Overview of the integrated deep learning framework. (A) Illustration of data source, which includes analyzing LC–MS data to generate data features (aligned peaks across samples), and mapping features to known metabolites. (B) Integrated information including feature abundance matrix along with clinical outcome, the known metabolic network structure and potential feature matching to metabolites. (C) A layer-by-layer gradual sparse neuron network, which takes feature expression data as input and samples classes as output. It comprises three parts: feature-metabolite embedding based on potential feature annotations, metabolic network embedding and layer-by-layer gradual sparsification, and fully connected layers. (D) Results of a trained model. The model can make classification, determine likely feature-metabolite matching, conduct metabolite and metabolic sub-network selection for functional analysis.