Multi-Omics Integration via Variational Graph Autoencoders for Biomarker Discovery

Abstract

The advent of high-throughput sequencing technologies has ushered in an era of multi-omics data availability, encompassing genomics, transcriptomics, epigenomics, and proteomics. While these diverse modalities offer complementary views of biological systems, integrating them to unravel complex disease mechanisms remains a significant computational challenge. Traditional integration methods often fail to capture the non-linear interactions and the underlying topological structure of biological networks. This paper proposes a novel framework utilizing Variational Graph Autoencoders (VGAE) for the robust integration of multi-omics data, specifically tailored for biomarker discovery in oncology. By constructing patient-similarity networks and leveraging the generative capabilities of variational inference, our approach effectively learns a low-dimensional latent representation that preserves both global graph structure and local feature attributes. We demonstrate that this method outperforms state-of-the-art matrix factorization and deep learning baselines in clustering accuracy and survival prediction. Furthermore, we introduce a gradient-based attribution mechanism to identify high-confidence biomarkers, validating their biological relevance against known pathway databases. Our results suggest that graph-based deep learning offers a scalable and mathematically rigorous path toward precision medicine.