A NOVEL METHOD TO INFER A BOOLEAN NETWORK FROM TIME-SERIES GENE EXPRESSION DATA

Abstract

Inferring a gene regulatory network from time-series gene expression data in systems biology is a challenging problem. A variety of methods have been proposed, most of which were not efficient in the large networks because they limited the number of regulatory genes or computed an approximated reliability of multivariate relations. Therefore, an improved method is needed to search regulatory genes efficiently and to predict the network dynamics accurately. To resolve these limitations, I propose a novel genetic algorithm with an update rule table and neural network based Boolean network inference method called GTNBNI. In the early stage, a mutual information-based Boolean network inference (MIBNI) is applied to find the optimal solutions. If it fails, GTNBNI is applied to find optimal solutions. All the results are integrated to construct a final Boolean network. I want to note that GTNBNI has two solutions, namely GTNBNI-Table and GTNBNI-NN. MIBNI has two subroutines: MIFS and SWAP. The MIFS subroutine selects a set of initial regulatory genes, and the SWAP subroutine tries to improve the gene-wise dynamics consistency by iteratively swapping between a set of selected regulatory genes and the set of un- selected regulatory genes. I conducted extensive simulations on artificial datasets and real datasets, the experimental result shows that MIBNI outperforms other methods significantly in terms of both structural and dynamics accuracy. MIBNI used a simple update rule functions based on only conjunction or disjunction functions. A genetic algorithm with an update rule table(GTNBNI-Table) is devised to improve the MIBNI by exploiting the GA, which searches an optimal solution. MIBNI is applied first to find the optimal solutions. When MIBNI fails to find an optimal solution, GTNBNI-Table is applied to find an optimal solution. I compared GTNBNI-Table with four well-known inference methods through extensive simulations on artificial and real gene expression datasets. The results demonstrated that GTNBNI-Table significantly outperforms them in terms of structural and dynamics accuracies. Despite the successful performance achieved in GTNBNI-Table, there is room for improvement because it used an incomplete update rule table. Therefore, a genetic algorithm-combined neural network (GTNBNI-NN) is devised to replace the update rule table by a neural network. GTNBNI-NN also exploits MIBNI in the early stage to find optimal solutions. If MIBNI fails, GTNBNI-NN is employed to find the optimal solutions. Through extensive simulations on artificial and real gene expression datasets, GTNBNI-NN significantly outperforms five other methods in terms of structural and dynamics accuracy.