Modelling slowly changing dynamic gene-regulatory networks

TL;DR

This paper introduces a novel method for estimating slowly changing dynamic gene-regulatory networks from high-dimensional time-course data, using penalized likelihood and convex optimization to identify stable interactions over time.

Contribution

The authors develop a new model that effectively captures gradual changes in dynamic gene networks, suitable for high-dimensional genomic datasets, with a novel optimization approach.

Findings

GL_Δ performs well in simulation studies.

The method successfully applied to T-cell gene data.

Efficient estimation of sparse, slowly changing networks.

Abstract

Dynamic gene-regulatory networks are complex since the number of potential components involved in the system is very large. Estimating dynamic networks is an important task because they compromise valuable information about interactions among genes. Graphical models are a powerful class of models to estimate conditional independence among random variables, e.g. interactions in dynamic systems. Indeed, these interactions tend to vary over time. However, the literature has been focused on static networks, which can only reveal overall structures. Time-course experiments are performed in order to tease out significant changes in networks. It is typically reasonable to assume that changes in genomic networks are few because systems in biology tend to be stable. We introduce a new model for estimating slowly changes in dynamic gene-regulatory networks which is suitable for a high-dimensional dataset, e.g. time-course genomic data. Our method is based on i) the penalized likelihood with $\ell_1$-norm, ii) the penalized differences between conditional independence elements across time points and iii) the heuristic search strategy to find optimal smoothing parameters. We implement a set of linear constraints necessary to estimate sparse graphs and penalized changing in dynamic networks. These constraints are not in the linear form. For this reason, we introduce slack variables to re-write our problem into a standard convex optimization problem subject to equality linear constraints. We show that GL$_\Delta$ performs well in a simulation study. Finally, we apply the proposed model to a time-course genetic dataset T-cell.