{"id":"5f3ec083-ddbd-426e-aa97-b2e6bb3b58cc","arxiv_id":"2502.02129","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"NeuralCPM learns a neural-network Hamiltonian for cellular Potts simulations from data, and reproduces self-organization dynamics beyond what analytical Hamiltonians achieve.","lead":"This paper introduces NeuralCPM, a cellular Potts model whose energy function is a neural network trained on observed cell configurations. It aims to simulate collective cell behaviors, such as embryo-like axis formation, that hand-crafted analytical Hamiltonians miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training relies on an approximate parallel sampler that violates detailed balance (App. B.1), so the learned Hamiltonian is not guaranteed to be a valid CPM Hamiltonian under the standard sequential Metropolis dynamics; the central claim may hold only for the biased sampler.","rationale":"I focused on the sampler because the method's theoretical grounding and the generality of the central claim both depend on Hθ being a Boltzmann Hamiltonian under the CPM's Metropolis dynamics. The reader's weakest_assumption identifies this same issue, and the paper's own Appendix B.1 admits the transition kernel is faulty. The concrete test would settle whether the bias is benign: if exact-sampler training gives materially different parameters or if the ApproxPCPM-trained Hamiltonian fails under exact simulation, then the learned object is not a valid CPM Hamiltonian and the headline claim is overstated. I also share the reader's secondary concern about training on synthetic final configurations for the Toda experiment, but that concern is compounded by, rather than independent of, the sampler issue. The paper does provide useful empirical evidence—parameter recovery is substantially better than the Gibbs and Gibbs-with-Gradients baselines, and the NH+closure variant stabilizes training—so the appropriate outcome is a conditional acceptance pending the proposed verification, not rejection.","tokens_in":17939,"tokens_out":11020,"duration_ms":127593,"concrete_test":"Retrain the §4.2 cell-sorting benchmark using the exact sequential CPM sampler (standard single-site Metropolis with the same boundary-based proposal but no parallel flips) to draw negative samples, on a small lattice (e.g., 64×64, 20 cells) so mixing is tractable. Compare the recovered contact energies to ground truth with the same T*-adjusted RMSE metric. If the exact-sampler-trained parameters deviate substantially from the reported values (0.021 for type A, 0.178 for type B), the ApproxPCPM bias is not benign. As a second check, simulate the Hθ trained with ApproxPCPM under the exact sequential CPM and measure the equilibrium axial-alignment or Classifier Score metrics; if these metrics collapse, the learned Hamiltonian is tied to the approximate sampler and is not a standalone CPM Hamiltonian.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 frames training as minimizing the NLL with gradient E_{p*}[∇Hθ] − E_{pθ}[∇Hθ] (Eq. 3). In practice, negative samples are drawn with Algorithm 2 (ApproxPCPM), which proposes P parallel flips, accepts/rejects each against the original state, and then combines all accepted flips. The paper states in Appendix B.1 that these transition probabilities 'do not satisfy detailed balance' and that 'we cannot guarantee that the system has a stationary distribution defined by Hθ(x)'. Therefore the estimated gradient in Eq. 6 is not the NLL gradient, and the learned Hθ is not guaranteed to be a valid Boltzmann Hamiltonian for the CPM as usually defined. The only validation is §4.2 parameter recovery, but that check also trains with ApproxPCPM, evaluates coefficients after a post-hoc temperature fit T = T*, and uses the same biased sampler to define the model, so it cannot detect sampler-specific overfitting. If the success of NeuralCPM depends on the approximate sampler's particular transition kernel, the central claim that the learned Hamiltonian accounts for dynamics that analytical Hamiltonians cannot is not established for the standard CPM. The paper's own limitation statement in Appendix B.1 is therefore load-bearing, not merely a caveat. A secondary concern, also raised by the reader, is that the Toda et al. experiment trains on synthetic final configurations with the target bi-polar arrangement prescribed; this further weakens the real-world demonstration but would not by itself invalidate the method if the sampler issue were resolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NeuralCPM, a cellular Potts model whose Hamiltonian is a neural network with built-in translation and cell-permutation symmetries. The Hamiltonian is trained from static snapshots of configurations by minimizing an energy-based negative log-likelihood objective, using a persistent contrastive divergence scheme with an approximate parallelized CPM sampler. The authors validate the approach in three settings: parameter recovery for a known analytical cell-sorting Hamiltonian, a synthetic 'Cellular MNIST' assembly task, and a real biological scenario of bi-polar axial organization from Toda et al. (2018). They find that a hybrid Hamiltonian combining an analytical cell-sorting term with the neural closure term yields the best trade-off between biological realism and collective structural accuracy, and they report that this hybrid model reproduces the temporal development of bi-polar organization.","tokens_in":18245,"tokens_out":3552,"duration_ms":37072,"significance":"If the central claim holds, NeuralCPM is a useful contribution to computational biology and machine learning: it provides a principled way to learn expressive CPM Hamiltonians directly from observational data while retaining the interpretable CPM simulation framework. The architecture is well motivated, the hybrid 'closure' formulation is a practical way to inject domain knowledge, and the temporal prediction against Toda et al. (2018) is an ambitious and nontrivial validation. The explicit comparison with analytical baselines and alternative neural architectures gives the paper a clear empirical message. However, the statistical grounding of the training objective is weakened by the approximate sampler that violates detailed balance, and the equilibrium bi-polar validation is partially circular. These issues affect the strength of the main claim as currently stated, but they are addressable with additional experiments or a reframing of the claims.","major_comments":[{"comment":"The training loss is presented as an estimate of the negative log-likelihood gradient (Eq. 3), but the negative samples are drawn with ApproxPCPM, whose transition probabilities, as the authors state, 'do not satisfy detailed balance' and for which 'we cannot guarantee that the system has a stationary distribution defined by Hθ(x)'. Therefore the expectation over pθ(x) in Eq. (3) is not estimated from the model distribution, and the learned Hamiltonian is not guaranteed to define the stationary distribution of the sampler used at inference. This is load-bearing for the central claim that NeuralCPM models dynamics that analytical Hamiltonians cannot, because the empirical success could be an artifact of the approximate sampler's particular transition kernel rather than of Hθ. Please either bound the bias introduced by the approximate sampler, validate the learned Hamiltonians under the standard sequential CPM Metropolis dynamics, or explicitly reframe the claims as holding for the ApproxPCPM dynamics.","section":"§3.2, Eq. (6); Appendix B.1, Algorithm 2"},{"comment":"The equilibrium bi-polar validation is partially circular. The synthetic training configurations were generated by prescribing the target location of each type-2 cell for a bi-polar arrangement, so the model is trained to produce the very outcome that is then compared with the real data in Figure 7a. The temporal-dynamics comparison in Figure 7b and the axial-alignment RMSE in Table 2 are not circular and provide the strongest evidence for the real-world claim; the paper should present them as the primary validation and treat the equilibrium overlap as a consistency check.","section":"§4.4, Figure 7a; Appendix A"},{"comment":"The parameter-recovery experiment reports the best RMSE at T = T*, a temperature selected post hoc to minimize the error. Because the same approximate sampler is used both for training and for defining the learned model, this check cannot detect systematic sampler-specific biases, and the T = 1 result for type B (RMSE 0.997) is considerably worse than the T* result (0.178). Please report the selection procedure for T*, provide repeated-seed variability or confidence intervals, and ideally evaluate the fitted Hamiltonian with a detailed-balance sampler to show that the recovered parameters transfer to the standard CPM.","section":"§4.2, Table 8 and Figure 4"}],"minor_comments":[{"comment":"The manuscript does not include a data or code availability statement; adding one would aid reproducibility, especially given the use of the shared Toda et al. (2018) microscopy data.","section":"General"},{"comment":"The notation 'j ∼ M(i)' is ambiguous; the paper should state that j is drawn uniformly from the set of neighboring sites belonging to a different cell.","section":"Appendix B.1, Algorithm 2"},{"comment":"The symbol N(L) is used for the set of neighboring lattice-site pairs in Eq. (4) and for the per-site neighbor set in Algorithm 2; this overloading should be resolved to avoid confusion.","section":"Equation (4) and Algorithm 2"},{"comment":"The caption states that 'almost all observations' lie within one standard deviation of the simulations, but it does not report the exact number of observations; please give the count (e.g., five of six) and specify the standard-deviation definition used for the shaded region.","section":"Figure 7 caption"},{"comment":"The word 'Parmameter' in the captions is a typo and should be corrected to 'Parameter'.","section":"Figures 4 and 8"},{"comment":"The word 'classs' is a typo and should read 'class'.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The detailed-balance concern raised in the stress-test is real and central: the authors themselves acknowledge in Appendix B.1 that ApproxPCPM does not define a stationary distribution consistent with Hθ, so Eq. (3) is not the NLL gradient as written. At the same time, the temporal prediction in Figure 7b is a genuinely non-circular empirical result and gives a viable path to revision. I would encourage the editor to require either a validation under standard sequential CPM dynamics or a clear reframing of the claims, and to ask for the parameter-recovery experiment to be reported with the temperature-selection procedure made explicit. The paper fits a machine-learning-for-science venue, but the methodological gap needs to be closed before the central claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously. NeuralCPM is the first approach I know that parameterizes a cellular Potts model Hamiltonian as a symmetry-aware neural network and trains it directly on observed configurations. That removes the hand-engineering bottleneck in CPM modeling, and the hybrid closure formulation that keeps a symbolic analytical Hamiltonian while letting the network fill in the gap is a sensible design that also stabilizes training. The Cellular MNIST experiments are a nice stress test of expressiveness, and the quantitative win on the bi-polar axial organization task (RMSE 37.3 vs. ~150 for analytical baselines) is genuinely striking, especially because the model predicts temporal dynamics and symmetry breaking that were not explicitly built into the training data.\n\nThat said, there are real soft spots. The load-bearing one is the approximate parallel sampler in Algorithm 2. The paper states outright that its transition probabilities do not satisfy detailed balance, so the chain is not guaranteed to have a stationary distribution defined by H_theta. That means the training objective in Eq. 6 is not the NLL gradient the paper claims it approximates, and the learned Hamiltonian is not guaranteed to be a valid Boltzmann Hamiltonian under the standard sequential CPM dynamics. The parameter-recovery experiment cannot detect this because it trains and evaluates with the same biased sampler, and it also fits a temperature T* post hoc. The real-data experiment is also weaker than it first appears: the model is trained on synthetic final configurations with the bi-polar arrangement prescribed, so the real Toda data are only used for validation. The temporal dynamics prediction is less circular and is the most convincing piece of evidence, but it is a single experiment.\n\nI would not call any of this fatal. The authors acknowledge the sampler issue explicitly and the empirical results are encouraging. But the central claim—that the learned Hamiltonian accounts for dynamics analytical Hamiltonians cannot—is only established for the approximate sampler, not for the standard CPM. A revision that bounds the sampler error, compares against a correctly balanced sampler on a small case, or releases code and data so others can check, would substantially raise my confidence.\n\nWho is this for? Researchers working on energy-based models for biological simulation, and computational biologists who use CPMs. They should read it. It deserves peer review—it is novel, clearly written, and the failures of the baseline samplers are honestly reported. I would send it out, but I would require the sampler issue to be seriously addressed before acceptance.","headline":"A genuinely new way to learn CPM Hamiltonians from data, with promising results on a real biological test case, but the training sampler breaks detailed balance and the real-data experiment trains on synthetic surrogates, so the method needs another pass before it can be fully trusted.","tokens_in":18800,"tokens_out":629,"would_cite":true,"duration_ms":7939,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-network Hamiltonian learned from snapshots of cell clusters reproduces bi-polar self-organization that analytical cellular Potts Hamiltonians cannot capture.","keywords":["cellular Potts model","neural Hamiltonian","energy-based models","collective cell dynamics","self-organization","bi-polar axial organization","MCMC sampling","symmetry-equivariant architecture"],"falsifier":"Simulate the learned bi-polar NeuralCPM Hamiltonian with an exact single-site Metropolis CPM sampler instead of the approximate parallel sampler and measure the axial-alignment RMSE over 12 hours; if the bi-polar organization disappears or the time course diverges from the real data, the reported success depended on the sampler's bias rather than on the learned energy function.","tokens_in":17735,"feed_emoji":"🧫","tokens_out":10771,"duration_ms":87442,"temperature":0.7,"pith_summary":"The paper proposes NeuralCPM, a cellular Potts model whose Hamiltonian—the energy function that drives stochastic cell movement and shape change—is a neural network rather than a hand-crafted physics formula. The authors' goal is to show that such a learned Hamiltonian can be fitted directly from observed multicellular configurations, avoiding the labor of engineering problem-specific Hamiltonians, whose rigid form also limits their accuracy. They demonstrate on a real experiment that a symmetry-aware neural Hamiltonian, combined with a standard analytical cell-sorting term as a closure model, reproduces the 12-hour bi-polar axial self-organization of a cell aggregate with an axial-alignment error of 37.3, whereas analytical baselines score around 150 and never form the two-pole structure. The same framework recovers known parameters of a cell-sorting Hamiltonian and assembles digit-like cellular structures in the synthetic Cellular MNIST test, supporting the claim that increased expressiveness translates into biological fidelity.","feed_headline":"Learned cell energy captures self-organization analytical laws miss","feed_subtitle":"Trained on static snapshots, it predicts 12-hour bi-polar self-organization that hand-built CPM energies cannot produce.","key_machinery":"The load-bearing object is the Neural Hamiltonian $H_\\theta(x)$, an architecture that assigns an energy to any lattice configuration $x$ (each site labeled by the cell occupying it) while respecting the two symmetries that hold for any CPM: invariance under translations of the lattice and under permutations of cell identities. It one-hot encodes each cell's occupancy as its own grid, feeds each grid through a per-cell convolutional network $\\phi_l$, aggregates all cells with a permutation-invariant sum into a global context lattice $A$, then passes each cell's representation together with $A$ through a second convolutional network $\\psi_l$, followed by pooling and an MLP that outputs the scalar energy. Training uses the energy-based-model maximum-likelihood gradient, where positive samples are observed configurations and negative samples come from an approximate parallel CPM sampler that proposes boundary spin flips; the hybrid model adds the standard analytical cell-sorting Hamiltonian as a weighted symbolic term, which the paper finds stabilizes training and enforces biological realism.","core_discovery":"The central claim is that a cellular Potts model whose Hamiltonian is parameterized by a translation- and permutation-equivariant neural network can be trained from static snapshots and can generate collective dynamics that analytical Hamiltonians cannot. Using the bi-polar axial-organization experiment of Toda et al. (2018) as the real-world test, the paper trains on synthetic final configurations in which type-2 cells cluster at two opposite poles, then simulates from a mixed cluster; the resulting trajectories match the observed time course and degree of axial organization, with over 95% of real observations within one standard deviation of the simulations and an axial-alignment RMSE of 37.3 for the hybrid model versus roughly 150 for analytical baselines. The paper explains this success through the Neural Hamiltonian's inductive biases—invariance to lattice translations and cell permutations—and through the biology-informed closure term, which prevents cell fragmentation and stabilizes energy-based training. Validation on parameter recovery for a known cell-sorting Hamiltonian (RMSE 0.021–0.178 at the optimal temperature) and on Cellular MNIST (Classifier Score 4.35 for the hybrid, beating all analytical baselines) supports the broader claim that NeuralCPM provides a general route from observational data to CPM dynamics.","pith_inferences":["Because the approximate sampler lacks detailed balance, the learned Hamiltonians may encode sampling bias; running the same Hamiltonian with an exact single-site CPM sampler would reveal how much of the reported behavior is due to the energy function itself.","The successful recipe—generate synthetic final states with a CPM simulator, train a NeuralCPM, validate on real time-lapse data—could transfer to other morphogenetic systems where only end states are observable.","The architecture's global receptive field and the equilibrium assumption limit NeuralCPM to small systems; scaling to tissue-level simulations will require locality and history-dependent dynamics, as the paper acknowledges in its limitations.","If the hybrid's symbolic component is kept fixed at its fitted parameters, the residual neural term may isolate the biological mechanisms the symbolic model misses, offering a tool for hypothesis discovery."],"forward_implications":["Biologists can fit a CPM directly to end-state observations, without needing full time-lapse trajectories or hand-designed energy terms.","Simulations with a learned Hamiltonian can extrapolate from synthetic training states to real experimental dynamics, including the temporal pathway of symmetry breaking.","Hybridizing a symbolic biological prior with the neural closure yields both biological plausibility (compact, unfragmented cells) and the expressiveness needed for patterns that analytical terms alone cannot form.","The learned energy function is a candidate quantitative description of the effective cell-cell interactions, so it can be interrogated to extract mechanisms rather than treated purely as a simulator."],"supporting_citations":[{"why":"Defines the cellular Potts model and the cell-sorting Hamiltonian that NeuralCPM fits and extends.","marker":"Graner & Glazier, 1992"},{"why":"Provides the real-world bi-polar axial organization experiment that NeuralCPM's simulations are validated against.","marker":"Toda et al., 2018"},{"why":"Provides the simulation environment used to generate the synthetic bi-polar training configurations.","marker":"Starruß et al., 2014"},{"why":"Introduces contrastive divergence, the basis of the maximum-likelihood training objective.","marker":"Hinton, 2002"},{"why":"Persistent contrastive divergence, used to keep MCMC chains across gradient steps.","marker":"Tieleman, 2008"},{"why":"Supplies the energy regularization term and sampling ideas used in the NeuralCPM loss.","marker":"Du & Mordatch, 2019"},{"why":"Provides the EBM training framework and gradient formula for neural energy functions.","marker":"Song & Kingma, 2021"},{"why":"Defines the type A and type B cell-sorting scenarios used to validate parameter learning.","marker":"Edelstein-Keshet & Xiao, 2023"}],"fun_headline_variants":["NeuralCPM learns cell energy from observations alone","Hybrid model beats analytical cell simulations","Deep network captures cell self-organization","Trained on snapshots, predicts collective cell motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training objective is a valid likelihood gradient only if the approximate parallel sampler's negative configurations come from the model's stationary distribution, but that sampler is not detailed-balance-correct, so the learned Hamiltonian rests on the unverified premise that its bias is small enough not to distort the gradient.","fun_headline_variants_meta":{"raw":{"variants":["NeuralCPM learns cell energy from observations alone","Hybrid model beats analytical cell simulations","Deep network captures cell self-organization","Trained on snapshots, predicts collective cell motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1438,"prompt_tokens":938,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":554,"tokens_out":500,"duration_ms":5468,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:12:54.528360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the learned bi-polar NeuralCPM Hamiltonian with an exact single-site Metropolis CPM sampler instead of the approximate parallel sampler and measure the axial-alignment RMSE over 12 hours; if the bi-polar organization disappears or the time course diverges from the real data, the reported success depended on the sampler's bias rather than on the learned energy function.","supporting_citations":[{"cited_title":"We distinguish the two scenarios a and b from Edelstein-Keshet & Xiao (2023), characterized by different contact energies between cells, which are laid out in tables 3 and","cited_arxiv_id":null,"evidence_quote":"Defines the cellular Potts model and the cell-sorting Hamiltonian that NeuralCPM fits and extends."}],"review_version":1}