A dual-scale fused hypergraph convolution-based hyperedge prediction model for predicting missing reactions in genome-scale metabolic networks

• DOI: 10.1093/bib/bbae383
• Publication Date: 2024-08-05T10:50:38Z
• Authors: Weihong Huang; Feng Yang; Qiang Zhang; Juan Liu
• Keywords: Hypergraph; Convolution (computer science)

Abstract Genome-scale metabolic models (GEMs) are powerful tools for predicting cellular and physiological states. However, there still missing reactions in GEMs due to incomplete knowledge. Recent gaps filling methods suggest directly responses without relying on phenotypic data. they do not differentiate between substrates products when constructing the prediction models, which affects predictive performance of models. In this paper, we propose a hyperedge model that distinguishes based dual-scale fused hypergraph convolution, DSHCNet, inferring effectively fill GEM. First, each as heterogeneous complete graph then decompose it into three subgraphs at both homogeneous scales. Then design two convolution-based to, respectively, extract features vertices scales, via attention mechanism. Finally, all further pooled generate representative feature hyperedge. The strategy decomposition DSHCNet enables engage message passing independently thereby enhancing capability information propagation making obtained product substrate more distinguishable. experimental results show average recovery rate by is least 11.7% higher than state-of-the-art methods, gap-filled our achieve best performance, demonstrating superiority method.