GENE REGULATORY NETWORK RECONSTRUCTION BY BAYESIAN INTEGRATION OF PRIOR KNOWLEDGE AND/OR DIFFERENT EXPERIMENTAL CONDITIONS

Adriano Velasque Werhli, Dirk HusmeierJournal of Bioinformatics and Computational Biology 2008 · DOI

TL;DR

This paper presents a novel Bayesian approach to reconstruct gene regulatory networks by integrating prior biological knowledge and data from multiple experimental conditions. By using a Markov Chain Monte Carlo (MCMC) framework, the method learns to balance information from existing knowledge and experimental data, and can also identify and account for differences in active pathways across various experimental settings.

What Problem Does This Paper Address?

Reconstructing gene regulatory networks (GRNs) from experimental data, such as microarray data, is a fundamental challenge in systems biology. Existing methods often struggle to accurately capture the complex interactions within these networks. A key limitation is the difficulty in effectively integrating prior biological knowledge (e.g., from databases like KEGG) and in handling data from different experimental conditions, which might reveal distinct active sub-pathways. Simply merging data from different conditions can obscure these crucial differences, while analyzing them separately can lead to a loss of statistical power. This paper aims to bridge this gap by developing a unified Bayesian framework that can leverage both prior knowledge and multi-condition data to improve GRN reconstruction.

Methods

The paper introduces two main methodological contributions:

  1. Bayesian Integration of Prior Knowledge: This approach builds upon previous work by expressing prior knowledge as an "energy function." This energy function is then used to define a prior distribution over possible network structures (specifically, directed acyclic graphs or DAGs) in the form of a Gibbs distribution. The key innovation is a Markov Chain Monte Carlo (MCMC) scheme that simultaneously samples both the network structures and the hyperparameters controlling the influence of prior knowledge versus experimental data. This allows the model to automatically learn how to best trade off information from these two sources.
  2. Bayesian Coupling Scheme for Multi-Condition Data: To address the challenge of integrating data from different experimental conditions, the authors propose a "Bayesian coupling scheme." This method learns separate networks for each experimental condition but "couples" them through a shared "hypernetwork." This hypernetwork acts as a soft constraint, encouraging the individual networks to be similar while still allowing for differences that reflect variations in active sub-pathways. The strength of this coupling is controlled by hyperparameters that are also learned from the data. This approach offers a compromise between treating data monolithically and analyzing each dataset in isolation.

Datasets

The proposed methods were evaluated using two types of data:

  • Synthetic Data:
    • Linear Gaussian data generated from a known "gold standard" Raf signaling pathway network.
  • Nonlinear synthetic data generated using the sigma-pi formalism and the Netbuilder software, also based on the Raf signaling pathway.
  • A modified version of the Raf network where specific edges were removed to simulate different active pathways.
  • A purely random data set to test robustness against corrupted data.
  • Real Data:
    • Intracellular multicolor flow cytometry experiments measuring protein concentrations related to the Raf pathway (from Sachs et al.). This data was subsampled to mimic the sample sizes typical of microarray experiments.

Key Insights and Ideas

  • Prior Knowledge as an Energy Function: Expressing biological prior knowledge in terms of energy functions provides a principled way to quantify its influence on network structure inference.
  • Simultaneous Sampling of Networks and Hyperparameters: The MCMC approach that jointly samples network structures and hyperparameters is crucial for automatically learning the optimal balance between prior knowledge and data.
  • Hypernetwork for Multi-Condition Data: The concept of a hypernetwork elegantly captures the idea that while biological systems adapt to different conditions, they retain a common underlying architecture. This allows for learning condition-specific differences while leveraging shared information.
  • Detecting Corrupted/Modified Data: The Bayesian coupling scheme proved effective in identifying and down-weighting the influence of corrupted (pure noise) or modified (different active pathways) data sets.
  • The Importance of Absence of Edges: For the Raf pathway data, prior knowledge about the absence of edges seemed to be more informative than knowledge about the presence of edges, suggesting that careful formulation of prior knowledge is critical.

Key Results

  • Improved Reconstruction with Prior Knowledge: The Bayesian integration of prior knowledge (from KEGG) significantly improved network reconstruction accuracy (measured by ROC curves for both directed and undirected graph evaluation) compared to methods using only data or only prior knowledge.
  • Effective Multi-Condition Learning: The Bayesian coupling scheme consistently outperformed monolithic (merging all data) and uncoupled (analyzing data separately) approaches in network reconstruction accuracy across synthetic and real data. The improvement was most pronounced on synthetic data where regulatory parameters varied across conditions.
  • Identification of Corrupted Data: The coupling scheme successfully identified and suppressed the influence of a purely random data set, as indicated by a significantly lower hyperparameter value associated with it.
  • Distinguishing Modified Networks: Data generated from a modified network structure was also distinguished, with its associated hyperparameter falling between those of the random and true networks.
  • Convergence Challenges: While the methods showed promise, MCMC convergence issues were noted, particularly for the coupled scheme, leading to less stable results in some simulations.

Limitations and Future Work

  • MCMC Convergence: The paper acknowledges that convergence and mixing of the MCMC simulations can be a challenge, especially for the coupled scheme. This can lead to less reliable results and requires careful diagnostics. Future work could explore more sophisticated sampling schemes, such as simulated annealing or parallel MCMC chains, and novel proposal moves that swap substructures between networks.
  • Prior Knowledge Representation: The authors suggest that more flexible ways of presenting prior knowledge, such as incorporating subgroups of nodes and edges, could be explored.
  • Gold Standard Accuracy: For real data, discrepancies between predicted and gold standard networks might stem from inaccuracies in the gold standard itself or the presence of unmodeled feedback loops.
  • Threshold Selection: The decision of whether an edge is present or absent in prior knowledge depends on a threshold, which was arbitrarily set. Sampling this threshold parameter from the posterior distribution could potentially lead to clearer performance enhancements.

Summary updated

Abstract

There have been various attempts to improve the reconstruction of gene regulatory networks from microarray data by the systematic integration of biological prior knowledge. Our approach is based on pioneering work by Imoto et al. where the prior knowledge is expressed in terms of energy functions, from which a prior distribution over network structures is obtained in the form of a Gibbs distribution. The hyperparameters of this distribution represent the weights associated with the prior knowledge relative to the data. We have derived and tested a Markov chain Monte Carlo (MCMC) scheme for sampling networks and hyperparameters simultaneously from the posterior distribution, thereby automatically learning how to trade off information from the prior knowledge and the data. We have extended this approach to a Bayesian coupling scheme for learning gene regulatory networks from a combination of related data sets, which were obtained under different experimental conditions and are therefore potentially associated with different active subpathways. The proposed coupling scheme is a compromise between (1) learning networks from the different subsets separately, whereby no information between the different experiments is shared; and (2) learning networks from a monolithic fusion of the individual data sets, which does not provide any mechanism for uncovering differences between the network structures associated with the different experimental conditions. We have assessed the viability of all proposed methods on data related to the Raf signaling pathway, generated both synthetically and in cytometry experiments.
BibTeX
@article{werhli2008gene,
  author = {Adriano Velasque Werhli and Dirk Husmeier},
  title = {{GENE REGULATORY NETWORK RECONSTRUCTION BY BAYESIAN INTEGRATION OF PRIOR KNOWLEDGE AND/OR DIFFERENT EXPERIMENTAL CONDITIONS}},
  journal = {Journal of Bioinformatics and Computational Biology},
  year = {2008},
  doi = {10.1142/s0219720008003539},
}
Figure