Aim: To address the limitations of current metabolomics analysis, including the neglect of metabolite interdependencies, poor interpretability of black-box models, and incomplete utilization of biological information due to pathway annotation limitations. This study aims to develop and validate a dual-branch biologically informed neural network (Dual-BINN) that integrates metabolic pathway hierarchy and molecular structural hierarchy for improved prediction and interpretability in metabolomics. Methods: We developed a Dual-BINN, which explicitly incorporates metabolic pathway hierarchy and molecular structural category hierarchy into the model architecture. Pathway and structure subnetworks were constructed and integrated via an adaptive fusion mechanism. SHAP was employed for interpretability analysis. The model was evaluated using multi-center plasma metabolomics data for gastric cancer and a breast cancer dataset. Results: On the independent gastric cancer test set, Dual-BINN achieved a recall of 0.937, which compares favorably with the recall of 0.905 reported by the 10-DM model on the same dataset split. Key metabolites were enriched in the tricarboxylic acid cycle, one-carbon metabolism, and energy metabolism pathways, while structurally concentrated in organic acids, amino acids, and nucleoside-related compounds. The model also demonstrated excellent classification performance on the breast cancer dataset, confirming strong cross-disease generalization ability. Conclusions: The proposed framework enhances predictive performance while providing biologically meaningful structured interpretations, offering a robust computational approach for metabolomic mechanism analysis and biomarker discovery in complex diseases.
Guo et al. (Sun,) studied this question.