Genome-scale metabolic network reconstructions are still built by automated gap-filling procedures that add reactions to a genome-derived draft network until it can perform required tasks such as biomass production. In the MIRAGE paper (Vitkin and Shlomi, Genome Biology 2012), which introduced functional-genomics-guided gap filling, the authors identified a cluster of unresolved problems that remain open today: when multiple alternative (not necessarily minimal) pathways could each complete the network, no gap-filling algorithm can determine which one(s) the organism actually uses, which plausibly explains the low recall of all tested methods; reconstruction does not enforce thermodynamic feasibility of the added pathways; and there is no principled way to weight heterogeneous evidence (sequence homology, protein-protein interactions, genomic context, metabolomics) when selecting reactions. The overarching objective is that automated reconstructions for many species approach the accuracy of expert-curated models without heavy manual curation. The 2014-2026 literature on mixed-integer-optimization, machine-learning, and database-curation approaches reports substantial partial progress but no resolution of the alternative-pathway identification core.

Background

A genome-scale metabolic model (GEM) is a formal inventory of all reactions, metabolites, and gene-protein-reaction associations (the links stating which genes encode the enzymes that catalyze each reaction) that an organism's metabolism is believed to comprise, usable with constraint-based methods such as flux balance analysis (FBA) — a linear-programming framework that predicts metabolic fluxes (reaction rates) subject to stoichiometric mass balance and capacity constraints. Because genome annotation is incomplete, a draft GEM built from an annotated genome contains gaps: dead-end metabolites (metabolites that no reaction in the draft can produce or consume) and missing capabilities such as inability to synthesize biomass precursors. Gap filling is the computational step that proposes adding reactions from a universal reference database so the draft model regains the required functions; the classic formulation (Satish Kumar, Dasika and Maranas, 2007) solves a mixed-integer linear program (MILP) that finds a minimal set of database reactions enabling a target such as growth. Because a minimal-cost solution is to a large extent arbitrary among many possible ones, and because the removed/missing reactions are unknown in real applications, gap filling frequently adds incorrect or biologically absent reactions. MIRAGE (MetabolIc Reconstruction via functionAl GEnomics; Vitkin and Shlomi, 2012) improved on minimal-set methods by weighting candidate reactions with functional-genomics evidence — gene-expression and phylogenetic-profile similarity (similarity of the presence-absence patterns of genes across genomes) between the candidate reaction's enzymes and the organism's genes — allowing non-minimal pathways supported by such evidence to be chosen, and it was applied to reconstruct models for 36 cyanobacteria. Nevertheless, its authors reported low recall and closed with the limitations that define this problem. Thermodynamic Metabolic Flux Analysis (TMFA; Henry, Broadbelt and Hatzimanikatis, 2007) is a constraint-based method that augments FBA with Gibbs-energy inequalities so that reaction directions and flux cycles respect the second law of thermodynamics; MIRAGE suggested, but did not implement, TMFA-style constraints inside the reconstruction consistency checks, noting that the required speedups were unavailable. Subsequent tool development — likelihood-based gene annotations (Benedict et al., 2014), fast MILP gap fillers (fastGapFill, Thiele et al., 2014), pattern/constraint hybrids (BoostGAPFILL, Oyetunde et al., 2017), topological completion (Meneco, Prigent et al., 2017), top-down carving from curated templates (CarveMe, Machado et al., 2018), taxonomically weighted reaction costs (Ong, Midford and Karp, 2020), evidence-scored gap filling with a thermodynamically curated database (gapseq, Zimmermann, Kaleta and Waschina, 2021), and topology-based deep learning (CHESHIRE, Chen, Liao and Liu, 2023) — has attacked pieces of the problem, but the central difficulty stated by MIRAGE persists: a completion task generally admits many alternative pathway solutions, and current methods cannot tell which alternative the organism actually contains.

Problem Statement

The problem, as posed at the close of the MIRAGE paper and still unresolved, is to make automated gap filling of genome-scale metabolic network reconstructions biologically faithful, in the following jointly linked sense. When a draft network derived from an annotated genome can be completed by several alternative pathways — not necessarily of minimal size — that each restore all required metabolic objectives (such as biomass production on a defined medium), determine which alternative pathway(s) the organism actually contains, by integrating functional-genomics signals (for example gene-expression and phylogenetic profiles, and more broadly protein-protein interaction and genomic-context evidence and metabolomics-defined metabolite cores) with metabolic flux analysis, instead of returning one arbitrary minimal-cost solution. In doing so, enforce thermodynamic feasibility of the reconstructed pathways, in the spirit of Thermodynamic Metabolic Flux Analysis, within the reconstruction consistency checks at tractable computational cost. The overarching objective uniting these requirements is that automatically reconstructed models for many species (MIRAGE's motivating scale is hundreds of organisms) approach the accuracy of expert-curated models, so that residual manual curation becomes a refinement rather than a correction step. The added reactions of the reconstruction must match the organism's true (curated-model or experimentally supported) reactions whenever the integrated evidence uniquely determines the true completion, including cases where the true completion is one of several alternative feasible pathways, and the reconstruction must correctly report non-identifiability, rather than an arbitrary completion, in the cases where the evidence does not determine the pathway; whether the remaining gap between automated and curated quality is due to fixable methodology or to an intrinsic identifiability limit of the available data lies outside the scope of the problem stated here.

References

  1. Vitkin E, Shlomi T (2012). MIRAGE: a functional genomics-based approach for metabolic network model reconstruction and its application to cyanobacteria networks. Genome Biology 13(11):R111. DOI: 10.1186/gb-2012-13-11-r111. https://doi.org/10.1186/gb-2012-13-11-r111
  2. Satish Kumar V, Dasika MS, Maranas CD (2007). Optimization based automated curation of metabolic reconstructions. BMC Bioinformatics 8:212. DOI: 10.1186/1471-2105-8-212. https://doi.org/10.1186/1471-2105-8-212
  3. Henry CS, Broadbelt LJ, Hatzimanikatis V (2007). Thermodynamics-Based Metabolic Flux Analysis. Biophysical Journal 92(5):1792-1805. DOI: 10.1529/biophysj.106.093138. https://doi.org/10.1529/biophysj.106.093138
  4. Benedict MN, Mundy MB, Henry CS, Chia N, Price ND (2014). Likelihood-Based Gene Annotations for Gap Filling and Quality Assessment in Genome-Scale Metabolic Models. PLoS Computational Biology 10(10):e1003882. DOI: 10.1371/journal.pcbi.1003882. https://doi.org/10.1371/journal.pcbi.1003882
  5. Thiele I, Vlassis N, Fleming RMT (2014). fastGapFill: efficient gap filling in metabolic networks. Bioinformatics 30(17):2529-2531. DOI: 10.1093/bioinformatics/btu321. https://doi.org/10.1093/bioinformatics/btu321
  6. Martínez VS, Quek LE, Nielsen LK (2014). Network Thermodynamic Curation of Human and Yeast Genome-Scale Metabolic Models. Biophysical Journal 107(2):493-503. DOI: 10.1016/j.bpj.2014.05.029. https://doi.org/10.1016/j.bpj.2014.05.029
  7. Oyetunde T, Zhang M, Chen Y, Tang Y, Lo C (2017). BoostGAPFILL: improving the fidelity of metabolic network reconstructions through integrated constraint and pattern-based methods. Bioinformatics 33(4):608-611. DOI: 10.1093/bioinformatics/btw684. https://doi.org/10.1093/bioinformatics/btw684
  8. Prigent S, Frioux C, Dittami SM, Thiele S, Larhlimi A, Collet G, Gutknecht F, Got J, Eveillard D, Bourdon J, Plewniak F, Tonon T, Siegel A (2017). Meneco, a Topology-Based Gap-Filling Tool Applicable to Degraded Genome-Wide Metabolic Networks. PLoS Computational Biology 13(1):e1005276. DOI: 10.1371/journal.pcbi.1005276. https://doi.org/10.1371/journal.pcbi.1005276
  9. Karp PD, Weaver D, Latendresse M (2018). How accurate is automated gap filling of metabolic models? BMC Systems Biology 12:73. DOI: 10.1186/s12918-018-0593-7. https://doi.org/10.1186/s12918-018-0593-7
  10. Latendresse M, Karp PD (2018). Evaluation of reaction gap-filling accuracy by randomization. BMC Bioinformatics 19:53. DOI: 10.1186/s12859-018-2050-4. https://doi.org/10.1186/s12859-018-2050-4
  11. Pan S, Reed JL (2018). Advances in gap-filling genome-scale metabolic models and model-driven experiments lead to novel metabolic discoveries. Current Opinion in Biotechnology 51:103-108. DOI: 10.1016/j.copbio.2017.12.012. https://doi.org/10.1016/j.copbio.2017.12.012
  12. Machado D, Andrejev S, Tramontano M, Patil KR (2018). Fast automated reconstruction of genome-scale metabolic models for microbial species and communities. Nucleic Acids Research 46(15):7542-7553. DOI: 10.1093/nar/gky537. https://doi.org/10.1093/nar/gky537
  13. Salvy P, Fengos G, Ataman M, Pathier T, Soh KC, Hatzimanikatis V (2019). pyTFA and matTFA: a Python package and a Matlab toolbox for Thermodynamics-based Flux Analysis. Bioinformatics 35(1):167-169. DOI: 10.1093/bioinformatics/bty499. https://doi.org/10.1093/bioinformatics/bty499
  14. Ong WK, Midford PE, Karp PD (2020). Taxonomic weighting improves the accuracy of a gap-filling algorithm for metabolic models. Bioinformatics 36(6):1823-1830. DOI: 10.1093/bioinformatics/btz813. https://doi.org/10.1093/bioinformatics/btz813
  15. Bernstein DB, Sulheim S, Almaas E, Segrè D (2021). Addressing uncertainty in genome-scale metabolic model reconstruction and analysis. Genome Biology 22:64. DOI: 10.1186/s13059-021-02289-z. https://doi.org/10.1186/s13059-021-02289-z
  16. Zimmermann J, Kaleta C, Waschina S (2021). gapseq: informed prediction of bacterial metabolic pathways and reconstruction of accurate metabolic models. Genome Biology 22:81. DOI: 10.1186/s13059-021-02295-1. https://doi.org/10.1186/s13059-021-02295-1
  17. Chen C, Liao C, Liu YY (2023). Teasing out missing reactions in genome-scale metabolic networks through hypergraph learning. Nature Communications 14:2375. DOI: 10.1038/s41467-023-38110-7. https://doi.org/10.1038/s41467-023-38110-7
  18. Huang W, Yang F, Zhang Q, Liu J (2024). A dual-scale fused hypergraph convolution-based hyperedge prediction model for predicting missing reactions in genome-scale metabolic networks. Briefings in Bioinformatics 25(5):bbae383. DOI: 10.1093/bib/bbae383. https://doi.org/10.1093/bib/bbae383
  19. Moyer DC, Reimertz J, Segrè D, Fuxman Bass J (2025). MACAW: a method for semi-automatic detection of errors in genome-scale metabolic models. Genome Biology 26:79. DOI: 10.1186/s13059-025-03533-6. https://doi.org/10.1186/s13059-025-03533-6
  20. Qu J, Wang K (2026). A Novel Topology-Based Candidate Reaction Prediction Approach for Gap-Fillings of Genome-Scale Metabolic Models. Metabolites 16(4):258. DOI: 10.3390/metabo16040258. https://doi.org/10.3390/metabo16040258