FluxDisco: Symbolic Regression for Stoichiometric Dynamical Systems via Monte Carlo Graph Search
Cassandra Durr, Alvaro Köhn-Luque, Chris Jewell, Lloyd A. C. Chapman
stat.ML · cs.LG · physics.data-an
Quick verdict
Claims first stoichiometry-aware SR allowing non-polynomial fluxes; convincing on 2-3 state benchmarks, but scaling and baseline-fairness concerns remain.
Summary
FluxDisco is a dynamical symbolic regression framework for flux-based stoichiometric ODE systems (ẋ = Sv(x)). It adapts Monte Carlo Graph Search: state merging via constant folding, a selection rule that never scores below the best sampled reward, Beta credible intervals replacing the conservative BKL bounds, and force-completion of expressions at a depth cap κ to keep rewards dense. Rewards combine a parsimony penalty η^C with exp(−MSE/τ) after least-squares constant optimisation and RK4 integration. On five simulated benchmarks (three SIR variants, Lotka-Volterra, stable/unstable Brusselator, Fairen-Velarde) across three noise levels, FluxDisco achieves trajectory NMSE below 2% with high coupled-term identification, while baselines (SPL, SINDy, PySR, ProGED, ODEFormer) fall below the naive mean-prediction baseline in some experiments; the Fairen-Velarde system is identified by no method. Ablations support state merging and grammar exclusions, but runtime exceeds alternatives by orders of magnitude. Key limitations: only 2–3 state systems with a single realisation each (no repetition-based uncertainty), per-system hyperparameter tuning of their own method (Fairen-Velarde: κ=12, τ=0.05) against baselines run at default settings, and domain-knowledge grammar exclusions that baselines cannot access. Code is released on GitHub.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Imposing a known stoichiometry improves both numerical accuracy and coupled-term identification compared to running the same search without it (ablation, Section 3.2).
- FluxDisco is, to the authors' knowledge, the first stoichiometric equation-discovery method supporting non-polynomial fluxes without mass-action kinetics (Section 1.2).
- FluxDisco achieves trajectory NMSE below 2% on all five benchmark systems across three noise levels, while general baselines (SPL, SINDy, PySR, ProGED, ODEFormer) exceed the naive mean baseline in some experiments.
- Modified MCGS components (Beta credible intervals, max-reward selection, state merging, grammar exclusions) each improve search efficiency, with gains diminishing as the computational budget grows (S4.1, S4.2).
Strengths
Open methodological niche
First (per authors, Section 1.2) stoichiometry-aware equation discovery supporting non-polynomial fluxes, contrasting with Reactive SINDy, SISR and KinFormer which are confined to mass-action polynomials; the square-root SIR and rational Fairen-Velarde benchmarks exercise this capability directly.
Well-argued algorithmic adaptations
Each MCGS modification is motivated: the max(u, r*) selection rule (Eq. 6) prevents soft-pruning of nodes with one excellent rollout; force-completion at depth κ (Section 2.1.2) keeps rewards dense versus SPL's zero-reward; episode-level propagation (2.1.5) cuts update cost; the stochastic-reward choice is justified in S1.1.
Transparent negative results and biases
Meaningful ablations
Reproducibility scaffolding
Weaknesses
No repetition-based uncertainty
Asymmetric evaluation protocol
Untested scalability
Heuristic interval construction
Under-specified efficiency comparison
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FluxDisco targets dynamical symbolic regression for a restricted but practically important model class: flux-based stoichiometric systems ẋ = Sv(x), where the stoichiometry matrix S is known a priori and the flux expressions v(x) must be discovered jointly from trajectory data. The authors position the work against the three existing stoichiometry-aware methods — Reactive SINDy [10,11], SISR [14], and KinFormer [15] — all of which assume mass-action kinetics and hence polynomial fluxes; the stated novelty (Section 1.2) is supporting non-polynomial fluxes (square-root SIR, rational Fairen-Velarde terms) under a known stoichiometry. Central claims: (1) imposing known stoichiometry improves accuracy and coupled-term identification versus the same search without it; (2) the adapted MCGS components (Beta credible intervals, max-reward selection, state merging, grammar exclusions) improve search efficiency; (3) FluxDisco recovers dynamics across five benchmark systems under three noise levels where general-purpose baselines frequently fail. Claim (1) is directly supported by the 'Ours (No Stoich.)' ablation (Section 3.2). Claim (3) is supported numerically with one important self-reported exception: Fairen-Velarde, identified by no method including theirs.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does FluxDisco solve?
It discovers the governing equations of ODE systems whose dynamics have flux form ẋ = Sv(x), where the stoichiometry matrix S is known but the flux expressions v(x) are unknown. Unlike earlier stoichiometric methods (Reactive SINDy, SISR, KinFormer), it does not require mass-action kinetics, so non-polynomial fluxes such as square-root or rational terms can in principle be recovered.
How is FluxDisco different from standard symbolic regression tools like PySR or SINDy?
It imposes the known stoichiometry so that fluxes are discovered jointly and the resulting equations automatically respect the reaction network structure. SINDy selects terms from a fixed polynomial library, and PySR, ProGED and similar tools search equations without enforcing flux coupling. The paper's ablation shows that imposing stoichiometry improves both accuracy and identification of coupled dynamics.
What are the main limitations of FluxDisco?
It was tested only on 2–3 state systems; joint flux discovery scales poorly because a single search node must contain the correct expressions for all fluxes; the stoichiometry must be exactly known (misspecification degrades estimates); runtimes exceed baselines by orders of magnitude; and the hardest benchmark (Fairen-Velarde) was not identified by any method, including FluxDisco itself.
Is code and data available for FluxDisco?
The paper links a public GitHub repository (github.com/CassandraDurr/FluxDisco). All benchmarks are synthetic, and the hyperparameters plus data-simulation settings are documented in Tables S1–S2 with full algorithms in the supplementary material, so experiments can be re-run on ordinary CPUs.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 4 and 6-10 present results only as plots; exact per-method NMSE, runtime and efficiency values cannot be read from the extracted text, so baseline-level quantitative comparisons rest on captions and prose.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.