Source-linked AI summary
Hybrid Approach of Relation Network and Localized Graph Convolutional Filtering for Breast Cancer Subtype Classification
Sungmin Rhee, Seokjun Seo, Sun Kim
TL;DR
Disease-subtype classification requires modeling complex biological relationships, while conventional deep learning is poorly matched to graph-structured data. The paper combines graph CNN and relation network components to learn localized cooperative-gene patterns and their associations, and reports better performance than existing methods in subtype classification and patient survival analysis.
Problem
Network biology needs disease-specific network knowledge that remains limited, while conventional deep learning is not designed mainly for graph-structured disease-subtype data.
Method
The model uses graph CNN to learn localized expression patterns of cooperative gene communities and relation network to learn associations between those patterns.
Results
The proposed method achieved significantly better performance than existing methods in breast cancer subtype classification and patient survival analysis.
Takeaways & Limitations
The approach captures cooperative gene patterns and their associations, supporting quantitative and qualitative characterization of breast cancer molecular characteristics.
Abstract
from arXiv · showhide
Network biology has been successfully used to help reveal complex mechanisms of disease, especially cancer. On the other hand, network biology requires in-depth knowledge to construct disease-specific networks, but our current knowledge is very limited even with the recent advances in human cancer biology. Deep learning has shown a great potential to address the difficult situation like this. However, deep learning technologies conventionally use grid-like structured data, thus application of deep learning technologies to the classification of human disease subtypes is yet to be explored. Recently, graph based deep learning techniques have emerged, which becomes an opportunity to leverage analyses in network biology. In this paper, we proposed a hybrid model, which integrates two key components 1) graph convolution neural network (graph CNN) and 2) relation network (RN). We utilize graph CNN as a component to learn expression patterns of cooperative gene community, and RN as a component to learn associations between learned patterns. The proposed model is applied to the PAM50 breast cancer subtype classification task, the standard breast cancer subtype classification of clinical utility. In experiments of both subtype classification and patient survival analysis, our proposed method achieved significantly better performances than existing methods. We believe that this work is an important starting point to realize the upcoming personalized medicine.
1 Introduction
Breast cancer subtyping is difficult because heterogeneous genetic, epigenetic, environmental, and gene-level interactions shape cancer phenotypes. The paper addresses this challenge with a hybrid graph CNN–relation network model for learning localized gene patterns and their associations.
- Breast cancer is multifactorial, with interacting genetic, epigenetic, and environmental risk factors influencing progression, development, and metastasis.
- PAM50 provides a clinically useful molecular subtype system for diagnosis decisions and treatment planning, but its biological mechanisms remain incompletely understood.
- Gene interactions and cooperative relationships must be considered because dysregulated genes jointly influence phenotypic differences through complexes, regulation, and signal transduction.
- Conventional deep learning is not straightforward for disease-subtype classification because it is designed mainly for grid-like rather than graph-structured data.
- The proposed hybrid model uses graph CNN to capture localized patterns of associating genes and a relation network to learn associations among those patterns.
- Experiments on synthetic and breast cancer subtype data evaluated classification performance and biological characteristics including survival hazard and subtype prognosis.
2 Related Work
Prior work extended neural networks to graph-structured data through recursive propagation, graph convolutions, attention, and relational reasoning. These approaches provide the technical background for modeling biological networks and entity relationships.
- Graph neural networks propagate node representations through edges until reaching a stable equilibrium, then use them for classification and regression.
- Graph CNN methods extended convolution beyond grid-like domains to manifolds and graphs, including spectral formulations based on Fourier-domain convolution properties.
- Chebyshev approximation improved the computational efficiency of spectral graph CNN while achieving higher experimental accuracy than existing spectral graph CNNs.
- Graph Attention Network evaluates masked self-attention across neighboring nodes to produce graph-node feature vectors and achieved state-of-the-art results across four graph-data experiments.
- Relational networks were proposed as a general neural-network solution for relational reasoning, while deductive approaches lack robust handling of fuzzy and variable inputs.
3 Methods
The proposed method combines graph convolution to capture localized gene-expression patterns with relation reasoning to learn associations among gene groups. It uses graph-based filtering and a modified relation network to process large biological networks and selected graph-node pairs.
- 3.1 Localized Pattern Representation by Graph Convolution Neural Network: Graph convolution captures localized patterns from gene-expression signals mapped onto a weighted gene graph.The method uses the graph Laplacian to define graph convolution over the network structure.
- 3.1 Localized Pattern Representation by Graph Convolution Neural Network: Chebyshev polynomial approximation makes graph convolution computationally scalable for large biological networks.The filtering operation is fully differentiable and has complexity O(K|E|F_inF_outP + K|V|).
- 3.1 Localized Pattern Representation by Graph Convolution Neural Network: The graph convolution output is pooled over neighboring nodes, with average pooling selected empirically for the proposed method.Neighboring nodes are identified using the Graclus algorithm.
- 3.2 Learning Relation Between Graph Entities Using Relation Network: The modified relation network selects the top κ edge-weighted object pairs to reduce model size and exclude pairs that are not highly relevant to classification.The selected pairs are processed with pair-specific functions rather than a single shared function.
- 3.3 Merging Graph Convolution Layer and Relation Network: The final architecture combines graph-convolution and relation-network representations, with h comprising graph convolution, pooling, and a fully connected layer.The last graph-convolution layer also supplies input to the relation network, and the model minimizes cross-entropy with Adam.
4 Results and Discussion
Experiments evaluated the hybrid model on synthetic graph-structured data and TCGA breast cancer expression profiles. The proposed method performed best in synthetic comparisons and showed subtype-, prognosis-, and survival-related structure in real data.
- Experimental setup: The evaluation used synthetic data with class-specific graph structures and breast cancer samples labeled by PAM50 subtype.The real dataset used 338 Luminal A, 265 Luminal B, 149 HER2, and 231 Basal-like samples, with topology from STRING.
- 4.1 Synthetic Experiment: The proposed method performed best in all synthetic experiments except one with the largest centroid distance.In that exception, accuracy was 99.35% for the proposed method and 99.36% for GCNN.
- 4.1 Synthetic Experiment: The synthetic task contained overlapping classes, making inference of relations among attributes important for classification.The experiment was designed to test dependencies among graph vertices and class overlap.
- 4.3 Comparison of Classification Performance: The proposed model achieved the best classification performance among the methods compared on breast cancer subtype data.The comparison included graph-based models and conventional baselines, while GCNN+RN showed the second-worst performance.
- 4.4 Consistency of tSNE Visualization and PAM50 Subtype Prognosis: The graph convolution feature map showed distinctive but not sharply separated patterns across the four breast cancer subtypes.The learned subtype ordering matched the known prognosis ordering, unlike the raw-expression tSNE visualization.
- 4.5 Survival Analysis: Feature-map clustering separated patients into groups with distinct survival patterns at p-value < 0.05, whereas raw-expression clustering failed.This survival analysis assessed whether learned representations captured patient prognosis alongside subtype classification.
5 Conclusion
The hybrid relation network and graph convolutional neural network approach learns cooperative gene patterns and their associations in breast cancer. It captured molecular characteristics in both quantitative and qualitative evaluations, while supporting future integration of heterogeneous data sources.
- 5 Conclusion: The hybrid model combines graph CNN and RN to perceive cooperative gene patterns and their associations.Graph CNN captures localized patterns, while RN learns relations between those patterns.
- 5 Conclusion: The method successfully captured breast cancer molecular characteristics in both quantitative and qualitative evaluations.
- 5 Conclusion: Future work will extend the model to heterogeneous sources such as sRNA sequencing, DNA methylation, and gene expression data.The authors propose incorporating multi-view learning and/or transfer learning.