iUMRG: multi-layered network-guided propagation modeling for the inference of susceptibility genes and potential drugs against uveal melanoma

Uveal melanoma (UM) affects approximately 5 per million adults and carries high risks of distant metastasis and poor survival, with median overall survival of 6–12 months in metastatic disease. Current treatments for localized tumors (enucleation, radiation) depend on tumor size and location, while effective systemic therapies for metastatic UM remain lacking. Conventional chemotherapy has shown limited efficacy, and targeted therapies against known mutations and pathways (e.g., MAPK, PI3K/Akt, Rho GTPase) have not yielded definitive benefits. Given the limitations of preclinical models to recapitulate clinical complexity, there is a critical need for computational approaches that leverage multi-omics and network medicine to identify novel susceptibility genes (SGs) and therapeutic candidates. This study aims to build and apply an integrated multi-layered molecular network and a propagation-based algorithm (iUMRG) to infer UM-related SGs and prioritize potential drugs, thereby advancing precision diagnostics, prognostics, and treatment.

Prior research highlights cytogenetic alterations (chromosomes 1p, 3, 6, 8) as prognostic indicators in UM and recurrent driver mutations in GNAQ/GNA11 and other GPCR-related genes influencing downstream MAPK, PI3K/Akt, and Rho GTPase signaling. Despite these insights, clinical trials with targeted therapies have not improved outcomes. Network medicine and multi-omics integration have emerged as powerful approaches for complex disease gene discovery via random walk and diffusion-based propagation across heterogeneous biological networks. However, such integrative network approaches have been underutilized in UM. The present work builds on these advances by integrating transcriptomic, ncRNA, regulatory, and interactome data to systematically infer UM susceptibility genes and reposition drugs.

Data integration and HMMN construction: Heterogeneous multi-layered molecular networks (HMMN) were assembled from public resources: lncRNA–mRNA (LncRNATarget, starBase), miRNA–mRNA and miRNA–lncRNA (miRTarBase, LncBase), TF–gene and TF–miRNA (TRANSFAC, TransmiR), and protein–protein interactions (HuRI). Known cancer gene sets were from COSMIC and MSigDB. The integrated HMMN comprised 18,231 nodes and 126,187 edges (as reported). Known UM-related SGs: 59 experimentally supported UM SGs (8 mRNAs, 39 miRNAs, 12 lncRNAs) were collected from CTD, HMDD, miRCancer, Lnc2Cancer, LncRNADisease, and Nc2Eye. Patient datasets: TCGA UM cohort (n=80) and GEO datasets GSE44295 (n=59), GSE22138 (n=63), and GSE84976 (n=28) provided expression and clinical data for validation. Propagation model (iUMRG): A random walk with restart (RWR) was implemented on the HMMN. Initialization set R_G^0 with seed nodes (the 59 SGs) at 1/59 and others 0. Iterative update: R_G^t = r * W * R_G^{t-1} + (1 - r) * R_G^0, where W is the HMMN adjacency matrix and r is the restart probability (optimized to 0.4). Convergence yields node equilibrium probabilities as similarity/risk scores relative to seed SGs. Performance evaluation: Leave-one-out cross-validation (LOOCV) used 58 SGs as seeds and held out 1 as test across 59 iterations. Ranked lists by risk score were aggregated to compute ROC curves and AUROC. Genome-wide prediction: Using all 59 SGs as seeds, iUMRG scored and ranked all HMMN nodes; the top 50 were designated candidate high-confidence SGs (HSGs). Functional and clinical validation: - KEGG enrichment via Metascape for co-expressed genes of lncRNAs, targets of risk miRNAs, and predicted risk protein-coding genes. - Differential expression between primary vs. metastatic tumors using limma. - Survival analyses (OS, PFS, DSS) via univariate/multivariate Cox models, Kaplan–Meier with log-rank tests, and optimal cutpoints (survminer). - Network topology analyses (mean distances, density, subnetwork construction, betweenness centrality) within the HMMN. Drug repositioning: Drug targets from DrugBank; small molecule–miRNA and drug–lncRNA associations from SM2miR and D-Lnc. A hypergeometric test assessed enrichment overlap between drug targets and candidate HSGs. Significant drugs were prioritized as repositioning candidates. Statistical analyses: Limma for differential expression; Cox regression and KM for survival; KEGG enrichment visualization with Metascape.

- HMMN integration resulted in a network with 18,231 nodes and 126,187 edges. - iUMRG performance: LOOCV achieved AUROC = 0.8825 at restart probability r = 0.4, outperforming other r values. - Candidate HSGs: Top 50 ranked genes included 22 protein-coding genes, 18 miRNAs, and 10 lncRNAs. - Functional enrichment: • LncRNA co-expressed genes enriched in calcium signaling, homologous recombination, neuroactive ligand–receptor interaction, and autophagy. • Targets of risk miRNAs enriched in viral carcinogenesis, miRNAs in cancer, pathways in cancer, and transcriptional misregulation in cancer. • Predicted risk protein-coding genes enriched in viral carcinogenesis. - Clinical associations: • In TCGA, 27/50 candidates were significantly or marginally associated with OS, PFS, or DSS. • In three additional cohorts, 11 (GSE44295), 6 (GSE22138), and 4 (GSE84976) candidates showed significant or marginal associations with outcomes. • Expression by stage (TCGA): MTUS2 (P=0.026), OIP5-AS1 (P=0.029), and hsa-miR-31 (P=0.007) increased with advancing clinical stage. • Primary vs metastatic tumors: 13 candidates were differentially expressed in at least one cohort; up-regulated in metastasis: EZH2, NEAT1, hsa-miR-200c, hsa-miR-200b, EGR1, IRF7, FBXW7; down-regulated: CTNNB1, KRTAP1-1, NAA50, PTEN, SNHG1, hsa-let-7c. - Network topology: The 50 candidates formed a subnetwork with 46 nodes and 201 edges, showing shorter mean distances and higher density than background. SNAI1 and RELA had the highest betweenness centrality. Higher SNAI1 and RELA expression associated with worse survival (TCGA: P=0.017 and P=0.0093; GSE44295: P≈0.089 and P≈0.07). - Drug repositioning: 17 candidate HSGs overlapped with targets of 10 drugs (7 FDA-approved). Examples: ASA predicted to target MYC and TP53; DHA, glucose, berberine derivative, and ATRA predicted to target 10 risk miRNAs (miR-141-3p, miR-181a-5p, miR-9-5p, miR-429, miR-200b-3p, miR-449b-5p, miR-200c-3p, miR-200a-3p, miR-7-5p, miR-27a-3p). Carboplatin+docetaxel, quercetin, isoprenaline, diamorphine, and panobinostat predicted to target lncRNAs GAS5, SNHG1, KCNQ1OT1, MEG3, and NEAT1. Overlaps included carboplatin+docetaxel, quercetin, isoprenaline, and panobinostat targeting GAS5 and MEG3; glucose and ATRA targeting miR-200b-3p and miR-200c-3p. Six drugs (ASA, DHA, ATRA, carboplatin+docetaxel, panobinostat, quercetin) have prior approval or use in UM contexts.

The study demonstrates that integrating heterogeneous molecular layers and applying propagation modeling can effectively prioritize UM susceptibility genes with strong cross-cohort clinical relevance and cancer-related functional enrichment. The high AUROC under LOOCV validates the robustness of the network-guided approach. Many predicted HSGs align with established UM biology, including EMT regulator SNAI1, NF-κB component RELA, and epigenetic regulator EZH2, underscoring the method’s ability to recover biologically meaningful drivers. Clinical correlations across TCGA and multiple GEO cohorts support the associations of these HSGs with progression, metastasis, and survival. Network analyses indicate that predicted HSGs occupy central, densely connected positions, consistent with disease-relevant modules. Drug repositioning analyses identified actionable candidates, several already used or trialed in UM, reinforcing the translational potential of iUMRG for therapy prioritization.

This work introduces iUMRG, a propagation-based framework on an integrated multi-layered molecular network to infer UM susceptibility genes and prioritize drugs. iUMRG achieved strong predictive performance (AUROC 0.8825), identified top 50 candidate HSGs with multi-cohort clinical and functional support, and proposed 10 drug candidates targeting 17 HSGs, including several with existing UM relevance. The framework can support precision medicine in UM by guiding biomarker discovery and therapeutic development. Future work should expand network layers (e.g., GWAS, Hi-C, disease similarity), optimize propagation models, and perform experimental validation of predicted HSGs and drug effects in vitro and in vivo.

- Network completeness: The integrated HMMN may be improved by incorporating additional data modalities such as GWAS, Hi-C chromatin conformation, and disease similarity networks. - Model optimization: Alternative or enhanced propagation models could further improve predictive performance. - Experimental validation: Predicted high-confidence susceptibility genes and drug candidates require in vitro and in vivo validation to confirm biological relevance and therapeutic potential.