Bioinformatics Approaches in Systems Biology

Abstract

Systems biology seeks to understand cellular behaviour through the integration of multi-omics data -- genomics, transcriptomics, proteomics, and metabolomics -- into computational models that capture the emergent properties of biological networks. However, the heterogeneity of omics data types, the sparsity of experimentally validated molecular interactions, and the computational complexity of genome-scale models remain critical bottlenecks. This study developed and benchmarked five bioinformatics integration strategies -- correlation-based network inference (WGCNA), mutual information networks (ARACNe), Bayesian network reconstruction (BNFinder), constraint-based metabolic modelling (flux balance analysis, FBA), and a proposed multi-omics graph neural network (MOG-Net) -- for reconstructing gene regulatory and metabolic networks in two model organisms: Saccharomyces cerevisiae (yeast) and Escherichia coli. Integration accuracy was evaluated against gold-standard experimentally validated interaction databases (YEASTRACT for yeast, RegulonDB for E. coli). MOG-Net achieved the highest area under the precision-recall curve (AUPRC 0.428 + 0.024 for yeast, 0.386 +- 0.018 for E. coli), significantly outperforming WGCNA (0.218 +- 0.032; p < 0.001), ARACNe (0.264 +- 0.028; p < 0.001), and BNFinder (0.312 +- 0.022; p = 0.002). FBA flux predictions correlated strongly with experimental 13C-flux measurements (Pearson r = 0.92 for yeast central metabolism). MOG-Net's graph attention mechanism identified 142 novel putative regulatory interactions in yeast not present in YEASTRACT, of which 38 (26.8%) were independently supported by published ChIP-seq or perturbation data. These results establish multi-omics graph neural networks as a powerful framework for integrative systems biology and provide a reproducible benchmark for network inference methods.