Narrowing 1,091 Metabolites Down to One Enzyme in MS

Why multiple sclerosis (MS) runs a mild course in one person and a severe one in another remains largely unexplained, and metabolomics has been an appealing avenue because metabolites reflect the combined output of genetic, environmental and immune processes. The difficulty is that observational metabolite associations cannot separate cause from consequence, a limitation the authors name directly with reference to their own earlier observational work. Noroozi and colleagues built a layered Mendelian randomization framework to address it. They tested 1,091 plasma metabolites and 309 metabolite ratios, measured by mass spectrometry in 8,299 participants of the Canadian Longitudinal Study on Aging, against the age-related MS severity score in 12,584 people with MS from the International MS Genetics Consortium, a cohort enriched for older patients with longer disease duration to make the disability phenotype more stable. The work follows the STROBE-MR reporting guidelines. It is a medRxiv preprint that has not been through peer review.

A Design Built to Narrow Rather Than to Scan
The instrument criteria are set out in full: a minor allele frequency above 0.01, SNP-based heritability above 5% for a metabolite to enter at all, association with circulating levels at P < 1×10⁻⁵, which the authors describe as a relaxed threshold chosen because the original metabolite GWAS sample size was modest, linkage clumping at r² = 0.01 within a 10,000 kb window, exclusion of the major histocompatibility region, and at least three instruments per metabolite. In practice each metabolite carried between 12 and 55 instruments, all with F-statistics above 10. Significance was defined at P < 0.05/73, or 6.8×10⁻⁴, using a previously estimated effective number of 73 independent metabolites. The primary screen returned 45 metabolites reaching nominal significance, 36 of them annotated and nine uncharacterized compounds, and every subsequent layer was built to remove candidates from that list rather than add to it.

The Sensitivity Work Comes Before the Conclusions
Of the 45, forty showed concordant effect directions under MR-Egger. Cochran's Q found evidence of heterogeneity for only two metabolites, glycochenodeoxycholate glucuronide and 1-stearoyl-2-docosahexaenoyl-GPC, and both were excluded from the multivariable analysis. MR-PRESSO flagged a single outlier variant for one of those two, and removing it did not change the significance of the association. Steiger filtering confirmed that the chosen instruments explain more variance in the metabolite than in MS severity, which is what validates the assumed direction rather than assuming it. Reverse Mendelian randomization found no evidence of reverse causation for any association except decadienedioic acid. And to address confounding directly, the authors checked whether instruments were associated with body mass index, smoking or years of schooling at genome-wide significance, queried a 250 kb window around each for high linkage with such variants, and report that excluding them did not materially alter the estimates.

Multivariable Analysis Sharpens the Lipid Signal
Because metabolites within a class are correlated, the authors ran multivariable models including all significant amino acids in one model and the lipids of the fatty acid sub-pathway in another, to estimate each metabolite's direct effect. Five survived with independent effects. Arachidonate (20:4n6) came in at β = −0.07 [−0.1, −0.04] with P = 7.78×10⁻⁷, and betaine at β = −0.07 [−0.1, −0.04] with P = 1.56×10⁻⁵, both associated with lower severity. Arginine went the other way at β = 0.07 [0.03, 0.1], P = 9.13×10⁻⁴, with propionylglycine and succinoyltaurine weaker. What matters about this step is the direction of travel: adjusting for correlation between related metabolites sharpened the lipid signal rather than dissolving it, which is the opposite of what happens when a screen result is an artifact of correlated instruments.

Colocalization, and What the Authors Say It Cannot Separate
To test whether the metabolite and severity signals are driven by the same variant rather than by correlated neighbours, colocalization was run at the enzyme-coding loci. On chromosome 11 the posterior probability for a shared causal variant exceeded 0.70 across three arachidonate-containing ratios. The authors are careful about what follows, because this is a well-established pleiotropic locus containing several overlapping genes: the signal colocalizes with FADS1, FADS2, MYRF and TMEM258 alike. rs174564 emerged as the putative shared causal variant for FADS1/2, while rs174537 and rs102274 came up for MYRF and TMEM258, and those sit in near-complete linkage disequilibrium with rs174564 at D′ = 1.0 and R² = 0.99. They state that they prioritized FADS1/2 on biological plausibility, since these encode the rate-limiting enzymes of polyunsaturated fatty acid metabolism, and then built the next layer to test that choice rather than assert it.

A Functional Variant, Followed Into Brain Cell Types
The variant chosen to proxy FADS1 perturbation is rs174546 C>T, in the gene's 3′ untranslated region, where the T allele has been shown to create a miR-149-5p binding site producing allele-specific repression. Expression data show the TT genotype with reduced FADS1 across multiple brain regions, liver and whole blood, while FADS2, MYRF and TMEM258 show no significant reduction and are modestly increased. Europeans carrying the T allele have a consistent fatty acid profile of raised Δ6-desaturase products and lowered Δ5-desaturase products, with experimental evidence of reduced enzymatic activity. Used as a single instrument, the metabolite pattern proved coherent across the biosynthetic cascade: upstream intermediates accumulating, downstream products falling, and greater MS severity at each step. The same variant was then used as an expression instrument across seven brain cell types in 474 individuals, where reduced FADS1 was associated with greater severity in astrocytes (β = 0.153, P = 1.3×10⁻²), oligodendrocytes (0.112, 1.2×10⁻²), excitatory neurons (0.101, 9.4×10⁻³) and inhibitory neurons (0.0766, 7.8×10⁻³), but not in endothelial cells, microglia or oligodendrocyte precursor cells, with colocalization supporting a shared causal variant in each relevant cell type.

A Second Locus, and What the Authors Say Comes Next
A separate signal appeared at CYP4F2, where colocalization between circulating succinoyltaurine and MS severity was moderate, with a posterior probability above 0.4. The shared variant is rs2108622, c.1297G>A (p.Val433Met), a missense change that lowers CYP4F2 protein levels rather than altering transcription or enzymatic activity, and genetically proxied reductions in that activity tracked with greater severity. CYP4F2 encodes a cytochrome P450 ω-hydroxylase that metabolizes long- and very-long-chain fatty acids including arachidonic acid, and prior genetic work has connected it to MS severity, optic neuritis susceptibility and fingolimod metabolism. The translational reading the authors offer is specific rather than general: people carrying low-activity FADS1/2 haplotypes may represent a subgroup in whom targeted dietary or supplementation approaches could be worth testing. Their stated limitations point the same way. The data come from individuals of European ancestry, and confirmation across diverse populations is needed; sex- and age-stratified analyses are necessary to define effects within subgroups; and the results warrant further validation through clinical and experimental research. As a preprint, it carries the note that it should not be used to guide clinical practice.

Disclaimer: This blog post is based on the cited preprint and is intended for informational purposes only. The preprint has not been certified by peer review and should not be used to guide clinical practice. It is not intended to provide medical advice. Please consult with a healthcare professional for any health concerns.

Reference:
Noroozi, R., Higgins Tejera, C., Chen, M., Briggs, F. B. S., Bhargava, P., & Fitzgerald, K. C. (2026). Integrative genetic analyses of lipid metabolism and multiple sclerosis severity using metabolome-wide and cis-Mendelian randomization. medRxiv. https://doi.org/10.64898/2026.05.27.26354239