Modelling Transcriptional Regulation with a Mixture of Factor Analyzers and Variational Bayesian Expectation Maximization

Kuang Lin, Dirk HusmeierEURASIP 2009 · DOI

TL;DR

This paper introduces a novel computational model, MFA-VBEM, that more accurately captures the intricate process of gene transcriptional regulation by accounting for how transcription factors (TFs) cooperate in complexes. This approach leads to improved reconstruction of TF activities, more biologically meaningful gene clusters, and more accurate prediction of regulatory networks compared to existing methods.

What Problem Does This Paper Address?

Understanding how genes are turned on and off (transcriptional regulation) is fundamental to systems biology. This process is orchestrated by transcription factors (TFs) that bind to DNA. However, current models often struggle to capture the full complexity:

  • TF Cooperation: TFs rarely act alone; they often form functional complexes to regulate genes, especially in higher organisms. Existing methods often treat TFs independently.
  • TF Activity vs. Expression: A TF's gene expression level doesn't always reflect its actual activity in regulating genes, as post-translational modifications can play a crucial role.
  • Data Noise: Gene expression and TF binding data are inherently noisy, making accurate modeling challenging.
  • Model Complexity: Developing models that are both accurate and computationally manageable is a significant hurdle.

This paper aims to bridge these gaps by proposing a model that explicitly accounts for TF cooperation, integrates gene expression and TF binding data, and uses a robust Bayesian framework for inference.

Methods

The core of this work is the Mixture of Factor Analyzers (MFA) model combined with Variational Bayesian Expectation Maximization (VBEM) for inference.

  • Mixture of Factor Analyzers (MFA): This probabilistic model treats each "component" of the mixture as a TF complex or module. Genes are then assigned to these modules, implying that they are regulated by a specific complex. This directly addresses the issue of TF cooperation.
  • Variational Bayesian Expectation Maximization (VBEM): This is a powerful Bayesian inference technique used to approximate complex posterior distributions. It allows for:
  • Learning Model Parameters: Estimating the strengths of TF-gene interactions and TF activities.
  • Model Selection: Automatically determining the optimal number of TF complexes (modules) present in the data, avoiding the need for manual tuning.
  • Handling Noise: The Bayesian framework inherently accounts for uncertainty and noise in the data.

The authors also propose a "birth and death" move scheme within the VBEM algorithm to efficiently explore different model complexities (number of TF modules) and avoid getting stuck in local optima.

Datasets

The authors evaluated their method using two types of datasets:

  • Synthetic Data: Generated to simulate gene expression and TF binding profiles under varying noise levels and lengths. This allowed for a direct comparison against known ground truth.
  • Real Data: Gene expression and TF binding data from Saccharomyces cerevisiae (yeast). This included:
  • ChIP-chip binding data from Lee et al. [34].
  • Microarray gene expression data from Spellman et al. [35] and Gasch et al. [36].
  • Transcriptional regulatory association data from YeastTract [38].

Key Insights and Ideas

  • TF Complexes are Key: The study strongly suggests that modeling TFs as cooperating complexes, rather than individual entities, significantly improves the accuracy of transcriptional regulation models.
  • Data Integration is Crucial: Combining gene expression data with TF binding information provides a more comprehensive view and leads to better predictions than using either data type alone.
  • Bayesian Inference for Robustness: The Bayesian approach, particularly with VBEM, offers robustness against noise and provides a principled way to perform model selection (determining the number of TF modules).
  • Parsimony Improves Inference: The MFA model, by capturing TF cooperation, is more parsimonious (uses fewer parameters) than traditional FA models, leading to more stable and less overfitted inference.

Key Results

  • Activity Profile Reconstruction: MFA-VBEM consistently outperformed Partial Least Squares (PLS) and Bayesian Factor Analysis (BFA-Gibbs) in reconstructing TF module activity profiles, achieving higher Pearson correlation coefficients, especially with longer expression profiles and lower noise levels.
  • Gene Clustering: When applied to gene clustering, MFA-VBEM produced clusters that were significantly enriched for known biological functions (Gene Ontology terms) more often than K-means, hierarchical clustering, COSA, and Plaid models, particularly when integrating TF binding data.
  • Regulatory Network Reconstruction:
    • On synthetic data, MFA-VBEM showed improved accuracy in predicting TF-gene interactions, especially with longer gene expression profiles and higher noise.
  • On yeast data, MFA-VBEM achieved a substantially higher Area Under the ROC Curve (AUC) for network reconstruction compared to BFA-Gibbs and Maximum Likelihood Factor Analysis (FA). It also demonstrated a better ability to identify TF complexes.

Limitations and Future Work

  • Computational Complexity: While MFA-VBEM is more scalable than some Gibbs sampling approaches, computational cost can still be a factor for very large datasets.
  • Model Selection for Number of Modules: The authors used a "birth and death" move scheme for model selection. They suggest that a more principled approach could involve treating the number of active transcriptional modules as a random variable itself, potentially using a Dirichlet process prior.
  • Prior Distributions: The choice of prior distributions can influence the results, and further exploration of these choices might be beneficial.
  • Future Directions: The authors propose exploring non-parametric Bayesian modeling (e.g., Dirichlet process priors) for inferring the number of TF modules and investigating the application of these ideas to other biological problems.

Summary updated

BibTeX
@inproceedings{lin2009modelling,
  author = {Kuang Lin and Dirk Husmeier},
  title = {{Modelling Transcriptional Regulation with a Mixture of Factor Analyzers and Variational Bayesian Expectation Maximization}},
  booktitle = {EURASIP},
  year = {2009},
  doi = {10.1155/2009/601068},
}
Figure