With the number of Parkinson's patients expected to rise due to an aging population, there is an increasing need to identify new diagnostic markers. These markers should be affordable and suitable for routine use to monitor the population, help stratify patients for treatment pathways, and provide new avenues for therapy. Genetic predisposition and familial forms account for approximately 10% of Parkinson's disease (PD) cases, leaving a large fraction of the population with minimal effective markers for identifying high-risk individuals. The establishment of population-wide omics and longitudinal health monitoring studies provides an opportunity to apply machine learning approaches to these unbiased cohorts to identify novel PD markers. In this study, we present the application of three machine learning models to identify protein plasma biomarkers of PD using plasma proteomic measurements from 43,408 UK Biobank subjects as the training and test set and an additional 103 samples from the Parkinson's Progression Markers Initiative (PPMI) as external validation. We identified a group of highly predictive protein plasma markers, including known markers Dopa decarboxylase (DDC) and Calbindin 2 (CALB2) as well as new markers involved in the JAK-STAT and PI3K-AKT pathways and hormonal signaling. We further demonstrated that these features are well correlated with UPDRS severity scores and stratified these into protective and risk-associated features that potentially contribute to the pathogenesis of PD.