Machine learning-driven diagnosis of multiple sclerosis from whole blood transcriptomics

Abstract

Multiple sclerosis (MS) is a neurological disorder characterized by immune dysregulation. It begins with a first clinical manifestation, a clinically isolated syndrome (CIS), which evolves to definite MS in case of further clinical and/or neuroradiological episodes. Here we evaluated the diagnostic value of transcriptional alterations in MS and CIS blood by machine learning (ML).

Deep sequencing of more than 200 blood RNA samples comprising CIS, MS and healthy subjects, generated transcriptomes that were analyzed by the binary classification workflow to distinguish MS from healthy subjects and the Time-To-Event pipeline to predict CIS conversion to MS along time. To identify optimal classifiers, we performed algorithm benchmarking by nested cross-validation with the train set in both pipelines and then tested models generated with the train set on an independent dataset for final validation.

The binary classification model identified a blood transcriptional signature classifying definite MS from healthy subjects with 97% accuracy, indicating that MS is associated with a clear predictive transcriptional signature in blood cells. When analyzing CIS data with ML survival models, prediction power of CIS conversion to MS was about 72% when using paraclinical data and 74.3% when using blood transcriptomes, indicating that blood-based classifiers obtained at the first clinical event can efficiently predict risk of developing MS.

Coupling blood transcriptomics with ML approaches enables retrieval of predictive signatures of CIS conversion and MS state, thus introducing early non-invasive approaches to MS diagnosis.