Plasma proteomics for Parkinson's disease classification: cross-cohort benchmarking of proteomic, transcriptomic, and multimodal models.

Blood-based biomarkers could facilitate early detection and severity monitoring for Parkinson's disease (PD), yet the relative diagnostic utility of proteomic versus transcriptomic signals and the added value of multimodal integration under cross-cohort transfer remain unclear. We trained diagnostic classifiers (PD versus healthy control) using targeted plasma proteomics (Olink) and whole-blood RNA sequencing from a development cohort under participant-grouped cross-validation with nested preprocessing, then externally validated models on an independent cohort. RNA-only models were benchmarked across multiple dimensionality-reduction and regularization strategies, and multimodal integration was evaluated using early fusion, balanced early fusion, late fusion, and stacking. A Proteomic Severity Index (PSI) was derived from baseline protein expression and assessed with linear and non-linear regressors. The proteomics-only Random Forest classifier achieved the strongest external validation (AUROC = 0.8724, 95% CI 0.8305-0.9097; AUPRC = 0.8989), whereas the best RNA-only configuration reached only 0.5978. No fusion strategy significantly improved discrimination beyond proteomics alone. Among 32 selected proteins, DDC showed the strongest severity association with total MDS-UPDRS (ρ = 0.61), and the linear PSI explained 28.2% of severity variance, though PSI residuals showed no independent longitudinal association. These findings support targeted plasma proteomics as the primary molecular modality for blood-based PD biomarker development.