A Novel Combined Embedding Model Based on Heterogeneous Network for Inferring Microbe-Metabolite Interactions

The metabolome serves as a crucial intermediary between the microbiome and its host, playing a key role in revealing their biological functions. Previous research has established connections between various microbiomes and metabolomes through correlation and association analyses. Although traditional statistical analysis methods have been used to quantify microbe-metabolite correlations, they do not fully elucidate the biological connections between these associated pairs. In recent years, some models based on networks have been proposed to predict microbe-metabolite interactions. However, relying solely on microbial abundance to reconstruct metabolomic profiles may overlook the complex and interactive synergistic relationships within the microbiome and metabolome. In this study, a novel Combined Embedding Model based on Heterogeneous Network (CEM_HN) is proposed for inferring microbe-metabolite interactions. First, we build a heterogeneous network, which consists of microbe-metabolite pairs, microbial internal interaction network, and metabolite internal interaction network. Then, we utilize paired embeddings obtained from an autoencoder to extract fine-grained pairwise information in microbe-metabolite pairs. This autoencoder helps capture the hidden biological associations between nodes. Finally, by fusing the node embeddings with paired embeddings, a combined embedding is obtained to infer microbe-metabolite interactions. The experimental results demonstrate that the proposed method has strong performance and biomedical interpretability in predicting microbe-metabolite interactions.