Metabolomics Pathway Prediction Using Enhanced-Graph Convolutional Networks with Graph Attention Networks

Abstract

Metabolomics, the comprehensive study of small molecules in biological systems, has a central role to play in the diagnosis of diseases, biomarker detection, and the design of new drugs. Although there have been major breakthroughs in analytical toolsets such as mass spectrometry (MS) coupled with chromatography, it is hard to predict metabolomics pathways because biochemical interactions are inherently complex. To meet this end, the current research suggests a deep learning-based approach using graph neural networks (GNN), which have shown high efficiency for graph-structured biological data. We specifically propose an enhanced graph convolutional network integrated with graph attention networks (EGCNGAT) to enhance pathway prediction performance. The hybrid framework employs graph convolutional networks (GCN) to represent molecular structural data and graph attention networks (GAT) to provide context-sensitive feature importance, thus improving the model’s capacity for learning complex pathway patterns. Comparative experiments against current deep learning approaches show that the introduced EGCN-GAT model obtains an accuracy of 98.90 percent, which is a 0.26 percent increase compared to the baseline MLGL-MP model. In addition, it demonstrates a 0.94 percent gain in precision as well as a slight gain in recall. The findings validate the performance of the proposed method and highlight its utility for developing pathway-level predictions in metabolomics studies.