## The Factor Graph Network Model for Biological Systems (2005)

Venue: | Proc. of RECOMB 2005 |

Citations: | 5 - 1 self |

### BibTeX

@INPROCEEDINGS{Gat-viks05thefactor,

author = {Irit Gat-viks and Amos Tanay and Daniela Raijman and Ron Shamir},

title = {The Factor Graph Network Model for Biological Systems},

booktitle = {Proc. of RECOMB 2005},

year = {2005},

pages = {31--47},

publisher = {Springer}

}

### OpenURL

### Abstract

Abstract. We introduce an extended computational framework for studying biological systems. Our approach combines formalization of existing qualitative models that are in wide but informal use today, with probabilistic modeling and integration of high throughput experimental data. Using our methods, it is possible to interpret genomewide measurements in the context of prior knowledge on the system, to assign statistical meaning to the accuracy of such knowledge and to learn refined models with improved fit to the experiments. Our model is represented as a probabilistic factor graph and the framework accommodates partial measurements of diverse biological elements. We develop methods for inference and learning in the model. We compare the performance of standard inference algorithms and tailor-made ones and show that hidden variables can be reliably inferred even in the presence of feedback loops and complex logic. We develop a formulation for the learning problem in our model which is based on deterministic hypothesis testing, and show how to derive p-values for learned model features. We test our methodology and algorithms on both simulated and real yeast data. In particular, we use our method to study the response of S. cerevisiae to hyperosmotic shock, and explore uncharacterized logical relations between important regulators in the system. 1

### Citations

7487 |
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...the edges of GR and additional edges from each Xi to the corresponding Yi. This can be easily seen from (2) by noting that in the acyclic case Z = 1 (the proof is as in Bayesian networks theory, e.g. =-=[18]-=-). When the model contains loops, the situation gets more complicated. For example, we note that according to the FGN model, P rM (Xi|P ai) does not necessarily equals the original beliefs θ(Xi, P ai)... |

1285 | Factor graphs and the sum-product algorithm
- Kschischang, Frey, et al.
- 2001
(Show Context)
Citation Context ...ntities (e.g., mRNAs, proteins, metabolites), and directly accommodates undelayed feedback loops which are essential in many biological systems. We formalize our model as a probabilistic factor graph =-=[14]-=-, and show how it can be developed naturally from basic assumptions on the biochemical processes underlying the regulatory network and on the information we have on it. Having established our methodol... |

810 | Using Bayesian Networks to Analyze Expression Data
- Friedman, Linial, et al.
- 2000
(Show Context)
Citation Context ...s and forms, and handle highly complex systems and extensive datasets. Recent studies on computational models for biological networks have attempted denovo reconstruction of a network on genes (e.g., =-=[6]-=-), used prior knowledge on network topology (e.g., [9, 11]), or combined transcription factor location and sequence datasto learn a clustered model for the genome-wide behavior of the system [1, 23, 2... |

607 |
G.F.(1990).The Computational Complexity of Probabilistic Inference Using Bayesian Belief Networks
- Cooper
- 1990
(Show Context)
Citation Context ... estimate the level of a certain metabolic intermediate. Solving instances of the inference problem is a pre-requisite for learning refined models. Inference in graphical models is an NP hard problem =-=[4]-=- that was extensively studied. A popular class of algorithms [28] uses an approximation of the posterior distribution assuming certain decomposition over independent variables or clusters of variables... |

442 | Constructing free-energy approximations and generalized belief propagation algorithms
- Yedidia, Freeman, et al.
- 2005
(Show Context)
Citation Context ... instances of the inference problem is a pre-requisite for learning refined models. Inference in graphical models is an NP hard problem [4] that was extensively studied. A popular class of algorithms =-=[28]-=- uses an approximation of the posterior distribution assuming certain decomposition over independent variables or clusters of variables. Algorithms from this class include loopy belief propagation (LB... |

336 |
Module networks: identifying regulatory modules and their conditionspecific regulators from gene expression data
- Segal, Shapira, et al.
- 2003
(Show Context)
Citation Context ...e.g., [6]), used prior knowledge on network topology (e.g., [9, 11]), or combined transcription factor location and sequence datasto learn a clustered model for the genome-wide behavior of the system =-=[1, 23, 2]-=-. Other studies built detailed models manually, utilizing existing biological knowledge [3, 5] but lacked computational methods for model reassessment in light of additional evidence. In this study we... |

229 | Learning the structure of dynamic probabilistic networks
- Friedman, Murphy, et al.
- 1998
(Show Context)
Citation Context ...s. Our model is unique in its handling of steady state feedback loops. It can be extended to handle slower temporal processes in a way analogous to the construction of dynamic Bayesian networks (DBN) =-=[7, 24]-=- from steady state Bayesian networks. As in DBNs, the algorithms for inference and learning can be naturally generalized from the steady state model to the dynamic model. Another possible extension is... |

187 |
Computational discovery of gene modules and regulatory networks
- Bar-Joseph, Gerber, et al.
- 2003
(Show Context)
Citation Context ...e.g., [6]), used prior knowledge on network topology (e.g., [9, 11]), or combined transcription factor location and sequence datasto learn a clustered model for the genome-wide behavior of the system =-=[1, 23, 2]-=-. Other studies built detailed models manually, utilizing existing biological knowledge [3, 5] but lacked computational methods for model reassessment in light of additional evidence. In this study we... |

151 | Introduction to Monte Carlo Methods
- MacKay
- 1999
(Show Context)
Citation Context ...on the performance of several inference algorithms. We implemented a Gibbs sampler, the LBP algorithm, and a modes-based instantiation inference algorithm. Our Gibbs sampler is a naive MCMC algorithm =-=[15]-=- that performs a random walk over the space of model states, based on sampling from local distributions. In our model, sampling is done only for the X variables (unobserved sensors do not affect other... |

111 |
Predicting gene expression from sequence
- Beer, Tavazoie
- 2004
(Show Context)
Citation Context ...e.g., [6]), used prior knowledge on network topology (e.g., [9, 11]), or combined transcription factor location and sequence datasto learn a clustered model for the genome-wide behavior of the system =-=[1, 23, 2]-=-. Other studies built detailed models manually, utilizing existing biological knowledge [3, 5] but lacked computational methods for model reassessment in light of additional evidence. In this study we... |

109 | Osmotic stress signaling and osmoadaptation in yeasts
- Hohmann
- 2002
(Show Context)
Citation Context ...kinase Hog1, and an array of transcription factors that coordinate a complex process of adaptation by transient growth repression and modifications to glycerol metabolism, membrane structure and more =-=[10]-=-. We have constructed an FGN model that represents known regulatory relations in the HOG system (Figure 6A) and used it to study the transcriptional program following treatment by variable levels of K... |

104 | Combining location and expression data for principled discovery of genetic regulatory network models - Hartemink, Gifford, et al. - 2002 |

84 |
Assigning numbers to the arrows: Parameterizing a gene regulatory network by using accurate expression kinetics
- Ronen, Rosenberg, et al.
(Show Context)
Citation Context ...the steady state model to the dynamic model. Another possible extension is the consideration of other classes of regulation functions (for example, we can consider continuous or ranked function as in =-=[26, 22, 12, 16]-=-).sFig. 1. An overview of the factor graph network model. A) Knowledge on the logical regulation functions is formalized as conditional probabilities. B) Continuous measurements and logical states are... |

76 | Tutorial on variational approximation methods
- Jaakkola
- 2001
(Show Context)
Citation Context ...is (see Figure 2) indicates that for estimation of single variable posteriors, LBP outperforms the other two algorithms (and also the mean field algorithm and a simple clustered variational algorithm =-=[13]-=-, data not shown). Also, when the prior is strong, MI provides reasonable accuracy. However, the distributionsA B C D E F G Fig. 2. Performance of different inference algorithms on a simulated model. ... |

65 |
Integrating high-throughput and computational data elucidates bacterial networks. Nature
- Covert, Knight, et al.
- 2004
(Show Context)
Citation Context ...ctor location and sequence datasto learn a clustered model for the genome-wide behavior of the system [1, 23, 2]. Other studies built detailed models manually, utilizing existing biological knowledge =-=[3, 5]-=- but lacked computational methods for model reassessment in light of additional evidence. In this study we describe a new algorithmic framework for representing biological knowledge and integrating it... |

61 | Inferring quantitative models of regulatory networks from expression data. Bioinformatics
- Nachman, Regev, et al.
- 2004
(Show Context)
Citation Context ...the steady state model to the dynamic model. Another possible extension is the consideration of other classes of regulation functions (for example, we can consider continuous or ranked function as in =-=[26, 22, 12, 16]-=-).sFig. 1. An overview of the factor graph network model. A) Knowledge on the logical regulation functions is formalized as conditional probabilities. B) Continuous measurements and logical states are... |

60 | Combining microarrays and biological knowledge for estimating gene networks via Bayesian networks
- Imoto, Higuchi, et al.
- 2004
(Show Context)
Citation Context ...ensive datasets. Recent studies on computational models for biological networks have attempted denovo reconstruction of a network on genes (e.g., [6]), used prior knowledge on network topology (e.g., =-=[9, 11]-=-), or combined transcription factor location and sequence datasto learn a clustered model for the genome-wide behavior of the system [1, 23, 2]. Other studies built detailed models manually, utilizing... |

39 | Bayesian network and nonparametric heteroscedastic regression for nonlinear modeling of genetic network
- Imoto, Kim, et al.
- 2003
(Show Context)
Citation Context ...the steady state model to the dynamic model. Another possible extension is the consideration of other classes of regulation functions (for example, we can consider continuous or ranked function as in =-=[26, 22, 12, 16]-=-).sFig. 1. An overview of the factor graph network model. A) Knowledge on the logical regulation functions is formalized as conditional probabilities. B) Continuous measurements and logical states are... |

39 | Evaluating functional network inference using simulations of complex biological systems
- Smith, Jarvis, et al.
- 2002
(Show Context)
Citation Context ...s. Our model is unique in its handling of steady state feedback loops. It can be extended to handle slower temporal processes in a way analogous to the construction of dynamic Bayesian networks (DBN) =-=[7, 24]-=- from steady state Bayesian networks. As in DBNs, the algorithms for inference and learning can be naturally generalized from the steady state model to the dynamic model. Another possible extension is... |

37 | Physical network models
- Yeang
- 2004
(Show Context)
Citation Context ...hs are widely used probabilistic graphical models that were originally applied to coding/decoding problems. Recently, factor graphs were used in computational biology, although in a different context =-=[27]-=-. When the dependency graph GR is acyclic, our FGN model is equivalent to a Bayesian network on the variables Xi and Yi, constructed using the edges of GR and additional edges from each Xi to the corr... |

36 |
Unique and redundant roles for HOG MAPK pathway components as revealed by wholegenome expression analysis
- O’Rourke, Herskowitz
- 2004
(Show Context)
Citation Context ....9 in all reported experiments. We collected published experimental data on each of the models. The data consisted of 23 conditions (cf. [8]) for the lysine model and 129 conditions for the HOG model =-=[17]-=-. Differential measurements from cDNA microarrays were transformed into absolute values as described in [8].sA B Fig. 3. Accuracy of learning regulation functions. Each figure is a ROC curve (X-axis: ... |

32 |
The transcriptional response of Saccharomyces cerevisiae to osmotic shock. Hot1p and Msn2p/Msn4p are required for the induction of subsets of high osmolarity glycerol pathway-dependent genes
- Rep, Krantz, et al.
- 2000
(Show Context)
Citation Context ...itivity, by slowing down the production of some central components in it. A second group of discrepancies involves genes that are targets of the Hog1 downstream regulators Sko1, Hot1, Msn1 and Msn2,4 =-=[19, 21, 20]-=-. In many cases, the literature does not specify the logical relations among the regulators and each of their regulatees, and this lack of knowledge is manifested as discrepancies. We thus used our mo... |

32 |
Osmotic stress-induced gene expression in Saccharomyces cerevisiae requires Msn1p and the novel nuclear factor
- Rep, Reiser, et al.
- 1999
(Show Context)
Citation Context ...itivity, by slowing down the production of some central components in it. A second group of discrepancies involves genes that are targets of the Hog1 downstream regulators Sko1, Hot1, Msn1 and Msn2,4 =-=[19, 21, 20]-=-. In many cases, the literature does not specify the logical relations among the regulators and each of their regulatees, and this lack of knowledge is manifested as discrepancies. We thus used our mo... |

21 |
Repressors and upstream repressing sequences of the stress-regulated ENA1 gene in Saccharomyces cerevisiae: bZIP protein Sko1p confers HOG-dependent osmotic regulation
- Proft, Serrano
- 1999
(Show Context)
Citation Context ...itivity, by slowing down the production of some central components in it. A second group of discrepancies involves genes that are targets of the Hog1 downstream regulators Sko1, Hot1, Msn1 and Msn2,4 =-=[19, 21, 20]-=-. In many cases, the literature does not specify the logical relations among the regulators and each of their regulatees, and this lack of knowledge is manifested as discrepancies. We thus used our mo... |

20 | Computational expansion of genetic networks
- Tanay, Shamir
- 2001
(Show Context)
Citation Context ...nalysis that predicts the system’s behavior in various conditions. Most importantly, our framework allows the learning of a refined model with improved fit to the experimental data. In previous works =-=[25, 8]-=- we have introduced the notions of model refinement and expansion and studied it when applied to discrete deterministic models. Here we study these problems in the more general settings of probabilist... |

11 | Modeling and analysis of heterogeneous regulation in biological networks
- Gat-Viks, Tanay, et al.
(Show Context)
Citation Context ...nalysis that predicts the system’s behavior in various conditions. Most importantly, our framework allows the learning of a refined model with improved fit to the experimental data. In previous works =-=[25, 8]-=- we have introduced the notions of model refinement and expansion and studied it when applied to discrete deterministic models. Here we study these problems in the more general settings of probabilist... |

6 | Modeling transcription programs: inferring binding site activity and dose-response model optimization
- Tanay, Shamir
- 2003
(Show Context)
Citation Context |

4 |
et al. Kinetic analysis of a molecular model of the budding yeast cell cycle. Molecular Biology of the Cell
- Chen
(Show Context)
Citation Context ...ctor location and sequence datasto learn a clustered model for the genome-wide behavior of the system [1, 23, 2]. Other studies built detailed models manually, utilizing existing biological knowledge =-=[3, 5]-=- but lacked computational methods for model reassessment in light of additional evidence. In this study we describe a new algorithmic framework for representing biological knowledge and integrating it... |