DMSG-VBM: Deep Multi-Source Graph-Enhanced Variational Bayesian Model for Predicting Higher-Order Microbe-Drug-Disease Associations

Abstract

Exploring potential relationships among drugs, microbes, and diseases is critical for advancing disease prevention and drug development. Several computational models have been developed to predict high-order interactions in multi-omics biological systems. However, negative sampling-based methods risk generating false negatives, whereas tensor decomposition approaches typically require extensive hyperparameter tuning and exhibit limited flexibility when handling complex data architectures. To address these limitations, we propose DMSG-VBM, a Bayesian Logistic Tensor Decomposition model incorporating multi-source graph deep learning. The model employs multi-source graph deep learning to generate prior expectations of latent variables while establishing an attention mechanism to facilitate the integration of multi-source features. Furthermore, DMSG-VBM integrates tensor-based Bayesian inference with multi-source graph deep learning within a probabilistic graphical model and adopts a variational expectation-maximization algorithm for adaptive inference of model parameters and latent variables. Experimental results across four scenarios on two benchmark datasets demonstrate that DMSG-VBM achieves superior performance in triplet prediction compared to other state-of-the-art methods, for both balanced and extremely imbalanced scenarios. Additionally, it exhibits higher hit rates in predictive tasks involving novel drugs, novel microbes, and novel diseases. Case studies further substantiate that DMSG-VBM effectively predicts potential associations among drugs, microbes, and diseases. The codes and dataset are available at: https://github.com/Mayingjun20179/DMSG-VBM.