Uncertainty-aware graph structure optimization with ensemble learning for enhanced cancer gene identification.

The heterogeneity and noise of biological networks reduce prediction accuracy. We propose NexusGene, a computational framework that integrates uncertainty-aware graph structure learning (UnGSL) with clustering-enhanced ensemble learning to improve cancer gene identification in complex biological networks. UnGSL refines the graph structure by quantifying uncertainty in node features and adjusting edge weights to mitigate the influence of noisy or unreliable data. Additionally, a clustering-enhanced ensemble learning strategy reduces prediction bias and false negatives by optimizing base learners based on gene distribution patterns and combining predictions from multiple models. NexusGene was benchmarked against six current methods across seven pan-cancer and 31 cancer-specific networks, consistently achieving the best overall performance, with particularly strong gains in heterogeneous settings. These results establish NexusGene as a robust and interpretable framework for cancer gene discovery across diverse cancer contexts.