Learning Genetic Regulatory Network Connectivity from Time Series Data

Abstract

Recent experimental advances facilitate the collection of time series data that indicate which genes in a cell are expressed. This information can be used to understand the genetic regulatory network that generates the data. Typically, Bayesian analysis approaches are applied which neglect the time series nature of the experimental data, have difficulty in determining the direction of causality, and do not perform well on networks with tight feedback. To address these problems, this dissertation presents an improved method, called the GeneNet algorithm, to learn genetic regulatory network connectivity which exploits the time series nature of experimental data to allow for better causal predictions on networks with tight feedback. More specifically, the GeneNet algorithm provides several contributions to the area of genetic network discovery. It finds networks with cyclic or tight feedback behavior often missed by other methods as it performs a more local analysis of the data. It provides the researcher with the ability to see the interactions between genes in a genetic network. It guides experimental design by providing feedback to the researcher as to which parts of the network are the most unclear. It is encased in an infrastructure that allows for rapid genetic network model creation and evaluation. The GeneNet algorithm first encodes the data into levels. Next, it determines an initial set of influence vectors for each species based upon the probability of the species’ expression increasing. From this set of influence vectors, it determines if any influence vectors should be merged, representing a combined effect. Finally, influence vectors are competed against each other to obtain the best influence vector. The result is a directed graph representation of the genetic network’s repression and activation connections. Results are reported for several synthetic networks showing significant improvements in both recall and runtime while performing nearly as well or better in precision over a dynamic Bayesian approach.

Type

Related