Identification of Robust Pathway Markers for Cancer through Rank-Based Pathway Activity Inference

Navadon Khunlertgit, Byung-Jun YoonAdv. Bioinformatics 2013 · DOI · OpenReview

TL;DR

This paper introduces a novel method for identifying robust pathway markers for cancer by integrating gene expression ranking with probabilistic pathway activity inference. This approach leads to more reliable and reproducible cancer classifiers, outperforming existing methods in predicting breast cancer metastasis.

What Problem Does This Paper Address?

Identifying reliable biomarkers for complex diseases like cancer is a significant challenge, especially with the small sample sizes typical in genomic studies. While gene expression profiles offer a wealth of information, directly using individual gene markers often results in poor interpretability and lack of reproducibility across different datasets. Existing approaches have tried to address this by aggregating gene activities into "pathway markers" or by using relative gene expression. However, there's a need for methods that are not only robust to variations in experimental conditions (like batch effects) but also yield interpretable and consistently performing classifiers. This paper aims to bridge this gap by proposing a method that leverages gene ranking within pathways to infer pathway activity more reliably.

Methods

The core of this paper is an enhanced pathway activity inference method that builds upon probabilistic approaches. The key innovation lies in using gene ranking instead of raw expression values to infer pathway activity. Here's a breakdown:

  • Gene Ranking: For each gene within a pathway and for each sample, the method ranks the gene's expression relative to other genes in the same pathway. This is done by creating a binary vector indicating whether a gene's expression is higher or lower than another gene's expression within a pair. This ranking is inherently robust to technical variations that affect absolute expression levels.
  • Probabilistic Pathway Activity Inference: The method then estimates the probability of observing a particular gene ranking under different phenotypic conditions (e.g., cancerous vs. non-cancerous). This is achieved by calculating the log-likelihood ratio (LLR) between these probabilities for each pair of genes within a pathway.
  • LLR Aggregation: The pathway activity level for a given sample is then computed by aggregating these normalized LLRs across all gene pairs within the pathway.
  • Feature Selection: To manage computational complexity and focus on informative gene pairs, the method employs mutual information to prescreen gene pairs, selecting the top 10% with the highest mutual information with the phenotype.
  • Normalization: To mitigate sensitivity to small variations in gene ranking, the LLRs are normalized by their mean and standard deviation.

This approach combines the robustness of relative gene expression analysis with the power of probabilistic modeling for pathway activity inference.

Datasets

The study utilized six independent breast cancer microarray gene expression datasets:

  • GSE2034 (USA)
  • NKI295 (The Netherlands)
  • GSE7390 (Belgium)
  • GSE1456 (Stockholm)
  • GSE15852
  • GSE9574

These datasets were profiled using the Affymetrix U133a platform, except for the NKI295 dataset which used a custom Agilent chip. The pathway data was sourced from the MSigDB 3.0 Canonical Pathways database. The study also explored the impact of different normalization methods (RMA, GCRMA, MAS5) on the results.

Key Insights and Ideas

  • Ranking is More Robust Than Raw Expression: The fundamental insight is that ranking genes within a pathway provides a more stable and reproducible measure of pathway activity compared to using raw expression values. This is because ranking is less susceptible to technical noise and batch effects that can alter absolute expression levels.
  • Probabilistic Framework Enhances Inference: By framing pathway activity inference probabilistically, the method can better capture the complex relationships between genes and their contribution to pathway function.
  • Pathway Markers Outperform Gene Markers: The study reinforces the idea that pathway-level analysis is more effective for biomarker discovery than focusing on individual genes, leading to more interpretable and robust results.
  • Cross-Dataset Reproducibility is Crucial: A significant focus is on demonstrating that markers identified on one dataset can reliably perform on independent datasets, even those from different platforms or processed with different normalization methods.

Key Results

  • Improved Discriminative Power: The proposed rank-based pathway activity inference method consistently showed superior discriminative power for pathway markers across all six datasets compared to mean, median, and the original LLR-based methods. This was measured using the absolute t-test score.
  • Robustness to Normalization: The method's performance was largely insensitive to the choice of normalization method (RMA, GCRMA, MAS5), indicating its inherent robustness.
  • High Cross-Dataset Reproducibility: Pathway markers identified using the proposed method retained their discriminative power well on independent datasets, even those from different platforms. This significantly outperformed previous methods in cross-dataset validation.
  • Better Classification Performance: When used to build classifiers (e.g., LDA), the proposed method typically yielded the best average classification performance (measured by AUC) across datasets, demonstrating its utility in building accurate and reproducible disease classifiers.
  • Overcomes LLR Method Limitations: The proposed method successfully addressed limitations of the previous LLR-based approach, particularly when dealing with datasets where certain normalization methods caused issues with gene variability or LLR values becoming infinite.

Limitations and Future Work

  • Computational Complexity: While the paper mentions prescreening gene pairs using mutual information to reduce complexity, the number of gene pairs in large pathways can still be substantial. The authors note that the 10% threshold for gene pairs might not be optimal and suggest exploring adaptive strategies for choosing this threshold.
  • Future Work: The authors suggest that their method can be extended to identify diagnostic subnetwork markers and build more accurate and robust disease classifiers. The robustness of the proposed scheme to "normalization mismatch" in cross-dataset experiments is highlighted as a key advantage for future applications.

Summary updated

Abstract

One important problem in translational genomics is the identification of reliable and reproducible markers that can be used to discriminate between different classes of a complex disease, such as cancer. The typical small sample setting makes the prediction of such markers very challenging, and various approaches have been proposed to address this problem. For example, it has been shown that pathway markers, which aggregate the gene activities in the same pathway, tend to be more robust than gene markers. Furthermore, the use of gene expression ranking has been demonstrated to be robust to batch effects and that it can lead to more interpretable results. In this paper, we propose an enhanced pathway activity inference method that uses gene ranking to predict the pathway activity in a probabilistic manner. The main focus of this work is on identifying robust pathway markers that can ultimately lead to robust classifiers with reproducible performance across datasets. Simulation results based on multiple breast cancer datasets show that the proposed inference method identifies better pathway markers that can predict breast cancer metastasis with higher accuracy. Moreover, the identified pathway markers can lead to better classifiers with more consistent classification performance across independent datasets.
BibTeX
@article{khunlertgit2013identification,
  author = {Navadon Khunlertgit and Byung-Jun Yoon},
  title = {{Identification of Robust Pathway Markers for Cancer through Rank-Based Pathway Activity Inference}},
  journal = {Adv. Bioinformatics},
  year = {2013},
  doi = {10.1155/2013/618461},
}
Figure