MvDE: Integrating Multi-View Similarities and Deep Ensemble Learning for Metabolite-Neurodegenerative Disease Association Prediction

Abstract

Identifying Metabolite–Disease Associations (MDAs) is essential for understanding disease mechanisms and discovering biomarkers. However, traditional experimental validation is costly and incomplete, necessitating advanced computational approaches to infer potential associations from biological data. We propose MvDE, a multi-view deep ensemble framework for MDA prediction. MvDE integrates multi-view similarity networks (semantic, structural, and interaction profiles) via local diffusion to enhance signal consistency. It utilizes a heterogeneous feature extraction module combining autoencoders, graph attention networks, and weighted non-negative matrix factorization to learn complementary representations. Finally, an ensemble classifier combining AdaBoost-CNN and LightGBM is employed to predict associations. Evaluated on a human metabolite–disease dataset containing 4,536 associations, MvDE achieved an Area Under the Precision-Recall curve (AUPR) of 0.9768 and an Area Under the Curve (AUC) of 0.9794 under five-fold cross-validation. The model consistently outperformed representative baselines, particularly in class-imbalanced settings, demonstrating superior stability and accuracy. The integration of multi-view similarities and ensemble learning allows MvDE to capture complex, nonlinear relationships while mitigating noise. Case studies on six neurodegenerative diseases, including Alzheimer’s and Parkinson’s, showed that many top-ranked novel predictions are supported by existing biological literature. MvDE provides a robust and scalable tool for MDA prediction. By effectively prioritizing candidate metabolites, it facilitates the discovery of diagnostic markers and therapeutic targets for neurodegenerative diseases.