Bayesian biclustering of gene expression data

Jiajun Gu, Jun S. LiuBMC 2008 · DOI

TL;DR

This paper introduces a novel Bayesian biclustering model (BBC) for analyzing gene expression data, offering a statistically rigorous approach to identify co-regulated genes across specific experimental conditions. The BBC model excels at handling complex data, including missing values, and demonstrates superior performance compared to existing methods in discovering biologically meaningful patterns in yeast gene expression.

What Problem Does This Paper Address?

Gene expression data, often generated by microarrays, provides a snapshot of which genes are active under various conditions. Traditional clustering methods group genes or conditions separately, but this can be limiting. Genes might participate in multiple biological pathways, and a specific biological process is usually active only under a subset of genes and experimental conditions. Biclustering aims to address this by simultaneously identifying groups of genes and conditions that exhibit similar expression patterns. However, many existing biclustering algorithms are heuristic or lack a strong statistical foundation, making them sensitive to noise, prone to overfitting, or unable to handle overlapping clusters effectively. This paper tackles the challenge of developing a robust, statistically sound biclustering method that can uncover biologically significant, potentially overlapping, gene-condition patterns in gene expression data.

Methods

The core of this paper is the Bayesian Biclustering (BBC) model. It's built upon a hierarchical Bayesian framework, treating biclustering as a statistical inference problem.

Here's a breakdown of the key methodological aspects:

  • Model Formulation: The BBC model assumes that gene expression values within a bicluster can be explained by a main effect of the cluster, gene-specific effects, condition-specific effects, and noise. A crucial innovation is how it handles multiple biclusters: instead of allowing arbitrary overlaps (which can lead to non-identifiability), it constrains overlaps to be in "one direction" – either genes can belong to multiple clusters but conditions only to one, or vice-versa. This is a significant departure from models like the "plaid model" which can result in extensive overlap.
  • Bayesian Inference: The model uses Gibbs sampling, a Markov Chain Monte Carlo (MCMC) method, for statistical inference. This allows them to draw samples from the posterior distribution of model parameters, including indicator variables that determine gene and condition membership in clusters.
  • Handling Missing Data: A major advantage highlighted is the BBC model's inherent ability to handle missing data through Monte Carlo imputation within the Gibbs sampling iterations. This is a common issue in microarray data and a significant practical benefit.
  • Model Selection: The Bayesian Information Criterion (BIC) is used to determine the optimal number of clusters (K) for a given dataset, providing a principled way to select the model complexity.
  • Normalization: The paper also investigates the impact of different data normalization techniques on biclustering results, proposing two new methods: Interquartile Range Normalization (IQRN) and Smallest Quartile Range Normalization (SQRN), which are shown to be more robust to outliers than standard methods.

Datasets

The authors evaluate their BBC algorithm using two main types of datasets:

  • Simulated Datasets:
    • Simple simulated data: Used to demonstrate the model's ability to identify clusters with varying degrees of overlap (common genes, common conditions, or both).
  • Plaid model simulated data: Used to compare the BBC algorithm against other established biclustering methods (ISA, SAMBA, CC, OPSMs, Plaid) in scenarios with single and multiple overlapping clusters.
  • Biologically realistic simulated data: This dataset was generated using a biochemistry model that simulates gene regulation by transcription factors, incorporating realistic noise from the Leukemia expression dataset. This benchmark is designed to test how well algorithms capture biologically motivated patterns.
  • Real Yeast Expression Data: A combined dataset from environmental stress and cell cycle experiments was used to demonstrate the practical utility of the BBC algorithm in discovering biologically significant biclusters.

Key Insights and Ideas

  • Statistical Rigor in Biclustering: The paper champions a model-based Bayesian approach, moving beyond heuristic methods. This provides a principled way to assess uncertainty and discover patterns with statistical confidence.
  • Controlled Overlap: The novel constraint on bicluster overlap (one-directional) is a key innovation that addresses a practical limitation of previous models, leading to more interpretable and less artificial results.
  • Robustness to Noise and Missing Data: The BBC model's ability to handle missing data and its demonstrated robustness in simulations with realistic noise are significant practical advantages for real-world biological data.
  • Biological Validation: The application to yeast data and the subsequent enrichment analysis (gene functions, experimental conditions, transcription factor binding sites) strongly suggest that the discovered biclusters are not just statistical artifacts but represent genuine biological relationships.

Key Results

  • Simulation Performance: The BBC algorithm consistently outperformed other tested biclustering methods (ISA, SAMBA, CC, OPSMs, Plaid) in terms of sensitivity and specificity across various simulated scenarios, especially when dealing with multiple overlapping clusters and biologically realistic data.
  • Yeast Data Analysis: When applied to yeast expression data, the BBC algorithm identified 57 biclusters. A significant proportion of these biclusters showed strong biological validation:
  • 36 biclusters had significant gene function enrichments.
  • 26 had significant transcription factor binding site (TFBS) enrichments.
  • 51 had significant experimental condition category enrichments.
  • 22 biclusters exhibited enrichment across all three categories.
  • Normalization Impact: The study highlighted that normalization is crucial. IQRN and SQRN were found to be robust and effective for the BBC model, outperforming other methods.

Limitations and Future Work

  • Pre-defined K: The BBC model requires setting the total number of clusters (K) beforehand. While BIC is used for model selection, determining the "optimal" K can still be a challenge in practice.
  • Computational Cost: Gibbs sampling, while powerful, can be computationally intensive, especially for very large datasets.
  • Future Directions: The authors suggest extending the BBC model to incorporate other types of biological data, such as promoter sequence information directly into the model, and integrating it into studies of gene transcription networks. The potential for handling more complex regulatory models is also implied.

Summary updated

Abstract

The BBC algorithm is shown to be a robust model-based biclustering method that can discover biologically significant gene-condition clusters in microarray data. The BBC model can easily handle missing data via Monte Carlo imputation and has the potential to be extended to integrated study of gene transcription networks.
BibTeX
@inproceedings{gu2008bayesian,
  author = {Jiajun Gu and Jun S. Liu},
  title = {{Bayesian biclustering of gene expression data}},
  booktitle = {BMC},
  year = {2008},
  doi = {10.1186/1471-2164-9-s1-s4},
}
Figure