MetDeeCINE: Deciphering Metabolic Regulation through Deep Learning and Multi-Omics

Abstract

Metabolism, the biochemical reaction network within cells, is crucial for life, health, and disease. Recent advances in multi-omics technologies, enabling the simultaneous measurement of transcripts, proteins, and metabolites, provide unprecedented opportunities to comprehensively analyze metabolic regulation. However, effectively integrating these diverse data types to decipher the complex interplay between enzymes and metabolites remains a significant challenge due to the extensive data requirements of kinetic modeling approaches and the limited interpretability of machine learning approaches. Here, we present MetDeeCINE, a novel explainable deep learning framework that predicts the quantitative relationship between each enzyme and metabolite from proteomic and metabolomic data. We demonstrate that our newly developed Metabolism-informed Graph Neural Network (MiGNN), a core component of MetDeeCINE that is guided by the stoichiometric information of metabolic reactions, outperforms other machine learning models in predicting concentration control coefficients (CCCs) using data obtained from kinetic models of E. coli. Notably, MetDeeCINE, even without explicit information on allosteric regulation, can identify key distant enzymes that predominantly control the steady-state concentrations of specific metabolites. Application of MetDeeCINE to mouse liver multi-omics experimental data further demonstrated its ability to generate biologically meaningful predictions through identifying a rate-limiting enzyme of gluconeogenesis associated with obesity, consistent with existing knowledge. MetDeeCINE offers a scalable and interpretable approach for deciphering complex metabolic regulation from multi-omics data, with broad applications in disease research, drug discovery, and metabolic engineering.