Leveraging Multi-Modal Attention Mechanisms for Interpretable Biomarker Discovery and Early Disease Prediction

Abstract

The identification of molecular biomarkers represents a critical challenge in precision medicine, where high-dimensional multi-omics data creates substantial analytical complexity. This study introduces a novel attention-based framework that integrates genomic, transcriptomic, and clinical data through multi-modal attention mechanisms to enable interpretable biomarker discovery and early disease prediction. The proposed architecture employs self-attention layers for feature-level representation learning and cross-modal attention for heterogeneous data integration, addressing the interpretability limitations of conventional black-box approaches. Evaluated on TCGA and UK Biobank datasets, the framework achieves superior predictive performance with AUC scores of 0.924 and 0.897 respectively, while identifying clinically validated biomarker candidates through attention weight visualization. The method demonstrates significant advantages over traditional feature selection techniques and existing deep learning approaches, providing actionable insights for clinical decision-making through transparent feature importance quantification.