Abstract
Metagenomics and metabolomics technologies have been widely used to investigate the microbe-metabolite interactions in vivo. However, the computational methods that accurately infer the microbe-metabolite interactions are lacking. We present a context-aware framework for graph representation learning, NetNiche, which predicts microbe-metabolite and microbe-microbe interactions in an accurate manner, by integrating their abundance data with prior knowledge. We applied NetNiche to datasets on gut and soil microbiome, and demonstrated that NetNiche can outperform the state-of-the-art methods, such as SParse InversE Covariance Estimation for Ecological Association Inference (SPIEC-EASI), Sparse Correlations for Compositional data (SparCC) and microbe-metabolite vectors (mmvec). NetNiche is an effective tool with wide applicability for the multi-omics study of human microbiome. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s43657-024-00168-8.