{"id":"37e907be-7928-49ab-9f1e-ca89eef36756","arxiv_id":"2507.00929","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A new integrated model couples DNA double-strand break simulations with a microdosimetric survival model, reproducing proton and X-ray cell survival data.","lead":"This paper combines two existing simulation tools, one for DNA damage and one for cell survival, to predict how radiation kills cells. The combined model is tested against lab data for X-rays and protons, and it reproduces the observed survival patterns across a range of radiation energies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cross-LET prediction rests on the a priori DBSCAN domain rule (Section 2.3): Eq. (2) treats each 1-um, same-chromosome DSB cluster as an independent well-mixed domain, and the paper never tests how RBE10 changes when that rule is varied.","rationale":"The paper is a serious integration effort, and the H460 design--fit once at 11.1 keV/um, predict the rest--is a legitimate out-of-sample test. The reader's conditional verdict is appropriate. My stress-test agrees with the reader's weakest-assumption identification: the definition of a 'domain' is the hinge. In GSM2, domains are the units within which sublethal lesions interact; replacing them with DBSCAN clusters makes cluster statistics the only LET-dependent input to Eq. (2). Because the 1 um and same-chromosome rules are set a priori and are not fitted, they are free modeling choices. A refit at the calibration LET can mask their effect at that point, but the variation across LET is exactly what the paper claims to predict. Without a sensitivity analysis, the excellent RBE10 agreement in Figure 3.4 is consistent with the model, but it does not establish that the biologically grounded domain definition is correct. The proposed epsilon scan would settle this: if RBE10 is insensitive to epsilon in the 0.5-2 um range, then the concern does not land and the current claim stands; if it is sensitive, the central conclusion is conditional on an arbitrary parameter and needs to be revised or the parameter independently constrained. This is not an objection to the modeling approach itself, but to the strength of the validation claimed for it. Other issues (HUVEC fit rather than prediction, missing error bars, excluded high-energy experiments) are noted by the reader and are secondary to this robustness gap.","tokens_in":14369,"tokens_out":6087,"duration_ms":69087,"concrete_test":"Re-run the full H460 proton workflow with DBSCAN epsilon set to 0.5 um and 2.0 um, keeping the same-chromosome rule and all other settings fixed; for each epsilon, refit a, b, r to the LET=11.1 keV/um SF data using the same NMSE objective, then recompute RBE10 at every LET. If the predicted RBE10 curve shifts by more than the Monte Carlo statistical uncertainty (or changes the ordering of LET points), the central cross-LET result depends on the untested a priori clustering radius, and the conditional verdict should be strengthened or the claim revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central predictive claim is that GSM2 parameters fitted at a single LET (11.1 keV/um) transfer across 4-20 keV/um once MINAS-TIRITH DSB coordinates are clustered by DBSCAN. In Eq. (2), survival is a product over clusters, so the only LET-dependent biological input is the number and size of clusters x_j produced by the rules in Section 2.3, especially the 1 um distance threshold and the same-chromosome restriction. These rules are not fitted, but they are also not varied or independently justified; Section 2.3 says the 1 um threshold is chosen a priori to align with microdosimetric conventions. A refit at 11.1 keV/um can absorb any change in cluster statistics at that one LET, so the cross-LET RBE10 prediction is a direct test of the clustering rule, not only of GSM2 parameters. Because the simulations are reported without statistical uncertainties and no sensitivity scan over epsilon or chromosome restriction is shown, the agreement in Figure 3.4 could in principle be produced by the domain definition rather than by a validated mechanism. This is the load-bearing soft spot.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents MT-GSM2, a multiscale model that feeds DSB coordinates and chromosome identities from the MINAS-TIRITH Geant4-DNA simulation into the GSM2 survival model. DBSCAN groups DSBs on the same chromosome within 1 um into clusters, which replace GSM2's artificial domains; Eq. (2) computes survival as a product over clusters. GSM2 parameters a, b, and r are fitted to HUVEC 220 kV X-ray survival data and to H460 proton survival at 11.1 keV/um, and the model is then used to predict H460 RBE10 across approximately 4-20 keV/um. The central quantitative claim is that this cross-LET prediction matches the Patel/Bronk experiments, supporting the clustering-based domain concept.","tokens_in":14699,"tokens_out":7366,"duration_ms":80044,"significance":"The design has real strengths: the H460 prediction is a genuine cross-LET test because a, b, and r are fixed at a single LET; the chromosome-aware DBSCAN domain definition is a conceptually important alternative to fitted MKM domains; the Geant4 phase-space replication and the residual comparison with TLK are useful benchmarking elements. If the clustering rule is robust, the framework would be a valuable mechanistic bridge between nanodosimetric damage patterns and cell survival. However, the paper does not currently demonstrate that robustness: the domain-defining parameters are not varied, no uncertainties are attached to predicted survival or RBE, and the HUVEC agreement is a fit rather than a prediction. These gaps must be closed before the \"excellent agreement\" claim can be accepted.","major_comments":[{"comment":"The DBSCAN clustering rule is load-bearing for the cross-LET prediction. Because a, b, and r are fit at 11.1 keV/um, the fit can absorb the cluster statistics at that single LET, so the predicted LET dependence is carried almost entirely by the number and size of the DBSCAN clusters. The 1 um threshold and the same-chromosome restriction are not fitted, but they are also not varied; Section 2.3 states the threshold is chosen a priori, and Section 2.5 calls the domain definition \"parameter-free,\" which is inaccurate. Please provide a sensitivity analysis over epsilon (e.g., 0.5-2 um) and over the chromosome restriction, showing RBE10 and SF curves, and state how the conclusions change. Without this, the agreement in Fig. 3.4 could be an artifact of the clustering rule.","section":"Section 2.3 and Eq. (2)"},{"comment":"No statistical uncertainties are reported for the predicted survival fractions, RBE10 values, or fitted parameters a, b, and r. Figures 3.3 and 3.4 show point predictions only, and Eq. (4) is minimized without confidence intervals. Since Ncells = 2000 and the phase-space sampling are stochastic inputs, the \"excellent agreement\" in Fig. 3.4 is not quantitatively assessable. Please add error bars or confidence intervals, e.g., by bootstrapping over simulated cells or by repeating the parameter fit on resampled data.","section":"Sections 3.3 and 3.6"},{"comment":"The HUVEC X-ray comparison is not an independent validation. The endothelial a, b, and r are fitted to the same 220 kV X-ray survival data shown in Fig. 3.3a, so the small residuals in Fig. 3.3b are a goodness-of-fit result, not a predictive test. The text should label this as calibration, and the claim in Section 4.2 that MT-GSM2 \"predicts the biological effects of photons... from first principles\" should be softened or supported by a withheld-dose or independent-photon test.","section":"Section 3.2 and Table 1"},{"comment":"The mapping from the DSB complexity index i to the number of sublethal lesions x_j is not written explicitly. The sentence \"DSB-type damages with i and i+1 account for the same number of sublethal lesions\" implies x_j = sum over DSBs of ceil((i+1)/2), but the paper never gives the formula. Because survival in Eq. (2) depends exponentially on x_j, this mapping is as important as the clustering rule. Please state the mapping in equation form and test its sensitivity to a plausible alternative (e.g., x_j = sum (i+1)).","section":"Section 2.4 and Eq. (2)"}],"minor_comments":[{"comment":"The abstract and several section headers contain rendering artifacts such as \"keV=—m\" and \"\\gsm\"; please fix the typography.","section":"Abstract"},{"comment":"\"cathegorized\" should be \"categorized\"; use consistent spelling of \"sublethal\" throughout.","section":"Section 2.2"},{"comment":"In the caption of Figure 3.3, \"MINAS-TIRTIH\" should be \"MINAS-TIRITH\".","section":"Section 3.2"},{"comment":"The quantity is labeled NMSE, but the expression is a dose-weighted squared error normalized by S(D); please clarify the normalization or rename the quantity.","section":"Equation (4)"},{"comment":"The claim of a \"strong correlation between the number of DSB clusters and RBE10\" is not directly shown by Figure 3.6, which plots cluster-to-total ratio versus LET; either add a direct RBE10-versus-clustering panel or rephrase the statement.","section":"Section 4.1"},{"comment":"The correlation between high clustering and reduced survival is built into Eq. (2), since survival is computed from the clusters; presenting it as independent support for the model's mechanism is misleading.","section":"Section 3.5 and Figure 3.7"},{"comment":"Please specify which proton experiments were excluded because their phase space exceeded the MINAS-TIRITH database limits, and whether Figure 3.4 includes all remaining experimental points.","section":"Section 2.6"},{"comment":"The row for HUVEC cells is labeled \"Endothelial\"; use the same cell-line name as in the text for consistency.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The cross-LET H460 prediction is the strongest part of the paper and justifies a revision rather than rejection. My main concern is that the authors may treat the HUVEC fit as validation; please ensure the revision distinguishes calibration from prediction. I would also request data/code availability for the DBSCAN pipeline, since the sensitivity analysis will otherwise be difficult to reproduce. The paper fits the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the paper's central predictive claim holds up. The authors fit GSM2's three parameters to H460 survival at a single LET (11.1 keV/µm) and then predict RBE10 across 4–20 keV/µm using DSB clusters from MINAS-TIRITH as the only LET-dependent input. That is a genuine cross-LET test, not a refit, and they pass it. This is the real contribution.\n\nWhat's new: coupling a nanodosimetric DSB simulator (MINAS-TIRITH) to a microdosimetric survival model (GSM2) by defining the model's \"domains\" as DBSCAN clusters of DSBs on the same chromosome within 1 µm. This is a biologically grounded replacement for the artificial domain structure in GSM2/MKM. Using chromosome identity to constrain clustering is new; the DBSCAN idea itself was used by Francis et al. 2011, and the authors cite it properly.\n\nWhat's good beyond the core test: the phase-space reproduction is careful, the comparison to the TLK model shows a substantial improvement in residuals for the X-ray case, and the paper is honest that the DBSCAN threshold is chosen a priori, not fitted.\n\nNow the soft spots. The HUVEC X-ray validation is not a validation—a, b, r for that cell line are fitted to those same survival data. It only shows the model can interpolate its own fit; the residual improvement over TLK is meaningful but not evidence of mechanism. The proton cross-LET result carries the weight, and it does carry it.\n\nThe bigger issue: the clustering rules themselves—the 1 µm threshold and the same-chromosome restriction—are never varied or tested for sensitivity. A refit at 11.1 keV/µm can absorb changes in cluster statistics at that LET, so the cross-LET prediction tests the combined machinery of the cluster definitions plus the GSM2 survival formula. Agreement with experiment is good, but without a sensitivity scan the reader cannot know whether the mechanism is robust or the cluster rule is tuned to fit. The simulation results also come without error bars or confidence intervals, which is unusual for a Monte Carlo pipeline and should be fixable.\n\nMinor: the text overclaims \"first principles\" and \"parameter-free\" for the domains—the DSB clustering does remove one class of fitted parameters, but a, b, r are still fitted, and the 1 µm rule is a convention. The exclusion of experiments above the MINAS-TIRITH energy limit is fine, but it means the validation covers only a slice of clinical proton energies.\n\nOverall: the central argument holds. This is a useful integration that deserves referee time. I'd ask the authors for a sensitivity analysis over epsilon and chromosome restriction, plus error bars on the simulated SF and RBE10, before accepting.\n\nFor you: worth a look if you work on RBE modeling. I'd bring it to reading group, and I'd accept it for peer review.","headline":"The cross-LET proton RBE10 prediction is a genuine non-circular test and it passes; the paper's main soft spot is the never-tested DBSCAN clustering rule, not the core integration.","tokens_in":15245,"tokens_out":2064,"would_cite":true,"duration_ms":22508,"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":"MT-GSM2 builds survival domains directly from DBSCAN clusters of simulated DNA double-strand breaks, predicting proton RBE10 across 4–20 keV/µm from parameters fit at a single LET.","keywords":["radiation-induced DNA damage","double-strand break clustering","microdosimetry","nanodosimetry","relative biological effectiveness","proton therapy","DBSCAN","cell survival modeling"],"falsifier":"Take a proton LET in the covered range that was not used for fitting, say 16 keV/µm, and compare MT-GSM2 predictions—with parameters fixed at LET = 11.1 keV/µm—against new H460 clonogenic survival data; a deviation larger than experimental uncertainty would falsify the fixed-parameter transferability claim.","tokens_in":1660,"feed_emoji":"🧬","tokens_out":5336,"duration_ms":117316,"temperature":0.7,"pith_summary":"The paper argues that cell survival after radiation can be predicted from the spatial positions and chromosome identities of double-strand breaks (DSBs), without relying on artificial subnuclear domains. It introduces MT-GSM2, which feeds DSB maps from the MINAS-TIRITH simulator into the GSM2 survival model and uses DBSCAN clustering to let simulated damage define the model's domains. The central test is proton RBE10 for H460 cells: parameters fit only at 11.1 keV/µm reproduce experimental survival across roughly 4 to over 20 keV/µm, and X-ray survival for HUVEC cells is also reproduced. If correct, the model offers a mechanistic bridge from nanodosimetric DNA damage patterns to clinical relative biological effectiveness, relevant to biologically optimized proton therapy.","feed_headline":"Clustered DNA breaks predict proton RBE without per-LET fitting","feed_subtitle":"One model links nanometer-scale break positions to survival, matching proton experiments from 4 to 20 keV/µm.","key_machinery":"The central object is the chromosome-aware DBSCAN cluster of DSBs, which replaces GSM2's artificial domains. Under the clustering rules, weighted DSBs within 1 µm of each other and on the same chromosome form a cluster, and the number of sublethal lesions in each cluster enters the GSM2 survival product $$S_n(z_n|D_{\\text{abs}}) = \\prod_j \\prod_x \\frac{r x}{(a+r)x + b x(x-1)}.$$ This turns spatial damage coordinates into biophysical domains with no fitted domain parameters, allowing the LET dependence of survival to emerge from the simulated DSB topology rather than from per-LET calibration.","core_discovery":"The paper's discovery is that treating each DBSCAN cluster of DSBs—restricted to DSBs on the same chromosome and within 1 µm of each other—as an independent GSM2 domain makes the fitted repair and interaction rates transferable across LET and across radiation type. DSBs weighted by their complexity index are grouped into clusters; each cluster then contributes a factor to cell survival through GSM2's reaction scheme, and averaging over a cell population gives the survival curve. With parameters fit only to H460 survival at LET = 11.1 keV/µm, the model reproduces RBE10 values over the full tested proton LET range, with predicted RBE10 exceeding the clinical reference of 1.1 and approaching experimental values at high LET. The authors present this as one of the first consistent multiscale models linking nanodosimetric and microdosimetric representations of radiation to cell survival.","pith_inferences":["The model's reliance on the 1 µm and same-chromosome rules implies a testable prediction: controlled changes in chromosome territory organization should shift survival for the same LET and dose, even when total DSB yield is unchanged.","If the cluster-independence assumption is right, pairwise DSB interactions are the dominant lethal channel at high LET; one could estimate the interaction rate b directly from time-resolved co-localization of repair foci.","Extending the framework to carbon ions would provide a sharper test, because MT-GSM2 would predict an RBE–LET curve for heavier ions with no new free parameters beyond the DBSCAN rules.","The fitted second-order rate b becoming comparable to the repair rate r suggests that approaches neglecting pairwise damage interactions may need artificial domain sizes to compensate; data-derived clusters could replace those domains more generally."],"forward_implications":["At LET around 20 keV/µm, predicted RBE10 exceeds 3, well above the fixed clinical value of 1.1; the paper argues that constant-RBE proton therapy underestimates biological effects near the distal edge of the spread-out Bragg peak.","Parameters fit only at LET = 11.1 keV/µm for H460 reproduce survival across the whole tested LET range, so the cluster-based domain structure, not per-LET recalibration, carries the LET dependence.","The ratio of clusters to total DSBs and the mean cluster size both grow with LET, and per-nucleus survival tracks clustering, supporting the claim that nanodosimetric damage topology drives the RBE rise.","Because the same framework reproduces 220 kV X-ray survival for HUVEC cells, a single parameter set can span photon and proton qualities, supporting use in mixed-field treatment planning.","Expanding the simulator's energy database to clinical proton energies, helium ions, and carbon ions is presented as the direct path to broader clinical applicability."],"supporting_citations":[{"why":"Supplies the GSM2 reaction scheme and the survival expression that MT-GSM2 inherits.","marker":"[11]"},{"why":"Derives the cell-survival computation via GSM2, including the domain-level survival formula used in the paper.","marker":"[21]"},{"why":"Describes MINAS-TIRITH, the simulator that produces spatially and chromosome-resolved DSB coordinates.","marker":"[26]"},{"why":"Provides the HUVEC 220 kV X-ray clonogenic survival data used for photon validation.","marker":"[33]"},{"why":"Gives H460 proton survival and RBE10 data across the LET range used for validation.","marker":"[34]"},{"why":"Adds high-throughput RBE mapping that documents proton RBE above 1.1 at high LET.","marker":"[35]"},{"why":"Introduces DBSCAN adapted to DNA damage clustering, the basis of the chromosome-aware clustering algorithm.","marker":"[45]"},{"why":"Establishes the microdosimetric domain concept and the 1 µm scale adopted for cluster proximity.","marker":"[47]"}],"fun_headline_variants":["DSB clusters link nanoscale damage to cell survival across LET","Multi-scale model predicts survival from DSB clusters without LET fitting","Clustered DNA breaks predict cell survival across proton LET","One model bridges nanometer-scale breaks to survival for proton therapy","DSB clusters predict RBE from 4 to 20 keV/µm without refitting"],"cache_read_input_tokens":17280,"weakest_assumption_plain":"The load-bearing premise is that only DSBs within 1 µm of each other and on the same chromosome can interact lethally, and that distinct DBSCAN clusters never interact; if this clustering rule is wrong, predicted survival changes even with identical fitted rates.","fun_headline_variants_meta":{"raw":{"variants":["DSB clusters link nanoscale damage to cell survival across LET","Multi-scale model predicts survival from DSB clusters without LET fitting","Clustered DNA breaks predict cell survival across proton LET","One model bridges nanometer-scale breaks to survival for proton therapy","DSB clusters predict RBE from 4 to 20 keV/µm without refitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000888,"raw_usage":{"total_tokens":3859,"prompt_tokens":997,"completion_tokens":2862,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2769}},"tokens_in":613,"tokens_out":2862,"duration_ms":21110,"temperature":1.0,"reasoning_tokens":2769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:02:27.853237+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a proton LET in the covered range that was not used for fitting, say 16 keV/µm, and compare MT-GSM2 predictions—with parameters fixed at LET = 11.1 keV/µm—against new H460 clonogenic survival data; a deviation larger than experimental uncertainty would falsify the fixed-parameter transferability claim.","supporting_citations":[{"cited_title":"Cordoni, M","cited_arxiv_id":null,"evidence_quote":"Supplies the GSM2 reaction scheme and the survival expression that MT-GSM2 inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the cell-survival computation via GSM2, including the domain-level survival formula used in the paper."},{"cited_title":"Thibaut, G","cited_arxiv_id":null,"evidence_quote":"Describes MINAS-TIRITH, the simulator that produces spatially and chromosome-resolved DSB coordinates."},{"cited_title":"Paget, M","cited_arxiv_id":null,"evidence_quote":"Provides the HUVEC 220 kV X-ray clonogenic survival data used for photon validation."},{"cited_title":"Patel, L","cited_arxiv_id":null,"evidence_quote":"Gives H460 proton survival and RBE10 data across the LET range used for validation."},{"cited_title":"Bronk, F","cited_arxiv_id":null,"evidence_quote":"Adds high-throughput RBE mapping that documents proton RBE above 1.1 at high LET."},{"cited_title":"Francis, C","cited_arxiv_id":null,"evidence_quote":"Introduces DBSCAN adapted to DNA damage clustering, the basis of the chromosome-aware clustering algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the microdosimetric domain concept and the 1 µm scale adopted for cluster proximity."}],"review_version":1}