Interpretable Machine Learning on Metabolomics Data Reveals Biomarkers for Parkinson’s Disease

Abstract

The use of machine learning (ML) with metabolomics provides opportunities for the early diagnosis of disease. However, the accuracy of ML and extent of information obtained from metabolomics can be limited owing to challenges associated with interpreting disease prediction models and analyzing many chemical features with abundances that are correlated and “noisy”. Here, we report an interpretable neural network (NN) framework to accurately predict disease and identify significant biomarkers using whole metabolomics data sets without a priori feature selection. The performance of the NN approach for predicting Parkinson’s disease (PD) from blood plasma metabolomics data is significantly higher than other ML methods with a mean area under the curve of >0.995. PD-specific markers that predate clinical PD diagnosis and contribute significantly to early disease prediction were identified including an exogenous polyfluoroalkyl substance. It is anticipated that this accurate and interpretable NN-based approach can improve diagnostic performance for many diseases using metabolomics and other untargeted ‘omics methods.

Prediction of motor and non-motor Parkinson’s disease symptoms using serum lipidomics and machine learning: a 2-year study

“…Binding of platelet-activating factor to the platelet-activating factor receptor can induce diacylglycerol production [35][36][37][38] . Serum diacylglycerol 16:0_18:3 (DG 34:3) was the #7 contributor to Schwab and England Activities of Daily Living Scale score, which is consistent with a recent metabolomics study which found plasma DG 34:3 was a top contributor to PD incidence prediction 27 . This study also found DG 16:0_18:3 (DG 34:3) contributed to other scales (UPDRS III, UPSIT, Geriatric Depression Scale and Hoehn and Yahr) but to a lesser extent (#62 -#78 on variable importance lists).…”

“…Despite the evidence indicating that lipids may play a role in PD, the predictive value that baseline blood lipids may have for both motor and nonmotor scales has not been explored longitudinally, although plasma lipids estimating cognitive and motor severity scores have been assessed crosssectionally 25 . Two previous studies have assessed baseline blood lipids for predicting a binary clinical outcome, namely, the future diagnosis of dementia 26 or PD 27 and have not investigated the degree to which lipids may predict motor and non-motor clinical scale scores, which are used as tools for assessing clinical trial drug effectiveness. One study of n = 43 PD subjects found that a panel of baseline serum lipids could discriminate between those that progressed to dementia and those that remained cognitively stable after 3 years 26 .…”

“…One study of n = 43 PD subjects found that a panel of baseline serum lipids could discriminate between those that progressed to dementia and those that remained cognitively stable after 3 years 26 . Another study of n = 39 each of PD and control subjects found that baseline plasma metabolites, which included lipids, contributed to the prediction of PD incidence after 15 years 27 . Given the range of motor and nonmotor symptoms in Parkinson's disease, comparatively assessing the ability of baseline serum lipids to predict future scale scores of olfaction, depression, motor function and quality of life in PD subjects could be valuable.…”

Discovering Biomarker Proteins and Peptides for Parkinson’s Disease Prognosis Prediction with Machine Learning and Interpretability Methods

“…Furthermore, the General Data Protection Regulation (GDPR) issued by the EU emphasizes the "right to explanation," which grants individuals the right to comprehend the logic behind automated decisions made by algorithms, especially in instances where these decisions significantly impact their lives [18]. In contemporary research, specifically on methods combining machine learning and CSF-MS, an increasing trend towards adopting and integrating these interpretable methods is evident, aiming to promote transparency and foster a better understanding of model behavior [14,19].…”

“…1. Regression problem-solving: Most existing studies perform classification [5,9,19,20] In this study, we perform feature selection using mutual information (MI). MI is a statistical concept measuring the degree of dependence between two random variables.…”