XCluSim: a visual analytics tool for interactively comparing multiple clustering results of bioinformatics data

Sehi L'Yi, Bongkyung Ko, DongHwa Shin, Young-Joon Cho, Jaeyong Lee, Bohyoung Kim, Jinwook SeoBMC Bioinformatics 2015 · DOI

TL;DR

Comparing the results of multiple clustering algorithms is crucial but cognitively demanding for bioinformatics researchers. XCluSim is a novel visual analytics tool that addresses this by providing interactive overviews and detailed visualizations to help users compare and explore numerous clustering outcomes efficiently.

What Problem Does This Paper Address?

In bioinformatics, cluster analysis is a standard technique for identifying patterns in large datasets, especially with the advent of high-throughput technologies like next-generation sequencing. However, selecting the "best" clustering method and its parameters is a significant challenge. Researchers often run multiple algorithms with varying parameters, generating a multitude of clustering results. Comparing these results to assess their quality and identify meaningful biological insights is a laborious and cognitively taxing process. Existing tools often struggle with scalability, lack dedicated visualizations for different clustering types, or don't offer intuitive ways to compare multiple results simultaneously. This paper aims to bridge this gap by providing a more effective and scalable solution for comparing and exploring these complex outputs.

Methods

XCluSim employs a visual analytics approach guided by the "overview first, zoom and filter, details-on-demand" mantra. Its core components include:

  • Parameter Information View: A tabular overview that displays the clustering algorithms and their parameter settings used, along with bar charts indicating the number of results generated for each combination.
  • Force-Directed Layout (FDL) Overview: Visualizes the similarity between clustering results as nodes in a graph, where closer nodes and thicker edges represent higher similarity (measured by F-measure). Pie charts within nodes show cluster sizes.
  • Dendrogram Overview: Presents clustering result similarities in a familiar dendrogram format, offering a space-efficient alternative to the FDL overview.
  • Enhanced Parallel Sets View: A key interactive component that visualizes the concordance and discordance among selected clustering results. It uses stacked bars for results and ribbons for "stable groups" (items clustered together across multiple results). Users can interactively reorder rows to adjust similarity views and rearrange bars/bands to reduce clutter.
  • Tabular List View: Provides detailed numerical information for selected clustering results, offering group-by (stable group or cluster) and heatmap modes.
  • Color Encoding (Tree Colors): A consistent color scheme is applied across all visualizations, derived from a hierarchical clustering of all clusters. This helps users intuitively assess cluster similarity.
  • Dedicated Visualizations for Algorithm Types: XCluSim offers specific visualizations tailored to hierarchical (dendrogram+heatmap), partitional (force-directed layout), and density-based (bar chart + parallel coordinate plot) clustering algorithms.

Datasets

The paper describes two case studies using real-world bioinformatics datasets:

  1. Cryptococcus neoformans var. grubii H99 strain: A dataset containing normalized expression levels of 6,980 genes, used to investigate the role of ferroxidase and the effect of fluconazole.
  2. Escherichia coli: A dataset of gene expression profiles for 169 genes, used to investigate tryptophan metabolism and identify biological relations.

No new benchmarks were introduced; existing datasets were used for evaluation.

Key Insights and Ideas

  • Scalable Comparison is Key: The core challenge is comparing many clustering results. XCluSim's overviews (FDL, dendrogram) are designed for this scalability, allowing users to grasp overall relationships before diving deeper.
  • "Stable Groups" as a Unifying Concept: Identifying items consistently clustered across multiple runs is a powerful way to find robust biological signals. XCluSim's enhanced parallel sets view makes these stable groups visually prominent.
  • Tailored Visualizations Matter: Different clustering algorithms produce results with distinct characteristics. Providing specialized visualizations (e.g., dendrogram for hierarchical, FDL for partitional) makes it easier to interpret these specific outputs.
  • Gestalt Principles Inform Design: The authors explicitly leverage Gestalt principles (similarity, proximity, connectedness, enclosure) to categorize and select effective visualization techniques, providing a structured approach to visualization design for clustering.
  • Interactivity is Crucial for Exploration: Beyond static views, XCluSim's interactive features (reordering, filtering, zooming, brushing, linking) are essential for users to actively explore the data and uncover insights.

Key Results

The paper presents qualitative results from two case studies with a bioinformatician. The participant provided positive feedback, highlighting XCluSim's ability to:

  • Identify stable groups across multiple clustering results.
  • Select and run diverse algorithms and compare them interactively.
  • Quickly shift attention to more interesting sets of results.
  • Confirm biological hypotheses (e.g., identifying genes related to azole resistance and tryptophan metabolism pathways).

The tool successfully helped the participant find optimal clustering parameters and validate biological relationships, demonstrating its practical utility.

Limitations and Future Work

  • Filtering Limitations: The participant noted that filtering of items was only available at the data manipulation step, suggesting it would be beneficial to allow interactive filtering within the visualization components themselves.
  • Handling Un-clustered Items: Currently, XCluSim treats un-clustered items as a single cluster. The authors plan to improve this by using different textures in parallel sets to distinguish them from normal clusters, making the visualization more informative.
  • Incorporating Validity Measures: The authors suggest that representing internal cluster validity indices (e.g., using grayscale intensity of stable group bands) would further enhance the reliability of XCluSim's insights.
  • Expanding Algorithm Support: While XCluSim integrates Weka algorithms and allows importing results, further integration of more algorithms could be beneficial.

Summary updated

BibTeX
@article{lyi2015xclusim,
  author = {Sehi L'Yi and Bongkyung Ko and DongHwa Shin and Young-Joon Cho and Jaeyong Lee and Bohyoung Kim and Jinwook Seo},
  title = {{XCluSim: a visual analytics tool for interactively comparing multiple clustering results of bioinformatics data}},
  journal = {BMC Bioinformatics},
  year = {2015},
  doi = {10.1186/1471-2105-16-s11-s5},
}
Figure