Parkinson's disease (PD) is a neurodegenerative disorder characterized by motor symptoms resulting from the loss of dopamine-producing neurons in the brain. Currently, there is no cure for the disease which is in part due to the heterogeneity in patient symptoms, trajectories and manifestations. There is a known genetic component of PD and genomic datasets have helped to uncover some aspects of the disease. Understanding the longitudinal variability of PD is essential as it has been theorised that there are different triggers and underlying disease mechanisms at different points during disease progression. In this paper, we perform longitudinal and cross-sectional experiments to identify which data modalities or combinations of modalities are informative at different time points. We use clinical, genomic, and proteomic data from the Parkinson's Progression Markers Initiative. We validate the importance of flexible data integration by highlighting the varying combinations of data modalities for optimal stratification at different disease stages in idiopathic PD. We show there is a shared signal in the DNAm signatures of participants with a mutation in a causal gene of PD and participants with idiopathic PD. We also show that integration of SNPs and DNAm data modalities has potential for use as an early diagnostic tool for individuals with a genetic cause of PD.

MOTIVATION: Heterogeneity in human diseases presents challenges in diagnosis and treatments due to the broad range of manifestations and symptoms. With the rapid development of labelled multi-omic data, integrative machine learning methods have achieved breakthroughs in treatments by redefining these diseases at a more granular level. These approaches often have limitations in scalability, oversimplification, and handling of missing data. RESULTS: In this study, we introduce Multi-Omic Graph Diagnosis (MOGDx), a flexible command line tool for the integration of multi-omic data to perform classification tasks for heterogeneous diseases. MOGDx has a network taxonomy. It fuses patient similarity networks, augments this integrated network with a reduced vector representation of genomic data and performs classification using a graph convolutional network. MOGDx was evaluated on three datasets from the cancer genome atlas for breast invasive carcinoma, kidney cancer, and low grade glioma. MOGDx demonstrated state-of-the-art performance and an ability to identify relevant multi-omic markers in each task. It integrated more genomic measures with greater patient coverage compared to other network integrative methods. Overall, MOGDx is a promising tool for integrating multi-omic data, classifying heterogeneous diseases, and aiding interpretation of genomic marker data. AVAILABILITY AND IMPLEMENTATION: MOGDx source code is available from https://github.com/biomedicalinformaticsgroup/MOGDx.