{"id":"d89d031d-fcf8-4d8c-a40e-ea432ca86a22","arxiv_id":"2411.19262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VBVarSel is an annealed variational Bayes algorithm for simultaneous clustering and variable selection that is claimed to be substantially faster than current methods with comparable accuracy.","lead":"This paper introduces VBVarSel, an annealed variational Bayes algorithm that clusters high-dimensional data while selecting the informative variables. It reports large speedups over existing clustering methods on simulated and real cancer 'omics data, with equal or better accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Within-cluster independence is the load-bearing weakness: realistic correlated omics data could invalidate the claimed state-of-the-art accuracy.","rationale":"The reader's weakest assumption is the same as the load-bearing concern identified here: the model's diagonal precision matrix assumes within-cluster independence, and robustness is tested only up to correlation 0.5. This is the right focal point because it directly connects to the abstract's claim that VBVarSel outperforms the state of the art on real biomedical data, where structured gene/protein correlations are the norm rather than the exception. The paper deserves credit for providing an open-source Python implementation, extensive simulation tables, and supplementary experiments on noise, initialization, and mild correlation. However, those experiments do not cover the realistic high-correlation regime, and the real-data analyses are largely qualitative, so the outperformance claim is not yet secured for the target setting. The verdict should remain CONDITIONAL: the method is promising and the implementation is reusable, but the central accuracy claim should be narrowed or further validated under realistic correlation structure before being accepted as state-of-the-art for omics subtyping. This does not require rejecting the paper; it requires either a new simulation experiment or a qualification of the claim.","tokens_in":24716,"tokens_out":5477,"duration_ms":55941,"concrete_test":"Run the same simulation protocol as the paper (Crook et al. 2019) with n=1000, J=200, and 10% relevant variables, but generate relevant variables from a block-diagonal covariance matrix with realistic gene-module structure: within-block correlations 0.7-0.95, several modules shared across clusters, and some high-correlation irrelevant modules. Use the paper's stated 'optimal' prior settings (Table 1 of the Supplement) and compare VBVarSel against VarSelLCM and sparse k-means over 20 repetitions. If VBVarSel's median ARI drops below roughly 0.9 or its irrelevant-variable selection rate exceeds roughly 0.1 while comparators are materially better, the state-of-the-art claim is not robust to realistic correlation. If VBVarSel maintains near-perfect ARI and selection accuracy, the concern does not land.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that VBVarSel is scalable, computationally efficient, and outperforms the current state of the art for high-dimensional biomedical clustering. The accuracy part of that claim rests on the conditional-independence assumption: given cluster assignment, covariates are independent (main text Eq. 2; Supplement Eq. 18: 'This factorisation is equivalent to having Λ_k as a diagonal matrix with diagonal entries τ_kj'). This is a strong misspecification for transcriptomic and proteomic data, where genes and proteins form highly correlated modules. The paper's robustness experiments only introduce correlations up to 0.5, either fixed, per-cluster, or randomly sampled (Supplement Tables 2-4); they do not test realistic block-structured correlation with high within-block values (0.7-0.95). Under such misspecification, the variational variable-selection indicators c_j and responsibilities r_nk can be distorted: irrelevant but correlated variables may be preferentially selected, relevant variables may be deselected because their signal is absorbed by correlated neighbours, and cluster assignments can be blurred. If this happens, the headline 'outperforms the current state of the art' is only established for near-diagonal settings, not for the real biomedical regime the paper claims. This is not an internal inconsistency; it is an unsecured generalization of the central claim. The open-source implementation and careful simulation work are real strengths, but they do not remove the need to test the method where its core modelling assumption is most strained.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VBVarSel, an annealed variational Bayes algorithm for simultaneous clustering and variable selection in high-dimensional Gaussian mixture models. The model assumes conditional independence of covariates within clusters (diagonal precision matrices), uses an overfitted mixture with a Dirichlet prior on weights to infer the number of clusters, and models variable selection through Bernoulli indicators with Beta priors. The authors derive mean-field variational updates, add an annealing temperature schedule, and validate the method on simulations following Crook et al. (2019), comparing runtime, variable-selection accuracy, and Adjusted Rand Index against hierarchical clustering, sparse K-means, SUGSVarSel, VSCC, and VarSelLCM. They also apply VBVarSel to TCGA breast cancer transcriptomics and TCPA proteomics data. The central claims are that VBVarSel is scalable, computationally efficient, and outperforms the current state of the art in accuracy and runtime.","tokens_in":24901,"tokens_out":5786,"duration_ms":54683,"significance":"If fully substantiated, VBVarSel would be a practically valuable addition to the toolkit for high-dimensional biomedical subtyping: it provides an open-source Python implementation, handles feature selection and cluster-number selection jointly, and shows large runtime advantages over MCMC- and IC-based alternatives in the reported experiments. The manuscript deserves credit for reporting quartiles over repeated simulation runs, for benchmarking on an externally generated simulation design, and for making code publicly available. However, the evidence as presented supports runtime advantages more strongly than accuracy advantages, and the central 'outperforms the state of the art' claim rests on settings in which the model's conditional-independence assumption is either exactly satisfied or only mildly violated.","major_comments":[{"comment":"The model's central factorization assumes that covariates are independent given cluster allocation, so that each within-cluster precision matrix is diagonal. The robustness experiments in Supplement Tables 2-4 only introduce correlations up to 0.5, and the main simulation benchmark uses exactly diagonal covariances. Real transcriptomic and proteomic data commonly exhibit block-structured correlation with high within-block values (0.7-0.95), under which the variational variable-selection indicators and cluster responsibilities can be distorted. Please add simulation scenarios with block-diagonal within-cluster covariance at realistic correlation levels and report ARI, selection proportions, and runtime, or explicitly restrict the claim of state-of-the-art accuracy to near-diagonal settings. This is load-bearing for the abstract and Discussion claims about biomedical applications.","section":"Eq. (2) / Supplement Eq. (18), Supplement Tables 2-4"},{"comment":"The simulation results do not support an accuracy advantage over VarSelLCM: in Tables 1-4, VarSelLCM matches VBVarSel's perfect or near-perfect ARI and variable-selection proportions in every setting, and SUGSVarSel also attains perfect results in the n=1000 rows of Tables 3 and 4. The text's statement that 'VBVarSel and VarSelLCM are the only two methods that consistently achieve perfect results' is therefore inaccurate. The supported advantage of VBVarSel over these methods is runtime, not accuracy. Please revise the accuracy claims and, ideally, benchmark against an additional recent method on real data where ground-truth subtypes are available for quantitative comparison.","section":"Results, Simulation study, Tables 1-4"},{"comment":"Algorithm 1 monitors the standard ELBO of Eq. (62) for convergence, but for T>1 the quantity being optimized is the annealed ELBO of Eq. (9), which does not provide a lower bound on the standard ELBO. Under fixed T>1 or during the early phase of a geometric/harmonic schedule, the standard ELBO need not be monotone, so the convergence criterion '0 < improve < epsilon' may fail to trigger or may stop at an iterate that does not optimize the annealed objective. Please state explicitly which objective is monitored, use the annealed objective for convergence and model selection, or provide a monotonicity result for the annealed updates.","section":"Algorithm 1; Eqs. (9) and (62); Supplement Section 1.5"},{"comment":"The manuscript itself reports that 'VBVarSel requires very careful tuning of b0j' and that 'even slight deviations from optimal would significantly and detrimentally impact the quality of the stratification', and the Discussion states that performance is sensitive to the temperature schedule. Given that the paper's selling point is a scalable, efficient, turnkey alternative to existing methods, this acknowledged sensitivity needs systematic treatment: report a sensitivity analysis over b0j, initial temperature, and schedule type, and provide explicit practical guidance on how these hyperparameters should be chosen in new applications.","section":"Supplement Section 3.1; Discussion"}],"minor_comments":[{"comment":"The abstract contains a typo: 'biomedeical' should be 'biomedical'.","section":"Abstract"},{"comment":"In Table 1, the SUGSVarSel runtime entry for n=100 is reported as '9.71 [9.98, 8.85]', with the median lying below the lower quartile and above the upper quartile; the quartile order appears to be reversed or the entries are mislabeled.","section":"Table 1, SUGSVarSel row"},{"comment":"In Eq. (37), the second term inside the sum should be a log-density, i.e. (1-c_j) ln f_j(x_nj | Phi_0j), not (1-c_j) f_j(x_nj | Phi_0j), as written.","section":"Supplement Eq. (37)"},{"comment":"The expression in Eq. (67) contains '2v' in the digamma argument, which should presumably be '2d0' as in Eq. (66); please correct the notation.","section":"Supplement Eqs. (66)-(67)"},{"comment":"The description of model selection says that multiple parameter initializations and temperature schedules are evaluated and the one maximizing the ELBO is chosen; because this selection is done on the same data later used to report performance, the manuscript should clarify whether any held-out or repeated-splitting procedure is used, or note that this constitutes tuning on the evaluation set.","section":"Performance evaluation section"}],"recommendation":"major_revision","confidential_remarks":"The negative comparison against SUGSVarSel is worth a transparency check: the simulation design is taken from Crook et al. (2019), whose author list includes P.D.W. Kirk, also an author of this manuscript. The paper should state that the competing method is evaluated in the simulation scenario proposed by its own authors and disclose the relationship, so that reviewers and readers can judge whether the settings favor one implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on the annealed variational mixtures paper. The useful core is real: mixing annealing with mean-field VI for finite mixtures and latent variable-selection indicators appears to be new. Ruffieux et al. did the annealing for regression; Tadesse et al. only waved at parallel tempering. The code is open-source, and the simulation study is careful: external generator, quartiles over 10 runs, several n and relevance settings. The headline result is a 90-2000x speedup over VarSelLCM while matching its accuracy. That alone is worth having.\n\nThe abstract's \"outperforms the current state of the art\" is too strong. On accuracy, VBVarSel and VarSelLCM are tied in most cells; the real differentiator is runtime. The annealing benefit also rests on selectively reported schedules; the supplement admits only the advantageous schedules are shown, and improvements over T=1 are often modest (e.g., ARI 0.97 to 1). The bigger caveat is the diagonal-precision assumption in Eq. 2/Supp. Eq. 18. Real transcriptomics and proteomics carry block-structured correlation well above 0.5, and the robustness tests stop at 0.5 without testing module structure. Under that misspecification, variable selection could latch onto correlated noise or drop redundant-but-relevant genes. This is not an internal inconsistency—the derivations are consistent—it is an unsecured generalization of the central claim. The supplement's own warning that b0j needs \"very careful tuning\" also tempers the turnkey impression.\n\nThe real-data sections are mostly qualitative, and the TCGA analysis doesn't benchmark against VarSelLCM on the same preprocessed data. The permutation-based semi-synthetic experiments are a nice touch and support the variable-selection mechanism.\n\nBottom line: this is a practical, mostly honest methods paper for model-based clustering on high-dimensional biomedical data. It deserves a serious referee, but the revision should rebalance the claims, add block-correlation robustness experiments, and discuss the tuning and schedule-selection burden openly. I'd cite it for the annealed VI + variable selection combination.","headline":"A practical annealed-VI + variable-selection combination that is genuinely new, but 'outperforms' should read 'matches accuracy with a big speedup,' and the diagonal-covariance assumption is untested where real omics data live.","tokens_in":25520,"tokens_out":4355,"would_cite":true,"duration_ms":39891,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H30","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that VBVarSel, an annealed variational Bayes algorithm, simultaneously clusters high-dimensional samples and selects the biomarkers that define the clusters, outperforming current methods in both speed and accuracy.","keywords":["variational Bayes","annealed variational inference","mixture models","variable selection","cancer subtyping","biomarker discovery","high-dimensional clustering","overfitted mixtures"],"falsifier":"Take a dataset with known cluster labels and a strong within-cluster correlation block, such as 200 simulated features in three clusters with pairwise correlation above 0.8 inside a co-expressed module plus 20 pure signal features, run VBVarSel, and compare its Adjusted Rand Index and variable-selection accuracy against ground truth; if either metric falls far below the near-perfect values reported for independent features, the central robustness claim is contradicted. The TCGA PAM50 analysis offers a real-data check: force the model to include correlated neighbours of the PAM50 genes and see whether PAM50 enrichment remains far above random.","tokens_in":24438,"feed_emoji":"🧬","tokens_out":7873,"duration_ms":66226,"temperature":0.7,"pith_summary":"The paper introduces VBVarSel, an annealed variational Bayes algorithm that fits high-dimensional Gaussian mixture models while simultaneously deciding which variables are informative for clustering and which are noise. The authors' central claim is that VBVarSel is both faster and more accurate than existing tools for this joint task, and they support it with simulations and two cancer datasets: breast-cancer gene expression and a pan-cancer protein panel. If the claim holds, disease subtyping and biomarker discovery could be run on full, unfiltered omics data without first reducing dimension or pre-selecting genes. The practical payoff is one open-source Python routine that returns cluster labels, a set of candidate biomarkers, and an inferred number of subtypes in minutes to about an hour.","feed_headline":"One algorithm clusters tumors and selects biomarkers at once","feed_subtitle":"Runs on full TCGA and TCPA panels without pre-filtering, recovering known cancer subtypes and PAM50 markers.","key_machinery":"The engine is an annealed mean-field variational Bayes routine over an overfitted finite Gaussian mixture with a diagonal per-cluster precision matrix. Three moving parts do the work: a temperature parameter multiplying the entropy term in the ELBO, which flattens the objective at high temperature and is reduced over time so the optimizer escapes shallow local optima; a latent binary indicator per variable that switches each covariate between its cluster-specific distribution and a common null distribution, so variable selection and clustering are solved in one objective; and a symmetric Dirichlet prior with concentration below one that lets surplus mixture components empty out, so the number of clusters is inferred rather than fixed. All conditional updates stay conjugate, so each iteration is a set of closed-form coordinate updates.","core_discovery":"The central claim is that combining variational inference, simulated-annealing temperature schedules, and latent binary feature-selection indicators makes simultaneous clustering and variable selection scalable enough for real biomedical data while matching or exceeding the accuracy of slower reference methods. On the Crook et al. simulation design, VBVarSel recovers the true three-cluster structure and the exact relevant-variable set in most settings, with runtimes of about 1 to 28 seconds for 100 to 1000 observations and 200 variables, against minutes to tens of thousands of seconds for competing implementations. On a 348 by 17373 TCGA breast-cancer transcriptomic matrix it converges in under an hour without preprocessing, selects roughly 39% of genes, and retains the PAM50 marker genes at a rate the paper reports as far above random and consistent with known breast-cancer subtypes. On a 5157 by 217 TCPA protein panel it produces clusters that largely match tissue of origin and groups related cancers together, keeping about 90% of the pre-selected proteins. The paper concludes that VBVarSel outperforms the current state of the art in accuracy and runtime, and credits the annealing component for improved robustness to poor initialization, correlated features, and noise.","pith_inferences":["This inference goes beyond the paper: the conditional-independence assumption is the natural boundary of the speed advantage, and on data with strong co-expression modules or long-range correlation, a sparse-precision or factor-analytic covariance variant would likely be needed.","This inference goes beyond the paper: the reported sensitivity to the scale hyperparameter and the empirical choice of temperature schedule suggest that automatic schedule and hyperparameter selection would be the highest-value usability improvement, though the paper does not propose one.","This inference goes beyond the paper: the binary selection indicator may be why the TCPA analysis retains 90% of variables, and a continuous saliency score could better capture subtle biomarker importance, which the authors mention as future work.","This inference goes beyond the paper: a direct scaling test on single-cell RNA-seq or multi-omics matrices would show whether the runtime advantage generalizes, since both real applications in the paper are bulk tissue with standard array technologies."],"forward_implications":["On the paper's simulations, VBVarSel recovers the true clustering and the relevant-variable set at near-perfect rates while running at least 2.5 times faster than the next-fastest method at 1000 observations, and it can complete the full TCGA breast-cancer study in under an hour.","The method removes the need for a separate feature-selection step: cluster assignment and biomarker identification come from the same fitted model, and the overfitted-mixture prior makes a pre-specified number of clusters unnecessary.","Annealing is what buys robustness: the paper reports that geometric or harmonic temperature schedules restore near-perfect accuracy under sub-optimal initialization, added Gaussian noise, and within-cluster correlations up to 0.5, where the non-annealed version degrades.","Because the objective is variational, the algorithm produces a full approximate posterior over allocations and selection indicators, not just point estimates, which is what allows direct probabilistic statements about biomarker relevance.","The two real-data applications suggest the same code can handle both wide transcriptomic data with 17,373 genes and 348 samples and tall proteomic data with 217 proteins and 5,157 samples, positioning it as a general tool for omics subtyping."],"supporting_citations":[{"why":"Supplies the simulation design of three spherical Gaussians with 200 variables and varying relevance, the SUGSVarSel baseline, and the original pan-cancer proteomics application.","marker":"(Crook et al, 2019)"},{"why":"Provides the sparse k-means sparcl baseline that VBVarSel must beat on both clustering accuracy and variable selection.","marker":"(Witten and Tibshirani, 2010)"},{"why":"Yields the VarSelLCM baseline, the only competitor that matches VBVarSel's accuracy on simulations but is hundreds to thousands of times slower.","marker":"(Marbac and Sedki, 2017)"},{"why":"Supplies the VSCC baseline, used to show the failure mode of methods that retain too many variables.","marker":"(Andrews and McNicholas, 2014)"},{"why":"Furnishes the annealed variational inference framework and geometric temperature schedule that VBVarSel adapts to clustering with variable selection.","marker":"(Ruffieux et al, 2020)"},{"why":"Provides the deterministic-annealing variant of variational Bayes and the rationale for fixed temperatures above one.","marker":"(Katahira et al, 2008)"},{"why":"Supports the exploration-exploitation argument behind the temperature schedule with variational tempering.","marker":"(Mandt et al, 2016)"},{"why":"Gives the overfitted-mixture theory that lets VBVarSel infer the number of clusters by shrinking surplus components.","marker":"(Rousseau and Mengersen, 2011)"},{"why":"Is the source of the TCGA data and pan-cancer context for the breast-cancer transcriptomic analysis.","marker":"(Weinstein et al, 2013)"},{"why":"Is the source of the TCPA protein expression data and the pre-selected protein panel used in the pan-cancer proteomic analysis.","marker":"(Akbani et al, 2014)"}],"fun_headline_variants":["VBVarSel: tumor subtyping and biomarker selection in one go","Joint clustering and variable selection for omics data","Annealed variational mixtures: cluster and select genes simultaneously","Scalable subtyping with built-in biomarker discovery","Outperforms state-of-the-art in clustering and marker selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim's load-bearing premise is that, within each cluster, the measured features are independent of one another, because the model uses a diagonal precision matrix and the paper only tests correlations up to 0.5 in simulation; if real omics data carry stronger or structured correlation, the cluster labels and biomarker lists could be systematically biased.","fun_headline_variants_meta":{"raw":{"variants":["VBVarSel: tumor subtyping and biomarker selection in one go","Joint clustering and variable selection for omics data","Annealed variational mixtures: cluster and select genes simultaneously","Scalable subtyping with built-in biomarker discovery","Outperforms state-of-the-art in clustering and marker selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2981,"prompt_tokens":910,"completion_tokens":2071,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1990}},"tokens_in":526,"tokens_out":2071,"duration_ms":15508,"temperature":1.0,"reasoning_tokens":1990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:22:24.207810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with known cluster labels and a strong within-cluster correlation block, such as 200 simulated features in three clusters with pairwise correlation above 0.8 inside a co-expressed module plus 20 pure signal features, run VBVarSel, and compare its Adjusted Rand Index and variable-selection accuracy against ground truth; if either metric falls far below the near-perfect values reported for independent features, the central robustness claim is contradicted. The TCGA PAM50 analysis offers a real-data check: force the model to include correlated neighbours of the PAM50 genes and see whether PAM50 enrichment remains far above random.","supporting_citations":[{"cited_title":"Statistical Applications in Genetics and Molecular Biology 18(6):20180065","cited_arxiv_id":null,"evidence_quote":"Supplies the simulation design of three spherical Gaussians with 200 variables and varying relevance, the SUGSVarSel baseline, and the original pan-cancer proteomics application."},{"cited_title":"Journal of the American Statistical Association 105(490):713--726","cited_arxiv_id":null,"evidence_quote":"Provides the sparse k-means sparcl baseline that VBVarSel must beat on both clustering accuracy and variable selection."},{"cited_title":"Statistics and Computing 27:1049--1063","cited_arxiv_id":null,"evidence_quote":"Yields the VarSelLCM baseline, the only competitor that matches VBVarSel's accuracy on simulations but is hundreds to thousands of times slower."},{"cited_title":"Journal of Classification 31(2):136--153","cited_arxiv_id":null,"evidence_quote":"Supplies the VSCC baseline, used to show the failure mode of methods that retain too many variables."},{"cited_title":"Variational Tempering","cited_arxiv_id":"1411.1810","evidence_quote":"Supports the exploration-exploitation argument behind the temperature schedule with variational tempering."},{"cited_title":"Journal of the Royal Statistical Society Series B: Statistical Methodology 73(5):689--710","cited_arxiv_id":null,"evidence_quote":"Gives the overfitted-mixture theory that lets VBVarSel infer the number of clusters by shrinking surplus components."},{"cited_title":"Nature genetics 45(10):1113--1120","cited_arxiv_id":null,"evidence_quote":"Is the source of the TCGA data and pan-cancer context for the breast-cancer transcriptomic analysis."},{"cited_title":"Nature communications 5(1):3887","cited_arxiv_id":null,"evidence_quote":"Is the source of the TCPA protein expression data and the pre-selected protein panel used in the pan-cancer proteomic analysis."}],"review_version":1}