pith. sign in

arxiv: 2604.25233 · v1 · submitted 2026-04-28 · 🧮 math.OC · q-bio.GN

A Combinatorial Optimisation Approach to Multi-factorial Gap-filling in Genome-scale Metabolic Models (GEMs)

Philip Kilby , Sevvandi Kandanaarachchi , Matthew J. Morgan , Amy M. Paten , Mariana Velasque , Andrew C. Warden , Juan P. Molina Ortiz This is my paper

Pith reviewed 2026-05-07 15:58 UTC · model grok-4.3

classification 🧮 math.OC q-bio.GN
keywords gap-fillinggenome-scale metabolic modelscombinatorial optimizationmetaheuristicslinear programmingmulti-factorialbacterial metabolism
0
0 comments X

The pith

A metaheuristic selects reaction subsets to gap-fill metabolic models across many media using only linear programs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper treats multi-factorial gap-filling of genome-scale metabolic models as the task of choosing one reaction subset that must work well on many different media conditions simultaneously. Conventional methods apply integer linear programming separately to each medium and then combine the results, which can produce models that behave unrealistically on some conditions and take a long time to compute. The new approach uses a metaheuristic search whose only subproblem is solving continuous linear programs to score each candidate subset on all media at once. On three bacterial strains and between nine and twenty-eight media, the method selects three to four thousand reactions from a pool of over eleven thousand and reports average gains of 7.3 percent in Kendall Tau rank correlation and 13.3 percent in root-mean-square error relative to the sequential baseline.

Core claim

The authors claim that a metaheuristic search over reaction subsets, guided by continuous linear-programming evaluations of biomass or flux matching on each medium, produces genome-scale metabolic models whose predictions match empirical data more closely across all tested conditions than models obtained by applying single-medium integer linear programs sequentially.

What carries the argument

A metaheuristic combinatorial optimizer that explores subsets of reactions from a large database, where each candidate subset is scored by solving one continuous linear program per medium to quantify how well the selected reactions reproduce measured growth or flux values.

Load-bearing premise

That a metaheuristic search guided only by continuous LP evaluations will locate reaction sets that generalize across media without excessive trapping in local optima or prohibitive run times.

What would settle it

Apply both the metaheuristic and the conventional sequential integer-programming method to a new bacterial genome-scale model together with twenty or more independent media conditions; if the metaheuristic yields higher root-mean-square error or lower Kendall Tau correlation with measured growth rates, the performance claim is falsified.

Figures

Figures reproduced from arXiv: 2604.25233 by Amy M. Paten, Andrew C. Warden, Juan P. Molina Ortiz, Mariana Velasque, Matthew J. Morgan, Philip Kilby, Sevvandi Kandanaarachchi.

Figure 1
Figure 1. Figure 1: Comparison of methods and performance before and after gap-filling. view at source ↗
Figure 3
Figure 3. Figure 3: A simple metabolic network, using reactions R1, R2 and R3, and metabolites A through G. Stoichiometric view at source ↗
read the original abstract

Genome-Scale Metabolic Models (GEMs) describe the interactions between genes, proteins, and the biochemical reactions that underpin an organism's metabolism aiming to computationally simulate functions at the cellular level. While many metabolic reactions can be inferred from genome analysis, constructing GEMs often involves incorporating reactions unsupported by genomic data to improve prediction accuracy. This is known as gap-filling, a process that can be performed manually (a time-consuming task) or computationally. Traditional computational gap-filling approaches aim to correct GEM predictions for a single environmental condition (medium) by solving a large Integer Linear Programming problem. Sequential application across multiple media can produce a more robust model, but often introduces unrealistic predictions in other media. They are also slow to run. In this paper, we study multi-factorial gap filling, which aims to gap-fill GEMs across typically 10 or more input media simultaneously, while improving their overall predictive accuracy and minimising unrealistic behaviour. We view the selection of the set of reactions as a combinatorial optimisation problem, and describe a method based on classic metaheuristic approaches which requires the solution of continuous Linear Programming problems only. This paper provides an introduction of this problem to an audience whose speciality lies outside biology, and suggests a practical first-cut solution method. We demonstrate the method gap-filling GEMs for three bacteria strains, selecting 3000 to 4000 reactions from a database of more than 11000 reactions, while attempting to match the empirically measured performance on 9 to 28 separate media conditions. We show that our method outperforms conventional approaches on multiple metrics, including Kendal Tau and RMS Error by an average of 7.3% and 13.3%, respectively.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper formulates multi-factorial gap-filling in GEMs as a combinatorial optimization problem over reaction subsets (3000–4000 chosen from >11000 candidates) and solves it with metaheuristics that invoke only continuous LP evaluations. The approach is applied to three bacterial GEMs on 9–28 media conditions and is reported to improve Kendall Tau by 7.3 % and RMS Error by 13.3 % on average relative to conventional single-medium ILP gap-filling.

Significance. A reliable multi-factorial method would reduce the computational cost and unrealistic cross-media predictions that arise from sequential single-medium gap-filling, offering a practical route to more robust GEMs for metabolic engineering and systems biology.

major comments (3)
  1. [Abstract, §4] Abstract and §4 (Results): the headline performance figures (7.3 % Kendall Tau, 13.3 % RMS Error) are presented as simple averages without statistical tests, confidence intervals, per-medium breakdowns, or explicit descriptions of how the baseline ILP implementations and media-specific predictions were aggregated; these omissions make it impossible to judge whether the reported gains are statistically meaningful or robust to implementation details.
  2. [§3, §4] §3 (Method) and §4: no convergence diagnostics, restart statistics, or parameter-sensitivity results are supplied for the metaheuristic, despite the stochastic search over a >11000-reaction space and the claim that the selected sets generalize across 9–28 media; without such evidence the central assumption that the metaheuristic reliably locates generalizable reaction sets remains unsupported.
  3. [§4] §4: the manuscript contains no comparison against exact ILP formulations on down-scaled instances or any other verification that the metaheuristic solutions are close to optimal; this leaves open the possibility that the observed gains are artifacts of local optima or media-specific overfitting rather than genuine multi-factorial improvement.
minor comments (2)
  1. [§3] The description of the metaheuristic control parameters (population size, iteration limits, acceptance thresholds) is incomplete; a table listing the exact values used for each GEM would improve reproducibility.
  2. [§2] Notation for the LP objective and constraint sets is introduced without a compact summary table; readers outside metabolic modeling would benefit from an explicit mapping between biological quantities and mathematical symbols.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments highlight important aspects of statistical rigor, algorithmic reliability, and solution quality that we will address through targeted revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (Results): the headline performance figures (7.3 % Kendall Tau, 13.3 % RMS Error) are presented as simple averages without statistical tests, confidence intervals, per-medium breakdowns, or explicit descriptions of how the baseline ILP implementations and media-specific predictions were aggregated; these omissions make it impossible to judge whether the reported gains are statistically meaningful or robust to implementation details.

    Authors: We agree that the current presentation of aggregate averages limits assessment of robustness and statistical significance. In the revised manuscript we will add per-medium performance tables, bootstrap-derived confidence intervals on the reported improvements, explicit descriptions of how ILP baselines and media predictions were aggregated, and paired statistical tests (e.g., Wilcoxon signed-rank) to evaluate whether the observed gains are significant. revision: yes

  2. Referee: [§3, §4] §3 (Method) and §4: no convergence diagnostics, restart statistics, or parameter-sensitivity results are supplied for the metaheuristic, despite the stochastic search over a >11000-reaction space and the claim that the selected sets generalize across 9–28 media; without such evidence the central assumption that the metaheuristic reliably locates generalizable reaction sets remains unsupported.

    Authors: We acknowledge that additional evidence of algorithmic stability is needed. The revised version will include convergence plots of the objective function across iterations, summary statistics (mean, standard deviation, best/worst) from multiple independent restarts, and a sensitivity analysis on key parameters such as population size and mutation rate. These additions will directly support the claim that the metaheuristic consistently identifies generalizable reaction sets. revision: yes

  3. Referee: [§4] §4: the manuscript contains no comparison against exact ILP formulations on down-scaled instances or any other verification that the metaheuristic solutions are close to optimal; this leaves open the possibility that the observed gains are artifacts of local optima or media-specific overfitting rather than genuine multi-factorial improvement.

    Authors: We recognize the value of verifying solution quality against exact optima. Although the full-scale combinatorial problem is intractable for exact ILP solvers, we will add experiments on down-scaled instances (reduced candidate reactions or fewer media) where exact solutions are computable. We will report the gap between metaheuristic and optimal objective values and discuss implications for local-optima or overfitting concerns. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes a metaheuristic combinatorial optimization method that selects reaction subsets from a large database by repeatedly solving continuous LP problems to match empirical growth rates across 9-28 media conditions for three GEMs. Reported gains (7.3% Kendall Tau, 13.3% RMS Error) are computed directly against external experimental measurements on those media, not against quantities defined inside the paper's own equations or fitted constants. No self-definitional steps, fitted-input predictions, or load-bearing self-citations appear in the abstract or described approach; the method is a standard application of metaheuristics with external benchmarking, rendering the derivation chain self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The approach rests on standard GEM assumptions and off-the-shelf optimization primitives; no new entities are postulated.

free parameters (1)
  • Metaheuristic control parameters (population size, iteration limits, acceptance thresholds)
    Typical tunable settings for metaheuristics that affect search behavior and must be chosen by the user.
axioms (2)
  • domain assumption Metabolic networks can be represented as linear programs whose feasible fluxes predict growth rates under a given medium.
    Core modeling assumption of flux-balance analysis used throughout GEM literature.
  • domain assumption A single set of added reactions can simultaneously improve predictive accuracy across multiple independent media without creating contradictions.
    The central modeling premise that justifies solving the problem jointly rather than sequentially.

pith-pipeline@v0.9.0 · 5646 in / 1409 out tokens · 61809 ms · 2026-05-07T15:58:27.298252+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references

  1. [1]

    A synthetic bacterium that degrades and assimilates poly (ethylene terephthalate).bioRxiv, pages 2025–09, 2025

    Dekel Freund, Kesava Phaneendra Cherukuri, Raul Mireles, Joseph Kippen, Maya Shossel, and Lianet Noda- García. A synthetic bacterium that degrades and assimilates poly (ethylene terephthalate).bioRxiv, pages 2025–09, 2025

    2025

  2. [2]

    Garland and Aaron L

    Jay L. Garland and Aaron L. Mills. Classification and Characterization of Heterotrophic Microbial Communities on the Basis of Patterns of Community-Level Sole-Carbon-Source Utilization.Applied and Environmental Microbiology, 57(8):2351–2359, August 1991

    1991

  3. [3]

    Daniel Hartleb, Florian Jarre, and Martin J. Lercher. Improved Metabolic Models for E. coli and Mycoplasma genitalium from GlobalFit, an Algorithm That Simultaneously Matches Growth and Non-Growth Data Sets.PLOS Computational Biology, 12(8):e1005036, 2016

    2016

  4. [4]

    The raven toolbox and its use for generating a genome-scale metabolic model for penicillium chrysogenum.PLoS computational biology, 9(3):e1002980, 2013

    Rasmus Agren, Liming Liu, Saeed Shoaie, Wanwipa V ongsangnak, Intawat Nookaew, and Jens Nielsen. The raven toolbox and its use for generating a genome-scale metabolic model for penicillium chrysogenum.PLoS computational biology, 9(3):e1002980, 2013

    2013

  5. [5]

    In silico cell factory design driven by comprehensive genome-scale metabolic models: Development and challenges.Systems Microbiology and Biomanufacturing, 3(2):207–222, 2023

    Jiangong Lu, Xinyu Bi, Yanfeng Liu, Xueqin Lv, Jianghua Li, Guocheng Du, and Long Liu. In silico cell factory design driven by comprehensive genome-scale metabolic models: Development and challenges.Systems Microbiology and Biomanufacturing, 3(2):207–222, 2023

    2023

  6. [6]

    High throughput genome scale modeling predicts microbial vitamin requirements contribute to gut microbiome community structure.Gut Microbes, 14(1):2118831, 2022

    Juan P Molina Ortiz, Mark Norman Read, Dale David McClure, Andrew Holmes, Fariba Dehghani, and Erin Rose Shanahan. High throughput genome scale modeling predicts microbial vitamin requirements contribute to gut microbiome community structure.Gut Microbes, 14(1):2118831, 2022

    2022

  7. [7]

    Disease-specific loss of microbial cross-feeding interactions in the human gut.Nature Communications, 14(1):6546, 2023

    Vanessa R Marcelino, Caitlin Welsh, Christian Diener, Emily L Gulliver, Emily L Rutten, Remy B Young, Edward M Giles, Sean M Gibbons, Chris Greening, and Samuel C Forster. Disease-specific loss of microbial cross-feeding interactions in the human gut.Nature Communications, 14(1):6546, 2023

    2023

  8. [8]

    M. R. Watson. Metabolic maps for the Apple II.Biochemical Society Transactions, 12(6):1093–1094, December 1984

    1984

  9. [9]

    The choice of the objective function in flux balance analysis is crucial for predicting replicative lifespans in yeast.Plos one, 17(10):e0276112, 2022

    Barbara Schnitzer, Linnea Österberg, and Marija Cvijovic. The choice of the objective function in flux balance analysis is crucial for predicting replicative lifespans in yeast.Plos one, 17(10):e0276112, 2022

    2022

  10. [10]

    Ravcheev, Malgorzata Nyga, Onyedika Emmanuel Okpala, Marcus Hogan, Stefanía Magnúsdóttir, Filippo Martinelli, Bram Nap, German Preciat, Janaka N

    Almut Heinken, Johannes Hertel, Geeta Acharya, Dmitry A. Ravcheev, Malgorzata Nyga, Onyedika Emmanuel Okpala, Marcus Hogan, Stefanía Magnúsdóttir, Filippo Martinelli, Bram Nap, German Preciat, Janaka N. Ediris- inghe, Christopher S. Henry, Ronan M. T. Fleming, and Ines Thiele. Genome-scale metabolic reconstruction of 7,302 human microorganisms for persona...

    2023

  11. [11]

    Ravcheev, Johannes Hertel, Malgorzata Nyga, Onyedika Emmanuel Okpala, Marcus Hogan, Stefanía Magnúsdóttir, Filippo Martinelli, German Preciat, Janaka N

    Almut Heinken, Geeta Acharya, Dmitry A. Ravcheev, Johannes Hertel, Malgorzata Nyga, Onyedika Emmanuel Okpala, Marcus Hogan, Stefanía Magnúsdóttir, Filippo Martinelli, German Preciat, Janaka N. Edirisinghe, Christopher S. Henry, Ronan M. T. Fleming, and Ines Thiele. AGORA2: Large scale reconstruction of the microbiome highlights wide-spread drug-metabolisi...

    2020

  12. [12]

    M. G. Kendall. A new measure of rank correlation.Biometrika, 30(1-2):81–93, 1938

    1938

  13. [13]

    Tabu Search

    Fred Glover and Manuel Laguna. Tabu Search. In C. R. Reeves, editor,Modern Heuristic Techniques for Combinatorial Problems, pages 70–150. Halsted Press, 1993

    1993

  14. [14]

    An Adaptive Large Neighborhood Search Heuristic for the Pickup and Delivery Problem with Time Windows.Transportation Science, 40(4):455–472, 2006

    Stefan Ropke and David Pisinger. An Adaptive Large Neighborhood Search Heuristic for the Pickup and Delivery Problem with Time Windows.Transportation Science, 40(4):455–472, 2006

    2006

  15. [15]

    Nikolaev and Sheldon H

    Alexander G. Nikolaev and Sheldon H. Jacobson. Simulated Annealing. In Michel Gendreau and Jean-Yves Potvin, editors,Handbook of Metaheuristics, volume 146, pages 1–39. Springer US, Boston, MA, 2010

    2010

  16. [16]

    Huangfu and J.A.J

    Q. Huangfu and J.A.J. Hall. Parallelizing the dual revised simplex method.Mathematical Programming Computation, 10(1):119–142, 2018

    2018

  17. [17]

    Genome-scale metabolic network analysis of the opportunistic pathogen pseudomonas aeruginosa pao1, 2008

    Matthew A Oberhardt, Jacek Puchałka, Kimberly E Fryer, Vítor AP Martins dos Santos, and Jason A Papin. Genome-scale metabolic network analysis of the opportunistic pathogen pseudomonas aeruginosa pao1, 2008

    2008

  18. [18]

    Prediction of microbial growth rate versus biomass yield by a metabolic network with kinetic parameters.PLoS computational biology, 8(7):e1002575, 2012

    Roi Adadi, Benjamin V olkmer, Ron Milo, Matthias Heinemann, and Tomer Shlomi. Prediction of microbial growth rate versus biomass yield by a metabolic network with kinetic parameters.PLoS computational biology, 8(7):e1002575, 2012

    2012

  19. [19]

    An experimentally validated genome-scale metabolic reconstruction of klebsiella pneumoniae mgh 78578, i yl1228.Journal of bacteriology, 193(7):1710–1717, 2011

    Yu-Chieh Liao, Tzu-Wen Huang, Feng-Chi Chen, Pep Charusanti, Jay SJ Hong, Hwan-You Chang, Shih-Feng Tsai, Bernhard O Palsson, and Chao A Hsiung. An experimentally validated genome-scale metabolic reconstruction of klebsiella pneumoniae mgh 78578, i yl1228.Journal of bacteriology, 193(7):1710–1717, 2011. 13

    2011

  20. [20]

    Aminobutyraldehyde Dehydrogenase

    Donovan H Parks, Maria Chuvochina, Christian Rinke, Aaron J Mussig, Pierre-Alain Chaumeil, and Philip Hugenholtz. GTDB: An ongoing census of bacterial and archaeal diversity through a phylogenetically consistent, rank normalized and complete genome-based taxonomy.Nucleic Acids Research, 50(D1):D785–D794, January 2022. A Technical details A.1 Calculation o...

    2022