Summary: Researchers have created an AI-based tool that can detect biochemical signs of Parkinson’s disease in blood samples years—up to 15 years—before clinical symptoms appear. Using machine learning to analyze complex patterns of metabolites, the system identifies combinations of molecules that serve as early biomarkers. Although larger validation studies are required, this approach reached up to 96% accuracy in the study cohort and is publicly available for researchers working with metabolomics data.
Key Facts:
- The CRANK-MS tool, developed by researchers at UNSW Sydney with collaborators at Boston University, analyzes blood metabolite profiles to detect signals associated with future Parkinson’s disease up to 15 years before diagnosis.
- In the published study cohort, the model achieved prediction accuracy up to 96%, uncovering distinctive metabolite patterns that may act as early warning markers.
- CRANK-MS is freely available to researchers and can be applied to metabolomics datasets for other diseases, offering a practical route to discover and validate new biomarkers.
Source: University of New South Wales
UNSW Sydney researchers and collaborators at Boston University have developed a promising AI tool for early detection of Parkinson’s disease.
Published in ACS Central Science, the study describes how a neural-network-based approach was trained to recognize metabolomic signatures in blood samples collected years before Parkinson’s symptoms emerged. The work focuses on subtle combinations of metabolites—small molecules produced by the body during metabolism—that are difficult to interpret with conventional statistical methods.
The team used blood samples from the Spanish cohort of the European Prospective Investigation into Cancer and Nutrition (EPIC). They examined samples from 39 participants who later developed Parkinson’s disease, comparing them with 39 matched controls who did not develop the condition. By applying their machine learning pipeline to the full set of metabolite features, the researchers identified combinations of molecular signals associated with future Parkinson’s risk.
UNSW researcher Diana Zhang and Associate Professor W. Alexander Donald built the machine learning platform CRANK-MS (Classification and Ranking Analysis using Neural network generates Knowledge from Mass Spectrometry). Unlike standard workflows that first reduce data dimensionality, CRANK-MS ingests the complete, unfiltered dataset and discovers which features and feature combinations best predict disease risk.
“Typical metabolomics studies test correlations between individual molecules and disease,” Zhang explains. “Our approach lets the model learn interdependencies among hundreds to thousands of metabolites, revealing patterns that would be hard to detect using traditional statistics.”
Associate Professor Donald adds that processing the complete dataset without preliminary feature elimination enables the model to uncover potentially important metabolites that might otherwise be discarded. The platform returns both predictions and an interpretable ranking of the chemical features that most strongly influence the model’s decisions.
Why this matters for Parkinson’s diagnosis
Parkinson’s disease is typically diagnosed based on motor symptoms such as tremor, rigidity, and bradykinesia. However, non-motor signs—sleep disorders, loss of smell, mood changes, and apathy—can appear years earlier. A reliable blood-based test that flags elevated risk could allow clinicians to monitor at-risk individuals earlier, begin interventions sooner, and stratify patients for preventive clinical trials.
The authors caution that these findings are preliminary and require replication in larger, geographically diverse cohorts before CRANK-MS could be used clinically. In the current, limited dataset, CRANK-MS achieved predictive accuracy as high as 96%, and it identified a small set of chemical markers that contributed most to accurate prediction.
“The results are encouraging on several fronts,” says A/Prof. Donald. “The predictive performance is high in advance of clinical diagnosis. The approach highlights biochemical markers linked to future disease. And several of the chemicals that drive our predictions have previously shown effects in laboratory studies, even if they had not yet been validated in human cohorts.”
Notable biochemical observations
Among the metabolite differences observed, the study notes lower circulating levels of triterpenoids in those who later developed Parkinson’s. Triterpenoids are plant-derived compounds with antioxidant and neuroprotective properties found in foods such as apples, olives, and tomatoes. The results raise the hypothesis—requiring careful testing—that dietary factors affecting triterpenoid intake might influence Parkinson’s risk.
The analysis also flagged signals consistent with polyfluorinated alkyl substances (PFAS) in participants who later developed Parkinson’s, suggesting a possible link to environmental chemical exposure. The authors stress that further chemical characterization is necessary to confirm PFAS identity and to explore causality.
Open access tool for researchers
CRANK-MS is openly available for researchers who wish to apply machine learning to metabolomics datasets. The developers designed the tool to be user-friendly and computationally efficient; Zhang notes that, on average, analyses can be completed in under ten minutes on a standard laptop.
“Applying CRANK-MS to Parkinson’s is just one example of how AI can advance biomarker discovery and disease monitoring,” Zhang says. “Because the tool works directly from mass spectrometry data, it can be adapted to many diseases where metabolomics provides useful insight.”
About this AI and Parkinson’s disease research news
Author: Lachlan Gilbert
Source: University of New South Wales
Contact: Lachlan Gilbert – University of New South Wales
Image: The image is credited to Neuroscience News
Original Research: The findings will appear in ACS Central Science