries and its direction for PHACTR1 expression. EDN1 was not tested here.

Negatives: CACNA1A, a known gene for a rare migraine form, did not show colocalization in 5 tissues. The CGRP pathway genes were not among the regions tested, so the positive control for this method could not be run. 12 of 17 raw hits failed the calibration test and were dropped. One of the three validation checks failed as written and was replaced by a documented calibration. No drug candidate was found.

Attempt 2 is a preliminary written report: its full result tables are not published here and it has not been independently replicated, so no chart is shown for it.

• Only FinnGen (the full 102k-case data are not public) and 8 tissues, with no trigeminal ganglion.
• Computational hypotheses. A real result needs an experiment or a clinic.