{"id":"06308e15-dfce-4c45-bf5b-5946611483c8","arxiv_id":"2504.12675","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proposes a self-supervised physics-balance loss plus a message-passing optimizer to estimate reaction fluxes from static features, without any supervision from observed flux values.","lead":"MPOCtrL is a physics-informed neural network that predicts metabolic fluxes from static single-cell data using a message-passing balancing algorithm. It claims to outperform existing flux estimators, but its training loss omits any direct data fit, so the reported success on synthetic data may rest on circular benchmarking.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 2's coherency-only loss cannot identify flux: any balanced function of D has zero loss, so the claimed learning-from-static-data result is underdetermined.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: identifiability from the coherency loss alone. I checked the formulation in Definition 2 and Eq. 2 and found no data-fitting term; the paper's own appendix acknowledges that product observations are needed for identifiability in the NESS case but provides no equivalent condition for the ESS/unobserved case used in experiments. The empirical results therefore cannot establish the headline claim. The synthetic target generation via MPO is a second, compounding concern, and the proposed invariance test would expose it. I agree with the reader's REJECT verdict; no adjustment is needed. The paper does provide a concrete optimizer, MPO, and a reproducible code link, and the real-world correlation is a positive sign, but these do not repair the missing identification step.","tokens_in":27507,"tokens_out":7540,"duration_ms":92722,"concrete_test":"Construct two synthetic datasets with identical observed features D but different ground-truth flux matrices on the same network: choose a nonzero matrix H(D) such that Gamma H(D) = 0 for every sample, for example a circulation around each cycle scaled by a smooth function of D, and set Y' = Y + H(D). Run the full MPOCtrL pipeline on (D, Y) and on (D, Y') with the same seed and hyperparameters. Because Eq. 2 never references Y or Y', the fitted network should be identical in both runs; if it is, the objective is invariant to the true flux, proving that the method cannot identify flux from static data. If the runs diverge, the divergence must come from code that uses flux labels outside the stated loss, and that use should be disclosed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Section 3.4 is not supported by its own objective. Eq. 2 is L = sum_j ||Gamma F(D_j, Theta)||^2 + L_p(F, Theta). The observed data D_j enter only as the argument to the neural networks; there is no term that compares F(D_j, Theta) to a measured flux, a measured derivative, or any other observed output. Every function F with Gamma F(D_j, Theta) = 0 for all j attains the same coherency loss, namely zero, so the true flux is not identifiable from the stated loss. Because the stoichiometric matrix Gamma has a nontrivial null space in any network with cycles, this is not a corner case. The paper supplies no additional assumption that connects the feature distribution to flux, such as a known kinetic law, and Appendix A.6 itself concedes that for NESS systems the product must be observed to ensure identifiability; the ESS/unobserved case used here has no analogous safeguard. The synthetic benchmark in A.3.2 additionally generates ground-truth flux by running MPO, the same balancing operator used at inference, so the high cosine similarities partly measure self-consistency with the algorithm rather than recovery of a data-driven mapping. The real-world PCC = 0.75 (p = 0.019) is suggestive but cannot compensate for the absence of a fitting term.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 'Constrained Learning,' a proposed extension of physics-informed neural networks (PINNs) that aims to learn dynamics, specifically flux rates, from static and partially observed data without time-course information. The method represents a dynamic system as a directed factor graph, defines a coherency loss based on mass conservation (||Gamma * F||^2), and combines it with a parsimony regularizer. The authors propose an optimization algorithm, MPOCtrL, that alternates between neural network updates and a message-passing balancing step (MPO). Experiments on synthetic networks and a real scRNA-seq dataset (GSE173433) compare MPOCtrL with scFEA and Compass, and the paper claims that MPOCtrL outperforms existing data-driven flux estimators. The main text and appendix provide algorithmic details, network diagrams, and an appendix section that classifies dynamic systems and proposes coherency losses for various sub-tasks.","tokens_in":27722,"tokens_out":3678,"duration_ms":36854,"significance":"If the central claim were supported, the framework would address a genuine and important limitation of PINNs, namely their reliance on time-course or fully observed data. The paper has some strengths: the implementation is publicly available (GitHub link provided), the appendix gives a systematic taxonomy of dynamic system sub-tasks, and the reported real-world PCC of 0.75 (p=0.019) is a potentially interesting observation. However, the significance is severely undermined by the identifiability issue in the core objective: the loss in Eq. (2) contains no data-fitting term, so any flux function lying in the null space of the stoichiometric matrix achieves zero coherency loss. The synthetic evaluation is also partly circular because the ground-truth fluxes are generated by the same MPO algorithm used during inference. Consequently, the experiments do not establish that the method recovers true flux rates, and the claim of outperforming all existing data-driven flux estimators is not supported.","major_comments":[{"comment":"Equation (2) defines the total loss as ||Gamma * F(D, Theta)||^2 + L_p(F, Theta). The observed data D enter only as inputs to the neural networks; there is no term comparing the predicted flux to a measured flux, a measured derivative, or any other observed output. Any function F with Gamma * F(D, Theta) = 0 for all samples achieves zero coherency loss, so the objective cannot distinguish the true flux from any other vector in the null space of the stoichiometric matrix. Since every network with cycles has a nontrivial null space, this is not a corner case. The paper supplies no additional assumption linking the feature distribution to the flux, such as a known kinetic law; Appendix A.6 itself states (Sub-task 3) that for NESS systems the product must be observed to ensure identifiability, and the ESS/unobserved setting used here has no analogous safeguard. Without an identifiability argument or a fitting term, the central claim of Section 3.4 is unsupported.","section":"3.4, Eq. (2)"},{"comment":"The synthetic ground truth is generated by running the MPO algorithm to balance randomly drawn positive numbers on the same directed factor graph that MPOCtrL uses during inference. This makes the benchmark measure self-consistency: the method is only asked to reproduce the fixed point of the same balancing operator used in its own pipeline. A data-driven flux estimator should be evaluated against fluxes that arise from an independent simulation or measurement process. As it stands, the cosine similarities in Table 1 cannot be interpreted as evidence that the method recovers underlying flux rates, and the comparison against scFEA and Compass is correspondingly weakened.","section":"A.3.2"},{"comment":"In the appendix's own taxonomy, identifiability for NESS systems requires that the reaction products (or time-course data) be observed; the loss L_NESS in Eq. (7) explicitly includes a supervised fitting term (X_{k,j} - sum_m F_{m,j})^2. The flux estimation experiments in Section 4 instead assume Vfa = X is completely unobserved (D ∩ Vfa = ∅) and rely solely on Eq. (2), which contains no such fitting term. Thus the paper's experimental setting is precisely the case the appendix indicates is not identifiable, and the claimed approximation of flux from static data is not justified.","section":"A.6, Sub-tasks 3 and 4"}],"minor_comments":[{"comment":"The word 'scarsity' in the abstract is a typo for 'scarcity'.","section":"Abstract"},{"comment":"The summation index in Eq. (2) runs over j = 1,...,n, but the number of samples is denoted m throughout the paper; this should be corrected.","section":"Eq. (2)"},{"comment":"The hyperparameters beta (MPO learning rate) and delta (imbalance loss threshold) are listed in Algorithm 1 but their numerical values or tuning procedure are not specified in the main text or the experiments; this makes reproducibility more difficult.","section":"Section 3.3 / Algorithm 1"},{"comment":"The paragraph beginning 'The message passing step is highly efficient' claims O(nK) time and space complexity, but the appendix (A.1.2) states O(MN). These complexity statements are inconsistent and should be reconciled.","section":"Section 3.3"},{"comment":"Figure 1 is referenced in the text but its sub-panels (a), (b), and (c) are not explained in sufficient detail; a reader cannot tell what the geometric illustration or the framework diagram depicts without referring to later sections.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"The central methodological problem is that the loss in Eq. (2) has no data-fitting term, so the objective is underdetermined. This is not a presentation issue; it undermines the main claim. A revision would need to either add a supervised fitting term (which changes the problem setting) or introduce explicit identifiability assumptions connecting the observed features to the flux. The synthetic benchmark's reliance on MPO-generated ground truth is another fundamental issue that would require a complete redesign of the evaluation. I therefore recommend rejection, despite the paper's useful appendix taxonomy and the authors' release of code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe punchline is that the central claim of this paper — that MPOCtrL learns dynamics from static data — doesn't hold up under inspection. The loss in Eq. 2 is pure constraint satisfaction: ||Γ F(D,Θ)||² plus a parsimony regularizer. The observed data D appear only as inputs to the neural networks; there is no term comparing F to a measured flux, derivative, or any other observed output. Any function F whose outputs lie in the null space of Γ achieves zero coherency loss, so the true flux is not identifiable from the stated objective. That is not a corner case: any network with cycles has a nontrivial null space.\n\nWhat the paper does well: the MPO message-passing balancer is concrete, with O(MN) complexity and a clear update rule. The authors ship code and compare against scFEA and Compass on several networks. The real-world PCC of 0.75 (p=0.019) on nine metabolites is suggestive, though based on a single small dataset.\n\nWhere it falls down: the synthetic evaluation is partly circular. The ground-truth fluxes are generated by running MPO, the same balancing algorithm used at inference, so high cosine similarity to the target partly reflects self-consistency. Appendix A.6 concedes that for NESS systems the product must be observed to ensure identifiability, but the ESS/unobserved case used in the main text has no analogous safeguard. That concession is close to an admission that the target setting is underdetermined. The abstract's claim that MPOCtrL outperforms all existing data-driven flux estimators is also stronger than the evidence supports, given two baselines and a circular benchmark.\n\nWho this is for: a researcher working on metabolic flux estimation might find the MPO heuristic and the scRNA-seq pipeline worth a look, but they should not treat the method as validated for learning dynamics from static data. The framing as 'Constrained Learning' is mostly a relabeling of existing flux-balance constraints.\n\nRecommendation: I would not accept this in current form. It deserves a serious referee, but the review must push for an identifiability argument or a non-circular synthetic experiment. If the authors can supply that, the MPO balancer could be a useful component; as it stands, the central claim is unsupported.","headline":"The paper's central claim doesn't survive its own objective: with no data-fitting term, any balanced flux function has zero loss, and the synthetic benchmark is circular; the MPO balancer and real-data pipeline are the genuinely useful pieces.","tokens_in":28317,"tokens_out":3181,"would_cite":false,"duration_ms":31384,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Static snapshots can replace time-course data in physics-informed learning, the paper claims, by deriving fluxes from conservation laws alone when measured derivatives are unavailable.","keywords":["physics-informed neural networks","constrained learning","metabolic flux analysis","message passing optimization","directed factor graph","non-time-course data","self-supervised learning","flux balance analysis"],"falsifier":"Take a fixed network and generate synthetic data where the input features are held fixed but the true balanced flux values are randomly permuted across samples, so no feature-flux link exists. If MPOCtrL's predictions do not change under this permutation while the permuted labels differ, the coherency loss alone cannot select the true flux, contradicting the identifiability claim.","tokens_in":27256,"feed_emoji":"🧬","tokens_out":4112,"duration_ms":45723,"temperature":0.7,"pith_summary":"The paper proposes a learning paradigm called Constrained Learning for estimating the rates or fluxes of a dynamical system from static, non-time-course, and partially observed data. Classical physics-informed neural networks need time-course data and full observation of variables; this work replaces the supervised time-derivative fit with a coherency loss that encodes physical laws, plus a parsimony penalty. For flux-balance systems the physics loss reduces to $\\|\\Gamma F(D,\\Theta)\\|^2$, the squared violation of mass conservation at every intermediate node. The authors present MPOCtrL and a message-passing optimizer, and report that on synthetic reaction networks and one real single-cell dataset the method matches or beats existing data-driven flux estimators.","feed_headline":"Static snapshots can replace time-course data in physics-informed learning","feed_subtitle":"A conservation-law loss plus message passing recovers metabolic fluxes from single-cell data, outperforming prior estimators.","key_machinery":"The central object is a Directed Factor Graph with factor nodes for variables and variable nodes for flux-carrying reactions, connected through a stoichiometric matrix $\\Gamma$ that encodes conservation laws. The argument is carried by the combined loss $L = L_\\Phi(F(D,\\Theta),\\Gamma) + L_p(F,\\Theta)$: the first term forces predicted fluxes to sit near the kernel of $\\Gamma$ (balanced flow), while the second keeps the network simple and interpretable via gating and L2 regularization. A belief-propagation-inspired Message Passing Optimization (MPO) step rebalances variable weights using messages from neighboring nodes, with claimed time and space complexity $O(nK)$.","core_discovery":"The central discovery claimed is that the time derivative of a dynamical system can be approximated without any time-course measurements, provided the system's physical laws are encoded as a constraint on a factor-graph representation. Concretely, for mass-carrying flux over a network, the steady-state conservation law gives the coherency loss $L_\\Phi = \\|\\Gamma \\{f_1,\\ldots,f_K\\}^T\\|^2$, where $\\Gamma$ is the stoichiometric weight matrix. Minimizing this loss together with a parsimony regularizer over feature-driven neural networks yields sample-wise flux estimates. The paper states that on metabolic flux analysis this Constrained Learning approach outperforms all existing data-driven flux estimators.","pith_inferences":["Because no term in the loss compares predicted flux to measured flux, the method's success on synthetic data likely relies on the constructed nonlinear link between features and balanced flux; when that link is weak, the coherency loss alone cannot distinguish among infinitely many balanced flux functions of the features.","A direct identifiability test would be to train on data where true flux is balanced but statistically independent of the input features; predictions should then be arbitrary balanced functions, exposing the hidden dependence on a feature-flux relationship.","The real-world validation, with nine metabolites and a single cell line, is narrow relative to the broad superiority claim; broader evaluation with isotope-based flux measurements across tissues would be needed to establish practical advantage.","MPO's update rules resemble belief propagation, so its convergence and fixed-point properties on cyclic graphs could be studied formally, potentially extending the framework to loopy networks with guarantees."],"forward_implications":["Physics-informed models could be applied to systems where temporal sampling is impossible, such as human tissue metabolism, using only static snapshots.","Cell-wise metabolic flux could be estimated from single-cell RNA-seq data without isotope tracing or time-resolved measurements.","The same coherency-loss formulation applies to other mass-carrying networks, including traffic flow, cash flow, and reaction-diffusion systems.","The MPO optimizer can serve as a general graph-balancing procedure independent of the neural network, with linear complexity in the number of nodes.","The proposed taxonomy of equilibrium versus non-equilibrium systems under steady or dynamic states suggests coherency losses for a broader class of dynamical models."],"supporting_citations":[{"why":"Defines the PINN framework that this work extends by removing the time-course data requirement.","marker":"[34]"},{"why":"Provides the scFEA baseline and the CMMRN network used for comparison on flux estimation.","marker":"[2]"},{"why":"Provides the Compass baseline, a static-data flux estimator that the paper compares against.","marker":"[43]"},{"why":"Supplies the principles of coherency and parsimony that ground the Constrained Learning loss design.","marker":"[27]"},{"why":"Supplies belief-propagation convergence results that motivate the message-passing optimization approach.","marker":"[17]"},{"why":"Establishes the factor-graph and sum-product machinery used to formulate the Directed Factor Graph representation.","marker":"[22]"},{"why":"Introduces flux balance analysis and its steady-state polytope, the physical constraint that defines the coherency loss.","marker":"[30]"},{"why":"Provides the belief-propagation background that the MPO message-passing updates are adapted from.","marker":"[49]"}],"fun_headline_variants":["Static snapshots, no time-course: physics constraints learn dynamics","Constrained learning derives motion from still observations","Physics-informed learning from static data without time series","MPOCtrL: flux estimation from static snapshots beats prior models","No time-course? Physics loss still recovers first-order dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the true flux rates are identifiable from the observed features and the conservation-law constraint alone, even though the loss never compares a predicted flux to any measured flux value.","fun_headline_variants_meta":{"raw":{"variants":["Static snapshots, no time-course: physics constraints learn dynamics","Constrained learning derives motion from still observations","Physics-informed learning from static data without time series","MPOCtrL: flux estimation from static snapshots beats prior models","No time-course? Physics loss still recovers first-order dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":1970,"prompt_tokens":892,"completion_tokens":1078,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":996}},"tokens_in":508,"tokens_out":1078,"duration_ms":11342,"temperature":1.0,"reasoning_tokens":996,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:25:49.017222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed network and generate synthetic data where the input features are held fixed but the true balanced flux values are randomly permuted across samples, so no feature-flux link exists. If MPOCtrL's predictions do not change under this permutation while the permuted labels differ, the coherency loss alone cannot select the true flux, contradicting the identifiability claim.","supporting_citations":[{"cited_title":"Alghamdi, W","cited_arxiv_id":null,"evidence_quote":"Provides the scFEA baseline and the CMMRN network used for comparison on flux estimation."},{"cited_title":"Wagner, C","cited_arxiv_id":null,"evidence_quote":"Provides the Compass baseline, a static-data flux estimator that the paper compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the principles of coherency and parsimony that ground the Constrained Learning loss design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies belief-propagation convergence results that motivate the message-passing optimization approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the factor-graph and sum-product machinery used to formulate the Directed Factor Graph representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces flux balance analysis and its steady-state polytope, the physical constraint that defines the coherency loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the belief-propagation background that the MPO message-passing updates are adapted from."}],"review_version":1}