Machine learning-based prediction of disease progression in primary progressive multiple sclerosis

Primary progressive multiple sclerosis (PPMS) affects 10-15% of multiple sclerosis patients and presents significant variability in the rate of disability progression. Identifying key biological features and patients at higher risk for fast progression is crucial to develop and optimize treatment strategies. Peripheral blood cell transcriptome has the potential to provide valuable information to predict patients' outcomes. In this study, we utilized a machine learning framework applied to the baseline blood transcriptional profiles and brain MRI radiological enumerations to develop prognostic models. These models aim to identify PPMS patients likely to experience significant disease progression and who could benefit from early treatment intervention. RNA-sequence analysis was performed on total RNA extracted from peripheral blood mononuclear cells of PPMS patients in the placebo arm of the ORATORIO clinical trial (NCT01412333), using Illumina NovaSeq S2. Cross-validation algorithms from Partek Genome Suite (www.partek.com) were applied to predict disability progression and brain volume loss over 120 weeks. For disability progression prediction, we analysed blood RNA samples from 135 PPMS patients (61 females and 74 males) with a mean ± standard error age of 44.0 ± 0.7 years, disease duration of 5.9 ± 0.32 years and a median baseline Expanded Disability Status Scale (EDSS) score of 4.3 (range 3.5-6.5). Over the 120-week study, 39.3% (53/135) of patients reached the disability progression end-point, with an average EDSS score increase of 1.3 ± 0.16. For brain volume loss prediction, blood RNA samples from 94 PPMS patients (41 females and 53 males), mean ± standard error age of 43.7 ± 0.7 years and a median baseline EDSS of 4.0 (range 3.0-6.5) were used. Sixty-seven per cent (63/94) experienced significant brain volume loss. For the prediction of disability progression, we developed a two-level procedure. In the first level, a 10-gene predictor achieved a classification accuracy of 70.9 ± 4.5% in identifying patients reaching the disability end-point within 120 weeks. In the second level, a four-gene classifier distinguished between fast and slow disability progression with a 506-day cut-off, achieving 74.1 ± 5.2% accuracy. For brain volume loss prediction, a 12-gene classifier reached an accuracy of 70.2 ± 6.7%, which improved to 74.1 ± 5.2% when combined with baseline brain MRI measurements. In conclusion, our study demonstrates that blood transcriptome data, alone or combined with baseline brain MRI metrics, can effectively predict disability progression and brain volume loss in PPMS patients.