{"id":"e992a13a-1bc2-4638-9ea6-15b10d7b9c6d","arxiv_id":"2507.10793","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A Markov chain framework couples personal gamma-distributed antibody response curves with event-to-event transition probabilities to track population antibody distributions after multiple infections and vaccinations.","lead":"This paper builds a mathematical model that tracks how antibody levels change in a population as people experience multiple infections and vaccinations. It offers public health analysts a way to interpret antibody survey data by combining personal antibody curves with the timing and sequence of immune events.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5/7 freeze each earlier event's gamma-shape contribution at the gap τ_j rather than evolving it with total time t, so the multi-event response asymptotes above αN; this contradicts the paper's own convolution rationale and propagates into all Section 4/5 conditional probabilities.","rationale":"The reader identified the additive independent-gamma assumption as the weakest point, focusing on biological dependence of the second response on current antibody level and on synthetic naive parameters. My stress-test found a more specific and more internal problem: even granting independence, the way Eq. (5) and Eq. (7) combine time-dependent components is inconsistent with the paper's own convolution argument. The first event's contribution is evaluated at the time of the second event, τ, and then held constant; it does not continue to evolve with total elapsed time t. Consequently the multi-event response does not return to the naive baseline, which undermines the claim of rigorously tracking post-event waning. This is partially an agreement with the reader because it is a flaw in the same additive gamma model, but it is distinct: it is an internal mathematical inconsistency rather than an empirical or biological counterfactual. The fix is straightforward in principle (replace τ_j with total time since that event for the additive terms), so the framework is salvageable, but the numerical fits and all conditional probabilities built on Eq. (5)-(8) would need to be redone. That supports the reader's CONDITIONAL verdict rather than changing it: the paper should not be accepted in its current form, but the central idea is not irreparably broken. The proposed concrete test—analytic re-derivation plus refitting with the corrected formula—would settle whether the frozen term materially changes the model's predictions and likelihood.","tokens_in":27577,"tokens_out":11934,"duration_ms":151010,"concrete_test":"Analytically, recompute the two-event shape by convolving Gamma(θ1 t/(1+φ1 t^k1), βN) with Gamma(θ2(t−τ)/(1+φ2(t−τ)^k2), βN); the result gives first-event term θ1 t/(1+φ1 t^k1), not θ1 τ/(1+φ1 τ^k1), so Eq. (5) is not the convolution the text invokes. Numerically, refit the Dataset 1 VI and VV models using the corrected shape θ1 t/(1+φ1 t^k1) + θ2(t−τ)/(1+φ2(t−τ)^k2) + αN with the same MLE procedure, and compare log-likelihoods. Also evaluate α2 at large t under Eq. (5) for the fitted parameters: if the limit exceeds αN by more than the naive distribution's spread, the model predicts permanently elevated antibody levels, contradicting the waning assumption that motivates the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The multi-event personal response model in Eq. (5) sets the first-event contribution to θ_z1 τ/(1 + φ_z1 τ^{k_z1}), where τ is the gap between events, rather than evaluating the single-event response at the total elapsed time t since the first event. Eq. (7) repeats this for every event j < M, freezing each earlier contribution at its gap τ_j. This is not merely a biological modeling choice: under the paper's own infinite-divisibility rationale (Section 3), if the first-event response at time t is independent of the second-event response at time t−τ, the convolution of their gamma components has shape [α1,z1(t) − αN] + [α1,z2(t−τ) − αN] + αN = θ_z1 t/(1 + φ_z1 t^{k_z1}) + θ_z2 (t−τ)/(1 + φ_z2 (t−τ)^{k_z2}) + αN. Eq. (5) instead holds the first term constant for all t > τ. As a result, the combined response never wanes back to the naive baseline; Eq. (37) explicitly gives lim_{t→∞} α_n(t) = Σ_{i=1}^{n−1} θ_i τ_i/(1 + φ_i τ_i^{k_i}) + αN > αN. The paper dismisses this as a small discrepancy in Section 8.5, but no quantitative support is given, and the error can be large when the second event occurs well before the first response has waned. Because every conditional probability in Sections 4.2 and 5.2 is built from R_{I,I}, R_{V,V}, and the general R_{z_m}, this frozen-response assumption is a load-bearing component of the central claim that the framework rigorously tracks population antibody distributions across multiple immune events.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-level modeling framework for population antibody kinetics. At the personal level, post-event antibody distributions are modeled as gamma densities whose shape parameter follows a peak-and-decline curve in time, extended to multiple immune events by adding shape contributions. At the population level, discrete-time Markov chains are used on a 13-state graph (two-event model, Section 4) and on a five-state graph (general time-inhomogeneous and time-homogeneous models, Section 5), yielding probabilities of immune-state trajectories that are convolved with personal response models to give conditional densities of antibody measurements on an absolute timeline. The framework is applied to two SARS-CoV-2 longitudinal datasets (Section 6) and to simulated population transmission (Section 7). The paper claims to be the first to simultaneously address antibody levels, prevalence, multiple immune-event classes, time dependence, and multiple immune events.","tokens_in":27976,"tokens_out":11018,"duration_ms":108829,"significance":"If the technical issues with the multi-event personal response model are resolved, the proposed framework would be a useful contribution, extending the authors' prior five-state Markov-chain antibody kinetics work to multiple immune events and providing explicitly enumerable conditional probabilities for the two-event case. The derivations in Sections 4.2 and 5.2 are careful applications of conditional probability and the law of total probability, and the paper is transparent about several of its limitations, including the non-vanishing asymptote of the multi-event shape (Eq. (37)) and the synthetic naive baseline used for Dataset 2. The graphical state models and closed-form conditional probabilities are likely to be usable by other researchers despite the empirical validation weaknesses.","major_comments":[{"comment":"The multi-event shape model evaluates each earlier event's gamma-shape contribution at the gap τ_j and freezes it for all later times t, rather than evaluating the single-event response at the total time elapsed since that event. This is inconsistent with the paper's own infinite-divisibility rationale in the paragraph following Eq. (3), which would give a combined shape of Σ_j [α1,zj(t_j) − αN] + αN with t_j = t − Σ_{i<j} τ_i. As written, Eq. (37) gives lim_{t→∞} α_n(t) = Σ_{i=1}^{n−1} θ_i τ_i/(1+φ_i τ_i^{k_i}) + αN > αN, so an individual with two events is predicted never to return to the naive baseline. The dismissal of this as a small discrepancy in Section 8.5 is unsupported; using the Dataset 1 infection parameters in Table 1, the frozen first-event contribution at τ = 100 days is about 11.9 excess shape units, whereas the single-event model would give about 3.0 at t = 400 days. Because every conditional probability in Sections 4.2 and 5.2 is built from these multi-event responses, this is a load-bearing assumption that needs either a corrected model or an explicit biological justification with quantitative sensitivity analysis.","section":"Section 3, Eqs. (5), (7); Section 8.5, Eq. (37)"},{"comment":"The empirical support is weaker than claimed because the same data are used for fitting and evaluation. Parameters in Tables 1 and 3 are obtained by MLE from the same trajectories whose measurements are later scored for contour coverage and 'high probability' thresholds in Table 2. These statistics are in-sample goodness-of-fit measures, not predictive validation, and the high percentages (e.g., 96% for VV in Dataset 1) are expected to be optimistic. Please either reframe Section 6 and the abstract's 'validate' language as goodness of fit, or provide an out-of-sample evaluation such as cross-validation or a held-out cohort.","section":"Section 6, Tables 1-3, Table 2"},{"comment":"The Dataset 2 analysis uses synthetic naive parameters α_n = 18.2 and β_n = 0.152, described as a 'reasonable guess' because no naive samples were collected. The single-event fits in Table 3 and the two-event VI fit, and therefore the Dataset 2 rows of Table 2, are all conditional on this unvalidated baseline. This should be presented as a substantial limitation rather than a minor detail, and a sensitivity analysis of the fitted parameters and coverage statistics to the naive parameters is needed to support the claimed validation.","section":"Section 6.2"}],"minor_comments":[{"comment":"The state-count formula N_s = 4(2N_e − 1) + 1 gives 37 for N_e = 5, contradicting the sentence 'with 125 states necessary to model 5 possible immune events'; the intended formula is evidently 4(2^{N_e} − 1) + 1, which matches both the 13-state two-event case and the 125-state five-event example.","section":"Section 8.1, Eq. (36)"},{"comment":"The display 'Prob(X_T = I) = Prob(r, T|X_{T−1}=N, X_T=I) = R(r,0) = N(r)' equates a state probability with a conditional density; the left-hand side should be a density or the notation should distinguish the two.","section":"Section 4.2, Eq. (14)"},{"comment":"The table and surrounding text refer to 'probability > 0.001' and 'probability above 0.2'; these thresholds are density contour levels, not probabilities of an interval, and the text should say so to avoid confusion.","section":"Table 2"},{"comment":"The product 'S(4)S(3)S(2)S(2)S(1)S(0)' contains a repeated S(2); it should be S(4)S(3)S(2)S(1)S(0).","section":"Appendix A, Eq. (A1)"},{"comment":"The text states that time is discrete throughout, but the gamma response models in Eqs. (3)-(8) use a continuous time variable; the relationship between the discrete Markov chain time step and the continuous personal timeline should be stated explicitly.","section":"Section 2.1"},{"comment":"The claim to be the first to simultaneously address topics (i)-(v) is difficult to verify; I suggest softening it or citing a broader comparison set.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically workable and the main mathematical machinery is sound, but the multi-event personal response model has an internal inconsistency that propagates into the central conditional probabilities; the validation is in-sample; and Dataset 2's synthetic baseline needs sensitivity analysis. I believe these can be addressed in revision. The code is promised only after acceptance, which limits reproducibility; I would encourage the authors to release it with the revision. I did not find any indication of misconduct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine extension of the authors' earlier single-event Markov chain framework to multiple immune events, and the conditional probability machinery is built carefully. The explicit 13-state two-event transition matrix, the additive gamma-shape personal response model, and the law-of-total-probability derivations in Sections 4 and 5 are all new relative to the cited literature. The application to two real SARS-CoV-2 cohorts is a real plus – the authors fit the model to actual longitudinal measurements and show contour plots and trajectory analyses. They also cite their own prior work appropriately and are unusually candid in the limitations section.\n\nThe main soft spots are three. First, the validation is entirely in-sample: the same data are used to fit the MLE parameters and then scored for contour coverage. That is a goodness-of-fit measure, not a prediction check, and the paper's use of the word 'validate' oversells it. Second, Dataset 2 leans on synthetic naive parameters (α_n = 18.2, β_n = 0.152) and a titration extrapolation applied to nearly 70% of the samples, so the empirical support from that dataset is weaker than the headline numbers suggest. These are disclosed, but they should temper the claims.\n\nThird, and most substantively, the stress-test note is correct: Eq. (5) and (7) freeze each earlier event's gamma-shape contribution at the inter-event gap τ_j rather than letting it continue to evolve with total time t. This contradicts the paper's own convolution rationale and means the multi-event response never returns to the naive baseline. The authors acknowledge this in Eq. (37) and say they expect the discrepancy to be small, but they give no quantitative support. The error can be substantial when a second event occurs before the first response has waned, and because every conditional probability in Sections 4.2 and 5.2 is built from these response densities, the bias propagates into the central claims. This is a fixable flaw – the natural correction would be to evaluate the first-event term at t, not τ – but as it stands, the multi-event predictions at longer times are on shaky ground.\n\nThe additive-independence assumption is also strong and untested, though the authors frame it as a modeling choice.\n\nOverall: worth developing, worth a serious referee, but not ready as is. A revision that fixes or quantifies the frozen-response issue and adds even a modest out-of-sample check would make this a solid contribution. I would send it to peer review, but I would make the multi-event response model the first thing the referee examines.","headline":"A sincere multi-event extension of the authors' own Markov-chain antibody kinetics framework, with coherent math but in-sample validation and a real, acknowledged modeling inconsistency that needs fixing before the predictions can be taken at face value.","tokens_in":28602,"tokens_out":2521,"would_cite":false,"duration_ms":34175,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J20","62P10","92D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims to be the first to compute the probability distribution of population antibody levels under arbitrary sequences of infection and vaccination events, by coupling a time-inhomogeneous Markov chain of immune states with…","keywords":["antibody kinetics","time-inhomogeneous Markov chain","gamma distribution","immune event sequences","SARS-CoV-2 serology","vaccination","prevalence estimation","personal trajectories"],"falsifier":"Take longitudinal data in which a person's antibody level is measured immediately before a second infection or booster, then repeatedly afterwards, and test whether the size or variability of the post-event rise depends on the pre-event level. If high pre-event titers systematically produce smaller rises than the additive-shape model predicts, or if the spread visibly changes after repeated events, the central factorization fails; a likelihood-ratio comparison against a model with pre-event-titer-dependent shape and scale would settle it.","tokens_in":27333,"feed_emoji":"💉","tokens_out":8628,"duration_ms":94093,"temperature":0.7,"pith_summary":"The paper tries to establish that population antibody measurements can be interpreted as a weighted mixture of personal immune histories: a time-inhomogeneous Markov chain supplies the probability of each sequence of infections and vaccinations, and a gamma-distributed personal response model supplies the antibody distribution after that sequence. This is the first mathematical treatment the authors know of that simultaneously covers antibody levels, prevalence, multiple event classes, time dependence, and multiple immune events per person. If correct, a random blood draw can be assigned a probability under each personal trajectory, which supports serosurveillance, missed-event detection, and booster-timing decisions. The framework is demonstrated on two longitudinal SARS-CoV-2 cohorts with documented infections and vaccinations, plus population-transmission simulations.","feed_headline":"Markov model tracks antibody levels through infections and shots","feed_subtitle":"Couples transmission odds with personal antibody curves to time boosters and spot missed infections.","key_machinery":"The machinery is a discrete-time, time-inhomogeneous Markov chain over immune states (naive, newly infected, newly vaccinated, previously infected, previously vaccinated) coupled to an additive gamma personal-response model. The transition matrix $S(T)$ gives the probability of moving between states at time step $T$, and products $H_T=S(T)S(T-1)\\cdots S(0)$ enumerate every possible sequence of immune events. On the personal side, Eq. (7) defines the gamma shape after a sequence of events as a sum of peak-and-decay terms (one per event) plus the naive shape, with the scale fixed, so the response to multiple events is the convolution of independent event responses. The paper uses this pair of objects to write conditional probabilities for antibody measurements, then specializes to a time-homogeneous chain whose stationary distribution has a closed form.","core_discovery":"The central claim is that the population-level distribution of antibody measurements at a given calendar time $T$ is a convolution of transition probabilities and personal response densities: for any immune state, $P(r,T\\mid X_T=I')$ equals a sum over possible event times of $R_I(r,T-t)$ multiplied by products of transition probabilities, and the same factorization holds for every multi-event trajectory. The multi-event personal response model is the load-bearing piece: after $M$ events the gamma shape parameter is the sum of event-specific terms, $\\alpha_{M,\\{z_m\\}}(t,\\tau)=\\sum_{j=1}^{M-1}\\theta_{z_j}\\tau_j/(1+\\phi_{z_j}\\tau_j^{k_{z_j}})+\\theta_{z_M}(t-\\sum\\tau_j)/(1+\\phi_{z_M}(t-\\sum\\tau_j)^{k_{z_M}})+\\alpha_N$, with the scale held at the naive value $\\beta_N$. Because independent gamma variables with a common scale add to a gamma, each conditional probability factors cleanly into a sequence probability times a gamma density. The authors fit the parameters by maximum likelihood to two SARS-CoV-2 cohorts and report that most personal trajectories fall within model contours; a time-homogeneous version has a stationary distribution via the Perron-Frobenius theorem.","pith_inferences":["Editorial extension: the additive-shape assumption in Eq. (7) could be tested directly with pre-event titer data; the paper does not run that test because its datasets document events but do not measure antibody level on the event day.","Editorial extension: interpret the conditional probabilities as posterior distributions over missed events: given a measurement and a documented event history, the model can rank hidden reinfections or breakthrough infections by probability. The paper's outlier-trajectory discussion points this way but does not formalize the inversion.","Editorial extension: the closed-form stationary distribution of the time-homogeneous chain suggests an inverse serosurveillance use: match observed cross-sectional antibody distributions to the model to estimate stabilized infection and vaccination incidence rates. The paper simulates forward transmission but does not solve this inverse problem.","Editorial extension: the same transition framework can be ported to influenza, RSV, or pertussis, where booster-timing questions could be recast as the first calendar time at which the predicted probability of falling below a protective antibody threshold crosses a chosen level."],"forward_implications":["Given incidence rates over time, the model outputs the full probability density of antibody levels at any calendar time, so a serosurvey sample can be compared with predicted population distributions rather than only mean prevalence.","Booster-timing decisions can be informed by the predicted fraction of the population below a chosen antibody threshold under alternative vaccination schedules.","An individual whose measured trajectory rises against the model's expected decay is flagged as a candidate for a missed infection or breakthrough event, as illustrated in the paper's outlier analyses.","Adding one more immune event to a personal history requires estimating only three new parameters, so long and varied event sequences remain tractable.","When incidence rates stabilize, the time-homogeneous chain has a stationary distribution that links long-run state prevalences to transition probabilities."],"supporting_citations":[{"why":"Supplies the prior single-event, time-dependent antibody prevalence framework that the multi-event model generalizes.","marker":"[20]"},{"why":"Provides the Markov-chain blueprint for transition matrices and prevalence estimation that this paper extends from single to multiple immune events.","marker":"[21]"},{"why":"Dataset 1 vaccination IgG measurements used for MLE fitting of single- and two-event personal response models.","marker":"[22]"},{"why":"Dataset 1 infection measurements used for fitting the infection response and validating personal trajectories.","marker":"[23]"},{"why":"Dataset 2 longitudinal serology used for validation, including the titration extrapolation procedure.","marker":"[24]"},{"why":"Biological basis for assuming the immune response mounts from its current distribution and wanes toward stabilization after a peak.","marker":"[1]"},{"why":"Closest existing multi-event antibody kinetics treatment (ferret influenza MCMC) that the paper positions its unified framework against.","marker":"[19]"},{"why":"Supplies the log-transformed measurement scale used to turn antibody readings into the variable r.","marker":"[26]"}],"fun_headline_variants":["Markov model predicts antibody levels to time boosters","New math tracks antibody curves for booster decisions","Probabilistic model forecasts immune response to events","Markov chains decode antibody kinetics for vaccination","Modeling antibody response to infection and vaccination with Markov chains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Antibody responses to multiple immune events are assumed to combine as independent additive contributions to the gamma shape, with the spread fixed at the naive value; if a second event's response depends on the current antibody level or changes the variability of the response, the model's conditional probabilities no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Markov model predicts antibody levels to time boosters","New math tracks antibody curves for booster decisions","Probabilistic model forecasts immune response to events","Markov chains decode antibody kinetics for vaccination","Modeling antibody response to infection and vaccination with Markov chains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1772,"prompt_tokens":1059,"completion_tokens":713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":675,"tokens_out":713,"duration_ms":9235,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:26:03.195815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take longitudinal data in which a person's antibody level is measured immediately before a second infection or booster, then repeatedly afterwards, and test whether the size or variability of the post-event rise depends on the pre-event level. If high pre-event titers systematically produce smaller rises than the additive-shape model predicts, or if the spread visibly changes after repeated events, the central factorization fails; a likelihood-ratio comparison against a model with pre-event-titer-dependent shape and scale would settle it.","supporting_citations":[{"cited_title":"Prevalence estimation and optimal classi- fication methods to account for time dependence in antibody levels","cited_arxiv_id":null,"evidence_quote":"Supplies the prior single-event, time-dependent antibody prevalence framework that the multi-event model generalizes."},{"cited_title":"Prevalence estimation methods for time- dependent antibody kinetics of infected and vaccinated individuals: A Markov chain approach","cited_arxiv_id":null,"evidence_quote":"Provides the Markov-chain blueprint for transition matrices and prevalence estimation that this paper extends from single to multiple immune events."},{"cited_title":"Trajectory of IgG to SARS-CoV-2 after vaccination with BNT162b2 or mRNA-1273 in an employee cohort and comparison with natural infection","cited_arxiv_id":null,"evidence_quote":"Dataset 1 vaccination IgG measurements used for MLE fitting of single- and two-event personal response models."},{"cited_title":"Distinct type 1 immune networks underlie the severity of restrictive lung disease after COVID-19","cited_arxiv_id":null,"evidence_quote":"Dataset 1 infection measurements used for fitting the infection response and validating personal trajectories."},{"cited_title":"Twelve- month longitudinal serology in SARS-CoV-2 na ¨ ıve and experienced vaccine recipients and unvaccinated COVID-19-infected individuals","cited_arxiv_id":null,"evidence_quote":"Dataset 2 longitudinal serology used for validation, including the titration extrapolation procedure."},{"cited_title":"SARS-CoV-2-infection- and vaccine-induced antibody responses are long lasting with an initial waning phase followed by a stabilization phase","cited_arxiv_id":null,"evidence_quote":"Biological basis for assuming the immune response mounts from its current distribution and wanes toward stabilization after a peak."},{"cited_title":"Characterising antibody kinetics from multiple influenza infection and vaccination events in ferrets","cited_arxiv_id":null,"evidence_quote":"Closest existing multi-event antibody kinetics treatment (ferret influenza MCMC) that the paper positions its unified framework against."},{"cited_title":"Classification under uncertainty: data analysis for diagnostic antibody testing","cited_arxiv_id":null,"evidence_quote":"Supplies the log-transformed measurement scale used to turn antibody readings into the variable r."}],"review_version":1}