acceptodds
Under review as a conference paper at ICLR 2027

SysBio-SR: Benchmarking Symbolic Regression for Trajectory Fidelity and Mechanism Recovery in Systems Biology

Abstract

Symbolic regression seeks to discover interpretable governing equations for dynamical systems from data. In systems biology, however, accurately reproducing system trajectories does not necessarily imply recovery of the underlying reaction mechanisms. Existing benchmarks evaluate numerical accuracy, expression complexity, or equation-level similarity, but not agreement between recovered reactions and reference mechanisms in biological systems. We introduce SysBio-SR, a benchmark of 198 literature-based biological ODE systems with reaction annotations spanning mass action, Michaelis–Menten, Hill, and complex kinetics. Trajectory fidelity is evaluated across all systems under in-distribution reconstruction and out-of-distribution (OOD) extrapolation, while mechanism recovery is assessed on the subset of 161 systems. Of 11 state-of-the-art symbolic-discovery methods, only 3 support mechanism recovery, as most output unconstrained expressions that cannot map to biological mechanisms. Joint analysis across systems highlights how severely trajectory-based evaluation overestimates true mechanism recovery. Motivated by these findings, we develop KineSelect, a framework that searches candidate reaction topologies under biologically grounded kinetic laws to select optimal mechanisms that balance trajectory fidelity with structural parsimony. Compared with the strongest baseline, KineSelect improves aggregate OOD extrapolation by 19.2% and overall reaction F1 score by 14.3%.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mechanism Recovery, Symbolic Regression, Dynamical Systems, Systems Biology, Scientific Machine Learning, Equation Discovery, Benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5E79OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies mechanism recovery in symbolic regression for systems biology and argues that trajectory fidelity alone may not adequately reflect the ability to recover the underlying mechanisms. The authors introduce the SysBio-SR benchmark and further propose KineSelect, which jointly considers trajectory fitting and mechanism identification. Experimental results show that KineSelect generally outperforms the compared methods in both trajectory prediction and reaction recovery.

Critical strengths
  1. The benchmark focuses on an important but often underappreciated issue: predictive accuracy is not equivalent to mechanistic correctness.

  2. The experimental evaluation is broad and includes informative failure-case analyses.

Critical weaknesses
  1. The novelty and methodological positioning of KineSelect are not sufficiently established with respect to recent structure-aware symbolic regression methods. Recent works such as RSSP and StruSR (see the next comment for specific references) have explicitly investigated how physics-informed and structural information can be incorporated into symbolic search. Although these methods target settings that are not identical to biological reaction-mechanism recovery, they are conceptually closely related to the central motivation of KineSelect: constraining symbolic search with scientifically meaningful structural information rather than relying solely on numerical fitting.

  2. The related-work coverage and methodological positioning appear incomplete, particularly with respect to recent structure-aware and physics-informed symbolic regression methods. The manuscript discusses sparse regression, evolutionary symbolic regression, neural symbolic approaches, and reaction-structured methods, but largely positions KineSelect only against conventional equation-discovery or reaction-structured baselines. Recent works such as Symbolic Regression with Physics-Informed Residual and Structural Sensitivity Pruning and StruSR: Structure-Aware Symbolic Regression with Physics-Informed Taylor Guidance explicitly investigate how structural and physical information can be incorporated into symbolic search rather than relying solely on numerical fitting. RSSP evaluates individual symbolic substructures using residual and structural sensitivities and uses these signals to guide evolutionary operators and pruning, while StruSR extracts Taylor-based structural priors from a trained surrogate and uses structural attribution to guide symbolic evolution. Although these methods address somewhat different problem settings from biological ODE mechanism recovery, they are conceptually close to the central motivation of KineSelect: constraining symbolic search with scientifically meaningful structural information instead of treating trajectory fit as the sole criterion. The absence of any discussion makes it difficult to assess which aspects of KineSelect are genuinely specific to biological mechanism recovery and which belong to the broader emerging paradigm of structure-aware, physics-informed symbolic regression. I therefore recommend that the authors substantially expand the Related Work section to discuss these methods and provide a mechanism-level comparison, clarifying differences in task definition, prior information, search representation, structural guidance, and evaluation objectives. If a direct empirical comparison is technically infeasible because RSSP and StruSR target different symbolic-discovery settings, this should be stated explicitly and justified; otherwise, including them or suitably adapted variants as baselines would considerably strengthen the experimental positioning.

  3. The reference reaction annotations appear to be reconstructed from ODE expressions rather than taken directly as independent biological ground truth. This raises an important question for me: does the “reaction mechanism ground truth” in SysBio-SR come directly from the original SBML reaction networks, or is it inferred from the algebraic form of the ODEs? This distinction is important because an algebraic decomposition of an ODE is not necessarily equivalent to an independently specified biological mechanism.

  4. The manuscript would benefit from further improvement in its overall organization and argumentation. The paper simultaneously emphasizes benchmark construction, the gap between trajectory fidelity and mechanism recovery, and the design of KineSelect, but the logical relationship among these three parts is not sufficiently tight. The correspondence between the problems introduced in the Introduction and the three components of KineSelect is not fully clear. In addition, the Related Work section remains somewhat method-listing-oriented and lacks a clear problem-driven progression. The coverage of related work is also incomplete, and I recommend incorporating the recent studies mentioned above.

Questions for authors

1.How do the authors position KineSelect relative to recent structure-aware and physics-informed symbolic regression methods such as RSSP and StruSR? In particular, which components of KineSelect are specific to biological reaction-mechanism recovery, and which belong to the more general paradigm of structure-guided symbolic search?

  1. The search procedure is stochastic, but the main experiments appear to use only a single base seed, and the noise experiments appear to use only one noise realization per noise level. Could the authors provide results over multiple random seeds and multiple noise realizations? This would help assess whether the reported OOD and reaction-F1 improvements are stable.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 5E79 — Methodology and Annotations

Thank you for recognizing our focus on the gap between trajectory prediction accuracy and mechanistic agreement, and the value of the benchmark experiments and failure-case analyses.

1. Methodological contributions of KineSelect (W1, Q1)

KineSelect connects reaction structure search with explicit regulatory roles, kinetic form identification conditional on reaction structure, and model selection that balances fit and parsimony. These designs give search candidates explicit reaction semantics and preserve each reaction's shared contribution to different state equations during fitting.

The three components address reaction connectivity, rate forms, and final model selection, respectively:

  • Reaction Structure Search (RSS): searching reaction topology and species roles. Building on SISR's reaction-level evolutionary representation, we introduce an explicit regulator role so that a species can affect a reaction rate without net consumption or production.
  • Kinetic Form Identification (KFI): selecting kinetic forms for a given reaction structure. Reaction connectivity alone does not determine the rate function. KFI includes kinetic forms in identification, constructing a candidate library with MA, Michaelis-Menten (MM), and Hill regulatory forms.
  • Fidelity-Parsimony Model Selection (FPMS): balancing fit and complexity to select the final model. For candidates with different fitting errors and structural complexities, FPMS jointly considers derivative error, trajectory error, and expression complexity. Final selection therefore reflects both fit and parsimony.

Evolutionary search, regression fitting, and balancing fit against complexity are general methodological ideas. Our specific designs for biological reaction dynamics concern species roles, coupling at the reaction level, kinetic-family identification conditional on structure, and a dedicated measure of kinetic complexity. Section 5 details these components.

2. Positioning relative to RSSP and StruSR (W1, W2, Q1)

In their primary partial differential equation (PDE) settings, RSSP and StruSR seek analytical solutions to known governing equations. Their physics-informed neural networks (PINNs) provide Taylor guidance, and subtree sensitivities guide search over expression trees. KineSelect identifies unknown ordinary differential equation (ODE) vector fields from trajectories, using reaction roles and MA, MM, and Hill families as priors. It searches reaction structures and kinetic forms while sharing each reaction's rate across the affected state equations.

Both works also report general SR experiments. RSSP allows data-fitting loss and surrogate-derived structural guidance without a known PDE. Such surrogate-derived priors describe local function behavior. KineSelect instead encodes reaction-level structure through species roles, stoichiometric coupling, and admissible MA, MM, and Hill kinetics.

We have not evaluated RSSP or StruSR on SysBio-SR. Their primary PDE formulations recover analytical solutions to known equations, whereas our task identifies unknown ODE vector fields from trajectories.

We have added both works and clarified their relationship to KineSelect in Section 2.3 (lines 150).

3. Origin and interpretation of the reference reaction annotations (W3)

We construct SysBio-SR's reference reaction annotations from the standardized ODEs of the BioModels source models. These ODEs define the dynamics that generate the benchmark trajectories, so we use their structure as the reference for equation discovery. We encode consumption/production directions, species dependencies, and kinetic forms in the same state representation used by the evaluated methods. This enables structural checks beyond trajectory accuracy.

The annotations provide ODE-derived structural references, not independent biological ground truth for the original SBML reaction network.

As detailed in §3.3 (lines 218) and Appendix A.5, we verify the annotations through symbolic and numerical reconstruction and manual consistency review.

AuthorsOpenReview ↗

Response to Reviewer 5E79 — Experimental Robustness and Organization

4. Stability across search seeds and noise seeds (Q2)

We have aggregated four clean-data runs, including the original paper's results, and expanded the noise evaluation to three noise seeds.

Results across search seeds with clean training data. The new experiments use search seeds 53, 25, and 81, with the same designated clean training trajectory in each run and OOD evaluation. The table summarizes four runs. We have added these results in Appendix C.6 (Tables 13-15; lines 1836).

Method and selectorOOD R²-score ↑Reaction F1 ↑
SISR / Alpha0.440 ± 0.0150.387 ± 0.008
SISR / FPMS0.386 ± 0.0370.430 ± 0.006
KineSelect / FPMS0.531 ± 0.0250.454 ± 0.007

Across all four runs, KineSelect/FPMS achieves higher OOD R²-score and reaction F1 than both SISR/Alpha and SISR/FPMS, consistent with our original findings.

Robustness across noise seeds. We evaluate KineSelect with its default FPMS selector, SISR with its original Alpha selector, and SISR with FPMS for a shared-selector comparison. For each noise level, we fit and evaluate models using different noise seeds. We have added the results in Appendix C.4.1 (Tables 10-11).

OOD R²-score ↑ (columns indicate noise levels)

Method and selector0.0010.010.050.1
SISR / Alpha0.436 ± 0.0160.426 ± 0.0180.417 ± 0.0060.376 ± 0.031
SISR / FPMS0.369 ± 0.0240.352 ± 0.0420.340 ± 0.0540.278 ± 0.061
KineSelect / FPMS0.532 ± 0.0010.491 ± 0.0080.419 ± 0.0040.363 ± 0.028

Reaction F1 ↑ (columns indicate noise levels)

Method and selector0.0010.010.050.1
SISR / Alpha0.388 ± 0.0150.386 ± 0.0100.362 ± 0.0120.318 ± 0.001
SISR / FPMS0.420 ± 0.0140.397 ± 0.0120.350 ± 0.0160.306 ± 0.028
KineSelect / FPMS0.454 ± 0.0010.423 ± 0.0040.354 ± 0.0130.309 ± 0.004

The results across noise seeds retain KineSelect/FPMS's advantage over SISR/Alpha on both metrics at 0.1% and 1% noise. Under shared FPMS, KineSelect has higher mean OOD R²-score than SISR at all four noise levels, while their reaction F1 means are close at 5% and 10% noise.

5. Connecting the benchmark, empirical findings, and method design (W4)

SysBio-SR combines trajectory accuracy and reaction-signature agreement in a common evaluation through standardized trajectories and reference reaction annotations. The empirical analysis shows the gap between these measures, motivating a methodological question: how can reaction priors and model selection improve recovery of reaction structure while maintaining trajectory prediction accuracy? KineSelect explores this question through RSS, KFI, and FPMS, which address candidate structure, rate forms, and final model selection. We evaluate the resulting models using both trajectory prediction accuracy and reaction-signature agreement.

We have clarified this connection in the Introduction and added a discussion of structural and physical priors in Section 2.3 (lines 150).

Reviewer_sKxaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors list three main contributions:

  1. They propose a new benchmark, called SysBio-SR, to evaluate recovery of ground truth ODE equations from trajectory data of mathematical models of biological systems
  2. They propose a new symbolic regression method, called KineSect, and show that it outperforms baseline approaches according to two metrics: trajectory fidelity (R²) and mechanism recovery (reaction F1 score)
  3. They “expose a fundamental limitation of symbolic regression in systems biology: high trajectory fidelity drastically overestimates real-world utility by failing to recover ground-truth mechanisms”

I am skeptical about all three contributions:

  1. While I agree it is interesting to benchmark symbolic regression approaches on biological systems (not well studied so far, see Figure 2-a), the systems used are already freely and openly available on BioModels: it is essentially a subset class of BioModels (deterministic systems with at most five variables, excluding systems whose dynamics depend on external events and current time t). I acknowledge authors efforts in curating and standardizing these subclass of BioModels in a specific format that enable their study, but I think that presenting it as “the first reaction-annotated benchmark suite enabling mechanism recovery evaluation for biological ODE discovery” should largely be attributed to the original BioModels database.
  2. Symbolic regression (SR) aims to find a symbolic mathematical expression (here ODE equations) that model data (here data points from trajectories of the unknown biosystem). The more little data you have, the more you are likely to overfit and not recover ground truth equations unless you have good priors baked into your models. As authors use very limited data, I assume that the method should propose some useful priors for biological ODE recovery, but it is unclear for me after reading the paper what are the key contributions in that regard (in particular as the proposed method seems very close to the baseline SISR). Moreover I have concerns on the methodology/evaluation and reported results (see next section).
  3. This is not a contribution but common knowledge: you train your model on a single trajectory of the dynamical system governed by ODE equations and show that trajectory recovery does not guarantee recovery of the ground truth ODE equations. This is called overfitting and a highly expected result given the under-defined nature of the problem you are trying to solve; not a main contribution.
Critical strengths

The authors put substantial efforts into curating the BioModels database and in benchmarking previous approaches (a total of 11 methods) on the proposed dataset.

There might also be some interesting components in the proposed KineSelect method, but as it remains unclear (see next section) I am not inclined in supporting acceptance.

Critical weaknesses
  1. Methodology and Evaluation: Authors use a single trajectory for training and a single trajectory for OOD testing (from a random perturbation of the default initial state). For training, this looks to me like a very under-defined problem unless you make assumptions about the class of dynamical systems or have some prior knowledge baked-in, and it is unclear after reading the paper where authors stand with respect to that matter. For evaluation, this also seems like a very brittle way to evaluate OOD trajectory recovery that is highly sensitive to the choice of the (only) tested initial condition. Moreover no statistical significance of the reported results are provided, a single seed is used both for training and evaluation. It is also unclear whether equivalent hyperparameter search was performed between the proposed KineSelect and baseline SISR. Authors are claiming state-of-the-art but the methodology for evaluation is questionable in my opinion.

  2. Contributions: it is not clear for me after reading the paper what are the important contributions of this work. I suggest authors think more about what it is that they want to contribute to. Whether a benchmark or approach for biological ODE recovery, it would be more informative to better understand the nature of the problem in biological ODE recovery in comparison to doing symbolic regression in other domains, what type of data do we typically have, why may we use stoichiometrically-informed versus foundation model based versus other types of approaches for this problem, and where does this paper stands in that landscape.

Questions for authors
  • Why use a single trajectory for training? For me this makes sense when the system is ergodic (e.g. Lorenz attractor) and a single trajectory can visit enough of the state space to provide a proper training dataset. This is not the case for BioModels. I assume in real-world biology it might be difficult to get data from varied initial conditions but if that is the case the problem is fundamentally under-defined unless you have strong inductive biases.
  • The inductive biases in your approach seem to be relatively efficient to recover mass-action kinetics but not for other families of models, why is this the case? Isn’t penalizing non-mass-action kinetic terms problematic to generalize to more complex families of biological systems?
  • Can you state more clearly what ingredients in your framework differ with respect to the SISR baseline and why do you believe these are important ingredients for biological ODE recovery?
  • Why do you believe that this kind of framework is suited to perform SR in biological systems with respect e.g. to foundation models that could incorporate prior knowledge about biological systems?
  • What is the F1 score reported in Figure 1-c? This is not explained and confusing for the reader as it seems on that plot that the stronger baseline is Reactive SINDy and not SISR.
  • Can you report better analysis of OOD generalization? Rather than looking at a single trajectory from a slightly perturbed initial state I would expect testing on a large number of initial conditions that have a higher chance of properly covering the state space of the dynamical system. While this is not easy as we do not know the reachable space a priori, one could at least sample N points in e.g. the hypercube [x_d^{min}, x_d^{max}] where min and max values are computed for the default trajectories with initial conditions provided in BioModels.
3 comments
AuthorsOpenReview ↗

Response to Reviewer sKxa — Summary and Contributions (Part 1/3)

Thank you for recognizing the value of the model curation and baseline evaluation, and for identifying the experiments and methodological explanations that need strengthening. Following your suggestion for OOD evaluation, we have completed additional tests on all 198 systems, with 20 new initial conditions per system and three search seeds. KineSelect's default configuration improves mean OOD R²-score over SISR's main configuration by 0.078, with a paired system-level 95% bootstrap interval of [0.036, 0.123]. We follow the order of your comments below to explain the benchmark design, method differences, and new evidence.

S1. What SysBio-SR adds to BioModels

SysBio-SR organizes literature models into a biological ODE discovery task with common observation inputs, reference representations, and scoring rules. BioModels provides the original biological models and reaction information, and we retain their provenance and attribution. Building on BioModels and SysBio-Traj, we provide three complementary elements.

  • Standardized, executable reference systems. The 198 systems contain 661 ODEs, with 1 to 5 state variables per system. We standardize state ordering and equation representation, and provide Python reference implementations and a simulation interface with configurable initial conditions and parameters. This allows controlled experiments using the same reference system.
  • Explicitly designed ID/OOD trajectory tasks. Each system provides a 512-point trajectory from its default initial condition for fitting. We change the initial condition to construct an OOD trajectory. We perturb initial conditions according to state scales and check integration validity, normalized trajectory differences, and amplitude bounds. We reject nearly coincident or numerically extreme candidates. The additional experiment in W1 tests the effect of using multiple initial conditions.
  • Reference term labels and evaluation aligned with standardized ODEs. For example, Tyson1991’s SBML file lists the reaction z→uz \rightarrow u, with rate k4z(k4′/k4+u2)k_4 z(k'_4/k_4 + u^2), and defines z=v−uz = v-u through an assignment rule. Rate rules govern the independent states uu and vv, so the corresponding contribution to u˙\dot{u} is k4(v−u)(k4′k4+u2)k_4(v-u)\left(\frac{k'_4}{k_4}+u^2\right). We therefore construct reference reaction signatures from the standardized ODEs, encoding consumption/production directions, species dependencies, and kinetic forms in a common representation for structural evaluation. We check consistency through symbolic and numerical reconstruction and manual review.

Trajectory evaluation covers all 198 systems, and projected reaction-signature evaluation covers the 161 Main/Challenge systems. The remaining 37 Stress systems retain reference annotations but contain kinetic forms outside the matching protocol. Sections 3.2 and 3.3 (lines 187 and 218) describe the source models, standardized trajectory task, and reference annotations. Appendices A.3 and A.5 give the trajectory settings and annotation construction and matching rules.

S2. Biological inductive biases in KineSelect

KineSelect makes reaction connectivity and species roles, reaction rate forms, and final selection of a complete model explicit parts of identification. Building on SISR's reaction-level evolutionary representation, Reaction Structure Search (RSS) introduces an explicit regulator role. Kinetic Form Identification (KFI) compares mass-action (MA), Michaelis-Menten (MM), and Hill candidates for a given reaction structure. Fidelity-Parsimony Model Selection (FPMS) balances fit and complexity. Q3 gives a component-wise comparison, and W1 reports results across search seeds with a shared final selector. We use this pipeline to test designs for biological reaction representations and their trade-offs.

S3. The empirical contribution of the trajectory/structure gap

We quantify how this gap affects method evaluation and model selection. On a common set of systems, we compare trajectory prediction with reaction agreement and analyze kinetic families and selection strategies.

For example, in Table 5, 18 of the 65 systems where KineSelect achieves OOD R² ≥ 0.90 still have reaction F1 below 0.50 (27.7%). The corresponding count for SISR is 17/56 (30.4%). High OOD trajectory accuracy can therefore coexist with low reaction F1. Evaluating both axes exposes structural differences that trajectory scores miss. Figure 9 further shows how selection rules change this trade-off on a fixed candidate archive.

We have clarified how these findings motivate the method design in the Introduction. Reaction-F1 conclusions concern projected-signature agreement as defined in Appendix A.5.

AuthorsOpenReview ↗

Response to Reviewer sKxa — Weaknesses (Part 2/3)

W1. Training setting, OOD evaluation, repeatability, and comparison conditions

W1-a. Fixed observation budget and training information

The task compares discovery methods under the same single-trajectory observation setting. Each system provides one fitting trajectory with all states observable. Model fitting and final selection use that trajectory. Trajectories from other initial conditions test the discovered model's predictions, and reference labels support structural evaluation. Q1 explains how this limited-observation task relates to training on multiple trajectories and to experimental design.

W1-b. Additional OOD evaluation across multiple initial conditions

We have completed the additional evaluation using your proposed state-range sampling scheme. For each system, we uniformly sample initial conditions from the box defined by each state's minimum and maximum along the default trajectory to obtain 20 valid reference trajectories. The table compares the same frozen models from search seeds 53, 25, and 81 on the original single OOD trajectory and the new initial conditions. Appendix C.7 (Tables 16-17; lines 1946) reports the main configurations. Here and in W1-c, we additionally include KineSelect/Alpha as a selector comparison.

Method and selectorOriginal single-trajectory OOD R²-score ↑Multiple-initial-condition OOD R²-score ↑
SISR / Alpha0.445 ± 0.0140.461 ± 0.007
KineSelect / Alpha0.640 ± 0.0040.608 ± 0.016
SISR / FPMS0.368 ± 0.0090.367 ± 0.008
KineSelect / FPMS0.538 ± 0.0240.538 ± 0.011

KineSelect achieves higher mean OOD R²-score than SISR under both Alpha and FPMS in both OOD evaluation settings.

W1-c. Results across search seeds with clean training data

KineSelect's gain over SISR recurs when both use FPMS. The table summarizes four runs with random seeds 42, 53, 25, and 81 on the original OOD trajectories. We have added the main configurations and shared-FPMS comparison in Appendix C.6 (Tables 13-15; lines 1836).

Method and selectorOOD R²-score ↑Reaction F1 ↑
SISR / Alpha0.440 ± 0.0150.387 ± 0.008
KineSelect / Alpha0.630 ± 0.0210.334 ± 0.007
SISR / FPMS0.386 ± 0.0370.430 ± 0.006
KineSelect / FPMS0.531 ± 0.0250.454 ± 0.007

Across all four runs, KineSelect exceeds SISR on both OOD R²-score and reaction F1 under shared FPMS, consistent with our original findings. Alpha minimizes complexity only among candidates satisfying its simulation-loss tolerance. With a tight threshold of α=1.05, this can exclude simpler models with slightly higher fitting loss and favor more complex equations.

W1-d. Reported configurations and parameter-selection rationale

Both SISR and KineSelect search over 2 to 12 reactions, with 2,000 individuals and up to 20 generations per reaction-count subpopulation. Appendices D.1-D.3 and Tables 18-20 document the common protocol and method-specific settings. We chose α=1.05 based on SISR’s benchmark performance (Appendix B.3, Figure 9). During identification, KineSelect applies the same observation-quantile rule for K and a shared Hill-exponent set across systems. Appendix B.2 gives these rules, and the new Appendix C.2.1 (Figure 8) reports reference-parameter coverage.

W2. Biological ODE discovery and the paper's overall positioning

We study identification of biological dynamical equations from time series of coupled states. Observations represent biological quantities such as species concentrations and cell or population counts, and the benchmark uses reference ODEs to generate their simulated time series. Input observations lie along system trajectories, and a single trajectory covers a limited part of state space. Candidate equations must describe consumption, production, and regulation between species, including a reaction's shared contributions to multiple state equations. The benchmark makes these questions concrete in low-dimensional, autonomous simulated systems with all states observable.

SysBio-SR provides standardized inputs and reference evaluation for this task. Our empirical analysis quantifies the gap between trajectory fidelity and mechanism recovery, using reaction F1 to assess agreement with the reference reaction signatures. Cross-equation coupling constrains KineSelect's candidate models, while reaction F1 evaluates their projected signatures. KineSelect tests specific designs for reaction roles, kinetic forms, and model selection. Explicit ODEs and reaction representations also give researchers candidates to review using domain knowledge.

AuthorsOpenReview ↗

Response to Reviewer sKxa — Questions (Part 3/3)

Q1. Why fit a single trajectory?

A single trajectory defines a fixed observation budget for studying inductive biases and model selection under the same data conditions. The 512 samples describe one trajectory. We constrain search through reaction roles, coupling across equations, and kinetic candidates, then test discovered models from other initial conditions. This setting does not assume that a single trajectory uniquely identifies the original reaction network.

Multiple trajectories can add identification constraints, and the value of additional observations also depends on their ability to distinguish candidate models. Vanlier et al. (2014) study observation design for distinguishing competing biochemical network models. Fasel et al. (2022) use predictive uncertainty in Ensemble-SINDy to select new initial conditions and add training data. Our task fixes the fitting observations to evaluate discovery methods. The open reference-simulation interface also supports future studies of training on multiple trajectories and adaptive observation acquisition.

Q2. Identifying MM/Hill kinetics and penalizing non-mass-action complexity

We include MM/Hill kinetics to capture saturation and regulation, while favoring simpler models when they fit comparably well. Over a limited observation range, some of these candidates can produce responses similar to MA terms. For example, kx/(K+x)≈(k/K)x when x≪ K, making the saturation parameter difficult to identify from these observations. FPMS evaluates this trade-off at the complete-model level, retaining the ability to select MM/Hill models when improved fit justifies their additional complexity.

FPMS balances derivative and trajectory fit against expression complexity and CspecC_{\mathrm{spec}}, the number of non-MA kinetic terms. Adding CspecC_{\mathrm{spec}} increases mean OOD R² from 0.4194 to 0.4373 and reaction F1 from 0.2784 to 0.2848 on the Challenge subset containing MM/Hill kinetics (Appendix C.3, Table 7).

Q3. Specific differences from SISR and their roles

KineSelect differs from SISR in how it represents regulation, identifies kinetic forms, and selects the final model.

  • RSS: Explicitly represents regulator roles, allowing the search to model species that influence reaction rates without net consumption or production.
  • KFI: Identifies the kinetic form of each reaction from MA, MM, and Hill candidates.
  • FPMS: Selects among complete candidate models by balancing fitting fidelity and model complexity.

Sections 5.1-5.3 describe these components. The shared-FPMS comparison in W1-c shows higher OOD R² and reaction F1 for KineSelect.

Q4. Reaction-structure priors and their relationship to pretrained models

Our framework encodes biological priors directly in the candidate models. Each reaction shares a rate law across the state equations it affects, with stoichiometry determining its contribution to each equation. Species roles and kinetic families further constrain the candidates. These restrictions provide explicit information for learning coupled dynamics from limited trajectory observations.

Foundation models can incorporate reaction-structure priors. Our trajectory evaluation includes ODEFormer and LLM-SR (Tables 2 and 13). One route is to train on synthetic reaction models paired with simulated trajectories, following the general strategy of synthetic ODE pretraining in ODEFormer. A central challenge is covering both reaction structures and the dynamics induced by parameters and initial conditions, including oscillatory regimes. KineSelect imposes reaction constraints directly during search, without requiring such pretraining. These constraints could also guide candidates proposed by a foundation model.

Q5. The F1 in Figure 1(c) and the ranking in that example

We have revised Figure 1 to clarify that panel (c) reports Symbolic-Term F1 for the single system of Munz et al. (2009) (caption, lines 65-67). We compute Symbolic-Term F1 by manually matching complete signed terms between the predicted and reference equations, ignoring parameter values. Our choice of SISR as the main baseline follows the benchmark-wide results in Tables 2 and 3.

Q6. Broader analysis of OOD generalization

We have completed the suggested OOD evaluation across multiple initial conditions. W1-b and Appendix C.7 (Tables 16-17) report the protocol and results.

Reviewer_ahPkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SysBio-SR collects 198 literature-derived biological ODE systems and evaluates trajectory fidelity across all systems and projected reaction-signature recovery on a 161-system subset. The paper also introduces KineSelect, combining reaction-topology search, kinetic-law identification, and fidelity-parsimony model selection (FPMS). Its reported aggregate OOD R2 increases from SISR's 0.426 to 0.508, and mean reaction F1 increases from 0.391 to 0.447. The experiments examine kinetic families, selection rules, and observation noise to show that accurate trajectories need not imply agreement with annotated mechanisms.

Critical strengths
  1. Separating trajectory fidelity from structural agreement addresses an important weakness in symbolic-regression evaluation. The curated systems, annotations, perturbed-initial-condition evaluation, and explicit integration-failure rules make this distinction concrete and useful.

  2. The fixed-archive selection experiments and kinetic-family breakdown provide informative diagnostics. They show that expanding the kinetic library can improve trajectories while reducing signature recovery, and that the final selector changes this trade-off. Reporting the low nonlinear-family scores and higher-noise ranking reversals helps establish the method's practical limits.

Critical weaknesses
  1. The central mechanism-recovery interpretation exceeds the resolution of the metric. Section 4 describes strict topology-and-law matching, whereas Appendix A.5 explicitly projects each reaction onto separate state equations, replaces nonzero stoichiometric coefficients by their signs, and pools signatures across reactions. The appendix acknowledges that the score measures projected-signature agreement; this loses cross-equation reaction coupling. For example, A -> B at rate kA and the split pair {A -> empty at rate kA; empty -> B regulated by A at rate k*A} have the same projected signatures under the stated rules. This is an analytical consequence of the written protocol, not an executed code test. The metric remains useful for projected kinetic-term agreement, but cannot establish recovery of a unique reaction network. Fitting one initial-condition trajectory also leaves identifiability unresolved: even complete ODE agreement need not identify the original reaction decomposition. The claims should reflect that equivalence, or stronger structural evaluation should be justified by identifiable targets.

  2. The additional method contribution and its reliability are less clear than the headline comparison suggests. Appendix C.3 reports SISR+FPMS F1 = 0.431 and OOD R2 = 0.440, leaving a KineSelect F1 margin of 0.016, approximately 3.7%, under the shared selector. The OOD gain remains meaningful. Section 6.3 correctly attributes nonlinear baseline zeros to restricted libraries; consequently, those comparisons do not isolate KFI's benefit from expressivity. Search variability is also unquantified, and Table 9 reverses the overall F1 ranking at 5% and 10% noise. Appendix B.3 states that the alpha-threshold default was chosen after inspecting benchmark outcomes, while an independent development/evaluation protocol for the full design is unclear. A shared-selector comparison with repeated searches and clearly separated configuration development would better establish the size and stability of the contribution.

Questions for authors
  1. Clarification: Under the implemented matcher, can the coupled reaction and split pair above be distinguished? Which claims concern projected-term recovery, and which concern identifiable reaction-network recovery?

  2. Clarification: Why are compatible generic-SR outputs excluded from the same projected-signature evaluation? If mapping requires additional information beyond equation terms, what information is required?

  3. Assessment-changing validation: Does the margin over SISR+FPMS persist across repeated searches on systems excluded from configuration development? Please distinguish settings chosen from benchmark outcomes from settings fixed independently.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ahPk — Weaknesses (Part 1/2)

Thank you for recognizing the value of evaluating trajectory prediction and structural agreement separately, and of our analyses using fixed candidate archives and failure cases. Your comments identify the evaluation scope and methodological gains that need clarification.

W1. Definition and interpretation of the reaction metric

Reaction F1 complements trajectory accuracy by assessing whether recovered models reproduce the reference species dependencies, consumption/production directions, and kinetic forms. As detailed in Appendix A.5, we construct and compare reaction signatures in the same standardized state-equation representation, so that the score measures equation-level structural agreement with the reference.

Matching uses unit stoichiometry and preserves mass-action orders, while excluding numerical rate and shape parameters. We compare equation terms independently of their original reaction grouping; Q1 explains this choice using your coupled/split example. Appendix A.5 (lines 1004-1024) gives the full construction and matching rules.

W2. Gains under a shared selector, expressivity, and performance under noise

With FPMS shared across methods, KineSelect exceeds SISR in both OOD R²-score and reaction F1 in each of the three new search seeds. Q3 provides the four-run summary, including the original paper's results, and discusses repeatability. These results support gains from the complete pipeline.

Your distinction between expressivity and the contribution of the KFI selection algorithm is valid. Our SISR baseline uses mass-action kinetics, whereas enabling KFI also introduces candidate kinetic forms such as MM and Hill. Comparisons on nonlinear kinetic families and ablations that disable KFI therefore change the representable model space. They do not isolate the contribution of the KFI selection algorithm alone. Appendix C.6 (Tables 14-15) reports the shared-FPMS comparison.

We have added noise results in Appendix C.4.1 (Tables 10-11). We summarize the main results below.

OOD R²-score ↑ (columns indicate noise levels)

Method and selector0.0010.010.050.1
SISR / Alpha0.436 ± 0.0160.426 ± 0.0180.417 ± 0.0060.376 ± 0.031
SISR / FPMS0.369 ± 0.0240.352 ± 0.0420.340 ± 0.0540.278 ± 0.061
KineSelect / FPMS0.532 ± 0.0010.491 ± 0.0080.419 ± 0.0040.363 ± 0.028

Reaction F1 ↑ (columns indicate noise levels)

Method and selector0.0010.010.050.1
SISR / Alpha0.388 ± 0.0150.386 ± 0.0100.362 ± 0.0120.318 ± 0.001
SISR / FPMS0.420 ± 0.0140.397 ± 0.0120.350 ± 0.0160.306 ± 0.028
KineSelect / FPMS0.454 ± 0.0010.423 ± 0.0040.354 ± 0.0130.309 ± 0.004

Under shared FPMS, KineSelect has higher mean OOD R²-score than SISR at all four noise levels, while their reaction F1 means are close at 5% and 10% noise (Appendix C.4.1). Noisy derivative estimates may affect candidate fitting and FPMS ranking, contributing to performance degradation at high noise. Alpha uses a simulation-loss tolerance followed by complexity minimization, without an explicit derivative-loss term at final selection. This difference offers one possible explanation for SISR/Alpha’s stronger high-noise performance relative to SISR/FPMS.

AuthorsOpenReview ↗

Response to Reviewer ahPk — Question (Part 2/2)

Q1. Can the implemented matcher distinguish a coupled reaction from the split pair?

The implemented matcher assigns the same signatures to the coupled and split representations in your example. Both produce the same ODE vector field:

A˙=−kA,B˙=kA.\dot A=-kA,\qquad \dot B=kA.

Our reaction-F1 gains show improved agreement with the reference species dependencies, consumption/production directions, and kinetic forms. These claims concern structural recovery under the matching rules in Appendix A.5, without asserting a unique reconstruction of the original reaction network. Recovered kinetic terms can provide a starting point for researchers to develop and refine mechanistic hypotheses using domain knowledge and additional observations.

Q2. Why does the same reaction evaluation exclude compatible generic SR outputs?

Our common trajectory evaluation already includes generic SR methods, and we report reaction F1 for methods with validated mappings from their outputs to standardized reaction signatures. In biological reaction models, stoichiometry specifies which species each reaction consumes or produces, while the kinetic law defines the rate expression. More general expressions require consistent normalization and kinetic-class recognition before matching to avoid differences caused solely by algebraic notation.

For example, the same saturation function can take the form x/(K+x)x/(K+x) or 1−K/(K+x)1-K/(K+x). Splitting these expressions directly into surface-level additive terms can produce different term counts and classes. Our reference annotations and scoring rules provide a basis for extending evaluation coverage, and compatible generic SR outputs can use the same metric. We have yet to validate adapters for these methods. We have clarified the evaluation scope and the need for expression normalization and signature mapping in Section 7 (Conclusion and Limitations).

Q3. Does the gain under shared FPMS persist across search seeds and on systems excluded from configuration development?

The gain under shared FPMS recurs in all three new search seeds. The table below summarizes four runs with random seeds 42, 53, 25, and 81. Each run uses the full denominators of 198 trajectory systems and 161 reaction systems. We have added these results in Appendix C.6 (Tables 13-15).

Method and selectorOOD R²-score ↑Reaction F1 ↑
SISR / FPMS0.386 ± 0.0370.430 ± 0.006
KineSelect / FPMS0.531 ± 0.0250.454 ± 0.007

In each of the three new search seeds under the current configuration, KineSelect/FPMS exceeds SISR/FPMS on both metrics.

We chose the Alpha default with reference to SISR's performance. Figure 9 reports the full threshold sweep and supports our choice of α=1.05, which yields a competitive SISR baseline. KineSelect's main configuration uses FPMS for final selection. We also report the shared-FPMS comparison.

KFI uses a small set of shape-parameter candidates to control the dictionary size for each reaction. We consulted benchmark parameter values when designing the candidate space; actual identification uses a common candidate-generation rule. For species with valid positive fitting observations, the KK candidates use the 5th percentile and median of those observations, representing a smaller and a typical observed scale. Hill candidates combine these KK values with the shared exponent set {1,2,4}\{1,2,4\}. Even when the exponent set contains the reference nn, the finite KK candidates do not guarantee that the resulting grid contains the corresponding reference (K,n)(K,n) pair. We must still infer reaction structure, kinetic forms, and rate constants from trajectories; we do not use the system's reference KK or nn as identification inputs. Appendix B.2 and Table 4 specify the candidate rules.

This is a finite parameter discretization. The new reference-parameter statistics describe the coverage of the fixed candidate rules. We have added this analysis in Appendix C.2.1 (Figure 8; lines 1339).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.