Eleven-Protein Plasma Panel Identifies Parkinson’s Disease Across Four Cohorts

The Search for Objective Biomarkers in Neurodegeneration

The clinical management of neurodegenerative disorders remains hindered by the lack of objective, minimally invasive diagnostic tools that can reliably distinguish specific pathologies from overlapping neurological presentations. While the etiology of these conditions is increasingly linked to complex interactions between neuroinflammation, protein aggregation, and systemic metabolic dysfunction, definitive diagnosis often requires longitudinal observation or expensive imaging [1, 2]. Recent evidence suggests that peripheral immune signaling and the microbiota-gut-brain axis (the bidirectional communication network between the gastrointestinal tract and the central nervous system) play significant roles in the progression of Parkinson’s and Alzheimer’s diseases [3, 4]. Furthermore, the emergence of reactive astrogliosis (the morphological and functional changes in astrocytes in response to injury) and microglial activation as early pathological hallmarks suggests that glial-derived proteins may serve as viable candidates for liquid biopsies [5, 6]. A recent study addressed this diagnostic gap by using high-throughput proteomics to identify an 11-protein plasma panel, including DDC, CXCL8, and APOH, which identified Parkinson’s disease with an area under the receiver operating characteristic curve of 0.939 in a cohort of 698 participants [7]. These findings suggest that systemic protein signatures can provide the specificity required to differentiate between complex neurodegenerative phenotypes in a clinical setting.

Proteomic Profiling and Model Architecture

To identify these systemic signatures, researchers utilized the Olink Explore 3072 assay to obtain comprehensive plasma proteomic profiles from a total of 698 participants. This high-throughput technology allows for the simultaneous measurement of thousands of proteins, providing a wide-angle view of the circulating proteome. To ensure the findings were specific to Parkinson’s disease rather than general neurological dysfunction, the total cohort included 149 Parkinson’s disease cases, 230 neurologically healthy controls, and 319 participants with other neurological conditions. This diverse control group is essential for clinical utility, as it reflects the diagnostic challenges physicians face when distinguishing between various movement disorders and neurodegenerative states. The study population was divided into a Training Set of 560 individuals and a Test Set of 138 individuals. Within the Training Set, the researchers analyzed data from 118 Parkinson’s patients, 184 healthy controls, and 258 individuals with other neurological disorders.

To narrow down the most relevant biomarkers from the thousands of proteins measured, the team applied the Boruta algorithm (a feature selection method that identifies all relevant variables in a dataset by comparing their importance against shuffled copies of the data). This rigorous selection process ensured that only the most robust predictors were included in the final diagnostic framework. Using the variables identified, the researchers developed a stacking ensemble machine learning model (a technique that combines multiple different algorithms to improve predictive performance by leveraging the strengths of each individual model). This computational architecture was trained on the Training Set to recognize a specific signature consisting of eleven proteins: APOH, ARG1, CCN1, CXCL1, CXCL8, DDC, GRAP2, IL1RAP, OSM, PRL, and SPRY2. By integrating these diverse markers, which include mediators of inflammation and lipid metabolism, the model provides a multi-dimensional biological snapshot that facilitates the differentiation of Parkinson’s disease from both healthy states and symptomatic mimics.

Diagnostic Performance and External Validation

The stacking ensemble model demonstrated high diagnostic accuracy when evaluated against the internal held-out Test Set of 138 participants, achieving an area under the receiver operating characteristic curve (AUROC) of 0.939. This metric, which represents the model’s ability to correctly distinguish between patients with Parkinson’s disease and those without the condition, suggests that the eleven-protein signature is highly sensitive and specific within the primary study population. To ensure these findings were not limited to a single dataset, the researchers conducted extensive external validation across three geographically and demographically diverse cohorts. This process is critical for clinicians, as it tests whether a diagnostic tool maintains its predictive power when applied to new patients in different clinical settings. Validation in the UK Biobank external cohort, which included a large sample of 43,969 individuals, yielded an AUROC of 0.789. While lower than the internal test set, this result remains significant given the massive scale and heterogeneity of the population.

The model’s performance was further confirmed in more targeted clinical datasets. In the Parkinson’s Disease Biomarkers Program external cohort of 138 participants, the model achieved an AUROC of 0.909. Similarly, evaluation within the Parkinson’s Progression Markers Initiative external cohort of 385 participants resulted in an AUROC of 0.816. The consistency of these findings across multiple independent cohorts underscores the high specificity and reliability of the proteomic model. For the practicing physician, the ability to maintain an AUROC between 0.789 and 0.909 in external populations suggests that these plasma biomarkers could eventually provide a standardized, objective supplement to the current clinical diagnosis of Parkinson’s disease. By successfully differentiating Parkinson’s cases from both healthy controls and other neurological mimics in diverse groups, the model addresses a major hurdle in neurodegenerative diagnostics: the need for a reliable biological signature that persists beyond the initial discovery group.

Biological Mechanisms and Clinical Interpretation

To ensure the diagnostic model provided biologically meaningful insights rather than acting as an opaque mathematical tool, the researchers utilized the Shapley Additive Explanations (SHAP) framework (a method used to explain the output of machine learning models by quantifying the contribution of each feature to the final classification). This was combined with network analysis to evaluate the biological relevance of each protein in the model. By conducting differential protein abundance analysis and pathway enrichment analysis, the authors were able to map the eleven proteins (APOH, ARG1, CCN1, CXCL1, CXCL8, DDC, GRAP2, IL1RAP, OSM, PRL, and SPRY2) to specific physiological processes. This rigorous approach allowed the team to verify that the model's predictive power was derived from established pathological mechanisms associated with neurodegeneration. The integration of network and pathway analyses revealed that the protein signature is closely linked to systemic activity involving inflammatory mediators and T-cell receptor signaling.

These findings suggest that the plasma proteomic profile captures the peripheral immune dysregulation often observed in Parkinson’s disease. Furthermore, the analyses identified the involvement of ErbB signaling (a pathway involved in cell growth and differentiation) as well as significant alterations in lipid metabolism. For the clinician, these results indicate that the eleven-protein panel reflects a broad spectrum of systemic pathology, extending the biological understanding of the disease beyond localized dopaminergic loss in the substantia nigra. By identifying these specific pathways, the study helps to elucidate potential novel risk factors and biological mechanisms that may drive disease progression. The model's reliance on proteins involved in ErbB signaling and lipid metabolism provides a more comprehensive diagnostic picture that could eventually assist in the differential diagnosis of Parkinson’s disease from other neurological conditions with overlapping motor symptoms. This shift toward a multi-pathway proteomic signature offers a more nuanced clinical interpretation of the disease state, potentially allowing for earlier identification of the systemic changes that precede or accompany traditional clinical markers.

References

1. Zhang W, Xiao D, Mao Q, Xia H. Role of neuroinflammation in neurodegeneration development. Signal Transduction and Targeted Therapy. 2023. doi:10.1038/s41392-023-01486-5

2. Chen L, Deng H, Cui H, et al. Inflammatory responses and inflammation-associated diseases in organs. Oncotarget. 2017. doi:10.18632/oncotarget.23208

3. Cryan JF, O’Riordan KJ, Cowan CS, et al. The Microbiota-Gut-Brain Axis. Physiological Reviews. 2019. doi:10.1152/physrev.00018.2018

4. Loh JS, Mak WQ, Tan L, et al. Microbiota–gut–brain axis and its therapeutic applications in neurodegenerative diseases. Signal Transduction and Targeted Therapy. 2024. doi:10.1038/s41392-024-01743-1

5. Gao C, Jiang J, Tan Y, Chen S. Microglia in neurodegenerative diseases: mechanism and potential therapeutic targets. Signal Transduction and Targeted Therapy. 2023. doi:10.1038/s41392-023-01588-0

6. Sofroniew MV, Vinters HV. Astrocytes: biology and pathology. Acta Neuropathologica. 2009. doi:10.1007/s00401-009-0619-8

7. Adewale B, Chia R, Moaddel R, et al. Machine learning model based on plasma proteomics for the identification of Parkinson's disease.. Brain : a journal of neurology. 2026. doi:10.1093/brain/awag140