Answers are grounded in this paper and cite their sources.
Introduction Glioblastoma (GBM) is a highly aggressive brain tumor with a 5-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. Methods We built a heterogeneous ceRNA graph over 2,075 differentially expressed loci, added gene-level copy number, and evaluated relational graph convolutional networks (RGCNs) against graph and non-graph baselines under 10-fold cross-validation. We then introduced a set of controls that separate the network signal from properties of the label set. Results The RGCN discriminated best among the graph architectures tested (AUCROC 0.874 ± 0.070 ), and a late-fusion ensemble with the CNV branch scored highest overall ( ( 0.883 ± 0.072 ) ). Positive labels are concentrated on miRNA nodes (26 of 31; 8.28% of miRNA nodes against 0.26% of mRNA nodes), and node type is recoverable from the relation structure. A logistic regression on one-hot node type alone, using no topology, reaches 0.824 ± 0.139 . An RGCN retrained on labels permuted within node type matches the observed PR-AUC ( ( p = 0.33 ) ) but not the observed AUCROC (0.799 against 0.876, p = 0.015 ), and within the miRNA subgraph, discrimination is weak and not distinguishable from that of the null model ( ( p = 0.10 ) ). The fusion gain has the same source: copy number is undefined at all 314 miRNA loci, the cnv observed flag alone reaches 0.697, and the copy-number value with that flag removed reaches 0.475 ± 0.007 . The five candidate loci remain differentially expressed when re-estimated from raw read counts and are supported by an alias-expanded literature screen, both independent of the model. Discussion Under this label set and graph construction, most of the apparent performance is carried by a node-type prior, with a modest residual network signal on AUCROC. The controls are inexpensive, require no additional data, and apply to any graph model trained on curated labels; we release the 26 graph, labels, and controls so that they can be run in related designs. The prognostic value of the candidate loci is not assessed here.
No takes yet. Share an insight, caveat, or question.
Khandelwal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: