NSINR: Neuro-symbolic interpretability noise-robust framework for disease ceRNA biomarker prediction

Abstract

Identifying biomarkers of cancers presents a persistent challenge due to insufficient interpretability in model decision. Current approaches only provide explainability that quantifies the contribution of input features or local substructures to predictions at the data level, yet they fail to uncover the inherent reasoning logic of the model and cannot deliver genuine interpretability, which severely restricts their ability to capture comprehensive biological mechanisms. To overcome these limitations, we introduce NSINR, Neuro-Symbolic Interpretability Noise-Robust Framework for Disease ceRNA Biomarker Prediction. Our framework supports dual-channel collaborative learning for disease-ceRNA biomarker prediction. The neural explainability channel constructs interpretable mask matrices to filter structural and feature noise, yielding compact graph topology and clear data-driven explanations. The symbolic interpretability channel models ceRNA regulation as a logical knowledge hypergraph and performs differentiable logical reasoning to provide transparent biological logic. An Attention fusion mechanism unifies neural learning outputs and symbolic reasoning results, ensuring coherent fusion of performance and interpretability. Comprehensive experiments on three tasks demonstrate that NSINR achieves superior performance, yielding mean AUCs of 0.808, 0.820, and 0.856, and mean AUPRs of 0.796, 0.812, and 0.842 for the MCA, CCA, and LCA tasks, respectively. Compared to the current state-of-the-art methods, NSINR achieves the expected predictive performance. Case studies on cancer biomarkers verify its capacity to discover credible and biologically meaningful candidates.