Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Scalable Bayesian structure learning of directed acyclic graphs via Laplace approximation, with an application to breast cancer gene expression networks

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read A closed-form Laplace score lets Bayesian DAG learning use heavier-tailed Normal–Gamma priors and still contract on the true skeleton at clinical sample sizes.

desk verdict Solid closed-form non-conjugate DAG score with honest diagnostics; the n-dependent α is a real but acknowledged soft spot, not a collapse of the math. read the letter →

arxiv 2607.10222 v1 pith:LXS5764G submitted 2026-07-11 stat.ME

classification stat.ME MSC 62H2262F1562P10
keywords BayesianstructurelearningdirectedacyclicgraphsLaplaceapproximationNormal–Gammapriorgeneralisedinverse-GaussianMarkovchainMonteCarloDAG-probitcausaldiscovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Bayesian structure learning of directed acyclic graphs has long been limited by two problems: the space of DAGs grows super-exponentially, and flexible non-conjugate priors make the node-wise marginal likelihood intractable. This paper shows that under a Normal–Gamma prior on the modified Cholesky factors of the precision matrix, the node-marginal integral is exactly a generalised inverse-Gaussian integral whose value is a modified Bessel function. The leading large-argument asymptotic of that Bessel function supplies a closed-form scoring function that can be plugged into a Metropolis–Hastings sampler over graphs. The same analysis yields an exact generalised-inverse-Gaussian posterior for each conditional variance, so those parameters can be sampled without Metropolis steps. The authors prove that the Laplace approximation is accurate at the per-sample rate o_P(1) and that the induced posterior over DAGs contracts on the true skeleton at rate sqrt(log q / n). In simulations at sample sizes typical of clinical cohorts the method improves on the conjugate Bayesian baseline and on PC, GES, NOTEARS and DAGMA; on the Wisconsin breast-cancer data the DAG-probit extension predicts malignancy with cross-validated ROC-AUC 0.94 from a sparse set of nuclear-morphometry features.

What carries the argument

The Laplace-approximated node-marginal score (Theorem 1, Eq. 8): the large-argument asymptotic of the modified Bessel function that arises as the exact GIG integral of the Normal–Gamma node model; it supplies both the practical MH score and the analytic control used for the contraction theorem.

What would settle it

On synthetic Gaussian DAGs with known ground truth, at moderate n and q, replace the n-dependent shape by a small fixed shape and check whether the Laplace score still recovers higher skeleton F1/MCC than the conjugate Normal–Inverse-Gamma baseline and the continuous-optimisation benchmarks; if the advantage disappears or the posterior fails to contract, the central practical claim fails.

Watch

Extended reading notes

Core claim

Under the non-conjugate Normal–Gamma prior on the modified Cholesky parameterisation, the node-marginal likelihood is of generalised inverse-Gaussian form and therefore equals a modified Bessel function of the second kind; its leading large-argument asymptotic is a closed-form Laplace score that can be used for Metropolis–Hastings structure learning, and the induced posterior contracts on the true skeleton at the near-optimal rate sqrt(log q / n).

Load-bearing premise

The default prior shape is allowed to grow with sample size so that the Bessel argument stays in the large-argument regime; if a fixed small shape is required, the approximation quality and Occam balance that the theory and defaults rely on are no longer guaranteed by construction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper develops a Laplace-approximated node-marginal score for Bayesian DAG structure learning under a non-conjugate Normal–Gamma prior on the modified Cholesky factors of the precision matrix. It shows that the exact node-marginal is a modified Bessel function of the second kind arising from a generalised inverse-Gaussian integral, derives the leading large-argument Laplace form (Theorem 1, Eq. 8), identifies exact GIG posteriors for conditional variances, and embeds the score in a Metropolis–Hastings sampler with a DAG-probit extension for binary outcomes. Asymptotic results give per-sample Laplace accuracy (Theorem 3) and skeleton posterior contraction at rate √(log q / n) (Theorem 4). Simulations compare against CPNIG, BGe, PC, GES, NOTEARS, and DAGMA; applications cover the Sachs protein network and a DAG-probit analysis of Wisconsin Diagnostic Breast Cancer (WDBC) nuclear morphometry features, reporting CV ROC-AUC 0.94.

Significance. If the technical claims hold, the paper supplies a practical closed-form score for a heavier-tailed non-conjugate coefficient prior that has previously required expensive numerical integration, together with uniform asymptotic control and exact GIG sampling of variances. Strengths include honest documentation of score non-equivalence (Table 1), mid-bin miscalibration (Table 11), local-move ESS collapse at larger q (Tables 12–14), mixed q=40 outcomes, and fair hyperparameter sweeps for continuous baselines (Table 5). The micro-benchmark establishing that exact Bessel evaluation is stable and essentially free (Table 2) is useful and correctly demotes Laplace to an analytic rather than computational device. The contribution is complementary to continuous-optimisation DAG learners and to order/partition MCMC, and the score is stated to transfer unchanged into those samplers.

major comments (3)
  1. Section 3 sets the default shape α = n + q − max_j p_j − 2 so that α^D_j ≍ n and the Bessel order ν_j stays commensurate with z_j ≍ n^{1/2}. Theorems 1 and 3 and the Occam balance of Eq. (8) rely on this large-argument regime; Appendix A.3 explicitly uses the n-dependent rule when controlling ν_j relative to z_j. The manuscript acknowledges that n-dependent α is a device for approximation quality rather than a fixed belief prior, and states only that results are “qualitatively unchanged” under fixed α = q+1. No table or figure reports F1/MCC/SHD, relative Laplace error, or posterior edge probabilities under fixed α at the same (q,n) cells as Tables 3–7 and 6. Because the reported gains over CPNIG/NOTEARS/DAGMA and the finite-sample behaviour of the score are load-bearing for the central claim, a quantitative fixed-α sensitivity analysis (same metrics and cells) is needed before the pract
  2. The title and keywords frame an application to “breast cancer gene expression networks,” and Section 7 (limitations) refers to “highlighted genes” and candidate biomarkers. The actual application in Section 6.2 is the WDBC nuclear morphometry features (radius, perimeter, concave points, etc.) under a DAG-probit model, not gene-expression data; Sachs (Section 6.1) is protein signalling. References 30–31 and 39–42 on breast-cancer gene signatures appear unused in the reported analyses. This is an internal inconsistency that misstates the empirical contribution and should be corrected throughout (title, abstract framing if needed, keywords, Discussion), with any gene-expression analysis either restored with full results or removed cleanly.
  3. Abstract and Introduction claim improvement over PC, GES, NOTEARS, and DAGMA “at sample sizes typical of clinical cohorts.” Tables 3–4 and 5 show this is only partly true: at q=20 nCPNG leads, but at q=40 NOTEARS attains higher F1 in several cells when it converges, and DAGMA is best at q=40, n=200. The paper reports these mixed outcomes in the body, which is commendable, but the abstract’s unqualified superiority claim should be aligned with the nuanced ranking (e.g., strongest at moderate n and smaller q; competitive rather than uniformly superior at q=40).
minor comments (6)
  1. Section 3.2 and Table 2 correctly recommend the exact Bessel score as default; ensure all reported simulation and real-data results state explicitly whether exact or Laplace was used (the text says exact for reported results, but Algorithm 1 still writes the Laplace form (8)).
  2. Table 1 and the score-equivalence discussion are valuable; consider moving a one-sentence pointer into the abstract or introduction so readers expecting BGe-style equivalence are not surprised.
  3. Figure 1’s plateau is well explained as sampler mixing at fixed budget; adding ESS or acceptance rate on the same panel would make the diagnosis self-contained.
  4. Proposition 2’s rate degradation by I_Φ is useful; a short numerical illustration at the WDBC prevalence (already computed as ≈0.58) could sit next to Table 18.
  5. Minor typography: missing spaces in several compound words in the front matter (e.g., “applicationtobreastcancer”, “node-marginallikelihood”); standardise “Normal–Gamma” vs “Normal-Gamma” hyphenation.
  6. Data availability is clear; if code for the nCPNG score and Algorithm 1 will be released, state the repository in the final version to support the reproducibility claims made for baselines.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the GIG/Bessel node-marginal and Laplace score follow from the stated Normal–Gamma prior and Gaussian likelihood; asymptotics and simulations are against external ground truth. The n-dependent α is a non-standard prior device, not a definitional loop.

full rationale

The derivation chain is self-contained. The node-marginal integral (Eq. 7) is obtained by integrating the Gaussian likelihood (1) against the Normal–Gamma prior (4)–(5); it is exactly GIG and equals a modified Bessel K_ν. Theorem 1 retains the leading large-argument asymptotic of that Bessel function to produce the closed-form score (8); the relative error is the standard O(z^{-1}) remainder, not a fitted residual. Theorems 3–4 then apply uniform Laplace control, KL separation between the true DAG and competitors, and a Chernoff bound on the structural prior (3) to obtain skeleton contraction at √(log q / n) against an external true graph D_0. Simulation data are generated from independent linear-Gaussian SEMs with random coefficients; real-data evaluation uses the external Sachs consensus network and held-out WDBC folds. Self-citations (conjugate CPNIG baseline, order-MCMC literature) are comparative or orthogonal, not load-bearing uniqueness claims. The only non-standard element is the default α = n + q − max_j p_j − 2, which the authors themselves flag as making the prior a device for keeping ν_j and z_j in the large-argument regime rather than a pure fixed belief prior. That choice conditions the quality of the Laplace expansion and Occam balance, but it does not make the score equal its target by construction, nor does any reported F1/MCC/AUC reduce to a fitted input renamed as prediction. Hence circularity is absent; the n-dependent hyperparameter is a methodological caveat, not a circular step.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central claims rest on standard Gaussian DAG factorisation and Bessel asymptotics, plus modelling choices (skeleton Bernoulli prior, n-dependent α, fixed g, local MH moves, probit normalisation) that are free or domain-level rather than derived. No new physical entities are postulated; ‘nCPNG’ and ‘DAG-probit’ are named model packages built from known pieces.

free parameters (5)
  • prior scale g
    Fixed at g=2 by default; controls Occam penalty via |M_j| and Bessel argument; rate-neutral asymptotically but finite-sample sensitive.
  • shape hyperparameter α (n-dependent default)
    Default α = n + q − max_j p_j − 2 is chosen so α^D_j ≍ n and the large-argument Bessel regime holds; authors admit this is not a pure fixed prior.
  • edge-inclusion probability π
    Bernoulli skeleton prior hyperparameter; set sparsely in applications (e.g. π=0.15 on WDBC) and affects prior mass on dense graphs.
  • continuous-baseline λ and magnitude thresholds
    NOTEARS/DAGMA ℓ1 penalty and post-hoc thresholds are swept for fair comparison; not free for the Bayesian method but affect claimed superiority margins.
  • median-probability threshold 0.5
    MAP/median graph reporting uses edge probability >0.5 plus acyclicity projection; standard but still a reporting choice.
assumptions (6)
  • domain assumption Observations follow a linear-Gaussian SEM with precision Markov w.r.t. a DAG (modified Cholesky form).
    Section 2, Eq. (1); all scores and contraction proofs assume Gaussian node-conditionals.
  • standard math Large-argument expansion of K_ν(z) is valid uniformly for the relevant (ν,z) regime with z ≍ n^{1/2}.
    Theorem 1 and Appendix A.1; underpins Laplace score error O(n^{-1/2}).
  • domain assumption True DAG has bounded maximum in-degree M < ∞ (or slowly growing under Remark 1).
    Theorem 4; needed for uniform Laplace control and KL separation.
  • ad hoc to paper Skeleton prior is orientation-invariant Bernoulli on undirected edges; orientation comes from likelihood/probit only.
    Eq. (3) and surrounding discussion; authors note equivalence-class mass imbalance as a genuine feature.
  • domain assumption Probit threshold θ_0 has flat improper prior; posterior propriety needs ≥1 success and failure.
    Proposition 1; used for DAG-probit applications.
  • standard math Local single-edge MH proposals with Hastings correction target the correct structure posterior.
    Algorithm 1; standard MCMC theory, but mixing is empirical.
invented entities (2)
  • nCPNG prior (non-conjugate Normal–Gamma on modified Cholesky) independent evidence
    purpose: Heavier-tailed coefficient marginal than conjugate Normal–Inverse-Gamma while retaining a tractable node score via GIG/Bessel/Laplace.
    Named packaging of known Normal–Gamma ingredients on a standard Cholesky DAG parameterisation; independent evidence is the simulation diagnostic attributing gains to coefficient tails (Section 5.5).
  • DAG-probit model independent evidence
    purpose: Couple latent Gaussian network to binary clinical outcome via thresholding X_1.
    Standard probit data-augmentation idea applied to the DAG; not a new physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Bayesian structure learning of directed acyclic graphs via Laplace approximation, with an application to breast cancer gene expression networks." pith.science (2026). https://pith.science/paper/LXS5764G

@misc{pith2026260710222,
  author       = {Pith},
  title        = {Pith review of: Scalable Bayesian structure learning of directed acyclic graphs via Laplace approximation, with an application to breast cancer gene expression networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXS5764G}},
  note         = {Machine review of arXiv:2607.10222}
}
abstract

Structure learning of directed acyclic graphs (DAGs) from observational data is a foundational task in causal discovery and is widely used to infer regulatory networks from medical and genomic measurements. The Bayesian formulation quantifies model uncertainty and admits prior biological knowledge, but its practical use has been hampered by the super-exponential growth of the DAG space and by the intractability of the node-marginal likelihood under flexible, non-conjugate priors. Existing closed-form solutions are largely confined to the conjugate Normal--Inverse-Gamma prior. We develop a Laplace-approximated Bayesian scoring function for the non-conjugate Normal--Gamma prior on the modified Cholesky parameterisation of the precision matrix, embed it in a Metropolis--Hastings sampler over DAGs, and couple the latent Gaussian network to a binary clinical outcome through a probit link. We show that the node-marginal integral is of generalised inverse-Gaussian form, so that its exact value is a modified Bessel function of the second kind and the proposed scoring function is its leading large-argument asymptotic; the posterior of each conditional variance is likewise generalised inverse-Gaussian and is sampled exactly. In simulation, the proposed prior improves on the conjugate baseline and on the PC, greedy-equivalence-search, NOTEARS, and DAGMA benchmarks at sample sizes typical of clinical cohorts. On two real datasets, the Sachs protein-signalling network, scored against its validated consensus graph, and the Wisconsin Diagnostic Breast Cancer data, the method recovers known structure and, through the DAG-probit extension, predicts malignancy from nuclear morphometry with a cross-validated ROC-AUC of $0.94$ using a sparse, interpretable set of direct predictors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 4 linked inside Pith

  1. [1]

    Oxford: Oxford University Press; 1996

    Lauritzen SL.Graphical Models. Oxford: Oxford University Press; 1996

  2. [2]

    Cambridge: Cambridge University Press; 2000

    Pearl J.Causality: Models, Reasoning, and Inference. Cambridge: Cambridge University Press; 2000

  3. [3]

    Being Bayesian about network structure.Mach Learn

    Friedman N, Koller D. Being Bayesian about network structure.Mach Learn. 2003;50:95-125

  4. [4]

    Causal protein-signaling networks derived from multiparameter single-cell data.Science

    Sachs K, Perez O, Pe’er D, Lauffenburger DA, Nolan GP. Causal protein-signaling networks derived from multiparameter single-cell data.Science. 2005;308(5721):523-529

  5. [5]

    In:BiomedicalImageProcessingandBiomedicalVisualization.Proc.SPIE1905.1993:861-870

    StreetWN,WolbergWH,MangasarianOL.Nuclearfeatureextractionforbreasttumordiagnosis. In:BiomedicalImageProcessingandBiomedicalVisualization.Proc.SPIE1905.1993:861-870

  6. [6]

    New York: Academic Press; 1973:239-273

    RobinsonRW.Counting labeledacyclicdigraphs.In:Harary F,ed.NewDirectionsin theTheory of Graphs. New York: Academic Press; 1973:239-273

  7. [7]

    A Bayesian method for the induction of probabilistic networks from data.Mach Learn

    Cooper GF, Herskovits E. A Bayesian method for the induction of probabilistic networks from data.Mach Learn. 1992;9:309-347

  8. [8]

    Parameter priors for directed acyclic graphical models and the charac- terization of several probability distributions.Ann Statist

    Geiger D, Heckerman D. Parameter priors for directed acyclic graphical models and the charac- terization of several probability distributions.Ann Statist. 2002;30(5):1412-1440

Show all 41 references
  1. [9]

    Learning Markov equivalence classes of directed acyclic graphs: an objective Bayes approach.Stat Med

    Castelletti F, Consonni G, Della Vedova ML, Peluso S. Learning Markov equivalence classes of directed acyclic graphs: an objective Bayes approach.Stat Med. 2020;39(30):4745-4766

  2. [11]

    Posterior graph selection and estimation consistency for high- dimensional Bayesian DAG models.Ann Statist

    Cao X, Khare K, Ghosh M. Posterior graph selection and estimation consistency for high- dimensional Bayesian DAG models.Ann Statist. 2019;47(1):319-348

  3. [12]

    BCDAG: an R package for Bayesian structure and causal learning of Gaussian DAGs

    Castelletti F, Mascaro A. BCDAG: an R package for Bayesian structure and causal learning of Gaussian DAGs. arXiv:2201.12003. 2022

  4. [13]

    Wishart distributions: Advances in theory with Bayesian applicationJournal of Multivariate Analysis

    Bekker A, van Niekerk J, Arashi M. Wishart distributions: Advances in theory with Bayesian applicationJournal of Multivariate Analysis. 2017;155:272-283

  5. [14]

    DAGs with NO TEARS: continuous optimization for structure learning

    Zheng X, Aragam B, Ravikumar P, Xing EP. DAGs with NO TEARS: continuous optimization for structure learning. In:Advances in Neural Information Processing Systems; 2018

  6. [15]

    In:Advances in Neural Information Processing Systems; 2020:17943-17954

    NgI,GhassamiA,ZhangK.OntheroleofsparsityandDAGconstraintsforlearninglinearDAGs. In:Advances in Neural Information Processing Systems; 2020:17943-17954

  7. [16]

    In:Advances in Neural Information Processing Systems; 2022

    BelloK,AragamB,RavikumarP.DAGMA:learningDAGsviaM-matricesandalog-determinant acyclicity characterization. In:Advances in Neural Information Processing Systems; 2022

  8. [17]

    DAGs with no curl: an efficient DAG structure learning approach

    Yu Y, Gao T, Yin N, Ji Q. DAGs with no curl: an efficient DAG structure learning approach. In: Proceedings of the 38th International Conference on Machine Learning; 2021:12156-12166

  9. [18]

    Truncated matrix power iteration for differentiable DAG learning

    Zhang Z, Ng I, Gong M, Liu Y, Gong C, Bello K. Truncated matrix power iteration for differentiable DAG learning. In:Advances in Neural Information Processing Systems; 2022

  10. [19]

    TriOpt: a scalable algorithm for linear causal discovery

    Joy RA, Zheleva E. TriOpt: a scalable algorithm for linear causal discovery. arXiv:2605.17465. 2026

  11. [20]

    Learning directed acyclic graphs via bootstrap aggregating

    Wang R, Peng J. Learning directed acyclic graphs via bootstrap aggregating. arXiv:1406.2098. 2014

  12. [21]

    DAGBagM: learning directed acyclic graphs of mixed vari- ables with an application to identify protein biomarkers for treatment response in ovarian cancer

    Chowdhury S, Wang R, Yu Q, et al. DAGBagM: learning directed acyclic graphs of mixed vari- ables with an application to identify protein biomarkers for treatment response in ovarian cancer. BMC Bioinformatics. 2022;23:321

  13. [22]

    Optimal structure identification with greedy search.J Mach Learn Res

    Chickering DM. Optimal structure identification with greedy search.J Mach Learn Res. 2002;3:507-554

  14. [23]

    A transformational characterization of equivalent Bayesian network structures

    Chickering DM. A transformational characterization of equivalent Bayesian network structures. In:ProceedingsoftheEleventhConferenceonUncertaintyinArtificialIntelligence(UAI).Morgan Kaufmann; 1995:87-98

  15. [24]

    Estimating high-dimensional directed acyclic graphs with the PC- algorithm.J Mach Learn Res

    Kalisch M, Bühlmann P. Estimating high-dimensional directed acyclic graphs with the PC- algorithm.J Mach Learn Res. 2007;8:613-636

  16. [25]

    Identifiability of Gaussian structural equation models with equal error variances.Biometrika

    Peters J, Bühlmann P. Identifiability of Gaussian structural equation models with equal error variances.Biometrika. 2014;101(1):219-228

  17. [26]

    2009;71(2):319-392

    RueH,MartinoS,ChopinN.ApproximateBayesianinferenceforlatentGaussianmodelsbyusing integrated nested Laplace approximations.J R Stat Soc Series B. 2009;71(2):319-392

  18. [27]

    NAZARIET AL 31

    SpokoinyV.Dimension-freeboundsfortheLaplaceapproximation.BayesianAnal.2025;20(1):1- 28. NAZARIET AL 31

  19. [28]

    Lecture Notes in Statistics, vol

    Jørgensen B.Statistical Properties of the Generalized Inverse Gaussian Distribution. Lecture Notes in Statistics, vol. 9. New York: Springer; 1982

  20. [29]

    Generating generalized inverse Gaussian random variates.Stat Comput

    Hörmann W, Leydold J. Generating generalized inverse Gaussian random variates.Stat Comput. 2014;24(4):547-557

  21. [30]

    Strong time dependence of the 76-gene prognostic signature for node-negativebreastcancerpatientsintheTRANSBIGmulticenterindependentvalidationseries

    Desmedt C, Piette F, Loi S, et al. Strong time dependence of the 76-gene prognostic signature for node-negativebreastcancerpatientsintheTRANSBIGmulticenterindependentvalidationseries. Clin Cancer Res. 2007;13(11):3207-3214

  22. [31]

    Gene expression profiling in breast cancer: understanding the molecularbasisofhistologicgradetoimproveprognosis.JNatlCancerInst.2006;98(4):262-272

    Sotiriou C, Wirapati P, Loi S, et al. Gene expression profiling in breast cancer: understanding the molecularbasisofhistologicgradetoimproveprognosis.JNatlCancerInst.2006;98(4):262-272

  23. [32]

    BeingBayesian aboutnetwork structure:aBayesian approachto structure discovery in Bayesian networks.Mach Learn

    Friedman N,KollerD. BeingBayesian aboutnetwork structure:aBayesian approachto structure discovery in Bayesian networks.Mach Learn. 2003;50:95-125

  24. [33]

    2004;5:549-573

    KoivistoM,SoodK.ExactBayesianstructurediscoveryinBayesiannetworks.JMachLearnRes. 2004;5:549-573

  25. [34]

    Partition MCMC for inference on acyclic digraphs.J Am Stat Assoc

    Kuipers J, Moffa G. Partition MCMC for inference on acyclic digraphs.J Am Stat Assoc. 2017;112(517):282-299

  26. [35]

    Addendum on the scoring of Gaussian directed acyclic graphical models.Ann Statist

    Kuipers J, Moffa G, Heckerman D. Addendum on the scoring of Gaussian directed acyclic graphical models.Ann Statist. 2014;42(4):1689-1691

  27. [36]

    2019;47(6):3413-3437

    LeeK,LeeJ,LinL.Minimaxposteriorconvergenceratesandmodelselectionconsistencyinhigh- dimensional DAG models based on sparse Cholesky factors.Ann Statist. 2019;47(6):3413-3437

  28. [37]

    Lasso meets horseshoe: a survey.Statist Sci

    Bhadra A, Datta J, Polson NG, Willard B. Lasso meets horseshoe: a survey.Statist Sci. 2019;34(3):405-427

  29. [38]

    arXiv:1109.4371

    Ben-DavidE,LiT,MassamH,RajaratnamB.High-dimensionalBayesianinferenceforGaussian directed acyclic graph models. arXiv:1109.4371. 2015

  30. [39]

    Interleukin-8 in breast cancer progression.J Interferon Cytokine Res

    Todorović-Raković N, Milovanović J. Interleukin-8 in breast cancer progression.J Interferon Cytokine Res. 2013;33(10):563-570

  31. [40]

    Recent advances reveal IL-8 signaling as a potential key to targeting breast cancer stem cells.Breast Cancer Res

    Singh JK, Simões BM, Howell SJ, Farnie G, Clarke RB. Recent advances reveal IL-8 signaling as a potential key to targeting breast cancer stem cells.Breast Cancer Res. 2013;15(4):210

  32. [41]

    2014;8(7):1278-1289

    CallariM,MusellaV,DiBuduoE,etal.Subtype-dependentprognosticrelevanceofaninterferon- induced pathway metagene in node-negative breast cancer.Mol Oncol. 2014;8(7):1278-1289

  33. [42]

    Prognostic characterization of OAS1/OAS2/OAS3/OASL in breast cancer.BMC Cancer

    Zhang Y, Yu C. Prognostic characterization of OAS1/OAS2/OAS3/OASL in breast cancer.BMC Cancer. 2020;20:575

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.