{"id":"6617c19d-0d29-4839-991d-b0fae5c52509","arxiv_id":"2502.08062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A pix2pix neural network, called RAPTOR, mimics CONDOR tissue-organisation simulations quickly and is fitted to match real glial, fibroblast, and corneal cultures.","lead":"Researchers trained a neural network to quickly mimic a slow biophysical simulation that predicts how cells align inside engineered tissues. The fast model could let engineers test many mould designs for growing cornea, nerve, or muscle tissue before doing lab experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experimental validation does not yet establish transferable predictive power: fitting three CONDOR parameters to only two scalar outputs (Eq. 6) is underdetermined, so the claimed 'excellent agreement' may reflect curve-fitting rather than parameters that transfer to new mould designs.","rationale":"The reader's weakest assumption identified exactly the load-bearing issue: fitting three CONDOR parameters to only width and area ratios (Eq. 6) cannot be assumed to yield physically transferable parameters. My stress-test reading of the manuscript strengthens this concern with concrete evidence from the text: Section III C 1 acknowledges that fitting only width and area may explain parameter variation between glial cultures; Section III C 3 reveals that the corneal fit used only a single observable because area was not reported; and Section III B documents that the high-Delta, low-kappa region is precisely where RAPTOR fails to match CONDOR. For corneal tissue, the fitted parameters lie in that failure region, and the reported width agreement is poor (RAPTOR 0.459, CONDOR 0.525, experiment 0.35). The paper's internal evidence therefore does not support the extrapolation from 'agreement on fitted quantities' to 'validated predictive power for new mould designs.' This is a genuine soft spot, but it is not fatal to the paper's core contribution: the RAPTOR surrogate for CONDOR is validated on 445 held-out simulations with high correlations, and the authors openly acknowledge the failure region. The correct verdict remains CONDITIONAL, as the reader stated: the emulation result is credible, but the experimental-parameter-fitting claim needs substantial additional validation before the method can be presented as a predictive design tool. I agree with the reader's assessment and recommend no change to the verdict; the conditions should include a non-circular test of parameter transferability and, ideally, release of code and trained weights to allow independent checking.","tokens_in":12335,"tokens_out":4074,"duration_ms":36041,"concrete_test":"Run an identifiability and transferability test on synthetic data. Choose 50 random parameter points p_true = (Delta, kappa_NNN, kappa_NNNN) inside the trained range. For each point, run CONDOR on two distinct mould geometries (e.g., the 4-pin rectangle and the I-shape) and compute their width and area ratios. Use the paper's Eq. 6 fitting procedure with RAPTOR, fitting to only the first geometry's ratios, and record the fitted p_fit. Then: (1) identifiability: if the fitted p_fit does not consistently recover p_true, or if the spread across repeated fits is comparable to the glial Delta standard deviation in Table I, the two-observable fit cannot determine three parameters; (2) transferability: use p_fit to predict the second geometry and compare the full density, alignment, and tension fields to the CONDOR truth for that geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that parameters fitted from an experiment are physically transferable to other mould geometries and to all output fields (density, alignment, tension). The fitting procedure in Section III C minimises Eq. 6, which contains only two residuals (area ratio and width ratio), for three unknown parameters (Delta, kappa_NNN, kappa_NNNN). A three-parameter model constrained by two scalars is underdetermined. The paper itself notes in Section III C 1 that parameter variation across the three glial cultures 'may arise since the parameter fit is only made for width and area, rather than the overall shape of the tissue'; the reported spread in Delta (mean 0.2227, sigma 0.06784) supports non-uniqueness. The corneal comparison is especially weak: only width was used in the fit because area was not reported, leaving one observable for three parameters, and the fitted Delta = 0.8298 lies in the high-Delta, low-kappa region the paper identifies as the failure region in Section III B. For that case RAPTOR predicts w/w0 = 0.459, CONDOR predicts 0.525, and the experimental value is 0.35, so neither model reproduces the measured contraction well. Quantitatively, the only experimental values compared are the fitted quantities themselves; no held-out mould geometry or full field comparison (density, alignment, tension) is validated experimentally. Thus the claim of 'predictions for arbitrary mould designs' rests on an assumption of transferability that is not tested. This does not undermine the RAPTOR-as-CONDOR-surrogate result, which is supported by the 445-case test set, but it does weaken the paper's broader claim of validated experimental prediction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents RAPTOR, a conditional pix2pix generative model trained on outputs of the CONDOR biophysical model, with the three CONDOR parameters Δ, κ_NNN, and κ_NNNN supplied as additional input channels. Training data consist of roughly 3,569 CONDOR simulations with random mould and tether layouts; the held-out test split of 445 simulations yields high Pearson correlations (0.96–0.99) for mean density, area, tension, and alignment quantities. The authors then exploit the speed of RAPTOR to fit CONDOR parameters to experimental glial, fibroblast, and corneal tissues by minimizing the two-residual objective in Eq. (6), and compare the resulting RAPTOR and CONDOR width and area ratios with measurements. They report excellent agreement and propose RAPTOR as a near-real-time tool for tethered mould design.","tokens_in":12653,"tokens_out":7674,"duration_ms":61478,"significance":"If the claims hold, RAPTOR would be a practically valuable extension of CONDOR-ML: it conditions a fast surrogate on physically relevant model parameters, and the systematic grid comparison in Sec. III B usefully characterises the surrogate's domain of validity. The held-out CONDOR test set and the explicit identification of the high-Δ/low-κ failure region are genuine strengths. However, the experimental validation does not yet establish transferable predictive power: the fitted parameters are not identifiable from the two scalar observables used, and no independent experimental outcomes are predicted. The central application claim therefore needs additional evidence before the paper can be accepted.","major_comments":[{"comment":"Section III C fits the three CONDOR parameters (Δ, κ_NNN, κ_NNNN) by minimising Eq. (6), whose residual contains only two scalar quantities: relative area and relative width. Three unknowns constrained by two scalars is an underdetermined inverse problem, so the parameters are not identifiable from this fitting procedure alone. The paper itself notes in Sec. III C 1 that part of the parameter variation across the three glial cultures 'may arise since the parameter fit is only made for width and area, rather than the overall shape of the tissue,' and the reported spread in Δ (mean 0.2227, σ 0.06784) is consistent with multiple nearly equivalent optima. This is load-bearing because the proposed use of RAPTOR for mould design assumes that parameters fitted to one tissue and mould transfer to other geometries and to all output fields (density, alignment, tension). Please demonstrate identifiability, for example through profile likelihoods, joint fitting to several moulds, or leave-one-out experimental validation, before claiming transferable parameters.","section":"III C, Eq. (6)"},{"comment":"The corneal comparison in Sec. III C 3 is the weakest experimental test and does not support the statement of 'excellent agreement'. Because the area was not reported in Ref. [17], the fit uses only one observable (w/w0 = 0.35) to determine three parameters. The optimised parameters give RAPTOR w/w0 = 0.459 and CONDOR w/w0 = 0.525, so neither model reproduces the measured contraction, and the fitted Δ = 0.8298 lies near the high-Δ, low-κ_NNNN/κ_NNN boundary of the region where Sec. III B documents systematic RAPTOR/CONDOR disagreement. This case therefore illustrates the identifiability problem rather than providing an independent validation of the method.","section":"III C 3, Table I"},{"comment":"The experimental validation is partly circular: the quantities reported in Table I (relative tissue area and width) are exactly the two residuals minimised in Eq. (6), so the agreement for those quantities is a consequence of the fitting procedure rather than an independent prediction. No experimental tissue is held out for a different mould geometry, and no full-field outputs (density, alignment, tension) are compared with experiment. The strong RAPTOR-vs-CONDOR test on the held-out CONDOR test set (Fig. 2) verifies emulation of the simulator but not the physical transferability of fitted parameters. Please add at least one independent experimental validation, such as fitting on one mould and predicting a second mould, comparing a measured alignment or density field not used in the loss, or performing leave-one-mould-out analysis across the three glial cultures.","section":"III C, Table I, Figs. 9–11"}],"minor_comments":[{"comment":"The heading 'T raining Data' and the duplicated 'of' in 'A total of of 3569 unique simulations' are typos that should be corrected.","section":"II B"},{"comment":"The text in Sec. III A defines N(P < 0.2) as the number of pixels with density less than 0.2, whereas the Fig. 2 caption says 'number of pixels with density exceeding 0.2'; please clarify which convention is intended and use it consistently.","section":"III A / Fig. 2"},{"comment":"The definition of N_c as 'P ip wi,p' is garbled; it should read N_c = Σ_{i,p} w_{i,p} (or an equivalent explicit expression), and the relationship between barred and unbarred field averages should be stated cleanly.","section":"II C, Eq. (5)"},{"comment":"Table I uses 'κ3n' and 'κ4n' in the header without definition; please define these abbreviations in the caption or use κ_NNN and κ_NNNN consistently with the text.","section":"Table I"},{"comment":"The phrases 'predictions for arbitrary mould designs' (Abstract) and 'allowing predictions for arbitrary choices of parameter values' (Sec. II B) are too strong given the failure region at high Δ and low κ documented in Sec. III B; please qualify the claim to the parameter range covered by the training data.","section":"IV / Abstract"},{"comment":"The Fig. 9 caption does not identify which experimental image corresponds to G1, G2, and G3 in Table I; please add labels so the reader can connect the visual comparisons to the tabulated values.","section":"Fig. 9 / Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a useful but incremental extension of the authors' own CONDOR-ML work, and the surrogate itself appears to be well tested on the held-out CONDOR set. The main scientific risk is the experimental parameter fitting: with three parameters and two scalar residuals, the fitted parameters may not be transferable to new mould designs, and the corneal case is particularly weak. A revision that adds an identifiability analysis or an independent experimental prediction would substantially strengthen the paper. The tenocyte/myoblast paragraph in Sec. III C 4 is a placeholder and should be labelled as such or removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RAPTOR is a useful extension of the authors' earlier CONDOR-ML: it feeds delta, kappa_NNN, and kappa_NNNN into the pix2pix input channels and adds a nonlinear least-squares parameter-fitting workflow. The core emulation claim is solid—on 445 held-out CONDOR simulations the mean-field correlations are 0.96–0.99, and they also test an I-shaped mould with a continuous tethering bar that has no analogue in the training set. That is a genuine extrapolation test and it mostly passes, with the expected breakdown at high delta and low kappa, which they acknowledge.\n\nThe soft spot is the experimental validation. The three CONDOR parameters are fitted by minimizing RSS over only width and area ratios (Eq. 6), so the later \"agreement\" is partly a curve-fit of the same scalars. Three parameters from two scalars is underdetermined; the glial parameter scatter they report supports that. The corneal case is worse: only width goes into the fit for three parameters, and the fitted delta=0.8298 sits in the high-delta, low-kappa region where they themselves show RAPTOR under-predicts contraction. The predicted width ratio of 0.459 versus the experimental 0.35 is not great agreement. Missing error bars on the experimental values also make comparisons hard to judge. Saying there is \"excellent agreement\" with cultured tissues overstates what the data support.\n\nThat said, the paper is honest about several of these limitations—they note that parameter variation may arise because the fit is only to width and area, and they flag the high-contraction failure region. The RAPTOR-as-CONDOR-surrogate result stands on its own and is reproducible in principle, though no code, data, or trained weights are released, which is a real barrier to independent verification.\n\nWho is this for? Anyone using CONDOR for tissue-mould design who wants fast parameter sweeps or optimization. It is a useful engineering contribution, not a major biophysical advance. With the experimental-validation language toned down and the fitting degeneracy discussed explicitly, it would be a solid methods paper. I would send it to review.","headline":"A genuinely useful fast surrogate for CONDOR simulations, with experimental validation that is partly curve-fitting and should be reframed.","tokens_in":13287,"tokens_out":1728,"would_cite":true,"duration_ms":15203,"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":"RAPTOR makes near-real-time predictions of tissue density, alignment, and tension for engineered corneal, glial, and fibroblast cultures, validated against laboratory-grown tissues.","keywords":["RAPTOR","pix2pix conditional GAN","CONDOR biophysical model","tissue engineering","cell alignment prediction","tethered mould design","parameter fitting","3D cell culture"],"falsifier":"A concrete test is to fit the three parameters to a glial tissue grown in one mould, then use RAPTOR (or CONDOR with those parameters) to predict the density and alignment maps for a glial tissue grown in a different mould whose geometry is absent from the training set, and compare the predicted fields pixel-by-pixel with the experimental culture. If the local fields deviate systematically, especially in the high-contraction regime where the paper already reports RAPTOR under-predicts contraction, the scalar width/area fit has not pinned down transferable parameters. Equally decisive would be a numerical check for parameter non-identifiability: two different parameter triples that give the same width and area ratios but different alignment and tension maps for the same mould would show that the fit is underdetermined.","tokens_in":12105,"feed_emoji":"🧫","tokens_out":9127,"duration_ms":71948,"temperature":0.7,"pith_summary":"This paper proposes RAPTOR, a machine-learning tool that predicts how cells organise in engineered 3D tissues grown in tethered moulds, and argues that it is accurate enough across cell types to be useful for mould design. RAPTOR is a pix2pix conditional generative adversarial network trained on simulations from the CONDOR biophysical model. Unlike the earlier CONDOR-ML proof of concept, RAPTOR takes the model parameters $\\Delta$, $\\kappa_{\\mathrm{NNN}}$ and $\\kappa_{\\mathrm{NNNN}}$ as inputs, so one trained network can make predictions for different cell and matrix types rather than a single fixed parameter set. The authors validate the network against laboratory-grown glial, fibroblast and corneal tissues, and show that fitting the three parameters to a tissue's measured width and area contraction lets RAPTOR and CONDOR reproduce the observed shapes. If the transferability of these fitted parameters holds, RAPTOR's sub-second predictions make near-real-time, high-throughput design of tethered moulds feasible.","feed_headline":"Neural net predicts engineered tissue organisation almost instantly","feed_subtitle":"RAPTOR maps mould shape and cell parameters to density, alignment, and tension maps in a fraction of a second.","key_machinery":"The load-bearing mechanism is parameter-conditioned image translation. RAPTOR is a pix2pix conditional GAN: a generator and discriminator trained adversarially to translate a $256\\times256$ five-channel input (mould depression, tether placement, plus constant maps of the dimensionless cell–matrix interaction $\\Delta$ and the spring constants $\\kappa_{\\mathrm{NNN}}$ and $\\kappa_{\\mathrm{NNNN}}$) into a $256\\times256$ eight-channel output (density, six orientation products, tension). The CONDOR model supplies the ground truth: an energy functional for a contractile network of bonds, minimised by simulated annealing, whose predictions are the cells' positions, orientations and bond tensions. Because the three parameters are part of the network input, the same trained network covers a range of tissue types, and because RAPTOR runs in a fraction of a second, the network can be embedded in a nonlinear least-squares loop (Eq. 6) that fits $\\Delta$, $\\kappa_{\\mathrm{NNN}}$ and $\\kappa_{\\mathrm{NNNN}}$ to the measured width and area ratios of an experimental tissue. The fitted parameters then feed back into either RAPTOR or CONDOR for design.","core_discovery":"The central discovery claimed is that a single parameter-conditioned GAN can reproduce the output fields of a biophysical tissue model across most of its parameter space and thereby predict the organisation of real cultured tissues. RAPTOR maps five input channels — mould depression, tether placement, and constant-valued maps of $\\Delta$, $\\kappa_{\\mathrm{NNN}}$ and $\\kappa_{\\mathrm{NNNN}}$ — to eight output channels: cell density, six products of the cell-orientation Q tensor ($S_x^2$, $S_y^2$, $S_z^2$, $S_xS_y$, $S_xS_z$, $S_yS_z$), and average bond tension. Trained on 3,124 CONDOR simulations augmented to 12,653 examples, it closely matches CONDOR on a held-out test set and in grid scans over the parameter space, with Pearson correlations above 0.96 for bulk properties; the clear exception is the high-contraction corner (large $\\Delta$, small $\\kappa$), where RAPTOR under-predicts contraction. The paper further claims that nonlinear least-squares fitting of the three parameters to the width and area ratios of cultured glial, fibroblast and corneal tissues yields parameter sets for which both CONDOR and RAPTOR reproduce the experimental tissue shapes, with fitted $\\Delta$ values that are lower for glial tissue and higher for fibroblast and corneal tissue, consistent with their different contractility.","pith_inferences":["The authors leave implicit that RAPTOR could be inverted for inverse design: rather than checking a mould shape, one could optimise the mould and tether layout directly against desired alignment and tension maps, since the forward map is cheap enough to embed in an evolutionary loop.","Because the parameter fit targets only two scalar quantities, the fitted triple ($\\Delta$, $\\kappa_{\\mathrm{NNN}}$, $\\kappa_{\\mathrm{NNNN}}$) may not be identifiable; our inference is that adding shape-based or spatially resolved residuals, such as local width profiles or alignment maps, to the fit would tighten parameter estimates and make transferability claims testable.","The failure mode at high $\\Delta$ and small $\\kappa$ is localised in a corner of parameter space, which suggests a targeted remedy the authors do not pursue: oversample that corner in the training data or use a separate network specialised to high-contraction cases.","The same parameter-conditioned image-translation design could in principle be applied to other biophysical tissue models, not just CONDOR, whenever a model can generate enough training simulations; this would let experimentalists choose the cheapest or most faithful simulator and still get real-time predictions."],"forward_implications":["RAPTOR predictions run in a fraction of a second, so mould designs can be screened against density, alignment, and tension criteria without running day-long CONDOR simulations, making high-throughput and automated design practical.","The fitted parameter values for a given cell type can be reused in subsequent CONDOR or RAPTOR runs, so a small number of calibration experiments could parameterise the model for new cell lines or matrix conditions.","The network generalises to at least one mould geometry absent from the training set (a long continuous tethering bar), indicating that predictions are not limited to memorised shapes; accuracy degrades only in the high-$\\Delta$, low-$\\kappa$ corner of parameter space.","The demonstration for glial, fibroblast, and corneal tissues suggests the same workflow — train once, then fit parameters per tissue type — could be applied to other engineered tissues that can be grown in tethered moulds."],"supporting_citations":[{"why":"Supplies the CONDOR energy model and simulated-annealing solver that generates the training data and all reference simulations.","marker":"[11]"},{"why":"Established the pix2pix-based CONDOR-ML approach and the random mould/tether training-data generation procedure that RAPTOR extends.","marker":"[12]"},{"why":"Supplies the pix2pix conditional GAN architecture used as the base of RAPTOR.","marker":"[14]"},{"why":"Defines the four-pin rectangular mould geometry used in the parameter-space grid comparisons.","marker":"[15]"},{"why":"Provides the glial tissue mould designs and cultured-tissue data used for validation.","marker":"[16]"},{"why":"Provides the corneal tissue width-contraction value used in the parameter fit.","marker":"[17]"},{"why":"Provides the fibroblast tissue images from which width and area ratios were estimated for the fit.","marker":"[18]"}],"fun_headline_variants":["RAPTOR: fast tissue organisation prediction via GAN","GAN predicts tissue organisation in a fraction of a second","Machine learning forecasts tissue architecture from mould design","Neural network predicts cell alignment and tension in engineered tissues","RAPTOR uses GAN to quickly predict tissue organisation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three cell and matrix parameters, fitted to only the width and area ratios of one experimental tissue, are physically transferable: the same values predict density, alignment, and tension in other mould geometries and across the whole tissue, rather than merely reproducing the two scalar numbers used in the fit.","fun_headline_variants_meta":{"raw":{"variants":["RAPTOR: fast tissue organisation prediction via GAN","GAN predicts tissue organisation in a fraction of a second","Machine learning forecasts tissue architecture from mould design","Neural network predicts cell alignment and tension in engineered tissues","RAPTOR uses GAN to quickly predict tissue organisation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000572,"raw_usage":{"total_tokens":2748,"prompt_tokens":1031,"completion_tokens":1717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1654}},"tokens_in":647,"tokens_out":1717,"duration_ms":30903,"temperature":1.0,"reasoning_tokens":1654,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:57:11.145280+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is to fit the three parameters to a glial tissue grown in one mould, then use RAPTOR (or CONDOR with those parameters) to predict the density and alignment maps for a glial tissue grown in a different mould whose geometry is absent from the training set, and compare the predicted fields pixel-by-pixel with the experimental culture. If the local fields deviate systematically, especially in the high-contraction regime where the paper already reports RAPTOR under-predicts contraction, the scalar width/area fit has not pinned down transferable parameters. Equally decisive would be a numerical check for parameter non-identifiability: two different parameter triples that give the same width and area ratios but different alignment and tension maps for the same mould would show that the fit is underdetermined.","supporting_citations":[{"cited_title":"Bajaj, R","cited_arxiv_id":null,"evidence_quote":"Supplies the CONDOR energy model and simulated-annealing solver that generates the training data and all reference simulations."},{"cited_title":"Ben-Arye and S","cited_arxiv_id":null,"evidence_quote":"Established the pix2pix-based CONDOR-ML approach and the random mould/tether training-data generation procedure that RAPTOR extends."},{"cited_title":"Jensen, C","cited_arxiv_id":null,"evidence_quote":"Supplies the pix2pix conditional GAN architecture used as the base of RAPTOR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the four-pin rectangular mould geometry used in the parameter-space grid comparisons."},{"cited_title":"Jensen and Y","cited_arxiv_id":null,"evidence_quote":"Provides the glial tissue mould designs and cultured-tissue data used for validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the corneal tissue width-contraction value used in the parameter fit."},{"cited_title":"Andrews, H.Dickinson, and J","cited_arxiv_id":null,"evidence_quote":"Provides the fibroblast tissue images from which width and area ratios were estimated for the fit."}],"review_version":1}