Gene regulatory network inference: A semi-supervised approach

Gene Regulatory Network (GRN) represents the regulatory interactions between genes. Experimental methods are capable of determining the nature of gene regulation in a given system, but are time-consuming and expensive. High-throughput technologies produce large number of gene expression data. These data along with additional information from heterogeneous data sources enable computational biologists to infer transcriptional gene regulations on genome-scale. Use of computational methods to infer gene regulatory networks from biological experimental data is a challenging task. Many studies have been conducted using unsupervised and supervised methods to achieve this task; however, such methods usually yield low prediction accuracy. The key issue of this problem is that, the negative training samples are not available for learning and the available positive training samples are insufficient. In this article, we propose a two-step semi-supervised method. First, a small number of reliable negative samples are extracted using clustering technique. Then, we utilize this negative samples for classification using an iterative procedure. We have applied this method to simulated as well as experimental gene expression dataset of Escherichia coli and evaluated the performance.