Metabolome-based signature of disease pathology in MS

Abstract

Background

Diagnostic delays are common for multiple sclerosis (MS) since diagnosis typically depends on the presentation of nonspecific clinical symptoms together with radiologically-determined central nervous system (CNS) lesions. It is important to reduce diagnostic delays as earlier initiation of disease modifying therapies mitigates long-term disability. Developing a metabolomic blood-based MS biomarker is attractive, but prior efforts have largely focused on specific subsets of metabolite classes or analytical platforms. Thus, there are opportunities to interrogate metabolite profiles using more expansive and comprehensive approaches for developing MS biomarkers and for advancing our understanding of MS pathogenesis.

Methods

To identify putative blood-based MS biomarkers, we comprehensively interrogated the metabolite profiles in 12 non-Hispanic white, non-smoking, male MS cases who were drug naïve for 3 months prior to biospecimen collection and 13 non-Hispanic white, non-smoking male controls who were frequency matched to cases by age and BMI. We performed untargeted two-dimensional gas chromatography and time-of-flight mass spectrometry (GCxGC-TOFMS) and targeted lipidomic and amino acid analysis on serum. 325 metabolites met quality control and supervised machine learning was used to identify metabolites most informative for MS status. The discrimination potential of these select metabolites were assessed using receiver operator characteristic curves based on logistic models; top candidate metabolites were defined as having area under the curves (AUC) >80%. The associations between whole-genome expression data and the top candidate metabolites were examined, followed by pathway enrichment analyses. Similar associations were examined for 175 putative MS risk variants and the top candidate metabolites.

Results

12 metabolites were determined to be informative for MS status, of which 6 had AUCs >80%: pyroglutamate, laurate, acylcarnitine C14:1, N-methylmaleimide, and 2 phosphatidylcholines (PC ae 40:5, PC ae 42:5). These metabolites participate in glutathione metabolism, fatty acid metabolism/oxidation, cellular membrane composition, and transient receptor potential channel signaling. Pathway analyses based on the gene expression association for each metabolite suggested enrichment for pathways associated with apoptosis and mitochondrial dysfunction. Interestingly, the predominant MS genetic risk allele HLA-DRB1*15:01 was associated with one of the 6 top metabolites.

Conclusion

Our analysis represents the most comprehensive description of metabolic changes associated with MS to date with the inclusion of genomic and genetic information. We identified atypical metabolic processes that differed between MS patients and controls, which may lead to the development of biological targets for diagnosis and treatment.

Introduction

Multiple sclerosis (MS) is an idiopathic autoimmune disease typified by inflammatory events in the central nervous system that trigger demyelination, exacerbating remyelination failure, and subsequent neurodegeneration. Most affected individuals will steadily accrue irreversible neurological disability over the disease course, but there is mounting evidence that earlier initiation of disease modifying therapies (DMTs) may mitigate long-term disability outcomes.¹ Unfortunately, MS currently requires a clinical diagnosis that largely depends on neurological symptoms and/or radiologic lesions that are disseminating in time and/or space. As a result, the initiation of DMTs may be delayed. Therefore, there is a need to develop non-invasive diagnostic tools to accelerate MS diagnosis. There is also significant need to further elucidate the patho-etiological mechanisms contributing to MS onset and accrual of disability in order to advance the development of novel therapeutics.

Multiple efforts to identify biological markers (biomarkers) for MS have been made, but most of these potential biomarkers are not MS-specific and many are only discriminatory in cerebrospinal fluid (CSF).^{2, 3} While CSF is among the most pertinent biological samples for the study of MS, the invasiveness and inherent risk of lumbar punctures makes CSF collection impractical for recurrent testing.⁴ Instead, biomarkers readily detected in the blood would be ideal for diagnostic and prognostic evaluation. While there are promising candidates for diagnosis, such as neurofilament light protein, the only clinically useful blood biomarkers for MS are antibodies against natalizumab and interferon β drugs; these inform drug efficacy and risk for adverse side-effects.^5-7

Increasingly, metabolomic assays are being used to develop disease biomarkers and to identify patho-etiologic processes. Improvements in instrumentation, analytical sensitivity, and consistency have advanced metabolomics-based biomarker development for diseases such as Type II diabetes and cardiovascular disease.^{8, 9} Most prior metabolomics work in MS has relied on a single assay/platform or had a scope limited to a particular metabolite class.^10-18 Only one prior study applied a global untargeted metabolomic approach (ultra-high performance liquid chromatography-tandem mass spectrometry) to MS, which allowed for interrogating a much larger number of metabolites.¹⁹ Therefore, by combining targeted and untargeted approaches there are opportunities to conduct more expansive and comprehensive metabolomic investigation in MS, particularly within the multiomic framework.

We profiled serum metabolites in MS cases and controls using untargeted two-dimensional gas chromatography and time-of-flight mass spectrometry (GCxGC-TOFMS) with targeted metabolomic approaches (i.e. lipid and amino acid profiles). GCxGC-TOFMS has been shown to discriminate and identify up to three times more serum metabolites than one-dimensional GC-TOFMS.^{20, 21} We then identified metabolites predictive of MS status using supervised machine learning and multivariable logistic regression. To explore possible etiologic processes, we integrated whole-genome expression data and genetic data for alleles encoded by human leukocyte antigen (HLA) genes (HLA-DRB1*15:01, HLA-A*02) and 175 putative MS risk single nucleotide polymorphisms (SNPs) outside the major histocompatibility complex (MHC).²² We identified 6 metabolites predictive of MS status and found evidence that pathways involved in apoptosis and mitochondrial dysfunction may be contributing to the altered metabolic profile in MS.

Materials and Methods

Study subjects

The Duke University Measurement to Understand the Reclassification of Disease of Cabarrus/Kannapolis (MURDOCK) longitudinal health study recruitment has been previously described.²³ MS cases (n=12) were participants in both the MURDOCK Study Community Registry and Biorepository (Pro00011196) and MURDOCK MS Cohort Study (Pro00023791), while controls (n=13) were drawn from the MURDOCK Study Community Registry and Biorepository (Supplementary Table 1). Participant recruitment and sample collection protocols were approved by the Duke University Institutional Review Board; written informed consent was obtained from all participants. The study population for the current analysis was non-Hispanic white, non-smoking males. MS cases were drug naïve for at least three months prior to biospecimen collection, and medical records were used to confirm a diagnosis of MS and disease-modifying therapy (DMT) history. Controls were age and body mass index (BMI) frequency matched to cases.

Sample collection and storage

The MURDOCK Study collected samples for extraction of serum, RNA, and DNA using standard collection criteria implemented by the MURDOCK Study.²³ Serum samples were stored at −80 °C by LabCorp and/or BioStorage Technologies until use in this project. Total RNA was extracted from PAXgene tubes using Qiagen PAXgene Blood RNA Kit (#762164) and stored at −80 °C. Genomic DNA was extracted from whole blood using QIAGEN Gentra Puregene Blood Kits (#158445) and stored at −20 °C.

Sample preparation and data generation

Metabolomic data

For GCxGC-TOFMS, serum aliquots were extracted with methanol/chloroform containing heptadecanoic acid and norleucine as internal standards, as previously described.²¹ Equal sample aliquots were pooled to create for quality control (QC) samples. Samples were dried under N₂, derivatized with methoxyamine in pyridine and then N-Methyl-N-(trimethylsilyl) trifluoroacetamide with 1% trimethylchlorosilane. The samples were analyzed on a LECO Pegasus 4D GCxGC-TOFMS in electron ionization mode. An Agilent (Agilent Technologies, Santa Clara, CA) DB-5ms UI GC column was the primary column and an Agilent DB-17ms was the secondary column. Pooled QC samples were analyzed after every nine samples to assess overall reproducibility and to correct for any observed variations. An alkane retention index standard (C10 – C40) was run at the beginning, middle and end of sample runs. GCxGC-TOFMS Peak lists were generated with LECO ChromaTOF software version 4.50.8.0 (LECO Corporation, St. Joseph, MI) using vendor-recommended parameters. NIST/EPA/NIH Mass Spectral Library 2011 (NIST11), LECO/Fiehn Metabolomics Library, and an in-house library were used as reference libraries for compound identification. MetPP software was used for peak merging and peak list alignment, and retention index (RI) matching was performed using iMatch algorithms with p-value threshold ≤ 0.001.^{20, 21}

For the targeted assay, we used the Biocrates AbsoluteIDQ p150 kit (Biocrates Life Sciences AG, Austria) which profiles specific amino acids, acylcarnitines, hexoses, and phospho- and sphingolipids. For the Biocrates kit, 10 μl serum aliquots were processed and analyzed using the Biocrates AbsoluteIDQ p150 kit protocol.

Gene expression data

Whole-genome expression data were generated from three Illumina HumanHT-12 v4.0 Gene Expression BeadChips for 24 of the 25 study participants (Supplementary Table 1). Biotin-labeled cDNA were generated and hybridized following the Illumina Whole Genome Gene Expression Direct Hybridization Assay. Case and control samples were randomized and sample duplicates included as assay replicate controls. Arrays were scanned with the Illumina iScan Microarray Scanner system. Sample replicates (1 per chip) and array assay controls (1 per assay) confirmed assay integrity and ruled out contamination with identical calls and predicted genotypes.

Genotypic data

Sequenom platform Iplex Gold Reagent Kit multiplexed genotyping assays were designed in ~25 SNP pools using the Assay Design Suite (Agena Biosciences) for 186 of 200 putative non-MHC MS risk variants.²² Linked discovery SNPs or tagging SNPs were substituted for effect SNPs refractory to multiplex design. Samples were spotted to Sequenom SpectroChip(s) and scanned with MALDI-TOFMS. Non-template negative controls, sample duplicates and control DNAs were included. HLA-DRB1*15:01 alleles were determined via TaqMan genotyping of SNP rs9271366 (assay ID: C__33416976_20). HLA-A*02 alleles were determined via TaqMan genotyping of the SNP rs2975033 (assay ID: C__15962692_20). Each SNP assay was reviewed for the genotype call quality based on PCR/Extension yield, proximity to cluster, and raw mass spectra.

Statistical analyses

There were 400 metabolite variables captured across the GCxGC-TOFMS and Biocrates p150 approaches (Supplementary Table 2). 325 metabolites with <30% missing observations were retained for analyses implemented in MetaboAnalyst (http://www.metaboanalyst.ca/).²⁴ ComBat was used to adjust for batch effects (Supplementary Figure 1A), and missing values were replaced using the default setting (by half of the minimum positive values detected in the data) in MetaboAnalyst.²⁵ Random forests, a supervised machine learning (non-parametric) algorithm that is well adapted for prediction and variable importance, fairly robust to the setting of tuning parameters, and capable of analyzing data where predictors outnumber sample size, was used to identify metabolites important (in contrast to a p-value) for MS classification.²⁶ The forest consisted of 5,000 trees and 100 randomly selected metabolites were used to determine classification at each node in a tree. A scree plot of the ranked variable importance scores was used to identify important (top-ranking) metabolites for MS classification (Supplementary Figure 2). Similar results were obtained for quantile normalized and pareto scaled metabolomics data (data not shown). Logistic regression models were used to assess the relationship between each important metabolite and MS status, including a random effect to account for any residual batch effects using the xtlogit function in STATA v13.1 (StataCorp, TX). The area under the curve (AUC) and 95% Bamber and Hanley confidence intervals were determined by non-parametric receiver operating characteristic (ROC) analyses using the roctab function in STATA for each metabolite. Metabolites with AUC>80% were considered our top metabolites.

Gene expression data were background subtracted and quantile normalized within GenomeStudio Software Gene Expression Module (Illumina). Expression values with detection p-values >0.05 and genes with at least 70% of observations were retained; 9067 genes passed QC. ComBat, as implemented in MetaboAnalyst, was applied to remove any chip effects (Supplementary Figure 1B) and data were quantile normalized and pareto scaled. The expression of CD19 (B lymphocyte marker); CD14 (monocyte marker); FCGR3A and NCAM1 (natural killer cell markers); and CD3E, CD3D, CD4, and CD8A (T lymphocyte markers) did not differ between cases and controls via Wilcoxon rank-sum tests, indicating no blood cellular composition differences between cases and controls (data not shown).²⁷ Therefore, we conducted the following multivariable linear mixed-effects regression models: each top metabolite was the outcome and each gene expression value was a predictor, as well MS status, and a random effect to capture residual chip effects using the xtreg function in STATA. The top 2.5% (p-value ranked) of genes associated with each metabolite were analyzed via the Ingenuity Pathway Analysis (IPA) for biological pathway enrichment. The Illumina HumanHT-12 v4.0 Gene Expression BeadChip gene set was used as the reference set. Interaction networks, including endogenous chemicals, and all node types were included in the analysis (35 molecules per network, 25 networks per analysis). Data sources were experimentally observed from mammalian species.

Genetic data were generated along with another 107 samples. QC was applied to the entire genetic data set and included removing samples with >50% missing genotypes; SNPs with >20% missing genotypes; samples with >10% genotypes; and SNPs with Hardy Weinberg Equilibrium p<0.01. After QC, the final genotype dataset included 19 subjects (Supplementary Table 1) and 175 putative non-HLA MS risk SNPs (Supplementary Table 3). Multivariable regression models with the top metabolites at the outcome and each SNP as a predictor, adjusting for MS status and a random effect capturing possible plate effects were conducted. Two HLA alleles were genotyped in all 25 samples, however genotypes for the HLA-DRB1*15:01 and HLA-A*02 tagging SNPs were successfully determined in 22 and 24 samples respectively. Similar regression models were conducted as described above.

A two-sided alpha<0.05 was considered statistically significant for all statistical comparisons.

Results

Identifying top metabolites associated with MS

Random forests identified 12 metabolites as informative for MS classification (Table 1, Supplementary Figure 2). The random forests’ out-of-bag error for classifying the 25 samples by MS status was 0.32 (0.42 for cases and 0.23 for controls; Supplementary Figure 3). Eight of these metabolites were associated with MS in logistic regression models (p<0.05), and 6 of them had an AUC >80% (Supplementary Figure 4). The top 6 were two glycerophospholipids, two metabolites associated with fatty acid metabolism, an amino acid, and a biothiol scavenger. All had higher levels in MS cases than controls (Figure 1).

Table 1.

Top ranked metabolites from Random forests analyses and logistic regression associations for MS status.

Metabolite	Random Forest Rank	Logistic regression p-value	AUC (95% CI)	Classa	Biological processesa	Sourcea
Pyroglutamate	1	0.033	0.85 (0.68, 1)	Amino acid derivative/Amino acid	Glutathione metabolism/Unknown	Endogenous, food
Laurate	2	0.035	0.86 (0.72, 1)	Medium chain fatty acid	Lipid and fatty acid metabolism	Endogenous, food
UnknownZ11	3	0.123	0.20 (0, 0.41)	Unknown	Unknown	Unknown
N-methoxy carbonylproline, methyl ester	4	0.261	0.26 (0.03, 0.5)	Amino acid derivative	Unknown	Unknown
Phosphatidylcholine PC ae C42:5	5	0.018	0.81 (0.64, 0.99)	Glycerophospholipid	Cellular membranes	Endogenous, food
C14:1 (Tetradecenoyl-L-carnitine)	6	0.018	0.82 (0.65, 0.99)	Acylcarnitine	Long-chain fatty acid oxidation	Endogenous, food
Unknown056	7	0.048	0.21 (0.01, 0.4)	Unknown	Unknown	Unknown
Phosphatidylcholine PC ae C40:5	8	0.016	0.82 (0.64, 1)	Glycerophospholipid	Cellular membranes	Endogenous, food
N-Methylmaleimide	9	0.015	0.82 (0.64, 1)	Electrophile	Biothiol scavenger	Unknown
Unknown022	10	0.71	0.61 (0.35, 0.87)	Unknown	Unknown	Unknown
Myo-Inositol	11	0.034	0.21 (0.03, 0.39)	Cyclohexanol	Galactose, inositol, and phosphatidylinositol metabolism	Endogenous, food
Linolic acid	12	0.051	0.70 (0.46, 0.94)	Long chain fatty acid	Lipid and fatty acid metabolism	Endogenous, food, environment

Abbreviations: 95% CI = 95% confidence interval; AUC = area under the curve

^a.Class, gene ontology, and source derived from the Human Metabolome Database (HMDB) and PubChem (NCBI).

Figure 1.

Box plots of the top metabolites by MS status.

Gene expression associations with top metabolites

We explored the relationships between expression for 9,067 genes and each of the top metabolites (Supplementary Table 4). Among MS-related genes, HLA-DRB1 was not associated with any metabolite, however the expression of several other HLA genes were associated with acylcarnitine C14:1 including HLA-DMA, HLA-DMB, HLA-DOA, HLA-DPA1, HLA-DPB1, HLA-DRA, HLA-DRB3, and HLA-DRB6. The expression of multiple genes proximal (+/− 25 kilobases) to the 200 putative non-MHC MS risk variants were associated with the metabolites (Supplementary Table 5).²² Of note are the associations for KPNB1, CLEC16A, PIK3R2, IKZF1, TXK and PHGDH, which were associated with at least three of the 6 metabolites (p<0.1; Supplementary Table 5).

Pathway enrichment analyses of the top 2.5% p-value ranked genes associated with each metabolite, after adjustment for MS status, were conducted and multiple pathways were enriched for each metabolite (Supplementary Table 6). The top-ranked pathways included iron homeostasis signaling (pyroglutamate), ceramide degradation (pyroglutamate), antigen presentation (acylcarnitine C14:1), vitamin D biosynthesis (PC ae 42:5), and mitochondrial dysfunction (PC ae 40:5 and N-Methylmaleimide) (Table 2). We cross-referenced all metabolite-associated enriched pathways with pathways enriched among genes associated with the putative 200 non-MHC risk variants by the International MS Genetics Consortium (Supplementary Table 6).²² A third of the metabolite-associated enriched pathways overlapped with the MS-related pathways. There were also several pathways enriched among the associated genes across metabolites (Supplementary Table 7), specifically apoptosis related processes (death receptor signaling, calcium-induced T lymphocyte apoptosis, lymphotoxin β receptor signaling, and TNFR1 signaling) and altered mitochondrial function (mitochondrial dysfunction, oxidative phosphorylation, and sirtuin signaling pathway).

Table 2.

Top 5 pathways enriched per metabolite based on gene expression analyses.

Metabolite	Biological Pathways	P-value	Genes
Pyroglutamate	Iron homeostasis signaling pathway	1.3E-04	HBZ, FTL, HSCB, CIAO1, BMP6, CUL1, SLC48A1
	Ceramide Degradation	1.5E-03	ACER3, NAAA
	Sphingosine and Sphingosine-1-phosphate Metabolism	2.6E-03	ACER3, NAAA
	Biotin-carboxyl Carrier Protein Assembly	2.6E-02	ACACB
	Glutathione Biosynthesis	2.6E-02	GCLM
Laurate	Ephrin B Signaling	7.2E-04	HNRNPK, GNB1, PXN, RAC2, EPHB1
	Inflammasome pathway	9.1E-04	MYD88, NLRP1, PYCARD
	Germ Cell-Sertoli Cell Junction Signaling	1.3E-03	IQGAP1, PAK2, PIK3CD, RHOB, PXN, TUBA4A, RAC2
	Melanocyte Development and Pigmentation Signaling	2.6E-03	PIK3CD, PLCG1, RPS6KA4, RPS6KC1, PRKAR1A
	SAPK/JNK Signaling	2.7E-03	PIK3CD, MAP4K2, HNRNPK, GNB1, RAC2
Phosphatidylcholine PC ae C42:5	Asparagine Biosynthesis I	1.1E-02	ASNS
	Role of Oct4 in Mammalian Embryonic Stem Cell Pluripotency	1.4E-02	FAM208A, IGF2BP1, NR2F6
	Dolichol and Dolichyl Phosphate Biosynthesis	2.2E-02	DOLK
	1,25-dihydroxyvitamin D3 Biosynthesis	3.3E-02	CYP2R1
	S-adenosyl-L-methionine Biosynthesis	3.3E-02	MAT2A
Phosphatidylcholine PC ae C40:5	Oxidative Phosphorylation	2.8E-05	NDUFB6, COX17, NDUFA7, NDUFB7, ATP5PF, COX5B, VPS9D1
	Mitochondrial Dysfunction	1.1E-04	CYB5R3, NDUFB6, COX17, NDUFA7, NDUFB7, ATP5PF, COX5B, VPS9D1
	Sirtuin Signaling Pathway	1.3E-03	NDUFB6, PARP1, NDUFA7, NDUFB7, ATP5PF, GADD45A, POLR1C, SLC2A1, ATG16L2
	Death Receptor Signaling	1.9E-03	PARP12, PARP1, DFFA, PARP14, BIRC3
	Calcium-induced T Lymphocyte Apoptosis	1.9E-03	HLA-DQB1, EP300, CD4, PPP3CC
Tetradecenoyl-L-carnitine (C14:1)	Antigen Presentation Pathway	4.3E-08	HLA-DPB1, HLA-DOA, HLA-DMA, HLA-DPA1, HLA-DMB, HLA-DRA, CIITA
	Allograft Rejection Signaling	1.5E-06	HLA-DPB1, HLA-DOA, HLA-DMA, HLA-DPA1, HLA-DMB, HLA-DRA
	OX40 Signaling Pathway	6.6E-06	HLA-DPB1, HLA-DOA, HLA-DMA, HLA-DPA1, HLA-DMB, HLA-DRA
	Cdc42 Signaling	1.9E-05	HLA-DPB1, HLA-DOA, HLA-DMA, HLA-DPA1, HLA-DMB, HLA-DRA, ARPC1A, VAV2
	B Cell Development	7.8E-05	HLA-DOA, HLA-DMA, HLA-DMB, HLA-DRA
N-Methylmaleimide	Oxidative Phosphorylation	6.0E-06	ATP5S, NDUFA4, COX7B, NDUFA9, NDUFA6, ATP5F1C, UQCRB, NDUFAB1
	Sirtuin Signaling Pathway	7.6E-06	NFKB1, GOT2, CLOCK, SP1, NDUFAB1, RELA, NDUFA4, PARP1, ATG3, NDUFA9, NDUFA6, ATP5F1C, GTF3C2
	Superpathway of Geranylgeranyldiphosphate Biosynthesis I (via Mevalonate)	2.0E-05	IDI1, HMGCR, FNTB, ACAT1
	Mitochondrial Dysfunction	3.6E-05	ATP5S, NDUFA4, COX7B, NDUFA9, NDUFA6, CASP3, ATP5F1C, UQCRB, NDUFAB1
	Mevalonate Pathway I	2.4E-04	IDI1, HMGCR, ACAT1

MS risk alleles associations with top metabolites

We assessed the relationship between genetic risk and MS-predictive metabolite levels. While no association was observed for HLA-A*02, we identified an association between HLA-DRB1*15:01 and acylcarnitine C14:1 (Table 3; Figure 2). We also observed multiple associations between metabolite levels and the 175 putative non-MHC risk variants that we genotyped (Table 4). Of note were five SNPs associated with both phosphatidylcholines PC ae 40:5 and PC ae 42:5, four of which were genic variants within ETS1, IL2RA and AFF1.

Table 3.

Genetic associations between HLA variants and top metabolites.

Metabolite	HLA-DRB1*15:01		HLA-A*02
	Effect Direction	P-value	Effect Direction	P-value
Pyroglutamate	−	0.66	−	0.31
Laurate	−	0.31	+	0.79
Phosphatidylcholine PC ae C42:5	+	0.80	+	0.68
Phosphatidylcholine PC ae C40:5	−	0.53	+	0.21
C14:1 (Tetradecenoyl-L-carnitine)	−	0.02	+	0.75
N-Methylmaleimide	−	0.10	+	0.29

Figure 2.

Box plot of C14:1 Tetradecemoyl-L-carnitine by HLA*DRB1*15:01 genotype.

Table 4.

Genic non-MHC MS risk variant associations with the top metabolites.

Metabolite	Associated SNPs	Location (GRCh38/hg38)	Minor Allele	Effect Direction	P-value	SNP Proximal Genes
Pyroglutamate	rs531612	chr11:65937961	C	−	0.013	DRAP1, TSGA10IP
	rs10230723	chr7:50200284	T	+	0.019	C7orf72, IKZF1
	rs2150879	chr17:59781849	A	−	0.024	VMP1
	rs34681760	chr5:6712721	T	+	0.033	LOC100505625, PAPD7
	rs9282641	chr3:122077921	A	+	0.047	CD86
Laurate	rs4766578	chr12:111466567	T	−	0.017	SH2B3
	rs2317231	chr1:157716547	T	+	0.019	FCRL3, FCRL2
	rs9913257	chr17:75324812	C	+	0.023	GRB2
	rs6032662	chr20:46105671	C	+	0.027	CD40, NCOA5
	rs2269434	chr11:47338861	C	+	0.034	MYBPC3
	rs12925972	chr16:79077400	T	−	0.043	WWOX
	rs17780429	chr6:137901451	A	+	0.045	LOC100130476
Phosphatidylcholine PC ae 42:5	rs4262739	chr11:128551280	A	−	0.003	ETS1
	rs12373588	chr2:111708688	G	+	0.007	ANAPC1, MIR4435-1HG
	rs12622670	chr2:68419404	C	+	0.014	FBXO48, PLEK
	rs2705616	chr4:86941244	G	+	0.018	AFF1
	rs2705618	chr4:86916393	T	+	0.018	AFF1
	rs62013236	chr15:78955140	T	+	0.020	CTSH, RASGRF1
	rs6589939	chr11:122647817	G	−	0.020	MIR100HG, UBASH3B
	rs116899835	chr14:88057144	T	+	0.021	LINC01146
	rs6738544	chr2:191124630	A	−	0.029	STAT4
	rs12722559	chr10:6028310	A	+	0.033	IL2RA
	rs2269434	chr11:47338861	C	−	0.035	MYBPC3
	rs9913257	chr17:75324812	C	−	0.047	GRB2
	rs6952809	chr7:2408858	T	+	0.048	CHST12
Phosphatidylcholine PC ae 40:5	rs1076928	chr6:36380912	C	+	0.001	ETV7
	rs4262739	chr11:128551280	A	−	0.004	ETS1
	rs12373588	chr2:111708688	G	+	0.007	ANAPC1, MIR4435-1HG
	rs1534422	chr2:12500615	A	−	0.016	LOC100506457
	rs883871	chr17:40096407	A	−	0.026	NR1D1
	rs719219	chr20:49818871	A	−	0.028	B4GALT5, SLC9A8
	rs12722559	chr10:6028310	A	+	0.034	IL2RA
	rs11852059	chr14:51839373	C	+	0.035	FRMD6, GNG2
	rs137956	chr22:39897459	C	+	0.036	ENTHD1, GRAP2
	rs9878602	chr3:71486187	G	+	0.038	FOXP1
	rs140522	chr22:50532837	T	−	0.041	ODF3B
	rs9992763	chr4:108137562	T	−	0.042	LEF1
	rs2705616	chr4:86941244	G	+	0.044	AFF1
	rs2705618	chr4:86916393	T	+	0.044	AFF1
	rs34695601	chr14:75547955	C	+	0.047	BATF
Acylcarnitine C14:1	rs2705616	chr4:86941244	G	−	0.001	AFF1
	rs2705618	chr4:86916393	T	−	0.001	AFF1
	rs12614091	chr2:203768138	T	+	0.006	CD28, CTLA4
	rs1177228	chr2:61015275	A	+	0.016	PUS10
	rs2269434	chr11:47338861	C	+	0.02	MYBPC3
	rs9909593	chr17:39813896	G	−	0.025	IKZF3
	rs2331964	chr3:121824051	T	−	0.035	IQCB1
	rs983494	chr1:160734175	A	+	0.045	CD48, SLAMF7
N-methylmaleimide	rs28703878	chr8:78504987	G	+	0.001	LOC102724874, PKIA
	rs3923387	chr8:143912625	T	+	0.002	EPPK1, PLEC
	rs12708716	chr16:11086016	G	−	0.004	CLEC16A
	rs11852059	chr14:51839373	C	−	0.005	FRMD6, GNG2
	rs6837324	chr4:48125245	G	+	0.007	TXK
	rs10245867	chr7:28102567	T	+	0.012	JAZF1
	rs719316	chr6:16672529	C	−	0.015	ATXN1
	rs438613	chr3:28030595	C	+	0.026	CMC1, EOMES
	rs35716097	chr5:177379635	T	+	0.031	RGS14
	rs483180	chr1:119724882	G	+	0.032	PHGDH
	rs12614091	chr2:203768138	T	−	0.041	CD28, CTLA4

Discussion

We combined exploratory open platform metabolomics with transcriptomics and risk variant genotyping to gain a multifaceted multiomic perspective on metabolic changes associated with MS. Our analyses identified 6 metabolites as promising MS predictive candidates involved in glutathione metabolism (pyroglutamate), fatty acid metabolism/oxidation (laurate, C14:1 acylcarnitine), cellular membrane composition (glycerophospholipids PC ae 42:5 and PC ae 40:5), and a transient receptor potential A1 (TRPA1) channel agonism (N-Methylmaleimide). Two of these metabolites have been previously associated with MS (pyroglutamate, laurate), while the remaining four have strong biologically plausible relevance.^{16, 17, 19, 28, 29} Integrated analyses with genomic and genetic data suggest distinct yet overlapping biological processes may be contributing to elevated levels of these metabolites in MS, including apoptotic processes and mitochondrial dysfunction.

Pyroglutamate (also known as 5-oxoproline) was the most important metabolite identified by random forest analysis. A natural cyclized derivative of glutamate that is generated during biosynthesis of the antioxidant glutathione, it has previously been associated with MS in other untargeted metabolomic analyses of CSF and plasma.^{16, 17, 19} Elevated pyroglutamate, as detected in our MS cohort, is associated with several rare metabolic disorders, including glutathione synthetase deficiency and 5-oxoprolinase deficiency. These disorders manifest neurological symptoms, including intellectual deficiency, microcephaly, and seizures. Murine experiments suggest that pyroglutamate may directly contribute to neurological symptoms, with peripheral pyroglutamate capable of crossing the blood brain barrier and triggering oxidative damage.^{30, 31} Pyroglutamate also stimulates sodium-dependent abluminal transport of amino acids across the blood brain barrier.³² As a marker of changes to glutathione metabolism and a source of oxidative damage, pyroglutamate may be both a marker of, and a contributor to, the environment of increased oxidative stress in MS.³³ Among the top-ranked pathways enriched among genes associated with this glutathione intermediate were glutathione biosynthesis, iron hemostasis signaling, and ceramide/sphigosine metabolism. The latter two pathways both contribute to apoptosis. Perturbed iron levels promote neuronal death, and abnormal iron deposition in the brain may facilitate neurodegeneration.³⁴ Interestingly, elevated ceramide levels have been observed near MS plaques, and sphingosine metabolism, which is essential for myelin synthesis and maintenance, has been reported to be abnormal in the MS brain.³⁵

The second most important metabolite identified by random forests was laurate (lauric acid), a medium-chain fatty acid that is elevated in the “Western Diet”.⁶ There is evidence supporting a role for this saturated fatty acid in MS pathology. In vitro treatment of mouse CD4+ T cells with laurate stimulated Th1 and Th17 differentiation and decreased Treg differentiation, while treatment of human CD4+ T cells promoted CD4+interferon-γ+ differentiation.^{28, 29} In a study of experimental autoimmune encephalomyelitis mice (a murine model of MS), increased dietary laurate promoted differentiation of Th1 and Th17 cells, which contributed to more severe disease.²⁹ Multiple pathways were associated with laurate, including Tec kinase signaling (essential for B and T cell development and activation), lymphotoxin β receptor signaling (apoptotic trigger that can result in the release of interleukin 8 [IL-8], and IL-8 contributes to Th17 development³⁷), IL-8 signaling, sphigosine-1-phosphate signaling, and axonal guidance (Supplementary Table 6).

Two phosphatidylcholines, the primary components of cellular membranes and myelin, were predictive of MS status: PC ae 40:5 and PC ae 42:5. The partial hydrolysis of phosphatidylcholines by phospholipase A2 enzymes generates an array of bioactive products, including eicosanoid pro-inflammatory signaling molecules and myelin-damaging lysophophatidylcholines. In our data, increased expression of a group IV phospholipase A2 enzyme, PLA2G4C, was significantly associated with increased PC ae 40:5 and PC ae 42:5 levels (Supplementary Table 4). PLA2G4C is one of the least well-characterized enzymes in the family, though it is important for mitochondrial function and genetic variants have been associated with schizophrenia and autism spectrum disorder.^{38, 39} PLA2G4C is expressed throughout the brain and protein levels are highest in nervous system tissues.^40,41,42 There is evidence PLA2G4C facilitates macrophage differentiation and polarization and CD4+ T cell transcripts have been found to be restricted to only PLA2G4A and PLA2G4C members of the family, suggesting this PLA2 enzyme may have particular importance in MS-relevant immune cells ^43,44.

There was little overlap in the pathways enriched among the genes whose expression was associated with either phosphatidylcholine, though MS risk variants in IL2RA, ETS1, and AFF1 were associated with both metabolites. ETS1 is a transcription factor that influences a wide array of cellular processes; interestingly, it is predicted to regulate PLA2G4C based on the presence of core consensus sequences in the promoter.⁴⁵ In our data, EST1 expression was associated with PC ae 42:5 levels (p=0.007; Supplementary Table 5). AFF1 is a core member of the transcription factor positive transcription elongation factor b (P-TEFb), which regulates RNA polymerase II-mediated transcription.⁴⁶ In our data, AFF1 expression was associated with PC ae 40:5 levels (p=0.017; Supplementary Table 5). Of note, variants near or within AFF1 have been associated with circulating lipid levels in genome-wide association studies, which supports an as yet uncharacterized role for AFF1 in lipid metabolism.^47-50

The fourth lipid-related metabolite predictive of MS in our analyses was acylcarnitine C14:1, an intermediate of fatty acid oxidation and a marker for the rare genetic metabolic disorder VCLAD (very long-chain acyl-dehydrogenase deficiency). In general, acylcarnitines are indicative of mitochondrial function and energy metabolism, which are both relevant to, and perturbed by, MS pathogenesis.^{51, 52} The sirtuin signaling pathway, which is associated with longevity and mitochondrial efficiency, was enriched among genes associated with acylcarnitine C14:1; it was also enriched among genes associated with PC ae 40:5 and N-methylmaleide (Supplementary Table 7).⁵³ Also enriched in association with C14:1 were genes involved with antigen presentation, including the expression of multiple class II HLA genes: HLA-DMA, HLA-DMB, HLA-DOA, HLA-DPA1, HLA-DPB1, HLA-DRA, HLA-DRB3, and HLA-DRB6 (Supplementary Table 4). Carriage of HLA-DRB1*15:01 risk alleles was also associated with C14:1, though not HLA-DRB1 expression (Table 3, Figure 2). The presentation of antigens by antigen presenting cells activates CD4+ helper T cells, which in turn triggers a switch in metabolic programming from fatty acid oxidation to glycolysis.⁵⁴ We hypothesize that a change in acylcarnitine levels paired with a change in class II HLA gene expression may reflect an activated immune environment promoting a transition in immune cell energy metabolism.

The final metabolite important for predicting MS status in these data was N-methylmaleimide, a transient receptor potential ankyrin 1 (TRPA1) channel agonist. TRPA1 is a nonselective membrane cation channel predominantly expressed in nociceptive sensory neurons but also astrocytes and other non-neuronal cells.^55-57 TRPA1 is activated by numerous endogenous and exogenous agonists and participates in hyperalgesia and neurogenic inflammation.^56-58 In mice with cuprizone-induced demyelination (a murine MS model that circumvents an autoimmune response), TRPA1 deficiency significantly reduced non-immune mediated demyelination by reducing apoptosis of mature oligodendrocyte.^{59, 60} As for N-methylmaleimide, little else is known about this electron deficient thiol-blocking agent. We did, however, associate this metabolite with an enrichment for pathways involved in mitochondrial functioning (mitochondrial dysfunction, oxidative phosphorylation, sirtuin signaling), cholesterol/steroid synthesis (mevalonate pathway 1), and apoptosis (Table 2, Supplementary Tables 6 and 7).

There are several strengths to this exploratory investigation of altered metabolic profiles in MS. First, we combined untargeted and targeted metabolomics approaches to sera from a homogeneous cohort of non-Hispanic white, non-smoking males with and without MS, who were frequency matched for age and BMI. Second, we focused on cases who were drug naïve for at least 3 months; thus, we are confident the metabolic signal is not modulated by drug activity. Third, we applied robust non-parametric methods to discern metabolites important for MS status, then used parametric models to determine metabolites most predictive of MS status. And lastly, we integrated multiomic data (transcriptomic and genetic) to explore possible patho-etiologic relationships contributing to variation in the metabolites of interest.

There are a few limitations we must acknowledge. First is the absence of a replication data set. Second, the use of prevalent and not incident MS cases was not ideal; as a result, the identified metabolites may not reflect processes conferring risk but ongoing pathophysiological processes. For example, we identified two phosphatidylcholines which are common components of myelin. Third, our cohort was relatively homogenous for sex, disease type, smoking status, BMI (most are overweight/obese), and disease duration (Supplementary Table 1) – while this was a strength in reducing confounding, it precluded us from investigating associations for these traits in this sample.

In conclusion, we applied targeted and untargeted discovery metabolomics approaches to identify 6 metabolites with elevated levels in MS cases compared with controls. Two of these MS-predictive metabolites have prior evidence for a role in MS, and there is a strong biological rationale for the remaining four peripheral blood metabolites being relevant to MS pathology.

Our integrated analyses suggested that perturbations in mitochondrial functioning, apoptosis, and energy metabolism contributed to the altered MS metabolic profiles. This study has generated a wealth of information that may lead to the development of new metabolic targets for improved MS diagnosis and treatment. Further research is warranted, particularly larger studies including diverse patient populations with considerations for sex, race and MS subtype. It would also be of great interest to investigate the relationships between the metabolites and MS progression (i.e. laurate was associated with disease severity in a murine MS model). And lastly, it would be important to determine if the metabolomics associations are specific to MS or represent inflammatory or neurodegenerative process that may be shared with other autoimmune and neurological disorders.

Highlights.

• Six metabolites were predictive for MS in males

• Results support lipidomic changes in MS

• Several MS risk variants were associated with the top predictive metabolites

• Pathways analyses suggest processes involved in MS pathology

Abbreviations

biomarkers

biological markers

GCxGC-TOFMS

two-dimensional gas chromatography and time-of-flight mass spectrometry

HLA

human leukocyte antigen

LD

linkage disequilibrium

MURDOCK

Duke University Measurement to Understand the Reclassification of Disease of Cabarrus/Kannapolis

QC

quality control

RR-MS

relapsing-remitting MS

SNP

single-nucleotide polymorphism

SP-MS

secondary-progressive MS