Bayesian Functional Data Clustering for Temporal Microarray Data

Ping Ma, Wenxuan Zhong, Yang Feng, Jun S. LiuInternational Journal of Plant Genomics 2008 · DOI

TL;DR

This paper introduces a novel Bayesian clustering method for temporal gene expression data. It uses a mixed-effects smoothing-spline model to capture gene expression patterns over time, automatically determines the optimal number of clusters, and gracefully handles missing data, offering a more robust approach than traditional methods.

What Problem Does This Paper Address?

Understanding gene function often relies on observing how their expression levels change over time or under different conditions. Microarray technology allows us to measure thousands of genes simultaneously, generating complex temporal datasets. A key challenge is to group genes that exhibit similar expression patterns, as these co-regulated genes are likely involved in the same biological processes. Traditional clustering methods like K-means and hierarchical clustering often fail to adequately account for the inherent temporal correlation in gene expression data. While multivariate Gaussian models can address correlation, they don't inherently consider the time order. Autoregression models are an option but often assume stationarity, which is rarely true for biological data. This paper aims to fill this gap by providing a flexible and statistically rigorous method for clustering temporal gene expression profiles.

Methods

The core of this paper is a Bayesian clustering approach built upon a mixed-effect smoothing-spline model. For each cluster, the average gene expression profile is modeled as a smoothing spline, allowing for flexible curve fitting. Individual gene variations within a cluster are captured by a parametric random effect. This approach is framed within a Reproducing Kernel Hilbert Space framework, enabling the representation of the mean expression curve as a linear combination of basis functions.

A key innovation is the use of a Gibbs sampler to efficiently explore the posterior distribution of model parameters. The method also incorporates a Bayesian Information Criterion (BIC) to automatically determine the optimal number of clusters, removing the need for manual selection. Furthermore, the model is designed to handle missing data inherently, a common issue in microarray experiments. When dealing with multiple clusters, a Gaussian mixture model is employed, with each component representing a cluster. The paper also discusses the use of a Dirichlet prior for cluster proportions and outlines the iterative conditional sampling steps for the Gibbs sampler.

Datasets

The paper evaluates its method on a microarray dataset of budding yeast (Saccharomyces cerevisiae). This dataset tracks gene expression changes under varying oxygen levels during aerobic conditions. The authors specifically mention analyzing a subset of 2388 differentially expressed genes identified by Lai et al. [11]. No new benchmark datasets are introduced.

Key Insights and Ideas

  • Temporal Correlation is Crucial: The paper strongly emphasizes that ignoring the time-ordered nature of gene expression data leads to suboptimal clustering and interpretation.
  • Bayesian Framework for Robustness: The Bayesian approach provides a natural way to incorporate prior knowledge and quantify uncertainty in the clustering results, making them more reliable.
  • Automatic Cluster Number Selection: The use of BIC to automatically determine the number of clusters is a significant practical advantage, simplifying the analysis pipeline.
  • Handling Missing Data: The inherent ability of the model to handle missing data is a major practical benefit for real-world microarray datasets.
  • Biologically Meaningful Clusters: The successful application to yeast data, resulting in clusters enriched for known biological functions, validates the method's biological relevance.

Key Results

The proposed Bayesian clustering algorithm was applied to a yeast microarray dataset, identifying 31 clusters that yielded the smallest BIC score. A functional enrichment analysis (using FunSpec) revealed that 23 out of these 31 clusters were significantly enriched for known biological functions. The paper presents examples of three clusters (A, B, and C) with their estimated mean expression profiles, demonstrating how these patterns correlate with specific biological roles:

  • Cluster A (40 genes): Upregulated as oxygen decreases, enriched for stress response, cell rescue, defense, and oxidoreductase functions.
  • Cluster B (92 genes): Rapid initial drop followed by a gradual increase in expression, enriched for protein synthesis and ribosome biogenesis.
  • Cluster C (68 genes): Involved in galactose fermentation, carbon utilization, and carbohydrate metabolism, showing an initial upregulation followed by downregulation as oxygen depletes.

Limitations and Future Work

The authors acknowledge that certain prior distributions and hyperparameters need to be input by the user. However, they found the clustering results to be robust to variations in these inputs. They suggest that their Bayesian clustering algorithm serves as a platform that can be extended to incorporate more biological knowledge. An open-source R code implementation of their method is available.

Summary updated

Abstract

We propose a Bayesian procedure to cluster temporal gene expression microarray profiles, based on a mixed-effect smoothing-spline model, and design a Gibbs sampler to sample from the desired posterior distribution. Our method can determine the cluster number automatically based on the Bayesian information criterion, and handle missing data easily. When applied to a microarray dataset on the budding yeast, our clustering algorithm provides biologically meaningful gene clusters according to a functional enrichment analysis.
BibTeX
@article{ma2008bayesian,
  author = {Ping Ma and Wenxuan Zhong and Yang Feng and Jun S. Liu},
  title = {{Bayesian Functional Data Clustering for Temporal Microarray Data}},
  journal = {International Journal of Plant Genomics},
  year = {2008},
  doi = {10.1155/2008/231897},
}
Figure