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

SungMin Rhee, Seokjun Seo, Sun Kim

Introduction

Breast cancer classification remains a critical challenge in oncology, as the disease manifests in distinct subtypes—Luminal A, Luminal B, HER2, and Basal-like—each requiring specific therapeutic strategies. While the current clinical standard, the "PAM50" gene set, provides a framework for subtyping, it often fails to capture the intricate, cooperative interactions between genes that drive cancer progression. This research introduces a hybrid deep learning architecture that integrates Graph Convolutional Neural Networks (GCNN) with Relation Networks (RN) to better model these biological complexities.

Methodology

The proposed model addresses the limitations of traditional deep learning by treating biological data as a graph rather than a grid. The architecture consists of two primary components designed to capture both localized and relational features within gene expression profiles.

Localized Graph CNN (GCNN)

The GCNN layer utilizes spectral filtering, specifically Chebyshev polynomial approximation, to "smooth" gene expression data over a known protein-protein interaction (PPI) network. This allows the model to identify localized patterns by analyzing how a gene's expression level relates to its immediate neighbors in the biological network.

The hybrid model architecture integrates (A) Graph Convolutional Layers and (B) a Relation Network to process gene expression and PPI data.

Modified Relation Network (RN)

Standard Relation Networks are computationally prohibitive for genomic data because they compare every possible pair of entities. To overcome this, the authors implemented a modified RN that sorts edges by weight and selects only the top $\kappa$ (e.g., 200 to 1,000) most significant interactions. This sparse approach enables the model to learn complex associations between specific gene groups without a combinatorial explosion in computation.

Detailed sequence of the Graph Convolutional component, including normalization, ReLU activation, and pooling layers.

Experimental Setup and Datasets

The model was validated using both synthetic and real-world data:

  • Synthetic Data: Generated via Multivariate Normal Distribution (MND) to simulate various graph structures, proving the model's ability to identify dependencies that traditional classifiers like SVMs overlook.
  • TCGA Data: RNA-seq expression profiles from 983 breast cancer patients in The Cancer Genome Atlas, focusing on 4,303 genes from "cancer hallmark" sets.
  • Biological Network: The STRING protein-protein interaction network provided the prior knowledge of gene connectivity used to structure the graph layers.

Key Results

The hybrid model demonstrated significant improvements over existing methods in both classification accuracy and biological relevance:

  1. Superior Accuracy: The model achieved a peak accuracy of 86.29% on the TCGA dataset, outperforming standalone GCNN (85.27%), Graph Attention Networks (85.64%), and traditional SVMs (77.53%).
  2. Prognostic Discovery: Although trained only for subtype classification, the model's learned features naturally clustered patients by survival hazard. A Kaplan-Meier analysis of these features yielded a highly significant p-value of 0.015, whereas clustering based on raw gene expression was not statistically significant (p=0.25).
  3. Robustness: Synthetic experiments confirmed that the model remains effective even when data classes have high overlap, provided their underlying graph dependencies differ.

Discussion and Future Work

A key innovation of this work is the scalability achieved through sparsity in the Relation Network. By restricting relational reasoning to the most significant edges of a biological graph, the authors successfully applied a module typically used for small-scale image objects to massive genomic datasets.

Future research will focus on extending this into a multi-view learning framework. By integrating multiple "omics" data types—such as DNA methylation and sRNA sequencing—the model could provide an even more holistic and accurate view of cancer biology, further refining personalized treatment strategies.

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