Hybrid Approach of Relation Network and Localized Graph Convolutional Filtering for Breast Cancer Subtype Classification

SungMin Rhee, Seokjun Seo, Sun Kim

Abstract

Breast cancer subtyping is critical for determining prognosis and treatment strategies. This study proposes a hybrid deep learning architecture that integrates Graph Convolutional Neural Networks (GCNNs) with Relation Networks (RNs). By combining localized feature extraction from protein-protein interaction (PPI) networks with relational reasoning between gene pairs, the model effectively captures the complex biological dependencies driving cancer phenotypes. Experimental results on synthetic and TCGA genomic data demonstrate that this hybrid approach outperforms standalone graph-based and traditional machine learning methods, achieving an average accuracy of 83.19% and providing clinically relevant patient stratifications.

Introduction

Breast cancer encompasses several molecular subtypes, including Luminal A, Luminal B, HER2-enriched, and Basal-like, each characterized by distinct gene expression profiles and clinical outcomes. Traditional machine learning approaches often fail to account for the underlying biological structure—specifically, how genes interact within functional modules and pathways. Network biology leverages protein-protein interaction (PPI) graphs to represent these relationships, but applying deep learning to such irregular data structures remains a challenge.

Existing methods typically focus on either local neighborhood features or global graph properties. However, cancer progression often involves both localized pathway dysregulation and complex, long-range interactions between different functional modules. This paper addresses these challenges by introducing a hybrid model that performs both localized graph filtering and relational reasoning.

Methodology

The proposed architecture consists of two primary components: a Localized Graph CNN and a modified Relation Network.

Localized Graph Convolutional Neural Network

The GCNN component serves as a feature extractor that operates on the PPI network. To overcome the computational limitations of spectral graph convolutions on large networks, the authors utilize a spectral formulation with Chebyshev polynomial approximation. This allows the model to capture gene expression patterns within a $K$-hop neighborhood, effectively identifying localized functional modules.

Modified Relation Network

While standard Relation Networks (RNs) are designed for relational reasoning, their $O(N^2)$ complexity makes them unsuitable for high-dimensional genomic data. The authors introduce two key modifications to adapt RNs for biological networks:

  1. Edge Selection: Instead of considering all possible gene pairs, the model only processes pairs corresponding to the top $\kappa$ edges (highest weights) in the STRING PPI network.
  2. Parameter Separation: To account for the diverse nature of biological interactions, the model employs separate multi-layer perceptrons (MLPs) for different gene pairs rather than a single shared MLP.
System architecture of the proposed hybrid model, showing the integration of Graph Convolutional Layers and the Relation Network.

The final classification is derived by integrating features from both the GCNN and RN paths, followed by a fully connected layer and a softmax output for subtype prediction.

Comparison of the original Relation Network architecture (a) with the modified version (bottom) featuring parameter separation for biological data.

Experimental Setup and Results

The model was evaluated using synthetic datasets and real-world RNA-seq data from The Cancer Genome Atlas (TCGA). The real-world dataset included 4,303 genes from cancer hallmark sets, with graph structures derived from the STRING database.

Classification Performance

The hybrid method achieved a peak accuracy of 86.29% (average 83.19%), significantly outperforming standalone GCNNs (82.39%), Graph Attention Networks (81.37%), and traditional classifiers like Random Forest (78.54%) and SVM (77.53%). This improvement underscores the value of explicitly modeling relational dependencies between genes.

Clinical and Survival Analysis

Beyond classification accuracy, the model's learned representations were found to be clinically meaningful. Visualization of the internal features using t-SNE revealed clusters that aligned with clinical prognosis: Basal-like (worst prognosis) $\rightarrow$ HER2 $\rightarrow$ Luminal B $\rightarrow$ Luminal A (best prognosis). Furthermore, survival analysis showed that patient groups identified by the model's features had statistically significant differences in survival outcomes ($p = 0.015$), whereas raw gene expression data failed to provide such distinction ($p = 0.25$).

Conclusion

This study demonstrates that a hybrid approach combining localized graph convolutions with relational reasoning provides a powerful framework for breast cancer subtyping. By incorporating biological priors from PPI networks and focusing on high-confidence interactions, the model achieves superior performance and clinical relevance. Future work will explore the integration of multi-omics data to further enhance the model's predictive capabilities across different cancer types.

Summary updated

BibTeX
@inproceedings{rhee2018hybrid,
  author = {SungMin Rhee and Seokjun Seo and Sun Kim},
  title = {{Hybrid Approach of Relation Network and Localized Graph Convolutional Filtering for Breast Cancer Subtype Classification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/490},
}
Figure