By 2028, your annual checkup might include a blood draw that screens for Parkinson's disease the way we currently screen for cholesterol - and the blueprint for that diagnostic was just published in Brain.

Here's the architectural embarrassment of modern neurology: we diagnose Parkinson's disease - the second most common neurodegenerative disorder, affecting 10 million people worldwide - by watching someone walk across a room. A neurologist evaluates your tremor, your gait, your facial expressions, and renders a judgment call that's accurate roughly 75% of the time in early stages. By the time those motor symptoms show up for inspection, somewhere between 50% and 80% of your dopaminergic neurons have already packed up and left. That's like calling in a building inspector after the east wing has collapsed and asking, "So, do you think there might be a structural issue?"

A team led by Boluwatife Adewale, Sonja Scholz, and colleagues across NIH, Johns Hopkins, and the University of Turin decided to stop squinting at the facade and start examining the load-bearing walls instead - specifically, 3,000 proteins circulating in human blood plasma (Adewale et al., 2025).

Surveying 3,000 Columns to Find the 11 That Matter

Using the Olink Explore 3072 assay - a platform that can measure roughly 3,000 proteins from a few microliters of plasma, which is basically a mass spectrometer's fever dream - the researchers profiled blood samples from 698 participants. That included 149 people with Parkinson's, 230 healthy controls, and 319 people with other neurological conditions (the tricky differential diagnoses that make PD so hard to pin down clinically).

Then came the demolition phase. The Boruta algorithm, a feature selection method that works like an honest architecture competition, created randomized "shadow" copies of every protein and asked: does this real protein predict PD better than its fake twin? Out of thousands of candidates, eleven proteins survived the judging panel: APOH, ARG1, CCN1, CXCL1, CXCL8, DDC, GRAP2, IL1RAP, OSM, PRL, and SPRY2.

If the feature selection was the site survey, the stacking ensemble model was the actual construction. Multiple ML algorithms - think of them as different structural engineering firms, each with their own specialty - generated independent predictions, and a meta-learner combined them into a unified diagnosis. The architectural logic here is elegant: no single algorithm carries all the load.

The Stress Test Results Are Stunning

The model posted an AUC of 0.939 on its held-out test set. Impressive, but plenty of ML models look gorgeous in the showroom. The real question is whether the structure holds when you ship it across town.

So they tested it on three external cohorts: the UK Biobank (n = 43,969), the Parkinson's Disease Biomarkers Program (n = 138), and the Parkinson's Progression Markers Initiative (n = 385). AUCs: 0.789, 0.909, and 0.816, respectively. The UK Biobank number is lower, which makes sense - it's a general population cohort, not a clinical one. That's like testing your earthquake-resistant skyscraper design in a neighborhood that mostly experiences light breezes. The fact that it still performed respectably is the point.

Reading the Blueprints with SHAP

To make the model's decisions interpretable - because "the algorithm said so" doesn't fly in clinical settings - the team used SHAP (SHapley Additive Explanations), a framework borrowed from game theory. SHAP assigns each protein a contribution score for every individual prediction, effectively letting you see which load-bearing walls are doing the most work for each patient. DDC (dopa decarboxylase, directly involved in dopamine synthesis) and several inflammatory markers emerged as key contributors. The pathway analysis revealed activity in inflammatory signaling, ErbB signaling, T-cell receptor pathways, and lipid metabolism - a finding that echoes recent large-scale proteomic work by Gan et al. (2025), who found lipid metabolism dysfunction detectable up to 15 years before PD diagnosis.

Why This Isn't Just Another Pretty Model

This study lands at an interesting moment. Hallqvist et al. (2024) recently showed a plasma protein panel could classify PD up to seven years before motor symptoms appear. The alpha-synuclein seed amplification assay just earned an FDA letter of support for clinical trials (Kluge et al., 2024). Multiple independent groups are converging on overlapping protein panels validated across the same cohorts - UK Biobank, PPMI, PDBP. The field is no longer asking "can blood tests detect Parkinson's?" but "which combination of blood biomarkers gets us to clinical utility fastest?"

What makes the Adewale et al. model architecturally sound is its clean sight lines: transparent feature selection, interpretable predictions, and validation across four independent datasets spanning tens of thousands of participants. The cantilever still needs reinforcement - prospective clinical trials, standardized assay protocols, cost-effectiveness analysis - but the structural integrity is there.

For a disease where early intervention could mean preserving neurons instead of mourning them, a reliable blood test isn't just convenient. It's the difference between renovation and rubble.

References

1. Adewale, B., et al. (2025). Machine learning model based on plasma proteomics for the identification of Parkinson's disease. Brain. DOI: 10.1093/brain/awag140 | PMID: 42015416

2. Hallqvist, J., et al. (2024). Plasma proteomics identify biomarkers predicting Parkinson's disease up to 7 years before symptom onset. Nature Communications. DOI: 10.1038/s41467-024-48961-3 | PMID: 38890280

3. Gan, Y.-H., et al. (2025). Large-scale proteomic analyses of incident Parkinson's disease reveal new pathophysiological insights and potential biomarkers. Nature Aging, 5(4), 642-657. DOI: 10.1038/s43587-025-00818-0 | PMID: 39979637

4. Kluge, A., et al. (2024). Detecting misfolded alpha-synuclein in blood years before the diagnosis of Parkinson's disease. Movement Disorders. DOI: 10.1002/mds.29766

5. Pillai, S., et al. (2024). Machine learning analysis of population-wide plasma proteins identifies hormonal biomarkers of Parkinson's disease. NPJ Parkinson's Disease. PMCID: PMC11703317

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.