{"id":"0220713c-e708-4b12-bfe9-3e5cc52a727d","arxiv_id":"1908.09653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The thesis proposes logistic PCA via non-convex singular value thresholding, generalized SCA for binary and quantitative data, and P-ESCA for multiple mixed-type data sets with structured sparsity to separate common and distinct variation.","lead":"This PhD thesis develops statistical methods for combining measurements taken at different biological levels and different scales, such as binary mutation status and continuous gene expression. It introduces penalized versions of PCA and simultaneous component analysis that handle mixed data types and separate shared from dataset-specific variation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The shared-subspace equality across data types is the least secure condition for the P-ESCA disentangling claim; it is assumed by construction but never stress-tested.","rationale":"The reader identified the shared-subspace equality as the weakest assumption, and I agree: P-ESCA's interpretation of a zero loading block as 'no shared variation' is only valid if the common score matrix A exactly spans the column spaces of all natural parameter matrices. This assumption is stated explicitly in Section 5.2.1, but the simulations in Section 5.6 generate data under exactly that model, so they provide no evidence about what happens when the assumption is violated. The real-data analysis is exploratory and cannot validate the decomposition either. This is the most load-bearing concern because the paper's central novelty is the disentangling of global, local common, and distinct variation; if the shared-subspace condition fails, the disentangling is an artifact of the misspecified model. While the paper also lacks shipped code, repeated-seed error bars, and has uneven baseline tuning, those are secondary to the structural assumption about the data-generating process. The proposed concrete test would directly probe this assumption by simulating a realistic departure and checking whether P-ESCA's pattern recovery and score estimation degrade. Until such a test is performed, the conditional verdict is appropriate: the method works when its key assumption holds, but its applicability to real heterogeneous data remains unverified.","tokens_in":54701,"tokens_out":8754,"duration_ms":99037,"concrete_test":"Generate two blocks from a single latent factor Z: X_1 (quantitative) with mean μ1 + Z γ1, and X_2 (binary) with log-odds ν2 + logit(Φ(α + β Z)) (a probit-like nonlinearity) or another saturating transform, so the true natural parameters do not lie in the same column subspace. Fit P-ESCA on 50 random seeds using the same CV model selection as Section 5.6, and measure (i) the RV coefficient between the estimated score A and the true Z, and (ii) the proportion of runs where the estimated sparsity pattern (which blocks are zero) matches the true pattern. If the estimated scores are distorted or truly shared blocks are zeroed out, the shared-subspace assumption is the cause. As a control, repeat with data generated exactly under the ESCA model (linear log-odds and linear mean) to confirm that the method recovers the intended pattern when the assumption holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"P-ESCA's central claim rests on the assumption (Section 5.2.1) that the natural parameter matrices {Θ_l} of all data sets lie in exactly the same column subspace spanned by the common score matrix A. On this assumption, a zero block in B_l is interpreted as an absence of shared variation between block l and the other blocks. If the true data-generating mechanisms are not exactly aligned on the linear-predictor scale — for instance, if binary observations arise from a probit liability process or a nonlinear threshold effect, or if a shared factor affects the log-odds of mutations and the mean expression through different scaling functions — then the Θ_l will not share a common A exactly. The group concave penalty will then shrink some blocks to zero to fit the misspecified model, and the resulting global/local/distinct pattern reflects the misspecification rather than the underlying biological structure. This is load-bearing because the paper's simulations (Section 5.6) and real-data illustration are the only support for the claim, and the simulations generate data exactly under the shared-subspace model, so they cannot reveal failure of this assumption. No sensitivity analysis or misspecification experiment is reported. The concern is not that the assumption is unusual (it is standard in factor analysis) but that the central claim is specifically about recovering an interpretable decomposition, and that recovery is only as valid as the exact subspace equality across heterogeneous scales.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a PhD thesis on fusing heterogeneous data sets in systems biology. It treats two types of heterogeneity: differences in data type (binary, quantitative, count) and differences in information content (global common, local common, distinct variation). Chapter 2 reviews PCA methods for binary genomics data; Chapter 3 develops a robust logistic PCA via non-convex singular value thresholding; Chapter 4 introduces GSCA for fusing one binary and one quantitative data set; Chapter 5 proposes P-ESCA, an exponential-family simultaneous component analysis with a group concave penalty on the loading matrix to disentangle global, local common, and distinct variation in multiple mixed-type data sets. The main claims are that the MM algorithms are monotonically decreasing, have closed-form updates, that a missing-value cross-validation procedure selects the model, and that simulations and a CLL data analysis support the disentangling claim.","tokens_in":54935,"tokens_out":3099,"duration_ms":36878,"significance":"If the central claims hold, P-ESCA is a potentially useful addition to the multi-block data fusion literature, extending SCA-type decompositions to mixed data types with a structured sparsity pattern. The manuscript gives creditworthy strengths: explicit majorization inequalities, analytic updates for all parameters, monotonicity arguments, a missing-value cross-validation implementation, and comparisons with MOFA and SLIDE in the supplementary material. The simulations are described in enough detail to be reproduced, and the real-data illustration on CLL data shows an applied context. The main weakness is that the disentangling claim rests on a strong shared-subspace assumption that is never stress-tested, and the simulation evidence is generated entirely under that assumption.","major_comments":[{"comment":"The model assumes that the natural parameter matrices of all data sets lie in exactly the same column subspace spanned by one common score matrix A, i.e., Θ_l = 1μ_l^T + A B_l^T for every block l. This exact equality is load-bearing for the interpretation of zero blocks in B_l as absence of shared variation. If the true data-generating process is not exactly aligned on the linear-predictor scale, for example if a shared biological factor affects the log-odds of binary measurements and the mean of quantitative measurements through different scaling or link functions, or if block-specific rotations are present, then the fitted group concave penalty can shrink blocks to zero to compensate for the misspecification rather than to reveal the true common/distinct structure. Section 5.6 generates data exactly under this model, so the simulations cannot detect such a failure. No misspecification or sensitivity experiment is reported. I recommend adding experiments with perturbed block subspaces, block-specific rotations, or alternative link functions to assess whether the decomposition remains interpretable.","section":"§5.2.1"},{"comment":"The model selection evaluation for P-ESCA is presented only as an example of CV-error-based selection, not as a systematic evaluation of whether the selected tuning parameters recover the true sparsity pattern. Since the central output of P-ESCA is a structured sparsity pattern that determines which components are global, local common, or distinct, the manuscript should report quantitative agreement between the selected pattern and the true pattern across simulation settings (e.g., varying signal-to-noise ratio, number of blocks, and strength of local common components). This is especially important because earlier chapters show that CV-based selection can behave badly for lasso-type penalties (Chapter 3, nuclear norm) and that the SCAD penalty performs poorly; the reader needs evidence that the group concave penalty does not inherit these problems in the multi-block setting.","section":"§5.6.2"}],"minor_comments":[{"comment":"The heading 'The majorization of the penalized negative lilkelihood' contains a typo: 'lilkelihood' should be 'likelihood'.","section":"§4.3.1"},{"comment":"In the sentence beginning 'using the GCSA model's CNA loadings', 'GCSA' should be 'GSCA' to match the model name used elsewhere.","section":"§4.5.2"},{"comment":"The section title 'Evaluation matrices and model selection' should probably read 'Evaluation metrics and model selection'.","section":"§5.5"},{"comment":"In the sentence about multiple score and loading vectors, 'vcectors' should be 'vectors'.","section":"Chapter 1"},{"comment":"The thesis is composed of chapters that are based on previously published or preprint papers. If this is submitted as a journal article, the new contribution of Chapter 5 should be clearly delineated from the published material in Chapters 2–4, and the manuscript should be restructured into a self-contained article.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a PhD thesis rather than a standard journal article, which affects its fit with the journal scope but does not by itself determine the technical verdict. The central derivation appears sound, but the load-bearing disentangling claim needs the misspecification and model-selection experiments described in the major comments. The author should also clarify how much of the thesis material is already published, since Chapters 2–4 are explicitly based on earlier papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The bottom line: this is a PhD thesis assembled from four previously released papers by the same author. The underlying methods—especially the P-ESCA model in Chapter 5—are legitimate technical contributions with careful MM derivations, but the thesis itself offers little new material as an arXiv submission. The load-bearing assumption of a single shared column subspace across all data types is never stress-tested outside simulations that generate data exactly under that assumption.\n\nWhat is genuinely good: the algorithmic work is real. Chapter 3's logistic PCA via non-convex singular value thresholding has explicit majorizations, closed-form block updates, and a monotonicity guarantee; the latent-variable interpretation giving an SNR definition for binary simulation is a nice piece of thinking. Chapter 5's group concave penalty is a sensible extension of the SLIDE group-lasso idea, and the author is commendably candid about the SCAD penalty's poor performance and standard logistic PCA overfitting. The simulations are described in enough detail to follow.\n\nSoft spots, in proportion: the shared-subspace equality (Section 5.2.1) is the crux. If the binary and quantitative natural parameters do not live in exactly the same linear subspace, the group penalty will shrink blocks to zero to fit a misspecified model, and the resulting global/local/distinct pattern becomes an artifact. The simulations cannot reveal this because they simulate under the same model. No sensitivity analysis or robustness check is reported. That is a genuine gap, not a fatal one—the same assumption underlies JIVE, MOFA, and most factor-analytic fusion methods—but for a method whose whole point is an interpretable decomposition, it should be tested. Also: no shipped code, single-run key comparisons in places, and the iClusterPlus baseline is not tuned with the same care as the GSCA model, so the 'better than iClusterPlus' claim is plausible but not fully nailed down.\n\nWho is this for? A method developer or systems biologist who wants a unified read of Song's contributions to binary PCA and mixed-type data fusion. As a thesis it is coherent and well-organized. As a journal submission it would be redundant with the four underlying papers.\n\nFor peer review: if this landed on a desk as a new paper, I would send it to an expert referee—the P-ESCA method deserves scrutiny, and the synthesis is clear—but I would tell the referee the novelty lies in the synthesis, not the results.","headline":"A coherent PhD thesis compiling four of the author's own methods papers; the P-ESCA chapter is the most original, but the shared-subspace assumption is never stress-tested.","tokens_in":55506,"tokens_out":2953,"would_cite":true,"duration_ms":28896,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A penalized exponential-family simultaneous component analysis model can fuse binary, count, and quantitative data sets and separate their global-common, local-common, and distinct variation through a group concave penalty on the loading…","keywords":["data fusion","exponential family PCA","simultaneous component analysis","common and distinct variation","group concave penalty","structured sparsity","mixed data types","majorization-minimization"],"falsifier":"Simulate two data sets with a known shared axis plus a data-set-specific axis, but generate the shared axis with different scaling or through a nonlinear transformation across a binary log-odds scale and a continuous mean scale; if P-ESCA cannot recover the specified common and distinct loading blocks under that generating process, the common-subspace assumption is the point of failure.","tokens_in":54455,"feed_emoji":"🧬","tokens_out":7548,"duration_ms":62292,"temperature":0.7,"pith_summary":"The thesis develops P-ESCA, a simultaneous component analysis model for multiple data sets that may be quantitative, binary, or count-valued. Its central claim is that imposing a nearly unbiased group concave penalty on the loading matrix induces a structured sparsity pattern that disentangles variation shared by all data sets, variation shared by subsets, and variation specific to each data set. The paper argues this works because exponential-family likelihoods place each data type on its natural scale while the shared score matrix couples the data sets, and it backs the claim with simulation studies and a chronic lymphocytic leukaemia example. If true, researchers can integrate heterogeneous omics measurements without first forcing them onto a common quantitative scale, and can read off which signals are global, local, or data-set-specific from the loading pattern alone.","feed_headline":"One penalty separates shared and unique signals in fused omics data","feed_subtitle":"P-ESCA fuses binary, count, and quantitative data while carving out global, local, and distinct components.","key_machinery":"The central object is the loading matrix of an exponential-family simultaneous component analysis model, organized in blocks $B_l$ for each data set. The mechanism is a group concave penalty on these blocks: it shrinks the $\\ell^2$ norm of each block nonlinearly, driving small blocks to exactly zero while leaving large blocks nearly unchanged. Those zero blocks create the structured sparsity pattern that identifies which latent components are global, local, or distinct. The fitting machinery is a Majorization-Minimization algorithm that majorizes the negative log-likelihood to a quadratic and the concave penalty to a weighted sum of singular values, yielding closed-form updates that decrease the objective monotonically.","core_discovery":"In the ESCA model, the natural parameter matrices of all data sets are assumed to lie in the same column subspace, spanned by a common score matrix $A$, with data-set-specific loadings $B_l$. P-ESCA adds a group concave penalty that acts on the $\\ell^2$ norm of each loading block, so whole blocks are shrunk to zero in a nearly unbiased way. The zero blocks mark components that are absent from a given data set, and that pattern is exactly how global, local-common, and distinct variation are read off. A Majorization-Minimization algorithm with closed-form updates fits the penalized likelihood, and a missing-value cross-validation procedure selects the tuning parameters. The paper reports that the method recovers simulated common-distinct structures across mixed data types and illustrates the decomposition on chronic lymphocytic leukaemia data.","pith_inferences":["If the common-subspace assumption holds only approximately, the group penalty may still recover the dominant shared axes, but weaker subset-shared components could be forced to zero; a residual-correlation diagnostic between fitted data sets would test this.","The same block-penalty construction could be combined with element-wise concave penalties to discover partially shared variables, not only wholly shared components, as the thesis itself outlines as a future direction.","Because each data type is modeled on its natural parameter scale, the estimated common score matrix could plausibly serve as an input to downstream supervised analyses, giving a principled way to use heterogeneous measurements for prediction."],"forward_implications":["Multiple omics data sets of different measurement scales can be fused in one model without dichotomizing or re-scaling away their distinct natures.","The loading matrix itself becomes the interpretable map of shared versus unique biology: a zero block means a component is not present in that data set.","Mixed binary, count, and quantitative data can be integrated for exploratory analysis, with model selection by cross-validation rather than by an ad hoc rank guess.","Because the algorithm is monotone and has closed-form updates, the approach is computationally tractable in the high-dimensional, low-sample-size settings typical of genomics."],"supporting_citations":[{"why":"supplies the exponential-family principal component analysis construction whose likelihood ESCA extends to simultaneous component analysis.","marker":"[68]"},{"why":"defines the simultaneous component analysis model that P-ESCA generalizes to mixed data types.","marker":"[65]"},{"why":"introduces the group-lasso structured sparsity approach on loadings that P-ESCA replaces with a nearly unbiased concave penalty.","marker":"[15]"},{"why":"provides the Bayesian group factor analysis formulation that motivates the common and distinct component decomposition.","marker":"[87]"},{"why":"offers the mixed-type Bayesian factor analysis benchmark that P-ESCA is compared against.","marker":"[89]"},{"why":"supplies the generalized double Pareto concave penalty used to achieve nearly unbiased group shrinkage.","marker":"[27]"},{"why":"provides the chronic lymphocytic leukaemia data sets used for the real-data illustration.","marker":"[92]"}],"fun_headline_variants":["Group penalty zeroes absent components in fused omics","P-ESCA carves global and local signals from mixed data","Penalty marks absent data blocks in multi-omics fusion","Concave penalty reveals distinct patterns in mixed omics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the natural parameter matrices of all data sets share exactly the same column subspace, spanned by one common score matrix $A$; if the shared signal is not linearly aligned across measurement scales, or some shared components are weak, the decomposition can be misspecified and the penalty may force the wrong zero blocks.","fun_headline_variants_meta":{"raw":{"variants":["Group penalty zeroes absent components in fused omics","P-ESCA carves global and local signals from mixed data","Penalty marks absent data blocks in multi-omics fusion","Concave penalty reveals distinct patterns in mixed omics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000696,"raw_usage":{"total_tokens":3099,"prompt_tokens":847,"completion_tokens":2252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":2185}},"tokens_in":463,"tokens_out":2252,"duration_ms":14810,"temperature":1.0,"reasoning_tokens":2185,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:29:40.113390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate two data sets with a known shared axis plus a data-set-specific axis, but generate the shared axis with different scaling or through a nonlinear transformation across a binary log-odds scale and a continuous mean scale; if P-ESCA cannot recover the specified common and distinct loading blocks under that generating process, the common-subspace assumption is the point of failure.","supporting_citations":[{"cited_title":"A generalization of princi- pal components analysis to the exponential family,","cited_arxiv_id":null,"evidence_quote":"supplies the exponential-family principal component analysis construction whose likelihood ESCA extends to simultaneous component analysis."},{"cited_title":"A structured overview of simultaneous component based data integration,","cited_arxiv_id":null,"evidence_quote":"defines the simultaneous component analysis model that P-ESCA generalizes to mixed data types."},{"cited_title":"Group factor analy- sis,","cited_arxiv_id":null,"evidence_quote":"provides the Bayesian group factor analysis formulation that motivates the common and distinct component decomposition."},{"cited_title":"Multi-Omics Factor Analysis—a framework for unsupervised integration of multi-omics data sets,","cited_arxiv_id":null,"evidence_quote":"offers the mixed-type Bayesian factor analysis benchmark that P-ESCA is compared against."},{"cited_title":"Generalized double Pareto shrink- age,","cited_arxiv_id":null,"evidence_quote":"supplies the generalized double Pareto concave penalty used to achieve nearly unbiased group shrinkage."},{"cited_title":"Drug-perturbation- based stratification of blood cancer,","cited_arxiv_id":null,"evidence_quote":"provides the chronic lymphocytic leukaemia data sets used for the real-data illustration."}],"review_version":1}