[Paper Review] A generalizable framework for unlocking missing reactions in genome-scale metabolic networks using deep learning

Xiaoyi Liu, Hongpeng Yang+6 more|arXiv (Cornell University)|Sep 20, 2024

Microbial Metabolic Engineering and BioproductionBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper introduces CLOSEgaps, a deep learning framework that models genome-scale metabolic networks as hypergraphs to predict missing reactions in draft GEMs without relying on phenotypic data. By integrating hypergraph convolutional networks and attention mechanisms, it achieves over 96% accuracy in gap-filling across diverse organisms, significantly improving phenotypic prediction and key metabolite production.

ABSTRACT

Incomplete knowledge of metabolic processes hinders the accuracy of GEnome-scale Metabolic models (GEMs), which in turn impedes advancements in systems biology and metabolic engineering. Existing gap-filling methods typically rely on phenotypic data to minimize the disparity between computational predictions and experimental results. However, there is still a lack of an automatic and precise gap-filling method for initial state GEMs before experimental data and annotated genomes become available. In this study, we introduce CLOSEgaps, a deep learning-driven tool that addresses the gap-filling issue by modeling it as a hyperedge prediction problem within GEMs. Specifically, CLOSEgaps maps metabolic networks as hypergraphs and learns their hyper-topology features to identify missing reactions and gaps by leveraging hypothetical reactions. This innovative approach allows for the characterization and curation of both known and hypothetical reactions within metabolic networks. Extensive results demonstrate that CLOSEgaps accurately gap-filling over 96% of artificially introduced gaps for various GEMs. Furthermore, CLOSEgaps enhances phenotypic predictions for 24 GEMs and also finds a notable improvement in producing four crucial metabolites (Lactate, Ethanol, Propionate, and Succinate) in two organisms. As a broadly applicable solution for any GEM, CLOSEgaps represents a promising model to automate the gap-filling process and uncover missing connections between reactions and observed metabolic phenotypes.

Motivation & Objective

• Address the persistent challenge of incomplete knowledge in genome-scale metabolic models (GEMs), which limits their accuracy and utility in systems biology and metabolic engineering.
• Overcome the limitations of traditional gap-filling methods that rely on costly, time-consuming phenotypic screening and manual curation.
• Develop a model-free, data-driven approach capable of identifying missing reactions in draft GEMs before experimental data or annotated genomes are available.
• Enable automated, scalable, and generalizable gap-filling for any organism, including non-model or uncultivable species.
• Improve the predictive power of GEMs by accurately recovering missing metabolic connections and enhancing flux balance analysis outcomes.

Proposed method

• Represent metabolic networks as hypergraphs, where hyperedges correspond to reactions and hypernodes represent metabolites, enabling topological modeling of complex reaction-metabolite relationships.
• Employ a hypergraph convolutional network (HGCN) with learnable parameters to extract hierarchical topological features from the hypergraph structure.
• Integrate an attention mechanism to dynamically weigh the importance of neighboring hyperedges and metabolites during feature propagation.
• Formulate gap-filling as a hyperlink prediction problem: predict the existence or absence of hypothetical reactions using a binary classification head on top of learned hyperedge representations.
• Use a similarity function (Equation 3) to compute attention weights between nodes and hyperedges, enabling context-aware representation learning.
• Apply a confidence threshold of 0.99999 to rank candidate reactions from the BiGG database, and iteratively add the top 200 reactions while detecting and eliminating energy-generating cycles (EGCs) via flux bound adjustments.

Experimental results

Research questions

• RQ1 Can a deep learning framework predict missing reactions in draft GEMs without requiring prior phenotypic data or manual curation?
• RQ2 How effective is the hypergraph-based representation learning approach in capturing topological features of metabolic networks for gap-filling?
• RQ3 To what extent does the proposed method improve phenotypic prediction accuracy in diverse GEMs across multiple organisms?
• RQ4 Can the framework identify biologically relevant missing reactions that enhance the production of key metabolites like lactate, ethanol, propionate, and succinate?
• RQ5 How generalizable is the approach across different organisms, including non-model and uncultivable species?

Key findings

• CLOSEgaps achieved over 96% accuracy in gap-filling for artificially introduced gaps across multiple genome-scale metabolic models, demonstrating high predictive reliability.
• The framework enhanced phenotypic prediction accuracy for 24 distinct GEMs, indicating improved model fidelity and functional coverage.
• In two organisms, CLOSEgaps significantly improved the in silico production of four crucial metabolites—lactate, ethanol, propionate, and succinate—by identifying and integrating missing reactions.
• The use of a stringent confidence threshold (0.99999) and iterative EGC detection ensured high-quality gap-filling while avoiding metabolic inconsistencies.
• The method successfully identified 15 energy dissipation reactions (e.g., for ATP, GTP) that were critical for eliminating spurious flux cycles in reconstructed models.
• The framework is broadly applicable to any GEM, regardless of organism type, and operates effectively even in the absence of experimental phenotypic data.