Constraint-Based, Score-Based and Hybrid Algorithms to Construct Bayesian Gene Networks in the Bovine Transcriptome.

Abstract

Simple SummaryIn this study, we investigated and compared six different Bayesian network algorithms from three different categories to identify hub genes critical to gene expression networks activated in response to progesterone in the bovine uterus. We observed many common hub genes identified between constraint-based algorithms (CBAs) and hybrid algorithms (HAs), while it appeared that score-based algorithm (SBA) methods led to more accurate and relevant predictions of core genes. The results revealed that the identification of hub genes was affected by the type of network reconstruction and by the subsequently used topological parameters. Two identified genes known to have roles during pregnancy are ISG15 and DGAT2. The identified hub genes are associated with biological processes such as amino acid metabolism, hormonal signaling pathways and the immune system. Our analysis revealed a role for miRNAs in the regulation of this system. The biological and physiological roles (enzymatic and hormonal effects) of unannotated identified hub genes should be functionally validated by further studies.Bayesian gene networks are powerful for modelling causal relationships and incorporating prior knowledge for making inferences about relationships. We used three algorithms to construct Bayesian gene networks around genes expressed in the bovine uterus and compared the efficacies of the algorithms. Dataset GSE33030 from the Gene Expression Omnibus (GEO) repository was analyzed using different algorithms for hub gene expression due to the effect of progesterone on bovine endometrial tissue following conception. Six different algorithms (grow-shrink, max-min parent children, tabu search, hill-climbing, max-min hill-climbing and restricted maximum) were compared in three higher categories, including constraint-based, score-based and hybrid algorithms. Gene network parameters were estimated using the bnlearn bundle, which is a Bayesian network structure learning toolbox implemented in R. The results obtained indicated the tabu search algorithm identified the highest degree between genes (390), Markov blankets (25.64), neighborhood sizes (8.76) and branching factors (4.38). The results showed that the highest number of shared hub genes (e.g., proline dehydrogenase 1 (PRODH), Sam-pointed domain containing Ets transcription factor (SPDEF), monocyte-to-macrophage differentiation associated 2 (MMD2), semaphorin 3E (SEMA3E), solute carrier family 27 member 6 (SLC27A6) and actin gamma 2 (ACTG2)) was seen between the hybrid and the constraint-based algorithms, and these genes could be recommended as central to the GSE33030 data series. Functional annotation of the hub genes in uterine tissue during progesterone treatment in the pregnancy period showed that the predicted hub genes were involved in extracellular pathways, lipid and protein metabolism, protein structure and post-translational processes. The identified hub genes obtained by the score-based algorithms had a role in 2-arachidonoylglycerol and enzyme modulation. In conclusion, different algorithms and subsequent topological parameters were used to identify hub genes to better illuminate pathways acting in response to progesterone treatment in the bovine uterus, which should help with our understanding of gene regulatory networks in complex trait expression.