Machine learning analysis reveals biomarkers for the detection of neurodegenerative diseases

Abstract

AbstractIt is critical to identify biomarkers for neurodegenerative diseases (NDDs) to advance disease diagnosis and accelerate drug discovery for effective treatment of patients. In this work, we retrieved genotyping and clinical data from 1223 UK Biobank participants to identify genetic and clinical biomarkers for NDDs, including Alzheimer’s disease (AD), Parkinson’s disease (PD), motor neuron disease (MND), and myasthenia gravis (MG). Using a machine learning modelling approach and Monte Carlo randomisation, we identified 16 informative clinical variables for predicting AD, PD, MND, and MG. In a multinomial model, these clinical variables could correctly predict the diagnosis of one of the four diseases with an accuracy of 88.3%. In addition to clinical biomarkers, we also explored genetic biomarkers. In a genome-wide association study of AD, PD, MND, and MG patients, we identified single nucleotide polymorphisms (SNPs) implicated in several craniofacial disorders such as apnoea and branchiootic syndrome. We found evidence for shared genetic risk loci across NDDs, including SNPs in cancer-related genes and SNPs known to be associated with non-brain cancers such as Wilms tumour, leukaemia, and pancreatic cancer. Our analysis supports current knowledge regarding the ageing-related degeneration/cancer shift.Significance statementThis study highlights the potential for hypothesis-free mathematical modelling of easily measured clinical variables to identify diagnostic biomarkers for neurodegenerative diseases (NDDs). Prior to this study, the focus in NDD research has surrounded toxic species such as amyloid beta and α-synuclein, but this approach has not enjoyed success at clinical trial. Here, we studied Alzheimer’s disease, Parkinson’s disease, motor neuron disease, and myasthenia gravis by constructing and inspecting a multinomial based on demographics and blood and urine biochemistry. Cognitive measures were important for the predictive power of the model. Model weights correctly indicated multiple trends reported in the literature. Separately, genome-wide association indicated a shared risk profile between NDD and cancer, which has also been reported in the literature.