Progressive fusion networks with adaptive graph structure learning for cancer subtype classification.

Classifying cancer subtypes is crucial for clinical diagnosis and therapy. Recently, graph neural network (GNN)-based methods have been explored for this task, but they often rely on predefined, explicit graph structures. In practice, such prior structures may be incomplete, noisy, or unavailable. To overcome these limitations, we propose Progressive Fusion Networks with Adaptive Graph Structure Learning (PFN-AGSL), aiming to optimize the graph structure and effectively fuse multi-view (omics) representations in a hierarchical manner for accurate subtype prediction. PFN-AGSL incorporates a graph structure learning module comprising global guidance and local refinement blocks, which preserves essential structural patterns while refining local connections. Based on the learned topology, the information aggregation module generates view-specific embeddings. These embeddings are then integrated through a progressive fusion strategy that continuously combines shallow-to-deep information, mitigating information loss during integration. Extensive experiments on three cancer datasets show that PFN-AGSL achieves superior performance compared with state-of-the-art methods. Ablation studies further confirm the crucial contributions of graph structure learning and progressive fusion. These results highlight the effectiveness of PFN-AGSL, demonstrating its potential as a promising tool for cancer subtype classification.