Inferring the dynamics of gene regulatory networks via optimized recurrent neural network and dynamic Bayesian network

Inferring gene regulatory networks (GRNs) from time-course expression data is a major challenge in systems biology and comprehensive understanding of its dynamics is difficult. Most temporal inference methods for the dynamics of GRNs assume linear dependencies among genes but this strong assumption of linearity among genes does not truly represent the dynamics of the GRNs which are inherently nonlinear. Other parametric and non-parametric methods for modeling nonlinear dynamical systems such as the S-systems and causal identification structure (CSI) have been proposed for modeling time-course nonlinearities in GRNs; however, these methods are statistically inefficient and analytically intractable especially in high dimensions. To overcome these problems, we propose an algorithm based on optimized recurrent neural network (RNN) and dynamic Bayesian (DBN) network called RNN-DBN. The inference algorithm for our DBN is based on nonlinear state space Elman recurrent neural network. Results on Drosophila Melanogaster nonlinear time-course benchmark dataset shows our method outperforms the G1DBN inference method based on linear model assumptions. The algorithm is further applied to time-course ovarian cancer dataset. The results show that the expression levels of three of five significant hub genes (flap structure-specific endonuclease 1, kinesin family member 11 and CDC6 cell division cycle 6 homolog (S. cerevisiae)) were decreased by oxaliplatin, but remained constant with cisplatin platinum drugs. These may therefore be potential drug candidates for ovarian cancer.