MetaGEM: Bottom-Up Reconstruction of Genome-Scale Metabolic Networks via Deep Enzyme-Metabolite Anchoring

TL;DR

MetaGEM is a novel bottom-up framework that leverages enzyme and metabolite data with deep learning to accurately reconstruct genome-scale metabolic networks, improving upon traditional methods.

Contribution

MetaGEM introduces a dual-tower deep learning architecture with contrastive learning for enzyme-metabolite prediction, enabling robust, automated metabolic network reconstruction.

Findings

Achieves AUROC of 0.9701 and MCC of 0.8033 on benchmark

Generates functional models for E. coli, B. subtilis, P. aeruginosa

Models align well with experimental phenotype and gene data

Abstract

Genome-scale metabolic models (GEMs) are essential tools for systems biology and rational chassis design, but conventional top-down reconstruction depends heavily on sequence homology and often leaves unknown enzymes and metabolic dark matter unresolved. Direct reconstruction from metabolomics is also difficult because mapping observed metabolites to reactions is an ill-posed inverse problem with combinatorial ambiguity and possible spurious networks. Here we present MetaGEM, a bottom-up framework that uses enzymes as physical anchors to convert system-level network inference into enzyme-metabolite interaction prediction. MetaGEM uses a multimodal dual-tower architecture that combines protein evolutionary semantics from a protein language model with three-dimensional metabolite representations. It further introduces contrastive learning with hard negative mining to separate structurally similar metabolites and reduce false positive interactions. On a de-homologized benchmark, MetaGEM achieves state-of-the-art enzyme-metabolite prediction performance, with AUROC of 0.9701 and MCC of 0.8033, and remains robust under low sequence identity splits. In downstream reconstruction, MetaGEM generates functional genome-scale metabolic models for Escherichia coli, Bacillus subtilis, and Pseudomonas aeruginosa. The reconstructed models improve network connectivity, capture promiscuous enzymes, and show strong agreement with experimental phenotype microarray and gene essentiality data. These results indicate that MetaGEM provides a practical route from metabolomic evidence to computable metabolic networks and offers a foundation for automated AI-driven virtual cell reconstruction.