Parkinson’s disease (PD) is increasing in prevalence, yet we lack readily available and non-invasive diagnostic biomarkers. An unbiased discovery proteomics study in plasma from PD patients and healthy controls (HC, phase 0) showed increased inflammation and decreased Wnt-signalling. To validate these findings, a targeted and multiplexed assay was developed and applied to a phase I cohort consisting of 99 de novo PD, 36 HC, and 18 subjects with polysomnography-confirmed isolated REM-sleep behaviour disorder (iRBD). In phase II, the multiplexed panel was refined and applied to a second cohort, focusing on the earliest stage of the disease by analysing 54 individuals with confirmed iRBD including longitudinal follow-up samples. A machine learning model, based on the expression of eight proteins, correctly identified all PD patients (phase I) and classified 79% of the iRBD as PD (phase II) up to 7 years before phenoconversion. Several of the identified biomarkers correlated with motor and non-motor symptom severity. This specific blood pattern was already evident in most iRBD cases, indicating pre-motor molecular events. This allows for the early identification of subjects at risk for PD in a blood test years before motor disease, which could support stratifying participants in future clinical trials aimed at preventing PD.