{"id":"f6557d83-1d9f-407a-b99d-5c07c7687cac","arxiv_id":"1908.00650","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A fully Bayesian gamma-negative binomial factor model provides an alternative to zero-inflated approaches for single-cell RNA-seq dimensionality reduction and clustering.","lead":"This paper introduces a hierarchical gamma-negative binomial model for single-cell RNA sequencing that captures overdispersion and dropout zeros without an explicit zero-inflation term. The model performs dimensionality reduction while adjusting for covariates, and the authors show it competes with ZINB-WaVE on several datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"hGNB's claim to obviate zero-inflation is unsupported: a single cell-level r_j fixes P(zero) as a function of mean, and the paper offers only silhouette scores, not zero-calibration checks.","rationale":"The most load-bearing condition for the central claim is that a negative binomial with a single cell-level dispersion parameter can replace an explicit zero-inflation term. This is precisely the reader's weakest_assumption, so agreement is 'agree.' I sharpened it with the explicit zero-probability formula, which shows the constraint the model imposes. The only evidence offered for this condition, silhouette widths and a mean-difference plot, tests cluster recovery and aggregate fit, not zero calibration, so it cannot distinguish the hGNB mechanism from a misspecified NB that happens to separate the simulated clusters. I also inspected the inference equations: eq. (14) prints the ARD precision updates without the required squares (e.g., sum of beta_vp/2 instead of sum of beta_vp^2/2) and with misdirected indices for eta_q and gamma_k; this is a real reproducibility defect but it is a fixable implementation typo rather than a challenge to the modeling claim, so I did not make it the headline. No code or convergence diagnostics further limit verification, and the V1 result is a stated failure case, but those reinforce the conditional verdict without changing it. The concern is strong enough that the paper should not be accepted as establishing the zero-inflation-free claim; it needs a direct zero-calibration validation, but the model and experiments are otherwise plausible, so I keep the reader's CONDITIONAL verdict.","tokens_in":10499,"tokens_out":9935,"duration_ms":90334,"concrete_test":"Generate ZINB data with cluster structure in factor scores, then add gene-specific dropout with probability partially decorrelated from mean (e.g., high dropout on a random half of high-expression genes). Fit hGNB and ZINB-WaVE. Compute a posterior predictive check: from hGNB posterior samples, produce replicate counts and form 95% intervals for the per-gene zero fraction; compare with observed zero fractions stratified by mean-expression decile. If intervals exclude observed zero fractions for the high-expression dropout-perturbed genes, hGNB has not captured zero inflation. Also report silhouette widths to show whether cluster recovery would have masked the miscalibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Under eqs. (1)-(3), for fixed r_j the hGNB zero probability is P(n_vj=0)=(1+mu_vj/r_j)^(-r_j), a one-parameter function of the mean. Thus within a cell, all genes with the same mean have the same predicted dropout rate; gene-specific dropout propensity beyond mean and covariates cannot be represented. Real scRNA-seq dropout is known to depend on gene-specific features and stochastic capture, not only on mean. The paper never proves the NB cell-dispersion mechanism can absorb such zeros; it only reports silhouette widths on ZINB-simulated data (Fig. 3) and an aggregate MD plot (Fig. 2). Silhouette width measures cluster recovery, not zero-count calibration; the MD plot averages over genes and can hide systematic errors. Therefore the central novelty, that hGNB 'obviates the need for explicit zero-inflation modeling,' is an assertion rather than a demonstrated property. If dropout is gene-specific beyond mean, hGNB would mis-specify zero probabilities while possibly still recovering clusters, so the empirical support is too indirect to carry the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes hGNB, a hierarchical gamma-negative binomial factor model for single-cell RNA sequencing counts. The model places a negative binomial distribution on each gene-cell count, with a cell-level dispersion parameter r_j that receives a gamma prior, and models the logit of the NB probability through a sum of cell-level covariates, gene-level covariates, and a bilinear factor term. Inference is performed with a Gibbs sampler that uses Chinese restaurant table augmentation for the dispersion parameters and Pólya-gamma augmentation for the regression and factor parameters. The paper reports an aggregated goodness-of-fit plot, silhouette-width comparisons against PCA, ZIFA, and ZINB-WaVE on three real datasets, a simulation study based on a ZINB-WaVE generative model, and a lineage-inference application on olfactory epithelium data. The central claims are that hGNB captures the high overdispersion of scRNA-seq data without explicit zero-inflation modeling, adjusts for gene- and cell-level covariates, and avoids preprocessing steps such as normalization.","tokens_in":10799,"tokens_out":9478,"duration_ms":89552,"significance":"If fully established, hGNB would be a useful fully generative alternative to ZINB-WaVE: it avoids an extra zero-inflation parameter, retains closed-form Gibbs updates through existing augmentation techniques (Zhou and Carin 2015; Polson et al. 2013), models counts directly, and incorporates covariates at both gene and cell levels. The conditional updates are mostly standard and the overall framework is coherent. However, the paper's distinctive claim, that a cell-level NB dispersion is sufficient to make explicit zero-inflation modeling unnecessary, is currently validated only indirectly: the reported experiments measure cluster recovery and an aggregated mean-difference fit rather than zero-count calibration. In addition, the model specification contains internal inconsistencies in Table 1 and algebraic errors in Eq. (14). These issues are local and fixable in principle, but they are load-bearing for the stated contributions.","major_comments":[{"comment":"The model specification is internally inconsistent. Table 1 constrains φ_vk to be positive with Σ_v φ_vk = 1 and θ_jk > 0, while Eqs. (4)-(5) and the full conditionals in Eqs. (10)-(13) treat φ_v and θ_j as unconstrained multivariate normal random variables. If the constraints are intended, the Gibbs updates do not sample from the correct posterior; if they are not, Table 1 must be corrected. This ambiguity directly affects the interpretation of the latent factor scores used in the clustering and lineage analyses.","section":"hGNB model, Table 1, and Eqs. (4)-(5)"},{"comment":"The conjugacy updates for α_p, η_q, and γ_k contain summation-index and algebraic errors. For α_p the rate should be f0 + (1/2)Σ_v β_vp^2 rather than f0 + Σ_v β_vp/2; for η_q it should be f0 + (1/2)Σ_j δ_jq^2 rather than f0 + Σ_v δ_jq/2; and for γ_k it should be f0 + (1/2)Σ_j θ_jk^2 rather than f0 + Σ_v θ_jk/2. As written, the expressions are not the correct normal-gamma conjugacy updates and involve undefined indices (j in the η_q line, v in the γ_k line).","section":"Eq. (14), precision and rate updates"},{"comment":"The assertion that hGNB 'obviates the need for explicitly modeling zero inflation' is not supported by the evidence presented. Figure 2 reports an aggregated mean-difference plot, which can mask gene- or cell-specific zero-count errors, and Figure 3 reports silhouette widths, which measure cluster separation rather than zero-probability calibration. For a fixed cell-level dispersion r_j, the model implies P(n_vj=0) = (r_j/(r_j + μ_vj))^{r_j}, so within a cell all genes with the same mean have the same predicted dropout probability; gene-specific dropout behavior beyond the mean, or beyond the modeled covariates, cannot be represented. To substantiate the central claim, the paper should provide zero-calibration diagnostics (for example, observed versus posterior-predictive zero fractions per gene and per cell) or a simulation in which dropout depends on gene-specific factors beyond the mean.","section":"Methods and Results, zero-inflation claim"},{"comment":"With the stated zero-mean normal priors on φ_v and θ_j, the bilinear term φ_v^T θ_j is invariant to orthogonal rotations of the factor space, and the paper does not impose or discuss identifying restrictions. The factor scores θ_j are subsequently fed into clustering and lineage-inference algorithms, so the manuscript should state which representation is being used and whether the Gibbs sampler mixes over equivalent modes. The constraints in Table 1 would partially address this issue, but they are inconsistent with the priors as noted above.","section":"Identifiability of the factor decomposition"}],"minor_comments":[{"comment":"In the row for δ_jq, the constraint column reads 'β_vp ∈ R'; this should be 'δ_jq ∈ R'.","section":"Table 1"},{"comment":"The notation φ_v ∼ N(φ_v; 0, I_K) should be N(0, I_K) or N(φ_v | 0, I_K) for consistency with the other prior specifications.","section":"Eq. (5)"},{"comment":"The declared output 'gene module membership matrix' is not what the model produces; the algorithm outputs posterior samples of factor scores, factor loadings, and regression coefficients. Please correct the algorithm description.","section":"Algorithm 1"},{"comment":"The point estimation procedure of retaining the posterior samples with the highest likelihood after burn-in is ad hoc; please justify it or use a standard estimator such as the posterior mean, and report MCMC diagnostics such as trace plots or effective sample sizes.","section":"Results, MCMC estimation"},{"comment":"The simulation results are reported as average silhouette widths without error bars or dispersion measures across the ten simulated datasets; statements such as 'identical to ZINB-WaVE' are hard to assess without this information.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal. I do not see circularity or misconduct concerns; the principal issue is that the headline zero-inflation claim is currently supported only by indirect experiments. The technical corrections to Eq. (14) and Table 1 are straightforward and should be addressable in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes hGNB, a hierarchical gamma-negative binomial factor model for scRNA-seq counts, and shows it can recover known cell types about as well as ZINB-WaVE on a few real datasets, with a silhouette-width edge on the S1/CA1 data. The modeling is a legitimate extension of the authors' earlier NB factor analysis: they add a cell-level gamma prior on the NB dispersion and throw in ARD priors on covariates. The Gibbs updates are standard applications of CRT and Polya-Gamma augmentation, and the derivations in the main text are mostly consistent. That is real, citable work for someone building on Bayesian count factor models.\n\nThe soft spots are real but not fatal. The stress-test note is on target: the claim that hGNB 'obviates the need for explicit zero-inflation modeling' is not backed by a calibration check. With a single cell-level r_j, the zero probability is a one-parameter function of the mean, so gene-specific dropout propensity beyond the mean is not represented. The paper only offers silhouette scores on ZINB-simulated data and one MD plot averaged over genes. Those are indirect; they show cluster recovery, not that the zero probabilities are correct. The identifiability concern about r_j versus the intercept in the logit is also real and unaddressed. A related internal slip: Table 1 constrains φ and θ to be nonnegative/positive while eq. (5) uses zero-mean normal priors, and eq. (14) has summation index typos (using V for the δ/γ updates). Minor, but they make the math look less careful than it is. No code, no convergence diagnostics, and only 2000 MCMC iterations with a 'highest likelihood' point estimate—that is thin for a Bayesian method paper.\n\nWho gets value: someone comparing count-based factor models for scRNA-seq, or a methodologist extending NB factor analysis with hierarchical dispersion. The central claim about zero inflation needs more work before I'd trust it, but the model itself is reasonable and the experiments are honest enough. I'd send this to a serious referee, with the expectation that they ask for zero-calibration diagnostics and a cleanup of the inconsistencies. It deserves peer review, not desk rejection, but it's a solid incremental paper, not a breakthrough.","headline":"A plausible Bayesian factor model for scRNA-seq that extends the authors' earlier NB factor analysis, but the central claim about obviating zero inflation is asserted rather than demonstrated.","tokens_in":11278,"tokens_out":595,"would_cite":false,"duration_ms":8069,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H25","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A hierarchical gamma-negative binomial model can capture the dropout-dominated counts of single-cell RNA sequencing without explicitly modeling zero inflation, while also adjusting for covariates and avoiding preprocessing normalization.","keywords":["single-cell RNA sequencing","negative binomial","hierarchical gamma prior","zero-inflation","dropout","dimensionality reduction","Bayesian Gibbs sampling","cell clustering"],"falsifier":"Generate synthetic data from a zero-inflated NB model with known latent structure and widely different per-gene dropout rates, then check whether hGNB's inferred factor scores recover the true clusters and whether its cell-level dispersion estimates track the fraction of zeros per cell. If hGNB's clustering silhouette drops below an explicit zero-inflation method at high zero rates, or if the posterior dispersion does not correlate with observed dropout proportions, the central claim fails.","tokens_in":10364,"feed_emoji":"🧬","tokens_out":6476,"duration_ms":61399,"temperature":0.7,"pith_summary":"This paper tries to establish that a fully generative hierarchical gamma-negative binomial (hGNB) model can capture the high over-dispersion and abundant zeros of single-cell RNA sequencing data without adding an explicit zero-inflation component. Such zeros, often called dropouts, are usually thought to require a separate zero-inflation term; the paper argues they can be absorbed by a cell-level dispersion parameter drawn from a gamma prior. If true, hGNB provides a single Bayesian model that simultaneously performs dimensionality reduction, adjusts for gene- and cell-level covariates, and skips common preprocessing steps like normalization. The experiments on simulated and real datasets support the claim by showing competitive or better cell-cluster separation than existing zero-inflated methods, plus a demonstration of lineage inference on olfactory epithelium cells.","feed_headline":"One model absorbs single-cell RNA dropouts without zero-inflation","feed_subtitle":"hGNB treats dropout zeros as cell-level dispersion, matching ZINB-WaVE on real and simulated clustering tasks.","key_machinery":"The load-bearing object is the hierarchical gamma prior on the negative binomial dispersion parameter, $r_j \\sim \\mathrm{Gamma}(e_0, 1/h)$, shared across all genes within a cell. This single cell-level dispersion is what lets the model absorb both biological over-dispersion and dropouts without an explicit zero-inflation component. The inference machinery consists of two data-augmentation identities: the Chinese Restaurant Table representation of the negative binomial, $(\\ell_{vj}\\mid n_{vj}, r_j)\\sim\\mathrm{CRT}(n_{vj}, r_j)$, which makes the dispersion update conjugate, and the Polya-Gamma augmentation $\\omega_{vj}\\sim\\mathrm{PG}(n_{vj}+r_j, \\psi_{vj})$, which turns the logit-link regression and factor terms into conditionally normal updates. Together these give closed-form Gibbs updates for all parameters.","core_discovery":"The paper's central discovery is that the excess zeros and over-dispersion of scRNA-seq counts need not be modeled through a separate zero-inflation mechanism. In hGNB, each count $n_{vj}$ follows a negative binomial distribution with cell-specific dispersion $r_j$, and a gamma prior on $r_j$ lets the dispersion vary flexibly across cells. Because deep-sequenced cells have fewer dropouts, the paper argues that $r_j$ encodes dropout prevalence, so a hierarchical NB likelihood alone can represent zero-inflated data. The logit of the NB probability is factorized into covariate regression terms and latent factor terms, enabling joint dimensionality reduction and covariate adjustment. Using compound-Poisson/CRT and Polya-Gamma augmentations, the model admits closed-form Gibbs updates, and experiments show it recovers clusters from data simulated under a zero-inflated NB model and separates known cell types on real datasets.","pith_inferences":["Editorial inference: the ARD priors on factor loadings could let the number of latent factors $K$ be chosen automatically by posterior shrinkage, but the paper fixes $K$ and does not test this.","Editorial inference: because the model puts one dispersion parameter per cell, it should be most successful when dropout is dominated by cell-level technical variation; a natural next test is to simulate gene-specific dropout rates and check whether $r_j$ alone still absorbs them.","Editorial inference: an identifiability analysis comparing hGNB with a zero-inflated NB on the same data would clarify when the two models are distinguishable; the paper provides empirical fits but no formal separation argument."],"forward_implications":["Raw scRNA-seq counts can be used directly: the model's cell-level covariates absorb sequencing depth and other technical effects, removing the need for a separate normalization step before dimensionality reduction.","The per-cell dispersion parameter $r_j$ provides a quantitative readout of dropout prevalence, since deep-sequenced cells should have fewer zeros and correspondingly different $r_j$ values.","The posterior factor scores $\\theta_j$ are sufficient for both clustering and trajectory inference, as demonstrated by the olfactory epithelium analysis where a minimum spanning tree recovered three developmental branches.","Explicitly zero-inflated methods may place too much weight on zeros; if hGNB's hierarchical dispersion suffices, the simpler NB model avoids that bias while matching or beating ZINB-WaVE in cluster separation."],"supporting_citations":[{"why":"Supplies the zero-inflated NB simulation scheme and the ZINB-WaVE baseline that hGNB is compared against for clustering and zero-inflation capture.","marker":"[9]"},{"why":"Provides the compound-Poisson/CRT augmentation that makes the cell dispersion parameter conjugate and updatable in closed form.","marker":"[13]"},{"why":"Provides the Polya-Gamma augmentation used for closed-form updates of regression coefficients and factor parameters.","marker":"[14]"},{"why":"Supplies the ZIFA zero-inflated dimensionality reduction baseline, used as a comparison and motivation for avoiding explicit zero-inflation.","marker":"[7]"},{"why":"The covariate-dependent negative binomial factor analysis that hGNB extends by adding the hierarchical gamma prior on cell dispersion.","marker":"[12]"},{"why":"One of the real-world scRNA-seq datasets (S1/CA1) on which hGNB's cluster separation is evaluated.","marker":"[22]"}],"fun_headline_variants":["Dropouts as dispersion: a gamma-NB model for scRNA-seq","Single-cell RNA zeros? Model them as dispersion, not inflation","No zero-inflation: hGNB handles scRNA-seq dropouts naturally","Gamma-NB model absorbs dropouts without zero-inflation","Zero-inflation? Not needed: hGNB uses dispersion instead"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that one per-cell dispersion number, shared across all genes, can absorb both the biological variability and the dropouts, so no separate zero-inflation mechanism is needed; the paper shows this empirically but gives no formal proof that the two are distinguishable.","fun_headline_variants_meta":{"raw":{"variants":["Dropouts as dispersion: a gamma-NB model for scRNA-seq","Single-cell RNA zeros? Model them as dispersion, not inflation","No zero-inflation: hGNB handles scRNA-seq dropouts naturally","Gamma-NB model absorbs dropouts without zero-inflation","Zero-inflation? Not needed: hGNB uses dispersion instead"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1533,"prompt_tokens":969,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":473}},"tokens_in":585,"tokens_out":564,"duration_ms":5178,"temperature":1.0,"reasoning_tokens":473,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:41:00.564750+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic data from a zero-inflated NB model with known latent structure and widely different per-gene dropout rates, then check whether hGNB's inferred factor scores recover the true clusters and whether its cell-level dispersion estimates track the fraction of zeros per cell. If hGNB's clustering silhouette drops below an explicit zero-inflation method at high zero rates, or if the posterior dispersion does not correlate with observed dropout proportions, the central claim fails.","supporting_citations":[{"cited_title":"Nature communications 9(1), 284 (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the zero-inflated NB simulation scheme and the ZINB-WaVE baseline that hGNB is compared against for clustering and zero-inflation capture."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence 37(2), 307–320 (2015)","cited_arxiv_id":null,"evidence_quote":"Provides the compound-Poisson/CRT augmentation that makes the cell dispersion parameter conjugate and updatable in closed form."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Polya-Gamma augmentation used for closed-form updates of regression coefficients and factor parameters."},{"cited_title":"Genome biology 16(1), 241 (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the ZIFA zero-inflated dimensionality reduction baseline, used as a comparison and motivation for avoiding explicit zero-inflation."},{"cited_title":"Bioinformatics 34(13), 61–69 (2018)","cited_arxiv_id":null,"evidence_quote":"The covariate-dependent negative binomial factor analysis that hGNB extends by adding the hierarchical gamma prior on cell dispersion."},{"cited_title":": Cell types in the mouse cortex and hippocampus revealed by single-cell rna-seq","cited_arxiv_id":null,"evidence_quote":"One of the real-world scRNA-seq datasets (S1/CA1) on which hGNB's cluster separation is evaluated."}],"review_version":1}