A Comparative Analysis of Deep Learning Models for Interpretable Multi-Omics Cancer Subtyping Using Integrated Gradients

Abstract

The integration of multi-omics data has advanced cancer subtype classification, yet the deep learning models employed often lack interpretability. Integrated Gradients (IG), an explainable AI (XAI) method, addresses this by attributing model predictions to input features, thereby facilitating biomarker discovery. While prior studies typically apply IG to a single architecture, this research systematically benchmarks IG across three distinct deep learning frameworks: MOGONET, MMDynamics, and HTML, using multi-omics data from The Cancer Genome Atlas (TCGA). We leverage IG to identify and compare top-ranked biomarkers, assessing their biological relevance and reproducibility across models. Our experiments, conducted with stratified K-fold cross-validation. This work provides practical guidance on selecting appropriate deep learning architectures and IG baselines for multi-omics research, aiming to foster more interpretable and clinically meaningful applications of AI in oncology.