Enhancing microbe-disease association prediction via multi-view graph convolution and latent feature learning

Bo Wang; Peilong Wu; Xiaoxin Du; Chunyu Zhang; Shanshan Fu; Tang Sun; Xue Yang

doi:10.1016/j.compbiolchem.2025.108581

Enhancing microbe-disease association prediction via multi-view graph convolution and latent feature learning

Comput Biol Chem. 2025 Jun 30:119:108581. doi: 10.1016/j.compbiolchem.2025.108581. Online ahead of print.

Authors

Bo Wang¹, Peilong Wu², Xiaoxin Du², Chunyu Zhang², Shanshan Fu², Tang Sun², Xue Yang²

Affiliations

¹ School of Computer and Control Engineering, Qiqihar University, Qiqihar, Heilongjiang 161006, China; Heilongjiang Key Laboratory of Big Data Network Security Detection and Analysis, Qiqihar University, Qiqihar, Heilongjiang 161006, China. Electronic address: bowangdr@qqhru.edu.cn.
² School of Computer and Control Engineering, Qiqihar University, Qiqihar, Heilongjiang 161006, China; Heilongjiang Key Laboratory of Big Data Network Security Detection and Analysis, Qiqihar University, Qiqihar, Heilongjiang 161006, China.

PMID: 40602045
DOI: 10.1016/j.compbiolchem.2025.108581

Abstract

Microbes play a crucial role in the onset, progression, and treatment of diseases. To address the challenges of missing information and insufficient feature fusion in microbe-disease association prediction, this paper proposes an innovative computational model named MVGCVAE. MVGCVAE is the first model to synergistically integrate multi-view graph convolutional networks (GCNs), variational autoencoders (VAEs), and dynamic kernel matrix weighting for microbe-disease association (MDA) prediction. First, we construct multiple similarity networks between microbes and diseases, using GCN to independently process the node features in each view. To better fuse information from different similarity views, we introduce an attention mechanism to assign different weights to each perspective, thereby generating an initial comprehensive feature representation of diseases and microbes. This enables the model to more effectively integrate features from various perspectives and enhances its sensitivity and discriminative ability for key features. Next, based on a heterogeneous network, we feed the fused node features into the GCN for further representation learning. After each layer of feature extraction, we use a Variational Autoencoder (VAE) for variational inference to optimize node representations and enhance adaptation to sparse data and nonlinear relationships. Then, we propose a dynamic weighted kernel matrix strategy. This strategy uses a multi-layer perceptron (MLP) to adaptively generate weights, flexibly integrating kernel matrices computed from different embeddings at each layer to optimize the feature fusion process. Finally, we combine the weighted matrix with the feature matrix using matrix multiplication to calculate the microbe-disease association, and further optimize the model's predictive capability through Laplacian Regularization. Experimental results show that MVGCVAE outperforms six existing comparison methods on multiple evaluation metrics. Additionally, case studies further validate the reliability of MVGCVAE in predictive tasks.

Keywords: Dynamic kernel matrix weighting; Feature fusion; Microbe-disease associations, Graph convolutional networks (GCNs); Variational autoencoder (VAE).