{"id":"4544b45a-f7d5-45a2-a93d-aa71a0a08390","arxiv_id":"2608.03927","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A physics-flavored CNN-transformer automatically fits stretched-exponential parameters to engineered skeletal muscle contraction curves, using synthetic pre-training and unsupervised real-data alignment.","lead":"This paper trains a CNN-transformer neural network to turn force-time recordings of engineered skeletal muscle contractions into a small set of physical model parameters describing contraction, fatigue, and relaxation. The method combines synthetic training data with unlabeled real recordings through a physics-flavored loss, and reports good curve fits on healthy, aged, and Duchenne Muscular Dystrophy muscle lines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-data evaluation is self-referential: fit quality under the same stretched-exponential model does not validate extracted kinetic parameters as biologically meaningful.","rationale":"The reader's weakest assumption correctly identifies the self-referential evaluation: no parameter ground truth exists, and the real-data metric is the same physics reconstruction loss used in training. My stress-test confirms this is the load-bearing issue rather than, say, architecture choice or synthetic-data volume. The concern is not that the fits look poor—they look excellent—but that excellent fits to the assumed model do not prove the extracted parameters are biologically meaningful or uniquely determined. The paper would need an independent, out-of-model validation of parameter recovery, or a direct identifiability analysis, to support the strong biological claims in Sec. 3.5. The reader's conditional verdict already reflects this, so no verdict change is needed; the concrete synthetic-out-of-model test would either resolve or confirm the concern.","tokens_in":18083,"tokens_out":3161,"duration_ms":40804,"concrete_test":"Generate a synthetic test set from a model outside the assumed family—for example, a double stretched exponential or a two-timescale compartment/twitch model with known rate constants—with realistic 50 fps noise and amplitude/fatigue variability, covering the same visual phenotypes as the real data. Run the fully trained PFNN on these curves and report the error between predicted and known generating parameters, alongside reconstruction MSE. If PFNN achieves low reconstruction MSE yet yields systematically biased parameters (e.g., folding a fast calcium-handling timescale into τ_r or β_d), the central parameterization claim is not established. Conversely, if the generating parameters are recovered accurately, the self-referential concern is substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PFNN extracts 'physically meaningful parameters' and yields biological insights (e.g., DMD/aged tissues show slowed kinetics, healthy controls show larger β_d heterogeneity). But the manuscript itself states in Sec. 2.1 that no experimental ground truth for the kinetic parameters exists. Every real-data performance metric in Tables 2 and 3 is the reconstruction MSE between the measured curve and F(p̂), where F is the very same piecewise stretched-exponential model used as the generative prior for synthetic training and as the unsupervised physics loss in Eq. (3). The hybrid model is explicitly trained to minimize L_fit on unlabeled real data, so low residual on real curves is partly a measure of how well the model has internalized its own assumed functional form, not independent evidence that the predicted τ, β, α values correspond to true biological kinetics.\n\nThis matters concretely because stretched-exponential parameters can be non-identifiable: different combinations of τ_r, β_r, τ_d, β_d, and fatigue parameters can produce nearly indistinguishable curves, especially with noise. If the real signals contain components outside the assumed model family, the network can absorb those components into shifted parameters while still achieving low reconstruction error. The L-BFGS baseline is a single 'average-curve' initialization (Fig. 7), so beating it only shows that the network escapes one local basin; it does not establish parameter correctness. Consequently, the cell-line differences in Sec. 3.5 and the hypothesized calcium-handling interpretation rest on unvalidated parameter values obtained through a self-referential loop.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a Physics-Flavored Neural Network (PFNN) that combines a CNN-Transformer architecture with a stretched-exponential model of engineered skeletal muscle (ESM) tetanus contractions. The model is trained on 300,000 synthetic curves generated from the same stretched-exponential function, plus unlabeled real contraction curves via an unsupervised reconstruction loss, and outputs ten physical parameters. The authors report that the network accurately predicts synthetic ground-truth parameters, improves reconstruction error on real data compared to a single-initialization L-BFGS baseline, generalizes to three independent donor cell lines, and yields biological insights such as slowed kinetics in Duchenne Muscular Dystrophy and aged tissues. The core methodological contribution is the hybrid synthetic-plus-unsupervised training paradigm, which is intended to overcome the absence of labeled biological data.","tokens_in":18431,"tokens_out":4210,"duration_ms":50152,"significance":"If the extracted parameters were shown to correspond to true biological kinetics, the PFNN would be a valuable tool for high-throughput functional phenotyping of engineered muscle, replacing simple peak-force metrics. The architecture is sensible, and the synthetic-to-real adaptation strategy is a pragmatic response to label scarcity. The integrated-gradients attribution analysis is a nice attempt to show that the network attends to physiologically relevant time windows. However, the current real-data evaluation is self-referential: the metric used to validate the model on real data, MSE(y_real, F(p_hat)), is the same physics-reconstruction loss used to train the model on unlabeled real data. Low reconstruction error therefore does not establish that the predicted parameters are biologically meaningful, especially since the paper explicitly states that no experimental ground truth exists. The biological conclusions in Section 3.5 rest on this unvalidated link.","major_comments":[{"comment":"The real-data evaluation is circular. In Eq. (3), the model is trained on unlabeled real data by minimizing L_fit(B_real) = MSE(y_real, F(p_hat_real)). The validation metrics in Tables 2 and 3 are exactly this same reconstruction MSE. A low value only shows that the network has internalized the stretched-exponential model F, not that the predicted tau, beta, alpha, etc. correspond to true biological kinetics. Section 2.1 confirms that no ground-truth parameters exist. To support the abstract's claim of 'high-fidelity parameterization' and the biological insights in Section 3.5, the authors need external validation: e.g., use the known pharmacological interventions from the primary dataset (ref. 26) to check that predicted parameters shift in expected directions, compare against an independent biophysical assay, or at minimum report a formal identifiability analysis on synthetic data with","section":"Section 3.2, 3.4; Eq. (3)"},{"comment":"The synthetic evaluation only tests whether the network can invert its own generator. Both training and test sets are generated from the same stretched-exponential model F, so high R² values (0.85–1.00) demonstrate identifiability within that family, but not that F adequately represents real ESM dynamics. The 'visual indistinguishability' claim in Fig. 3 is qualitative, and Fig. 4B already shows systematic deviations in the decay regime on real data. The authors should quantify the domain shift, e.g., by comparing the distribution of residuals between synthetic and real curves, or by testing on synthetic data generated from perturbed/out-of-distribution parameters. Without this, the statement that 'synthetic data is a strong basis' is not fully supported.","section":"Section 3.1, Fig. 4"},{"comment":"The L-BFGS baseline is not a strong comparator. The initialization is a single 'average curve' (Fig. 7), and the claim that L-BFGS 'became trapped in local minima' is not established because only one starting point was tried. Classical curve fitting routinely uses multi-start or random restarts, especially for non-convex problems. The authors should compare against multi-start L-BFGS (e.g., 50–100 random initializations, reporting best-of-N) or a global optimizer. Beating a single-initialization solver shows only that the network provides a better warm start, not that the network is more accurate than classical fitting. This directly affects the headline result 'surpasses L-BFGS in terms of accuracy'.","section":"Section 3.4, Table 2, Fig. 7"},{"comment":"The biological parameter comparisons across cell lines are based on point estimates without uncertainty quantification or statistical significance testing. The PCA clustering in Fig. 9 could arise from differences in noise, curve shape outside the model family, or non-identifiable parameter combinations, rather than from true physiological differences. The authors should provide confidence intervals for the predicted parameters (e.g., via ensembling or dropout) and perform proper statistical tests (e.g., mixed-effects models accounting for multiple curves per tissue) before claiming that DMD and aged tissues show 'significantly decayed' kinetics or that beta_d heterogeneity differs. The manuscript's own caveat that only three cell lines were studied does not address this more fundamental issue.","section":"Section 3.5, Fig. 9"}],"minor_comments":[{"comment":"Equation (2) contains a typo: the last term should be MSE(y_real, F(p_hat_real)), not MSE(y_real, p_hat_real). Equation (3) is correct, but the inconsistency is confusing.","section":"Eq. (2), Section 2.3.2"},{"comment":"The integrated-gradient attribution scores are described qualitatively. Please specify how scores were normalized across samples, how the average was computed, and whether any statistical threshold was used to define 'sharp' peaks.","section":"Figure 6"},{"comment":"The Limitations section discusses only computational complexity and sampling-rate flexibility; it does not mention the absence of ground-truth validation for real-data parameters or the potential non-identifiability of the stretched-exponential parameters. Given that this is the central risk to the biological interpretation, it should be explicitly acknowledged.","section":"Section 5, Limitations"},{"comment":"There are several typographical errors: 'intragration' (Introduction), 'unsuperwised' (Fig. 2), 'paramters' (Fig. 9), 'appraoch' (Conclusion). A careful proofread is needed.","section":"General"},{"comment":"A, t0, and C are listed as 'Unconstrained' although the physical model likely expects A and C to be non-negative. Please clarify or add appropriate transformations.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising architecture and a clear practical motivation, but the central claim of 'physically meaningful parameterization' is not yet supported by the evidence. The real-data evaluation is self-referential, and the biological conclusions in Section 3.5 go beyond what can be inferred from reconstruction error alone. The authors should be encouraged to add external validation using the known pharmacological perturbations from their prior work, or an identifiability analysis, plus a stronger L-BFGS baseline and uncertainty quantification. These are substantial but feasible additions that would determine whether the approach is a genuine advance or merely a good curve-fitting engine."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this if you care about ML methods for biophysical time-series, but keep your hand on the salt. The paper is a solid engineering contribution: a CNN-Transformer that predicts stretched-exponential parameters of muscle contraction curves, trained on synthetic data and then self-aligned on unlabeled real recordings. The architecture is an assembly of known pieces—CNN features, ViT-style parameter tokens, physics-based reconstruction loss—but the combination is new and it clearly works in the narrow sense: the model fits real curves well and the hybrid training beats synthetic-only by a factor of 2.5 in MSE. The synthetic R2 numbers (0.85–1.00) show the network can invert its own generator, which is necessary but not sufficient. The attribution analysis is a nice touch and supports the claim that the network attends to the right temporal regions.\n\nThe soft spot is the validation loop. Because no ground-truth parameters exist for real data (the paper says so in Sec. 2.1), every real-data metric is the same reconstruction MSE used in training. Low residual means the model has internalized its assumed functional form, not that the predicted τ, β, α correspond to biology. The L-BFGS baseline is a single \"average curve\" initialization, so beating it only shows the network escapes one local basin. The validation set was also used for hyperparameter tuning, and no code or data are released, which limits reproducibility. The cell-line comparisons and the calcium-handling interpretation rest on these unvalidated parameter values; the authors are appropriately cautious about the small cohort, but the headline claim of \"high-fidelity parameterization\" overreaches.\n\nNone of this is fatal. The fix is straightforward: use a held-out test set for the primary data, compare against multi-start L-BFGS or a properly initialized optimizer, and ideally validate parameters against an independent biological readout (e.g., known drug effects or calcium imaging). Releasing code and data would also help a lot.\n\nWho should read it: people working on muscle tissue engineering or automated phenotyping, and ML researchers interested in physics-informed networks for small biological datasets. It deserves a serious referee, not a desk reject, but the referee will need to push on the validation. I'd bring it to a reading group if you want a concrete case study of self-referential evaluation in PINNs.\n\nRecommendation: send to peer review with major-revision expected.","headline":"A competent engineering pipeline for automated kinetic phenotyping, but the real-data validation is self-referential and the biological claims outrun the evidence.","tokens_in":18913,"tokens_out":3631,"would_cite":false,"duration_ms":36200,"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":"The paper claims a physics-flavored neural network can turn raw force-time curves of engineered muscle into a small set of biologically meaningful stretched-exponential parameters, and that it does so more accurately than conventional itera","keywords":["physics-informed neural network","stretched exponential","skeletal muscle tissue engineering","tetanus contraction","kinetic phenotyping","self-supervised learning","Duchenne muscular dystrophy","force-time analysis"],"falsifier":"Simultaneously record calcium transients and pillar force in the same engineered muscle contraction, compute the network's predicted τ_d and β_d on the force curve, and compare them with the independently measured decay of cytosolic calcium; if the predicted relaxation kinetics do not track the calcium signal, the parameters are curve-shape descriptors rather than biological kinetics.","tokens_in":17967,"feed_emoji":"💪","tokens_out":6406,"duration_ms":71827,"temperature":0.7,"pith_summary":"Engineered skeletal muscle tissues are usually characterized by a single number, peak force, which throws away the shape of the contraction. This paper tries to establish that a hybrid CNN-Transformer can instead read the entire force-time profile and output kinetic parameters — effective timescales, heterogeneity exponents, fatigue onset and slope — that capture the biology of contraction. The key move is to use the same stretched-exponential physical model twice: first to generate hundreds of thousands of synthetic labeled curves for pre-training, then as a differentiable reconstruction loss that lets unlabeled real recordings be absorbed into training. If the claim holds, contraction phenotyping becomes automated, scalable, and rich enough to distinguish disease and aging states from curve shape alone.","feed_headline":"Neural net extracts muscle kinetics from force curves, besting L-BFGS","feed_subtitle":"A physics-flavored model separates healthy, Duchenne, and aged engineered muscle from force-time profiles alone.","key_machinery":"The central object is a piecewise stretched-exponential force model: a rise phase F_rise(t)=A(1−exp(−((t−t0)/τ_r)^β_r)), a smoothed fatigue factor Φ(t), and a decay phase F_decay(t)=C+(A_c−C)exp(−((t−t1)/τ_d)^β_d). This model serves double duty as the generator of synthetic training data and as the physics-based reconstruction loss. The network architecture is a CNN-Transformer in which ten learnable parameter tokens, prepended to a convolutionally extracted embedding, query the whole signal via self-attention and are mapped to the ten relaxed physical parameters. The physics loss forces the tokens to align with the kinetic constants of the model, filtering out non-physical noise while prese","core_discovery":"The paper claims that a CNN-Transformer network, trained with a hybrid objective, can take raw force-time recordings of engineered skeletal muscle tetani and output the parameters of a piecewise stretched-exponential contraction model — the aggregate timescale τ, the stretching exponents β for rise and decay, the fatigue onset and slope, and the absolute force scale — with enough fidelity that the reconstructed curves match held-out experimental recordings better than a standard trust-region fitter. The authors further claim that because the physics model is used both to generate a large synthetic training corpus and as a reconstruction loss on unlabeled real data, the network self-aligns to","pith_inferences":["The self-training recipe is not muscle-specific: any biosignal with a trusted phenomenological model could be pre-trained synthetically and self-aligned on unlabeled recordings, so the strategy plausibly transfers to cardiac twitch or calcium-transient analysis.","A direct test of biological validity would compare the network's relaxation parameters against an orthogonal measurement, such as simultaneous calcium imaging or stiffness measurements, because no ground-truth parameters exist for real tissues.","The β_d shift between healthy and Duchenne lines suggests a quantitative marker of relaxation heterogeneity, but with only three cell lines this is a hypothesis that needs a larger cohort to confirm."],"forward_implications":["Kinetic phenotyping no longer requires hand-picked initial guesses: the network supplies a learned prior that lands near the optimum, and post-hoc iterative refinement adds only marginal gains.","Unlabeled contraction data from new experiments can be pulled into training automatically through the reconstruction loss, so the pipeline improves as more measurements accumulate without manual labeling.","The extracted parameter vectors, rather than peak force alone, become the comparison space for drug screens and disease models, with cell lines separating cleanly in principal-component space.","Because the architecture handles variable-length signals, recordings of different durations can be analyzed without cropping or interpolation, easing integration across experimental setups.","Generalization to three independent human donor lines, including a Duchenne muscular dystrophy line, indicates the parameterization transfers across biological contexts without retraining."],"supporting_citations":[{"why":"Supplies the piecewise stretched-exponential contraction model, the pharmacological dataset of 370 tetani, and the prior demonstration that these parameters capture drug-specific kinetics.","marker":"26"},{"why":"Describes the pillar-based force-recording system and the AB1167 healthy cell line used as a generalization test set.","marker":"36"},{"why":"Introduces physics-informed neural networks, the methodological template for the physics-based reconstruction loss.","marker":"27"},{"why":"Provides the Vision Transformer design with prepended tokens, adapted here for parameter regression.","marker":"34"},{"why":"Establishes self-attention and positional encodings that let the model attend to parameter-relevant temporal windows.","marker":"58"},{"why":"Defines the Kohlrausch stretched-exponential function underlying the contraction model.","marker":"42"},{"why":"Gives the representation of stretched-exponential relaxation as a continuous sum of exponentials, justifying the heterogeneity interpretation.","marker":"50"},{"why":"Supplies integrated gradients, the attribution method used to show temporal co-localization of parameter predictions.","marker":"60"}],"fun_headline_variants":["Physics-flavored transformer beats L-BFGS on muscle curves","Physics-AI bests classical fitter on muscle force-time data","AI extracts muscle kinetics from force curves, beats old fitter","Physics-flavored AI nails muscle contraction parameters","Physics-aware neural net outperforms classical muscle fitter"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The piecewise stretched-exponential equation is assumed to describe real tetanus contraction dynamics so faithfully that minimizing reconstruction error within that same model yields biologically meaningful kinetic parameters.","fun_headline_variants_meta":{"raw":{"variants":["Physics-flavored transformer beats L-BFGS on muscle curves","Physics-AI bests classical fitter on muscle force-time data","AI extracts muscle kinetics from force curves, beats old fitter","Physics-flavored AI nails muscle contraction parameters","Physics-aware neural net outperforms classical muscle fitter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002078,"raw_usage":{"total_tokens":7919,"prompt_tokens":743,"completion_tokens":7176,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":7093}},"tokens_in":487,"tokens_out":7176,"duration_ms":50940,"temperature":1.0,"reasoning_tokens":7093,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:33:25.533575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simultaneously record calcium transients and pillar force in the same engineered muscle contraction, compute the network's predicted τ_d and β_d on the force curve, and compare them with the independently measured decay of cytosolic calcium; if the predicted relaxation kinetics do not track the calcium signal, the parameters are curve-shape descriptors rather than biological kinetics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the piecewise stretched-exponential contraction model, the pharmacological dataset of 370 tetani, and the prior demonstration that these parameters capture drug-specific kinetics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the pillar-based force-recording system and the AB1167 healthy cell line used as a generalization test set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces physics-informed neural networks, the methodological template for the physics-based reconstruction loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Kohlrausch stretched-exponential function underlying the contraction model."},{"cited_title":"Stretched Exponential Relaxation","cited_arxiv_id":"1808.00881","evidence_quote":"Gives the representation of stretched-exponential relaxation as a continuous sum of exponentials, justifying the heterogeneity interpretation."}],"review_version":1}