Reverse engineering of genetic networks with time delayed recurrent neural networks and clustering techniques

Abstract

In the iterative process of experimentally probing biological networks and computationally inferring models for the networks, fast, accurate and flexible computational frameworks are needed for modeling and reverse engineering biological networks. In this dissertation, I propose a novel model to simulate gene regulatory networks using a specific type of time delayed recurrent neural networks. Also, I introduce a parameter clustering method to select groups of parameter sets from the simulations representing biologically reasonable networks. Additionally, a general purpose adaptive function is used here to decrease and study the connectivity of small gene regulatory networks modules. In this dissertation, the performance of this novel model is shown to simulate the dynamics and to infer the topology of gene regulatory networks derived from synthetic and experimental time series gene expression data. Here, I assess the quality of the inferred networks by the use of graph edit distance measurements in comparison to the synthetic and experimental benchmarks. Additionally, I compare between edition costs of the inferred networks obtained with the time delay recurrent networks and other previously described reverse engineering methods based on continuous time recurrent neural and dynamic Bayesian networks. Furthermore, I address questions of network connectivity and correlation between data fitting and inference power by simulating common experimental limitations of the reverse engineering process as incomplete and highly noisy data. The novel specific type of time delay recurrent neural networks model in combination with parameter clustering substantially improves the inference power of reverse engineered networks. Additionally, some suggestions for future improvements are discussed, particularly under the data driven perspective as the solution for modeling complex biological systems.