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

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Accurate cancer subtype classification is vital for effective clinical diagnosis and treatment strategies.
• Existing graph neural network (GNN) methods often require predefined graph structures, which can be unreliable or unavailable in real-world scenarios.
• Limitations of current GNN approaches necessitate novel methods for robust cancer subtype identification.

Purpose of the Study:

• To develop a novel method, Progressive Fusion Networks with Adaptive Graph Structure Learning (PFN-AGSL), for accurate cancer subtype prediction.
• To address the limitations of predefined graph structures in GNNs by incorporating adaptive graph structure learning.
• To effectively fuse multi-view omics data in a hierarchical manner for enhanced classification performance.

Main Methods:

• PFN-AGSL employs a graph structure learning module with global guidance and local refinement to preserve and refine network topology.
• An information aggregation module generates view-specific embeddings based on the learned graph structure.
• A progressive fusion strategy integrates these embeddings hierarchically, minimizing information loss.

Main Results:

• PFN-AGSL demonstrated superior performance over state-of-the-art methods across three independent cancer datasets.
• Ablation studies confirmed the significant contributions of both the adaptive graph structure learning and progressive fusion components.
• The method effectively handles incomplete or noisy prior graph structures.

Conclusions:

• PFN-AGSL offers a powerful and effective approach for cancer subtype classification by integrating adaptive graph learning and progressive data fusion.
• The proposed method shows significant potential as a valuable tool in clinical oncology for improving diagnostic accuracy.
• This work highlights the importance of learning graph topology and hierarchical data integration for complex biological data analysis.