Detection of prodromal Parkinson’s disease using a urine proteomics panel and machine learning

Abstract

AbstractParkinson’s disease is a progressive neurodegenerative disorder and idiopathic REM-sleep behaviour disorder (iRBD) has been identified as its single most specific early symptom. To facilitate the screening of individuals at high risk to develop Parkinson’s disease, we developed a multiplexed panel of urine proteomics using machine learning and targeted mass spectrometry to detect iRBD. Random urine samples from clinically and genetically well characterized patients with iRBD, idiopathic and hereditary forms of Parkinson’s disease, and matching controls, collected in two academic centres, were analysed in a standardized way.First, a biomarker discovery and exploratory comparison of samples from randomly selected idiopathic Parkinson’s patients and age/sex-matched healthy controls were proteomically profiled and quantified (> 2500 proteins). The most differentially expressed biomarkers were combined into a high-throughput, multiplexed assay using targeted proteomics designed for use of tandem mass spectrometers for potential translation into clinical practise. This was then validated on independent patient and control samples (n=184). After detecting a major influence of sex on the proteome, we focused subsequent analyses on the larger group of available male samples (n= 114) and report results based on iRBD (n= 14), idiopathic (n= 35) and young-onset Parkinson’s disease (n= 15), carriers of LRRK2 (n= 10), PARKIN-gene mutations (n= 5), and healthy control subjects (n= 35). After establishing excellent compatibility between the two study sites, orthogonal partial least squares discriminant analysis (OPLS-DA) excluded a relevant effect of aging, but detected significant differences between iRBD and healthy controls (ANOVA-CVP= 0.002), as well as the combination of iPD/iRBD and healthy controls (ANOVA-CVP= 0.01).Uni- and multivariate analyses detected a shared expression pattern for the protein biomarkers UBC, NCAM1, MIEN1, SPP2, REG1B, ITIH2, BCHE and C3 between iRBD and idiopathic Parkinson’s disease. Utilizing split train/test-datasets in a multiple-regression classifier model resulted in a mean accuracy of 78% to detect iRBD, matching iRBD’s conversion rate to Parkinson’s disease. Hierarchical clustering revealed greater similarities in urine proteomic changes between iRBD and idiopathic than monogenic Parkinson’s disease. Several proteins identified correlated either with clinical severity (e.g. VCAM1, MSN, HPX), or risk for future conversion to Parkinson’s disease (VCAM1, MSN, MYO10, HSPAIL).This demonstrates the power of machine learning and urine biomarkers to identify iRBD patients. As we develop new therapies and interventions, the ability to detect individuals at-risk of neurodegeneration in very early disease stage will be invaluable for treatment success.