XAI-Enabled Ensemble Learning for Reliable Multi Omics Biomarker Discovery in Cancer

Abstract

Early discovery of cancer biomarkers in a consistent manner still represents an important challenge, because these are obtained from multi-omics data whose high dimensionality, heterogeneity and noise. Although deep learning outperforms conventional models for cancer analysis prediction, single-model methods usually lack robustness and interpretability, and are not directly conducive to clinical applications. In this paper, we introduce EnsembleNet an explainable ensemble deep learning model for robust multiomics biomarker discovery across multiple cancer types. EnsembleNet achieves this by incorporating six distinct neural network architectures-FusionNet, TransformerNet, Hybrid-Net, SENet, SuperNet and FusionNetV2-trained on the same data splits and then combined using soft-voting and stacked generalization with a logistic-regression meta-learner. The framework is tested on four TCGA cohorts (BRCA, CESC, LUAD and PRAD) with transcriptomics and proteomics. Model robustness is evaluated by cross-validation, noise perturbation and missing-feature simulations, as well as rigorous statistics testing for significance. Among multiple datasets, EnsembleNet performs significantly better and stably than the other single deep learning and classical machine-learning baselines. We have shown through statistical analysis (Friedman and Wilcoxon signed-rank test followed by effect-size measure) that the ensemble strategy is robust. To learn the computed models that they can be biologically interpretable, a Shapley Additive explanation (SHAP)-based explainability pipeline is used to obtain model-level and ensemble-level consensus biomarker profiles. EnsembleNet is an accurate, robust and interpretable AI methodology for multi-omics biomarker discovery in cancer. Through the integration of ensemble deep learning with explainable AI and rigorous statistical validation, the new framework provided here forms a robust base for translational cancer research and precision medicine use cases.