Metabolomics-based diet assessment and diet-specific biomarker metabolites identification are becoming ubiquitous. Existing studies offer a limited understanding of the underlying biochemical dynamics due to a lack of information on the holistic metabolic system changing the metabolite concentrations. Moreover, small cohort sizes of feeding trials inhibit the applicability of automated representation learning-based empirical performance improvement. In this work, we integrate prior knowledge of the human metabolic system, specifically from a genome-scale metabolic model, with metabolomic concentrations to draw novel insights into diet-related metabolism and improve dietary intake assessment. We propose multiple feature design approaches utilizing such integration - including the construction and analysis of a heterogeneous knowledge network. The proposed features offer novel hypotheses for a deeper understanding of the underlying diet-specific metabolism - such as prospective metabolic reactions and metabolic subsystems involved in the biomechanism. Our proposed features also often exceed or match baseline empirical performances of diet assessment, when used alone or together with metabolite concentrations.