Plasma Metabolite Signature Classifies Male LRRK2 Parkinson's Disease Patients

Abstract

Parkinson's disease (PD) is a progressive neurodegenerative disease, causing loss of motor and nonmotor function. Diagnosis is based on clinical symptoms that do not develop until late in the disease progression, at which point the majority of the patients' dopaminergic neurons are already destroyed. While many PD cases are idiopathic, hereditable genetic risks have been identified, including mutations in the gene for LRRK2, a multidomain kinase with roles in autophagy, mitochondrial function, transcription, molecular structural integrity, the endo-lysosomal system, and the immune response. A definitive PD diagnosis can only be made post-mortem, and no noninvasive or blood-based disease biomarkers are currently available. Alterations in metabolites have been identified in PD patients, suggesting that metabolomics may hold promise for PD diagnostic tools. In this study, we sought to identify metabolic markers of PD in plasma. Using a ¹H-¹³C heteronuclear single quantum coherence spectroscopy (HSQC) NMR spectroscopy metabolomics platform coupled with machine learning (ML), we measured plasma metabolites from approximately age/sex-matched PD patients with G2019S LRRK2 mutations and non-PD controls. Based on the differential level of known and unknown metabolites, we were able to build a ML model and develop a Biomarker of Response (BoR) score, which classified male LRRK2 PD patients with 79.7% accuracy, 81.3% sensitivity, and 78.6% specificity. The high accuracy of the BoR score suggests that the metabolomics/ML workflow described here could be further utilized in the development of a confirmatory diagnostic for PD in larger patient cohorts. A diagnostic assay for PD will aid clinicians and their patients to quickly move toward a definitive diagnosis, and ultimately empower future clinical trials and treatment options.

Keywords: Parkinson’s disease; biomarker; leucine; machine learning; metabolite.

Figure 1

Differential metabolites correlated with male LRRK2 PD. (A) Volcano plot of the metabolite resonances extracted from the plasma of male LRRK2 PD patients and non-PD controls, with orange dots indicating significantly different resonances (p < 0.05). (B) Violin plots of the significant resonances identified in (A). The resonance at 3.77 and 74.26 ppm in the ¹H and ¹³C dimensions indicated by the green box passed the FDR p-value of <0.05.

Figure 2

Dysregulated Male LRRK2 PD Metabolic Pathways. (A) Altered metabolites in male LRRK2 PD are enriched for amino acids and metabolites associated with glycolysis and gluconeogenesis. (B) Altered metabolites glucose, alanine, and lactate are central components of the glucose-alanine and Cori cycle important for supporting gluconeogenesis in the liver. Detected differential metabolites are highlighted in orange boxes. Double arrows represent multiple enzymatic transformational steps. Abbreviations: G6P; Glucose-6-phosphate, F6P; Fructose-6-phosphate, F16BP; Fructose 1,6-biphosphate, DHAP; dihydroxyacetone phosphate, GA3P; Glyceraldehyde-3-phosphate, 13BP; 1,3-Biphosphoglycerate, 3PG; 3-Phosphoglycerate, 2PG; 2-Phosphoglycerate, PEP; Phosphoenolpyruvate.

Figure 3

Classifier workflow. Following sample preparation and NMR filtering, raw spectra were processed, resonance peaks assigned to the database-matching metabolites, and the resonances across samples normalized. Modeling comprised feature selection and 10× cross-validation of OPLS-DA and RF methods. The model with the highest stability and overall performance was selected as the champion model.

Figure 4

Cross-validation performance of champion machine learning algorithm for LRRK2 PD classification. The data with KW-only and “Known + Unknown” features was split 70/30 into training and test data, 10× cross-validated using an OPLS-DA model.

Figure 5

Olaris BoR score has high accuracy to classify male LRRK2 PD. (A) Waterfall plot of BoR score for all data with predicted cases scoring over 0.5 and predicted controls scoring under 0.5. True cases are colored orange and true controls are colored blue. (B) Receiver operator characteristic (ROC) analysis with 0.830 area under the curve (AUC). Red lines mark corresponding 81.3% sensitivity and 78.6% specificity using a score cutoff at 0.5.