Three Statistical Models Converge on GRIN2A in MS Brain Volume
Genome-wide studies have been good at finding who develops multiple sclerosis (MS). They have been much less successful at explaining why the disease looks so different from one patient to the next once it starts. Strijbis, Inkster and colleagues took up that second question, and they did so with a method designed for its particular difficulty: MRI phenotypes in MS are several correlated measurements at once, not one outcome, and the standard approach of testing one variant against one measurement at a time handles that badly. They applied sparse reduced-rank regression, a multivariate penalised method, to 326 patients from the GeneMSA consortium, 262 with relapsing-remitting and 64 with secondary progressive disease, scanned at the University Hospital in Basel and the VU University Medical Centre in Amsterdam. The genes chosen were those involved in glutamate signalling, because glutamate biology sits at the junction of the neuroinflammation and neurodegeneration that MS combines. The authors describe the work as a first exploratory study and ask for independent confirmation, and that is the right way to read it.
Why a Multivariate Method, and How They Guarded It
Mass-univariate linear modelling fits every variant-phenotype pair separately, which with thousands of variants and several phenotypes means a correction for every test and no way to capture effects that several variants produce together. Sparse reduced-rank regression models all variants and all phenotypes jointly, treats the problem as variable selection rather than hypothesis testing, and ranks variants by how often they are selected across repeated resampling of the data. Earlier simulation work had shown it more powerful than the univariate approach. The design choice worth crediting is what the authors did next: rather than trusting a new method on its own, they ran it alongside Lasso regression and conventional univariate modelling, and set agreement across all three as their criterion for a result being reliable. A signal that only one method finds is treated as weaker than one that all three reach independently.
What the Multivariate Model Selected
The analysis covered 3,810 variants: 1,292 from 34 glutamate-related genes, 2,517 control variants chosen specifically because they showed no association with brain volume, and a tag for HLA-DRB1*15:01. Six variants in GRIN2A, the gene encoding the 2A subunit of the NMDA receptor, reached a selection probability of 0.50 or higher, led by rs3859123 at 0.64. Variants in GRM7, GRM1 and GRIK4 followed just below. The phenotypes the model selected are as informative as the genes: normalized brain volume at 0.815, normalized grey matter volume at 0.757, and normalized white matter volume at 0.736, all ahead of the lesion measures. Whatever these glutamate variants are doing, the model ties them to brain volume rather than to lesion formation.
Where the Three Methods Agree, and Where They Usefully Differ
rs3859123 was selected by all three approaches: Lasso ranked it second for brain volume with a selection probability of 0.581, and the univariate model ranked it first for brain volume at P = 0.000306, with the GG genotype carrying lower brain and white matter volume than AA. A second GRIN2A variant, rs9927924, came first in the univariate analysis of white matter volume. That convergence is the main result. The divergence is instructive too. Applying the per-phenotype Bonferroni correction the authors used within the glutamate group gives a threshold of 0.05/1,292, or 3.87×10⁻⁵. The one univariate result clearing it is a GRM5 variant, rs7483764, with change in T2 lesion load at P = 2.33×10⁻⁵, which the multivariate model ranks only 32nd. The GRIN2A variants lead the univariate ranking but sit 8 to 238 times above that threshold. So each approach promotes a signal the other would set aside, which is the paper's methodological argument made visible: the multivariate model brings out combined, multi-phenotype effects that one-at-a-time correction discards. It is also why confirmation in independent data, rather than any one of these methods, has to decide between them.
A Useful Negative From the Strongest MS Risk Allele
The HLA-DRB1*15:01 tag rs3135388 was included because earlier work had suggested a modest link to T2 lesion load. It showed no relevant association with any of the seven phenotypes: ranked 3,516th by the multivariate model, given a Lasso selection probability of 0.105 for lesion change, and reaching only P = 0.19 in the univariate analysis. That null supports a point the discussion makes directly. In Alzheimer's disease, APOE shapes both who develops the disease and how it progresses, but no comparable gene has emerged in MS, and the genetics of how MS expresses itself in the brain may be a largely separate problem from the genetics of who develops it. Studies like this one are how that separate problem gets opened up.
Why Glutamate and the NMDA Receptor Make Biological Sense
The candidate choice rests on convergent evidence the authors lay out carefully. Glutamate is raised in the cerebrospinal fluid of MS patients and associated with disease severity and course; spectroscopy shows altered glutamate in acute lesions and in white matter that appears normal on conventional imaging; and post-mortem tissue shows changed expression of glutamate transporters, metabotropic glutamate receptors and glutamate-metabolising enzymes. NMDA receptors sit on neurons, astrocytes, oligodendrocyte processes and the compact myelin wrapping axons in white matter, and when excessively activated they raise calcium inside the cell and drive neuronal and oligodendrocyte death. An association between GRIN2A and white matter volume in particular fits that picture well. The five leading variants lie in introns and have no established function yet; the authors suggest they may influence how much GRIN2A transcript is made, possibly through intronic microRNAs, and present that as a hypothesis to test.
What Would Carry This Forward
The authors map the next steps themselves. As the first use of this method on an MS dataset, there are no earlier results to compare against directly. Without a healthy age- and sex-matched control group, they cannot yet say whether these associations are specific to MS or would appear in other neurodegenerative conditions and in ageing, and they state that openly. Confirmation in independent datasets is the step they identify as necessary. A confirmation study would also do well to carry forward a few features of this design: patients were scanned on two 1.5 T systems at two centres, and relapsing and secondary progressive disease were analysed together, so both are worth examining as factors in their own right. And the intronic GRIN2A variants invite functional work on transcript levels. What this paper contributes is a multivariate framework guarded by two independent methods, a GRIN2A signal that all three reach, and a concrete shortlist for the field to test. The work received no specific grant funding and the authors declare no conflicts of interest.
Disclaimer: This blog post is based on the cited research article and is intended for informational purposes only. It is not intended to provide medical advice. Please consult with a healthcare professional for any health concerns.
Reference:
Strijbis, E. M. M., Inkster, B., Vounou, M., Naegelin, Y., Kappos, L., Radue, E.-W., Matthews, P. M., Uitdehaag, B. M. J., Barkhof, F., Polman, C. H., Montana, G., & Geurts, J. J. G. (2013). Glutamate gene polymorphisms predict brain volumes in multiple sclerosis. Multiple Sclerosis Journal, 19(3), 281–288. https://doi.org/10.1177/1352458512454345
