Polygenic risk score association with multiple sclerosis susceptibility and phenotype in Europeans

Abstract

Polygenic inheritance plays a pivotal role in driving multiple sclerosis susceptibility, an inflammatory demyelinating disease of the CNS. We developed polygenic risk scores (PRS) of multiple sclerosis and assessed associations with both disease status and severity in cohorts of European descent.

The largest genome-wide association dataset for multiple sclerosis to date (n = 41 505) was leveraged to generate PRS scores, serving as an informative susceptibility marker, tested in two independent datasets, UK Biobank [area under the curve (AUC) = 0.73, 95% confidence interval (CI): 0.72–0.74, P = 6.41 × 10⁻¹⁴⁶] and Kaiser Permanente in Northern California (KPNC, AUC = 0.8, 95% CI: 0.76–0.82, P = 1.5 × 10⁻⁵³).

Individuals within the top 10% of PRS were at higher than 5-fold increased risk in UK Biobank (95% CI: 4.7–6, P = 2.8 × 10⁻⁴⁵) and 15-fold higher risk in KPNC (95% CI: 10.4–24, P = 3.7 × 10⁻¹¹), relative to the median decile. The cumulative absolute risk of developing multiple sclerosis from age 20 onwards was significantly higher in genetically predisposed individuals according to PRS. Furthermore, inclusion of PRS in clinical risk models increased the risk discrimination by 13% to 26% over models based only on conventional risk factors in UK Biobank and KPNC, respectively. Stratifying disease risk by gene sets representative of curated cellular signalling cascades, nominated promising genetic candidate programmes for functional characterization. These pathways include inflammatory signalling mediation, response to viral infection, oxidative damage, RNA polymerase transcription, and epigenetic regulation of gene expression to be among significant contributors to multiple sclerosis susceptibility. This study also indicates that PRS is a useful measure for estimating susceptibility within related individuals in multicase families. We show a significant association of genetic predisposition with thalamic atrophy within 10 years of disease progression in the UCSF-EPIC cohort (P < 0.001), consistent with a partial overlap between the genetics of susceptibility and end-organ tissue injury. Mendelian randomization analysis suggested an effect of multiple sclerosis susceptibility on thalamic volume, which was further indicated to be through horizontal pleiotropy rather than a causal effect.

In summary, this study indicates important, replicable associations of PRS with enhanced risk assessment and radiographic outcomes of tissue injury, potentially informing targeted screening and prevention strategies.

Shams et al. show that polygenic risk scores identify individuals at high risk of developing multiple sclerosis and who may benefit from preventive interventions. A strong genetic predisposition also associates with hallmarks of disease progression, advancing understanding of biochemical pathways related to disease inheritance.

Introduction

Multiple sclerosis is a chronic disease of the CNS with established genetic susceptibility footprints. Leveraging genome-wide genotype data from 47 429 multiple sclerosis cases and 68 374 controls, the International Multiple Sclerosis Genetics Consortium (IMSGC) developed a dataset that yielded statistical evidence for the association of 200 autosomal susceptibility variants outside the major histocompatibility complex (MHC), 32 within the extended MHC region, and one in chromosome X.¹ These associations together with an additional 416 highly suggestive, albeit not genome-wide significant variants explain 48% of multiple sclerosis heritability, and collectively highlight gene networks operating in the adaptive and innate arms of immune response, as well as enrichment of genes expressed in microglia.¹

The polygenic mode of multiple sclerosis inheritance provided the rationale for developing aggregated genetic burden scores including all identified genome-wide significant susceptibility variants, in an attempt to better predict the cumulative effects of genetic liability.^2,3 Polygenic risk scores (PRS) combine all genetic effects into a single metric of inherited susceptibility with multiple important potential applications: they can be used to assess genetic heritability, measure genetic overlap between different traits, improve screening, assist in identifying biomarkers of complex diseases, stratify patients in clinical trials, and adjust treatment strategies.^4–9 However, the early application of this tool in multiple sclerosis was compromised by gaps in the number of known risk loci and imperfect estimation of allelic weights, leading to poor sensitivity.^10–12 Methodological and analytical advances, and increasing sample sizes in genome-wide association studies (GWAS) have facilitated genetic discoveries and provided adequate power for detecting variations with small effects on the target phenotype, thus enabling more compelling studies of the risk distribution in a given population. Specifically, the LDPred algorithm considerably enhanced the prediction accuracy of the genetic load for autoimmune diseases including type 1 diabetes, Crohn’s disease, and rheumatoid arthritis.¹³ Recent updates of the program (LDPred2) provide even higher performance and more accurate effect size adjustments compared to both the old version and other competitive methods, particularly for causal variants in complex long-range linkage disequilibrium (LD) regions such as the MHC.¹⁴ LDPred2 also identifies variants with true zero effect size, significantly reducing the number of variants used in the PRS calculations without affecting its predictive performance. Here, we developed both genome-wide and pathway-specific PRS for well-characterized, independent multiple sclerosis datasets to assess the aggregated genetic burden effects on disease risk and activity.

Materials and methods

Polygenic risk score derivation

The post-quality control (QC) IMSGC Discovery summary statistics based on 14 802 multiple sclerosis cases and 26 703 controls including 8 589 719 variants was used for PRS derivation to predict multiple sclerosis risk in people of European ancestry. Please refer to IMSGC for details of multisite cohorts, inclusion criteria, demographics and analytical methods.¹ Datasets used in this study are summarized in Supplementary Table 1. All duplicate single nucleotide polymorphisms (SNPs), multi-allelic markers, variants with low imputation quality (INFO < 0.6), rare variants with minor allele frequencies of <1%, variants with genotype missingness >10%, and variants deviating from Hardy–Weinberg equilibrium (P < 1 × 10⁻⁶) were excluded. For specific details about each dataset please refer to the Supplementary material.

The LDPred2 algorithm was used to obtain posterior mean effect sizes by adjusting a prior probability and accounting for LD.¹⁴ As recommended by recent studies,^14,15 only HapMap3 variants that passed rigorous QC in our validation and replication datasets, namely UK Biobank (UKBB), Kaiser Permanente in Northern California (KPNC), and UCSF-EPIC, were included in PRS development (n = 800 702). This set contained established disease-associated variants and other loci. The advantage of using HapMap3 variants is that they have passed extensive QC and have a good coverage of the whole genome.

The underlying population structure of both UKBB and KPNC datasets is homogeneous and strictly European, while 6.6% of UCSF-EPIC subjects were identified as Admixed Americans (AMR), which was demonstrated by performing principal component analysis (PCA) of each dataset with 1000 Genomes Phase 3 populations as a reference (Supplementary Fig. 1A–H). The LD reference used was based on 362 320 European individuals in UKBB. The tuning hyperparameters, sparsity (p) and heritability (h²), were optimized in UKBB Phase 1 (UKBB1) (validation set) comprising 601 multiple sclerosis cases and 109 990 unaffected subjects. Heritability values tested in the model included the one estimated by LD score regression (h² = 0.38) as well as that multiplied by 0.7, and 1.4. A total of 17 P-values between 1 × 10⁻⁴ and 1 evenly spaced on a logarithmic scale were examined. Discriminative capacity of PRS was assessed as the maximum area under the receiver-operating curve (AUC) and 95% confidence intervals (CI) obtained from 10 000 non-parametric bootstrap replicates. The sparse option was enabled in LDPred2 to allow for computing true zero effect sizes. A total of 455 902 (57% of included variants) had non-zero effect sizes β = {β₁, β₂, …, β_n} and PRS of the j-th individual was computed by taking the weighted sum of risk alleles, $PR{S}_j={\sum }_{i=1}^n{\beta }_{i}SN{P}_{ij}$. All PRS scores are standardized to have zero mean and unit variance. The best performing model out of 61 in the validation dataset (P = 0.018, h² = 0.38, AUC = 70%, 95% CI: 68% to 72%) was then tested in UKBB Phase 2 (UKBB2), consisting of 1354 multiple sclerosis cases and 252 065 unaffected subjects after QC. The top model was then tested in the KPNC dataset. Scores were adjusted for the first 20 principal components of ancestry. The flowchart of PRS development is shown in Fig. 1.

Figure 1.

Flow diagram of the MS-PRS derivation. The post-QC discovery IMSGC summary statistics were used to construct a predictive model of multiple sclerosis risk. Multiple genetic models comprising significant variants and their associated effect sizes were optimized over a range of parameters using LDpred2. Performance of genetic models was validated in the UKBB1 and independently tested in UKBB2 and KPNC datasets.

Since the UCSF-EPIC cohort was part of the IMSGC meta-analysis, two sets of PRS scores were developed for this dataset. In the first set, PRS is computed using the same effect sizes as described above, which are based on the summary statistics incorporating all IMSGC-GWAS datasets. These scores are referred to as MS-PRS for all cohorts used in this study. Additionally, a second set of scores for UCSF-EPIC was developed for which a summary statistics based on only SNPs contributed by all the IMSGC sites except UCSF (n = 14) was used (PRS_LOO). The PRS_LOO also included SNPs that passed QC in all datasets, as described above. However, before computing the effect sizes in LDPred2, variants with SD_ss < 0.5.SD_val or SD_ss > 0.1 + SD_val or SD_ss < 0.1 or SD_val < 0.05, in which SD_ss is the standard deviations from the summary statistics and SD_val is the standard deviations of genotypes of subjects in the validation set, are removed as suggested by Privé et al.¹⁴ Applying this condition using the leave-one-out summary statistics resulted in removing 5320 SNPs from PRS_LOO.

To develop familial PRS, DNA samples of 135 individuals from 35 families with an unaffected parent, a co-affected parent-child pair, and a discordant sib-pair were genotyped using a custom Illumina Chip.¹⁶ Due to the limited overlap between variants in the familial dataset and the HapMap3 markers (∼18%), the familial PRS was computed based on the common variants present in this dataset, the IMSGC summary statistics, and UKBB1, which was used as the validation set (n = 921 610). Chromosome-specific scores were computed by summing over variants within each chromosome multiplied by their corresponding effect sizes, which were estimated genome-wide, as described above.

Gene set selection and derivation of pathway-specific risk scores

Pathway-specific risk scores were developed for the 2852 gene sets identified in the Canonical Pathway subset of Molecular Signatures Database (MSigDB version 7.2) consisting of curated subsets representing biological pathways using PRSice2.¹⁷ The LD pruning was carried out using r² of 0.1 and a 1 mb window. To generate competitive P-value estimates for each pathway, 10 000 permutations of sample labels were implemented. Nagelkerke’s pseudo R² value was adjusted for an estimated prevalence of 0.00127,¹⁸ gender, age and the first 20 principal components. The gene sets surviving the Bonferroni correction for multiple testing (corrected P = 1.75 × 10⁻⁵) in UKBB1 were further replicated in UKBB2. Cumulative scores of 132 gene sets were significantly associated with multiple sclerosis risk in the UKBB1 (Bonferroni corrected P < 1.75 × 10⁻⁵). A total of 85 associations were replicated in UKBB2 (Supplementary Table 2).

Polygenic risk score association with neuroimaging phenotypes

Calibrated volumetric measurements of the total brain (BV), white matter (WMV), peripheral grey matter (pGMV), CSF, and three compartments of the deep grey matter, namely thalamus, caudate and putamen in 467 UCSF-EPIC participants with 10 years of annual follow-up were investigated as hallmarks of disease progression.¹⁹ First, the association of both PRS_LOO and MS-PRS with each metric at baseline was examined in a linear model including sex, age and disease duration as covariates. To replicate associations in UKBB, only cases from both phases with available neuroimaging data were included in the analysis (n = 132).²⁰

Longitudinal associations were then examined by computing percentage change in neuroimaging phenotypes with respect to the baseline at each annual visit within a 10-year follow-up period. Using linear mixed modelling for repeated measurements (MMRM), the associations between polygenic risk scores and longitudinal per cent change of volumetric measurements of BV, WMV, pGMV, thalamus, caudate and putamen were investigated. To avoid potential biases imposed by inflated MS-PRS in UCSF-EPIC, analyses were only performed with PRS_LOO. Years six and seven were excluded due to the high proportion of missing data. Specifically, fixed effects in the model included PRS_LOO, baseline phenotype values, age at baseline, disease duration at baseline, sex and visit as nominal variables, interaction terms between visit and PRS_LOO (visit × PRS_LOO), visit and age at baseline (visit × age), and visit and sex (visit × sex). PRS_LOO was treated as a continuous variable. Antedependent covariance structure was used in all models.

For details on datasets used in this study, the conventional risk factors, coefficient of determination, absolute risk stratification, and Mendelian randomization, please refer to the online Supplementary material.

Data availability

This study includes no data deposited in external repositories. All the data supporting the findings of this study are available through application to UK Biobank or request from the corresponding author.

Results

Polygenic susceptibility of multiple sclerosis

The MS-PRS performance in UKBB2 evaluated by the AUC (AUC = 0.73, 95% CI: 0.72–0.74; Supplementary Fig. 2A) was higher than that in UKBB1 (AUC = 0.7, 95% CI: 0.68–0.72). Increased MS-PRS of multiple sclerosis cases compared to unaffected subjects in UKBB2 was confirmed by unpaired two-sample Wilcoxon test (P = 1.29 × 10⁻¹⁹⁰) shown in Fig. 2A and Supplementary Fig. 2B. The risk of multiple sclerosis across equal strata of increasing MS-PRS was estimated by odds ratios relative to the median decile (OR_Med) shown in Fig. 2B as well as relative to the remainder of the population, corrected for the first 20 principal components (PCs), age and sex. Both measures indicated that individuals at the tail of the MS-PRS distribution were at markedly higher risk of developing the disease. Specifically, individuals in the top 5% and 10% of MS-PRS in UKBB2 were at more than 6- (95% CI: 5.7–6.5, P = 9.9 × 10⁻¹³⁴) and 5-fold (95% CI: 4.8–5.4, P = 6.41 × 10⁻¹⁴⁶) increased risk, respectively, compared to the rest of the population (Supplementary Fig. 2C). Similarly, individuals at the top MS-PRS decile were at greater risk relative to the median decile (OR_Med > 5.3, 95% CI: 4.7–6, P = 2.8 × 10⁻⁴⁵), shown in Fig. 2B. Prevalence of multiple sclerosis notably increased according to the MS-PRS percentile (Fig. 2C). These results confirm that genetic predisposition is an important component of multiple sclerosis risk in a population-level cohort.

Figure 2.

Discriminative power of MS-PRS in the UKBB2 and KPNC datasets. (A) The MS-PRS of subjects affected by multiple sclerosis (mean = 1, IQR = 1.770, min = −2.575, max = 4.561) is greater than unaffected individuals (mean = −0.005, IQR = 1.484, min = −3.979, max = 5.3) in UKBB2 (***P = 1.29 × 10⁻¹⁹⁰). (B) The OR_Med at higher MS-PRS deciles relative to the median decile is increased more than 5-fold in UKBB2. The y-axis is on a log scale. Bars are standard errors. (C) Prevalence of multiple sclerosis in UKBB2 increased according to the MS-PRS percentile. (D) Consistent with UKBB2, the MS-PRS of cases (mean = 0.585, IQR = 1.064, min = −1.474, max = 2.945) is higher than controls (mean = −0.579, IQR = 1.076, min = −3.193, max = 2.514) in KPNC (***P = 1.5 × 10⁻⁵³). (E) The OR_Med in the top MS-PRS decile of KPNC is higher than 15-fold relative to the median decile. (F) A linear increase of the multiple sclerosis prevalence with respect to MS-PRS is observed in the KPNC cohort. IQR = interquartile range; min = minimum; max = maximum.

The discriminative power of MS-PRS in the well-curated, case-control KPNC cohort was enhanced (Supplementary Fig. 2D; AUC = 0.8, 95% CI: 0.76–0.82) compared to UKBB2. The scores of KPNC cases were greater than controls (P = 1.5 × 10⁻⁵³), as shown in Fig. 2D. Similar to UKBB2, the proportion of KPNC subjects with increased MS-PRS had higher OR values relative to the remainder of this dataset (Supplementary Fig. 2E and F). The top MS-PRS decile relative to the median decile indicated a 15-fold higher OR (95% CI: 10.4–24, P = 3.7 × 10⁻¹¹), as shown in Fig. 2E. The multiple sclerosis prevalence in the KPNC dataset according to MS-PRS percentile linearly increased, reflecting the size and higher proportion of multiple sclerosis cases in this dataset (Fig. 2F).

Splitting MS-PRS by chromosome demonstrated the higher statistical significance of chromosome 6, containing the extended MHC region, in predicting the multiple sclerosis disease status in both UKBB2 and KPNC datasets (Supplementary Fig. 3A and B). The predictive power of collective scores based on all autosomes excluding chromosome 6, marked by ‘A-6’ in Supplementary Fig. 3A and B, was almost equal to that of the scores based on chromosome 6 alone, showing that excluding chromosome 6 in both datasets significantly affected P-values of cumulative MS-PRS. The MS-PRS distributions in males and females were not significantly different in KPNC, while the unaffected female MS-PRS was slightly higher than that of males (P < 0.01) in UKBB2 (Supplementary Fig. 3C and D). Such difference was driven by the outliers and became insignificant when those were excluded from the UKBB2 analysis.

To assess the prediction accuracy of MS-PRS, we used R² on the liability scale $(R_l^2)$, which was suggested to be directly comparable to heritability by Lee et al.²¹ Please refer to the Supplementary material for prevalence and derivation details. Assuming that multiple sclerosis liability has a normal distribution, coefficients of determination on the liability scale in the UKBB2 were $R_{l,0.00127}^2$, 7.7% (95% CI: 6.8–8.7%), $R_{l,0.0019}^2$, 8.3% (95% CI: 7.3–9.2%), and $R_{l,0.0069}^2$, 10.7% (95% CI: 9.4–11.8%), while R² values corrected for the ascertainment bias in KPNC were increased to $R_{l,0.00127}^2$, 12.5% (95% CI: 11.5–13.5%), $R_{l,0.0019}^2$, 13.3% (95% CI: 12.3–14.4%), and $R_{l,0.0069}^2$, 16.9% (95% CI: 15.7–18.2%) shown in Table 1. Under the logistic distribution assumption, $R_l^2$ in UKBB2 and KPNC were 16.23% (95% CI: 16.2–16.3%) and 30% (95% CI: 28.1–31.9%), respectively. The UKBB is a population-level cohort, whereas KPNC is a case-control study in which cases are identified and matched to controls arising in the same population. Nonetheless, $R_l^2$ was comparable between these datasets, particularly under the normal distribution assumption, in agreement with Lee et al.²¹ Limiting the analysis to only first, fifth, and 10th MS-PRS deciles combined, increased the prediction accuracy by 30% (P = 6.97 × 10⁻¹⁴⁵) and 60% (P = 1.01 × 10⁻³⁸) in UKBB2 and KPNC datasets, repectively.

Table 1.

Prediction accuracy of MS-PRS in the test datasets

MS prevalence	$R_l^2$ (95% CI)
	UKBB2	KPNC
0.00127	7.7 (6.8–8.7)	12.5 (11.5–13.5)
0.0019	8.3 (7.3–9.2)	13.3 (12.3–14.4)
0.0069	10.6 (9.4–11.8)	16.9 (15.7–18.2)

The regression P-values were 3.13 × 10⁻¹⁷⁰ and 1.56 × 10⁻⁵⁷ for UKBB2 and KPNC datasets, respectively. $R_l^2$= coefficient of determination on the liability scale.

We also computed Nagelkerke’s pseudo-R²$(R_{\mathrm{Nag}}^2)$ on the observed scale in order to compare the MS-PRS performance with a previous case-control study by IMSGC.² Genetic burden scores in this study showed 3% association with susceptibility estimated by $R_{\mathrm{Nag}}^2$. The case-control ratio in this study (∼0.38) was significantly higher than that in UKBB2 (∼0.005), but lower than KPNC (∼0.85), while having a substantially sparser genotyping density than both UKBB and KPNC. The $R_{\mathrm{Nag}}^2$ was 8% (P = 7.26 × 10⁻²²⁴) and 33% (P = 9.4 × 10⁻⁶⁰) in UKBB2 and KPNC, respectively. Although $R_{\mathrm{Nag}}^2$ captured the improved predictive power of MS-PRS in independent test sets, it is relatively sensitive to dataset composition and thus may not serve as an accurate measure for comparing PRS models between population-level and case-control cohorts.

The cumulative incidence of multiple sclerosis as a function of age was used to assess the ability of PRS to refine risk estimates in the UK population. Risk stratification was primarily driven by PRS as indicated by significantly diverging risk trajectories after age 20 (Supplementary Fig. 4A). Multiple sclerosis clinical onset age typically ranges between 20 and 40. The cumulative risk of individuals in the UKBB2 within the top 5% of MS-PRS for developing multiple sclerosis up to age 40 was more than 8- and 30-fold higher than those within the 30–60% of MS-PRS (P < 0.0001; Supplementary Table 2) and the bottom 5% of MS-PRS percentile (P < 0.0001), respectively. Coarser MS-PRS strata were used for this analysis in the KPNC dataset to account for the smaller sample size. Significant divergence of risk trajectories after the age of 20 and higher cumulative incidence for KPNC subjects in the top 20% of MS-PRS up to age 40 (fold change compared to the lowest 20% MS-PRS, 6.5; P < 1 × 10⁻⁴) is shown in Supplementary Fig. 4B and Supplementary Table 3.

Combining polygenic risk score of multiple sclerosis with conventional risk factors

Classification accuracies of basic models, consisting of age, sex, and established conventional risk factors (CRF) of multiple sclerosis were compared against models in which MS-PRS was also included (Supplementary Table 4). An increase in the AUC, net reclassification index (NRI), and integrated discrimination index (IDI) are reported as measures of risk discrimination improvement for MS-PRS-included versus basic models (Table 2). For more details, please refer to the Supplementary material.

Table 2.

Risk model improvement upon PRS addition in UKBB2 and KPNC datasets

Prediction model	AUC (95% CI)	NRI (95% CI)	IDI (95% CI)	P a
UKBB2
ȃsex + age	0.64 (0.62–0.65)	NA	NA	NA
ȃsex + age + PRS	0.77 (0.75–0.78)	0.729 (0.67–0.788)	0.007 (0.006–0.008)	<0.001
ȃsex + age + HLA.DRB1*15:01 (rs3135391)	0.69 (0.68–0.71)	0.47 (0.407–0.533)	0.002 (0.002–0.003)	<0.001
ȃsex + age + mono	0.64 (0.63–0.66)	NA	NA	NA
ȃsex + age + mono + PRS	0.77 (0.75–0.78)	0.727 (0.668–0.786)	0.007 (0.006–0.008)	<0.001
ȃsex + age + mono + smoking	0.65 (0.64–0.67)	NA	NA	NA
ȃsex + age + mono + smoking + PRS	0.78 (0.76–0.79)	0.726 (0.667–0.785)	0.007 (0.007–0.008)	<0.001
ȃsex + age + mono + smoking + BMI	0.66 (0.64–0.67)	NA	NA	NA
ȃsex + age + mono + smoking + BMI + PRS	0.77 (0.76–0.79)	0.719 (0.658–0.781)	0.007 (0.006–0.008)	<0.001
KPNC
ȃsex + age	0.62 (0.58–0.65)	NA	NA	NA
ȃsex + age + PRS	0.80 (0.77–0.83)	0.838 (0.718–0.958)	0.238 (0.21–0.266)	<0.001
ȃsex + age + HLA.DRB1*15:01 (rs3135391)	0.72 (0.68–0.75)	0.64 (0.517–0.762)	0.104 (0.083–0.124)	<0.001
ȃsex + age + mono	0.62 (0.58–0.65)	NA	NA	NA
ȃsex + age + mono + PRS	0.80 (0.77–0.83)	0.859 (0.74–0.978)	0.238 (0.21–0.266)	<0.001
ȃsex + age + mono + smoking	0.64 (0.60–0.67)	NA	NA	NA
ȃsex + age + mono + smoking + PRS	0.81 (0.78–0.84)	0.874 (0.756–0.993)	0.234 (0.206–0.262)	<0.001
ȃsex + age + mono + smoking + family history	0.67 (0.63–0.70)	NA	NA	NA
ȃsex + age + mono + smoking + family history + PRS	0.82 (0.79–0.85)	0.844 (0.725–0.964)	0.214 (0.187–0.241)	<0.001
ȃsex + age + mono + smoking + family history + PRS + PRS*family history	0.82 (0.80–0.85)	0.813 (0.692–0.933)	0.222 (0.195–0.25)	0.001
ȃsex + age + mono + smoking + family history + overweight as a child	0.67 (0.64–0.71)	NA	NA	NA
ȃsex + age + mono + smoking + family history + overweight as a child + PRS	0.82 (0.80–0.85)	0.844 (0.725–0.964)	0.212 (0.185–0.239)	<0.001
ȃsex + age + mono + smoking + family history + overweight as a child + BMI	0.71 (0.68–0.75)	NA	NA	NA
ȃsex + age + mono + smoking + family history + overweight as a child + BMI + PRS	0.83 (0.80–0.86)	0.814 (0.687–0.941)	0.184 (0.157–0.211)	0.001
ȃsex + age + mono + smoking + family history + overweight as a child + BMI20	0.68 (0.60–0.75)	NA	NA	NA
ȃsex + age + mono + smoking + family history + overweight as a child + BMI20 + PRS	0.85 (0.79–0.92)	1.051 (0.805–1.297)	0.202 (0.142–0.262)	<0.001

Basic models are shown in bold. NA = not applicable; BMI = body mass index; BMI20 = BMI at 20; mono = mononucleosis infection.

Significance of AUC difference, NRI and IDI.

Inclusion of PRS in all basic models consistently enhanced model performance in UKBB2 (Table 2). An alternative genetic model consisting of only a single SNP tagging HLA-DRB1*15:01, considered as the major genetic contributor to multiple sclerosis susceptibility,²² achieved an AUC of 0.69 (95% CI: 0.68–0.71), lagging the PRS-included model by 8% (P < 0.001). The NRI of the PRS-included model was 0.73 (95% CI: 0.67–0.79, P < 0.001), higher than that of a single SNP model (0.47, 95% CI: 0.41–0.53, P < 0.001). The best-performing model in this dataset included PRS, mono, smoking history, sex, and age (AUC = 0.78, 95% CI: 0.76–0.79, P < 0.001).

Other established risk factors, such as family history and being overweight as a child are available in KPNC, thus additional models are presented for this dataset. Consistent with the UKBB2 results, the single SNP tagging HLA-DRB1*15:01 achieved an AUC of 0.69 (95% CI: 0.65–0.72), while the AUC of PRS-included model reached 0.8 (95% CI: 0.77–0.83). Adding PRS to the model including mono, smoking history, family history, current body mass index (BMI), being overweight as a child, sex, and age increased the AUC by 12% to 0.83 (95% CI: 0.80–0.85), with an NRI of 0.818 (95% CI: 0.698–0.938, P < 0.001), IDI of 0.207 (95% CI: 0.146–0.268), and a false positive rate of <1%. Replacing the current body mass index (BMI) by BMI at 20s, which was only available in the KPNC dataset increased the AUC to 86% (95% CI: 0.80–0.85). Nevertheless, due to the high proportion of missing data (Supplementary Table 4), this result is likely biased. Altogether, addition of PRS to CRF notably improved model performance in both UKBB2 and KPNC.

This outcome was further tested in the UCSF-EPIC cohort.²³ The prediction accuracy of PRS_LOO, described in the ‘Materials and methods’ section, is negatively impacted by both re-adjusted effect sizes and the number of included SNPs. Nevertheless, addition of PRS_LOO to CRFs significantly enhanced model performance, increasing the AUC of the basic model including mononucleosis infection, smoking, family history, current BMI, being overweight as a child, sex, and age by 16% (NRI = 0.792, 95% CI: 0.662–0.922, Supplementary Table 5) and decreasing the false positive rate by 6%. The interaction term between PRS and family history did not affect model performance in either KPNC or EPIC. Repeating this analysis excluding the AMR subjects from UCSF-EPIC (6.6%) increased the AUC by 1% (Supplementary Table 5). Including MS-PRS in EPIC overestimated the prediction accuracy of the full model (AUC > 0.9), as expected.

Pathway-specific polygenic risk scores

To investigate whether PRS can identify genetic circuits underlying multiple sclerosis risk, pathway-specific risk scores were computed. Out of 85 risk-associated pathway-based scores replicated in UKBB2, several were related to adaptive immune response, such as IL-5 (R² = 1.8%, P = 2.03 × 10⁻⁶⁴) and IL-12 signalling (R² = 2.1%, P = 7.89 × 10⁻⁷⁷), T cell receptor (TCR) signalling (R² = 2.1%, P = 5 × 10⁻⁸⁶), MHC class II antigen presentation (R² = 2%, P = 9.7 × 10⁻⁷⁸), interferon gamma signalling (R² = 2%, P = 2.7 × 10⁻⁷⁶), and complement cascade (R² = 1.4%, P = 3 × 10⁻⁵⁴). Viral and parasite infection response pathways also emerged as significantly associated with multiple sclerosis risk (P < 1 × 10⁻⁵⁸). Signature pathways for other autoimmune chronic conditions such as lupus (R² = 2.2%, 'P = 9.84 × 10⁻⁸⁶), Hashimoto’s thyroiditis (R² = 2%, P = 1.78 × 10⁻⁷⁸), and diabetes type I (R² = 2%, P = 3.39 × 10⁻⁸¹) appeared common with multiple sclerosis susceptibility pathways. Gene sets involved in cell adhesion (R² = 2%, P = 1.54 × 10⁻⁷⁷) and extracellular matrix (ECM) organization (R² = 1.6%, P = 1.16 × 10⁻⁵⁶) and protein glycosylation (R² = 1%, P = 6.95 × 10⁻³⁸) were also among the top pathway scores. Other signalling cascades such as the VEGF (R² = 2%, P = 4.83 × 10⁻⁷⁵) and several NOTCH pathways (R² < 1.3%, P < 1 × 10⁻³⁶) also contribute to multiple sclerosis susceptibility. The risk of the top 10% relative to the remainder of the population ranged between 2 (95% CI: 0.92–1.06) and 3 (95% CI: 2.8–3.2), according to the pathway-specific scores.

Multiple sclerosis risk and parental genetic load

The best-performing PRS for the UCSF Multi-case Quartets dataset, validated in UKBB1, reached an AUC of 0.66. For details on the dataset and PRS derivation, please refer to the Supplementary material. The PRS distribution of cases was significantly higher than controls (P = 0.0004), as shown in Fig. 3A and Supplementary Fig. 5A. The statistical significance and the predictive power of the familial PRS compared to that of MS-PRS in UKBB2 and KPNC datasets was negatively impacted by several factors, including the limited number of subjects, the kinship among individuals, and the limited overlap of variants with those included in the HapMap3-based MS-PRS. Consistently, the chromosome-based PRS highlighted the importance of chromosome 6 in risk discrimination in affected families (Supplementary Fig. 5B). Moreover, the PRS distribution of siblings was not impacted by sex stratification, as indicated in Supplementary Fig. 5C. Families were then classified according to parents’ risk scores, i.e. lower or higher than the median PRS (PRS_Median), and lower or higher than 25% and 75% quantiles of PRS of all subjects, to further elucidate patterns of inheritance. The average PRS of mothers (M), fathers (F), affected (AS) and unaffected siblings (US) in each subgroup are shown in Fig. 3B. For comparisons across subgroups, we included 12 additional siblings data available for nine families to enhance statistical power. The PRS scores of all affected siblings (n = 40) were higher than all unaffected siblings (n = 42, P < 0.05), shown in the ‘All’ subgroup of Fig. 3B. In only four families, both parents were at higher risk than PRS_Median. The most common pattern in the dataset, according to the subgroups defined in Fig. 3B, was mothers’ scores higher and fathers’ scores lower than PRS_Median (n = 15). Therefore, the affected sibling scores in this subgroup compared to all unaffected siblings reached the highest statistical significance across all subgroups (P = 0.001). Although the statistical power was reduced in other subgroups due to the small number of families that met the criteria, the mean PRS of affected siblings was consistently higher (P < 0.05) than all unaffected siblings in families of one high and one low parental risk (Fig. 3B). The AUC for predicting disease status among siblings was 0.65 (Fig. 3C), which is comparable to overall AUC in this dataset.

Figure 3.

The PRS analysis of families with multiple sclerosis. (A) The PRS distribution of multiple sclerosis cases (mean = 0.265, IQR = 1.364, min = −1.990, max = 2.362) in families with co-affected parent-child pairs is higher than that of controls (mean = −0.27, IQR = 1.414, min = −2.180, max = 2.655; ***P = 0.0004). (B) The discriminative power of PRS among siblings (AUC = 0.65) is equal to the overall AUC of 0.66 in this dataset. (C) Families are stratified according to parents’ risk scores and average PRS of affected siblings in each subgroup is compared to that of all unaffected siblings (US in the ‘All’ subgroup). The median PRS and the 25% and 75% PRS quantiles are used for grouping families. For families in which one parent is at high and one at low risk relative to both median and top and bottom 25% PRS quantiles, affected siblings’ PRS scores in subgroups are consistently higher than all unaffected siblings. Subgroups: ‘All’ = all mothers, fathers, affected and unaffected siblings in the dataset; ‘Mother Higher Father Lower (Med)’ = mothers and fathers having higher and lower PRS scores relative to the median score, respectively; ‘Father Higher Mather Lower (Med)’ = fathers and mothers having higher and lower PRS scores relative to the median score, respectively; ‘Mother Higher Father Lower (Qt)’ = mothers and fathers having higher and lower PRS scores relative to the top and bottom 25% PRS quantiles, respectively; ‘Father Higher Mother Lower (Qt)’ = fathers and mothers having higher and lower PRS scores relative to the top and bottom 25% PRS quantiles, respectively. (D) Genetic correlations between family members depicted in the matrix scatter plot. Density plots across the diagonal are PRS distributions of M, F, US and AS. Linear (green) and loess (purple) regressions of pairwise PRS graphs are shown in the lower half of the matrix plot, shadings reflect data spread. The highest positive correlation is between PRS of fathers and that of AS (P < 0.001) followed by the correlation between US and AS (P < 0.001). AS = affected siblings; Corr = correlation; F = fathers; M = mothers; Med = median; Qt = quantile; US = unaffected siblings.

Pairwise correlations of MS-PRS between family members are depicted in Fig. 3D. Diagonal panels are PRS distributions of parents and siblings. Mothers constituted 65% of affected parents in this dataset, and the mother and affected siblings' distributions were skewed towards higher PRS values, while fathers’ distribution was shifted towards lower scores. The unaffected siblings' distribution was relatively symmetric. The correlation indices shown in the upper-half panels were obtained from the data-points in the lower-half of the matrix plot. The negative correlation between mothers and fathers reflected that only one spouse was affected in each family. The strongest correlation was observed among fathers and affected children (0.5, P < 0.005), while PRS of mothers better correlated with unaffected children (0.41, P < 0.05), suggesting that genetic predisposition of fathers is an important risk factor for children. The correlation among siblings was the second highest (0.46, P < 0.005).

Polygenic risk score association with disease progression and activity

Changes in CNS volumes represent a quantifiable surrogate of tissue loss and long-term disease progression in multiple sclerosis.^24–26 Assessing the proportion of phenotypic variations at baseline explained by MS-PRS and PRS_LOO showed modest associations of BV (β_LOO = −0.09, $R_{LOO}^2=1.29\text{\%}$; β_MS-PRS = −0.11, $R_{MS\text{-}PRS}^2=1.87\text{\%}$), WMV (β_LOO = −0.10, $R_{LOO}^2=1.2\text{\%}$; β_MS-PRS = −0.13, $R_{MS\text{-}PRS}^2=2\text{\%}$), thalamus (β_MS-PRS, −0.11, $R_{MS\text{-}PRS}^2=1.46\text{\%}$), putamen (β_LOO = −0.10, $R_{LOO}^2=1.16\text{\%}$; β_MS-PRS = −0.11, $R_{MS\text{-}PRS}^2=1.38\text{\%}$), and CSF (β_LOO = 0.12, $R_{LOO}^2=1.77\text{\%}$; β_MS-PRS = 0.11, $R_{MS\text{-}PRS}^2=1.48\text{\%}$) with at least one of the risk scores (Supplementary Table 6). Replicating these observations in UKBB showed consistency in terms of directionality of associations with EPIC as shown in Supplementary Table 6, but only associations with thalamic (β = −0.15, R² = 3.2%) and putamen (β = −0.15, R² = 3%) volumes remained significant in UKBB (P < 0.05). Furthermore, only CSF volume association with MS-PRS emerged as nominally significant (β = 0.1, R² = 0.3%) in high-risk unaffected subjects in UKBB (top 5% MS-PRS, n = 1495) as shown in Supplementary Table 7.

Next, we investigated the association of PRS with longitudinal per cent change of regional brain volumes at each annual visit relative to the baseline in the UCSF-EPIC dataset over a 10-year follow-up period. Association of PRS_LOO with peripheral grey (β = −0.26) and thalamic atrophy (β = −0.53) remained significant upon Bonferroni correction for multiple testing across phenotypes (P < 0.007, Supplementary Table 8). Model predictions showed the highest atrophy rate in thalamus (Supplementary Table 9). An elevated PRS_LOO was associated with increased thalamic volume loss at each visit (Fig. 4A). Of note, fluctuations in thalamic volume varied over time mainly due to the measurement noise, while group means decreased monotonically (Supplementary Fig. 6). Adjusting for treatment did not affect model predictions since most individuals (>55%) were on ‘platform therapy’ or ‘other’, grouped as one, in each follow-up year (Supplementary Table 10). However, associations may have been partially masked by longer periods of treatment and/or increased frequency of high potency therapy among patients throughout the 10-year course.

Figure 4.

Genetic risk association with clinical and radiographic phenotypes. (A) Thalamic atrophy at each visit against PRS_LOO in the 10-year follow-up period. The blue line is a fit through the raw data at each visit and the shading shows the CI. The black line is the MMRM predictions including measurements from all visits. Different shapes for data-points are used to mark treatment. (B) The forest plot shows Mendelian randomization estimates (standard deviation change in imaging phenotype per doubling in odds of multiple sclerosis) in ENIGMA-CHARGE and UKBB and 95%CI. Inverse variance weighted Mendelian randomization analysis shows a weak effect of multiple sclerosis susceptibility on thalamic volume only in the ENIGMA-CHARGE cohort (P = 0.02). (C) Association of PRS_LOO with disease activity represented by presence of relapses is significant (***P = 0.0002; W Relapse, mean = 0.651, IQR = 1.450, min = −1.886, max = 3.442; W/O Relapse, mean = 0.278, IQR = 1.519, min = −2.474, max = 2.978), but not with EDSS worsening (EDSS Worsening, mean = 0.429, IQR = 1.360, min = −2.348, max = 3.069; No EDSS Worsening, mean = 0.410, IQR = 1.545, min = −2.474, max = 3.442) from baseline to Year 5. EDSS = Expanded Disability Status Scale; IQR = interquartile range; min = minimum; max = maximum; PRESSO = Pleiotropy RESidual Sum and Outlier.

We additionally tested the existence of a putative causal effect of liability to multiple sclerosis on the same baseline imaging phenotypes, except CSF volume, in the ENIGMA-CHARGE cohort (n = 37 741) and broad unaffected UKBB population (n = 31 968) within a two-sample Mendelian randomization framework.^27,28 Summary statistics for left and right parts of thalamus, caudate and putamen volumes as well as BV and WMV were available in UKBB, while only total thalamus, caudate and putamen were available in ENIGMA-CHARGE. For these analyses, we only used genome-wide significant non-MHC disease-associated SNPs (P < 5 × 10⁻⁸) as instrumental variables, to reduce the likelihood of weak instrument bias and distortion from horizontal pleiotropy.^29,30 We found weak evidence of an effect of liability to multiple sclerosis (scaled per doubling in odds) and thalamic volume in the main inverse variance weighted Mendelian randomization analysis in the ENIGMA-CHARGE cohort (β = −0.22, P = 0.02), but not in the UKBB (Fig. 4B and Supplementary Tables 11 and 12). Sensitivity analyses revealed that the association in ENIGMA-CHARGE was likely driven by horizontal pleiotropy (Mendelian randomization-Egger intercept = −0.009, 95% CI: −0.016 to −0.001, P = 0.04) and the effect did not persist using a pleiotropy-robust method (Supplementary Table 11). Results for BV and WMV were not significant in UKBB (Supplementary Table 12). Despite substantial heterogeneity (Q statistic 144 to 207), sensitivity analyses were consistent with a null causal effect.

To further assess the association of polygenic scores with disease activity, UCSF-EPIC participants (n = 464) were divided into two groups according to whether they had experienced one or more relapses within a 5-year interval from the baseline visit, regardless of disease worsening within the same timeframe based on Expanded Disability Status Scale (EDSS) scores. Since no phenotypic information was utilized for the risk score calculations, we examined phenotypic associations with both PRS_LOO and MS-PRS. An increase in both risk scores was associated with relapse activity, which remained significant after correcting for age, sex, and disease duration (β_LOO = 0.34, P = 0.002; β_MS-PRS = 0.26, P = 0.002; Fig. 4C). On the other hand, when EPIC patients were stratified according to EDSS worsening regardless of relapse co-occurrence (Supplementary Table 13), no significant difference was observed between the two groups (Fig. 4C). Multiple sclerosis age of onset is negatively associated with PRS in all datasets, but only significant in UKBB2 (β = −0.006, P = 0.01).

Discussion

Understanding the genetic architecture of polygenic diseases like multiple sclerosis requires accommodating significant variability in the number, relative weight, and ontological type of risk variants each individual carries. Polygenic risk scores developed and tested in different cohorts substantially outperformed previous susceptibility scores, including conventional genetic burdens based on a subset of 233 bona fide multiple sclerosis susceptibility variants.^12,31,32 The ability of PRS to serve as a predictive biomarker for high-risk individuals in population-based cohorts has been demonstrated for multiple diseases,^33–35 and was suggested to facilitate early diagnosis and implementation of preventive or therapeutic interventions.^6,36,37 Interestingly, a recent study showed that polygenic risk profiling can assist in prioritizing individuals with low PRS for identification of rare pathogenic variant heterozygotes.³⁸

Our results demonstrated that including PRS improves risk stratification of basic models including age, sex and established conventional multiple sclerosis risk factors, increasing AUC by up to 0.25 (Table 2 and Supplementary Table 4). Furthermore, comparing similar metrics between population-level UKBB cohort and the KPNC case-control study suggested that despite fundamental differences in the dataset design and compositions, individuals at significantly increased risk could be identified via MS-PRS in both cohorts. The MS-PRS scores explained 7–11% and 12–16% of multiple sclerosis liability, and even higher in extreme PRS deciles, in UKBB2 and KPNC, respectively. UKBB is a remarkable resource, steadily making progress in linking the diagnostic data with other health records in the UK. New disease cases are periodically added to this dataset. Nonetheless, lack of hospitalization records for all multiple sclerosis subjects and possibility of inaccurate self-reports may have resulted in the presence of false negatives in this dataset, affecting modestly the overall precision of MS-PRS. On the other hand, multiple sclerosis cases in the KPNC case-control study are neurologist-diagnosed and the controls are sex, age and locality matched, resulting in the absence of false negatives and improved performance of MS-PRS. Despite these differences affecting the overall AUC of the top MS-PRS, results were consistent between these datasets. Thus, although genetic prediction of future disease status in the general population is not sensitive enough due to the low prior probability to multiple sclerosis, if the target sample is at higher risk according to conventional risk factors and/or those experiencing suggestive symptoms, PRS can assist the diagnosis and the choice of management strategy. Of note, in the EPIC dataset, clinically isolated syndrome (CIS) patients' PRS scores were similar to multiple sclerosis patients but significantly different from healthy controls (Supplementary Fig. 7).

Stratifying disease risk by cellular pathways may provide insights into pathological mechanisms and unravel important biological overlap between different disorders. Pathway-based scores related to regulation of immune response showed the most significant association with disease status, confirming the central role of adaptive immunity in driving multiple sclerosis risk. TCR signalling and MHC class II antigen presentation are among the top disease-related pathways, highlighting the role of canonical antigen presentation processes in multiple sclerosis pathogenesis.³⁹ Dysregulation in specific interleukin-mediated pathways may also contribute to possible imbalance toward pro-inflammatory signals in subjects at risk of multiple sclerosis. For example, our analysis pinpoints IL-12, a master regulator of Th1 responses, and IL-5, a potent chemoattractant and differentiation factor for eosinophils and basophils^40,41 as key underlying processes of risk. In this context, the importance of cell adhesion and ECM organization for lymphocyte extravasation and CNS infiltration is also highlighted.⁴² Regulatory pathways underlying gene expression such as epigenetic processes are also emerging as essential.⁴³ Our data suggest that post-translational regulatory processes, such as defective protein glycosylation, might be equally involved in multiple sclerosis pathogenesis. Notably, aberrant glycosylation patterns can modulate the self/non-self identification of multiple cellular proteins as well as switching antibodies from protective to autoreactive.⁴⁴ The significance of the NOTCH signalling cascade in our analysis strengthens the interface between genetic predisposition and both neurodegeneration and immune response.^45–47 Another significant pathway, the VEGF cascade, similarly modulates both CNS inflammation and neuronal survival in autoimmune demyelination.⁴⁸ In addition, the association of oxidative stress pathways and multiple sclerosis susceptibility may support the role of biological ageing in multiple sclerosis pathology.^49–52 Finally, a robust body of data supports an aetiological and pathological contribution of viral infection, mainly the Epstein–Barr virus, to multiple sclerosis,^53–55 and the results presented here are consistent with an association of viral infection pathway with multiple sclerosis risk.

A higher aggregation of susceptibility variants in multi-case compared to single-case multiple sclerosis families has been previously reported.¹⁰ Yet, these studies showed a limited power in predicting the case-control status.¹² To further our understanding of the heritability patterns within families, we studied a multi-case familial dataset in which one parent and at least one child were diagnosed with multiple sclerosis and incorporated all variants identified by a custom genotyping array for multiple sclerosis.¹⁶ We confirmed that a greater PRS in families of disease-discordant parents is associated with an increased risk of multiple sclerosis among all subjects (AUC = 66%), as well as just among the siblings (AUC = 65%). In this study, the PRS of the affected siblings were significantly higher if either or both parents were at high risk compared to the rest of the cohort. Unaffected siblings at high risk may especially benefit from this knowledge. These results suggest that polygenic profiling provides a compelling opportunity to forecast multiple sclerosis within sibships but needs to be further tested in larger familial cohorts.

Associations between MS-PRS and relapses and regional brain volumes were modest, yet important. Our results suggest a robust association of longitudinal peripheral and deep grey matter atrophy with high genetic predisposition, the strongest association being with thalamic atrophy. Thalamic atrophy is an important marker of multiple sclerosis progression occurring early and declining consistently throughout the course of multiple sclerosis and across clinical subtypes.⁵⁶ Thalamic volume loss in multiple sclerosis patients is associated with decreased neuroperformance in all scales.⁵⁷ Evidence on genetic correlation does not necessarily imply direct genetic modulation of the CNS tissue. Indeed, ours and prior studies in multiple sclerosis described heritability enrichment mainly in immune-related tissues.⁵⁸ The Mendelian randomization analysis in ENIGMA-CHARGE replicated an effect of multiple sclerosis liability on thalamic volume, although the results indicated that this was through horizontal pleiotropy rather than a causal effect. Conversely, in UKBB where there was little evidence of pleiotropy, no Mendelian randomization association was observed. Differences between the two cohorts could be due to averaging bilateral structures in ENIMGA-CHARGE, and its inclusion of case-control studies with psychiatric diagnoses. Taken together, our results indicate that genetic liability to multiple sclerosis is unlikely to cause global or regional subcortical volume changes in the general adult population; rather, the association between MS-PRS and peripheral and deep grey matter atrophy is specific to those with multiple sclerosis. Lastly, we observed a negative association between MS-PRS and age of onset, consistent with a recent report on the genetic underpinning of early disease onset.⁵⁹

In summary, PRS has the advantage of being accessible at any time and incorporating it in the current clinical risk models can be a promising basis for intensive monitoring, reducing modifiable risk factors that may delay the disease onset, and promoting early diagnosis or inform treatment options in cohorts at a higher prior probability, e.g. individuals with suggestive symptoms or those with family history asshown for other diseases.^60,61 Also, considering that multiple sclerosis clinical onset typically occurs between 20 and 40 years of age, implementing preventive strategies for those at higher risk in early adolescence could be an effective strategy to control the rising global incidence of multiple sclerosis and its detrimental consequences. It is noteworthy that diverse population-level GWAS screening, for example in African Americans and Hispanic Americans, is a pressing need in multiple sclerosis genetics and essential for utilizing polygenic profiling in non-European populations. Given the increasing incidence rate of multiple sclerosis, PRS can play an important role in future public health as a part of multifactorial predictive models along with modifiable lifestyle factors, family history, and rare variations. Therefore, this study is an important step towards translating GWAS studies into relevant biology and clinically meaningful outcomes.

Appendix 1

UCSF-EPIC Team in alphabetical order: Jessa Alexander, Riley Bove, Sergio Baranzini, Bruce A. C. Cree, Eduardo Caverzasi, Richard Cuneo, Stacy J. Caillier, Tiffany Cooper, Ari J. Green, Chu-Yueh Guo, Jeffrey M. Gelfand, Refujia Gomez-O’shea, Sasha Gupta, Jill Hollenbach, Meagan Harms, Roland G. Henry, Stephen L. Hauser, Myra Mendoza, Jorge R. Oksenberg, Nico Papinutto, Sam Pleasure, Kyra Powers, Adam Renschen, Adam Santaniello, Joseph J. Sabatino Jr., William A. Stern, Michael R. Wilson, Scott S. Zamvil.

Funding

This study was supported primarily by grants from the National Multiple Sclerosis Society RG-1707-28775 to R.G.H. and J.R.O., National Multiple Sclerosis Society RFA-2104-37474 to J.R.O., and National Institutes of Health R35NS111644 and the Valhalla Foundation to S.L.H., as well as National Institutes of Health R01NS099240 to S.E.B. and National Institutes of Health R01ES017080, R01AI076544, R01NS049510 to L.B. Also, N.A.P. is supported by RG-1707-28657 and JF-1808-32223, Harry Weaver Award from the National Multiple Sclerosis Society. H.S. is supported by a postdoctoral fellowship from National Multiple Sclerosis Society (FG-1807-31603). This publication is also supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through UCSF-CTSI Grant Number TL1 TR001871 to H.S. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the NIH.

Competing interests

The authors have no conflicts of interest to declare.

Supplementary material

Supplementary material is available at Brain online.