{"id":"fd6d68f0-cb2b-4d1f-9476-7cbc445d0fa4","arxiv_id":"2501.08356","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"A population of about 1,180 Shannon-model variants, each fitted to the full fluorescence voltage waveform of a single rabbit ventricular myocyte, reproduces action potential durations cell-by-cell and reveals weak parameter correlations.","lead":"Researchers fit a detailed heart cell model to voltage recordings from 1,228 rabbit cells, creating a population of nearly 1,180 cell-specific models. The work shows that standard fitting methods can scale to thousands of cells, which matters for drug safety testing and personalized heart models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The population-level 'phenotype' claim rests on single non-prepaced AP fits, yet the authors' own §3.5 check shows prepacing shifts parameters by factors up to 32 (λ_tos=32, λ_Ks=0.2); until this protocol dependence is shown to be cell-independent, M cannot be regarded as a random sample of…","rationale":"The reader's weakest assumption—that the single-waveform, fixed-initial-condition protocol makes M a reflection of the fitting protocol rather than of cell phenotype—is also the most load-bearing concern I can find, and the paper's own Section 3.5 provides the sharpest evidence: mean parameter shifts of factors 32, 0.20, 2.87 and 0.52 under 600-beat prepacing. I considered three other candidate concerns. (1) Circularity: the cell-by-cell biomarker match (R² up to 0.99) is mostly a restatement of the waveform fit, but the paper is transparent about this ('as expected,' Supplementary Figure 8) and it does not affect the feasibility claim. (2) The post hoc acceptance threshold γ=0.3: this only affects the ~4% rejected cells and is secondary. (3) The invalid i.i.d.-noise assumption behind the chi-squared p-values: the authors flag the autocorrelation themselves and the p-values are not the primary evidence. None of these is as damaging to the abstract's central 'phenotype' framing as the prepacing effect. I give credit where it is due: the synthetic-data test shows parameter recovery with relative errors O(10⁻²), the Bayesian test on nine cells confirms MAP/MLE agreement, and the deposited code and data make the prepacing check I propose directly executable. My recommendation is UNCHANGED because the reader's CONDITIONAL verdict already captures exactly this concern; my analysis reinforces it rather than moving it. The paper should be published only with the population-as-phenotype claim either supported by the proposed random-subset prepacing check or reframed as a property of the no-prepacing fitting protocol.","tokens_in":25003,"tokens_out":12145,"duration_ms":120838,"concrete_test":"Using the deposited Zenodo code and data (doi:10.5281/zenodo.11191649), take a random subset of at least 30 accepted fits (not the 9 hand-picked cells) and re-run the Section 3.5 prepacing test exactly as reported: refit each cell with 600 conditioning beats at 2 Hz and compute per-cell ratios λ_i = α̂_prep,i/α̂_i for all eight estimands. The concern is settled by the spread of these ratios: if the λ_tos (or λ_Ks, λ_Kr) values across the random subset show large cell-to-cell variability (e.g., interquartile range spanning more than a factor of 3, or coefficient of variation above 0.5), then the multiplicative correction in Eq. (16) is invalid, the reported population M is an artifact of the no-prepacing protocol, and the 'random sample from phenotype' claim fails. If the ratios cluster tightly around the Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: (i) full AP waveforms can be fitted cell-by-cell at scale, and (ii) the resulting population M is 'a random sample from the phenotype of healthy rabbit ventricular myocytes' (abstract; also §3.4). Part (i) is well supported by the synthetic-data test (§3.2), the Bayesian check (§3.3), and the deposited Zenodo artifacts. Part (ii) is the load-bearing weakness. The experimental recordings are steady-state: cells are field-stimulated for 5 min at 2 Hz and 5 APs are averaged (§2.2). Fitting, however, uses a single non-prepaced AP starting from baseline Shannon initial conditions, with all non-estimated initial states fixed to baseline values (§2.3–2.4). The authors themselves quantify the resulting bias in §3.5 and Supplementary Table 7: refitting nine cells with 600-beat prepacing changes mean estimates by λ_tos=32, λ_Ks=0.20, λ_Kr=2.87, λ_Clb=0.52, λ_NaK=0.25 (Eq. 16). A factor of 32 on G_tos means the non-prepaced fits are absorbing the initial-condition transient rather than describing the cell's steady-state phenotype. The proposed remedy, multiplying all estimates by the mean λ-ratios, assumes the ratios are cell-independent; yet it is validated only on nine hand-picked cells chosen 'so as to have AP waveforms ranging from relatively short to relatively long,' which is not a random sample and cannot support a population-wide correction. All distributional results (Table 1, Figs. 5–7) and the forward predictions (e.g., the Ca2+ biomarkers in Table 2) use the uncorrected estimates. The authors are transparent about this limitation in §3.5 and in the Conclusion, but the abstract's phenotype claim is unqualified. Whether the concern lands therefore hinges on whether the λ-ratios are approximately constant across a random sample of cells.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a maximum-likelihood pipeline for fitting the Shannon et al. (2004) rabbit ventricular myocyte action potential model to fluorescence voltage recordings, estimating eight scaled ionic conductances/pump currents plus a noise variance for each cell. The method is applied to 1228 myocytes, of which 1180 fits are accepted using a goodness-of-fit threshold p > 0.3, producing a population M of cell-specific model variants. The authors report population-level parameter distributions, pairwise correlations, marginal dependencies of APD biomarkers on parameters, and cell-by-cell agreement of APD30/APD50/APD90 (R² = 0.854, 0.975, 0.990). They interpret M as a random sample from the phenotype of healthy rabbit ventricular myocytes. Validation includes a synthetic-data recovery test, a Bayesian MAP comparison on nine cells, and deposition of code and data at Zenodo.","tokens_in":25473,"tokens_out":4112,"duration_ms":45586,"significance":"If the population-level interpretation is accepted, the paper would be a substantial step toward combining sample-specific and population-based modelling, showing that full action potential waveforms can be fitted at scale and that the resulting population reproduces measured biomarker distributions and individual-cell values. The paper has notable strengths: machine-checkable open code and data, a synthetic-data test with parameter recovery errors of order 1%, a Bayesian cross-check of point estimates, and a large experimental dataset. The central analytical claim, however — that M is a random sample from the healthy myocyte phenotype — is not currently supported, because the fits use a single non-prepaced action potential while the measurements are taken from prepaced cells, and the authors' own prepacing check shows shifts up to a factor of 32 in one estimated conductance. With the claim appropriately reframed or the protocol-dependence corrected, the contribution remains valuable for computational electrophysiology.","major_comments":[{"comment":"The prepacing check directly undermines the abstract's central claim that M is 'a random sample from the phenotype of healthy rabbit ventricular myocytes.' The experimental cells are field-stimulated for five minutes at 2 Hz before recording (§2.2), but the fits start from baseline Shannon initial conditions and use a single non-prepaced AP (§2.3–2.4). The authors' own re-fits of nine cells with 600-beat prepacing give mean ratios λ_tos = 32, λ_Ks = 0.2, λ_Kr = 2.87, λ_Clb = 0.52, λ_NaK = 0.25. A factor of 32 on G_tos means the non-prepaced estimates are largely absorbing the initial-condition transient rather than describing a steady-state cell phenotype. The proposed multiplicative correction assumes these ratios are cell-independent, but they are estimated from nine hand-picked cells chosen 'so as to have AP waveforms ranging from relatively short to relatively long,' which is not a random sample and cannot support a population-wide correction. All distributional statements (Table 1, Figs. 5–7) and forward predictions inherit this problem. I request either a substantially larger and randomized prepacing validation, or a clear reframing of M as conditional on the non-prepaced protocol, with the 'random sample from phenotype' claim removed or heavily qualified.","section":"§3.5, Eq. (16)"},{"comment":"The noise-model assumption of independence is contradicted by the paper's own residual analysis: Supplementary Figure 3(b) shows non-negligible autocorrelation of residuals over 10 to 15 lags. The likelihood (Eq. 4), the chi-square goodness-of-fit p-values (Eq. 8), the standard errors (Eq. 6), and the acceptance threshold all rely on this assumption. The authors acknowledge the violation and suggest an autoregressive noise model as a future refinement, but they do not quantify how the violation affects the reported p-values, standard errors, or membership in M. Because the acceptance of 1180 of 1228 fits and the uncertainty bars in Figures 1 and 5 depend on these quantities, the impact of the autocorrelation should be assessed, for example by fitting an AR(1) or ARMA noise model on a subset of cells and reporting changes in estimates, standard errors, and acceptance rates.","section":"§2.1, Eq. (3); §3.1"},{"comment":"The headline cell-by-cell match of APD30, APD50, and APD90 (R² = 0.854, 0.975, 0.990) is partly by construction: these biomarkers are deterministic summaries of the full voltage waveform used in the fitting objective. The high R² therefore demonstrates internal consistency of the fits rather than an independent prediction. The contrast with earlier population-calibration studies, which used only APD90 to calibrate, is valid as a statement about fitting the full waveform, but not as evidence of predictive superiority. To substantiate the 'also match experimental biomarker values on a cell-by-cell basis' claim as a predictive advance, the authors should validate the fitted models against data not used in fitting, such as a held-out portion of the waveform, a second AP from the same cell, or independently measured biomarkers like calcium transients or APD restitution.","section":"§3.4, Figure 8"},{"comment":"The acceptance threshold γ = 0.3 is described as 'selected by comparison with the goodness-of-fit values of the fits shown in Figure 1,' i.e., chosen post hoc from nine examples. Since the chi-square p-values themselves are affected by the noise-model misspecification noted above, the composition of M (1180 cells) depends on an ad hoc threshold. The paper should report the sensitivity of the population statistics (Table 1, Figs. 5–7) to the choice of γ over a plausible range, and ideally prespecify an acceptance rule or justify the threshold on statistical grounds rather than by visual inspection of nine cells.","section":"§3.4"}],"minor_comments":[{"comment":"The fluorescence-to-voltage mapping is anchored to the baseline model's plateau (0 mV) and rest (−86 mV) values. The perturbation test was performed on only one cell with ±1 mV standard deviations; the population-level effects of this mapping choice remain unquantified. Reporting the same sensitivity on a random subset of cells would strengthen the uncertainty analysis.","section":"§3.5, procedural uncertainty"},{"comment":"The synthetic-data test uses a single randomly drawn parameter vector and a single noise realization. The reported relative errors of order 10⁻² would be more convincing if repeated over multiple synthetic cells with different parameter draws and noise realizations, so that the distribution of recovery errors could be described rather than a single example.","section":"§3.2"},{"comment":"The text states that 'on average, six out of the eight estimates are obtained with small uncertainty,' but Figure 1 shows that different parameters have large standard errors in different cells. It would be useful to report, for the full population, the fraction of cells for which each parameter is estimated with relative standard error below some threshold, since this bears on identifiability claims.","section":"§3.1, Figure 1"},{"comment":"The uni-variate regressions have R² ≈ 0.1 and the multivariate regressions R² ≈ 0.6, yet the text and figures emphasize p-values 'indicating high statistical significance.' With N = 1180, statistical significance is expected for small effects; the discussion would be clearer if it focused on effect sizes and prediction error rather than p-values.","section":"§3.4, Figure 7"},{"comment":"There are occasional typographical errors, e.g., 'wafevorms' in §2.2 and 'rererences' in the caption of Table 2. These do not affect the science but should be corrected in revision.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has real value as a large-scale, openly documented fitting study, and the synthetic and Bayesian checks are commendable. My concern is not with the fitting pipeline itself but with the population-level interpretation: the non-prepaced fitting protocol, combined with the authors' own prepacing sensitivity results, means that M cannot currently be regarded as a random sample from the healthy myocyte phenotype. This is fixable by reframing the claims and/or strengthening the prepacing validation, so I recommend major revision rather than rejection. I would also encourage the editor to ensure that the revised version distinguishes fitting quality from predictive validation when discussing the cell-by-cell biomarker agreement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the genuinely good news: this paper demonstrates that you can fit full action potential waveforms to a standard rabbit ventricular model cell-by-cell at a scale nobody has done before. 1180 accepted fits out of 1228 cells, with synthetic-data recovery to ~1% and Bayesian MAP agreeing with MLE. That is a real methodological milestone, and the code/data deposit makes it reproducible.\n\nThe soft spot that matters is the population-level interpretation. The abstract calls M 'a random sample from the phenotype of healthy rabbit ventricular myocytes.' But the fits are to a single non-prepaced action potential starting from baseline state variables, whereas the recordings are steady-state after 5 minutes of 2 Hz pacing. The authors themselves quantify the damage in §3.5: refitting nine cells with 600 prepacing beats changes mean estimates by λ_tos=32, λ_Ks=0.2, λ_Kr=2.87, λ_Clb=0.52. A factor of 32 on G_tos says the non-prepaced fits are absorbing the initial-condition transient, not describing steady-state phenotype. Their proposed remedy—multiply all estimates by the mean λ ratios—assumes the ratios are cell-independent, but the nine cells were hand-picked to cover short-to-long AP waveforms, not sampled at random. So the distributional results in Table 1 and Figures 5–7, and the forward Ca²⁺ predictions, rest on uncorrected estimates that are demonstrably protocol-dependent. That doesn't kill the paper's central claim—you can fit waveforms at scale—but it does mean the 'population as phenotype' claim is not supported.\n\nSecondary issues: the acceptance threshold γ=0.3 is post hoc, picked after looking at nine fits; the noise independence assumption is violated (residual autocorrelation across 10–15 lags), so the χ² p-values are more lenient than reported; and the R²=0.99 for APD90 is a circularity, since APD90 is a summary of the voltage trace you fitted. The latter is fine as a consistency check, but it shouldn't be headline 'matching'.\n\nCredit where due: they are transparent about the prepacing limitation, they ran the check, and they deposit code and data. That is better than most papers in this area.\n\nRecommendation: send to serious peer review. The right outcome is probably major revision: reframe or re-support the phenotype claim, ideally by running prepaced fits on a random subset of cells or explicitly conditioning the population on the protocol. The fitting pipeline and scale are solid and worth publishing.","headline":"A genuinely scalable cell-by-cell fitting pipeline, but the 'random sample of phenotype' claim is not yet supported—the paper deserves serious review with a required fix.","tokens_in":26103,"tokens_out":2388,"would_cite":true,"duration_ms":23070,"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":"This paper reports that a population of 1,180 cell-specific action-potential models, each fitted to a single fluorescence voltage trace from a rabbit ventricular myocyte, reproduces measured biomarker values cell by cell, matching the…","keywords":["cellular excitability","rabbit ventricular myocytes","fluorescence voltage measurements","action potential waveform","parameter estimation","population of models","inter-cell variability","maximum likelihood"],"falsifier":"Refit a random sample of the 1,180 accepted cells with 600-conditioning-beat prepacing and compare the two parameter populations: if the prepaced estimates differ from the reported ones by the order of the mean ratios in Eq. (16) ($\\lambda_{\\mathrm{tos}} \\approx 32$, $\\lambda_{\\mathrm{Ks}} \\approx 0.2$, $\\lambda_{\\mathrm{Kr}} \\approx 2.87$, $\\lambda_{\\mathrm{Clb}} \\approx 0.519$), then the fitted population is protocol-dependent and the claim that it samples the healthy myocyte phenotype fails.","tokens_in":24768,"feed_emoji":"⚡","tokens_out":14695,"duration_ms":121882,"temperature":0.7,"pith_summary":"The paper seeks to show that whole action-potential waveforms can be fitted at scale: for each of 1,228 rabbit ventricular myocytes, a standard ionic model is calibrated by adjusting eight ion-current parameters plus the measurement noise, producing 1,180 accepted cell-specific model variants. Earlier population approaches reproduced only aggregate biomarker histograms, but here every fitted model is tied to one recorded cell, and the resulting models match measured action-potential durations at 30%, 50% and 90% repolarisation on a cell-by-cell basis ($R^2 = 0.854$, $0.975$, $0.990$). The authors interpret the population of parameter estimates as a random sample from the phenotype of healthy rabbit ventricular myocytes and report that the parameters are only weakly correlated, so action-potential duration does not depend strongly on any single ion current. If the claim holds, the practical payoff is a route to cell-specific cardiac models for studying inter-cell variability, drug responses, and unmeasured quantities such as intracellular calcium.","feed_headline":"Nearly 1,200 heart-cell models match each cell's action potential","feed_subtitle":"One recording per cell recovers cell-specific ion-current settings and matches repolarisation timing to R² = 0.99.","key_machinery":"The machinery is a Gaussian maximum-likelihood fit of the Shannon et al. (2004) model: for each trace of 5,000 voltage samples the algorithm searches over nine quantities $\\theta = (G_{\\mathrm{Kr}}, G_{\\mathrm{Ks}}, G_{\\mathrm{K1}}, G_{\\mathrm{tos}}, G_{\\mathrm{CaL}}, G_{\\mathrm{Clb}}, I_{\\mathrm{NaK}}, I_{\\mathrm{NaCa}}, \\sigma)$ using the covariance-matrix-adaptation evolution strategy, then reports each estimate with a standard error from the Jacobian of the fit and accepts it when the chi-squared goodness-of-fit probability exceeds 0.3. The union of accepted point estimates is the population $M$. Because every accepted model is paired with one biological cell, this machinery is what converts raw fluorescence recordings into cell-by-cell biomarkers and parameter distributions, and it is the basis for all downstream statements about phenotype variability.","core_discovery":"On its own terms, the paper's central claim is that maximum-likelihood estimation can recover a cell-specific version of the Shannon et al. (2004) rabbit ventricular myocyte model from a single noisy fluorescence voltage trace, and that doing this for thousands of cells yields a population that matches experiments both in aggregate and per cell. The estimands are the maximal conductances of $I_{\\mathrm{Kr}}$, $I_{\\mathrm{Ks}}$, $I_{\\mathrm{K1}}$, $I_{\\mathrm{tos}}$, $I_{\\mathrm{CaL}}$ and $I_{\\mathrm{Clb}}$, the maximal densities of $I_{\\mathrm{NaK}}$ and $I_{\\mathrm{NaCa}}$, and the noise standard deviation $\\sigma$; all other model parameters and initial conditions stay at baseline. Fitting 1,228 cells and accepting 1,180 fits with goodness-of-fit $p > 0.3$ gives action-potential durations at 30%, 50% and 90% repolarisation with coefficients of determination $R^2 = 0.854$, $0.975$ and $0.990$ against experimental values. The paper validates the pipeline on synthetic data with known parameters and, for nine cells, by Bayesian posterior sampling. It concludes that the accepted population is a random sample from the healthy rabbit ventricular myocyte phenotype and that fitting entire action-potential waveforms at scale is feasible.","pith_inferences":["The prepacing check the paper reports (mean ratios $\\lambda_{\\mathrm{tos}} = 32$, $\\lambda_{\\mathrm{Ks}} = 0.2$, $\\lambda_{\\mathrm{Kr}} = 2.87$, $\\lambda_{\\mathrm{Clb}} = 0.519$) implies the fitted population may encode the single-waveform protocol as much as the cell phenotype, so the 'random sample from a healthy phenotype' reading should be verified before drug-response use.","A direct test would be to refit the same cells with a train of paced beats or with paired pre-drug/post-drug traces; if estimates shift by the reported factors, additional conditioning protocols are needed to separate protocol from phenotype.","The same pipeline could be applied to alternative rabbit myocyte models, and a model-selection comparison would show which parameter variations survive across model structures and which are artefacts of the chosen baseline.","The paired dofetilide-response measurements announced but not analysed here make this population a natural substrate for inferring drug pharmacodynamics once the initial-condition ambiguity is resolved."],"forward_implications":["Whole-waveform fitting at scale becomes feasible: roughly 30 minutes per cell on 96 threads, so populations of thousands of cell-specific models can be built from multi-cell recordings.","Population-level biomarker ranges and distributions are reproduced even for biomarkers such as the duration at 30% repolarisation that earlier histogram-calibration approaches did not target.","Each fitted model can predict unmeasured cellular quantities, including ionic current densities and intracellular calcium biomarkers that fall within published experimental ranges.","Weak correlations among the estimated parameters imply that action-potential duration is spread across multiple currents, so a single conductance will not explain most of the observed variability.","The synthetic-data and Bayesian checks give confidence that reported fits are accurate even where classical standard errors are conservative."],"supporting_citations":[{"why":"Supplies the baseline rabbit ventricular myocyte model whose eight parameters are estimated for every cell.","marker":"Shannon et al. (2004)"},{"why":"Provides the fluorescence recording protocol and the earlier sensitivity analysis that selected the eight estimands.","marker":"Lachaud et al. (2022)"},{"why":"Defines the covariance-matrix-adaptation evolution strategy used to maximise the likelihood for each cell.","marker":"Hansen (2006)"},{"why":"Supplies the probabilistic inference software used for optimisation and Bayesian sampling.","marker":"Clerx et al. (2019)"},{"why":"Provides the CVODES integrator used to solve the model in every likelihood evaluation.","marker":"Hindmarsh et al. (2005)"},{"why":"Supplies the adaptive MCMC sampler used in the Bayesian validation of nine fits.","marker":"Bardenet et al. (2015)"},{"why":"Represents the earlier population-calibration approach that the paper contrasts with cell-by-cell matching.","marker":"Britton et al. (2013)"},{"why":"Exemplifies calibrating populations of models to data density, the approach this study extends with cell-specific constraints.","marker":"Lawson et al. (2018)"}],"fun_headline_variants":["1,180 heart-cell models match each cell's repolarisation","1,180 myocyte models replicate individual action potentials","Population of 1,180 models fits rabbit heart cells cell-by-cell","Fluorescence voltages yield 1,180 cell-specific cardiac models","Single traces fit 1,180 cell-specific action potential models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that one beat per cell, recorded without first conditioning the model into a steady rhythm, contains enough information to identify each cell's ionic phenotype, even though its own re-fits with 600 conditioning beats change some parameters by factors from 0.2 to 32.","fun_headline_variants_meta":{"raw":{"variants":["1,180 heart-cell models match each cell's repolarisation","1,180 myocyte models replicate individual action potentials","Population of 1,180 models fits rabbit heart cells cell-by-cell","Fluorescence voltages yield 1,180 cell-specific cardiac models","Single traces fit 1,180 cell-specific action potential models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1600,"prompt_tokens":1059,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":454}},"tokens_in":675,"tokens_out":541,"duration_ms":5421,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:36:34.153117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit a random sample of the 1,180 accepted cells with 600-conditioning-beat prepacing and compare the two parameter populations: if the prepaced estimates differ from the reported ones by the order of the mean ratios in Eq. (16) ($\\lambda_{\\mathrm{tos}} \\approx 32$, $\\lambda_{\\mathrm{Ks}} \\approx 0.2$, $\\lambda_{\\mathrm{Kr}} \\approx 2.87$, $\\lambda_{\\mathrm{Clb}} \\approx 0.519$), then the fitted population is protocol-dependent and the claim that it samples the healthy myocyte phenotype fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the baseline rabbit ventricular myocyte model whose eight parameters are estimated for every cell."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fluorescence recording protocol and the earlier sensitivity analysis that selected the eight estimands."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the covariance-matrix-adaptation evolution strategy used to maximise the likelihood for each cell."},{"cited_title":"L., Ghosh, S., Mirams, G","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic inference software used for optimisation and Bayesian sampling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive MCMC sampler used in the Bayesian validation of nine fits."}],"review_version":1}