Where MS Lesions Sit, Treated as a Genetic Trait

Where lesions sit in the multiple sclerosis (MS) brain, and what shape they take, has long been recognized as a neuroradiological marker in diagnosis, yet it has been studied far less quantitatively than the traditional measures of T₂ hyperintense or T₁ hypointense lesion volume. Gourraud and colleagues asked whether that spatial arrangement is heritable. They reconstructed the three-dimensional topology of lesions at 1 mm³ voxel resolution from 3 T T₁-weighted scans in a University of California, San Francisco cohort of 484 people with MS, then drew on genome-wide genotypes already published for the same individuals and ran association analyses in the 284 who met their stringent phenotype criteria: clinically isolated syndrome or relapsing remitting disease, duration under ten years, onset after age 20. The authors set this out as a general strategy: intermediate phenotypes that are more precise, more quantitative and closer to the genotype than a physiological state of health or disease are potentially a more informative target for association analysis.

The Registration Problem Comes Before Any Genetics
Comparing a voxel across people requires every brain to sit in one common frame, and lesions themselves interfere with getting there. All imaging came from a single 3 T scanner using a three-dimensional inversion recovery spoiled gradient-recalled echo sequence with 1-mm³ isometric voxels across 180 slices, so acquisition was uniform to begin with. Lesions were segmented by pixel intensity threshold with manual editing, and each lesion was then inpainted, its voxel intensity swapped for that of the surrounding normal white matter, before the brain was non-linearly registered to a single reference scan from a healthy 42-year-old woman. Affine registration tolerates lesions but only fits the global shape of the brain, while non-rigid methods are sensitive to their presence. Inpainting first improves registration over cost-function masking, particularly in lesional areas. The authors state that placing every subject's lesion mask in a common reference frame is critical to performing genetic associations with lesion distribution across a population.

The Probability Map Alone Shows Two Effects on Distribution
Across the full 484 patients and roughly 67,000 voxels, the probability of a lesion peaks around the periventricular areas and falls off markedly elsewhere, which the authors note is as expected. Two splits of the cohort then produced differences in the shape of the distribution rather than its size. Carriers of at least one HLA-DRB1*15:01 allele showed a moderate but significantly different distribution of lesional voxels from non-carriers (P = 2.3×10⁻⁴, matched Wilcoxon test). Men showed more voxels with high lesion probability than women (P = 2.2×10⁻¹⁶). The authors set out the competing reading of that second result themselves, allowing that it may reflect a sex-dependent effect on the number or arrangement of lesional voxels, or the larger white matter volume typical in men.

Why Voxel-Wise Association Fails, and What Was Built Instead
A direct genetic test on each voxel is not workable, because the probability of any given voxel carrying a lesion is very small and most voxels therefore hold no information. The authors tried two ways around this. The first defined subgroups by counting how many of the most frequently affected voxels a patient had lesions in, tuned so the split was roughly balanced; three such combinations were taken to GWAS, where several markers were nominally significant but none survived correction. The second defined a new trait. Lesional voxels that touch along a side, an edge or a corner, 26 neighbours in three dimensions, were merged into clusters the authors call cluxels, which cuts the number of quantitative variables by orders of magnitude and gives a trait that is independent of total lesion volume. Two patients in the cohort carried almost the same lesion volume, 4,737 mm³ and 4,984 mm³, arranged into 13 cluxels in one and 79 in the other, and across the whole cohort the correlation between cluxel count and lesion volume runs at only R = 0.344.

Principal Components Separate Load From Shape
GWAS on cluxel number and on the average minimum distance between cluxels found nothing at genome-wide significance, and average cluxel size came in borderline. So the authors changed the decomposition. Principal component analysis on the variance in lesional voxels put almost 10% of total variance into the first component, which tracks lesion volume closely (R² = 88%). None of the remaining orthogonal components correlated with any other variable they had: total lesion volume, EDSS, age of onset, T₁ and T₂ lesion load, total lesional voxels, age, disease duration, sex, HLA-DRB1*15:01, treatment, or MS subtype. The authors describe this as allowing them to identify the proportion of variance explained by lesion volume itself and then focus on the remaining orthogonal variance, in which lesion topology rather than load is more represented. They ran GWAS on components two through ten, controlling false discovery by permuting the positions of lesional voxels 100 times while leaving each individual's lesion load intact, which removes spatial structure without changing burden.

All Thirty-One Hits Land on a Single Component
Thirty-one associations reached an FDR-corrected P of 0.01, all of them with the eighth component. Thirteen passed P < 10⁻⁷, against an average of 0.64 such associations in the permuted data for that component and 2.32 across all ten, and no other component yielded anything significant. The strongest were rs16861690 near CDCA7 (P = 1.80×10⁻¹¹) and rs13410351 near TMEM182 (P = 6.12×10⁻¹¹), while the top variant falling inside a gene was rs10119179 in SYK (P = 3.65×10⁻⁹), followed by rs6983731 in TRAPPC9 (P = 1.09×10⁻⁸), with SLITRK6, MYT1L, RIC3, GRIK4, NLRP11 and a non-synonymous variant in MRPS15 also on the list. The authors assemble published support for SYK, MYT1L, TRAPPC9, SLITRK6 and RIC3 in the development and distribution of white matter lesions. They give SYK as one of the most relevant associations, a kinase that phosphorylates myelin basic protein and also α-synuclein, preventing its aggregation, which would give it an anti-neurodegenerative role, and whose DNA methylation is dynamically regulated in human cerebral cortex across the lifespan. Taken together the hits map onto either neural functions such as axonogenesis and transmission of nerve impulse, or immune ones such as NFκB signalling and T-cell regulation.

A Network Result, and the Limits the Authors Set
To ask whether these genes work together, the authors ran protein interaction network analysis, scoring each gene product by its most strongly associated SNP and growing sub-networks from the resulting map. The top-ranked module held 48 interacting proteins, most of them expressed in inflammatory cells or in the central nervous system, and gene ontology showed enrichment both in immune response and in CNS development and function, with myelination, axon ensheathment, regulation of axonogenesis and neuron recognition among the terms. Permuting P-values among all elements in the protein interaction network produced a top-scoring network with no significant enrichments. Of the genes with modest associations that physically interact with the top hits, 19 of 30 are expressed in the brain, and the authors single out SEMA3A, RTN4R, GRM7, LRRC4C and FYN for their role in axon guidance during development. They also set out what is missing. There is no independent confirmation cohort, which they call the most reliable evidence a genetic association can have, and they explain why it was out of reach: a similarly sized or larger group with both genome-wide genotypes and high-resolution structural imaging was something very few centres could assemble, and the multicentre cohorts that exist bring the unsolved problem of comparing images acquired at different sites. They add that the study may be underpowered for unequivocal signals, and that most associated SNPs were uncommon in this cohort, at around 5%, with regional association plots showing several SNPs in high linkage disequilibrium with index markers that carried no association of their own. Their stated conclusion is that they have identified an MRI-based quantitative trait associated with common variants in MS patients, and that the data reduction methods proposed substantially improve the potential for prospective studies to test these candidate modifier genes.

Disclaimer: This blog post is based on the cited study 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:
Gourraud, P.-A., Sdika, M., Khankhanian, P., Henry, R. G., Beheshtian, A., Matthews, P. M., Hauser, S. L., Oksenberg, J. R., Pelletier, D., & Baranzini, S. E. (2013). A genome-wide association study of brain lesion distribution in multiple sclerosis. Brain, 136(4), 1012–1024. https://doi.org/10.1093/brain/aws363