Developing neural network algorithms to improve Gene Regulatory Network (GRN) inference over the state-of-the-art algorithms

Abstract

Gene regulatory networks (GRNs) entail complex and nonlinear interactions which are captured by systems of ordinary differential equations (ODEs). Repressilator, the first artificial GRN to be engineered, is taken as the GRN of this study. For the repressilator, different parameter values for the ODEs lead to vastly different dynamics. This is tackled by inferring a posterior probability distribution as a part of Bayesian inference, incorporating prior beliefs and observed data. The state-of-the-art algorithms for this are slow, owing to the high dimensionality of the models. Moreover, they need to be re-run whenever new observations are available, keeping the average inference time per observed dataset high. A faster alternative with acceptable accuracy is explored via a neural network-based inference pipeline. Results from this pipeline demonstrate a reduction in the average inference time, or amortisation, below that obtained from the current approaches. This holds promise for a deeper understanding of natural GRNs, while enabling robust engineering of artificial ones.