Primerjava metod za inferenco genskih regulatornih omrežij

The main goal of our work was to systematically evaluate and compare different methods for gene regulatory network (GRN) inference. GRN inference is a challenging task due to the high dimensionality of gene expression data and the complex regulatory interactions between transcription factors (TFs) and their target genes. We studied several modern methods for GRN inference, including PANDA, OTTER, GENIE3, ARACNe, and CLR. For each method, we described the required input data, the computational approach used, and the results obtained. We evaluated the effectiveness of individual methods by comparing reconstructed networks with a reference network, using metrics such as MSE, accuracy, sensitivity, and F1. In addition, we analysed how noise and sample size affect the performance of each method. Since the “true” underlying network of gene regulation that could serve as a gold standard for verifying the results is not available, we decided to create our own simulated test networks. The first, smaller network was used to test the impact of sample size and noise on the performance of individual methods. The second, larger network is based on the knowledge network of gene interactions in model plant Arabidopsis thaliana and represents realistic biological data. Based on these networks, we simulated gene expression using a multivariate normal distribution. In doing so, we ensured that characteristic correlations between TFs and their target genes were preserved, including activation and inhibition relationships and self-regulation. Our work contributes to a better understanding of the effectiveness of different methods for GRN reconstruction. It also represents progress, as we test the methods on more complex organisms, multicellular plants such as Arabidopsis thaliana, and not just on bacteria or yeast, as has been common in comparative evaluations of methods such as in the DREAM5 challenges. Our results thus enable the comparison of methods under controlled conditions, assess their robustness at different noise levels and sample sizes, and provide guidelines for selecting appropriate approaches for the analysis of real biological data.