MetabOmics: Metabolism-Oriented Omics Data Integration

Biological processes arise from complex interactions across multiple molecular layers, yet bridging the gap between disparate omics data types remains a significant challenge. This paper introduces MetabOmics, a comprehensive metabolism-oriented integrated multi-omics analysis method structurally designed to accommodate genomics, transcriptomics, proteomics, and metabolomics datasets. Our methodology centers on the construction of an integrated multi-omic interaction network that incorporates a wide range of biological interactions, including gene expression, translation, transcription factor activity, and post-transcriptional regulation via microRNAs. To capture the cascading effects of molecular changes, we map measured biological entities onto this network and utilize information diffusion models, such as Linear Threshold Diffusion, to propagate fold-changes throughout the system. These propagated measurements are then used to update the lower and upper bounds of metabolic reactions within a genome-scale metabolic model in a personalized manner. Finally, we apply an extended metabolic flux analysis algorithm to compute reaction and pathway differentiation scores.To demonstrate the empirical efficacy of this framework, we evaluated our approach using paired transcriptomics and metabolomics data across six different cancer cohorts. To further validate the framework's capacity for deep multi-omics integration, we additionally applied MetabOmics to the MayoRNASeq Progressive Supranuclear Palsy (PSP) cohort, successfully integrating transcriptomics, metabolomics, and proteomics. Our results demonstrate that our network-based integration achieves highly competitive classification performance compared to unconstrained multi-omics baselines, and significantly outperforms single-omics approaches. Crucially, we quantitatively establish that MetabOmics produces vastly more stable and biologically concordant feature selections; it achieves robust literature concordance across all evaluated cohorts, whereas simple data concatenation frequently fails to identify disease-relevant pathways.