Enhanced metabolite-disease associations prediction via Neighborhood Aggregation Graph Transformer with Kolmogorov–Arnold Networks

Abstract

Metabolites are essential products of cellular chemical reactions, critical for sustaining life and reproduction. Research shows that metabolite concentrations in patients differ from those in healthy individuals, making metabolite-based disease prediction crucial for diagnosis and treatment. To address the limitations of current computational methods in accuracy and interpretability, we propose a novel Neighborhood Aggregation Graph Transformer method (AGKphormer). This method enhances link relationships by optimizing the minimum nuclear norm using the Alternating Direction Method of Multipliers (ADMM) and incorporates Fast Kolmogorov–Arnold Networks (FastKAN) to improve both accuracy and interpretability. We first construct a heterogeneous network based on the correlation and similarity between metabolites and diseases. Then, we utilize the ADMM algorithm to enhance link relationships by solving the minimum nuclear norm, reducing sparse relationships between nodes and providing richer features for neural network learning. For the features learned by the graph convolutional network (GCN), we employ a Graph Transformer augmented with FastKAN to learn long-range dependencies. This approach enables global feature embedding and addresses GCN’s smoothness issue while enhancing interpretability. Through five-fold cross-validation, AGKphormer achieved average AUC and AUPR values of 97.32% and 97.34%, respectively, outperforming most methods and demonstrating its effectiveness in predicting disease-related metabolites. Additionally, case studies further confirm that AGKphormer is a reliable tool for discovering potential metabolites.