{"id":"cc7ff764-4df5-4cde-a605-f6d83e664abc","arxiv_id":"2506.06233","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CoxPH-SuSiE adds credible sets and calibrated posterior inclusion probabilities to Bayesian variable selection for censored time-to-event outcomes, demonstrated in genetic fine-mapping simulations and UK Biobank asthma data.","lead":"This paper extends the SuSiE Bayesian variable selection method to Cox proportional hazards models, producing posterior inclusion probabilities and credible sets for time-to-event genetic fine-mapping. Simulations show its PIPs are better calibrated than three existing TTE variable selection methods, and it is applied to asthma age-of-onset in UK Biobank.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulation data in Appendix D do not satisfy proportional hazards, so the calibration evidence for the central claim rests on a misspecified data-generating model.","rationale":"The reader's weakest_assumption identified two fragile premises: the Laplace approximation under gIBSS, and the simulation model that is not truly proportional hazards. I agree that both are worth scrutiny, but the simulation issue is the most load-bearing for the central claim because it directly undermines the empirical evidence that the method produces calibrated PIPs and 95% credible sets. If the data are generated from an additive hazards model, the Cox partial likelihood is a working model, and the 'causal' status of a SNP is defined by coefficients in a model that differs from the likelihood used for inference. The paper's explicit assertion that the simulation satisfies PH is incorrect, which is a concrete, checkable error rather than a matter of taste. The Laplace/gIBSS concern is also real, but the authors acknowledge the heuristic nature of gIBSS and the simulations are intended to validate it; if the simulations themselves are invalid, that validation disappears. The real-data application is interesting but does not provide ground truth for calibration, so it cannot substitute for corrected simulations. I recommend keeping the reader's CONDITIONAL verdict: the method may be sound, but the evidence needs to be repaired. The proposed concrete test—rerunning under a true PH model and checking whether calibration and coverage persist—would settle whether the concern lands. If the results hold, the central claim is supported; if not, the paper would need to temper its claims or address the misspecification head-on.","tokens_in":28941,"tokens_out":4574,"duration_ms":44537,"concrete_test":"Re-run the fine-mapping simulations using a true proportional-hazards generator, e.g., T_i ~ Expon(exp(b0 + x_i^T b)), while keeping the genotype matrices, effect sizes, censoring rates, and method settings identical to the paper; then recompute the PIP calibration curves (Figure 3) and CS coverage (Figure 5). If the results change materially (e.g., coverage drops below 90% in some scenarios), the current simulation evidence does not support the central claim. Also inspect the released simulation code to confirm whether it uses exp() or the linear rate as written in Appendix D.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CoxPH-SuSiE delivers well-calibrated PIPs and 95% credible sets for TTE fine-mapping is validated almost entirely by the simulations in Section 5. Those simulations are described in Appendix D. Steps 3–4 set λ_i^s = b0 + x_i^T b and then simulate T_i ~ Expon(λ_i^s). This is an additive hazards model, not a Cox proportional hazards model, which would require λ(t; x) = λ0(t) exp(x^T b). The paper states that data simulated this way 'satisfies the proportional hazard assumption', but it does not. Moreover, in the GTEx simulation setting the effect variance is listed as 1 (Table 1), so sampled coefficients can be large enough to make λ_i^s negative for some genotype values, which is invalid for an exponential rate. This suggests either the text misdescribes the code or the simulation itself is flawed. Because the method's calibration and CS coverage are demonstrated only under this misspecified process, the evidence for the central claim is compromised: good performance under additive hazards does not automatically transfer to the Cox PH model the method is designed for. A concrete check is to rerun the simulations under a true PH data-generating process (for example, T_i ~ Expon(exp(b0 + x_i^T b)) with the same genotype matrices, effect sizes, and censoring rates) and compare Figures 3–5. If calibration and coverage hold, the concern is resolved; if they degrade, the central claim needs to be qualified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the \"Sum of Single Effects\" (SuSiE) approach to Bayesian variable selection to time-to-event outcomes under the Cox proportional hazards model, calling the resulting method CoxPH-SuSiE. The method approximates single-variable Cox partial likelihood Bayes factors using a Laplace approximation, embeds these in a SuSiE-style iterative algorithm (gIBSS), and outputs posterior inclusion probabilities and credible sets. The authors validate the Laplace Bayes factor against numerical quadrature, compare PIP calibration and credible set coverage in simulations against three existing TTE variable-selection methods, and apply the method to fine-mapping eight asthma loci in UK Biobank, reporting 14 asthma risk SNPs in credible sets. The central claim is that CoxPH-SuSiE provides well-calibrated PIPs and 95% credible sets in high-LD, large-sample settings, and outperforms existing TTE fine-mapping methods.","tokens_in":29292,"tokens_out":3480,"duration_ms":37210,"significance":"If the central claim holds, CoxPH-SuSiE would be the first TTE fine-mapping method to deliver both calibrated posterior inclusion probabilities and credible sets in settings with very strong linkage disequilibrium and large sample sizes. The paper is commendable for making its code and data-analysis pipelines publicly available, for validating the Laplace Bayes factor against numerical quadrature, and for explicitly acknowledging the heuristic nature of gIBSS. The real-data application is a useful demonstration of the method, with plausible biological follow-up (e.g., the 10p14/GATA3 regulatory sequence and the FLG stop-gain variants). The main risk to significance is that the simulation evidence, which underpins the central claim, may be generated under an additive hazards model rather than the Cox proportional hazards model the method is designed for.","major_comments":[{"comment":"The simulation procedure used to generate all TTE fine-mapping data (and used implicitly in Section 4's Bayes factor comparisons) does not generate data under the Cox proportional hazards model. Step 3 sets λ_i^s = b0 + x_i^T b, and Step 4 simulates T_i ~ Expon(λ_i^s). The hazard is therefore λ(t; x_i) = b0 + x_i^T b, an additive hazards model, not the Cox PH form λ(t; x_i) = λ0(t) exp(x_i^T b). The statement in Appendix D that data simulated this way \"satisfies the proportional hazard assumption\" is incorrect. Moreover, with the GTEx simulation setting in Table 1 (effect variance 1) and coefficients sampled from N(0,1), the linear predictor b0 + x_i^T b can easily be negative for genotype values 1 or 2 when a coefficient is negative, making the exponential rate invalid. Because Figures 3-5 and the PIP calibration and credible set coverage claims rest entirely on this procedure, the evidence for the central claim is currently compromised. I ask the authors to rerun the simulations under a true PH data-generating process (for example, T_i ~ Expon(exp(b0 + x_i^T b)) with the same genotype matrices, effect sizes, and censoring behavior) and compare the resulting calibration and coverage plots; if calibration and coverage are preserved, the concern is resolved, otherwise the central claim needs to be qualified.","section":"Appendix D, steps 3-4; Table 1"},{"comment":"The paper honestly acknowledges that gIBSS is not known to optimize any objective function and has no convergence guarantee. However, the numerical validation that is used to support the method's approximate posterior inferences is the same simulation procedure criticized above. Thus the two key approximations—the Laplace BF and the gIBSS iterates—are jointly validated only under an additive hazards model, not under the Cox PH model. The repeated application of the Laplace approximation to conditional models with offsets updated from previous effects could accumulate error, and the lack of an objective function means the algorithm's output may depend on initialization, iteration count, or stopping rule. I request additional evidence that the method converges in practice (e.g., sensitivity analyses with different initializations and larger maximum iteration counts) and that the calibration conclusions hold under a true Cox PH DGP. This is compatible with the paper's stated reliance on numerical experiments, but the experiments must target the actual model.","section":"Section 3.3 (gIBSS), Section 4, Section 5"}],"minor_comments":[{"comment":"The displayed expression for the censoring rate contains notational imprecision: the expectation should be over the sum of Bernoulli indicators Pr(C_i < T_i), not over the indicator itself as written; please rewrite for clarity.","section":"Appendix D, eq. (28)"},{"comment":"The text refers to \"median MAS\" but MAS is defined as the mean absolute correlation within a credible set; please clarify which summary is plotted, or use a consistent term.","section":"Figure 5"},{"comment":"There is a typo: \"associations onlyl\" should be \"associations only.\"","section":"Section 6, paragraph after eq. (22)"},{"comment":"The stopping criterion is referred to in the loop but never specified; please state the convergence criterion used in the implementation (e.g., change in elbo-like quantity or maximum absolute change in PIPs).","section":"Algorithm 1"},{"comment":"The text says the simulations in Section 4 use a \"single-SNP CoxPH model,\" but if the same Appendix D procedure is used, that model is additive hazards; please either align the terminology with the corrected simulation design or specify the exact hazard form used for the BF comparison.","section":"Section 4, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"The simulation issue is load-bearing: the paper's central calibration and coverage claims are demonstrated only under a misspecified additive-hazards DGP, and the effect variance setting can produce invalid negative rates. This likely requires rerunning the main simulations under a true proportional hazards process, which is a substantial but feasible revision. The paper is otherwise well structured, reproducible, and honest about the heuristic status of gIBSS, so I view this as a major revision rather than a rejection, provided the authors can confirm that the method's calibration and coverage properties survive under the intended Cox PH model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. CoxPH-SuSiE is the first TTE variable-selection method to give credible sets alongside PIPs, and the Laplace-BF device for single-variable Cox Bayes factors is a concrete, testable idea. The paper ships code, runs 640 simulated datasets, and is upfront that gIBSS is a heuristic with no convergence proof. That's honest and useful.\n\nThe soft spot is real. Appendix D says data are simulated with λ_i^s = b0 + x_i^T b and T_i ~ Expon(λ_i^s). That is additive hazards, not Cox proportional hazards; the paper's assertion that it satisfies PH is wrong. With effect variance 1 and b0=1, the rate can be negative, which is invalid for an exponential. So the calibration and coverage evidence in Figures 3–5 comes from a misspecified DGM. This doesn't necessarily kill the method—the Laplace BF is checked against quadrature and the real-data results look sensible—but the central claim about well-calibrated PIPs and 95% CSs is not yet supported for the actual target model. A rerun under true PH (e.g., Expon(exp(b0+x_i^T b))) would settle it.\n\nAlso, the abstract's \"14 asthma risk SNPs\" overstates: many are just sentinel SNPs in CSs with modest PIPs. That's a minor wording issue.\n\ngIBSS's lack of an objective is a real limitation but the authors own it, and the empirical performance is encouraging.\n\nWho is this for? Statisticians and geneticists doing fine-mapping of TTE outcomes, especially in biobank-scale data. It's a genuinely new combination and the execution is mostly careful.\n\nFor peer review: yes, send it out. A good referee will catch the simulation issue; the paper is worth revision, not rejection.","headline":"Useful new method for TTE fine-mapping, but the calibration evidence is undercut by an additive-hazards simulation that is mislabeled as proportional hazards.","tokens_in":29770,"tokens_out":3475,"would_cite":true,"duration_ms":33012,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62N01","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"CoxPH-SuSiE gives time-to-event fine-mapping calibrated posterior inclusion probabilities and credible sets.","keywords":["Bayesian variable selection","time-to-event data","Cox proportional hazards","SuSiE prior","genetic fine-mapping","posterior inclusion probability","credible sets","UK Biobank"],"falsifier":"Simulate time-to-event data under the true proportional-hazards model with event rate $\\lambda_i(t) = \\lambda_0(t)\\exp(x_i^\\top b)$ rather than the linear-hazard scheme used in Appendix D, plant a single causal variant, run CoxPH-SuSiE, and check whether its 95% credible sets contain the causal variant in at least 95% of replicates and whether SNPs with PIP near 0.95 are causal about 95% of the time.","tokens_in":28748,"feed_emoji":"🧬","tokens_out":11002,"duration_ms":98562,"temperature":0.7,"pith_summary":"This paper introduces CoxPH-SuSiE, a Bayesian variable-selection method for time-to-event outcomes that pairs the Cox proportional hazards model with the 'Sum of Single Effects' (SuSiE) prior. The aim is genetic fine-mapping: given a genomic region, find which variants affect the age at which a disease occurs, with posterior inclusion probabilities (PIPs) and 95% credible sets that quantify uncertainty honestly. The paper argues that CoxPH-SuSiE does this where earlier survival-data variable-selection methods did not: in simulations built on real GTEx and UK Biobank genotypes, its PIPs were the best calibrated among the methods compared, and its credible sets reached or nearly reached their 95% target coverage. Applied to asthma age-of-diagnosis data in UK Biobank, the method found 14 risk SNPs in 8 loci, 6 with PIP above 50%, including two known pathogenic variants in the filaggrin gene.","feed_headline":"Time-to-event fine-mapping gains calibrated credible sets","feed_subtitle":"CoxPH-SuSiE adds the SuSiE prior to survival models, so PIPs and 95% credible sets finally match their claims.","key_machinery":"The load-bearing object is the Laplace-approximate Bayes factor (Eq. 12) for a single-variable CoxPH regression with an offset, computed from the maximum partial-likelihood estimate $\\hat b$, its standard error $s$, and the likelihood ratio. This approximation turns each single-effect update into a Gaussian integral, allowing the SuSiE prior — $b = \\sum_{l=1}^L b_l \\gamma_l$, where each $\\gamma_l$ is a one-hot vector — to be fitted by generalized Iterative Bayesian Stepwise Selection (gIBSS), which cycles through $L$ single effects, updates offsets to remove already-found effects, and recomputes posterior inclusion probabilities. Credible sets are then formed by sorting variants by PIP and adding them until the cumulative probability exceeds the target level, with optional pruning by 'purity', the smallest absolute correlation within the set.","core_discovery":"The central claim is that the SuSiE machinery — representing the regression coefficients as a sum of L single-effect vectors and iterating single-variable regressions — transfers to the Cox partial likelihood, and that the transfer only works if the single-variable Bayes factors are computed with the Laplace approximation (12) rather than the commonly used asymptotic Bayes factor. With that choice, CoxPH-SuSiE produces posterior inclusion probabilities that are better calibrated than those from R2BGLiMS, BVSNLP, survival.svb, and SuSiE-RSS in both GTEx and UK Biobank simulations, and 95% credible sets whose coverage reaches or comes very close to the nominal level. The paper also shows that the method scales to hundreds of thousands of samples and thousands of highly correlated SNPs, and that it can expose secondary association signals, as illustrated by the asthma fine-mapping analysis.","pith_inferences":["Beyond the paper: if the observed calibration holds more generally, CoxPH-SuSiE PIPs could be aggregated across biobank traits to prioritize variants whose effects are shared across multiple time-to-event phenotypes.","The same template — a SuSiE prior plus a Laplace-approximated single-variable Bayes factor — could be carried to logistic or Poisson regression, giving case-control and count fine-mapping the same calibrated credible sets; the paper notes the generality but does not test it.","A stress test worth running: replace the quadratic approximation inside gIBSS with numerical quadrature on small data sets; if credible sets and PIPs barely change, the Laplace step is not the main source of error, and if they change a lot, the calibration observed here may depend on the approximation.","Because the simulation scheme in Appendix D generates event times with rates linear in the genotype, the proportional-hazards assumption is only approximately satisfied; rerunning with strictly exponential hazards would reveal whether the good calibration is robust to the true data-generating mechanism."],"forward_implications":["Time-to-event fine-mapping can now report both PIPs and credible sets, quantities that none of the earlier survival-based variable-selection methods provided.","A SNP with PIP above 0.95 should be a true causal variant roughly 84% of the time in UK Biobank-scale data, compared with at most 35% for the other methods tested.","The Laplace Bayes factor should replace the asymptotic Bayes factor in single-variable survival regressions, since it is more accurate at nearly the same computational cost.","The method's cost is $O(npL)$ per iteration and parallelizes over covariates, making biobank-scale fine-mapping of age-at-onset traits practical.","The asthma application shows that credible-set fine-mapping can reveal multiple independent signals at one locus, as at 2q12.1 and 10p14, which marginal association tests alone would miss."],"supporting_citations":[{"why":"Introduces the SuSiE prior, the SER model, the IBSS algorithm, and the credible-set definition that CoxPH-SuSiE extends.","marker":"Wang et al. (2020)"},{"why":"Defines the proportional hazards model and the partial likelihood used as the likelihood in CoxPH-SuSiE.","marker":"Cox (1972)"},{"why":"Provides the Bayesian justification for treating the partial likelihood as a likelihood, which grounds the whole posterior computation.","marker":"Kalbfleisch (1978)"},{"why":"Defines the asymptotic Bayes factor that the paper compares against and finds less accurate than the Laplace BF.","marker":"Wakefield (2009)"},{"why":"Supplies the Bayes-factor framework and the Laplace approximation that Eq. (12) is a variant of.","marker":"Kass and Raftery (1995)"},{"why":"The survival package used to compute $\\hat b$, $s$, and the likelihood ratio needed for each single-variable Cox fit.","marker":"Therneau and Grambsch (2000)"},{"why":"Introduces SuSiE-RSS, the summary-statistics comparator that CoxPH-SuSiE outperforms in coverage and power.","marker":"Zou et al. (2022)"},{"why":"Provides R2BGLiMS, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE.","marker":"Newcombe et al. (2017)"},{"why":"Provides BVSNLP, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE.","marker":"Nikooienejad, Wang and Johnson (2020)"},{"why":"Provides survival.svb, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE.","marker":"Komodromos et al. (2022)"}],"fun_headline_variants":["CoxPH-SuSiE: calibrated credible sets for survival fine-mapping","Survival fine-mapping gets calibrated credible sets via SuSiE","Laplace approximation fixes Cox fine-mapping credible sets","Scalable survival fine-mapping with SuSiE prior","CoxPH-SuSiE: survival fine-mapping with honest credible sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the quadratic (Laplace) approximation to each single-variable Cox partial log-likelihood stays accurate when applied over and over inside the iterative gIBSS routine, which the authors themselves describe as a heuristic with no proven convergence guarantee.","fun_headline_variants_meta":{"raw":{"variants":["CoxPH-SuSiE: calibrated credible sets for survival fine-mapping","Survival fine-mapping gets calibrated credible sets via SuSiE","Laplace approximation fixes Cox fine-mapping credible sets","Scalable survival fine-mapping with SuSiE prior","CoxPH-SuSiE: survival fine-mapping with honest credible sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001308,"raw_usage":{"total_tokens":5335,"prompt_tokens":954,"completion_tokens":4381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":4289}},"tokens_in":570,"tokens_out":4381,"duration_ms":31797,"temperature":1.0,"reasoning_tokens":4289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:58:04.259956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate time-to-event data under the true proportional-hazards model with event rate $\\lambda_i(t) = \\lambda_0(t)\\exp(x_i^\\top b)$ rather than the linear-hazard scheme used in Appendix D, plant a single causal variant, run CoxPH-SuSiE, and check whether its 95% credible sets contain the causal variant in at least 95% of replicates and whether SNPs with PIP near 0.95 are causal about 95% of the time.","supporting_citations":[{"cited_title":", Sarkar , Abhishek A","cited_arxiv_id":null,"evidence_quote":"Introduces the SuSiE prior, the SER model, the IBSS algorithm, and the credible-set definition that CoxPH-SuSiE extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the proportional hazards model and the partial likelihood used as the likelihood in CoxPH-SuSiE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian justification for treating the partial likelihood as a likelihood, which grounds the whole posterior computation."},{"cited_title":"( 2009 )","cited_arxiv_id":null,"evidence_quote":"Defines the asymptotic Bayes factor that the paper compares against and finds less accurate than the Laplace BF."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayes-factor framework and the Laplace approximation that Eq. (12) is a variant of."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The survival package used to compute $\\hat b$, $s$, and the likelihood ratio needed for each single-variable Cox fit."},{"cited_title":"Sum of Single Effects","cited_arxiv_id":null,"evidence_quote":"Introduces SuSiE-RSS, the summary-statistics comparator that CoxPH-SuSiE outperforms in coverage and power."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides R2BGLiMS, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE."},{"cited_title":", Wang , Wenyi W","cited_arxiv_id":null,"evidence_quote":"Provides BVSNLP, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE."},{"cited_title":", Aboagye , Eric O E","cited_arxiv_id":null,"evidence_quote":"Provides survival.svb, one of the existing time-to-event BVSR baselines whose PIP calibration is compared with CoxPH-SuSiE."}],"review_version":1}