Not all roads lead to the immune system: The Genetic Basis of Multiple Sclerosis Severity Implicates Central Nervous System and Mitochondrial Involvement

Abstract

Multiple sclerosis (MS) is a leading cause of neurological disability in adults. Heterogeneity in MS clinical presentation has posed a major challenge for identifying genetic variants associated with disease outcomes. To overcome this challenge, we used prospectively ascertained clinical outcomes data from the largest international MS Registry, MSBase. We assembled a cohort of deeply phenotyped individuals with relapse-onset MS. We used unbiased genome-wide association study and machine learning approaches to assess the genetic contribution to longitudinally defined MS severity phenotypes in 1,813 individuals. Our results did not identify any variants of moderate to large effect sizes that met genome-wide significance thresholds. However, we demonstrate that clinical outcomes in relapse-onset MS are associated with multiple genetic loci of small effect sizes. Using a machine learning approach incorporating over 62,000 variants and demographic variables available at MS disease onset, we could predict severity with an area under the receiver operator curve (AUROC) of 0.87 (95% CI 0.83 - 0.91). This approach, if externally validated, could quickly prove useful for clinical stratification at MS onset. Further, we find evidence to support central nervous system and mitochondrial involvement in determining MS severity.

Full text

1
Not all roads lead to the immune system: The Genetic
Basis of Multiple Sclerosis Severity Implicates Central
Nervous System and Mitochondrial Involvement
Vilija G. Jokubaitis1,2,3,4*, Omar Ibrahim1, Jim Stankovich1, Pavlina Kleinova5, Fuencisla
Matesanz6, Daniel Hui7, Sara Eichau8, Mark Slee9, Jeannette Lechner-Scott10,11, Rodney
Lea12, Trevor J Kilpatrick4,13, Tomas Kalincik4,14, Philip L. De Jager15, Ashley Beecham16,
Jacob L. McCauley16, Bruce V. Taylor17, Steve Vucic18, Louise Laverick3, Karolina
Vodehnalova5, Maria-Isabel García-Sanchéz19, Antonio Alcina6, Anneke van der
Walt1,2,3,4, Eva Kubala Havrdova5, Guillermo Izquierdo8,20, Nikolaos Patsopoulos7, Dana
Horakova5#, Helmut Butzkueven1,2,3,4#
*corresponding author
#contributed equally
1Department of Neuroscience, Central Clinical School, Monash University, Melbourne, Australia
2Department of Neurology, Alfred Health, Melbourne, Australia
3Department of Medicine, University of Melbourne, Melbourne, Australia
4Department of Neurology, Melbourne Health, Melbourne, Australia
5Department of Neurology and Centre of Clinical Neuroscience, First Faculty of Medicine, Charles
University and General University Hospital, Prague, Czech Republic
6Instituto de Parasitología y Biomedicina López Neyra, CSIC, Granada, Spain
7Brigham and Women’s Hospital, Harvard Medical School, MA, USA
8Hospital Universitario Virgen Macarena, Sevilla, Spain
9College of Medicine and Public Health, Flinders University, Adelaide, Australia
10Department of Neurology, John Hunter Hospital, Newcastle, Australia
11School of Medicine and Public Health, University of Newcastle, Newcastle, Australia
12Genomics Research Centre, Centre of Genomics and Personalised Health, Queensland University of
Technology, Australia
13Melbourne Neuroscience Institute, University of Melbourne, Melbourne, Australia
14CORe, Department of Medicine, University of Melbourne, Australia
15Multiple Sclerosis Center and the Center for Translational & Computational Neuroimmunology,
Department of Neurology, Columbia University, New York, NY, USA
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.

2
16John. P. Hussman Institute for Human Genomics, Miller School of Medicine, University of Miami, FL,
USA
17Menzies Institute for Medical Research, University of Tasmania, Hobart, Australia
18Westmead Institute, University of Sydney, Sydney, Australia
19UGC Neurología. Hospital Universitario Virgen Macarena, Nodo Biobanco del Sistema Sanitario
Público de Andalucía, Sevilla, Spain
20Fundación DINAC, Sevilla, Spain
* Corresponding author: Dr Vilija Jokubaitis
Department of Neuroscience, Central Clinical School, Monash University
The Alfred Centre, Level 6, 99 Commercial Rd, Melbourne, VIC 3004, Australia
vilija.jokubaitis@monash.edu
Abstract count: 187
Word count: 4000 (excluding methods)
References main body: 68
References total (including online methods): 86
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

3
Abstract
Multiple sclerosis (MS) is a leading cause of neurological disability in adults. Heterogeneity in MS
clinical presentation has posed a major challenge for identifying genetic variants associated with
disease outcomes. To overcome this challenge, we used prospectively ascertained clinical outcomes
data from the largest international MS Registry, MSBase. We assembled a cohort of deeply
phenotyped individuals with relapse-onset MS. We used unbiased genome-wide association study
and machine learning approaches to assess the genetic contribution to longitudinally defined MS
severity phenotypes in 1,813 individuals. Our results did not identify any variants of moderate to
large effect sizes that met genome-wide significance thresholds. However, we demonstrate that
clinical outcomes in relapse-onset MS are associated with multiple genetic loci of small effect sizes.
Using a machine learning approach incorporating over 62,000 variants and demographic variables
available at MS disease onset, we could predict severity with an area under the receiver operator
curve (AUROC) of 0.87 (95% CI 0.83 – 0.91). This approach, if externally validated, could quickly
prove useful for clinical stratification at MS onset. Further, we find evidence to support central
nervous system and mitochondrial involvement in determining MS severity.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

4
Multiple sclerosis (MS), a complex trait disease, is a leading cause of non-traumatic neurological
disability in adults. It affects approximately 2.8 million people worldwide, predominantly females.1
Rates of disability progression and long-term outcomes are highly heterogeneous amongst people
with relapse-onset MS (RMS).2 At present, the ability to predict a person’s likely long-term disease
outcome at onset is very limited, but highly desirable, in order to stratify individuals for initiation with
the most appropriate disease-modifying therapy.
To-date, over 230 common variants have been linked to MS risk.3 The only replicated genetic modifier
of MS phenotype is carriage of the principal risk allele, the human leukocyte antigen (HLA)
DRB1*15:01. In European populations, carriage of the HLA-DRB1*15:01 allele confers younger age of
onset.4 However, large studies have shown that this allele has no effect on MS progression after
onset.5, 6 Further, there is strong evidence to suggest that currently known risk variants, aside from
HLA-DRB1*15:01, play no major role in determining MS severity.7-9
A genetic influence on MS outcome is, however, plausible, in particular relating to the severity of
secondary inflammation (e.g. development of slowly expanding, or chronic rim-active lesions),
resilience to neuroaxonal injury, or remyelination capacity. Indeed, preliminary genome-wide
association study (GWAS) evidence suggests that susceptibility and severity likely involve distinct
biological processes and pathways.10-12
The best evidence to-date for a genetic contribution to disease outcomes comes from a small number
of cross-sectional GWAS dedicated to a search for severity signals associated with the MS severity
scale (MSSS) score13, or age at onset.9, 10, 14-17 However, these signals failed to reach significance at the
genome-wide level, possibly due to inclusion of populations with both relapse-onset and progressive-
onset clinical courses. As the genetic architecture underlying worsening in relapse-onset MS and
progressive-onset MS is possibly distinct,18 it could be important to study these populations
separately. Further, use of limited cross-sectional phenotypic MSSS data to assess disease severity
limits accurate severity phenotyping due to both major ceiling effects and instability in RMS.13
The heterogeneity in MS severity, both between individuals, and within individuals over time, is large.
Therefore, analysis of longitudinally acquired clinical trajectories over many years is likely to be more
reliable for accurate severity assignation. Given that preliminary evidence suggests that genetic
variation influences severity outcomes, we used both unbiased genome-wide association, and
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

5
machine learning approaches to examine 1) whether prospectively ascertained, longitudinally-defined
RMS phenotypes could reveal novel genetic variants associated with disease severity, 2) whether a
machine learning model with multi-single nucleotide variant (SNV) inclusion has sufficient positive
predictive value to potentially be used at the time of MS diagnosis to guide clinical and treatment
decisions. Our secondary analyses further interrogated SNV signals derived from our primary analyses,
and also aimed to replicate previously reported suggestive markers of MS severity using a targeted
approach.
Results
Cohort characteristics
The cohort comprised of 5,851 people with relapse-onset MS from Australia, the Czech Republic, and
Spain (Figure S1). Those who met study minimum inclusion criteria (Figure S2), represented 63,072
patient-years of follow-up. Of these, 1,984 (33.9%) people were genotyped, of whom 1,813 (91.4%),
representing 22,884 patient-years of follow-up, passed additional filtering and genotyping quality
control (QC; Table S1). The clinical and demographic characteristics of the cohort based on
longitudinal age-related MS severity scale19 (l-ARMSS) scores (Table 1), and longitudinal MSSS (l-MSSS;
Table S2) are shown. Per-country cohort characteristics are provided in Table S3. Individual
phenotypes based on continuous l-ARMSS and l-MSSS, binary l-ARMSS and l-MSSS, Age at Onset
(AAO), and MS susceptibility weighted genetic risk scores (wGRS) are available in Table S4. The
correlation between l-ARMSS and l-MSSS was strong (r=0.90, p<0.0001, Table S4). l-ARMSS and l-
MSSS scores in individual disease trajectories are shown in Figure 1.
Primary analyses
Genome-wide association search for SNVs associated with longitudinal severity measures
We first performed a genome-wide association analysis to identify novel variants associated with l-
ARMSS continuous (Table S5) or binary (Table S6) phenotypes. Cohort characteristics described in
Table 1 demonstrated that those in the severe l-ARMSS cohort had longer follow up (12.5 years v 11.2
years, Cohen’s d = 0.32), longer symptom duration (22.2 years v 16.2 years, Cohen’s d=0.61), a younger
age at onset (27.2 years vs 32.3 years, Cohen’s d =0.58), a higher annualised relapse rate (0.14 v 0.10,
Cohen’s d=0.43), a lower cumulative proportion of time exposed to disease-modifying therapy (60.6%
vs 79.9%, Cohen’s d=0.22), and a higher wGRS (2.90 vs 2.71, Cohen’s d=0.23) relative to the mild
cohort. Therefore, all regression analyses of l-ARMSS phenotypes were a priori adjusted for the first 5
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

6
principal components (PCs), the number of HLA-DRB1*15:01 alleles carried, percentage of time
exposed to disease-modifying therapy and imbalanced variables as above. Fixed-effects meta-analysis
results from six groups (two from each country) did not identify any single nucleotide variants (SNVs)
that surpassed genome-wide significance (p<5x10-8) for any of the above phenotypic outcomes (Table
S5; Figure S3). Similarly, adjusted I-MSSS analyses (Table S2) did not identify any significant
associations, related to continuous (Table S7; Figure S4), or binary (Table S8) phenotypes. Assessment
of the genomic locations of SNVs with p<1x10-5 for the l-ARMSS phenotype demonstrated that 55.1%
(p=6.63x10-3 for enrichment) of the signals were in intergenic regions and 34.7% were intronic (Figure
S5a). This was numerically different to the I-MSSS endpoint analysis, where 47.9% of SNVs were
intronic (p=5.41x10-3 for enrichment) and 36.8% were intergenic (p=0.0193; Figure S5b).
A summary of the top variants associated with the continuous l-ARMSS and l-MSSS analyses are shown
in Table 2. The top signal in the continuous l-ARMSS analysis was rs7289446 (b=-0.4882, p=2.73x10-
7), intronic to SEZ6L, a gene associated with dendritic spine density and arborization.20 The top signal
in the continuous l-MSSS analysis also implicated a variant intronic to SEZ6L, rs1207401 (b=-0.4841,
p=2.88x10-7). These two SEZ6L associated SNVs are in perfect linkage disequilibrium (R2=1, D’=1; Table
S9; Figure S6).
Heritability analyses
To estimate the extent to which the variability of l-ARMSS or l-MSSS-defined severity could be
explained by genetic architecture, we calculated narrow-sense heritability estimates (h2g) for our
cohort (n=1,813). The overall heritability estimate for the l-ARMSS phenotype was h2g 0.19 (SE 0.15,
p=0.02) using the GREML by GCTA method. Similar estimates were achieved using alternate
heritability estimate tools (Table S10). The overall l-MSSS h2g heritability estimate was slightly greater
than for l-ARMSS (h2g 0.29; SE 0.14, p=0.001). However, alternate heritability estimates for l-MSSS
proved highly inconsistent (Table S10).
Machine Learning
Given that, as expected, our unbiased GWAS approach did not identify any SNVs that surpassed GWAS
significance thresholds, we implemented a non-linear, xgboost21 machine learning (ML) algorithm to
determine whether a non-linear ML model could find genetic associations with severity as compared
to traditional GWAS approaches. We input all SNVs with an l-ARMSS GWAS p<0.01, accounting for
62,351 variants. However, no single variant was given a weight of greater than 0.005, confirming that
no genetic variant contributed appreciably to MS severity.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

7
Prediction of clinical course using machine learning
Recognising that our ML model did not further illuminate the underlying genetic architecture of MS
severity, we further sought to determine whether it could be used to predict severity, based on l-
ARMSS outcome extremes (n=447 mild, n=464 severe). Our ML algorithm trained on 70% (n=638 l-
ARMSS) of the cohort, then tested the remaining 30% (n=273 l-ARMSS). Our ML classification
algorithm had high predictive accuracy, with an area under the receiver operating characteristic curve
(AUROC) 0.85 (95% CI 0.80 – 0.89). The addition of AAO together with MS susceptibility wGRS (Table
S4) further boosted the ML AUROC to 0.87 (95% CI 0.83 – 0.91; Figure 2). Our classification algorithm
had 86% sensitivity, and 88% specificity, with a positive predictive value (PPV) of 89% and negative
predictive value (NPV) of 85%. Severity classification based on l-MSSS phenotype was weaker, with an
AUROC of 0.85 (95% CI 0.80-0.89), 98% sensitivity, but only 68% specificity; with a PPV of 76% and
NPV of 97% (Figure 2).
Restricting the ML algorithm to just those SNVs with p<1x10-5 in the l-ARMSS GWAS (n=336) decreased
predictive accuracy to AUROC =0.79 (95% CI 0.74 – 0.84) confirming the polygenic nature of the
genetic architecture underlying MS severity.
Secondary analyses
Sex-stratified Genome-wide association search for SNVs associated with longitudinal severity
measures
Given our primary analyses did not identify signals of genome-wide significance, we performed sex-
stratified analyses to determine whether any variant effects were potentially sex-associated. Table 2
summarises the SNVs nearest the top 5 gene regions for each sex. The top hit in the female l-ARMSS
analysis, rs1219732 intronic to CPXM2 (bfemale =0.5693, p=6.48x10-08), approached genome-wide
significance (Table S11). This variant also approached significance in association with l-MSSS (bfemale
=0.5447, p=1.89x10-07, Table S12). We also found rs10967273, an intergenic variant, was associated
with l-MSSS-defined severity in females (bfemale =0.8289, p=3.52x10-08, Table S12; Figure S7). However,
this variant did not surpass significance thresholds (bfemale =0.7994, 1.17x10-07) in the l-ARMSS analysis.
In males, the top hit in the l-ARMSS analysis was rs7315384, intronic to STAB2 (bmale= 1.04, p=1.29x10-
07), followed by rs7070182, intronic to TCF7L2 (bmale =-1.11, p=3.65x10-07; Table S13). The l-MSSS
analysis in males identified rs698805 intronic to CAMKMT (bmale =-1.5395, p=4.35x10-08) as associated
with severity (Table S14; Figure S8). This variant did not surpass GWAS thresholds in the l-ARMSS
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

8
analysis (bmale =-1.4199, p=5.64x10-06, Table S13). The variants identified in our sex stratified analyses
were not associated with severity in the opposite sex (Table S15). We did not identify any novel
genetic associations with age at onset (Table S16; Figure S9).
Pathway analyses
To identify potential biological processes overrepresented in our analyses, we analysed the top
suggestive SNVs (p< 1x10-5) using gene-set enrichment analyses. We used FUMA22 to assign SNVs to
genes and tissues, and genes to functions. Tissue enrichment implemented in FUMA revealed an over
representation of cerebellar cortex-expressed genes for both l-ARMSS (cerebellar hemisphere
p=0.071; cerebellum p=0.077; Figure S10a) and l-MSSS (cerebellar hemisphere p=0.017; cerebellum
p=0.023); Figure S10b). In contrast, whole blood-associated genes were not enriched in our analyses
with either l-ARMSS (p=0.75) or l-MSSS (p=0.82) outcomes. Gene set enrichment analyses of the l-
ARMSS phenotype implicated endothelial cell development (b=0.43; p=8.25x10-05), pseudopodium
assembly (b=0.84; p=2.37x10-04), response to progesterone (b=0.35; p=2.74x10-04) and NMDA
receptor activity (b=1.10; p=2.79x10-04). The l-MSSS phenotype was additionally enriched for Wnt
signalling pathways (b=0.24; p=2.02x10-04; Table S17). We also examined gene set enrichment using
Panther. Here we corroborated an overrepresentation of heteromeric G-protein signalling pathways
associated with l-ARMSS (p = 4.98x10-05, FDR = 8.23x-10-03) and l-MSSS (p = 1.00x10-04, FDR = 1.67x10-
02) phenotypes. The AAO phenotype was associated with endothelin (p = 2.51x10-04, FDR = 4.18x10-02),
and cadherin signalling pathways (p = 2.90x10-04, FDR = 2.42x10-02).
Survival Analyses
Given our primary GWAS analyses did not reveal SNVs that surpassed the genome-wide level of
statistical significance, we assessed whether 30 of the top signals (Table 2) might play a role in severity
modulation using an alternative definition of severity, making use of our longitudinal dataset. Here
we assessed the time to reach the hard disability milestones of irreversible expanded disability status
scale (EDSS) score 3 (irEDSS3) and irreversible EDSS 6 (irEDSS6) in both univariable and adjusted
analyses (Table S18). We identified four SNVs that were associated with time to reach irreversible
EDSS 3 and 6 in both unadjusted and adjusted analyses including: rs7289446 (intronic to SEZ6L),
rs295254 (intronic to RCL1), rs111399831 (nearest to SUCLA2), rs61578937 (nearest to NCOA2). These
SNVs were then combined in multivariable analyses to determine whether they could independently
predict time to disability milestones (Table S19). Three SNVs remained independently predictive of
both time to irreversible EDSS 3 and 6 (Figure 3), including rs7289446 (SEZ6L: irEDSS3 adjusted HR
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

9
(aHR) 0.77, p=0.008, Figure 3a; irEDSS6 aHR 0.72, p=4.85x10-4, Figure 3b), rs295254 (RCL1: irEDSS3
aHR 1.33, p=9.10x10-4, Figure 3c; irEDSS6 aHR 1.32, p=7.37x10-4, Figure 3d) and rs111399831 (irEDSS3
aHR 0.77, p=0.036, Figure 3e; irEDSS6 aHR 0.62, p=2.82x10-4, Figure 3f).
In the sex-stratified analyses, we identified rs9643199 (intronic to MTSS1) and rs2776741 (nearest to
RCAN3AS) as consistently associated with time to irreversible EDSS 3 and 6 in females (Figure 4), but
not males (Table S18, Figure S9). The independent hazards of time to reach irreversible EDSS 3 and 6
for these variants were: rs9643199 (MTSS1: irEDSS3 aHR 1.35, p=0.006, Figure 4a; irEDSS6 aHR 1.46,
p=7.09x10-4, Figure 4b) and rs2776741 (irEDSS3 aHR 0.77, p=0.010, Figure 4c; irEDSS6 aHR 0.74,
p=0.005, Figure 4d; Table S19). rs7070182 intronic to TCF7L2 (Figure 4) was the only variant
consistently associated with time to irreversible EDSS 3 (aHR 0.59, p=0.013, Figure 4e) and 6 (aHR
0.56, p=0.005, Figure 4f) in males, with no effect in females (Table S18; Figure S11).
MS susceptibility allele association with severity phenotypes
We sought to determine whether there was an association between known MS susceptibility
variants (wGRS), and our severity phenotypes of interest. We found weak positive correlations
between the MS susceptibility wGRS (Table S20) and l-ARMSS (r=0.07, p=0.003, Figure S12a); l-MSSS
(r=0.03, p=0.19 Figure S12b); and a weak negative correlation with AAO (r=-0.08, p=0.0005) (Figure
S12c). We did not find an association between l-AMRSS or l-MSSS and the known non-HLA
autosomal risk variants3 that were directly genotyped (198/200), p>1x10-3 (Table S20).
The distribution, per phenotype, of HLA MS susceptibility tagging SNVs including HLA-DRB1*15:01,
HLA-DRB1*03:01, HLA-DRB1*130:01, HLA-DRB1*08:01, HLA-DQB1*03:02, and protective alleles:
HLA-A*02:01, HLA-DQA1*01:01, HLA-B*44:02 and HLA-B*55:01 is described in Table S21. We
confirmed that HLA-DRB1*15:01 homozygosity was associated with an earlier AAO (rs3135388, 29.2
years v 30.4 years, p=0.005). However, homozygosity at HLA-DRB1*15:01 was not associated with
disease severity as per l-ARMSS, nor l-MSSS, nor was any other SNV-genotyped HLA allele (Table
S21).
Validation assessment of previously published putative severity SNVs
In addition to the main European DRB1*15:01 tagging SNV, rs3135388, we tested 116 putative non-
HLA SNV associations with cross-sectional MSSS measures, disease severity, and AAO. We were able
to replicate the association between rs868824, intronic to IMMP2L on chromosome 7, with AAO (b =
-1.0935 years; p=4.31x10-4), however, no other putative severity variant met or surpassed the
Bonferroni-corrected replication threshold (p=4.31x10-4, Table S22).
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

10
Finally, we tested the association between a variant intronic to LRP2 (rs12988804), previously
reported to be associated with relapse risk,23 and annualised relapse rate (ARR). However, we were
unable to find an association between this variant and ARR in our targeted analysis (p = 0.925).
Discussion
Two of the fundamental, unanswered questions with respect to relapsing-remitting MS are first, what
is the source of the marked clinical disease heterogeneity? That is, why do some people with RMS
have a rapidly progressing, severely disabling disease course, whilst others do not? And second, can
we utilise genetic and other information to predict MS outcomes?
Here, through a series of analyses that took advantage of a unique, multicentre, prospectively
ascertained, longitudinal, clinical dataset,24 we can shed some light on the genetic architecture that
underpins MS clinical heterogeneity. Our primary, unbiased GWAS analyses demonstrate that there
are no common variants with moderate to large effect sizes that contribute to MS severity. With time,
and very large cohorts, we will likely confirm that MS severity is at least partially determined by
polygenic mechanisms of small effect size. Alternatively, we may find that variants which influence
severity may be time-variable, rather than having a constant effect.25 Importantly, our results suggest
that disease outcomes are not under strong genetic control. Indeed our study results demonstrated
that common genetic variants explained only 20% of severity heritability, with wide error margins.
Therefore, suggesting that, as clinical experience shows, outcomes are, to an extent, modifiable with
appropriate and early disease-modifying therapy (DMT) intervention.26-28 This is further underscored
in the modern era where, with the introduction of DMT, rates of disability accumulation have slowed,
and fewer disabling cases are being seen, relative to historical cohorts.29-33 Future pharmacogenomic
studies34 may prove to be invaluable to guide precise prescription practices to further slow
progression or modify disease outcomes.
The complex interplay between genes and the environment likely additionally plays a significant role
in outcome modulation. It has been shown that disability accumulation may be modified by additional
lifestyle factors such as pregnancy27 and smoking cessation.35 Epigenetic studies may therefore shed
further light on relevant, modifiable mechanisms that regulate MS outcomes.
The application of machine learning to GWAS data is considered by some36 to be the “last hope” to
gain meaningful insights for complex diseases where no variants meet significance thresholds.
Our machine learning algorithm was unable to provide additional biological insights into the
underlying genetic architecture of MS severity, instead reinforcing that common SNVs independently
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

11
contribute miniscule weights towards determining MS severity. Regardless, machine learning was
able to predict non-linear effects and large SNV clusters that can accurately classify MS outcomes, and
may prove to be of prognostic utility. Whilst we retained the MS susceptibility wGRS in our algorithm,
it was not a major contributor to our predictive model, again consistent with our above findings, and
past reports demonstrating that MS risk variants have little influence on severity outcomes.7 The
classification accuracy of our machine learning algorithm increased with the addition of age at onset.
In fact, age at onset was one of the strongest predictors of outcome in our machine learning models,
consistent with past reports.27 Our machine learning algorithm was designed with internal checks to
prevent data over fitting. We used a slow learning rate, with a 70/30 training/testing set and internal
bootstrapping over 20,000 learning iterations. We achieved positive predictive values for outcome
assignation of between 0.889 – 0.844 and negative predictive values in the range of 0.851 – 0.853. To
our knowledge, this is a world first for MS genetic studies. Whilst a previous ML study successfully
predicted MS severity,37 this was predicated on health records, and data that take years to decades to
obtain e.g. change in clinical parameters between years ‘x’ and ‘y’ to predict ‘z’. The variables included
in our classification algorithm are readily available at disease onset. With the rapid decrease in the
cost of beadchip genotyping, and high PPV and NPV we achieved, our machine learning algorithm
could readily translate into clinical practice upon validation in an independent cohort.
In our secondary analyses, we replicated the association between the main MS risk allele HLA-
DRB1*15:01 and age at onset.4, 6 Further, ours is the first study to replicate rs868824, intronic to
IMMP2L,10 as being associated with age at onset in a targeted analysis. IMMP2L, an inner
mitochondrial membrane protease, has been associated with cellular senescence,38 ovarian aging via
oxidative stress and estrogen-mediated pathways,39 together with neurological disorders.40-43 Recent
evidence points to accelerated cellular senescence and biological aging in people MS,44-46 and suggests
that these factors may reduce remyelination capacity.45 The validation of the association of rs868824
with age at onset, is a first step towards understanding the potential biological mechanisms underlying
accelerated cellular senescence in MS.
The integration of the top SNVs identified in our de novo GWAS analyses into hard EDSS disability
milestone survival analyses, again identified variants intronic to or near genes implicated in
mitochondrial function: rs111399831, nearest to SUCLA2, and rs9643199 intronic to MTSS1; as well
as variants intronic to genes implicated in CNS function: rs7289446 intronic to SEZ6L, rs295254
intronic to RCL1, rs9643199 intronic to MTSS1, rs2776741 nearest to RCAN3AS, and rs7070182
Intronic to TCF7L2. The latter three having sex-specific effects. The hazard ratios associated with
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

12
reaching irreversible EDSS 3 or 6, conferred by carriage of the minor allele at each SNV, were
consistent with the effect sizes of these variants in our GWAS analyses: that is, effect sizes were small,
but significant in survival analyses. Most importantly, these variants identify highly biologically and
clinically plausible leads for potential replication of clinical heterogeneity in similar or larger cohorts.
SUCLA2 is expressed in the brain and muscle, and encodes the beta-subunit of succinate-CoA ligase,
an enzyme required for the maintenance of mitochondrial DNA.47 Variations in MTSS1 have been
reported to associate with mitochondrial complex 1 deficiency in ClinVar (SCV001137705.1,
SCV001137706.1). The variants we describe here are intronic to, or near to these genes, and are likely
to be tagging rather than causal. However, together with IMMP2L, we describe three variants
associated with mitochondrial function, where mitochondrial dysfunction is a recognised
pathophysiological hallmark of CNS injury in MS.48, 49
Whilst MTSS1 has been reported to associate with mitochondrial function, it has primarily been
described in the context of B-cell mediated immunity,50 and various CNS pathologies.51,52 Most
relevant perhaps to MS outcomes, is the association between MTSS1, cortical volume, and Purkinje
cell dendritic arborization.53, 54 SEZ6L, the top signal in both l-ARMSS and l-MSSS analyses, is
implicated in dendritic spine density variation, and arborization in the hippocampus and
somatosensory cortex.20 Disruptions in SEZ6L cause neurodevelopmental, psychiatric, and
neurodegenerative conditions, as well as having a role in motor function.20, 55, 56 Copy number
variation in RCL1, has also been associated with severe psychiatric disease,57 and depression.58
Progressive synaptic loss, or synaptopathy, is a hallmark of MS pathology;59, 60 evident in both
acutely active demyelinating lesions,61 as well as chronic inactive lesions.62 It has been shown that
loss of synaptic density is associated with network dysfunction,60 implicating a failure of synaptic
plasticity to compensate for immune-mediated neural damage. It is therefore plausible that the
variants identified in this study implicate a genetic susceptibility to impaired compensatory
mechanisms, or impaired neural survival in those with severe MS. This of course requires
independent validation but raises an intriguing new line of enquiry.
Interestingly, we identified a variant intronic to TCF7L2 as associated with severity in males. TCF7L2
is a transcription factor involved in Wnt signalling pathways, and associated with
neurodevelopmental disorders.63 Critically in the context of our study, TCF7L2 has been shown to
maintain oligodendrocyte progenitor cells in the progenitor state, acting as a molecular switch that
can inhibit Wnt signalling to promote oligodendroglial differentiation.64 Why this variant was
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

13
associated with severity only in males in our study is unclear. However, the role of TCF7L2 in MS
severity requires further investigation. Interestingly, gene set enrichment analyses revealed that
both Wnt signalling pathway components together with progesterone response pathways were
enriched in our analyses. Progesterone is a known regulator of myelin development, as well as
having neuroprotective effects,65 lending support to the notion that genetic susceptibility to
impaired remyelination predisposes to more severe MS.
Our tissue enrichment analyses specifically pointed towards genes enriched in cerebellar function.
The cerebellum plays a key role in motor coordination as well as cognition.66 It has long been held26
and recently confirmed,67 that cerebellar signs and symptoms are a predictor of poor prognosis in
MS. The results of our analyses therefore point to highly relevant and biologically plausible genetic
explanations for clinically observed disease heterogeneity.
Our multicentre study was conducted using rigorously defined and prospectively collected
longitudinal clinical and treatment data from the MSBase Registry, making our cohort globally
unique. Due to the nature of this cohort, we were unable to validate our results in an equivalent
dataset, therefore, the data presented herein require independent validation. We did try to
overcome this limitation by testing top SNV signals using alternate definitions of disease severity,
namely survival analyses of time to irreversible disability milestones. Similarly, our ML analyses were
performed using a conservative 70/30 training/testing split relative to the traditional 80/20 split,
accompanied with internal bootstrapping. Our efforts to expand our cohort for future analyses are
ongoing.
Here we report an important milestone in our progress towards understanding the clinical
heterogeneity of MS outcomes, implicating functionally distinct mechanisms to MS risk. We
demonstrate that common genetic variants of moderate to large effect sizes do not contribute to MS
severity. In secondary sex-stratified analyses, we identified two genetic loci that surpassed GWAS
significance thresholds, providing evidence of sex dimorphism in MS severity. We identified a further
six variants with strong evidence for regulating clinical outcomes. We observed an overrepresentation
of genes expressed in CNS compartments generally, and specifically in the cerebellum. These involved
mitochondrial function, synaptic plasticity, cellular senescence, calcium and g-protein receptor
signalling pathways. Importantly, we demonstrate that machine learning using common SNV clusters,
together with clinical variables readily available at diagnosis can improve prognostic capabilities at
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

14
diagnosis, that which goes beyond T2 MRI lesion load,68 and with further validation has the potential
to translate to meaningful clinical practice change.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

15
Methods
Study population:
Participants were recruited from eight tertiary-referral MS-specialist centres, from 3 countries
(Australia, Spain, and Czech Republic), participating in the MSBase Registry. MSBase is an
international, prospective, observational, MS clinical outcomes registry, registered with the World
Health Organization International Clinical Trials Registry Platform, ID ACTRN12605000455662.24 Data
are entered by neurologists in, or near real-time including: participant demographics, disease
phenotype, expanded disability status scale (EDSS) scores, relapse information, and disease modifying
therapy use. Clinical assessments occur on average every 6 months.
Ethics approvals:
This study was approved by the Melbourne Health Human Research Ethics Committee, and by
institutional review boards at all participating centres. All participants gave written informed consent
for participation in the MSBase Registry, together with additional informed consent to participate in
genetic research (HREC/13/MH/189 and per local approvals elsewhere).
Study Inclusion Criteria:
People with MS (pwMS) of European ancestry with clinically definite, relapse-onset MS, based on
McDonald criteria69-71 and participating in MSBase. Further, minimum inclusion criteria comprised:
sex, birthdate, age at onset, ³5 years of symptom duration; ³5 years prospective follow-up in MSBase;
³3 EDSS scores recorded in the absence of a relapse (defined as EDSS scores recorded within 30 days
of relapse onset date); availability of relapse and treatment history. Symptom duration was calculated
based on the most recently recorded EDSS visit.
Phenotyping, severity assignation and recruitment:
Data used for phenotyping pwMS were extracted from the registry on 4 September 2019. EDSS scores
recorded in the absence of a relapse were used to calculate an age-related MS severity (ARMSS)19
score and MS Severity Scale13 (MSSS) score. It has been demonstrated that at the individual level, that
cross-sectional MSSS cannot be used for prognostication,13 but that longitudinal MSSS scores may be
less noisy in individual prognostics.72 Therefore, for each participant, we calculated the median
longitudinal ARMSS (l-ARMSS) and median longitudinal MSSS (l-MSSS) using each available ARMSS or
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

16
MSSS score (minimum 3 scores, Tables 1 & S2). Median relapse-independent l-ARMSS and l-MSSS
scores were then divided into quintiles to stratify the cohort for severity. The top and bottom quintiles
were defined as outcome extremes. Our definitions of mild and severe disease were based on all
individuals meeting minimum inclusion criteria (n=5,851; Figure S1). Participant recruitment was then
enriched for those at outcome extremes. Lists of study participants in the top and bottom decile of
severity were sent to centre PIs to ensure accurate diagnosis. In cases of diagnostic uncertainty, or re-
classification (e.g. ADEM, primary progressive MS), these pwMS were excluded from our study. A
further age at onset (AAO) phenotype was defined as age at first symptom onset.
Symptom duration was defined as the number of years between first symptom onset and the most
recently recorded clinical visit reported by a neurologist in MSBase. Follow-up was defined as the
number of years between first neurologist recorded visit in MSBase, and the most recent clinical visit.
Percentage of time exposed to disease-modifying therapy (DMT) was defined as the total time
exposed to any approved MS DMT as a percentage of symptom duration, as recorded in MSBase.
Statistical analyses:
Data processing and statistical analyses were performed in Stata v17 (Stata Corp, College Station, TX)
or R (http://R-project.org). Monash high performance computing infrastructure through MASSIVE was
used for big data manipulation and computationally extensive analyses.73 Continuous variables were
assessed for normality using the Shapiro-Wilk normality test, and described as mean and standard
deviation (SD) or median with interquartile range (IQR), as appropriate. Categorical variables were
described using frequencies. Standardised differences between key demographic and clinical variables
were assessed using the Cohen’s d statistic. Correlations between l-ARMSS and AAO and the weighted
genetic risk score (wGRS) were assessed using Pearson’s correlation coefficients. All analyses were 2-
tailed.
Genotyping, imputation and quality control:
Detailed methodology can be found in supplementary materials and methods. Briefly, whole blood
gDNA was genotyped using the Illumina MegaEx BeadChip array. This array was customized with an
additional 3,000 single nucleotide variants (SNVs) of interest including: known MS risk SNVs, a suite of
tag SNVs to classical HLA alleles,11 previously published putative severity SNVs,10, 14, 15, 17 and others of
interest, including SNVs previously associated with neurodegeneration in other diseases (see
supplementary materials). Samples were genotyped in two tranches. Each tranche containing DNA
from Australia, the Czech Republic and Spain. Genome-wide data was therefore organised into six
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

17
data sets. Following rigorous per-data-set quality control (supplementary methods, Table S1), we
imputed all samples using the Haplotype Reference Consortium panel (r1.1)74 on the Michigan
imputation server, resulting in 22,469,259 SNVs. After further per-SNP QC, this resulted in 5,985,626
(final) SNVs with a minor allele frequency of at least 5%.
Association testing:
PLINK (v 1.9 and v 2.0)75 were used to conduct association testing and meta-analyses. For continuous
traits, we used linear regression to analyse each of the 6 data sets, adjusted for the first 5 principal
components (PCs), weighted genetic risk scores76, disease-modifying therapy use, together with
variables identified to have a standardised difference greater than 15% between severity extremes.
Combined data set results of all 6 groups were then analysed using fixed-effects meta-analyses to
identify statistically independent SNVs.
For binary traits, all 6 groups were analysed jointly, due to lower sample numbers. We used logistic
regression, adjusting for group ID, the first 5 PCs together with covariates with Cohen’s d >0.15.
For replication analyses, the Bonferroni-deflated p-value to meet replication threshold was set to p £
4.31x10-4 (0.05/116 replication SNVs). The de novo genome-wide association study (GWAS) p-value
threshold was set to p<5x10-8. SNVs meeting 1x10-8< p <5x10-5 were considered to have nominal
evidence of association with the trait of interest. Weighted genetic risk scores (wGRS)76 were
calculated based on directly genotyped SNVs described by the IMSGC (supplementary material).3
Calculations estimated a sample size of 915 individuals per group was required to achieve 80% power
to detect an SNV with MAF 0.2 and an odds ratio of 1.3, based on binary severity outcomes.
Survival analyses:
We assessed time to reach the hard disability milestones of irreversible EDSS 3 and 6 for those
individuals who had not yet reached these at first MSBase-recorded clinic entry. Where disability
milestones were not met during study observation, data were censored at the most recent clinic visit.
Survival analyses were based on carriage of the minor allele for SNVs at the top 10 nearest genes
identified in our l-ARMSS and l-MSSS de novo association analyses; and top 5 nearest genes identified
in our sex-stratified l-ARMSS and l-MSSS analyses. Cox proportional hazards regression (implemented
in Stata v17) was used to calculate hazard ratios (HRs) with 95% confidence intervals (CI). Multivariable
models were adjusted for AAO, MS susceptibility wGRS, percentage of follow-up exposed to DMT and
sex. The Schoenfeld residuals global test was used to detect a violation of the Cox proportional hazards
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

18
assumption. Where the proportionality assumption was violated, a Weibull regression approach was
applied. Survival data were visualised using Kaplan-Meier curves.
Heritability analysis:
Narrow sense heritability (hsnp) was estimated from individual-level GWAS data using a genome-based
restricted maximum likelihood (GREML) approach implemented using GCTA software,77 and BOLT-
LMM.78 BOLT-LMM was additionally used to estimate per-chromosome heritability estimates. We
further employed a summary statistic approach using a linkage disequilibrium (LD) score regression
(LDSC) implemented in LDSC software.79
Enrichment analyses:
We used FUMA22 to assign SNVs (p <1x10-5) to Genes, Tissues and functions using as per online
instructions. The Panther database80 was used to further confirm gene pathways/ontologies that were
over-represented and enriched in the variants with top association hits (p <1x10-5). Tissue expression
of variants with p<1x10-05 was further validated using both the TissueEnrich package81 and
Geneanalytics database82.
Machine Learning:
We chose to implement non-linear machine learning (ML) models for severity classification as linear
ML models, that do not account for interaction between genetic variants, have been found to perform
no better than simple linear regression in the context of common variant-based disorders.83 All SNVs
that had a p-value of 0.01 or less in the de novo meta-analyses were used to generate datasets
compatible with gradient boosting algorithms (xgboost21). A total of 62,351 SNVs were included, with
binary l-ARMSS score severity as the outcome. A training set of 70% of the pwMS was randomly
selected, ensuring a balanced representation of severe and mild MS outcomes. After training with
internal bootstrapping of 0.7 for 10k iterations, the model was tested on the 30% of the remaining
cohort, i.e. those datasets never encountered by the algorithm. We were cautious to avoid overfitting
our models by using 70%-30% cut off for the train/test data sets; a more conservative approach than
others that tend to use a 80%-20% cut off.84 Further, a slow learning rate (eta = 0.01) was
implemented to avoid overfitting.85, 86 The algorithm calculated a prediction score for each new
individual regarding their severity group membership. Accuracy of prediction was compared to the
clinically-informed grouping of each individual.
Using the prediction values generated on the test set for each model, as well as the true membership
values of each sample, a confusion matrix was generated along with accuracy, sensitivity, specificity,
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

19
and Kappa statistics using the confusion matrix function of Caret package. Furthermore, to evaluate
the prediction accuracy and performance of the models, the Receiver Operator Characteristic (ROC)
curve was plotted to explore the relationship between false positives and negatives, and the Area
Under the Curve (AUC) for each model was calculated.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

20
Acknowledgements
We thank all the people with MS who participated in this research without whom this work would not
be possible. We would also like to acknowledge the patients and the Biobank Nodo Hospital Virgen
Macarena (Biobanco del Sistema Sanitario Público de Andalucía) integrated in the Spanish National
biobanks Network (PT20/00069) supported by ISCIII and FEDER funds, for their collaboration in this
work.
The authors would like to acknowledge Prof David Booth, research nurses Ms Jo Baker, Ms Jodi
Haartsen, Ms Sandra Williams, Ms Lisa Taylor for assisting with sample collection for this study, and
Ms Malgorzata Krupa for research assistance. Further, the authors acknowledge the Center for
Genome Technology within the University of Miami John P. Hussman Institute for Human Genomics
for generating all the MegaEx array genotype data for this project and specifically Anna Konidari for
overseeing the genotyping efforts and assistance with the Illumina custom design process.
This work was supported by a Research Fellowship awarded to Dr Vilija Jokubaitis from Multiple
Sclerosis Research Australia (16-0206), and research grant support from the Royal Melbourne Hospital
Home Lottery Grant (MH2013-055), Charity Works for MS (2012 Project grant), MSBase Foundation
Project Grant, and Monash University. EH, DH, PK are supported by the Czech Ministry of Education,
project PROGRES Q27/LF1. FM and AA receive support from the Agencia Española de Investigación
(AEI)-Fondos Europeos de Desarrollo Regional (FEDER) (PID2019-110487R-C21) and Junta de
Andalucía (P18-RT-2623).
Author contributions
VGJ conceived and designed the study, obtained data, performed data analysis, interpreted the
data, and drafted the manuscript.
OI contributed to study design, performed data analysis, interpreted the data, and substantively
revised the manuscript
JS contributed to study design, performed data analysis, interpreted the data, and substantively
revised the manuscript
PK obtained study data, interpreted the data, and substantively revised the manuscript
FM contributed to study design, obtained study data, and substantively revised the manuscript
DH performed data analysis, and substantively revised the manuscript
SE obtained study data, interpreted the data, and substantively revised the manuscript
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

21
JLS obtained study data, interpreted the data, and substantively revised the manuscript
MS obtained study data, interpreted the data, and substantively revised the manuscript
RL interpreted the data, and substantively revised the manuscript
TJK obtained study data, interpreted the data, and substantively revised the manuscript
TK obtained study data, and substantively revised the manuscript
PDJ obtained study data, and substantively revised the manuscript
AB contributed to data analysis, and substantively revised the manuscript
JLM contributed to data analysis, and substantively revised the manuscript
BVT obtained study data, interpreted the data, and substantively revised the manuscript
SV obtained study data, and substantively revised the manuscript
LL contributed to study analysis, and substantively revised the manuscript
KV obtained study data, and revised the manuscript
MIGS obtained study data, and revised the manuscript
AA obtained study data, and revised the manuscript
AvdW obtained study data, and substantively revised the manuscript
EKH obtained study data, and substantively revised the manuscript
GI obtained study data, interpreted the data, and substantively revised the manuscript
NP contributed to study design, performed data analysis, interpreted the data, and substantively
revised the manuscript
DH obtained study data, interpreted the data, and substantively revised the manuscript
HB conceived and designed the study, obtained data, interpreted the data, and drafted the
manuscript.
Competing interests
The authors report no competing interests
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

22
References:
1. Walton C, King R, Rechtman L, et al. Rising prevalence of multiple sclerosis
worldwide: Insights from the Atlas of MS, third edition. Mult Scler. 2020 Nov
11:1352458520970841.
2. Lizak N, Lugaresi A, Alroughani R, et al. Highly active immunomodulatory therapy
ameliorates accumulation of disability in moderately advanced and advanced multiple
sclerosis. J Neurol Neurosurg Psychiatry. 2017 Mar;88(3):196-203.
3. International Multiple Sclerosis Genetics C. Multiple sclerosis genomic map
implicates peripheral immune cells and microglia in susceptibility. Science. 2019 Sep
27;365(6460).
4. Van der Walt A, Stankovich J, Bahlo M, et al. Heterogeneity at the HLA-DRB1 allelic
variation locus does not influence multiple sclerosis disease severity, brain atrophy or
cognition. Multiple sclerosis. 2011 Mar;17(3):344-52.
5. Barcellos LF, Oksenberg JR, Begovich AB, et al. HLA-DR2 dose effect on susceptibility
to multiple sclerosis and influence on disease course. American journal of human genetics.
2003 Mar;72(3):710-6.
6. Masterman T, Ligers A, Olsson T, Andersson M, Olerup O, Hillert J. HLA-DR15 is
associated with lower age at onset in multiple sclerosis. Annals of neurology. 2000
Aug;48(2):211-9.
7. George MF, Briggs FB, Shao X, et al. Multiple sclerosis risk loci and disease severity in
7,125 individuals from 10 studies. Neurol Genet. 2016 Aug;2(4):e87.
8. Jensen CJ, Stankovich J, Van der Walt A, et al. Multiple sclerosis susceptibility-
associated SNPs do not influence disease severity measures in a cohort of Australian MS
patients. PLoS One. 2010 Apr 2;5(4):e10003.
9. IMSGC. Genome-wide association study of severity in multiple sclerosis. Genes and
immunity. 2011 Dec;12(8):615-25.
10. Baranzini SE, Wang J, Gibson RA, et al. Genome-wide association analysis of
susceptibility and clinical phenotype in multiple sclerosis. Human molecular genetics. 2009
Feb 15;18(4):767-78.
11. Moutsianas L, Jostins L, Beecham AH, et al. Class II HLA interactions modulate
genetic risk for multiple sclerosis. Nature genetics. 2015 Oct;47(10):1107-13.
12. Jokubaitis VG, Butzkueven H. A genetic basis for multiple sclerosis severity: Red
herring or real? Mol Cell Probes. 2016 Dec;30(6):357-65.
13. Roxburgh RH, Seaman SR, Masterman T, et al. Multiple Sclerosis Severity Score:
using disability and disease duration to rate disease severity. Neurology. 2005 Apr
12;64(7):1144-51.
14. DeLuca GC, Ramagopalan SV, Herrera BM, et al. An extremes of outcome strategy
provides evidence that multiple sclerosis severity is determined by alleles at the HLA-DRB1
locus. Proceedings of the National Academy of Sciences of the United States of America.
2007 Dec 26;104(52):20896-901.
15. Brynedal B, Wojcik J, Esposito F, et al. MGAT5 alters the severity of multiple
sclerosis. Journal of neuroimmunology. 2010 Mar 30;220(1-2):120-4.
16. Jackson KC, Sun K, Barbour C, et al. Genetic model of MS severity predicts future
accumulation of disability. Ann Hum Genet. 2020 Jan;84(1):1-10.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

23
17. International Multiple Sclerosis Genetics C, Wellcome Trust Case Control C, Sawcer S,
et al. Genetic risk and a primary role for cell-mediated immune mechanisms in multiple
sclerosis. Nature. 2011 Aug 10;476(7359):214-9.
18. Kalincik T, Vivek V, Jokubaitis V, et al. Sex as a determinant of relapse incidence and
progressive course of multiple sclerosis. Brain. 2013 Dec;136(Pt 12):3609-17.
19. Manouchehrinia A, Westerlind H, Kingwell E, et al. Age Related Multiple Sclerosis
Severity Score: Disability ranked by age. Mult Scler. 2017 Dec;23(14):1938-46.
20. Nash A, Aumann TD, Pigoni M, et al. Lack of Sez6 Family Proteins Impairs Motor
Functions, Short-Term Memory, and Cognitive Flexibility and Alters Dendritic Spine
Properties. Cereb Cortex. 2020 Apr 14;30(4):2167-84.
21. Chen T, Guestrin C, editors. Xgboost: A scalable tree boosting system. Proceedings of
the 22nd acm sigkdd international conference on knowledge discovery and data mining;
2016.
22. Watanabe K, Taskesen E, van Bochoven A, Posthuma D. Functional mapping and
annotation of genetic associations with FUMA. Nat Commun. 2017 Nov 28;8(1):1826.
23. Zhou Y, Graves JS, Simpson S, Jr., et al. Genetic variation in the gene LRP2 increases
relapse risk in multiple sclerosis. J Neurol Neurosurg Psychiatry. 2017 Oct;88(10):864-8.
24. Butzkueven H, Chapman J, Cristiano E, et al. MSBase: an international, online registry
and platform for collaborative outcomes research in multiple sclerosis. Multiple sclerosis.
2006 Dec;12(6):769-74.
25. Fuh-Ngwa V, Zhou Y, Charlesworth JC, et al. Developing a clinical-environmental-
genotypic prognostic index for relapsing-onset multiple sclerosis and clinically isolated
syndrome. Brain Commun. 2021;3(4):fcab288.
26. Jokubaitis VG, Spelman T, Kalincik T, et al. Predictors of disability worsening in
clinically isolated syndrome. Ann Clin Transl Neurol. 2015 May;2(5):479-91.
27. Jokubaitis VG, Spelman T, Kalincik T, et al. Predictors of long-term disability accrual in
relapse-onset multiple sclerosis. Ann Neurol. 2016 Jul;80(1):89-100.
28. Iaffaldano P, Lucisano G, Butzkueven H, et al. Early treatment delays long-term
disability accrual in RRMS: Results from the BMSD network. Mult Scler. 2021 Apr
26:13524585211010128.
29. Beiki O, Frumento P, Bottai M, Manouchehrinia A, Hillert J. Changes in the Risk of
Reaching Multiple Sclerosis Disability Milestones In Recent Decades: A Nationwide
Population-Based Cohort Study in Sweden. JAMA Neurol. 2019 Jun 1;76(6):665-71.
30. Hua LH, Hersh CM, Tian F, Mowry EM, Fitzgerald KC. Clinical characteristics of a large
multi-center cohort of people with multiple sclerosis over age 60. Mult Scler Relat Disord.
2021 Jan;47:102637.
31. Simonsen CS, Flemmen HO, Broch L, et al. The course of multiple sclerosis rewritten:
a Norwegian population-based study on disease demographics and progression. J Neurol.
2021 Apr;268(4):1330-41.
32. University of California SFMSET, Cree BA, Gourraud PA, et al. Long-term evolution of
multiple sclerosis disability in the treatment era. Ann Neurol. 2016 Oct;80(4):499-510.
33. Tintore M, Cobo-Calvo A, Carbonell P, et al. Effect of Changes in MS Diagnostic
Criteria Over 25 Years on Time to Treatment and Prognosis in Patients With Clinically
Isolated Syndrome. Neurology. 2021 Sep 14.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

24
34. Zhong M, van der Walt A, Campagna MP, Stankovich J, Butzkueven H, Jokubaitis V.
The Pharmacogenetics of Rituximab: Potential Implications for Anti-CD20 Therapies in
Multiple Sclerosis. Neurotherapeutics. 2020 Oct;17(4):1768-84.
35. Ramanujam R, Hedstrom AK, Manouchehrinia A, et al. Effect of Smoking Cessation
on Multiple Sclerosis Prognosis. JAMA Neurol. 2015 Oct;72(10):1117-23.
36. Nicholls HL, John CR, Watson DS, Munroe PB, Barnes MR, Cabrera CP. Reaching the
End-Game for GWAS: Machine Learning Approaches for the Prioritization of Complex
Disease Loci. Frontiers in Genetics. 2020;11:350.
37. Pinto MF, Oliveira H, Batista S, et al. Prediction of disease progression and outcomes
in multiple sclerosis with machine learning. Scientific Reports. 2020
2020/12/03;10(1):21038.
38. Yuan L, Zhai L, Qian L, et al. Switching off IMMP2L signaling drives senescence via
simultaneous metabolic alteration and blockage of cell death. Cell Res. 2018 Jun;28(6):625-
43.
39. He Q, Gu L, Lin Q, et al. The Immp2l Mutation Causes Ovarian Aging Through ROS-
Wnt/beta-Catenin-Estrogen Pathway: Preventive Effect of Melatonin. Endocrinology. 2020
Sep 1;161(9).
40. Liu C, Li X, Lu B. The Immp2l mutation causes age-dependent degeneration of
cerebellar granule neurons prevented by antioxidant treatment. Aging Cell. 2016
Feb;15(1):167-76.
41. Leblond CS, Cliquet F, Carton C, et al. Both rare and common genetic variants
contribute to autism in the Faroe Islands. NPJ Genom Med. 2019;4:1.
42. Lucas SEM, Zhou T, Blackburn NB, et al. Rare, potentially pathogenic variants in 21
keratoconus candidate genes are not enriched in cases in a large Australian cohort of
European descent. PLoS One. 2018;13(6):e0199178.
43. Vinas-Jornet M, Esteba-Castillo S, Baena N, et al. High Incidence of Copy Number
Variants in Adults with Intellectual Disability and Co-morbid Psychiatric Disorders. Behav
Genet. 2018 Jul;48(4):323-36.
44. Buhring J, Hecker M, Fitzner B, Zettl UK. Systematic Review of Studies on Telomere
Length in Patients with Multiple Sclerosis. Aging Dis. 2021 Aug;12(5):1272-86.
45. Nicaise AM, Wagstaff LJ, Willis CM, et al. Cellular senescence in progenitor cells
contributes to diminished remyelination potential in progressive multiple sclerosis. Proc
Natl Acad Sci U S A. 2019 Apr 30;116(18):9030-9.
46. Si Z, Sun L, Wang X. Evidence and perspectives of cell senescence in
neurodegenerative diseases. Biomed Pharmacother. 2021 May;137:111327.
47. El-Hattab AW, Scaglia F. Mitochondrial DNA depletion syndromes: review and
updates of genetic basis, manifestations, and therapeutic options. Neurotherapeutics. 2013
Apr;10(2):186-98.
48. Duncan GJ, Simkins TJ, Emery B. Neuron-Oligodendrocyte Interactions in the
Structure and Integrity of Axons. Front Cell Dev Biol. 2021;9:653101.
49. Witte ME, Mahad DJ, Lassmann H, van Horssen J. Mitochondrial dysfunction
contributes to neurodegeneration in multiple sclerosis. Trends Mol Med. 2014
Mar;20(3):179-87.
50. Sarapulov AV, Petrov P, Hernandez-Perez S, et al. Missing-in-Metastasis/Metastasis
Suppressor 1 Regulates B Cell Receptor Signaling, B Cell Metabolic Potential, and T Cell-
Independent Immune Responses. Front Immunol. 2020;11:599.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

25
51. Li Z, Zhang L, Liu Z, et al. miRNA-182 regulated MTSS1 inhibits proliferation and
invasion in Glioma Cells. J Cancer. 2020;11(19):5840-51.
52. Brown AS, Meera P, Altindag B, et al. MTSS1/Src family kinase dysregulation
underlies multiple inherited ataxias. Proc Natl Acad Sci U S A. 2018 Dec 26;115(52):E12407-
E16.
53. Minkeviciene R, Hlushchenko I, Virenque A, et al. MIM-Deficient Mice Exhibit
Anatomical Changes in Dendritic Spines, Cortex Volume and Brain Ventricles, and Functional
Changes in Motor Coordination and Learning. Front Mol Neurosci. 2019;12:276.
54. Kawabata Galbraith K, Fujishima K, Mizuno H, et al. MTSS1 Regulation of Actin-
Nucleating Formin DAAM1 in Dendritic Filopodia Determines Final Dendritic Configuration
of Purkinje Cells. Cell Rep. 2018 Jul 3;24(1):95-106 e9.
55. Causevic M, Dominko K, Malnar M, et al. BACE1-cleavage of Sez6 and Sez6L is
elevated in Niemann-Pick type C disease mouse brains. PLoS One. 2018;13(7):e0200344.
56. Pigoni M, Wanngren J, Kuhn PH, et al. Seizure protein 6 and its homolog seizure 6-
like protein are physiological substrates of BACE1 in neurons. Mol Neurodegener. 2016 Oct
5;11(1):67.
57. Brownstein CA, Smith RS, Rodan LH, et al. RCL1 copy number variants are associated
with a range of neuropsychiatric phenotypes. 2021 May;26(5):1706-18.
58. Amin N, de Vrij FMS, Baghdadi M, et al. A rare missense variant in RCL1 segregates
with depression in extended families. Mol Psychiatry. 2018 May;23(5):1120-6.
59. Mandolesi G, Gentile A, Musella A, et al. Synaptopathy connects inflammation and
neurodegeneration in multiple sclerosis. Nat Rev Neurol. 2015 Dec;11(12):711-24.
60. Di Filippo M, Portaccio E, Mancini A, Calabresi P. Multiple sclerosis and cognition:
synaptic failure and network dysfunction. Nat Rev Neurosci. 2018 Oct;19(10):599-609.
61. Dutta R, Chang A, Doud MK, et al. Demyelination causes synaptic alterations in
hippocampi from multiple sclerosis patients. Ann Neurol. 2011 Mar;69(3):445-54.
62. Vercellino M, Marasciulo S, Grifoni S, et al. Acute and chronic synaptic pathology in
multiple sclerosis gray matter. Mult Scler. 2021 Jun 14:13524585211022174.
63. Dias C, Pfundt R, Kleefstra T, et al. De novo variants in TCF7L2 are associated with a
syndromic neurodevelopmental disorder. Am J Med Genet A. 2021 Aug;185(8):2384-90.
64. Sock E, Wegner M. Using the lineage determinants Olig2 and Sox10 to explore
transcriptional regulation of oligodendrocyte development. Dev Neurobiol. 2021 Sep 3.
65. Kipp M, Amor S, Krauth R, Beyer C. Multiple sclerosis: neuroprotective alliance of
estrogen-progesterone and gender. Front Neuroendocrinol. 2012 Jan;33(1):1-16.
66. Clark SV, Semmel ES, Aleksonis HA, Steinberg SN, King TZ. Cerebellar-Subcortical-
Cortical Systems as Modulators of Cognitive Functions. Neuropsychol Rev. 2021
Sep;31(3):422-46.
67. Le M, Malpas C, Sharmin S, et al. Disability outcomes of early cerebellar and
brainstem symptoms in multiple sclerosis. Mult Scler. 2021 Apr;27(5):755-66.
68. Tintore M, Arrambide G, Otero-Romero S, et al. The long-term outcomes of CIS
patients in the Barcelona inception cohort: Looking back to recognize aggressive MS. Mult
Scler. 2020 Nov;26(13):1658-69.
69. Polman CH, Reingold SC, Banwell B, et al. Diagnostic criteria for multiple sclerosis:
2010 revisions to the McDonald criteria. Annals of neurology. 2011 Feb;69(2):292-302.
70. Polman CH, Reingold SC, Edan G, et al. Diagnostic criteria for multiple sclerosis: 2005
revisions to the "McDonald Criteria". Annals of neurology. 2005 Dec;58(6):840-6.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

26
71. Thompson AJ, Banwell BL, Barkhof F, et al. Diagnosis of multiple sclerosis: 2017
revisions of the McDonald criteria. Lancet Neurol. 2018 Feb;17(2):162-73.
72. Hughes S, Spelman T, Trojano M, et al. The Kurtzke EDSS rank stability increases 4
years after the onset of multiple sclerosis: results from the MSBase Registry. Journal of
neurology, neurosurgery, and psychiatry. 2012 Mar;83(3):305-10.
73. Goscinski WJ, McIntosh P, Felzmann U, et al. The multi-modal Australian ScienceS
Imaging and Visualization Environment (MASSIVE) high performance computing
infrastructure: applications in neuroscience and neuroinformatics research. Frontiers in
Neuroinformatics. 2014 2014-March-27;8(30).
74. Das S, Forer L, Schonherr S, et al. Next-generation genotype imputation service and
methods. Nat Genet. 2016 Oct;48(10):1284-7.
75. Purcell S, Neale B, Todd-Brown K, et al. PLINK: a tool set for whole-genome
association and population-based linkage analyses. Am J Hum Genet. 2007 Sep;81(3):559-
75.
76. De Jager PL, Chibnik LB, Cui J, et al. Integration of genetic risk factors into a clinical
algorithm for multiple sclerosis susceptibility: a weighted genetic risk score. Lancet Neurol.
2009 Dec;8(12):1111-9.
77. Yang J, Lee SH, Goddard ME, Visscher PM. GCTA: a tool for genome-wide complex
trait analysis. Am J Hum Genet. 2011;88(1):76-82.
78. Loh P-R, Tucker G, Bulik-Sullivan BK, et al. Efficient Bayesian mixed-model analysis
increases association power in large cohorts. Nature genetics. 2015;47(3):284-90.
79. Bulik-Sullivan BK, Loh P-R, Finucane HK, et al. LD Score regression distinguishes
confounding from polygenicity in genome-wide association studies. Nature Genetics. 2015
2015/03/01;47(3):291-5.
80. Mi H, Thomas P. PANTHER pathway: an ontology-based pathway database coupled
with data analysis tools. Methods Mol Biol. 2009;563:123-40.
81. Jain A, Tuteja G. TissueEnrich: Tissue-specific gene enrichment analysis.
Bioinformatics. 2018;35(11):1966-7.
82. Ben-Ari Fuchs S, Lieder I, Stelzer G, et al. GeneAnalytics: An Integrative Gene Set
Analysis Tool for Next Generation Sequencing, RNAseq and Microarray Data. OMICS.
2016;20(3):139-51.
83. Romagnoni A, Jégou S, Van Steen K, et al. Comparative performances of machine
learning methods for classifying Crohn Disease patients using genome-wide genotyping
data. Scientific Reports. 2019 2019/07/17;9(1):10351.
84. Park JH, Shin SD, Song KJ, et al. Prediction of good neurological recovery after out-of-
hospital cardiac arrest: A machine learning analysis. Resuscitation. 2019 Sep;142:127-35.
85. Friedman JH. Greedy Function Approximation: A Gradient Boosting Machine. The
Annals of Statistics. 2001;29(5):1189-232.
86. Jerome F, Trevor H, Robert T. Additive logistic regression: a statistical view of
boosting (Withdiscussion and a rejoinder by the authors). The Annals of Statistics. 2000
4/1;28(2):337-407.
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

27
Figures
Figure 1a: EDSS-age trajectories for included participants classified into mild (brown; n=1,180),
intermediate (green; n=3,559) or severe (purple; n=1,112) groups based on median longitudinal
ARMSS scores. 1b: EDSS-symptom duration trajectories for included participants classified into mild
(brown; n=1,232), intermediate (green; n=3,512), or severe (purple; n=1,107) groups based on median
longitudinal MSSS scores.
Figure 1a: EDSS-age trajectories for included participants classified into mild
(brown; n=1,180), intermediate (green; n=3,559) or severe (purple; n=1,112)
groups based on median longitudinal ARMSS scores. 1b: EDSS-symptom duration
trajectories for included participants classified into mild (brown; n=1,232),
intermediate (green; n=3,512), or severe (purple; n=1,107) groups based on
median longitudinal MSSS scores.
a
b
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

28
Figure 2: Machine learning algorithm classification of mild and severe cases. a: l-ARMSS +wGRS + AAO
ML (n=447 mild, n=464 severe); AUROC 0.87 (95% CI 0.83-0.91) b: l-MSSS +wGRS + AAO ML (n=585
mild, n=466 severe); AUROC 0.85 (95% CI 0.80-0.89) c: Feature importance for l-ARMSS +wGRS + AAO
ML model d: Feature importance for l-MSSS +wGRS + AAO ML model e: l-ARMSS +wGRS + AAO ML
confusion matrix (30% cohort) f: l-ARMSS +wGRS + AAO ML confusion matrix (30% cohort).
ab
c d
e f
Figure 2: Machine learning algorithm classification of mild and severe cases. a: l-ARMSS
+wGRS + AAO ML AUROC 0.87 b: l-MSSS +wGRS + AAO ML AUROC 0.85 c: Feature importance
for l-ARMSS +wGRS + AAO ML model d: Feature importance for l-MSSS +wGRS + AAO ML
model e: l-ARMSS +wGRS + AAO ML confusion matrix f: l-ARMSS +wGRS + AAO ML confusion
matrix
Reference
Prediction Mild Severe
Mild 121 15
Severe 20 115
Longitudinal ARMSS Longitudinal MSSS
Reference
Prediction Mild Severe
Mild 156 49
Severe 3105
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

29
Figure 3: Kaplan-Meier survival curves showing time to irreversible EDSS milestones based on the
presence (1,2) or absence (0) of the minor allele at each locus a: rs7289446 intronic to SEZ6L time to
irreversible EDSS 3 (aHR 0.77, p=0.008) b: rs7289446 intronic to SEZ6L time to irreversible EDSS 6 (aHR
0.72, p=4.85x10-4) c: rs295254 intronic to RCL1 time to irreversible EDSS 3 (aHR 1.33, p=9.10x10-4) d:
rs295254 intronic to RCL1 time to irreversible EDSS 6 (aHR 1.32, p=7.37x10-4) e: rs11399831 nearest
to SUCLA2 time to irreversible EDSS 3 (aHR 0.77, p=0.036) f: rs11399831 nearest to SUCLA2 time to
irreversible EDSS 6 (aHR 0.74, p=2.82x10-4).
ab
c d
e f
Figure 3: Kaplan-Meier survival curves showing time to irreversible EDSS milestones based on
the presence (1,2) or absence (0) of the minor allele at each locus a: rs7289446 intronic to
SEZ6L time to irreversible EDSS 3 (aHR 0.77, p=0.008) b: rs7289446 intronic to SEZ6L time to
irreversible EDSS 6 (aHR 0.72, p=4.85x10-4) c: rs295254 intronic to RCL1 time to irreversible
EDSS 3 (aHR 1.33, p=9.10x10-4) d: rs295254 intronic to RCL1 time to irreversible EDSS 6 (aHR
1.32, p=7.37x10-4) e: rs11399831 nearest to SUCLA2 time to irreversible EDSS 3 (aHR 0.77,
p=0.036) f: rs11399831 nearest to SUCLA2 time to irreversible EDSS 6 (aHR 0.74, p=2.82x10-4).
Irreversible EDSS 3 Irreversible EDSS 6
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

30
Figure 4: Kaplan-Meier survival curves showing time to irreversible EDSS milestones based on the
presence (1,2) or absence (0) of the minor allele at each locus a: rs9643199 intronic to MTSS1 time to
irreversible EDSS 3 in females (aHR 1.35, p=0.006) b: rs9643199 intronic to MTSS1 time to irreversible
EDSS 6 in females (aHR 1.46, p=7.09x10-4) c: rs2776741 nearest to RCAN3AS time to irreversible EDSS
3 in females (aHR 0.77, p=0.010) d: rs2776741 nearest to RCAN3AS time to irreversible EDSS 6 in
females (aHR 0.74, p=0.005) e: rs7070182 intronic to TCF7L2 time to irreversible EDSS 3 in males (aHR
0.59, p=0.013) f: rs7070182 intronic to TCF7L2 time to irreversible EDSS 6 in males (aHR 0.56, p=0.005).
ab
c d
e f
Figure 3: Kaplan-Meier survival curves showing time to irreversible EDSS milestones based on
the presence (1,2) or absence (0) of the minor allele at each locus a: rs9643199 intronic to
MTSS1 time to irreversible EDSS 3 in females (aHR 1.35, p=0.006) b: rs9643199 intronic to
MTSS1 time to irreversible EDSS 6 in females (aHR 1.46, p=7.09x10-4) c: rs2776741 nearestto
RCAN3AS time to irreversible EDSS 3 in females (aHR 0.77, p=0.010) d: rs2776741 nearest to
RCAN3AS time to irreversible EDSS 6 in females (aHR 0.74, p=0.005) e: rs7070182 intronic to
TCF7L2 time to irreversible EDSS 3 in males (aHR 0.59, p=0.013) f: rs7070182 intronic to
TCF7L2 time to irreversible EDSS 6 in males (aHR 0.56, p=0.005).
Irreversible EDSS 3 Irreversible EDSS 6
Females Only
Males Only
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

31
Table 1: Cohort Characteristics by longitudinal ARMSS (l-ARMSS) categorisation
Assessed
Genotyped – Passed QC
Characteristics
All
n=5,851
Mild
n=1,180
Severe
n=1,167
All
n=1,813
Mild
n=447
Severe n=464
Cohen’s d
(mild vs
severe
genotyped)
Population
n(%)
Australian
Czech
Spanish
1993 (34.1)
2664 (45.5)
1194 (20.4)
625 (53.0)
346 (29.3)
209 (17.7)
375 (32.1)
497 (42.6)
295 (25.3)
676 (37.3)
716 (39.5)
421 (23.2)
209 (46.8)
156 (34.9)
82 (18.3)
161 (34.7)
164 (35.3)
139 (30.0)
l-ARMSS Score
Median (IQR)
4.53
(2.79, 6.55)
1.49
(0.93, 1.97)
8.24
(7.63, 8.91)
4.13
(2.43, 7.21)
1.49
(0.96, 2.03)
8.55
(7.92, 9.12)
10.40
Range
0.08 – 9.99
0.08 – 2.39
7.13 – 9.99
0.14 – 9.98
0.14 – 2.38
7.13 – 9.98
Disease course
n(%)
RRMS
5075 (86.7)
1150 (97.5)
743 (63.7)
1471 (81.1)
438 (98.0)
226 (48.7)
SPMS
776 (13.2)
30 (2.5)
424 (36.3)
342 (18.9)
9 (2.0)
238 (51.3)
Sex n(%)
Female
4,261 (72.8)
889 (75.3)
809 (69.3)
1,313 (72.4)
337 (75.4)
316 (68.1)
Male
1,590 (27.2)
291 (24.7)
358 (30.7)
500 (27.6)
110 (24.6)
148 (31.9)
Age at onset
(AAO)
Median (IQR)
29.2
(23.4, 36.5)
34.3
(27.7, 41.9)
26.0
(20.9, 32.1)
28.3
(22.7, 35.7)
32.3
(26.4, 39.6)
27.2
(22.1, 33.1)
0.58
Age at most
recent visit
Median (IQR)
47.2
(39.7, 56.4)
50.5
(42.8, 58.5)
47.4
(40.0, 56.1)
48.1
(40.9, 57.2)
51.2
(43.5, 58.6)
51.1
(43.6, 58.5)
0.003
Range
17.4 – 87.6
20.2 – 87.6
17.4 – 80.1
17.4 – 84.5
25.2 – 83.1
17.4 – 80.1
Median (IQR)
10.1
9.2
10.9
11.7
11.2
12.5
0.32
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

32
Follow up in
MSBase (years)
(7.5, 13.3)
(7.0, 12.1)
(8.3, 15.0)
(9.7, 15.2)
(9.4, 14.1)
(10.2, 16.2)
Range
5.0 – 54.8
5.0 – 32.2
5.0 – 47.0
5.1 – 32.2
5.1 – 32.2
5.2 – 29.1
Number of EDSS
assessed
Median (IQR)
18 (11, 31)
13 (9, 20)
18 (11, 31)
21 (14, 34)
17 (12, 26)
19 (12, 30)
0.10
Range
3 – 91
3 – 91
3 – 85
3 – 91
3 – 91
3 – 85
Symptom
duration (years)
Median (IQR)
16.0
(10.7, 23.1)
13.8
(9.7, 20.2)
20.6
(14.3, 26.9)
18.1
(13.2, 24.4)
16.2
(11.9, 22.0)
22.2
(17.5, 28.0)
0.61
Range
5.1 – 66.6
5.1 – 58.5
5.3 – 57.1
5.6 – 55.4
5.9 – 55.4
6.7 – 50.4
% time exposed
to DMT
Median (IQR)
82.9
(36.1, 97.6)
82.04
(14.6, 97.3)
67.4
(18.3, 93.3)
79.7
(37.6, 96.6)
79.9
(29.8, 97.8)
60.6
(17.3, 91.2)
0.22
Range
0-100
0-100
0-100
0-100
0-100
0-100
ARR
Median (IQR)
0.16 (0, 0.38)
0.10 (0, 0.19)
0.17 (0, 0.44)
0.17 (0.06,
0.36)
0.10 (0, 0.22)
0.17 (0.06,
0.39)
0.43
Range
0 – 2.45
0 – 1.28
0 – 2.45
0 – 1.62
0 – 1.01
0 – 1.54
wGRS
Mean (SD)
-
-
-
2.85 (0.79)
2.71 (0.80)
2.90 (0.78)
0.23
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

33
Table 2: Results of fixed effects meta-analyses: top SNVs for the 10 nearest genes for l-ARMSS and l-MSSS phenotypes & top SNVs for the 5 nearest genes
in sex-stratified analyses.
rsID
Chr
ranked order
SNV
ranked order
nearest gene
Nearest gene
Minor Allele
MAF
adj p# ^
β
l-ARMSS#, n=1813
rs7289446
22
1
1
SEZ6L
G
0.27
2.73E-07
-0.488
rs1207401
22
2
1
SEZ6L
A
0.27
2.90E-07
-0.490
rs7758683
6
6
2
EPHA7
T
0.23
1.88E-06
-0.477
rs56089601
4
7
3
STK32B
C
0.10
2.69E-06
0.655
rs12953974
18
10
4
CTIF
A
0.12
3.64E-06
-0.607
rs56194930
9
11
5
ACO1
G
0.11
3.69E-06
0.648
rs73091975
7
14
6
CCDC129
G
0.09
3.84E-06
-0.680
rs2274333
1
16
7
CA6
G
0.29
4.60E-06
-0.429
rs295254
9
20
8
RCL1
G
0.38
5.64E-06
0.388
rs11057374
12
21
9
DNAH10
G
0.35
5.79E-06
-0.390
rs111399831
13
29
10
SUCLA2
A
0.21
7.34E-06
-0.468
l-MSSS^, n=1813
rs1207401
22
1
1
SEZ6L
A
0.27
2.88E-07
-0.484
rs7289446
22
4
1
SEZ6L
G
0.27
3.35E-07
-0.479
rs9643199
8
6
2
MTSS1
A
0.26
1.90E-06
0.465
rs2725556
15
22
3
UNC13C
A
0.08
2.39E-06
0.709
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

34
rs61578937
8
23
4
NCOA2
G
0.13
2.86E-06
-0.562
rs7758683
6
27
5
EPHA7
T
0.23
3.61E-06
-0.456
rs56089601
4
32
6
STK32B
C
0.10
4.65E-06
0.631
rs4495680
1
33
7
RGS13
G
0.41
5.65E-06
-0.390
rs56363129
10
42
8
SEPHS1
G
0.15
6.56E-06
0.516
rs111399831
13
47
9
SUCLA2
A
0.15
7.07E-06
-0.534
l-ARMSS Females only, n=1313
rs1219732
10
1
1
CPXM2
T
0.35
6.48E-08
0.569
rs10967273
9
3
2
LOC100506422
T
0.13
1.16E-07
0.799
rs4572384
16
15
3
CRISPLD2
A
0.40
2.77E-06
0.487
rs2776741
1
16
4
RCAN3AS
A
0.31
2.98E-06
-0.516
rs61873874
10
18
5
MKI67
A
0.05
3.39E-06
1.007
l-ARMSS Males only, n=500
rs7315384
12
1
1
STAB2
C
0.24
1.29E-07
1.041
rs7070182
10
2
2
TCF7L2
C
0.18
3.65E-07
-1.112
rs11845242
14
4
3
LINC00520
G
0.41
5.63E-07
-0.800
rs3885012
12
5
4
PHLDA1
G
0.06
1.19E-06
-1.499
rs11665069
18
7
5
FHOD3
C
0.41
1.34E-06
0.755
l-MSSS Females only, n=1313
rs10967273
9
1
1
LOC100506422
T
0.13
3.52E-08
0.830
rs9643199
8
2
2
MTSS1
A
0.26
6.54E-08
0.631
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint

35
rs17169210
5
4
3
SLC25A48
C
0.23
1.25E-07
-0.621
rs1219732
10
7
4
CPXM2
T
0.35
1.70E-07
0.550
rs61873874
10
50
5
MKI67
A
0.05
1.45E-06
1.051
l-MSSS Males only, n=500
rs698805
2
1
1
CAMKMT
G
0.07
4.35E-08
-1.540
rs7315384
12
2
2
STAB2
C
0.24
8.00E-08
1.020
rs3885012
12
14
4%
PHLDA1
G
0.06
3.69E-07
-1.312
rs7070182
10
35
5
TCF7L2
C
0.18
3.86E-07
-1.054
rs28442172
18
51
6
FHOD3
G
0.13
2.13E-06
-1.105
Fixed-effects meta-analyses (n=6) adjusted for the first 5 principal components (PCs)
# l-ARMSS analyses adjusted for: % time on therapy since disease onset, weighted genetic risk score (wGRS), number of DRB1*1501 alleles, and imbalanced
variables: follow-up time in MSBase (years), symptom duration (years), Annualised Relapse Rate (ARR)
^ l-MSSS analyses additionally adjusted for % time on therapy since disease onset, wGRS, number of DRB1*1501 alleles, and imbalanced variables including:
age at most recent visit, number of EDSS assessments, symptom duration (years)
% 3rd closest gene hit had >80% heterogeneity
. CC-BY-NC-ND 4.0 International licenseIt is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint The copyright holder for thisthis version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270362doi: medRxiv preprint