Two Cohorts, Three Classifiers: Metabolomic Signatures of MS Severity

Villoslada and colleagues built their study around a feature most metabolomics work in multiple sclerosis (MS) lacks: repeated sampling of the same people. Their prospective cohort of 61 patients and 41 matched healthy controls gave serum every three months for two years, which allows a question a single blood draw cannot answer, namely whether a metabolite difference is stable in an individual over time or an artifact of the day it was drawn. A second, retrospective cohort of 238 patients and 74 controls across five Spanish centres supplied the numbers needed for between-person variability. Both cohorts were phenotyped prospectively for two years, with relapses and disability recorded on the Expanded Disability Status Scale (EDSS). Serum went through ultra-high-performance liquid chromatography coupled to mass spectrometry, with samples randomised before extraction and analysed blind to diagnosis. The methods also carry an admission worth reading before the results: sample sizes came from a convenience dataset, no power calculation was performed, and the authors call the work exploratory.

Separating Patients From Controls, and a Drug Effect in the Middle of It
Unsupervised analysis of patients against controls gave a goodness of fit of 0.247 and a predictive ability of 0.128. The supervised model did better, reaching 0.481 for fit and 0.348 for predictive ability. The more interesting result came from the time series. Across two years of repeated sampling, 29 spectral peaks were consistently associated with MS (p < 0.01), and identifying them pointed to sphingomyelin and a lysophosphatidylethanolamine as the core of the signature. That is exactly what serial sampling is for, since a molecule that holds its difference across eight or nine draws per person is a different kind of finding from one that shows up once. One problem arrives immediately: the change in lysophosphatidylethanolamine over time was associated with interferon-β use (p = 0.01). Half of the headline signature moves with the treatment.

One Classifier Works Well, Another Does Almost Nothing
The study built three severity models, and their performance differs enough that averaging them would mislead. Separating patients who stayed below EDSS 3.0 from those who reached above 4.5 produced the best result, with a fit of 0.756 and predictive ability of 0.555. Separating relapse-free patients from those who relapsed was weaker, at 0.425 and 0.231. Predicting the change in EDSS across the two years gave a fit of 0.222 and a predictive ability of 0.036. Predictive ability near zero means the model forecasts nothing, and the paper nonetheless describes this group of results collectively as showing medium to high accuracy. The distinction is worth preserving, because telling apart a patient who is already disabled from one who is not is a far easier problem than working out who will become disabled, and only the first of those worked here.

The Molecules That Survived Both Cohorts
In the retrospective cohort, 116 metabolites differed between stable patients (n = 77) and those with active disease (n = 134) at uncorrected p < 0.05, which on its own is close to what multiple testing would produce. The authors then applied the filter that makes the result worth reading: they kept only metabolites that survived correction for multiple testing and were also significant in the prospective cohort. For relapse-free status, seven survived, led by 13S-hydroxy-octadecadienoic acid at p = 1.94E-09 and the lysophosphatidylcholine LysoPC (17:0/0:0) at p = 0.0001, alongside three further lysophosphatidylcholines, arachidonic acid, and a diacylglycerophosphocholine. For change in disability, the surviving set was cortisol, glutamic acid, tryptophan, eicosapentaenoic acid, the same oxidised fatty acid, and two more lysophospholipids. Requiring a molecule to appear in two independently collected cohorts is a stronger standard than most single-cohort studies apply.

Cortisol Shows What These Numbers Mean in Practice
The paper singles out cortisol as a candidate predictor of disability increase two years later, and the accompanying figures are the ones a clinician would actually need. Area under the curve was 0.6, sensitivity 0.38, and specificity 1.0. Specificity of 1.0 with sensitivity of 0.38 describes a test that essentially never flags someone who will not worsen, and also misses about two-thirds of those who will. An area under the curve of 0.6 sits nearer a coin flip than to a usable test. There is also a reporting wrinkle: the table lists cortisol at p = 0.027 for the disability comparison while the text reports p = 0.051 as a trend, the two coming from different tests. None of this makes cortisol uninteresting as biology, but it does leave the finding well short of clinical use, which the authors acknowledge when they write that the accuracy of their disability classifiers was not very high.

What the Lipid Pattern Points At
The signature amounts to an imbalance between phospholipids and sphingolipids, with shifts in two amino acids on top. Each component has a plausible route into MS biology. Sphingomyelins are among the main lipid classes in myelin and signal through the sphingosine-1-phosphate receptor, which an existing class of MS drugs already targets. Phosphatidylethanolamine modulates immune responses by activating CD300 receptors. Phosphatidylcholine is the main phospholipid of cell membranes and governs proliferative growth and programmed cell death. Arachidonic acid acts as a second messenger for phospholipase activation and as an inflammatory mediator in its own right. Glutamic acid activates P2X7 purinergic receptors on macrophages and contributes to oligodendrocyte death in white matter, and tryptophan sits at an immune checkpoint controlled by indoleamine 2,3-dioxygenase. The authors favour a single reading over a pathway-specific one: these serum changes reflect chronic immune activation, and greater severity corresponds to greater activation. They set it against a competing account from brain tissue, where normal-appearing grey matter in MS shows higher phospholipid and lower sphingolipid content, consistent with sphingolipids being diverted into phospholipid production.

The Validation They Planned and Could Not Run
The sharpest limitation is one the authors state themselves. Holding two independent cohorts, the informative design was available to them: use one cohort to pick features, train a classifier with internal cross-validation, then test it in the other and report the area under the curve. They could not, because the prospective cohort's platform returned spectral peaks rather than named metabolites. What the paper reports instead is that certain molecules reached significance in both cohorts, which is a weaker claim than a classifier that transfers between them. Several other constraints sit alongside it. Disease-modifying drugs were permitted, so drug effects and disease activity cannot be fully separated, though most treated patients received interferon-β, which reduces that heterogeneity. Disease activity was defined clinically without MRI, so patients with new lesions in clinically silent regions may have been counted as stable. Oligoclonal band status was not analysed. The sample was a convenience dataset. The metabolomic study was designed and its analysis supervised by employees of a metabolomics company, which the paper discloses. The study that follows is the one the authors describe and could not perform: both cohorts run on a platform that names its metabolites, a classifier trained in one, and its performance reported in the other.

Disclaimer: This blog post is based on the cited research article and is intended for informational purposes only. It is not intended to provide medical advice. Please consult with a healthcare professional for any health concerns.

Reference:
Villoslada, P., Alonso, C., Agirrezabal, I., Kotelnikova, E., Zubizarreta, I., Pulido-Valdeolivas, I., Saiz, A., Comabella, M., Montalban, X., Villar, L., Alvarez-Cermeño, J. C., Fernandez, O., Alvarez-Lafuente, R., Arroyo, R., & Castro, A. (2017). Metabolomic signatures associated with disease severity in multiple sclerosis. Neurology: Neuroimmunology & Neuroinflammation, 4(2), e321. https://doi.org/10.1212/NXI.0000000000000321