Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Biological Pathway Guided Gene Selection Through Collaborative Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read BioMARL: a pathway-guided multi-agent reinforcement learning framework that selects 100-gene cancer signatures with higher held-out AUC and stronger biological enrichment than eight standard feature-selection methods.

desk verdict A genuinely novel two-stage gene-selection framework that earns serious review, but the central MARL learning signal rests on an unvalidated surrogate and the 'significantly improves' claim is not backed by error bars or significance tests. read the letter →

arxiv 2505.24155 v1 pith:JOQJQP3O submitted 2025-05-30 cs.LG

Ehtesamul Azim , Dongjie Wang , Tae Hyun Hwang , Yanjie Fu , Wei Zhang This is my paper
classification cs.LG
keywords geneselectionmulti-agentreinforcementlearningbiologicalpathwaysKEGGfeatureTCGAbiomarkerdiscoverygraphneuralnetworks
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

This paper claims that gene selection for cancer outcome prediction can be improved by explicitly using biological pathway knowledge inside a reinforcement-learning selector. It proposes BioMARL, a two-stage framework: first, a statistical pre-filter whose gene scores are boosted by the predictive performance of classifiers trained on the gene's KEGG pathways; second, a multi-agent DQN in which each gene is an agent that learns whether to be selected, guided by a reward combining estimated prediction impact, gene centrality within pathways, and pathway coverage. On six TCGA datasets, the paper reports that BioMARL achieves the highest mean AUC in five and yields gene sets with more significantly enriched Gene Ontology terms than eight baseline feature selectors. The core wager is that collaborative, pathway-aware decision-making produces signatures that are simultaneously more predictive and more biologically interpretable.

What carries the argument

The load-bearing machinery is the two-stage BioMARL pipeline. Stage one computes an integrative score per gene: a weighted mix of chi-squared, random-forest, and SVM importance scores, with method weights proportional to validation performance, multiplied by a pathway bonus $1 + \beta \log(1 + \bar{S}_p(g))$ derived from mean classifier performance on the KEGG pathways containing $g$; genes above $\mu + 2\sigma$ pass to stage two. Stage two treats each surviving gene as an agent in a DQN whose state is a GNN embedding with edge weights $E_{ij} = \rho C_{ij} + (1-\rho) J_{ij}$ blending expression correlation and pathway-membership Jaccard similarity. Each agent's reward is $r_i = \omega r_{\mathrm{base}} + \xi \Delta\phi_i + \zeta \Delta\psi_i$, where $r_{\mathrm{base}}$ comes from an ensemble meta-learner estimating per-gene performance impact, $\Delta\phi_i$ is differential pathway centrality, and $\Delta\psi_i$ is differential pathway coverage. A centralized critic provides a global baseline, and a synergy-matrix shared memory biases action selection toward previously successful gene pairs.

What would settle it

On a held-out TCGA dataset, take a random sample of single-gene flips in the pre-filtered gene pool, compute the true change in held-out AUC when each gene is added or removed, and correlate those true changes with the meta-learner's estimates $\hat{\Delta R}$; if the Spearman correlation is below about $0.3$, the reward driving the agents is dominated by estimation error. A second, simpler check is to include the pre-filtered top-$k$ genes alone as a ninth baseline; if BioMARL's AUC advantage vanishes against it, the multi-agent stage is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that BioMARL, a two-stage pipeline combining a KEGG-pathway-boosted statistical pre-filter with a multi-agent deep Q-network selector, produces 100-gene signatures that are simultaneously more predictive and more biologically interpretable than those of eight established feature-selection methods. Concretely, it reports the highest mean AUC in five of six TCGA classification tasks (ER, HER2, PR, and triple-negative breast cancer status, plus survival-stratified lung adenocarcinoma and ovarian cancer), lower variance than most baselines, and gene sets with more significantly enriched Gene Ontology terms; selected genes are further shown to be literature-documented cancer genes and to stratify patients by survival in log-rank tests.

Load-bearing premise

The agents' reward is computed from a surrogate model's estimate of how much adding or removing each gene changes prediction accuracy, and the method assumes that estimate is unbiased even though it is never validated against the true single-gene performance change.

Editorial extensions

If this is right

  • Held-out AUC gains on five of six datasets mean the 100-gene signatures generalize beyond the training cohort, so candidate biomarkers produced this way are plausible starting points for wet-lab validation.
  • Because the selected genes are enriched for known GO terms and map to KEGG pathways, the signatures come with a built-in mechanistic story, which is what a clinician or biologist needs to trust a biomarker panel.
  • The ablation results imply each of the three collaborative mechanisms (personalized reward, centralized critic, and shared memory) earns its keep; removing any one lowers mean AUC on the four datasets tested.
  • The downstream-model-agnostic result suggests the selected genes carry signal that is not specific to a random-forest evaluator, so the signatures are likely to transfer to other classifiers used in practice.
  • The significant Kaplan-Meier separation for genes like C1GALT1C1 in breast cancer and PIK3CD in lung cancer indicates the selected markers are associated with overall survival, not just class labels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A test the paper does not run: if the pre-filtered top-100 genes alone were added as a baseline, and it matched BioMARL's AUC, then the multi-agent stage contributes interpretability but little predictive power; the architecture's authors should report this comparison.
  • The reward surrogate assumption could be retired: on the small pre-filtered pool (roughly a few hundred genes), true leave-one-out classifier evaluations are cheap enough to replace the meta-learner, giving exact rewards and removing the paper's weakest premise.
  • The synergy-matrix shared memory is a portable idea: any feature-selection setting with known pairwise structure (co-expression, protein interaction, drug-target networks) could plug this collaboration bias into an RL selector.
  • The enrichment counting at $p \le 0.01$ without multiple-testing correction is likely to overstate biological interpretability; an FDR-adjusted comparison across the eight baselines is the natural follow-up.
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

4 major / 5 minor

Summary. The paper proposes BioMARL, a two-stage gene selection framework that first applies a pathway-guided statistical pre-filtering step using KEGG pathway information and then models genes as collaborative reinforcement learning agents. Each agent learns a binary selection action using a DQN, with state representations built from a GNN over correlation/pathway edges, a reward that combines a meta-learner's predicted performance impact, gene centrality, and pathway coverage, and additional mechanisms including a centralized critic and shared memory. The authors evaluate BioMARL on six TCGA-derived classification tasks and report that it outperforms eight baseline feature selection methods in five of six datasets, with additional analyses of ablation variants, hold-out proportions, GO enrichment, downstream classifier robustness, case-study genes, survival analysis, and a real-world gastric cancer dataset.

Significance. The proposed integration of biological pathway knowledge into a multi-agent reinforcement learning framework for gene selection is a timely and potentially useful contribution, especially given the emphasis on HDLSS genomic data. The paper ships code and preprocessed data, which supports reproducibility. If the central claims were fully validated, the work could offer a practical tool for biomarker discovery. However, the current evidence for both predictive superiority and biological interpretability is weakened by missing statistical rigor in the main comparison, an unvalidated surrogate reward model, and an interpretability analysis that is partly self-referential because the reward explicitly maximizes the same pathway-coverage and centrality quantities that are later tested for enrichment.

major comments (4)
  1. [§4.2, Figure 2] The abstract and Section 4.2 claim that BioMARL 'significantly improves' prediction accuracy, but Figure 2 reports only average AUC over 10 runs, with no error bars, confidence intervals, or significance tests. The text also states that BioMARL exhibits 'relatively low variance in most cases,' but no variance measure is shown. To support the central predictive claim, the authors should report per-run standard deviations or confidence intervals and perform paired statistical tests (e.g., paired t-test or Wilcoxon signed-rank test) across the 10 runs for each dataset.
  2. [§3.3, 'Personal Performance Impact Estimator'; Algorithm 1 lines 28–32] The reward that drives all agent learning is computed from the meta-learner f_meta, but f_meta is never validated against true leave-one-out or retrained performance changes of the downstream Random Forest. The text says that evaluating model performance for each perturbation state 'yields labels y_t' and then computes ΔR_t = f_meta(D_t) − R_t, yet the online update buffer stores only (a_t, R_t) pairs, not per-perturbation labels, leaving the training of f_meta ambiguous. If f_meta is biased, the agents optimize a faulty reward, and the reported 5/6 AUC advantage could be driven by the pre-filtering stage rather than by the learned MARL selection. The ablation BioMARL−Rwd in Table 2 removes only the pathway-based reward components and not f_meta, so it does not isolate this issue. The authors should clarify how f_meta is trained from y_t and provide a validation of f_meta against actual performance changes, especially for rare-class datasets such as LUAD with 21 positive samples.
  3. [§4.2, 'Gene Enrichment Analysis' and §3.3, 'Reward Mechanism'] The biological interpretability claim is substantially self-referential: the reward explicitly maximizes KEGG pathway coverage and gene centrality, and then GO enrichment is used as evidence of biological relevance. Because the pathway-guided pre-filtering and the reward both favor genes with known pathway annotations, enrichment among selected genes is expected by construction. The GO analysis in Figure 5 also counts terms with p-value ≤ 0.01 without correcting for multiple testing, and no comparison is made against the pre-filtered candidate pool or against random gene sets of the same size. The authors should report enrichment relative to an appropriate null (e.g., random subsets of the pre-filtered genes) and use corrected significance thresholds or effect-size measures.
  4. [§4.2, 'Clear Separation' and 'Case Study II' / Figure 7] The survival analysis is presented as validation of biological relevance, but the log-rank p-values are computed for genes selected using the same data, without an independent test set or multiple-testing correction. The heatmap claim of 'perfect separation' is also not quantified. Since the genes were chosen partly because they separate the classes, the KM curves and p-values are likely optimistic. The authors should either perform survival analysis on held-out patients or state explicitly if the analysis is exploratory, and they should correct for the number of genes tested.
minor comments (5)
  1. [§3.2, §3.3, Algorithm 1] The notation is inconsistent: β is used for the pathway bonus scaling factor in Section 3.2, for the synergy bias weight in Algorithm 1 line 24, and for a reward weight in Algorithm 1 line 42, while the text in Section 3.3 uses (ω, ξ, ζ) for the reward weights. The authors should unify the notation throughout.
  2. [§3.3, 'Reward Mechanism'] The formula for r_base includes '+ I_t 1' in the text, but Algorithm 1 line 32 omits the global improvement term. Please clarify the exact definition and whether I_t is part of every agent's reward.
  3. [§4.2, 'Gene Enrichment Analysis'] The text states 'BioMARL demonstrated superior performance on both datasets,' but Figure 5 reports results for three datasets (BRCA(PR), OV, LUAD). Please correct the wording and the figure caption.
  4. [Table 3] The column header contains a typo ('Desccription') and the table does not compare the literature evidence for BioMARL-exclusive genes with the evidence for genes selected by baseline methods, so the case study does not by itself demonstrate that BioMARL selects more relevant genes.
  5. [Abstract and §4.2] The abstract's claim of 'significantly improves' is not supported by any significance test in the paper; please revise the wording to match the actual statistical evidence.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the interpretability result is largely a restatement of the KEGG-coverage reward; held-out AUC remains an external benchmark.

  1. fitted input called prediction [Section 3.3 'Reward Mechanism' / 'Pathway Coverage Measure'; Section 4.2 'Gene Enrichment Analysis']
    "The final reward for each gene i combines the base performance estimate with the pathway-based measures through a weighted sum: r_i = ω·r_base + ξ·Δφ_i + ζ·Δψ_i. ... Pathway Coverage Measure ... ψ_p(S) = |S∩G_p|/|G_p| ... We examined the enrichment of Gene Ontology (GO) terms [3] among the top-ranked 100 genes identified by each method ... BioMARL demonstrated superior performance."

    The biological-relevance validation is not independent of the selection objective. BioMARL's reward explicitly maximizes KEGG pathway coverage Δψ_i and gene centrality Δφ_i, so the selected genes are concentrated in KEGG-annotated pathways by construction. The subsequent GO-enrichment comparison reports that BioMARL's genes are enriched in GO terms; since GO and KEGG are overlapping biological annotation resources with strong correlation for well-studied cancer pathways, a selector rewarded for pathway coverage will tend to dominate methods that never see pathway annotations on this metric. The GO-enrichment 'improvement' is therefore largely a restatement of the reward already applied, rather than an external confirmation of biological interpretability.

full rationale

Most of the derivation chain is not circular: the multi-agent Q-learning is trained with a perturbation-based performance estimator and evaluated on held-out AUC against eight baselines, an external benchmark; the ablation removes components and measures AUC under the same hold-out protocol. No load-bearing self-citation or imported uniqueness theorem is used; prior work by the authors appears only as related work or general genomics support. The unvalidated f_meta surrogate (Section 3.3 and Algorithm 1 lines 29–32) is a correctness and robustness risk, not a circularity, because the downstream evaluation is independent of f_meta. The one genuine circular element is the interpretability/GO-enrichment evidence: the method directly rewards KEGG pathway coverage and centrality, and the paper then evaluates 'biological relevance' with GO-term enrichment, an overlapping annotation resource. This makes the biological-interpretability headline partially self-referential. Score 4 reflects partial circularity confined to the interpretability claim, while the primary accuracy claim retains independent content.

Assumptions & free parameters 8 free parameters · 5 assumptions · 3 invented entities

The central claim rests on a large set of manually chosen hyperparameters (reward weights, graph weights, critic weights, synergy bias, memory decay), on the assumption that KEGG annotations are reliable for these cancers, and on the untested accuracy of the surrogate meta-learner that generates every reward. The method introduces several internal computational constructs (synergy matrix, centralized critic, personal performance impact estimator) that have no external validation. None of these are physics-style invented entities, but they are load-bearing assumptions that the paper does not justify independently.

free parameters (8)
  • Reward weights (ω, ξ, ζ) = 0.5, 0.25, 0.25
    Balance prediction performance, gene centrality, and pathway coverage in the final reward; set by hand in Appendix 8.1 with no sensitivity analysis.
  • Pathway bonus scaling factor β = 0.2
    Controls the strength of the pathway performance adjustment in pre-filtering; chosen manually in Appendix 8.1.
  • Graph edge correlation weight ρ (α in appendix) = 0.7
    Mixes gene expression correlation with pathway Jaccard similarity to build the GNN adjacency matrix; set by hand.
  • Critic weights λa and λb = 0.7 and 0.3
    Balance local Q-learning target with the centralized critic value estimate; set by hand.
  • Synergy bias weight η (β in pseudocode) = 0.08 increasing linearly to 0.3
    Controls how strongly the shared-memory synergy matrix biases action selection; set by hand.
  • Shared memory decay factor = 0.99
    Exponential decay of collaboration success records and synergy values; set by hand.
  • DQN discount factor γ = 0.85
    Standard RL discount, but its value affects how much future reward is weighed during selection; set by hand.
  • Replay buffer capacity and exploration steps = 1700 and 3000
    Chosen in Appendix 8.1; these affect the stability and convergence of the DQN training.
assumptions (5)
  • domain assumption Pre-filtering threshold ŝ_g > μ + 2σ assumes the adjusted meta-scores are approximately normally distributed; otherwise the 2σ cutoff has no probabilistic justification.
    Section 3.2 defines the filtered set G_pre exactly as genes with adjusted scores above mean plus two standard deviations, but provides no distributional check.
  • domain assumption KEGG pathway annotations are accurate, complete, and relevant for the studied cancer types; pathway membership is a valid proxy for functional gene relationships.
    Section 3.2 and 3.3 use KEGG pathways both to adjust pre-filtering scores and to compute centrality/coverage rewards; if annotations are wrong or incomplete, all biological guidance degrades.
  • domain assumption The meta-learner f_meta accurately estimates performance changes ΔR_t for arbitrary single-gene perturbations, and these estimates are reliable enough to serve as RL rewards.
    Section 3.3 'Personal Performance Impact Estimator' and Algorithm 1 lines 28-31 use f_meta to compute the reward for every agent; no validation of f_meta accuracy is reported.
  • domain assumption The GNN state representation, with edge weight ρ=0.7, captures the gene interactions that matter for selection; the learned policies depend on this fixed graph.
    Section 3.3 'State Representation' constructs edges as a weighted combination of correlation and pathway similarity; the choice of ρ is manual and not varied.
  • standard math Standard DQN convergence assumptions hold in this multi-agent setting with a centralized critic and shared memory.
    The method relies on Q-learning convergence, target networks, and prioritized experience replay as in standard deep RL literature, without additional guarantees for the multi-agent interaction.
invented entities (3)
  • Synergy matrix M
    purpose: Records pairwise feature synergies based on past successful selections and biases action selection through a synergy term.
    The matrix is learned from training data on the same datasets; it has no external falsifiable handle and its values are not compared to any reference.
  • Centralized critic V(s)
    purpose: Provides a global state-value estimate that is mixed into DQN target values to reduce variance and coordinate agents.
    An internal model component that is trained and evaluated only within the proposed framework; no external prediction or benchmark.
  • Personal performance impact estimator f_meta
    purpose: Approximates the per-gene performance impact (ΔR_t) for reward computation without full model retraining.
    A meta-learner trained on random selection vectors from the same data; its accuracy is never validated against true leave-one-out performance changes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Biological Pathway Guided Gene Selection Through Collaborative Reinforcement Learning." pith.science (2026). https://pith.science/paper/JOQJQP3O

@misc{pith2026250524155,
  author       = {Pith},
  title        = {Pith review of: Biological Pathway Guided Gene Selection Through Collaborative Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JOQJQP3O}},
  note         = {Machine review of arXiv:2505.24155}
}
read the original abstract

Gene selection in high-dimensional genomic data is essential for understanding disease mechanisms and improving therapeutic outcomes. Traditional feature selection methods effectively identify predictive genes but often ignore complex biological pathways and regulatory networks, leading to unstable and biologically irrelevant signatures. Prior approaches, such as Lasso-based methods and statistical filtering, either focus solely on individual gene-outcome associations or fail to capture pathway-level interactions, presenting a key challenge: how to integrate biological pathway knowledge while maintaining statistical rigor in gene selection? To address this gap, we propose a novel two-stage framework that integrates statistical selection with biological pathway knowledge using multi-agent reinforcement learning (MARL). First, we introduce a pathway-guided pre-filtering strategy that leverages multiple statistical methods alongside KEGG pathway information for initial dimensionality reduction. Next, for refined selection, we model genes as collaborative agents in a MARL framework, where each agent optimizes both predictive power and biological relevance. Our framework incorporates pathway knowledge through Graph Neural Network-based state representations, a reward mechanism combining prediction performance with gene centrality and pathway coverage, and collaborative learning strategies using shared memory and a centralized critic component. Extensive experiments on multiple gene expression datasets demonstrate that our approach significantly improves both prediction accuracy and biological interpretability compared to traditional methods.

Figures

Figures reproduced from arXiv: 2505.24155 by the authors.

Figure 1
Figure 1. Framework Overview. BioMARL consists of two key parts: (a) pathway-guided meta-selection combining multiple statistical methods [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall performance comparison of BioMARL with eight state-of-the-art baselines: Best performing baseline highlighted in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The influence of different hold-out percentages in terms [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Enrichment analysis of selected genes for breast, ovarian [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Heatmap of the expression profiles of 100 marker genes [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Robustness evaluation of BioMARL across classification [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LLM-ML Teaming: Integrated Symbolic Decoding and Gradient Search for Valid and Stable Generative Feature Transformation

    cs.LG 2025-06 conditional novelty 4.0 of 10

    A product-of-experts decoder that blends a fine-tuned LLM's token probabilities with a gradient-searched sequence decoder produces more valid and stable feature transformations than either alone.

Reference graph

Works this paper leans on

61 extracted references · 56 canonical work pages · cited by 1 Pith paper

  1. [1]

    Khandakar Tanvir Ahmed, Jiao Sun, Sze Cheng, Jeongsik Yong, and Wei Zhang

    work page

  2. [2]

    Fuad Al Abir, SM Shovan, Md Al Mehedi Hasan, Abu Sayeed, and Jungpil Shin

    work page

  3. [3]

    Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. 2000. Gene ontology: tool for the unification of biology. Nature genetics 25, 1 (2000), 25–29

    work page 2000

  4. [4]

    Molecular Omics 18, 7 (2022), 652–661

    Biomarker identification by reversing the learning mechanism of an autoen- coder and recursive feature elimination. Molecular Omics 18, 7 (2022), 652–661

    work page 2022

  5. [5]

    Oluleye H Babatunde, Leisa Armstrong, Jinsong Leng, and Dean Diepeveen. 2014. A genetic algorithm-based feature selection. (2014)

    work page 2014

  6. [6]

    Ehtesamul Azim, Dongjie Wang, Kunpeng Liu, Wei Zhang, and Yanjie Fu. 2024. Feature Interaction Aware Automated Data Representation Transformation. In Proceedings of the 2024 SIAM International Conference on Data Mining (SDM) . SIAM, 878–886

    work page 2024

  7. [7]

    Sudipto Baul, Khandakar Tanvir Ahmed, Joseph Filipek, and Wei Zhang. 2022. omicsGAT: Graph attention network for cancer subtype analyses. International Journal of Molecular Sciences 23, 18 (2022), 10220

    work page 2022

  8. [8]

    Nolan Bard, Jakob N Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, et al. 2020. The hanabi challenge: A new frontier for ai research. Artificial Intelligence 280 (2020), 103216

    work page 2020

Show all 61 references
  1. [9]

    Brigham & Women’s Hospital and others. 2012. Comprehensive molecular portraits of human breast tumours. Nature 490, 7418 (2012), 61–70

    2012

  2. [10]

    William W Bivin, Oleksandr Yergiyev, Mark L Bunker, Jan F Silverman, and Uma Krishnamurti. 2017. GRB7 expression and correlation with HER2 amplifica- tion in invasive breast carcinoma. Applied Immunohistochemistry & Molecular Morphology 25, 8 (2017), 553–558

    2017

  3. [11]

    Cancer Genome Atlas Research Network. 2014. Comprehensive molecular profil- ing of lung adenocarcinoma. Nature 511, 7511 (2014), 543

    2014

  4. [12]

    Cancer Genome Atlas Research Network. 2013. The cancer genome atlas pan- cancer analysis project. Nat. Genet 45, 10 (2013), 1113–1120

    2013

  5. [13]

    Han-Yu Chuang, Eunjung Lee, Yu-Tsueng Liu, Doheon Lee, and Trey Ideker

  6. [14]

    Cancer Genome Atlas Research Network and others. 2011. Integrated genomic analyses of ovarian carcinoma. Nature 474, 7353 (2011), 609

    2011

  7. [15]

    Aileen I Fernandez, Matthew Liu, Andrew Bellizzi, Jane Brock, Oluwole Fadare, Krisztina Hanley, Malini Harigopal, Julie M Jorns, M Gabriela Kuba, Amy Ly, et al. 2022. Examination of low ERBB2 protein expression in breast cancer tissue. JAMA oncology 8, 4 (2022), 607–610

    2022

  8. [16]

    Thomas Gabel and Martin Riedmiller. 2007. On a successful application of multi- agent reinforcement learning to operations research benchmarks. In 2007 IEEE international symposium on approximate dynamic programming and reinforcement learning. IEEE, 68–75

    2007

  9. [17]

    Marc-André Dittrich and Silas Fohlmeister. 2020. Cooperative multi-agent system for production control using reinforcement learning. CIRP Annals 69, 1 (2020), 389–392

    2020

  10. [18]

    Pilar García-Díaz, Isabel Sánchez-Berriel, Juan A Martínez-Rojas, and Ana M Diez-Pascual. 2020. Unsupervised feature selection algorithm for multiclass cancer classification of gene expression RNA-Seq data. Genomics 112, 2 (2020), 1916–1925

    2020

  11. [19]

    Nanxu Gong, Wangyang Ying, Dongjie Wang, and Yanjie Fu. 2025. Neuro- symbolic embedding for short and effective feature selection via autoregressive generation. ACM Transactions on Intelligent Systems and Technology 16, 2 (2025), 1–21. KDD ’25, August 3–7, 2025, Toronto, ON, C...

    2025

  12. [20]

    Consolata Gakii, Paul O Mireji, and Richard Rimiru. 2022. Graph based feature selection for reduction of dimensionality in next-generation rna sequencing datasets. Algorithms 15, 1 (2022), 21

    2022

  13. [21]

    Ryan J Hartmaier, Lee A Albacker, Juliann Chmielecki, Mark Bailey, Jie He, Michael E Goldberg, Shakti Ramkissoon, James Suh, Julia A Elvin, Samuel Chiac- chia, et al. 2017. High-throughput genomic profiling of adult solid tumors reveals novel insights into cancer pathogenesis....

    2017

  14. [22]

    Amin Hashemi, Mohammad Bagher Dowlatshahi, and Hossein Nezamabadi-pour

  15. [23]

    Isabelle Guyon, Jason Weston, Stephen Barnhill, and Vladimir Vapnik. 2002. Gene selection for cancer classification using support vector machines. Machine learning 46 (2002), 389–422

    2002

  16. [24]

    Minoru Kanehisa and Susumu Goto. 2000. KEGG: kyoto encyclopedia of genes and genomes. Nucleic acids research 28, 1 (2000), 27–30

    2000

  17. [25]

    Tzu-Jen Kao, Chung-Che Wu, Nam Nhut Phan, Yen-Hsi Liu, Hoang Dang Khoa Ta, Gangga Anuraga, Yung-Fu Wu, Kuen-Haur Lee, Jian-Ying Chuang, and Chih- Yang Wang. 2021. Prognoses and genomic analyses of proteasome 26S subunit, ATPase (PSMC) family genes in clinical breast cancer. Ag...

    2021

  18. [26]

    International Journal of Machine Learning and Cybernetics 13, 1 (2022), 49–69

    Ensemble of feature selection algorithms: a multi-criteria decision-making approach. International Journal of Machine Learning and Cybernetics 13, 1 (2022), 49–69

    2022

  19. [27]

    Nichola Johnson, Paolo De Ieso, Gabriele Migliorini, Nick Orr, Peter Broderick, Daniel Catovsky, Athena Matakidou, Timothy Eisen, Christy Goldsmith, Frank Dudbridge, et al. 2016. Cytochrome P450 allele CYP3A7* 1C associates with adverse outcomes in chronic lymphocytic leukemia...

    2016

  20. [28]

    Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. 2021. Las- sonet: A neural network with feature sparsity. Journal of Machine Learning Research 22, 127 (2021), 1–29

    2021

  21. [29]

    Yuanyuan Li, Kai Kang, Juno M Krahn, Nicole Croutwater, Kevin Lee, David M Umbach, and Leping Li. 2017. A comprehensive genomic pan-cancer classification using The Cancer Genome Atlas gene expression data. BMC genomics 18 (2017), 1–13

    2017

  22. [30]

    Yunchuan Kong and Tianwei Yu. 2018. A graph-embedded deep feedforward network for disease outcome classification and feature selection using gene expression data. Bioinformatics 34, 21 (2018), 3727–3737

    2018

  23. [31]

    Patrick Lebok, Melike Öztürk, Uwe Heilenkötter, Fritz Jaenicke, Volkmar Müller, Peter Paluchowski, Stefan Geist, Christian Wilke, Eicke Burandt, Annette Lebeau, et al. 2016. High levels of class III 𝛽-tubulin expression are associated with aggressive tumor features in breast c...

    2016

  24. [32]

    Catalina Lodillinsky, Laetitia Fuhrmann, Marie Irondelle, Olena Pylypenko, Xiao- Yan Li, Hélène Bonsang-Kitzis, Fabien Reyal, Sophie Vacher, Claire Calmel, Olivier De Wever, et al. 2021. Metastasis-suppressor NME1 controls the invasive switch of breast cancer by regulating MT1...

    2021

  25. [33]

    Andrew Mulcahy, Stephanie Rennane, Daniel Schwam, Reid Dickerson, Lawrence Baker, and Kanaka Shetty. 2025. Use of Clinical Trial Characteristics to Estimate Costs of New Drug Development. JAMA Network Open 8, 1 (2025), e2453275– e2453275

    2025

  26. [34]

    Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li. 2019. Automating feature subspace exploration via multi-agent reinforcement learning. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 207–215

    2019

  27. [35]

    Kunpeng Liu, Haibo Huang, Wei Zhang, Ahmad Hariri, Yanjie Fu, and Kien Hua

  28. [36]

    Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang. 2017. Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games. arXiv preprint arXiv:1703.10069 (2017)

    2017 arXiv

  29. [37]

    Daniel R Rhodes, Bushra Ateeq, Qi Cao, Scott A Tomlins, Rohit Mehra, Bharathi Laxman, Shanker Kalyana-Sundaram, Robert J Lonigro, Beth E Helgeson, Ma- haveer S Bhojani, et al. 2009. AGTR1 overexpression defines a subset of breast cancer and confers sensitivity to losartan, an ...

    2009

  30. [38]

    Borja Seijo-Pardo, Iago Porto-Díaz, Verónica Bolón-Canedo, and Amparo Alonso- Betanzos. 2017. Ensemble feature selection: Homogeneous and heterogeneous approaches. Knowledge-Based Systems 118 (2017), 124–139

    2017

  31. [39]

    Sungsoo Park, Bonggun Shin, Won Sang Shim, Yoonjung Choi, Kilsoo Kang, and Keunsoo Kang. 2019. Wx: a neural network-based feature selection algorithm for transcriptomic data. Scientific reports 9, 1 (2019), 10500

    2019

  32. [40]

    Hanchuan Peng, Fuhui Long, and Chris Ding. 2005. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min- redundancy. IEEE Transactions on pattern analysis and machine intelligence 27, 8 (2005), 1226–1238

    2005

  33. [41]

    Robert Tibshirani. 1996. Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society Series B: Statistical Methodology 58, 1 (1996), 267– 288

    1996

  34. [42]

    Laura J Van’t Veer, Hongyue Dai, Marc J Van De Vijver, Yudong D He, Augusti- nus AM Hart, Mao Mao, Hans L Peterse, Karin Van Der Kooy, Matthew J Marton, Anke T Witteveen, et al. 2002. Gene expression profiling predicts clinical outcome of breast cancer. nature 415, 6871 (2002)...

    2002

  35. [43]

    Meng Wang, Guangxin Zhang, Yue Zhang, Xuelian Cui, Shuaibin Wang, Song Gao, Yicun Wang, Ying Liu, Jeeyoo H Bae, Wei-Hsiung Yang, et al. 2020. Fibrinogen alpha chain knockout promotes tumor growth and metastasis through integrin– AKT signaling pathway in lung cancer. Molecular ...

    2020

  36. [44]

    Maria Amélia Lopes Silva, Sérgio Ricardo de Souza, Marcone Jamilson Freitas Souza, and Ana Lúcia C Bazzan. 2019. A reinforcement learning-based multi- agent framework applied for solving routing and scheduling problems. Expert Systems with Applications 131 (2019), 148–171

    2019

  37. [45]

    Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Kor- jus, Juhan Aru, Jaan Aru, and Raul Vicente. 2017. Multiagent cooperation and competition with deep reinforcement learning. PloS one 12, 4 (2017), e0172395

    2017

  38. [46]

    Meng Xiao, Dongjie Wang, Min Wu, Pengfei Wang, Yuanchun Zhou, and Yanjie Fu. 2023. Beyond discrete selection: Continuous embedding space optimization for generative feature selection. arXiv preprint arXiv:2302.13221 (2023)

    2023 arXiv

  39. [47]

    Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun Wang

  40. [48]

    Yiming Yang, Jan O Pedersen, et al. 1997. A comparative study on feature selection in text categorization. In icml, Vol. 97. Citeseer, 35

    1997

  41. [49]

    Xinyuan Wang, Dongjie Wang, Wangyang Ying, Rui Xie, Haifeng Chen, and Yanjie Fu. 2024. Knockoff-Guided Feature Selection via A Single Pre-trained Reinforced Agent. arXiv preprint arXiv:2403.04015 (2024)

    2024 arXiv

  42. [50]

    Yixin Wang, Jan GM Klijn, Yi Zhang, Anieta M Sieuwerts, Maxime P Look, Fei Yang, Dmitri Talantov, Mieke Timmermans, Marion E Meijer-van Gelder, Jack Yu, et al. 2005. Gene-expression profiles to predict distant metastasis of lymph- node-negative primary breast cancer. The Lance...

    2005

  43. [51]

    Ke Zhang, Fang He, Zhengchao Zhang, Xi Lin, and Meng Li. 2020. Multi-vehicle routing problems with soft time windows: A multi-agent reinforcement learning approach. Transportation Research Part C: Emerging Technologies 121 (2020), 102861

    2020

  44. [52]

    Wei Zhang, Jeremy Chien, Jeongsik Yong, and Rui Kuang. 2017. Network-based machine learning and graph theory algorithms for precision oncology. NPJ precision oncology 1, 1 (2017), 25

    2017

  45. [53]

    Wei Zhang, Nicholas Johnson, Baolin Wu, and Rui Kuang. 2012. Signed network propagation for detecting differential gene expressions and DNA copy number variations. In Proceedings of the ACM conference on bioinformatics, computational biology and biomedicine. 337–344

    2012

  46. [54]

    Zishuang Zhang and Zhi-Ping Liu. 2021. Robust biomarker discovery for hepa- tocellular carcinoma from high-throughput data by multiple feature selection methods. BMC medical genomics 14 (2021), 1–12

    2021

  47. [55]

    Wangyang Ying, Dongjie Wang, Haifeng Chen, and Yanjie Fu. 2024. Feature selection as deep sequential generative learning. ACM Transactions on Knowledge Discovery from Data 18, 9 (2024), 1–21

    2024

  48. [56]

    Wangyang Ying, Dongjie Wang, Xuanming Hu, Ji Qiu, Jin Park, and Yanjie Fu. 2024. Revolutionizing Biomarker Discovery: Leveraging Generative AI for Bio-Knowledge-Embedded Continuous Space Exploration. In Proceedings of the 33rd ACM International Conference on Information and Kn...

    2024

  49. [61]

    Xiaosa Zhao, Kunpeng Liu, Wei Fan, Lu Jiang, Xiaowei Zhao, Minghao Yin, and Yanjie Fu. 2020. Simplifying reinforced feature selection via restructured choice strategy of single agent. In 2020 IEEE International conference on data mining (ICDM). IEEE, 871–880. 8 Appendix 8.1 Im...

    2020

  50. [2007]

    Molecular systems biology 3, 1 (2007), 140

    Network-based classification of breast cancer metastasis. Molecular systems biology 3, 1 (2007), 140

    2007

  51. [2018]

    In International conference on machine learning

    Mean field multi-agent reinforcement learning. In International conference on machine learning. PMLR, 5571–5580

  52. [2021]

    In Proceedings of the 2021 SIAM International Conference on Data Mining (SDM)

    Multi-armed bandit based feature selection. In Proceedings of the 2021 SIAM International Conference on Data Mining (SDM) . SIAM, 316–323

    2021

  53. [2022]

    Bioinfor- matics 38, 1 (2022), 179–186

    Multi-omics data integration by generative adversarial network. Bioinfor- matics 38, 1 (2022), 179–186

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.