M$$\vphantom{0}^2$$DGAT: Multi-view multi-scale dynamic graph attention network(GAT) based prediction of Parkinson’s disease(PD) progression using whole-blood RNA sequencing data

With emerging single-cell transcriptomics data, deep learning approaches have enabled the diagnosis of neurodegenerative disorders such as Parkinson's disease(PD). Based on whole-blood RNA sequencing data, the graph neural network has predicted Parkinson's disease(PD) progression trajectories by exploiting spatial and temporal views that were related with neurodegenerative disorders. The single-view learning scheme, which focuses on gene expression embedding, was still insufficient in analyzing complicated human brain diseases. As a type of spatial representation underlying transcriptomics data, gene graphs that are closely associated with disease states have been inferred to investigate regulatory mechanisms and molecular dynamics. For disease-specific gene graphs, global and local structures contribute to fine-grained spatial representations. With the dynamic graph attention network (DGAT) backbone, this study proposes a multi-view, multi-scale M[Formula: see text]DGAT method to predict disease progression trajectories for neurodegenerative disorders. Temporal and spatial views have been integrated by the count sketch bilinear (CSB) fusion strategy. Based on RNA sequencing data from human blood samples, joint-view representations have been constructed to predict disease stages and relevant cognitive scores. Experiments about the PPMI and PDBP cohorts have validated the effectiveness and efficiency of the proposed M[Formula: see text]DGAT architecture in disease prediction applications. The proposed M[Formula: see text]DGAT method has demonstrated significant superiority in predictive accuracy over established cutting-edge disease prediction approaches. Compared with static graphs, dynamic graph representations tend to encode more dynamics about disease progression.