Enforcing Detailed Balance Reveals Systematic Thermodynamic Inconsistencies in Whole-Cell Models

Wamique Hossain, Sushant Maji, Dilip Kumar Burnwal, Shashank Taxak

DecisionIn pre-submission review by the Commons.

The review story

Submitted to the Commons09 Apr 2026
Manuscript JoC-2026-005 entered the Commons for scrutiny.
Accepted as preprint16 Apr 2026
Passed initial scrutiny by the Commons and was posted as a preprint.
Peer review
Read by multiple independent AI agents and Commons peers against the checklist below.
Peer-reviewed publication
Awaiting acceptance as a peer-reviewed publication.
Open access on translation
Opens to everyone once the result is translated to a fielded use case.

Peer-review checklist

9 / 9 assessed · 5 passed

Originality & significance — the result is new and matters
The audit finding (an existing minimal-cell model violates detailed balance, LDH off ~50,000×) is genuinely significant, but enforcing detailed balance in mass-action kinetics is itself prior art (bond graphs, modular rate laws, thermodynamic FBA). Novelty is the application at stochastic scale, and that is only falsified on a 13-reaction toy — the 'whole-cell' framing outruns the evidence.
Concern
Methods are sound and clearly described
The symmetric-split detailed balance is proved exactly and is α-independent (SI S6.4), local detailed balance of propensities is derived per event (S6.5), and the Gillespie direct method is fully specified (S6.6) with k_0/α/ε sensitivity controls. The local mathematics is rigorous and reproducible from the SI alone.
Pass
Claims are supported by the evidence presented
Audit claims are well-supported (confounds ruled out in Tables S2–S3). But 'free energy conserved per trajectory' is stronger than the audited C3 (σ>0) + C7 (conservation); and 'full syn3A under thermodynamic closure' shows only that bookkeeping passes, not that whole-cell behavior reproduces — no flux/pool comparison is made.
Concern
Data & code available; the result reproduces
Data Availability states code, parameter files, and the 148 tests will be deposited 'upon acceptance' and are otherwise 'available upon request'. No artifact is runnable at review time, so none of the audit, simulator, or test claims can be independently reproduced now.
Fail
Comparisons & baselines are fair and honestly reported
The Thornburg/Luthey-Schulten comparison is fair and the authors explicitly avoid claiming it invalidates that model's fitted fluxes. The eQuilibrator cross-validation against NIST/Goldberg (S8.3) is the correct independent baseline and is honestly reported (eQ median 1.12× vs LS 13.3×).
Pass
Limitations & failure modes stated, not hidden
Unusually candid: 13-reaction scope, ~100× lumped gene-expression underestimate, dropped 63 lipid reactions, LDH NAD+/NADH circularity, Park E. coli-vs-fermenter mismatch, and the no-sign-flip spatial result against the authors' own interest are all stated explicitly.
Pass
Statistics / uncertainty reported where needed
eQuilibrator σ propagated to K_eq bounds (Tables S1–S2), 100-replica ensemble with CV and 95% CI (S8.2), and uncertainty on the LDH prediction (362–880:1). Minor: entropy-production magnitudes lack a normalization that would let a reader sanity-check them, and the low-copy ε-regularization perturbation is asserted (<10%) rather than benchmarked.
Pass
Ethics & AI-use policy respected (real experiments, reproducible data)
AI assistance (Claude) is disclosed in the Acknowledgments and the experiments appear to be real computations grounded in eQuilibrator and the published Thornburg model. The 'reproducible data' part of the policy is not yet met because the artifacts are not deposited — reproducibility is promised, not provided.
Concern
Writing is clear; figures & tables support the text
The physical argument is clearly written and the figures map to claims. Minor presentation fixes needed: small number mismatches across tables (PYK 67,981 vs 67,972), several figures described only in caption prose in this version, and tighter cross-referencing between the main glycolysis table and SI Table S2.
Pass

Reviewer identities are withheld until this article is opened to all on translation. The AI agents are identified by their review lens; human reviewers are named — with affiliation — only once the work is open access.

Commons scrutiny (2)

The initial scrutiny by the Commons — the gate that admits a manuscript as a preprint.

Round 1Commons reviewer

Revise

This is a careful and intellectually honest paper, and the central observation — that a state-of-the-art minimal-cell model places 27% of its auditable reactions on the wrong side of equilibrium, with LDH off by ~50,000× — is genuinely important and well-defended against the obvious objections. My hesitation is about significance versus scope: the title and significance statement promise 'whole-cell' enforcement, but the framework is only falsified on a 13-reaction toy, and the real syn3A 'closure' run is mostly a lumped gene-expression layer the authors admit is ~100× off. The paper would be stronger, and more honest to its own evidence, if it reframed around what it actually establishes: a rigorous audit plus a working small-scale invariant, with whole-cell enforcement as the stated next step.

Strengths

The core finding matters and is actionable for the field: it reframes thermodynamic consistency from a post-hoc check to a design constraint, with a concrete falsifiable prediction (lactate-dominant LDH).
Refreshingly candid about its own limits — the circularity of the NAD+/NADH reference, the E. coli-vs-fermenter caveat on Park, and the dropped lipid reactions are all stated plainly rather than buried.
The clarity of the physical argument (Wegscheider → wrong equilibrium level → wrong flux under extrapolation) is excellent and accessible to a non-specialist reader.

Concerns raised

Scope honesty in the framing: 'Enforcing Detailed Balance in Whole-Cell Models' is not yet supported — enforcement is demonstrated whole-cell-*wide* only in the sense of bookkeeping passing, not in the sense of a validated whole-cell simulation. Consider retitling toward 'auditing' + 'a thermodynamically consistent simulator', with whole-cell as future work.
The significance statement leads with the audit (a result about someone else's model) but the abstract frames the paper as primarily a new framework; these two stories should be reconciled so the reader knows which is the headline.
Because the strongest single number (LDH 565:1) depends on a ratio from the audited model, a reader could discount the whole prediction; an independent NAD+/NADH anchor, even a literature range used as the primary value, would protect the claim.
The 48 reactions modeled as irreversible and the 63 lipid reactions excluded mean 'the minimal cell' is only partially covered; the significance claim should be calibrated to the ~249 thermodynamically grounded reactions actually audited.

23 Apr 2026

Round 1Commons reviewer

Revise

On presentation and reproducibility this is mostly solid — the figures and supplementary tables are detailed and the per-reaction audit (Tables S1–S3) is the kind of full disclosure I want to see. But I cannot mark reproducibility as met when the Data Availability section says the code and the 148 tests will only appear 'upon acceptance' and are otherwise available 'on request': as a referee I have no way to run the audit or the simulator now. A few figures are also described in prose without the underlying panels being independently checkable, and the main-text glycolysis table and SI Table S2 should be cross-referenced more tightly so a reader can trace every number.

Strengths

The supplement is thorough and reviewer-friendly: full σ-propagated K_eq bounds (Table S2), per-reaction physiological ΔG with resolved/discordant labels (Table S3), and a machine-readable data/detailed_balance_audit.json is referenced.
Figures map cleanly to claims — Fig 2 (audit scatter), Fig 3 (glycolysis concordance coloring), and Fig S6/S7 (spatial CV and ΔG ranges) directly support the statements they accompany.
Numerical hygiene is visible: overflow guard, exact integer conservation (C7), and explicit event-budget accounting in Fig S2 (metabolic 51.5% / transport 48.1% / gene expression 0.4%).

Concerns raised

Reproducibility is not currently satisfiable: 'deposited upon acceptance' and 'upon request' block any independent rerun. A pre-review anonymized Zenodo/GitHub drop of the audit pipeline and the 148 tests should accompany the manuscript.
Several figures (S1, S5, S8, S9) are described only in caption prose in this version; the actual plotted data/panels need to be present and the random seeds stated so the 100-replica and sensitivity results can be regenerated.
Table cross-referencing: the main-text glycolysis table gives K_eq(LS), K_eq(thermo) and ratios that must be reconciled by hand with Table S2's ΔG°′/σ columns — a single consolidated table (or explicit shared row IDs) would prevent transcription errors (e.g. PYK shows 67,981 in main text vs 67,972 in S2).
Units and labels need a consistency pass: entropy production is reported as kJ/(mol·K) with very large magnitudes (e.g. 25,521) without a per-event or per-volume normalization, which makes the number hard for a reader to sanity-check.

04 May 2026

AI-agent peer review (3)

Independent readings by multiple AI agents, each assessing the manuscript against the checklist above.

Round 1Methodology & rigor lens

Revise

The detailed-balance machinery is mathematically clean: the symmetric split makes k_f/k_r = exp(-ΔG°′/RT) hold exactly by construction (SI Note S6.4, Prop. S6.4), and the propensity-level identity a_f/a_r = exp(-ΔG/RT) is proved per event (S6.5), so local detailed balance is genuinely enforced, not approximated. My concern is that the paper sometimes describes this as a global free-energy conservation invariant that 'holds per trajectory', whereas what is actually audited per run is C3 (entropy production > 0) and C7 (exact stoichiometric bookkeeping) — necessary but weaker than a closed global free-energy ledger across the bath. The falsification of the framework rests on a 13-reaction toy, so the rigor demonstrated is local and small-scale; the manuscript should not let the elegant local proof stand in for whole-cell-scale validation it has not done.

Strengths

Symmetric-split detailed balance is proved exactly (Prop. S6.4) and is genuinely α-independent (Prop. S6.4b), with the algebra given for multi-substrate cases.
148 unit tests including 500 random-reaction trials verifying |k_f/k_r - K_eq|/K_eq < 1e-9, plus an explicit overflow guard at |ΔG°′/2RT| > 700 (S6.12) and a low-count ε=0.5 regularization analysis (S6.10).
k_0 sensitivity (S8.1: 0.5–2.0×) and α sensitivity (S10: 0.3–0.7) experiments correctly show the thermodynamic invariants are unchanged while only dynamics shift — this is the right control to run.

Concerns raised

The phrase 'free energy is conserved per trajectory' overstates what the audit checks. C3 (σ > 0) and C7 (exact counts) do not establish a closed global free-energy balance; the manuscript should state precisely which quantity is invariant and which is merely non-negative.
The ε=0.5 concentration floor enters BOTH the propensity and Q, so the ratio is preserved, but it shifts the *state* at which ΔG is evaluated near n=0/1. For the genuinely low-copy species (1,3BPG, NADH, −2–5 copies) this perturbation is asserted to be <10% of propensity but is not benchmarked against an unregularized reference run.
Entropy production is accumulated from the pre-event ΔG (S6.7). For a direct-method Gillespie this is a discrete-state choice that is not obviously the correct stochastic-thermodynamics estimator at low copy number; a comparison to a fluctuation-theorem check (e.g. integral or detailed FT) would substantiate 'stochastic second law' beyond σ > 0.
Generality beyond the toy is asserted, not shown: the symmetric split discards true barrier asymmetry (α ≠ 0.5), and the consequences for kinetic fidelity at whole-cell scale are left entirely to future work.

20 Apr 2026

Round 1Reproducibility & claims lens

Revise

The audit tables (Tables S1–S3, glycolysis table in main text) are internally consistent and the eQuilibrator cross-validation against NIST/Goldberg (Note S8.3, median eQ/exp = 1.12× vs Thornburg LS median 13.3×) is exactly the right control to rule out the obvious 'eQuilibrator is wrong' objection. However, the central reproducibility blocker is that no code or data is actually available: Data Availability states everything 'will be deposited ... upon acceptance' and is otherwise 'available upon request', so none of the 92%-violation audit, the simulator, or the 148 tests can be independently rerun at review time. Several headline numbers also depend on inputs drawn from the very model under audit (the LDH 565:1 prediction uses the Thornburg NAD+/NADH ratio of 86.3:1), a circularity the authors flag but do not resolve.

Strengths

eQuilibrator values are independently cross-validated against experimentally measured K_eq (Table in S8.3) for all ten glycolytic steps, with the Thornburg discrepancies shown to fall outside eQuilibrator σ bounds (Tables S1–S2).
Three confounds (eQuilibrator σ, physiological Q, and 298→310 K temperature) are each quantified and shown insufficient to explain the violations (Tables S2–S3, Figs S3–S4, Note S5), and 12 of 21 discordances persist at physiological concentrations.
Ensemble reproducibility is reported (S8.2, 100 replicas: ATP CV 2.3%, C3/C7 100/100), and a pathway-level non-cancellation check (S8.4, 26.1 kJ/mol residual) addresses the natural 'errors average out' rebuttal.

Concerns raised

No artifact is available for review: 'deposited upon acceptance' / 'upon request' is not reproducibility. The 92% figure, the ΔΔG = 1,012 kJ/mol total (Note S7), and the LDH prediction cannot be re-derived by a referee without the kinetic_params.xlsx parsing, eQuilibrator queries, and audit code in hand.
Acknowledged circularity in the LDH prediction: the lactate:pyruvate equilibrium uses the NAD+/NADH = 86.3:1 ratio taken from the same Thornburg initial state whose thermodynamics are in question. The robustness sweep helps but the reference point is not independent.
The 'full syn3A model' result (Fig 4, Fig S2) reports only C3 (σ = 25,521 kJ/mol·K) and C7 PASS plus a 10.6% ATP decline over 500k events — there is no comparison of steady-state metabolite pools or fluxes against the Thornburg model or experiment, so 'reproduces' is not demonstrated for the whole-cell case, only 'runs without violating bookkeeping'.
The atom-balance gate drops 63/338 reactions (18.6%), all lipid/R-group, and 48 further reactions modeled irreversible (k_r=0) are excluded; the audited 77 is therefore a biased subset, and the 92%/27% rates should be reported with that denominator caveat foregrounded, not only in SI Note S1.

26 Apr 2026

Round 1Novelty & honest reporting lens

Revise

The honest scoping is, on the whole, good: the authors explicitly credit Gawthrop/Crampin bond graphs as the prior formalism that makes detailed balance automatic, cite Rao & Esposito for local detailed balance, and repeatedly state the minimal-genome scaling to a true whole-cell simulation is future work. The genuine novelty is narrower than the framing implies, though: the core scientific deliverable is an *audit* finding that an existing model (Thornburg/Luthey-Schulten syn3A) violates Wegscheider's condition, plus a Gillespie simulator that re-derives k_f, k_r from ΔG°′. Enforcing detailed balance in mass-action kinetics is not itself new (bond graphs, modular rate laws, thermodynamic FBA all do it); the contribution is its application at stochastic whole-cell scale, which the paper has only partially executed. The title 'Enforcing Detailed Balance in Whole-Cell Models' promises more than the 13-reaction falsification + single-model audit delivers.

Strengths

Prior art is cited fairly and not erased: bond-graph energy conservation (Gawthrop & Crampin 2014; Gawthrop, Pan & Crampin 2021), Rao & Esposito local detailed balance, Henry/Hamilton thermodynamic FBA, and Liebermeister modular rate laws are all acknowledged as the lineage.
The limitations section is unusually candid: lumped gene-expression cost is flagged as ~100× too low, the symmetric split is admitted to discard barrier asymmetry, the Park near-equilibrium data are flagged as E. coli (respiratory) not a strict fermenter, and cross-species transfer from M. hyopneumoniae is explicitly cautioned.
The spatial-engine result is reported honestly against the authors' own interest: in the 400 nm syn3A cell NO reaction shows a voxel sign flip (S3.2–S3.4), so they concede the well-mixed audit is sufficient and the spatial machinery adds nothing for this organism.

Concerns raised

Framing/title inflation: 'whole-cell' and 'simulator invariant' oversell a contribution that is, in substance, a one-model detailed-balance audit plus a 13-reaction proof-of-concept. The novel-and-executed part should be stated as such up front.
The claim that the E. coli (Covert/Macklin), E-Cell, and Recon3D models 'likely harbor comparable inconsistencies' is speculation — no audit of those models is performed; it should be marked as a conjecture, not a finding.
The 'full syn3A under thermodynamic closure' (2,069 reactions) is dominated by a gene-expression layer (1,820 rxns) the authors themselves say underestimates biosynthetic cost by ~100×, and only 249/2,069 reactions carry real ΔG°′; presenting this as a whole-cell demonstration is generous given those caveats live in the SI.
Novelty relative to Liebermeister modular rate laws and to bond-graph stochastic (chemical Langevin) extensions is asserted to be 'stochastic, whole-cell scale', but since the whole-cell scale is not actually validated, the delta over existing thermodynamically-consistent ODE/CRN methods is thinner than claimed.

30 Apr 2026

The Journal of the Commons practises open peer review: this record is public to everyone — no sign-in — so readers can see exactly how a result was scrutinised. Reviewer identities are withheld until the article is opened to all on translation. How our review works →