Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Atherosclerosis through Hierarchical Explainable Neural Network Analysis

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ATHENA joins a clinical patient-similarity graph with patient-specific protein networks to classify subclinical atherosclerosis and to split each imaging subtype into two molecular clusters.

desk verdict A sensible hierarchical GNN architecture applied to a real cohort, but two under-specified evaluation steps could let the outcome label leak into both feature selection and graph construction, so the headline gains are not yet trustworthy. read the letter →

arxiv 2507.07373 v2 pith:5FARXATS submitted 2025-07-10 cs.LG cs.AI

Irsyad Adam , Steven Swee , Erika Yilin , Ethan Ji , William Speier , Dean Wang , Alex Bui , Wei Wang
show 2 more authors
Karol Watson Peipei Ping
This is my paper
classification cs.LGcs.AI
keywords hierarchicalgraphneuralnetworksubclinicalatherosclerosispatientsimilarityprotein-proteininteractiontranscriptomicssubtypingexplainableAIdeeplearning
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

The paper claims that a two-level graph, called ATHENA, can classify subclinical atherosclerosis better than either clinical-feature similarity or molecular networks used alone. On the outer level, patients are nodes connected by clinical similarity; on the inner level, each patient's transcriptomics is mapped onto a protein–protein interaction network whose learned embedding becomes that patient's node feature. The paper reports consistent gains across three GNN backbones, with AUC up to 0.843 and F1 up to 0.839, and it reports that explainability-driven subgraph clustering splits each imaging-defined subtype into two molecularly distinct patient groups. The value of the claim, if true, is that a single architecture can capture both cohort-wide clinical patterns and patient-specific molecular mechanisms, enabling earlier risk stratification and biologically grounded patient subtyping.

What carries the argument

The central object is the two-level hierarchical graph: an outer cohort graph whose nodes are patients and whose weighted edges come from a Gaussian (RBF) kernel on clinical features followed by k-nearest-neighbor sparsification, and whose node features are graph-level embeddings produced by an inner GNN encoder over each patient's transcriptomics-injected protein–protein interaction network. The inner encoder compresses each patient's molecular network to one vector, the outer GNN then propagates those vectors across clinically similar patients, and the final patient representation is classified by atherosclerotic phenotype. This two-level message passing is what carries the claim: the paper's ablations attribute the gain to the combination of the two networks, not to either network alone, and the GNNExplainer-derived subgraphs provide the molecular interaction patterns that are subsequently clustered into patient subtypes.

What would settle it

Run ATHENA with the outcome labels stripped from the clinical feature vector and with gene selection nested inside each training fold, then compare AUC and F1 to the reported values; if the gap to the ablation baselines shrinks, the reported hierarchy gains are partly label leakage. Independently, permute the outer patient-similarity edges and re-measure; if performance is unchanged, the clinical graph contributes nothing.

Watch

Extended reading notes

Core claim

ATHENA's core claim is that embedding patient-specific molecular interaction networks inside a clinical patient-similarity graph produces a representation that is better for subclinical atherosclerosis classification than either network alone, and that this representation supports discovery of molecular heterogeneity within clinical subtypes. On the PESA cohort (391 patients, whole-blood transcriptomics), ATHENA reports mean AUC 0.805–0.812 and macro F1 0.718–0.725 across GCN, GAT, and GraphSage, with ablation-average AUC improvements of 0.057–0.065 and F1 improvements of 0.041–0.042 over the patient-similarity-only setting. On the arterial-tissue validation cohort (104 patients), the hierarchical model reaches AUC 0.837–0.843 and macro F1 0.832–0.839. The paper further claims that applying GNNExplainer to the trained model and clustering the extracted molecular subgraphs yields two clusters within each imaging-defined subtype of the PESA cohort and two clusters in the validation cohort, with cluster signatures dominated by inflammatory versus structural and remodeling pathways.

Load-bearing premise

The reported gains are only trustworthy if no outcome label leaks into the model: the clinical features used to build patient similarity must not include the PESA or CAC score, and gene selection must be redone inside each cross-validation fold rather than once on the whole dataset.

Editorial extensions

If this is right

  • Subclinical atherosclerosis classification improves across all three evaluated GNN backbones when the hierarchical representation is used, with AUC up to 0.843 and F1 up to 0.839 on the validation cohort.
  • On the PESA cohort, per-subtype AUC remains in the 0.795–0.835 range for Generalized, Intermediate, and Focal atherosclerosis, indicating the gain is not concentrated in one phenotype.
  • The ablation shows that using the patient similarity network alone or the PPI network alone is worse than their combination, so the hierarchy itself, not just deep learning, is claimed to drive the improvement.
  • Within each imaging-defined atherosclerosis subtype, ATHENA separates patients into two clusters with distinct molecular pathway signatures, suggesting that imaging-identical patients can have different underlying biology.
  • The same two-cluster structure appears in the independent arterial-tissue cohort, supporting the claim that the discovered molecular heterogeneity is not specific to one tissue or dataset.

Reading between the lines

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

  • If the two-cluster structure is followed over time in a longitudinal cohort, it could be tested whether the inflammatory cluster precedes the structural-remodeling cluster or vice versa; the present cross-sectional design cannot order them.
  • The two-level design is transportable: any disease with per-patient molecular profiles and clinical covariates could be represented the same way, making this a general subtyping template rather than an atherosclerosis-specific model.
  • A direct attack on the mechanism would be to shuffle the outer graph's edges while keeping inner PPI encoders and labels fixed; if accuracy does not fall, the cohort-level clinical graph is not the active ingredient in the reported gains.
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 / 5 minor

Summary. The paper introduces ATHENA, a hierarchical graph neural network that combines patient-specific STRING PPI networks, whose node features are transcriptomic expression values, with a patient similarity graph built from clinical feature vectors via an RBF kernel and k-nearest-neighbor edges. Each patient node in the cohort graph is initialized with the graph-level embedding of that patient's PPI network, and a second GNN propagates information across clinically similar patients for classification of subclinical atherosclerosis subtypes. The authors report consistent AUC and F1 improvements over linear, MLP, and graph-only baselines on the PESA cohort and a secondary arterial-tissue cohort, and they use GNNExplainer-derived subnetworks plus t-SNE to propose two molecular clusters within each imaging-defined subtype. The core equations are standard GNN operations and are internally consistent.

Significance. If the reported results are valid, the hierarchical representation is a meaningful contribution to multi-modal biomedical graph learning, since it jointly models molecular interaction topology and cohort-level clinical similarity. The paper has notable strengths: evaluation across three GNN backbones, an ablation study separating the contribution of each network level, and validation on an external cohort. However, the central numerical claims depend on the absence of label leakage, and the manuscript does not establish two necessary protocol conditions: that supervised ANOVA/Bonferroni feature selection is nested inside cross-validation, and that the clinical features used to construct the patient similarity graph exclude the outcome label. Because either violation would directly inflate the reported AUC/F1 gains and the cluster separation, the current evidence does not yet support the paper's central claims.

major comments (3)
  1. [§3.1 and §4.1] The transcriptomic preprocessing in §3.1 applies iterative one-way ANOVA with a Bonferroni-adjusted p-value threshold of 0.01 and states that the selected features 'establish a baseline for interpreting differential gene expression across the four PESA scores.' This is supervised feature selection driven by the outcome labels. Section 4.1 describes 5-fold nested cross-validation for model training but never states that this ANOVA/Bonferroni selection is recomputed inside each training fold. If the full 391-patient cohort is used to select the 645 genes before the outer CV splits, the held-out labels influence the retained gene set, and all downstream AUC/F1 numbers in Tables 2 and 3 are optimistically biased. The authors must either confirm that feature selection is nested within each training fold and describe the procedure, or re-run the experiments with properly nested selection.
  2. [§3.2] The clinical feature vector c_i is defined as an m-dimensional vector, and the RBF patient similarity kernel is built from it, but the components of c_i are never enumerated. If c_i includes the PESA score, CAC score, or any imaging-derived subtype, then the patient similarity graph H directly transmits label information through the GCN/GAT/GraphSage message passing in Eq. (3), which could produce exactly the uniform performance gains reported in Tables 2 and 3 and the two clusters in §4.4. The authors must list all clinical features used to build the similarity graph and explicitly state that outcome-related variables are excluded, or the comparison against non-hierarchical baselines is invalid.
  3. [§4.4] The claim that ATHENA discovers two molecularly distinct patient clusters within each imaging-defined subtype rests solely on t-SNE visualizations of GNNExplainer-derived subnetworks. No cluster-quality metrics (e.g., silhouette score, stability across folds), no statistical test for differential gene expression between the clusters, and no external validation of the proposed biomarker pathways are provided. Since the subnetworks are extracted from the model's own predictions, the apparent separation could reflect artifacts of the explainability method or of t-SNE rather than genuine biological subtypes. Quantitative cluster evaluation and, ideally, validation in independent data are needed to support the subtyping claim.
minor comments (5)
  1. [§4.1] The metric 'Macro F1' is used but never defined; the authors should clarify that it is the unweighted mean of per-class F1 scores and explain how the four PESA phenotypes are aggregated in the reported numbers.
  2. [§4.1] The text says 'area under the receiver operator curve'; the correct term is 'receiver operating characteristic curve.'
  3. [§3.3 / Eq. (3)] The cohort-level GNN in Eq. (3) uses the same notation for AGGREGATE and the same layer index as the PPI encoder in Eq. (1), but the two networks may use different aggregators and hidden dimensions; please distinguish the notation or specify the shared configuration.
  4. [§4.1] The grid-search parameter space is referenced as Supplementary Tables 3 and 4, but these tables are not included in the manuscript; they should be provided or a public repository citation added.
  5. [Throughout] The dataset from Sánchez-Cabo et al. is referred to both as 'Sanchez-Cabo et al.' and 'Sanchez et al.' in the text and figures; standardize the citation style.

Circularity Check

1 steps flagged · score 5.0 of 10

Outcome-supervised gene selection likely precedes cross-validation, making ATHENA's reported AUC/F1 gains a fitted-input artifact rather than an independent prediction.

  1. fitted input called prediction [Section 3.1 (Datasets and Preprocessing), Section 4.1 (Evaluation Metrics), Tables 2-3]
    "For preprocessing transcriptomics, we utilized a variance threshold of 0.1, followed by an iterative one-way analysis-of-variance (ANOVA) with a Bonferroni-adjusted p-value threshold of 0.01 to filter out transcriptomics features with minimal variability across patients. The selected transcriptomics features establish a baseline for interpreting differential gene expression across the four PESA scores."

    The gene set fed into every model is selected by testing differential expression across the four PESA outcome classes, i.e., the prediction target. Section 4.1 then describes 5-fold nested cross-validation only for model training, with no statement that this outcome-based ANOVA/Bonferroni filter is re-fit inside each training fold. If the filter is applied once to the full 391-patient cohort before the cross-validation splits (as the section ordering implies), then the held-out labels have already determined which genes survive the filter. The 'predictions' in Tables 2-3 are therefore produced by features fitted to the target, and the reported AUC/F1 improvements are partly forced by that fit rather than by genuine generalization. This is the classic fitted-input-called-prediction pattern.

full rationale

The paper's contribution is an empirical pipeline rather than a formal derivation, so most components are not circular by construction: the hierarchical graph construction, GNN message passing, and the PPI/clinical feature integration are defined independently of the target labels. The ablation study compares the same GNN backbones with and without the hierarchy, and the reported uniform gains are not logically forced by the architecture itself. The subtype clusters are post-hoc explanations of the trained model and are not presented as the model's training objective; they are exploratory and not circular in the strict sense. No load-bearing self-citation chain is present: the dataset citations are external, and the GNNExplainer / STRING citations are standard tools. The one genuine circularity-type concern is the transcriptomics feature filter in Section 3.1: it is an outcome-supervised ANOVA/Bonferroni selection, and the paper never states that this filter is nested inside each training fold. If the filter is applied to the full cohort before the 5-fold cross-validation described in Section 4.1, then the held-out PESA labels have influenced which genes are used as inputs to all models, and the reported AUC/F1 improvement is partly a fitted-input artifact rather than evidence of generalization. Because the paper's ordering of preprocessing before evaluation and its wording 'across the four PESA scores' support this reading, this is partial circularity. The patient-similarity clinical feature vector c_i is also never enumerated; however, without a quoted list of its components, whether it contains the outcome label remains a leakage risk rather than a demonstrated circular step.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper relies on standard domain assumptions about imaging labels, blood transcriptomics, and external PPI databases, which are acceptable. The critical unstated assumption is the leakage-free evaluation protocol: the paper never confirms that feature selection is confined to training folds or that the clinical features used for the patient similarity graph exclude the outcome. The central claim depends on this assumption, and the reported numbers do not otherwise justify the headline improvements.

free parameters (6)
  • RBF kernel bandwidth sigma = Not reported (tuned via grid search)
    Controls patient similarity edge weights; central to cohort graph construction and information flow.
  • kNN neighborhood size k = 50 (Sanchez-Cabo), 20 (Steenman)
    Chosen by hand; determines graph sparsity and the degree of label or feature diffusion.
  • Transcriptomics variance threshold = 0.1
    Filtering cutoff chosen without stated justification; affects which genes enter the PPI graphs.
  • ANOVA Bonferroni p-value threshold = 0.01
    Outcome-based feature selection threshold; affects all downstream results and may introduce leakage if applied outside cross-validation.
  • STRING confidence threshold = 0.7
    Edge filter for PPI networks; chosen by convention and affects the topology of every patient-specific molecular graph.
  • Model hyperparameters (layers, hidden dims, learning rate, weight decay) = Not fully reported; grid search in missing Supplementary Tables 3 and 4
    Training setup affects all model comparisons and the final performance estimates.
assumptions (5)
  • domain assumption PESA score from imaging is a valid ordinal ground truth for atherosclerosis burden.
    Labels are clinician-assigned from ultrasound and CAC; the model learns to predict these labels.
  • domain assumption Whole-blood transcriptomics reflects atherosclerotic processes in the vessel wall.
    First dataset uses peripheral blood RNA; the paper acknowledges tissue may differ.
  • domain assumption STRING high-confidence interactions (score >= 0.7) capture the true molecular interaction topology relevant to atherosclerosis.
    PPI topology is taken as ground truth from an external database without tissue or disease specificity.
  • domain assumption GNN message passing and GNNExplainer subgraphs reveal disease mechanisms rather than model artifacts.
    Used for subtype discovery; no independent validation of the clusters against orthogonal data.
  • ad hoc to paper The evaluation protocol has no label leakage: feature selection is nested in cross-validation and the similarity graph excludes the target label.
    The paper does not specify these safeguards, yet the validity of the performance claim depends on them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Atherosclerosis through Hierarchical Explainable Neural Network Analysis." pith.science (2026). https://pith.science/paper/5FARXATS

@misc{pith2026250707373,
  author       = {Pith},
  title        = {Pith review of: Atherosclerosis through Hierarchical Explainable Neural Network Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FARXATS}},
  note         = {Machine review of arXiv:2507.07373}
}
read the original abstract

In this work, we study the problem pertaining to personalized classification of subclinical atherosclerosis by developing a hierarchical graph neural network framework to leverage two characteristic modalities of a patient: clinical features within the context of the cohort, and molecular data unique to individual patients. Current graph-based methods for disease classification detect patient-specific molecular fingerprints, but lack consistency and comprehension regarding cohort-wide features, which are an essential requirement for understanding pathogenic phenotypes across diverse atherosclerotic trajectories. Furthermore, understanding patient subtypes often considers clinical feature similarity in isolation, without integration of shared pathogenic interdependencies among patients. To address these challenges, we introduce ATHENA: Atherosclerosis Through Hierarchical Explainable Neural Network Analysis, which constructs a novel hierarchical network representation through integrated modality learning; subsequently, it optimizes learned patient-specific molecular fingerprints that reflect individual omics data, enforcing consistency with cohort-wide patterns. With a primary clinical dataset of 391 patients, we demonstrate that this heterogeneous alignment of clinical features with molecular interaction patterns has significantly boosted subclinical atherosclerosis classification performance across various baselines by up to 13% in area under the receiver operating curve (AUC) and 20% in F1 score. Taken together, ATHENA enables mechanistically-informed patient subtype discovery through explainable AI (XAI)-driven subnetwork clustering; this novel integration framework strengthens personalized intervention strategies, thereby improving the prediction of atherosclerotic disease progression and management of their clinical actionable outcomes.

Figures

Figures reproduced from arXiv: 2507.07373 by the authors.

Figure 1
Figure 1. Transcriptomics-injected patient-specific PPI net [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pipeline for multi-modal hierarchical graph deep learning for patient classification. Graph encoders are applied to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Model performance against individual subtypes of atherosclerosis for Sanchez-Cabo et al. [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Model performance against cohort of atheroscle [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 48 canonical work pages

  1. [1]

    Anu Aggarwal, Courtney L Jennings, Emily Manning, and Scott J Cameron. 2023. Platelets at the vessel wall in non-thrombotic disease.Circulation research132, 6 (2023), 775–790

    work page 2023

  2. [2]

    Shohreh Amini, Marcus Saunders, Kimberly Kelley, Kamel Khalili, and Bassel E Sawaya. 2004. Interplay between HIV-1 Vpr and Sp1 modulates p21WAF1 gene expression in human astrocytes.Journal of Biological Chemistry279, 44 (2004), 46046–46056

    work page 2004

  3. [3]

    Bo Bai, Yan Liang, Cheng Xu, Mary YK Lee, Aimin Xu, Donghai Wu, Paul M Vanhoutte, and Yu Wang. 2012. Cyclin-dependent kinase 5–mediated hyperphos- phorylation of sirtuin-1 contributes to the development of endothelial senescence and atherosclerosis.Circulation126, 6 (2012), 729–740

    work page 2012

  4. [4]

    Fotios Barkas, Yusuf Ziya Sener, Pelin Arabacilar Golforoush, Azin Kheirkhah, Elena Rodriguez-Sanchez, Jan Novak, Maria Apellaniz-Ruiz, Ralph Kwame Akyea, Vanessa Bianconi, Alexandr Ceasovschih, et al . 2024. Advancements in risk stratification and management strategies in primary cardiovascular prevention. Atherosclerosis395 (2024), 117579

    work page 2024

  5. [5]

    Guillaume Bompard, Juliette van Dijk, Julien Cau, Yoann Lannay, Guillaume Marcellin, Aleksandra Lawera, Siem van Der Laan, and Krzysztof Rogowski. 2018. CSAP acts as a regulator of TTLL-mediated microtubule glutamylation.Cell reports25, 10 (2018), 2866–2877

    work page 2018

  6. [6]

    Xin Guo, Sohsuke Yamada, Akihide Tanimoto, Yan Ding, Ke-Yong Wang, Shohei Shimajiri, Yoshitaka Murata, Satoshi Kimura, Takashi Tasaki, Atsunori Nabeshima, et al. 2012. Overexpression of peroxiredoxin 4 attenuates atheroscle- rosis in apolipoprotein E knockout mice.Antioxidants & redox signaling17, 10 (2012), 1362–1375

    work page 2012

  7. [7]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

  8. [8]

    Göran K Hansson. 2005. Inflammation, atherosclerosis, and coronary artery disease.New England journal of medicine352, 16 (2005), 1685–1695

    work page 2005

Show all 52 references
  1. [9]

    Yuchen Jing, Bai Gao, Zhiyang Han, Lifang Xia, and Shijie Xin. 2021. The protective effect of HOXA5 on carotid atherosclerosis occurs by modulating the vascular smooth muscle cell phenotype.Molecular and Cellular Endocrinology 534 (2021), 111366

    2021

  2. [10]

    Muhammad Raheel Khan, Zunaib Maqsood Haider, Jawad Hussain, Farhan Hameed Malik, Irsa Talib, and Saad Abdullah. 2024. Comprehen- sive Analysis of Cardiovascular Diseases: Symptoms, Diagnosis, and AI Innovations.Bioengineering11, 12 (2024), 1239

    2024

  3. [11]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

    2016 arXiv

  4. [12]

    Kazuo Kitagawa, Masayasu Matsumoto, Tsutomu Sasaki, Hiroyuki Hashimoto, Keisuke Kuwabara, Toshiho Ohtsuki, and Masatsugu Hori. 2002. Involvement of ICAM-1 in the progression of atherosclerosis in APOE-knockout mice.Atheroscle- rosis160, 2 (2002), 305–310. ATHENA: Atheroscleros...

    2002

  5. [13]

    Zhiqiang Li, Yifan Fan, Jing Liu, Yan Li, Chongmin Huan, Hai H Bui, Ming-Shang Kuo, Tae-Sik Park, Guoqing Cao, and Xian-Cheng Jiang. 2012. Impact of sphin- gomyelin synthase 1 deficiency on sphingolipid metabolism and atherosclerosis in mice.Arteriosclerosis, thrombosis, and v...

    2012

  6. [14]

    Peter Libby. 2012. Inflammation in atherosclerosis.Arteriosclerosis, thrombosis, and vascular biology32, 9 (2012), 2045–2051

    2012

  7. [15]

    Yude Lou, Peng Hui Li, Xiao Qi Liu, Tian Xiang Wang, Yi Lan Liu, Chen Chen Chen, and Kun Ling Ma. 2024. ITGAM-mediated macrophages contribute to basement membrane damage in diabetic nephropathy and atherosclerosis.BMC nephrology25, 1 (2024), 72

    2024

  8. [16]

    Guangming Luo, Junwon Kim, and Kiwon Song. 2014. The C-terminal domains of human neurofibromin and its budding yeast homologs Ira1 and Ira2 regulate the metaphase to anaphase transition.Cell Cycle13, 17 (2014), 2780–2789

    2014

  9. [17]

    Stephanie Maiwald, Peter-Paul Zwetsloot, Suthesh Sivapalaratnam, and Geesje M Dallinga-Thie. 2013. Monocyte gene expression and coronary artery disease. Current Opinion in Clinical Nutrition & Metabolic Care16, 4 (2013), 411–417

    2013

  10. [18]

    Seth S Martin, Aaron W Aday, Norrina B Allen, Zaid I Almarzooq, Cheryl AM Anderson, Pankaj Arora, Christy L Avery, Carissa M Baker-Smith, Nisha Bansal, Andrea Z Beaton, et al. 2025. 2025 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Hear...

    2025

  11. [19]

    Daniel H O’leary, Joseph F Polak, Richard A Kronmal, Teri A Manolio, Gregory L Burke, and Sidney K Wolfson Jr. 1999. Carotid-artery intima and media thickness as a risk factor for myocardial infarction and stroke in older adults.New England Journal of Medicine340, 1 (1999), 14–22

    1999

  12. [20]

    Shraddha Pai, Shirley Hui, Ruth Isserlin, Muhammad A Shah, Hussam Kaka, and Gary D Bader. 2019. netDx: interpretable patient classification using integrated patient similarity networks.Molecular systems biology15, 3 (2019), e8497

    2019

  13. [21]

    Nikolaos Papageorgiou, Alexandros Briasoulis, Emmanuel Androulakis, and Dimitris Tousoulis. 2017. Imaging subclinical atherosclerosis: where do we stand?Current cardiology reviews13, 1 (2017), 47–55

    2017

  14. [22]

    Sanne AE Peters, Hester M den Ruijter, Michiel L Bots, and Karel GM Moons

  15. [23]

    Bastian Pfeifer, Anna Saranti, and Andreas Holzinger. 2022. GNN-SubNet: disease subnetwork detection with explainable graph neural networks.Bioinformatics 38, Supplement_2 (2022), ii120–ii126

    2022

  16. [24]

    Oona Rainio, Jarmo Teuho, and Riku Klén. 2024. Evaluation metrics and statistical tests for machine learning.Scientific Reports14, 1 (2024), 6086

    2024

  17. [25]

    Nishant Rajadhyaksha and Aarushi Chitkara. 2023. Graph contrastive learning for multi-omics data.arXiv preprint arXiv:2301.02242(2023)

    2023 arXiv

  18. [26]

    Russell Ross. 1999. Atherosclerosis–An inflammatory disease N Engl J Med.1999 115 (1999), 10

    1999

  19. [27]

    Fátima Sánchez-Cabo, Valentín Fuster, Juan Carlos Silla-Castro, Gema González, Erika Lorenzo-Vivas, Rebeca Alvarez, Sergio Callejas, Alberto Benguría, Eduardo Gil, Estefanía Núñez, et al . 2023. Subclinical atherosclerosis and accelerated epigenetic age mediated by inflammatio...

    2023

  20. [28]

    Roman Schulte-Sasse, Stefan Budach, Denes Hnisz, and Annalisa Marsico. 2021. Integration of multiomics data with graph convolutional networks to identify new cancer genes and their associated molecular mechanisms.Nature Machine Intelligence3, 6 (2021), 513–526

    2021

  21. [29]

    Clay F Semenkovich et al . 2006. Insulin resistance and atherosclerosis.The Journal of clinical investigation116, 7 (2006), 1813–1822

    2006

  22. [30]

    Marja Steenman, Olivier Espitia, Blandine Maurel, Beatrice Guyomarch, Marie- Françoise Heymann, Marc-Antoine Pistorius, Benjamin Ory, Dominique Hey- mann, Rémi Houlgatte, Yann Gouëffic, et al . 2018. Identification of genomic differences among peripheral arterial beds in ather...

    2018

  23. [31]

    Damian Szklarczyk, Katerina Nastou, Mikaela Koutrouli, Rebecca Kirsch, Farrokh Mehryary, Radja Hachilif, Dewei Hu, Matteo E Peluso, Qingyao Huang, Tao Fang, et al. 2025. The STRING database in 2025: protein networks with directionality of regulation.Nucleic Acids Research53, D...

    2025

  24. [32]

    Songqing Tang, Taoyong Chen, Zhou Yu, Xuhui Zhu, Mingjin Yang, Bin Xie, Nan Li, Xuetao Cao, and Jianli Wang. 2014. RasGRP3 limits Toll-like receptor-triggered inflammatory response in macrophages by activating Rap1 small GTPase.Nature communications5, 1 (2014), 4657

    2014

  25. [33]

    Siyi Tang, Amara Tariq, Jared A Dunnmon, Umesh Sharma, Praneetha Elugunti, Daniel L Rubin, Bhavik N Patel, and Imon Banerjee. 2023. Predicting 30-day all-cause hospital readmission using multimodal spatiotemporal graph neural networks.IEEE Journal of Biomedical and Health Info...

    2023

  26. [34]

    Jaclyn N Taroni, Peter C Grayson, Qiwen Hu, Sean Eddy, Matthias Kretzler, Peter A Merkel, and Casey S Greene. 2018. MultiPLIER: a transfer learning framework reveals systemic features of rare autoimmune disease.bioRxiv(2018), 395947

    2018

  27. [35]

    Carolina Torrijo-Belanche, Belén Moreno-Franco, Ainara Muñoz-Cabrejas, Naiara Calvo-Galiano, José Antonio Casasnovas, Carmen Sayón-Orea, and Pilar Guallar- Castillón. 2024. High serum phosphate is associated with cardiovascular mortality and subclinical coronary atherosclerosi...

    2024

  28. [36]

    P Tudzynski, U Stahl, and K Esser. 1980. Transformation to senescence with plasmid like DNA in the ascomycete Podospora anserina.Current Genetics2 (1980), 181–184

    1980

  29. [37]

    Aftab Ullah, Asif Jan, Hasan Naeem Kareem, Wahby Mohammed Ahmed Babaresh, Abdur Rahim, Syed Shaukat Ali, Waheed Ali Shah, Salim K Hajwal, Alaa Hamza Hermis, Mustafa Kareem Jawad, et al. 2025. Prevalence and Deter- minants of Coronary Artery Calcification in Adults With Metabol...

    2025

  30. [38]

    Toru Uyama, Kazuhito Tsuboi, and Natsuo Ueda. 2017. An involvement of phospholipase A/acyltransferase family proteins in peroxisome regulation and plasmalogen metabolism.FEBS letters591, 18 (2017), 2745–2760

    2017

  31. [39]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks.arXiv preprint arXiv:1710.10903(2017)

    2017 arXiv

  32. [40]

    Valerie Z Wall, Shelley Barnhart, Jenny E Kanter, Farah Kramer, Masami Shimizu- Albergine, Neeta Adhikari, Thomas N Wight, Jennifer L Hall, and Karin E Born- feldt. 2018. Smooth muscle glucose metabolism promotes monocyte recruitment and atherosclerosis in a mouse model of met...

    2018

  33. [41]

    Tongxin Wang, Wei Shao, Zhi Huang, Haixu Tang, Jie Zhang, Zhengming Ding, and Kun Huang. 2021. MOGONET integrates multi-omics data using graph con- volutional networks allowing patient classification and biomarker identification. Nature communications12, 1 (2021), 3445

    2021

  34. [42]

    Gregory P Way and Casey S Greene. 2018. Extracting a biologically relevant latent space from cancer transcriptomes with variational autoencoders. InPACIFIC SYMPOSIUM on BIOCOMPUTING 2018: Proceedings of the Pacific Symposium. World Scientific, 80–91

    2018

  35. [43]

    Ming-Yu Xie, Shaowu Zhang, Tong Zhang, Yan Li, and Xiaodong Cui. 2025. MTGCL: Multi-Task Graph Contrastive Learning for Identifying Cancer Driver Genes from Multi-omics Data.IEEE Transactions on Computational Biology and Bioinformatics(2025)

    2025

  36. [44]

    Hiroyuki Yamamoto, Naofumi Yoshida, Toshiro Shinke, Hiromasa Otake, Masaru Kuroda, Kazuhiko Sakaguchi, Yushi Hirota, Takayoshi Toba, Hachidai Takahashi, Daisuke Terashita, et al. 2018. Impact of CD14++ CD16+ monocytes on coronary plaque vulnerability assessed by optical cohere...

    2018

  37. [45]

    Zhou Yin, Mischa Machius, Eric J Nestler, and Gabby Rudenko. 2017. Activator Protein-1: redox switch controlling structure and DNA-binding.Nucleic acids research45, 19 (2017), 11425–11436

    2017

  38. [46]

    Rex Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec. 2019. GNN Explainer: A Tool for Post-hoc Explanation of Graph Neural Networks. CoRRabs/1903.03894 (2019). arXiv:1903.03894 http://arxiv.org/abs/1903.03894

    2019 arXiv

  39. [47]

    Hiroyuki Yoshida, Shintaro Inoue, and Yasunori Okada. 2025. New molecules in- dispensable for hyaluronan degradation, HYBID (CEMIP/KIAA1199) and TMEM2 (CEMIP2): Differential roles in physiological and pathological non-neoplastic conditions.Proceedings of the Japan Academy, Ser...

    2025

  40. [48]

    Silvio Zaina, Holger Heyn, F Javier Carmona, Nuray Varol, Sergi Sayols, En- ric Condom, José Ramírez-Ruz, Antonio Gomez, Isabel Gonçalves, Sebastian Moran, et al. 2014. DNA methylation map of human atherosclerosis.Circulation: Cardiovascular Genetics7, 5 (2014), 692–700

    2014

  41. [49]

    Zunlan Zhao, Shouhang Chen, Hongzhao Wei, Weile Ma, Weili Shi, Yixin Si, Jun Wang, Liuyi Wang, and Xiqing Li. 2024. Online application for the diagnosis of atherosclerosis by six genes.Plos one19, 4 (2024), e0301912

    2024

  42. [50]

    Yu Zhou, Haixia Zheng, Xin Huang, Shufeng Hao, Dengao Li, and Jumin Zhao

  43. [2012]

    Improvements in risk stratification for the occurrence of cardiovascular disease by imaging subclinical atherosclerosis: a systematic review.Heart98, 3 (2012), 177–184

    2012

  44. [2022]

    Received 10 September 2025

    Graph neural networks: Taxonomy, advances, and trends.ACM Transactions on Intelligent Systems and Technology (TIST)13, 1 (2022), 1–54. Received 10 September 2025

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.