RRHGE: A Novel Approach to Classify the Estrogen Receptor Based Breast Cancer Subtypes

Ashish Saini, Jingyu Hou, Wanlei ZhouThe Scientific World JOURNAL 2014 · DOI

TL;DR

This paper introduces RRHGE, a novel algorithm that leverages gene interaction networks to classify estrogen receptor (ER)-positive and ER-negative breast cancer subtypes with improved accuracy. By focusing on reliable gene interactions and identifying "hub genes" within subnetworks, RRHGE offers a more robust approach to subtype classification, crucial for personalized treatment strategies.

What Problem Does This Paper Address?

Breast cancer, a leading cause of mortality in women, is heterogeneous, with subtypes like ER-positive (ER+) and ER-negative (ER-) responding differently to treatments. Accurate classification is paramount for effective therapy. While gene signatures derived from gene expression data have shown promise, existing methods often suffer from instability and poor reproducibility across different datasets. A key challenge is the presence of "false-positive" interactions in protein-protein interaction (PPI) databases, which can bias the identification of reliable gene signatures. This paper aims to address these limitations by developing a method that not only extracts gene signatures but also rigorously assesses the reliability of the underlying gene interactions.

Methods

The core of the RRHGE (Robust Reliability Based Hub Gene Expression) algorithm lies in its multi-faceted approach to identifying reliable gene signatures:

  • Reliability Metrics: The paper introduces three reliability measures for gene interactions:
  • Data Source-Based Reliability ($\beta_1$): Assesses reliability based on the number of independent databases reporting an interaction.
  • Experimental Method-Based Reliability ($\beta_2$): Evaluates reliability based on the variety of experimental methods used to detect an interaction.
  • Interaction Level-Based Reliability ($\beta_3$): Considers the "neighborhood" of interacting genes, assuming interactions with more connected genes are more biologically relevant.
  • Weighted Reliability Measure ($\Omega$): These three metrics are combined using a multivariate linear regression model to create a single, weighted reliability score for each gene interaction.
  • Gene Expression Metrics: Gene expression values are summarized, and then merged with the weighted reliability scores to create "reliable gene expressions."
  • Subnetwork Identification: The algorithm identifies subnetworks of highly correlated and reliable gene expressions.
  • Hub Gene Identification: Within these subnetworks, "hub genes" (those with the most interactions) are identified. The paper posits that hub genes are crucial for driving biological processes and are thus important for classification.
  • Gene Signature Construction: The RRHGE algorithm constructs gene signatures by selecting subnetworks and their associated hub genes, aiming for robustness and biological relevance.

Datasets

The study utilized several publicly available datasets:

  • Protein-Protein Interaction (PPI) Datasets: Six PPI datasets were integrated: BIOGRID, INTACT, MINT, DIP, BIND, and HPRD.
  • Breast Cancer Gene Expression Datasets: Five datasets were used: GSE7390, GSE6532, GSE21653, GSE11121, and the van de Vijver dataset.
  • Four datasets (GSE7390, GSE6532, GSE21653, and GSE11121) were integrated to form a larger training set.
  • The van de Vijver dataset was used as an independent testing set.
  • The Desmedt dataset (GSE7390) was also used as a testing set.

Key Insights and Ideas

  • Reliability is Key: The paper strongly emphasizes that simply using raw PPI data leads to unreliable gene signatures. Rigorously assessing the reliability of gene interactions is crucial for accurate classification.
  • Hub Genes as Drivers: The hypothesis that hub genes within subnetworks are critical for identifying biologically relevant gene signatures is a central theme. These genes are thought to be involved in essential biological processes driving cancer.
  • Subnetwork Approach Superiority: The study suggests that gene signatures derived from subnetworks are more reproducible and perform better than those derived from simple gene lists.
  • Multi-Platform Integration: By integrating data from multiple microarray platforms, the study aims to reduce dataset bias and improve the generalizability of the gene signatures.

Key Results

  • High Classification Accuracy: The RRHGE algorithm achieved superior classification performance compared to several existing methods on two independent testing datasets (Desmedt and van de Vijver).
  • Matthews Correlation Coefficient (MCC): On the Desmedt dataset, RRHGE achieved an MCC of 0.868, significantly outperforming other methods. On the van de Vijver dataset, it achieved an MCC of 0.692.
  • Robustness: RRHGE demonstrated stable classification performance across both testing datasets, unlike some other algorithms that showed significant variations.
  • Gene Signature Size: The final RRHGE gene signature consists of 471 genes (326 for ER+ and 145 for ER− subtypes).
  • Biological Validation: Gene ontology (GO) analysis and pathway analysis confirmed that the genes in the RRHGE signature are associated with biological processes known to be disrupted in cancers, such as apoptosis, cell death, and DNA damage response.
  • Survival Analysis: Kaplan-Meier survival analysis on the Desmedt dataset showed that the RRHGE gene signature effectively separates ER+ (good prognosis) and ER− (poor prognosis) patient groups based on distant metastasis-free survival (DMFS) and overall survival (OS).

Limitations and Future Work

  • Computational Cost: While RRHGE-TSN (a variant focusing on the top subnetwork) offers a computationally less intensive option, the full RRHGE algorithm is presented as the most accurate, implying a trade-off between performance and computational resources.
  • Dataset Bias: The paper acknowledges that while integrating multiple datasets helps, further increasing the number of training sets and incorporating more platform types could further enhance the overlap and robustness of subnetwork-based gene signatures.
  • Integration of Other Data Types: The authors suggest investigating how classification performance might be improved by integrating other data types, such as DNA copy number variation.
  • Improving Reliability Metrics: The study proposes that further research could focus on enhancing the reliability metrics by incorporating more biologically relevant information.

Summary updated

Abstract

The evaluation on real breast cancer samples showed that our RRHGE algorithm achieved higher classification accuracy than the existing approaches.
BibTeX
@article{saini2014rrhge,
  author = {Ashish Saini and Jingyu Hou and Wanlei Zhou},
  title = {{RRHGE: A Novel Approach to Classify the Estrogen Receptor Based Breast Cancer Subtypes}},
  journal = {The Scientific World JOURNAL},
  year = {2014},
  doi = {10.1155/2014/362141},
}
Figure