Graph Neural Network Architectures for Multi-Omics-Based Cancer Classification with Emphasis on Interpretability and Biomarker Discovery

Abstract

Cancer describes a class of diseases in which malignant cells form inside the human body due to genetic change. These cells divide indiscriminately upon development, extend throughout the organs, and in many cases, they can result in loss of life. Cancer is the second leading cause of mortality globally after cardiovascular illnesses. Recent studies on integrating multiple omics data highlighted the potential to advance our understanding of the cancer disease process.

Graph neural networks (GNNs) have emerged as powerful computational models for cancer classification tasks, particularly when applied to high-dimensional and heterogeneous multi-omics datasets. GNNs differ from classic neural models MLPs, CNNs, RNNs through their capability to handle complex biological network relationships by mapping biological entities as graph nodes which they analyze using network structure information. They perfectly suit PPI networks or gene regulatory networks because they can effectively capture the natural biological interactions present in these networks. GNNs address key challenges in multi-omics data analysis, including data sparsity and complexity, by learning node embeddings that integrate both omics features and topological information. Attention-based GNNs have advanced both model interpretability and predictive accuracy which leads to more precise biomarker and cancer type classification. These advantages make GNNs as effective approaches to optimizing precision oncology especially when they use integrated omics data as input features.

Graph Attention Networks (GATs) improve attention-based GNNs by implementing dynamic weights for neighboring nodes which depend on their relevance to the model learning process. The selective attention mechanism proves highly effective when analyzing multiomics data because different biological relationships have varying degrees of informative value. Building upon the strengths of GATs in emphasizing important interactions, the Graph Kolmogorov–Arnold Network (GKAN) introduces new interpretability through its combination of Kolmogorov–Arnold representation theorem with graph structures. The univariate functions of GKAN provide effective non-linear modeling capacity for multi-omics data structures which maintain their network connections.

Our work introduces three key innovations: (1) LASSO-MOGAT, a novel Graph Attention Network that integrates LASSO-based feature selection with multi-omics graph learning, demonstrating superior performance in classifying 31 cancer types; (2) An interpretable Graph Kolmogorov–Arnold Network (GKAN) that identifies pan-omics biomarker signatures through learnable activation functions; and (3) A systematic comparison of graph construction methods, proving that multi-omics correlation networks outperform single-omics approaches.