Inference of gene regulatory networks incorporating multi-source biological knowledge via a state space model with L1 regularization.

Takanori Hasegawa,Masao Nagasaki,Satoru Miyano,Seiya Imoto,Rui Yamaguchi,Frank Emmert-Streib

doi:10.1371/journal.pone.0105942

Abstract

Comprehensive understanding of gene regulatory networks (GRNs) is a major challenge in the field of systems biology. Currently, there are two main approaches in GRN analysis using time-course observation data, namely an ordinary differential equation (ODE)-based approach and a statistical model-based approach. The ODE-based approach can generate complex dynamics of GRNs according to biologically validated nonlinear models. However, it cannot be applied to ten or more genes to simultaneously estimate system dynamics and regulatory relationships due to the computational difficulties. The statistical model-based approach uses highly abstract models to simply describe biological systems and to infer relationships among several hundreds of genes from the data. However, the high abstraction generates false regulations that are not permitted biologically. Thus, when dealing with several tens of genes of which the relationships are partially known, a method that can infer regulatory relationships based on a model with low abstraction and that can emulate the dynamics of ODE-based models while incorporating prior knowledge is urgently required. To accomplish this, we propose a method for inference of GRNs using a state space representation of a vector auto-regressive (VAR) model with L1 regularization. This method can estimate the dynamic behavior of genes based on linear time-series modeling constructed from an ODE-based model and can infer the regulatory structure among several tens of genes maximizing prediction ability for the observational data. Furthermore, the method is capable of incorporating various types of existing biological knowledge, e.g., drug kinetics and literature-recorded pathways. The effectiveness of the proposed method is shown through a comparison of simulation studies with several previous methods. For an application example, we evaluated mRNA expression profiles over time upon corticosteroid stimulation in rats, thus incorporating corticosteroid kinetics/dynamics, literature-recorded pathways and transcription factor (TF) information.

Highlights

Transcriptional regulation, which is controlled by several factors, plays essential roles to sustain complex biological systems in cells
Comparison Results To show the effectiveness of the proposed method, we compared it with other gene regulatory networks (GRNs) inference methods, i.e., a state space model (SSM) [21,63], a general vector auto-regressive (VAR) model using the LARS-LASSO algorithm [30,61], GeneNet [31,32] based on an empirical graphical Gaussian model (GGM), dynamic Bayesian networks using first order conditional dependencies (G1DBN) [33], GLASSO [34] based on sparse GGM and the mutual information-based network inference algorithms: ARACNE [3], CLR [35] and MRNET [36]
We proposed a novel method for inference of gene regulatory networks incorporating existing biological knowledge and time-course observation data

Summary

Introduction

Transcriptional regulation, which is controlled by several factors, plays essential roles to sustain complex biological systems in cells. One strategy to elucidate transcriptional regulation using observational data is to apply an ordinary differential equation (ODE)-based approach, which can represent the dynamic behavior of biomolecular reactions based on biologically reliable models, e.g., the Michaelis-Menten equation [4] or the S-system [5], which are described by differential equations. This approach can recapitulate the complex dynamic behavior of biological systems [6,7]. Under this statistically efficient paradigm [15], this approach cannot be applied to ten or more genes to infer regulatory structures if the missing information is extensive [10]

Methods

Results

Conclusion