Integration of structural constraints into the gene regulatory network inference

Abstract

The inference of gene regulatory networks (GRNs) from the gene expression data remains a challenging task. The number of genes is significantly larger than the number of experiments, where each experiment contains a noise component. We impose structural constraints on the inferred gene regulatory network based on the structure of reference GRNs. Our idea is motivated by the fact that GRNs contain a vast number of patterns, i.e. motifs, that are significantly more common than in randomized networks. We impose these constraints by modifying the weights of genes contributing to the joint loss function in the regression problem. We modify weights iteratively with gradient descent. Our approach is based on the already established partial correlation method dubbed SPACE. By extracting the expected number of regulatory genes, gene degree distribution and motifs from the reference network, we have improved by a small margin the inference accuracy, precision, recall and F1 score in the inference of GRNs derived from the GRN of the E. coli bacteria.