Plasma proteomics identify biomarkers predicting Parkinson's disease up to 7 years before symptom onset

Parkinson's disease (PD) is a heterogeneous and increasingly prevalent neurodegenerative disorder marked by progressive motor and non-motor symptoms driven by α-synuclein aggregation and Lewy body formation. Failures of neuroprotective trials are partly attributed to disease heterogeneity and a lack of objective biomarkers, especially in early or pre-motor stages. Current highly specific assays such as α-synuclein seed amplification assays (SAA) are robust in cerebrospinal fluid but rely on invasive lumbar puncture, limiting scalability. Peripheral blood biomarkers are attractive due to ease of collection and potential for repeated monitoring. Although serum neurofilament light chain (NFL) correlates with progression, it is not disease-specific. Emerging evidence points to early peripheral inflammatory processes in PD and in isolated REM sleep behaviour disorder (iRBD), a prodromal stage within the neuronal α-synuclein disease (NSD) spectrum. The study's objective was to discover and validate blood-based proteomic biomarkers that identify PD and predict progression from prodromal iRBD years before motor onset, enabling early detection and stratification for prevention trials.

The authors highlight key prior findings: CSF α-synuclein SAA is the most specific indicator for NSD in prodromal stages (sensitivity ~74%, specificity ~93%) but is invasive and not yet robust in blood. Peripheral inflammatory markers (e.g., C-reactive protein, IL-6) and α-synuclein-specific T cells associate with motor and cognitive decline in PD and are detectable in iRBD. NFL in serum increases with progression but lacks disease specificity. Prior multiplex antibody/aptamer technologies have yielded inconsistent proteomic findings due to panel selection and cross-reactivity biases. Mass spectrometry-based proteomics provides an unbiased approach more likely to reveal disease pathways. The NSD framework integrates iRBD, PD, and DLB into a staged biological spectrum, motivating identification of early blood biomarkers for staging and risk stratification.

Study design included three phases:

• Phase 0 (Discovery): Untargeted, bottom-up proteomics of plasma from de novo PD (n=10) and matched healthy controls (HC; n=10). Samples were depleted of abundant proteins, fractionated by 2D-LC, and analyzed by label-free QTOF MSE. Proteins were identified using stringent criteria; 895 proteins passed quality filters, and 47 were differentially expressed (nominal p<0.05). Pathway analysis indicated enrichment in inflammatory pathways.
• Phase I (Targeted validation): A targeted, multiplexed LC–MS/MS assay (MRM on triple quadrupole) was developed based on discovery findings, prior PD/AD/ageing proteomics, and literature on neuroinflammation. The panel targeted 121 proteins (167 peptides). Cohorts: de novo PD (n=99), HC (n=36), other neurological diseases (OND; n=41), and iRBD (n=18). Of 121 targets, 32 proteins were consistently quantified. Statistics included Mann–Whitney U with Benjamini–Hochberg FDR=5%. Multivariate analyses used PCA and OPLS-DA (significance by ANOVA and permutation tests). Machine learning: Linear SVM with cross-validated recursive feature elimination; 70/30 train/test split with stratification, z-scoring, and handling of missing by median imputation; ROC and PR curves computed on test set; repeated cross-validation (6-fold, 40 repetitions) for robustness metrics.
• Phase II (Refinement and clinical evaluation): Refined rapid targeted LC–MS/MS method quantifying reliably detectable proteins (n=32). Independent replication cohort: longitudinal serum samples (n=146) from iRBD individuals (n=54), with up to 10 years follow-up; 16 phenoconverted to PD (n=11) or DLB (n=5). The Phase I OPLS-DA and SVM models (trained on PD vs HC) were applied to these samples. Additional analyses: linear mixed-effects models for longitudinal trends, and correlations between proteomic markers and clinical measures (MMSE, Hoehn & Yahr, UPDRS I–III, total) using Spearman with BH correction. Sample collection and processing followed standardized protocols (fasting morning draws, EDTA plasma/serum processing within 30 min, -80°C storage). Quality metrics included internal standards, digestion control (yeast ENO1), retention time alignment, and pooled QC evaluation. Plasma–serum correlation was good; correlations with CSF were limited, consistent with prior work.

• Discovery (Phase 0): Among 895 quantified proteins, 47 differed between PD and HC (nominal p<0.05). Enriched pathways included acute phase response, complement and coagulation, glucocorticoid receptor signaling, and ER stress/unfolded protein response, indicating early inflammatory signatures.
• Targeted validation (Phase I): Of 121 targeted proteins, 32 were consistently detected; 23 differed significantly between PD and HC (BH FDR 5%). PD and iRBD showed upregulation of SERPINA3, SERPINF2, SERPING1, and complement C3, and downregulation of granulin (GRN). OND and PD shared upregulation of PTGDS, CST3, VCAM1, and PLD3 vs HC. PCA separated PD and HC; iRBD clustered between them.
• OPLS-DA (Phase I): Highly significant model (p=2.3E-27; permutations p<0.001). Key contributors: GRN, DKK3, C3, SERPINA3, HPX, SERPINF2, CAPN2, SERPING1, SELE. In PD/HC classification, all samples were correctly classified; predicting iRBD yielded 72% PD-like; OND distributed between groups.
• SVM classification (Phase I): Recursive feature elimination selected 8 proteins: GRN, MASP2, HSPA5 (BiP), PTGDS, ICAM1, C3, DKK3, SERPING1. Test set accuracy was 100%. Combined panel achieved ROC AUC=1.0 and PR AUC=1.0; individual ROC AUCs ranged 0.53–0.92, PR 0.79–0.96. Repeated cross-validation yielded mean±SD: precision 0.87±0.09, recall 0.87±0.08, F1 0.86±0.09, balanced accuracy 0.82±0.12. SVM classified 94% of iRBD as PD-like (small iRBD set in Phase I).
• Independent longitudinal iRBD (Phase II): Applied to 146 serum samples from 54 iRBD subjects, OPLS-DA predicted 70% samples as PD; SVM predicted 79% as PD. Among 16 phenoconverters (11 PD, 5 DLB), earliest correct classification occurred 7.3 years before diagnosis; latest 0.9 years (mean 3.5±2.4 years). In 27/40 individuals with multiple timepoints, all samples were consistently predicted PD by SVM.
• Clinical correlations (Phase I): Higher motor/cumulative severity (UPDRS II, III, total; H&Y) associated with lower GRN, DKK3, PPP3CB, SELE and higher CST3, PTGDS, C3, HSPA5, HSPA1L, SERPINA3, SERPINF2, SERPING1, HPX. MMSE correlated negatively with C3 and SERPINs and PTGDS, and positively with lower severity markers. Detailed p values in Table 2.
• Longitudinal trends (Phase II): BCHE decreased significantly over time in iRBD (p=0.01); in SVM-predicted PD-like iRBD with ≥2 timepoints, BCHE decline remained significant (p=0.01); TUBA4A nominally increased (p=0.04, not FDR-significant). H&Y and UPDRS subscores increased over time (post-BH significant). Cholesterol associated with time and positively correlated with several proteins (e.g., HSPA8, APOE, MASP2).

The study demonstrates that a blood-based multiplex proteomic signature reflects early pathophysiological events in PD, including heightened inflammation (complement activation; elevated SERPINs), ER stress/unfolded protein response (elevated HSPA5/HSPA1L), and downregulation of Wnt signaling components (DKK3, PPP3CB), alongside reduced progranulin (GRN) suggesting diminished neuroprotection. These molecular alterations distinguish de novo PD from controls with high accuracy and are detectable in prodromal iRBD years prior to clinical diagnosis, addressing the clinical need for non-invasive, scalable early detection tools. Complement C3 and SERPINs correlated with motor severity and cognitive decline, linking peripheral inflammatory activation to clinical phenotype. The refined eight-protein panel enables robust classification and may facilitate population screening and stratification for neuroprotective trials. While CSF SAA remains highly specific, its invasiveness limits widespread use; a peripheral blood panel can complement CSF-based diagnostics and help identify high-risk individuals earlier. Mass spectrometry-based targeted assays on triple quadrupole platforms are clinically implementable and upgradable, supporting translation.

This work establishes a targeted blood-based proteomic panel combined with machine learning that accurately distinguishes de novo PD from controls and identifies prodromal iRBD individuals up to 7 years before motor/cognitive symptom onset. The eight-biomarker panel (GRN, MASP2, HSPA5, PTGDS, ICAM1, C3, DKK3, SERPING1) captures a biologically meaningful signature of inflammatory activation, ER stress, and impaired Wnt signaling, many markers correlating with symptom severity. The approach is scalable on widely available LC–MS/MS platforms and can aid subject selection/stratification for prevention trials. Future directions include: validation in independent and earlier at-risk cohorts (e.g., hyposmia), refinement of the panel for sensitivity/specificity and technical performance, integration with established progression markers (e.g., serum NFL, DAT imaging), and expansion to distinguish among NSD syndromes (PD, DLB, MSA) and to develop progression biomarkers as trial endpoints.

• Not all iRBD phenoconverters were predicted as PD-like; proteomic patterns may change over time and prediction may depend on sampling relative to conversion.
• Plasma/serum proteins correlated poorly with CSF protein levels, limiting direct inference about central proteomic changes.
• Discovery-to-validation discrepancies (common in proteomics) reduced the number of reliably quantifiable targets (32/121), necessitating panel refinement.
• Heterogeneity within OND group limited disease-specific interpretations; some OND samples were classified as PD-like.
• Longitudinal biomarker levels did not uniformly change post-classification; the panel served more as a state/risk classifier than a dynamic progression marker in this study.
• Sample sizes in discovery and certain sub-analyses were modest; broader multi-center validation is needed.