On the selection of appropriate distances for gene expression data clustering

Pablo A. Jaskowiak, Ricardo J. G. B. Campello, Ivan G. CostaBMC Bioinformatics 2014 · DOI

TL;DR

Choosing the right distance metric is crucial for effective gene expression data clustering, but it's often overlooked. This paper systematically evaluates 15 different distance measures across four common clustering algorithms on 52 real-world datasets, providing much-needed guidance for researchers in cancer and gene time-series analysis.

What Problem Does This Paper Address?

Gene expression data analysis often involves clustering to identify patterns, such as subtypes of cancer or groups of genes with similar temporal behavior. While many clustering algorithms exist, the choice of distance metric—how we quantify similarity between data points—is a critical parameter that significantly impacts the quality of the clustering results. However, there's a lack of clear guidance on which distance metric to use for gene expression data, especially considering the distinct characteristics of different applications like cancer sample clustering versus gene time-series clustering. This paper aims to fill this gap by providing a comprehensive empirical evaluation of various distance measures.

Methods

The authors evaluated 15 distance measures, categorized into three groups:

  • Correlation coefficients: Pearson, Goodman-Kruskal, Kendall, Spearman, Rank-Magnitude, and Weighted Goodman-Kruskal.
  • Traditional proximity measures: Cosine, Euclidean, Manhattan, and Supreme (Minkowski family).
  • Time-series specific measures: Jackknife, Short Time-Series Dissimilarity (STS), Local Shape-based Similarity (LSS), YS1, and YR1.

These distances were tested with four common clustering methods: k-medoids, and three hierarchical methods (Complete-Linkage, Average-Linkage, and Single-Linkage). The evaluation was performed across three scenarios: fixed number of clusters, variable number of clusters, and estimated number of clusters (using Silhouette). For cancer datasets, results were validated using the Adjusted Rand Index (ARI) against known ground truth. For gene time-series data, where ground truth is often unavailable, a biological validation using Gene Ontology (GO) enrichment analysis was employed. Robustness to noise was also assessed.

Datasets

The study utilized a total of 52 gene expression datasets:

  • Cancer datasets: 35 datasets from a benchmark set, comprising 14 cDNA and 21 Affymetrix microarray datasets.
  • Gene time-series datasets: 17 datasets from a benchmark set, all derived from yeast (Saccharomyces cerevisiae) and produced using cDNA microarrays.

Key Insights and Ideas

  • Application-dependent distances: The "best" distance metric is not universal; it strongly depends on whether you're clustering cancer samples or gene time-series data.
  • Distance matters more than you think: Even with a suitable clustering algorithm, the choice of distance can make the difference between meaningful and poor clustering outcomes.
  • Time-series complexity: Gene time-series clustering is more intricate, suggesting a need for better coupling between distance metric biases and clustering algorithm biases due to the problem's smaller dimensionality.
  • Specialized measures can underperform: Some distances specifically designed for gene time-series (like LSS and STS) performed poorly, suggesting their formulation might not be optimal for the evaluated scenarios.
  • Robustness vs. Accuracy Trade-off: While some measures (like Spearman) showed better robustness to noise, others (like Rank-Magnitude) offered a better balance between accuracy and robustness.

Key Results

  • Cancer Sample Clustering: For cancer data, Jackknife and Pearson generally provided the best accuracy (ARI), with Cosine also performing well. Rank-Magnitude emerged as a strong contender, offering a good balance of accuracy and robustness to noise. Single-Linkage consistently performed poorly across all distances.
  • Gene Time-Series Clustering: For gene time-series data, YS1 and YR1 (which incorporate correlation with other time-series information) showed excellent results with Complete-Linkage. Jackknife also performed well with k-medoids and Average-Linkage.
  • Robustness to Noise: Spearman demonstrated the best overall robustness to noise, while EUC showed the worst. Rank-Magnitude offered a good balance.
  • Clustering Method Performance: k-medoids was generally recommended for cancer data. For hierarchical methods, Average-Linkage was preferred over Complete-Linkage. Single-Linkage was not recommended for any scenario.

Limitations and Future Work

  • Computational Cost: Jackknife, while effective, has a quadratic computational complexity, which can be prohibitive for large datasets.
  • Time-Series Specific Measures: The underperformance of some specialized time-series measures (LSS, STS) suggests further research is needed to refine or develop more effective metrics for this domain.
  • Future Work: The authors suggest that their work provides a compendium of alternatives and guidelines, implying that practitioners should use these findings to select appropriate distances for their specific gene expression clustering tasks. The paper also highlights the importance of considering the interplay between distance measures and clustering methods.

Summary updated

BibTeX
@article{jaskowiak2014selection,
  author = {Pablo A. Jaskowiak and Ricardo J. G. B. Campello and Ivan G. Costa},
  title = {{On the selection of appropriate distances for gene expression data clustering}},
  journal = {BMC Bioinformatics},
  year = {2014},
  doi = {10.1186/1471-2105-15-s2-s2},
}
Figure