A Graph Convolutional Network Framework Integrating Nuclear Norm Minimization and Conditional Random Fields for Microbe–Disease Association Prediction

Abstract

Microbiota dysbiosis is closely associated with a wide range of human diseases, yet wet-lab validation remains costly and time-consuming. Therefore, this study aims to develop an efficient framework for predicting potential microbe–disease associations. We propose a novel predictive model, NNGCFCAE, which integrates Nuclear Norm Minimization (NNM), an enhanced Graph Convolutional Network (GCF) with an energy-based Conditional Random Field (CRF) smoothing mechanism, and a multi-channel convolutional autoencoder (CAE) with residual connections to effectively infer latent microbe–disease associations. First, we construct a heterogeneous network by integrating known microbe–disease associations with Gaussian Interaction Profile (GIP) kernels and Hamming Interaction Profile (HIP) features. Subsequently, nuclear norm minimization is applied to complete the initial association matrix, yielding a preliminary prediction score matrix. The GCF module then extracts spatial structural features of microbe and disease nodes from the network, while the CAE module further learns attribute-based representations of these nodes. Finally, the prediction score matrix, topological features, attribute features, and multi-source information are fused to form a joint representation matrix, which is used to compute the final association scores between microbes and diseases. Experiments conducted on datasets such as HMDAD and Disbiome demonstrate that NNGCFCAE significantly outperforms several state-of-the-art methods in terms of AUC and AUPR. Ablation studies and case analyses on obesity, asthma, and ulcerative colitis further demonstrate its biological plausibility, highlighting its potential for uncovering latent microbe–disease associations.