Methods for Inferring Gene Regulatory Networks from Time Series Expression Data

Nguyen, Phan

doi:https://doi.org/10.21985/n2-rspz-7k79

Work

Methods for Inferring Gene Regulatory Networks from Time Series Expression Data

Public

Download PDF

Gene regulatory networks (GRNs) are important abstractions of the complex regulatory interplays between genes, proteins, metabolites, and other molecular-level entities. Comprehensive GRNs provide high-level overviews of the topology of gene-gene interactions and their purposes, thereby enabling a comprehensive understanding of their role in phenotypic variation, disease mechanisms, and other biological processes and how they may be perturbed for therapeutic purposes. With pathway analysis methods, GRNs can be used to glean mechanistic insights and derive meaning from high-throughput measurements in a knowledge-driven manner, allowing for the identification of pathways underlying processes and complex diseases as well as potential treatment targets. However, many GRNs are only partially known, and constructing accurate GRNs from gene expression data remains a challenge, complicated by problems such as small sample sizes, gene expression stochasticity, and incomplete characterizations of the gene regulatory dynamics. Here we present two methods for GRN reconstruction from time-course gene expression data. For the first, we develop a semi-supervised method that enables the synthesis of information from partially known networks with time-course data. This approach adapts PLS-VIP for time-course data and uses reference networks to simulate expression data from which null distributions of VIP scores are generated to estimate edge probabilities for input expression data. By using simulated dynamics to generate reference distributions, this approach incorporates previously known regulatory relationships and links the network to the dynamics to derive posterior networks and discover novel and anomalous connections. For the second approach, we adapt the time-lagged Ordered Lasso, a regularized regression method with temporal monotonicity constraints, for de novo and semi-supervised reconstruction by assuming that the regulatory strength of a gene diminishes with increasing temporal distance. We show that this constraint imbues favorable properties to improve prediction accuracy and that these methods can discover novel regulatory dependencies in existing pathways and produce accurate networks subject to the dynamics and assumptions of the time-lagged Ordered Lasso. In the Appendix, we introduce preliminary work for a method to identify significantly altered pathways based on their spectral properties. We describe an approach that assesses the alterations to a GRN across different phenotypic conditions based on the changes in the propagation of a gene expression signal and network-wide connectivity that are encoded by the graph Laplacian. Importantly, this method allows graphs to be signed by using equivalent unsigned representations that are compatible with the graph Laplacian to analyze GRNs based on the magnitudes of association between genes as well as the nature of those associations. We present theoretical results and preliminary applications of this method and discuss future directions to enhance the approach.

Creator