Relational Graph Convolutional Networks for Glioblastoma Biomarker Discovery via ceRNA and Copy Number Variation Analysis
- 1University of Cincinnati
Abstract
Glioblastoma (GBM) is a highly aggressive brain tumor with a five-year survival rate of 6.9%, attributable in substantial part to the shortage of reliable biomarkers. Competing endogenous RNA (ceRNA) and copy number variation (CNV) analyses each carry biomarker-identification potential, but existing work treats them separately and does not integrate multiple regulatory mechanisms. We therefore applied relational graph convolutional networks (RGCNs) to ceRNA and CNV knowledge graphs under a late-fusion ensemble architecture. Across 10-fold cross-validation the RGCN discriminated best among the graph architectures tested (AUCROC 0.874 {+/-} 0.070), significantly exceeding graph convolutional, graph attention and relational attention networks. Combining the ceRNA and CNV branches at the decision level gave the best overall performance (AUCROC 0.883 {+/-} 0.072; PR-AUC 0.208 {+/-} 0.152) and improved on the ceRNA-only model in precision--recall terms, although that improvement does not survive correction for multiple comparisons and we therefore report it as suggestive. Screening the late-fusion ranking against the existing glioma literature left five candidates that are absent from the curated glioblastoma biomarker set and the subject of at most one prior glioma report, among them hsa-miR-203b and hsa-miR-5683, each differentially expressed by more than fivefold on a log_2 scale. All five are computational predictions. Relational graph learning over a ceRNA network, combined with genomic dosage at the decision level, is thus a workable framework for biomarker prioritization, and the five loci give targeted experimental work a place to start.
Related
- Integrating CSF and EV Proteomes Enhances Biomarker Discovery by Capturing Shared Variation Across Compartments
- ALSmiR: a curated evidence resource for ALS-associated microRNAs for biomarker discovery
- An alignment-last approach enables rapid transcriptomic biomarker discovery in large cohorts
- Integrating multi-omics with machine learning: from genomic variant functional annotation to pharmacogenomics and biomarker discovery
- The cGAS-STING pathway in acute myocardial infarction: Biomarker discovery and mechanistic insights
BibTeX
@misc{khandelwal2026relational,
title = {{Relational Graph Convolutional Networks for Glioblastoma Biomarker Discovery via ceRNA and Copy Number Variation Analysis}},
author = {S. Khandelwal and J. Zhan and N. Jarvis},
howpublished = {bioRxiv},
year = {2026},
month = aug,
doi = {10.64898/2026.08.16.744525},
url = {https://www.biorxiv.org/content/10.64898/2026.08.16.744525v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - Relational Graph Convolutional Networks for Glioblastoma Biomarker Discovery via ceRNA and Copy Number Variation Analysis
AU - S. Khandelwal
AU - J. Zhan
AU - N. Jarvis
PY - 2026
DA - 2026/08/16
PB - bioRxiv
DO - 10.64898/2026.08.16.744525
UR - https://www.biorxiv.org/content/10.64898/2026.08.16.744525v2
AB - Glioblastoma (GBM) is a highly aggressive brain tumor with a five-year survival rate of 6.9%, attributable in substantial part to the shortage of reliable biomarkers. Competing endogenous RNA (ceRNA) and copy number variation (CNV) analyses each carry biomarker-identification potential, but existing work treats them separately and does not integrate multiple regulatory mechanisms. We therefore applied relational graph convolutional networks (RGCNs) to ceRNA and CNV knowledge graphs under a late-fusion ensemble architecture. Across 10-fold cross-validation the RGCN discriminated best among the graph architectures tested (AUCROC 0.874 {+/-} 0.070), significantly exceeding graph convolutional, graph attention and relational attention networks. Combining the ceRNA and CNV branches at the decision level gave the best overall performance (AUCROC 0.883 {+/-} 0.072; PR-AUC 0.208 {+/-} 0.152) and improved on the ceRNA-only model in precision--recall terms, although that improvement does not survive correction for multiple comparisons and we therefore report it as suggestive. Screening the late-fusion ranking against the existing glioma literature left five candidates that are absent from the curated glioblastoma biomarker set and the subject of at most one prior glioma report, among them hsa-miR-203b and hsa-miR-5683, each differentially expressed by more than fivefold on a log_2 scale. All five are computational predictions. Relational graph learning over a ceRNA network, combined with genomic dosage at the decision level, is thus a workable framework for biomarker prioritization, and the five loci give targeted experimental work a place to start.
ER -