{"id":"66d3fe89-d797-4a35-9ad1-f13c3b51abff","arxiv_id":"2412.12960","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A stochastic model plus likelihood for TCID50 data shows measured infectious dose equals actual infectious virions times establishment probability, improving parameter estimation.","lead":"This modeling paper shows that the randomness in endpoint dilution (TCID50) assays can be folded into a new likelihood function for estimating virus infection parameters, and that measuring virus in infectious virions rather than infection-causing doses makes parameters easier to interpret. It also suggests that most of the variability seen between replicate in vitro infections comes from the measurement assay, not from the infection itself.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central IV-to-SIN conversion and the resulting identifiability claims rest on the acknowledged but false assumption that the A549 infection experiments and the MDCK ED assays share all infection parameters; a cell-type-specific establishment probability would bias the reported IV-unit estimates.","rationale":"The reader's weakest-assumption analysis identifies the A549/MDCK shared-parameter assumption as the main vulnerability, and I agree. This is the single most load-bearing concern because it sits directly between the model's IV variable and the experimentally measured SIN data: Eqn. (2) defines a cell-condition-specific conversion, but the paper applies a conversion computed under the wrong cell-type parameters. The authors are transparent about the assumption and call it 'most critical' in the Discussion, but transparency does not remove the bias; it makes the empirical parameter estimates conditional on an untested equivalence. The paper's methodological contributions—the LED likelihood, the use of IV units, and the physical constraints—remain interesting and internally coherent, but the demonstration on real data is not robust to a plausible violation of this assumption. A simple multiplicative-factor test would quantify the sensitivity without requiring new experimental data, and a paired ED assay in A549 cells would settle whether the factor is actually 1. Since the paper is explicitly conditional on this assumption and the reader's verdict is already CONDITIONAL, I do not recommend changing the verdict.","tokens_in":40231,"tokens_out":4393,"duration_ms":48181,"concrete_test":"Re-estimate the model with a free multiplicative factor h = P_V→Establishment(MDCK)/P_V→Establishment(A549) multiplying the IV-to-SIN conversion in the ED likelihood, with a log-uniform prior over [0.1, 10] (or, better, fix h to a value measured by performing paired ED titrations of the same A549-derived samples on both MDCK and A549 cells). Then compare the MAP and 95% credible intervals of γ, β, ρ, B, R0, and the noise-attribution conclusion. If the posterior for h deviates substantially from 1 and the parameter intervals shift by more than their reported widths, the shared-parameter assumption is load-bearing and the reported IV-unit estimates require revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claims—Eqn. (2) and the IV-unit parameter estimates—require that the establishment probability used to convert model-predicted infectious virions into measured SIN is the establishment probability in the ED assay. In Section II.D and again in the Discussion (third limitation), the authors set C_model(t) = [V(t)/s] * P_V→Establishment(π) using parameters π estimated from the A549 SC/MC infections, with only the cell number and supernatant volume changed to the MDCK ED-assay values. The paper itself states: 'we had to assume that the ED assay and infection experiments were conducted under the same conditions, and as such shared all infection parameters ... This was not the case.' If P_V→Establishment in MDCK cells differs from that in A549 cells—which is likely, since the assays use different cell types and the authors themselves give an example where 1 SIN in MDCK could correspond to 2 IV while 1 SIN in A549 corresponds to 10 IV—then the conversion factor in the likelihood is wrong by a multiplicative constant. This directly biases every parameter estimate that is anchored by the infectious-titre data: γ, β, ρ, and the derived B, R0, tinf, and P_V→Establishment. The identifiability advantage of IV units (Sections II.D and III) and the conclusion that ED-assay noise explains inter-replicate variability (Section II.E, Figure 13) both inherit this bias because they rely on the same MAP parameter set. The concern is not about internal mathematical consistency; Eqn. (2) and the LED likelihood derivation are sound. It is about the validity of the empirical application: without a correct cell-type-specific establishment probability, the reported numerical estimates and the identifiability claims are not empirically supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a stochastic framework for interpreting TCID50 endpoint dilution (ED) assays in terms of infectious virions (IV) for in vitro virus infection models. The central result is Eqn. (2): the experimentally measured infection-causing dose concentration (SIN/ml) equals the actual infectious virion concentration (IV/ml) multiplied by the establishment probability P_V→Establishment of an infection initiated with a single infectious virion. The authors introduce a new likelihood (LED) based on the raw ED assay outcome (number of infected wells per dilution), compare it to a Gaussian residual likelihood (LNR), express the model virus variable in IV rather than SIN units with physical constraints, and compare ODE versus stochastic model predictions. Using published A549 influenza infection data with MDCK-based ED assays, they estimate parameters under four model variants and argue that ED assay stochasticity explains the inter-replicate variability observed in experimental infections. The paper includes substantial methodological derivations, including the extinction probability, the ED likelihood, and the infecting-time distribution, all validated against stochastic simulations.","tokens_in":40617,"tokens_out":5018,"duration_ms":51581,"significance":"If the results hold, the paper offers an important conceptual advance: it gives a principled, parameter-dependent conversion between a commonly measured infectivity unit (SIN/TCID50) and the biologically meaningful number of infectious virions, and it introduces a likelihood that correctly handles below-detection ED outcomes. The methodological strengths include careful stochastic derivations checked against simulation (e.g., Eqn. (1), the infecting-time distribution in Sec. IV.H), a deliberate comparison of likelihood choices, and physically motivated constraints in IV units that can improve identifiability of parameters such as γ and ρ. The paper is also refreshingly explicit about its limitations. However, the empirical demonstration of the framework rests on an acknowledged incorrect assumption that A549 infection experiments and MDCK ED assays share all infection parameters, which biases the quantitative parameter estimates and the conclusions built on them. The methodological components remain valuable and potentially correct, but the paper's current claims about IV-unit parameter values and the source of inter-replicate variability require revision.","major_comments":[{"comment":"The conversion C_model(t) = [V(t)/s] · P_V→Establishment(π) for the ED assay (Eqn. (2) and its use in the LED likelihood) relies on parameters π estimated from the A549 SC/MC infections being applied to the MDCK ED assay. The manuscript itself states, in the Discussion, that this shared-parameter assumption 'was not the case.' Because P_V→Establishment depends on cell-type-specific parameters (notably γ and β), any difference in establishment probability between A549 and MDCK cells multiplies the IV-to-SIN conversion by a constant. This directly biases the IV-unit estimates of γ, β, ρ, and all derived quantities (B, R0, tinf, P_V→Establishment) reported in Tables I and II, and consequently the identifiability discussion in §II.D and the noise-attribution conclusion in §II.E. To make the empirical claims load-bearing, the authors should either re-analyze the data with ED assays performed in the same cell type as the infections, or explicitly propagate the uncertainty in the cell-type-specific establishment probability and demonstrate that the qualitative conclusions are robust to plausible differences between A549 and MDCK.","section":"§II.D and Discussion (third limitation)"},{"comment":"The claimed individual identifiability of γ and ρ when [V] = IV relies on the physical constraint that IV cannot exceed vRNA, which in practice is imposed by a single data pair: the post-rinse SC infectious titre (10^7 SIN/ml, measured by MDCK ED) and the post-rinse total vRNA (10^7.76 vRNA/ml), giving P_V→Establishment ≥ 0.174. This constraint mixes MDCK-based SIN measurements with A549-based vRNA measurements, so it is exactly the type of quantity affected by the A549/MDCK parameter mismatch. If the true MDCK establishment probability differs from the A549-derived value, the inferred IV scale and hence the posterior bounds on γ and ρ change, weakening the conclusion that IV units resolve the γ–ρ degeneracy. The authors should show how the γ and ρ posteriors change under a range of plausible P_V→Establishment values in the ED assay, or obtain a same-cell-type measurement to anchor this constraint.","section":"§II.D, Figure 8O and Table IV"},{"comment":"The conclusion that ED assay stochasticity alone explains the experimentally observed inter-replicate variability is based on simulating ED assays from SM-predicted IV time courses using the MAP parameter set (Table I, SM,IV,LED). Since that parameter set is obtained under the acknowledged A549/MDCK sharing assumption, the simulated ED noise band in Figure 13B is not an unbiased representation of the actual ED assay noise for the experimental samples. If P_V→Establishment in the MDCK ED assay differs from the A549-based value, the simulated SIN values and their dispersion shift, and the visual match with the experimental triplicates in Figure 13C is no longer meaningful. The authors should either verify this conclusion with same-cell-type data or temper the claim to state that the observed variability is consistent with ED noise under the (unverified) assumption of shared parameters.","section":"§II.E, Figure 13"}],"minor_comments":[{"comment":"The notation C_model(t|π) is used both for the model prediction in SIN/ml (when [V]=SIN) and for the model prediction converted to SIN/ml via the establishment probability (when [V]=IV). The text explains the distinction, but the equations would be clearer if the two quantities were written as C_model^SIN(t) and C_model^IV(t)·P_V→Establishment.","section":"§II.C, Eqns. (4)–(5)"},{"comment":"The p-values are described as one-tailed fractions of pairwise comparisons, which is not a standard hypothesis test. This should be stated in the caption so that readers do not interpret them as conventional two-sided p-values.","section":"Table II"},{"comment":"The derivation of the ED likelihood uses the small-x approximation ln(1−x) ≈ −x. The approximation is well justified because p = C_actual·V_vir with V_vir ≈ 5.2×10^−16 ml, but the paper should state the numerical magnitude explicitly so that the approximation's validity is transparent.","section":"§IV.F, Eqn. (17)"},{"comment":"The grey band in panel (B) represents the 95% range of 10,000 simulated ED outcomes at each time point for the SM-predicted IV concentration, while the experimental points are three independent infection replicates. The text should note that this comparison does not include infection-to-infection variability beyond the mean IV trajectory, only ED sampling noise.","section":"§II.E, Figure 13"}],"recommendation":"major_revision","confidential_remarks":"The paper is a methodologically rich contribution with careful stochastic derivations and an unusually candid self-assessment of its central limitation. The main blocker is the acknowledged A549/MDCK parameter-sharing assumption, which undermines the quantitative IV-unit estimates and the derived conclusions about identifiability and noise attribution. This is fixable within the paper's scope: a sensitivity analysis over plausible establishment-probability differences, or re-analysis with same-cell-type data, would address the concern. I would not reject the manuscript, but the current version overstates the empirical support for the framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for Eqn. (2) and the LED likelihood. The central identity—measured SIN/ml equals IV/ml times the establishment probability of a single IV—is correct, and the likelihood for raw well counts is a genuine extension of your own midSIN work. The paper is also honest about its biggest problem, right in the Discussion: the A549 infections and MDCK ED assays do not share infection parameters, so the conversion factor used in the likelihood is likely wrong by a cell-type-specific constant. The paper says this is incorrect and quantifies an example (1 SIN in MDCK = 2 IV, in A549 = 10 IV), but then still uses the A549-derived establishment probability to convert MDCK ED outcomes and to impose the vRNA bound on IV. Every IV-unit parameter estimate—γ, ρ, β, P_Est—inherits that bias, and so does the claim that ED assay noise explains inter-replicate variability. That is a genuinely soft spot, not a manufactured one.\n\nWhat's genuinely new and valuable: the systematic comparison of LNR vs LED and SIN vs IV units is overdue, and the physical constraints (γ ≤ 1 cell/IV, IV ≤ vRNA) are a clever way to partially identify parameters that otherwise are degenerate. The stochastic vs ODE comparison is less interesting in itself—the RNS trick works, but the result that stochasticity doesn't matter under high MOI is expected—though the derivation of the SM infecting time distribution is a nice piece of work.\n\nThe p-value table is a problem too. The paper claims LED is better than LNR, but none of the parameter differences reach significance, and the only significant finding (tinf) is about an ODE-vs-SM artifact, not the likelihood choice. For a paper recommending 'should be widely adopted,' the lack of a clear statistical win weakens the sales pitch.\n\nThere's no code or data shipped, which makes the 500,000-step MCMC and LED implementation harder to verify. That's fixable.\n\nOverall: this is a paper for modelers who fit viral time courses from TCID50 data. The framework is worth having, but the empirical application is not trustworthy as-is. I'd send it to a serious referee with a request for a same-cell dataset or at least a sensitivity analysis, plus code/data sharing. The theoretical core is sound, so the right outcome is major revision, not rejection.","headline":"Solid methodological core, undermined by an acknowledged but unresolved cell-type mismatch in the empirical application; still worth a peer review with major revisions.","tokens_in":41170,"tokens_out":3604,"would_cite":true,"duration_ms":34423,"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":"TCID50 numbers are not virion counts; a new formula converts between them.","keywords":["TCID50","endpoint dilution assay","infectious virion concentration","establishment probability","stochastic infection model","likelihood-based parameter estimation","influenza A virus","assay noise"],"falsifier":"Measure the establishment probability directly: inoculate many replicate wells with a dilution expected to contain about one infectious virion, count the fraction that become infected, and compare with Eqn (1) evaluated at the estimated parameters. If the measured fraction differs from the predicted $P_{V\\to\\text{Establishment}}$ by more than the assay's sampling error, the conversion underlying every IV-unit estimate is wrong; a simpler version is to repeat the authors' parameter estimation with the ED assay performed in the same A549 cells used for the infections and check whether $\\gamma$, $\\rho$, and $P_{V\\to\\text{Establishment}}$ shift.","tokens_in":40055,"feed_emoji":"🦠","tokens_out":4966,"duration_ms":47174,"temperature":0.7,"pith_summary":"The paper establishes that the concentration of infection-causing doses measured by a TCID50 endpoint dilution assay (SIN/ml) is not the concentration of infectious virions (IV/ml), but that concentration multiplied by the probability that an infection seeded by one infectious virion establishes rather than goes extinct. Because that establishment probability depends on the very parameters being estimated, the assay's undercounting is not a fixed calibration factor. The paper introduces a likelihood for the raw endpoint-dilution outcome that handles below-detection measurements correctly, and shows that expressing viral titre in IV units, with physical constraints, makes parameters such as gamma and rho individually identifiable. Re-analysis of influenza A infection data suggests that the stochasticity of cell-virus interactions has little effect on the infection time course itself, while the random noise of the TCID50 assay explains the scatter seen between replicate infections. This matters because it provides a route from routine infectivity readings to biologically meaningful virion-level parameters.","feed_headline":"A TCID50 reading hides how many infectious virions are in the sample","feed_subtitle":"New likelihood connects assay reads to virion counts, explains replicate scatter, and makes key flu parameters identifiable.","key_machinery":"The load-bearing object is the extinction probability $P_{V\\to\\text{Extinction}}$ in Eqn (1), derived from a stochastic model with Erlang-distributed eclipse and infectious phases, negative-binomial per-cell virion production, and loss of infectivity and cell entry. Its complement, $P_{V\\to\\text{Establishment}} = 1 - P_{V\\to\\text{Extinction}}$, converts infectious virions to measured specific infections as in Eqn (2). The new likelihood $L_{ED}$ (Eqn 5) is a product over dilution columns of binomial infection probabilities, with no Gaussian error assumption, and it naturally includes measurements below the detection limit. Parameter estimation uses MCMC, and for the stochastic model the random number seed is added as a parameter so each parameter set maps to a single deterministic trajectory.","core_discovery":"The paper's central claim is expressed as Eqn (2): the measured infection-causing dose concentration (SIN/ml) equals the actual infectious virion concentration (IV/ml) multiplied by $P_{V\\to\\text{Establishment}} = 1 - P_{V\\to\\text{Extinction}}$, the probability that an infection initiated with a single infectious virion establishes. That probability is parameter-dependent and given by Eqn (1), so the conversion factor between SIN and IV is not a constant but part of what the model must estimate. Building on this, the paper proposes a likelihood, $L_{ED}$ (Eqn 5), that uses the full endpoint-dilution outcome rather than a Gaussian residual assumption, and it re-expresses the model's infectious titre in IV units rather than SIN units. In IV units, physical constraints (one IV cannot infect more than one cell; IV cannot exceed total vRNA) break the degeneracies that otherwise plague parameter estimation. Applied to experimental influenza A infections, the framework yields statistically similar phase durations and production rates to standard approaches, but with the added ability to estimate quantities such as the probability of infection establishment and the number of IV entry events per successful cell infection.","pith_inferences":["The authors stop short of saying so, but the same logic implies that comparing viral titres across cell types, or between an assay and an infection done in different cells, carries an unknown conversion factor: equal TCID50 readings can correspond to very different numbers of infectious virions in the two settings.","A testable extension is to run the identical infection experiment with the ED assay performed in the same cell line as the infection; if the shared-parameter assumption is correct, parameter posteriors should match the authors' results, and if not, the discrepancy would quantify cell-type effects on establishment.","A concrete assay-design improvement follows directly from the paper: increasing the number of replicate wells per dilution and using finer dilution steps should narrow the measurement noise identified as the dominant source of inter-replicate scatter, and this can be verified in silico before spending reagents.","The framework should port to plaque and focus-forming assays, which share the 'one dose causes one infection' assumption; expressing those titres in IV units would require the same establishment-probability conversion, possibly with assay-specific parameters."],"forward_implications":["Measurements below the assay's detection limit (no infected wells at any dilution) are handled by $L_{ED}$ directly, instead of being set to the limit of detection as in Gaussian-residual likelihoods.","Expressing virus in IV units with the constraints $\\gamma \\le 1$ cell/IV and IV $\\le$ vRNA removes the $(\\gamma, \\rho)$ degeneracy, giving finite, meaningful posterior intervals for each parameter.","The stochastic model and the ODE give statistically indistinguishable parameter estimates for the in vitro infections considered, so the 15-fold extra computation buys little for such well-inoculated infections.","Simulated ED assay noise alone reproduces the scatter among experimental triplicates, implying that replicate variability in such experiments comes largely from the measurement assay and could be reduced by more wells per dilution or finer dilution steps.","The estimated parameters imply that 1 SIN corresponds to roughly 1.1 to 5.8 infectious virions under the conditions studied, and that about 1 to 5 infectious virion entry events are needed for one successful cell infection."],"supporting_citations":[{"why":"Supplies the experimental influenza A virus single-cycle and multi-cycle infection data and the base ODE parameters that the paper re-estimates and extends.","marker":"[34]"},{"why":"Provides the midSIN estimator used to convert ED assay outcomes to SIN/ml and the endpoint-dilution likelihood from which $L_{ED}$ is derived.","marker":"[25]"},{"why":"Provides the stochastic model and the extinction probability $P_{V\\to\\text{Extinction}}$ in Eqn (1) that underlies the SIN-to-IV conversion.","marker":"[33]"},{"why":"Earlier estimate of $\\gamma$ in cell/TCID50 units that the paper contrasts with its IV-unit estimate.","marker":"[37]"},{"why":"Stochastic theory of early viral infection used for comparison of burst production and extinction probabilities.","marker":"[27]"},{"why":"Monte Carlo simulation of TCID50 and the conversion factor 1 TCID50 $\\approx$ 0.56 SIN used in setting base parameters.","marker":"[23]"}],"fun_headline_variants":["TCID50 counts infections, not virions - a new model fixes that","New likelihood links TCID50 reads to true virion counts","Random cell-virus hits skew TCID50—new estimation helps","IV units break parameter deadlocks in virus assays","Why TCID50 replicates vary—and how to model it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire conversion and all IV-unit parameter estimates rest on treating the ED assay and the infection experiments as sharing the same infection parameters, even though the infections used A549 cells and the ED assay used MDCK cells; the paper states explicitly that this assumption was not true.","fun_headline_variants_meta":{"raw":{"variants":["TCID50 counts infections, not virions - a new model fixes that","New likelihood links TCID50 reads to true virion counts","Random cell-virus hits skew TCID50—new estimation helps","IV units break parameter deadlocks in virus assays","Why TCID50 replicates vary—and how to model it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000725,"raw_usage":{"total_tokens":3348,"prompt_tokens":1141,"completion_tokens":2207,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":757,"completion_tokens_details":{"reasoning_tokens":2119}},"tokens_in":757,"tokens_out":2207,"duration_ms":16046,"temperature":1.0,"reasoning_tokens":2119,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:33:00.735957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the establishment probability directly: inoculate many replicate wells with a dilution expected to contain about one infectious virion, count the fraction that become infected, and compare with Eqn (1) evaluated at the estimated parameters. If the measured fraction differs from the predicted $P_{V\\to\\text{Establishment}}$ by more than the assay's sampling error, the conversion underlying every IV-unit estimate is wrong; a simpler version is to repeat the authors' parameter estimation with the ED assay performed in the same A549 cells used for the infections and check whether $\\gamma$, $\\rho$, and $P_{V\\to\\text{Establishment}}$ shift.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the experimental influenza A virus single-cycle and multi-cycle infection data and the base ODE parameters that the paper re-estimates and extends."},{"cited_title":"Cresta, D","cited_arxiv_id":null,"evidence_quote":"Provides the midSIN estimator used to convert ED assay outcomes to SIN/ml and the endpoint-dilution likelihood from which $L_{ED}$ is derived."},{"cited_title":"Stochastic failure of cell infection post viral entry: Implications for infection outcomes and antiviral therapy","cited_arxiv_id":"2208.00637","evidence_quote":"Provides the stochastic model and the extinction probability $P_{V\\to\\text{Extinction}}$ in Eqn (1) that underlies the SIN-to-IV conversion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier estimate of $\\gamma$ in cell/TCID50 units that the paper contrasts with its IV-unit estimate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Stochastic theory of early viral infection used for comparison of burst production and extinction probabilities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Monte Carlo simulation of TCID50 and the conversion factor 1 TCID50 $\\approx$ 0.56 SIN used in setting base parameters."}],"review_version":1}