Multi-view Graph Learning based Progression Modeling of Parkinson’s Disease Using Whole-blood Transcriptomics Data

Abstract

In human disease modeling, blood transcriptomics data has played a crucial role in revealing regulatory abnormality. As patterns underlying blood transcriptomics data are subtle, it is insufficient to rely on temporal expression features to predict future stages of brain diseases. Spatial features have encoded meaningful information about gene-level and cell-level interactions. Progression of neurological disorder lasts 10-15 years, indicating that disease-specific gene graphs are changing. In order to employ dynamic spatial features, this work proposes a novel dynamic spatiotemporal graph learning (DST-GNN) to conduct disease progression prediction. The DST-GNN method aims to integrate expression patterns and dynamic gene graphs. Validation experiments about benchmark whole-blood RNA-seq datasets from the AMP-PD platform have demonstrated the effectiveness and advantages of the DST-GNN.