CeRNA–CNV graph fusion has been ported nowhere: cross-tumour evidence stops at component-level validation

Question: How have multi-omics graph fusion approaches combining microRNA-target networks with copy number variation data been evaluated for cancer biomarker prioritization in other tumor types?

Answer

They have largely not been. Across the retrieved literature, exactly one study performs decision-level fusion of a microRNA-target (ceRNA) graph with a CNV graph for biomarker prioritization, and it does so in glioblastoma only [1]. What exists in other tumour types is component-level validation: CNV-to-miRNA dosage coupling is characterised in ovarian cancer [4], miRNA-target network modules are resolved in breast cancer [7] and adrenocortical carcinoma [11], and miRNA plus CNV layers are folded into multi-omics classifiers or survival models in pancreatic, lung, renal and gynaecological cohorts [12] [10] [13]. None of these evaluates the specific endpoint that matters for the fusion claim — whether adding genomic dosage to a miRNA-target graph improves the ranking of candidate biomarkers. Where graph fusion is evaluated outside glioblastoma, the target is patient-level prediction (subtype, survival, drug response) and the marginal gain over single-omics is frequently not statistically secure [6] [2] [1]. Taken together, the papers support a narrower conclusion than the field's framing implies: the biological premise (CNV shapes miRNA output) and the methodological premise (graph learning over regulatory networks can rank candidates) have each been tested in several tumours, but their combination has never been transferred and independently benchmarked.

The only end-to-end evaluation is in glioblastoma, and its fusion gain is reported as suggestive rather than demonstrated

The single study directly matching the query combined a ceRNA knowledge graph and a CNV knowledge graph under a late-fusion RGCN ensemble, with all five novel candidates explicitly described as computational predictions [1]. Fusion gave the best overall discrimination (AUCROC 0.883) but the precision–recall improvement over the ceRNA-only branch did not survive correction for multiple comparisons, and the authors report it as suggestive [1]. Candidate evaluation was performed by screening the fused ranking against a curated glioblastoma biomarker set and prior glioma literature, retaining loci with at most one prior report [1]. That is a novelty filter, not a validated benchmark, and no independent cohort, knockout, or perturbation assay is presented.

In other tumours the CNV→miRNA branch is validated as a dosage–expression relationship, not as a fusion input

Integrative analysis of 8q24.3-encoded miRNAs in high-grade serous ovarian cancer found that copy-number gain broadly associated with increased miRNA expression, but that the effect was not uniform across candidates and that intronic miRNAs coupled variably with host genes [4]. Isoform-resolved analysis did not strengthen the copy-number associations relative to total miRNA measurements, and the prognostic miRNAs reported (miR-937, miR-4664, miR-6849) were derived from expression–survival associations rather than from a fused network ranking [4]. At pan-cancer scale, event-level modelling of somatic copy-number alterations identified 460 loci under selection enriched for known oncogenes and tumour suppressors [3] — establishing that CNV carries prioritizable signal, but with no miRNA layer attached. The coupling assumption behind the fusion design therefore has partial, tumour-specific support at best, and the ovarian result argues against treating CNV dosage as a uniform proxy for miRNA activity.

miRNA-target networks in other tumour types are used for module discovery, not for ranking against genomic dosage

Community-resolved analysis of paired miRNA–mRNA profiles from 1,161 breast cancers with validated interactions and protein networks identified 16 functional miRNA modules, and found most inferred miRNA functions to be strongly subtype-dependent [7]. In adrenocortical carcinoma, ceRNA networks reconstructed from TCGA-ACC and GTEx yielded four candidate biomarkers via multi-layer overlap filtering, with panel classification accuracy reported at AUC 0.99 [11]. A pancreatic adenocarcinoma study integrated copy number alterations, methylation, gene expression and miRNA expression with transcription-factor interactions and predicted roughly 2,000 miRNA–DDR gene pairs, several enriched among experimentally validated interactions [12]. Edge-based miRNA enrichment exploiting miRNA–gene interactions outperformed node-based methods across multiple cancer types and was demonstrated mechanistically in bladder cancer [5]. In each case the miRNA network is an interpretive or discovery scaffold; none of these designs propagates a CNV signal through the network to re-rank candidates.

Where graph or decision-level fusion has been evaluated for other tumours, the endpoint is prediction accuracy and the fusion benefit is often not significant

Similarity network fusion of RNA, methylation and CNV in TCGA-BRCA improved subtype recovery over concatenation (NMI 0.495 vs 0.175) but its survival advantage over RNA-only did not exclude zero (Δ = +0.049, CI −0.036–0.144) [6]. A pathway-centric graph network conditioning expression on mutation and CNV improved survival prediction across 10 TCGA cancer types and 5,650 patients [2], and confidence-guided multi-omics graph learning improved subtype classification across four single-cancer tasks and transferred from BRCA to KIRC without fine-tuning [14]. Feature-level fusion of methylation, miRNA and CNV outperformed other modality combinations in BRCA, CESC and UCEC survival models [10], and hypergraph-based prioritization produced compact panels in LUAD, HNSC, KIRC and COAD [13]. Decision-level fusion across multiple biological networks has also been shown to outperform representation-level consistency in cancer driver gene identification on three cancer datasets [21]. All of these measure predictive performance; none measures whether fusion improves biomarker ranking quality.

The methodological premises of graph fusion are themselves contested, which weakens any cross-tumour claim

A benchmarking study found that none of several GNN architectures outperformed a structure-agnostic multilayer perceptron on PPI graphs for KIRC and breast cancer subtype prediction [17]. Gene-essentiality prediction gains from GNNs degraded substantially under randomized topology and were not preserved by degree-constrained rewiring, indicating that biological structure — not graph machinery — carries the signal [18]. Data leakage is described as a prevalent and underappreciated flaw producing inflated performance estimates and poor generalization in biomarker studies [15]. A re-audited PDAC transcriptomic signature achieved internal C-index 0.630 but external C-index 0.527 with a fixed five-gene model, and no association survived multiple-testing correction across cohorts [16]. Unbalanced functional standards also introduce hidden evaluation bias in network benchmarking, motivating module-resolved precision–recall metrics [19].

Conflicts and gaps

Two claims are in direct tension. The glioblastoma study concludes that relational graph learning over a ceRNA network with decision-level genomic dosage "is a workable framework for biomarker prioritization" [1], while its own fusion improvement fails multiple-comparison correction and its candidates remain untested. Several papers assert transferability without testing it — a GBM driver-pathway framework is described as readily extendable to BRCA, OV and LUAD [9], and pan-cancer claims are made from women's-cancer-only cohorts [10]. There is also unresolved disagreement on whether integrating CNV adds information at all: one study reports that three-to-five-modality integration improved stability but did not surpass the best two-modality combinations [22], and the ovarian analysis found CNV–miRNA coupling non-uniform [4]. Finally, no retrieved paper reports a cross-cancer ablation isolating the contribution of the miRNA-target graph topology from the contribution of CNV features alone — the exact comparison the question turns on.

What would change the answer

A pre-registered port of the late-fusion RGCN design to at least three additional TCGA cohorts (BRCA, LUAD, KIRC, OV) with a factorial ablation — ceRNA-only, CNV-only, early concatenation, late fusion — would settle whether fusion adds ranking value beyond its components; the design should include a degree-preserving shuffled ceRNA graph as a topology control, following the randomized-network test that showed GNN gains vanish without biological structure [18]. Evaluation should move off literature-novelty screening onto held-out ground truth: time-stamped splits of curated biomarker sets so that candidates cannot be scored against the same literature used to build the graph, plus module-resolved precision–recall rather than a single aggregate metric [19]. External transport is essential, on the model of network-propagation subtyping validated in ICGC CCA and PCAWG without retraining [20], of co-essentiality prioritization checked against clinical trial records [8], and of cross-cancer transfer from BRCA to KIRC [14]. Leakage auditing should be built in from the start [15], and at least a subset of top-ranked miRNAs should be taken to perturbation or target-inhibition assays in cell models, as done for network-predicted targets in SOX2-dependent lung cancer [23].

References

[1] Khandelwal, S., Zhan, J., Jarvis, N.. "Relational Graph Convolutional Networks for Glioblastoma Biomarker Discovery via ceRNA and Copy Number Variation Analysis". bioRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.16.744525

[2] Di Wang, Chupei Tang, Junxiao Kong, Jixiu Zhai, Moyu Tang, Tianchi Lu. "PathMoG: A Pathway-Centric Modular Graph Neural Network for Multi-Omics Survival Prediction". arxiv. 2026-04-28T00:00:00.000Z. URL: https://arxiv.org/abs/2604.24371

[3] Kaufmann, T. L., Streck, A., Markowetz, F., Van Loo, P., Schwarz, R. F.. "Deciphering selection patterns of somatic copy-number events". bioRxiv. 2026-03-03T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.01.708809

[4] Filipek, K., Merelli, I., Chiappori, F., Penzo, M.. "Landscape of 8q24.3-Encoded microRNAs and Their Prognostic Impact in Ovarian Cancer". bioRxiv. 2026-03-11T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.06.710032

[5] Zhang, Z., Lai, X.. "miREA: a network-based tool for microRNA-oriented enrichment analysis". bioRxiv. 2026-03-06T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.27.708509

[6] Taha Ahmad. "Graph-Based Multi-Omics Integration Improves Subtype Recovery and Survival Prediction Over Classical Integration Strategies in TCGA-BRCA". arxiv. 2026-03-03T00:00:00.000Z. URL: https://arxiv.org/abs/2603.00212

[7] Newby, E., Davis, E., Dhawan, A.. "Multiomic network analysis reveals conserved and subtype-specific cooperative microRNA regulators in breast cancer". bioRxiv. 2026-05-28T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.07.669176

[8] Lee, K., Kim, D., Kim, I., Lee, J., Ha, D., Lim, S., Kim, E., Im, S.-H., Shin, K., Kim, S.. "A Co-essentiality Network of Cancer Driver Genes Better Prioritizes Anticancer Drugs". bioRxiv. 2024-12-22T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.20.629566

[9] Zhang, J.. "Integrative multi-omics framework identifies phenotypically impactful driver pathways in glioblastoma multiforme". bioRxiv. 2026-08-19T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.28.640792

[10] Nguyen, R., Vafaee, F.. "Multi-Omics Prognostic Marker Discovery and Survival Modeling: A Case Study on Pan-Cancer Survival Analyses in Women's Cancers". medRxiv. 2025-04-08T00:00:00.000Z. URL: https://doi.org/10.1101/2025.01.08.25320212

[11] Omidi, J.. "Therapy-Associated Oncogenic Axes of Core Cellular and Metabolic Programs Revealed by Multi-Omics Regulatory Network Analysis in Adrenocortical Carcinoma". medRxiv. 2026-02-05T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.21.25342239

[12] Bose, S., Upama, P. B., Ahamed, S. I.. "Unraveling miRNA-Driven DNA Damage Response Networks in Pancreatic Adenocarcinoma: A Multi-Omics and Machine Learning Approach". bioRxiv. 2025-12-12T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.09.693279

[13] Gupta, M. K., Paul, M., Pati, S. K.. "HR-FRGS: A Novel Biomarker Discovery Protocol using Dual Layer Hypergraph Learning for NGS RNA Sequence Data". bioRxiv. 2026-09-25T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.21.753079

[14] Boyang Fan, Hengchuang Yin, Siyu Yi, Yifan Wang, Zhicheng Li, Leijiyu Zhou, Jiancheng Lv, Wei Ju. "CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification". arxiv. 2026-10-09T00:00:00.000Z. URL: https://arxiv.org/abs/2604.24201

[15] An, L., Finney, C. A., Shvetcov, A., The Global Neurodegeneration Proteomics Consortium,, Vogel, J.. "Performance of protein panels is inflated across many biomarker studies". medRxiv. 2026-09-26T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.02.26362037

[16] Markarian, M. B., Houssigian, G. V.. "Leakage-Safe Cross-Cohort Transcriptomic Survival Analysis in Pancreatic Ductal Adenocarcinoma: A Corrected Reanalysis". bioRxiv. 2026-09-12T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.14.688421

[17] Schirmacher, J., Maurer, M. C., Metsch, J. M., Ploesch, S., Chereda, H., Blumenthal, D. B., Hauschild, A.-C.. "Benchmarking Graph Neural Networks for Multi-Omics Cancer Subtyping using Methylation and Gene Expression Profiles". bioRxiv. 2026-08-25T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.21.745839

[18] Rahimi, S., Bonner, S., Afzal, A., Milo, M., Morrissey, E., Petsalaki, E.. "Biological Network Organization, Not Generic Graph Topology, Drives Graph-Based Gene Essentiality Prediction". bioRxiv. 2026-07-09T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.30.735480

[19] Demirtas, T., Shaw, A., Billmann, M.. "pFLEX – a Python library for fast functional evaluation of genetic networks at the biological module-level". bioRxiv. 2026-06-14T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.11.731557

[20] Kwon, T., Park, Y.-G., Choi, J.-G.. "VARION: A Network Propagation Framework for Individual Patient Somatic Mutation Interpretation in Cancer Molecular Subtyping". bioRxiv. 2026-09-12T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.11.751075

[21] Bang Chen, Lijun Guo, Houli Fan, Wentao He, Rong Zhang. "Soft-Evidence Fused Graph Neural Network for Cancer Driver Gene Identification across Multi-View Biological Graphs". arxiv. 2025-10-09T00:00:00.000Z. URL: https://arxiv.org/abs/2510.0629

[22] La Ode Aman, Arfan Arfan, Aiyi Asnaw, Purnawan Pontana Putra, Hamsidar Hasan, Dizky Ramadani Putri Papeo. "Prediction of PLX-4720 Sensitivity in Cancer Cell Lines through Multi-Omics Integration and Attention-Based Fusion Modeling". arxiv. 2025-12-16T00:00:00.000Z. URL: https://arxiv.org/abs/2512.12113

[23] Hwang, W., Kottmann, D., Guo, W., MacMahon, M., Correia, L., Harris, R., McCaughan, F., Han, N.. "Utilising multi-modal data-driven network analysis to identify monotherapy and combinational therapy targets in SOX2-dependent squamous cell lung cancer". bioRxiv. 2025-02-04T00:00:00.000Z. URL: https://doi.org/10.1101/2025.01.30.635634

How this survey was made

This survey answers one question: “How have multi-omics graph fusion approaches combining microRNA-target networks with copy number variation data been evaluated for cancer biomarker prioritization in other tumor types?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 23 papers whose work this text draws on, chosen from the 131 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 21, 2026 and covered work from the previous two years.