Inference of Boolean Networks Using Sensitivity Regularization
© Wenbin Liu et al. 2008
Received: 22 November 2007
Accepted: 9 April 2008
Published: 25 May 2008
The inference of genetic regulatory networks from global measurements of gene expressions is an important problem in computational biology. Recent studies suggest that such dynamical molecular systems are poised at a critical phase transition between an ordered and a disordered phase, affording the ability to balance stability and adaptability while coordinating complex macroscopic behavior. We investigate whether incorporating this dynamical system-wide property as an assumption in the inference process is beneficial in terms of reducing the inference error of the designed network. Using Boolean networks, for which there are well-defined notions of ordered, critical, and chaotic dynamical regimes as well as well-studied inference procedures, we analyze the expected inference error relative to deviations in the networks' dynamical regimes from the assumption of criticality. We demonstrate that taking criticality into account via a penalty term in the inference procedure improves the accuracy of prediction both in terms of state transitions and network wiring, particularly for small sample sizes.
The execution of various developmental and physiological processes in cells is carried out by complex biomolecular systems. Such systems are dynamic in that they are able to change states in response to environmental cues and exhibit multiple steady states, which define different cellular functional states or cell types.
The massively parallel dynamics of complex molecular networks furnish the cell with the ability to process information from its environment and mount appropriate responses. To be able to stably execute cellular functions in a variable environment while being responsive to specific changes in the environment, such as the activation of immune cells upon exposure to pathogens or their components, the cell needs to strike a balance between robustness and adaptability.
Theoretical considerations and computational studies suggest that many types of complex dynamical systems can indeed strike such an optimal balance, under a variety of criteria, when they are operating close to a critical phase transition between an ordered and a disordered dynamical regime [1–3]. There is also accumulating evidence that living systems, as manifestations of their underlying networks of molecular interactions, are poised at the critical boundary between an organized and a disorganized state, indicating that cellular information processing is optimal in the critical regime, affording the cell with the ability to exhibit complex coordinated macroscopic behavior [4–8]. Studies of human brain oscillations , computer network traffic and the Internet , financial markets , forest fires , neuronal networks supporting our senses , and biological macroevolution have also revealed critical dynamics .
A key goal in systems biology research is to characterize the molecular mechanisms governing specific cellular behaviors and processes. This typically entails selecting a model class for representing the system structure and state dynamics, followed by the application of computational or statistical inference procedures for revealing the model structure from measurement data . Multiple types of data can be potentially used for elucidating the structure of molecular networks, such as transcriptional regulatory networks, including genome wide transcriptional profiling with DNA microarrays or other high-throughput technologies, chromatin immunoprecipitation-on-chip (ChIP-on-chip) for identifying DNA sequences occupied by specific DNA binding proteins, computational predictions of transcription factor binding sites based on promoter sequence analysis, and other sources of evidence for molecular interactions . The inference of genetic networks is particularly challenging in the face of small sample sizes, particularly because the number of variables in the system (e.g., genes) typically greatly outnumbers the number of observations. Thus, estimates of the errors of a given model, which themselves are determined from the measurement data, can be highly variable and untrustworthy.
Any prior knowledge about the network structure, architecture, or dynamical rules is likely to improve the accuracy of the inference, especially in a small sample size scenario. If biological networks are indeed critical, a key question is whether this knowledge can be used to improve the inference of network structure and dynamics from measurements. We investigated this question using the class of Boolean networks as models of genetic regulatory networks.
Boolean networks and the more general class of probabilistic Boolean networks are popular approaches for modeling genetic networks, as these model classes capture multivariate nonlinear relationships between the elements of the system and are capable of exhibiting complex dynamics [51619–21]. Boolean network models have been constructed for a number of biomolecular systems, including the yeast cell cycle , mammalian cell cycle , Drosophila segment polarity network , regulatory networks of E. coli metabolism , and Arabidopsis flower morphogenesis [27–29].
At the same time, these model classes have been studied extensively regarding the relationships between their structure and dynamics. Particularly in the case of Boolean networks, dynamical phase transitions from the ordered to the disordered regime and the critical phase transition boundary have been characterized analytically for random ensembles of networks [30–34]. This makes these models attractive for investigating the relationships between structure and dynamics .
In particular, the so-called average sensitivity was shown to be an order parameter for Boolean networks . The average sensitivity, which can be computed directly from the Boolean functions specifying the update rules (i.e., state transitions) of the network, measures the average response of the system to a minimal transient perturbation and is equivalent to the Lyapunov exponent . There have been a number of approaches for inferring Boolean and probabilistic Boolean networks from gene expression measurement data [202137–44].
We address the relationship between the dynamical regime of a network, as measured by the average sensitivity, and the inference of the network from data. We study whether the assumption of criticality, embedded in the inference objective function as a penalty term, improves the inference of Boolean network models. We find that for small sample sizes the assumption is beneficial, while for large sample sizes, the performance gain decreases gradually with increasing sample size. This is the kind of behavior that one hopes for when using penalty terms.
This paper is organized as follows. In Section 2, we give a brief definition of Boolean Networks and the concept of sensitivity. Then in Section 3, three measures used in this paper to evaluate the performance of the predicted networks are introduced, and a theoretical analysis of the relationship between the expected error and the sensitivity deviation is presented. Based on this analysis, an objective function is proposed to be used for the inference process in Section 4, while the simulation results are presented in Section 5.
2. Background and Definitions
2.1. Boolean Networks
A Boolean network is defined by a set of nodes , and a set of Boolean functions , . Each node represents the expression state of the gene , where means that the gene is OFF, and means it is ON. Each Boolean function with specific input nodes is assigned to node and is used to update its value. Under the synchronous updating scheme, all genes are updated simultaneously according to their corresponding update functions. The network's state at time is represented by a vector and, in the absence of noise, the system transitions from state to state in a deterministic manner.
Sensitivity is in fact a global dynamical parameter that captures how a one-bit perturbation spreads throughout the network and its expectation under the random Boolean network model is equivalent to the well-known phase transition curve .
3. Error Analysis
3.1. Performance Measures
There are several ways to measure the performance of an inference method by comparing the state transitions, wiring, or sensitivities with the original network. In this paper, we will use three measures that are described below.
3.1.1. The State Transition Error
where denotes the normalized error of the predicted function . Additionally, and are extended such that they are functions of all the variables (instead of variables) by adding fictitious (i.e., dummy) variables.
3.1.2. The Receiver Operating Characteristic (ROC)
where TP and FP represent true positive and false positive instances, respectively, while and represent the total positive and negative instances, respectively. We will use the ROC distributions to evaluate the accuracy of "wiring" (i.e., the specific input nodes assigned to each node) for the inferred network.
3.1.3. The Sensitivity Error
3.2. Analysis of Expected Error
In this case, will have a linear relationship with . This indicates that will surely introduce additional error . Our simulations indicate that the inference method we use (best-fit, see below) yields a network with in most cases.
4. Inference Method
To infer a Boolean network, for each target node we need to apply some optimization criterion to each set of input variables and Boolean function on those input variables and then choose the variable set and corresponding Boolean function that minimizes the objective function. The first step in our proposed procedure is to find variable sets and Boolean functions that provide good target prediction. Based upon time-series observations, given a target node and an input-node vector the best predictor, , minimizes the error, , among all possible predictors . Finding the best predictor for a given node means finding the minimal error among all Boolean functions over all input-variable combinations. We consider three variable combinations. Since we will optimize via an objective function containing a sensitivity penalty term, we will select a collection of input-variable sets and select the minimal-error Boolean function over each input-variable set. This is accomplished by using the plugin (resubstitution) estimate of the error , which is given by the number of times in the data divided by the number of times, the pair is observed in the data. This procedure is equivalent to the best-fit extension method . We make use of an efficient algorithm for solving the best-fit extension problem and finding all functions having error smaller than a given threshold . We then select the four best variable sets and corresponding Boolean functions as candidates. The limitation of four candidates is based on computational considerations; in principle, there is no such limitation.
Compute the resubstitution errors of all Boolean functions using the full sample data set.
Select the four input-variable sets whose corresponding functions possess the lowest .632 bootstrap estimates.
This procedure is the same as the one used in  to evaluate the impact of different error estimators on feature-set ranking. It was demonstrated there that the bootstrap tends to outperform cross-validation methods in choosing good feature sets. While it was also observed that bolstering tends to be slightly better than bootstrap, bolstering cannot be applied in discrete settings, so it is not a viable option in our context.
where represents the bootstrap-estimated error of the previously selected Boolean function and is the sensitivity error. The first item represents the prediction error, while the second represents the "structural error" associated with general network dynamics. Our hypothesis is that a better inference should have a small error in both state transition and sensitivity, and consequently, the value of its objective function should be minimal. Of the four input-variable sets selected via prediction error for a target node, we use the one with minimal objective function for network construction.
5. Simulation Results
All simulations are performed for random Boolean networks with and . For a given BN, we randomly generate pairs of input and output states. We also consider the effect of noise, with 5% noise added to the output states of each gene by flipping its value with probability 0.05.
From the perspective of network inference, performance is best characterized via a distance function between the ground-truth network and the inferred network, more specifically, by the expected distance between the ground-truth and inferred network as estimated by applying the inference to a random sample of ground-truth networks . In our case, we have chosen the normalized state-transition error as the distance between the networks.
When the actual sensitivity is 1, the method helps for small samples and the performances become close for large samples, analogous to Figure 6. When the true network deviates from the modelling assumption, , the proposed method helps for small samples and results in some loss of performance for large samples. This kind of behavior is what one would expect with an objective function that augments the error. In effect, the sensitivity is a penalty term in the objective function that is there to impose constraint on the optimization. In our case, when the true sensitivity is not equal to 1, the sensitivity constraint yields smaller sensitivity error than the best-fit method in small sample situations, while the sensitivity error of the best-fit method is smaller for large samples. In sum, the constraint is beneficial for small samples.
Sensitivity is a global structural parameter of a network which captures the network's operating dynamic behavior: ordered, critical, or chaotic. Recent evidence suggests that living systems operate at the critical phase transition between ordered and chaotic regimes. In this paper, we have proposed a method to use this dynamic information to improve the inference of Boolean networks from observations of input-output relationships. First, we have analyzed the relationship between the expectation of the error and the deviation of sensitivity, showing that these quantities are strongly correlated with each other. Based on this observation, an objective function is proposed to refine the inference approach based on the best-fit method. The simulation results demonstrate that the proposed method can improve the predicted results both in terms of state transitions, sensitivity, and network wiring. The improvement is particularly evident in small sample size settings. As the sample size increases, the performance of both methods becomes similar. In practice, where one does not know the sensitivity of the true network, we have assumed it to be 1, the critical value, and investigated inference performance relative to its robustness to the true sensitivity deviating from 1. For small samples, the kind we are interested in when using such a penalty approach, the proposed method continues to outperform the best-fit method.
For practical applications, one can apply an optimization strategy, such as genetic algorithms, to attain suboptimal solutions instead of the brute force searching strategy used in this paper. As the final chosen function for each gene generally lies within the top three candidates in our simulations, one can just select from a few top candidate functions for each gene instead of using all of the possible candidates. Finally, it should be noted that the ideas presented here could also be incorporated into other inference methods, such as the ones in .
Support from NIGMS GM072855 (I.S.), P50-GM076547 (I.S.), NSF CCF-0514644 (E.D.), NCI R01 CA-104620 (E.D.), NSFC under no. 60403002 (W.L.), NSF of Zhejiang province under nos. Y106654 and Y405553 (W.L.) is gratefully acknowledged.
- Langton CG: Computation at the edge of chaos: phase transitions and emergent computation. Physica D 1990,42(1–3):12-37.MathSciNetView ArticleGoogle Scholar
- Krawitz P, Shmulevich I: Basin entropy in Boolean network ensembles. Physical Review Letters 2007,98(15):-4.View ArticleGoogle Scholar
- Packard NH: Adaptation towards the edge of chaos. In Dynamic Patterns in Complex Systems. Edited by: Kelso JAS, Mandell AJ, Shlesinger MF. World Scientific, Singapore; 1988:293-301.Google Scholar
- Rämö P, Kesseli J, Yli-Harja O: Perturbation avalanches and criticality in gene regulatory networks. Journal of Theoretical Biology 2006,242(1):164-170. 10.1016/j.jtbi.2006.02.011MathSciNetView ArticleMATHGoogle Scholar
- Kauffman SA: The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press, New York, NY, USA; 1993.Google Scholar
- Shmulevich I, Kauffman SA, Aldana M: Eukaryotic cells are dynamically ordered or critical but not chaotic. Proceedings of the National Academy of Sciences of the United States of America 2005,102(38):13439-13444. 10.1073/pnas.0506771102View ArticleGoogle Scholar
- Serra R, Villani M, Semeria A: Genetic network models and statistical properties of gene expression data in knock-out experiments. Journal of Theoretical Biology 2004,227(1):149-157. 10.1016/j.jtbi.2003.10.018MathSciNetView ArticleGoogle Scholar
- Nykter M, Price ND, Aldana M, et al.: Gene expression dynamics in the macrophage exhibit criticality. Proceedings of the National Academy of Sciences of the United States of America 2008,105(6):1897-1900. 10.1073/pnas.0711525105View ArticleGoogle Scholar
- Linkenkaer-Hansen K, Nikouline VV, Palva JM, Ilmoniemi RJ: Long-range temporal correlations and scaling behavior in human brain oscillations. Journal of Neuroscience 2001,21(4):1370-1377.Google Scholar
- Valverde S, Solé RV: Self-organized critical traffic in parallel computer networks. Physica A 2002,312(3-4):636-648. 10.1016/S0378-4371(02)00872-5View ArticleMATHGoogle Scholar
- Fukuda K, Takayasu H, Takayasu M: Origin of critical behavior in Ethernet traffic. Physica A 2000,287(1-2):289-301. 10.1016/S0378-4371(00)00452-0View ArticleGoogle Scholar
- Lux T, Marchesi M: Scaling and criticality in a stochastic multi-agent model of a financial market. Nature 1999,397(6719):498-500. 10.1038/17290View ArticleGoogle Scholar
- Malamud BD, Morein G, Turcotte DL: Forest fires: an example of self-organized critical behavior. Science 1998,281(5384):1840-1842.View ArticleGoogle Scholar
- Kinouchi O, Copelli M: Optimal dynamical range of excitable networks at criticality. Nature Physics 2006,2(5):348-352. 10.1038/nphys289View ArticleGoogle Scholar
- Sneppen K, Bak P, Flyvbjerg H, Jensen MH: Evolution as a self-organized critical phenomenon. Proceedings of the National Academy of Sciences of the United States of America 1995,92(11):5209-5213. 10.1073/pnas.92.11.5209View ArticleGoogle Scholar
- Shmulevich I, Dougherty ER Princeton Series in Applied Mathematics. In Genomic Signal Processing. Princeton University Press, Princeton, NJ, USA; 2007.View ArticleGoogle Scholar
- Ramsey SA, Klemm SL, Zak DE, et al.: Uncovering a macrophage transcriptional program by integrating evidence from motif scanning and expression dynamics. PLoS Computational Biology 2008,4(3, e1000021):1-25.MathSciNetView ArticleGoogle Scholar
- Lähdesmäki H, Rust AG, Shmulevich I: Probabilistic inference of transcription factor binding from multiple data sources. PLoS ONE 2008,3(3):e1820. 10.1371/journal.pone.0001820View ArticleGoogle Scholar
- Kauffman S: Homeostasis and differentiation in random genetic control networks. Nature 1969,224(5215):177-178. 10.1038/224177a0View ArticleGoogle Scholar
- Shmulevich I, Dougherty ER, Kim S, Zhang W: Probabilistic Boolean networks: a rule-based uncertainty model for gene regulatory networks. Bioinformatics 2002,18(2):261-274. 10.1093/bioinformatics/18.2.261View ArticleGoogle Scholar
- Shmulevich I, Dougherty ER, Zhang W: From Boolean to probabilistic Boolean networks as models of genetic regulatory networks. Proceedings of the IEEE 2002,90(11):1778-1792. 10.1109/JPROC.2002.804686View ArticleGoogle Scholar
- Li F, Long T, Lu Y, Ouyang Q, Tang C: The yeast cell-cycle network is robustly designed. Proceedings of the National Academy of Sciences of the United States of America 2004,101(14):4781-4786. 10.1073/pnas.0305937101View ArticleGoogle Scholar
- Davidich MI, Bornholdt S: Boolean network model predicts cell cycle sequence of fission yeast. PLoS ONE 2008,3(2):e1672. 10.1371/journal.pone.0001672View ArticleGoogle Scholar
- Fauré A, Naldi A, Chaouiya C, Thieffry D: Dynamical analysis of a generic Boolean model for the control of the mammalian cell cycle. Bioinformatics 2006,22(14):e124-e131. 10.1093/bioinformatics/btl210View ArticleGoogle Scholar
- Albert R, Othmer HG: The topology of the regulatory interactions predicts the expression pattern of the segment polarity genes in Drosophila melanogaster . Journal of Theoretical Biology 2003,223(1):1-18. 10.1016/S0022-5193(03)00035-3MathSciNetView ArticleGoogle Scholar
- Samal A, Jain S: The regulatory network of E. coli metabolism as a Boolean dynamical system exhibits both homeostasis and flexibility of response. BMC Systems Biology 2008., 2, article 21:Google Scholar
- Espinosa-Soto C, Padilla-Longoria P, Alvarez-Buylla ER: A gene regulatory network model for cell-fate determination during Arabidopsis thaliana flower development that is robust and recovers experimental gene expression profiles. The Plant Cell 2004,16(11):2923-2939. 10.1105/tpc.104.021725View ArticleGoogle Scholar
- Mendoza L, Alvarez-Buylla ER: Dynamics of the genetic regulatory network for Arabidopsis thaliana flower morphogenesis. Journal of Theoretical Biology 1998,193(2):307-319. 10.1006/jtbi.1998.0701View ArticleGoogle Scholar
- Mendoza L, Thieffry D, Alvarez-Buylla ER: Genetic control of flower morphogenesis in Arabidopsis thaliana : a logical analysis. Bioinformatics 1999,15(7-8):593-606.View ArticleGoogle Scholar
- Derrida B, Pomeau Y: Random networks of automata: a simple annealed approximation. Europhysics Letters 1986,1(2):45-49. 10.1209/0295-5075/1/2/001View ArticleGoogle Scholar
- Shmulevich I, Kauffman SA: Activities and sensitivities in Boolean network models. Physical Review Letters 2004,93(4):-4.View ArticleGoogle Scholar
- Aldana M, Coppersmith S, Kadanoff LP: Boolean dynamics with random couplings. In Perspectives and Problems in Nonlinear Science. Edited by: Kaplan E, Marsden JE, Sreenivasan KR. Springer, New York, NY, USA; 2002:23-89.Google Scholar
- Luque B, Solé RV: Lyapunov exponents in random Boolean networks. Physica A 2000,284(1–4):33-45.View ArticleGoogle Scholar
- Luque B, Solé RV: Phase transitions in random networks: simple analytic determination of critical points. Physical Review E 1997,55(1):257-260. 10.1103/PhysRevE.55.257View ArticleGoogle Scholar
- Nochomovitz YD, Li H: Highly designable phenotypes and mutational buffers emerge from a systematic mapping between network topology and dynamic output. Proceedings of the National Academy of Sciences of the United States of America 2006,103(11):4180-4185. 10.1073/pnas.0507032103View ArticleGoogle Scholar
- Nykter M, Price ND, Larjo A, et al.: Critical networks exhibit maximal information diversity in structure-dynamics relationships. Physical Review Letters 2008,100(5):-4.View ArticleGoogle Scholar
- Lähdesmäki H, Shmulevich I, Yli-Harja O: On learning gene regulatory networks under the Boolean network model. Machine Learning 2003,52(1-2):147-167.View ArticleMATHGoogle Scholar
- Zhou X, Wang X, Dougherty ER: Gene prediction using multinomial probit regression with Bayesian gene selection. EURASIP Journal on Applied Signal Processing 2004,2004(1):115-124. 10.1155/S1110865704309157MathSciNetView ArticleMATHGoogle Scholar
- Zhao W, Serpedin E, Dougherty ER: Inferring gene regulatory networks from time series data using the minimum description length principle. Bioinformatics 2006,22(17):2129-2135. 10.1093/bioinformatics/btl364View ArticleGoogle Scholar
- Marshall S, Yu L, Xiao Y, Dougherty ER: Inference of a probabilistic Boolean network from a single observed temporal sequence. EURASIP Journal on Bioinformatics and Systems Biology 2007, 2007:-15.Google Scholar
- Pal R, Ivanov I, Datta A, Bittner ML, Dougherty ER: Generating Boolean networks with a prescribed attractor structure. Bioinformatics 2005,21(21):4021-4025. 10.1093/bioinformatics/bti664View ArticleGoogle Scholar
- Tabus I, Astola J: On the use of MDL principle in gene expression prediction. EURASIP Journal on Applied Signal Processing 2001,2001(4):297-303. 10.1155/S1110865701000270MathSciNetView ArticleMATHGoogle Scholar
- Akutsu T, Miyano S, Kuhara S: Inferring qualitative relations in genetic networks and metabolic pathways. Bioinformatics 2000,16(8):727-734. 10.1093/bioinformatics/16.8.727View ArticleGoogle Scholar
- Akutsu T, Miyano S, Kuhara S: Algorithms for identifying Boolean networks and related biological networks based on matrix multiplication and fingerprint function. Journal of Computational Biology 2000,7(3-4):331-343. 10.1089/106652700750050817View ArticleGoogle Scholar
- Boros E, Ibaraki T, Makino K: Error-free and best-fit extensions of partially defined Boolean functions. Information and Computation 1998,140(2):254-283. 10.1006/inco.1997.2687MathSciNetView ArticleMATHGoogle Scholar
- Efron B: Estimating the error rate of a prediction rule: improvement on cross-validation. Journal of the American Statistical Association 1983,78(382):316-331. 10.2307/2288636MathSciNetView ArticleMATHGoogle Scholar
- Efron B, Tibshirani RJ Monographs on Statistics and Applied Probability. In An Introduction to the Bootstrap. Volume 57. Chapman & Hall, New York, NY, USA; 1993.View ArticleGoogle Scholar
- Sima C, Attoor S, Brag-Neto U, Lowey J, Suh E, Dougherty ER: Impact of error estimation on feature selection. Pattern Recognition 2005,38(12):2472-2482. 10.1016/j.patcog.2005.03.026View ArticleGoogle Scholar
- Dougherty ER: Validation of inference procedures for gene regulatory networks. Current Genomics 2007,8(6):351-359. 10.2174/138920207783406505MathSciNetView ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.