Predictive Utility of Serum Protein Biomarkers from the Octave MSDA Panel for Optic Neuritis Events in People with Multiple Sclerosis
Kian Jalaleddini1, Dejan Jakimovski2, Anisha Keshavan1, Shannon McCurdy1, Ferhan Qureshi1, Ati Ghoreyshi1, Niels Bergsland2, Michael Dwyer2, Murali Ramanathan3, Bianca Weinstock-Guttman4, Ralph Benedict3, Robert Zivadinov2
1Octave Bioscience, 2Buffalo Neuroimaging Analysis Center, University at Buffalo, 3University At Buffalo, 4Department of Neurology, University At Buffalo
To assess the predictive power of serum protein biomarkers measured from the Octave Multiple Sclerosis Disease Activity (MSDA) immunoassay panel for predicting optic neuritis (ON) events in people with MS (pwMS) using machine learning models.
ON, a common manifestation in multiple sclerosis (MS), can substantially affect patients' quality of life. The prediction of ON using serum biomarkers could offer clinicians valuable insights into disease progression and potentially guide therapeutic strategies, however the efficacy of biomarkers in forecasting ON in MS remains underexplored.
The proteomic analysis was performed in a subset of 109 patients from the larger CEG-MS initiative with ON prevalence of 50.5%. Patients either reported ON events 1.1(1.4) years after the baseline proteomic measurement or never reported ON events. Serum biomarker measurements from the Octave MSDA panel were extracted and pre-processed using proximity extension assay methodology on the Olink platform. Three machine learning models (Support-Vector-Machine, Logistic-Regression, and Linear-Discriminant-Analysis) were trained to predict ON from baseline biomarkers and demographics in a leave-one-out cross-validation approach to optimize the F1-score.
The top-performing model was constructed using a sequence of preprocessing with StandardScaler and feature selection, and a Linear-Discriminant-Analysis model with a validation F1-score of 0.67. The most influential features in the model were OPN, CXCL13, CXCL9, TNFSF13B, and age. Further inspection of the protein concentration changes suggested that elevations in the concentrations of OPN (p<0.05), CXCL13 (p<0.001), and CXCL9 (p<0.0001) were predictive of ON in the next 5 years.
Multianalyte proteomic models that include key mediators in inflammatory processes are able to stratify pwMS regarding their risk of future ON occurrence. This suggests that heightened immune activation might precede and set the stage for the onset of clinical manifestations, offering a window of opportunity for early intervention.