Genome-Wide Scale-Free Network Inference for Candida albicans

Abstract

Discovery of essential genes in pathogenic organisms is an important step in the development of new medication. Despite a growing number of genome data available, little is known about C. albicans, a major fungal pathogen. Most of the human population carries C. albicans as commensal, but it can cause systemic infection that may lead to the death of the host if the immune system has deteriorated. In many organisms central nodes in the interaction network (hubs) play a crucial role for information and energy transport. Knock-outs of such hubs often lead to lethal phenotypes making them interesting drug targets. To identify these central genes via topological analysis, we inferred gene regulatory networks that are sparse and scale-free. We collected information from various sources to complement the limited expression data available. We utilized a linear regression algorithm to infer genome-wide gene regulatory interaction networks. To evaluate the predictive power of our approach, we used an automated text-mining system that scanned full-text research papers for known interactions. With the help of the compendium of known interactions, we also optimize the influence of the prior knowledge and the sparseness of the model to achieve the best results. We compare the results of our approach with those of other state-of-the-art network inference methods and show that we outperform those methods. Finally we identify a number of hubs in the genome of the fungus and investigate their biological relevance.

Keywords: Candida albicans; LASSO; hubs; linear regression; network inference; prior knowledge; reverse engineering; scale-free.

Figure 1

F-measure of the LASSO inference for the 503 gold genes in which the gold standard and the expression data overlap. We exploited the BIND prior knowledge. The different graphs represent different values of ϵ and therefore different weighting of prior knowledge. It indicates that a higher influence of prior knowledge yields better results concerning the F-measure.

Figure 2

F-measures and number of interactions for different values of c for the LASSO-based genome-wide inference that used all genes and all available prior knowledge. The maximum F-measure 0.0018 is reached at c = 0.2 with a network of 6866 interactions between 6,167 genes.

Figure 3

F-measure obtained by LASSO-based genome-wide network inferences (left) with or without prior knowledge (FAC, PPI, TRANS, BIND) and with all four prior knowledge sources (ALL). The three bars on the right show the results of the mutual information-based networks.

Figure 4

Distribution of degrees for the LASSO-based inference without prior knowledge. The red line represents the fitted power-law. The correlation coefficient of the logarithmical data is 0.95.

Figure 5

Distribution of degrees for the ARACNE-based inference. The red line represents the fitted power-law. The correlation coefficient of the logarithmical data is 0.08.

Figure 6

Venn diagram of the four different prior knowledge sources and the gold standard. Empty fields contain no common interaction. There is little overlap between the sources of prior knowledge as well as between the prior knowledge and the gold standard.

Figure 7

Subnetwork GAL-genes using expression data and the full set of prior knowledge that does not contain these predicted relations. The connection of GAL10, GAL1, and GAL7 is well studied in many yeast forms. The network predictions also contain this relationship.

Figure 8

Predicted hub PSA2. The labels on the edges tell which inference and prior knowledge predicted this interaction. LASSO for the LASSO inference without prior knowledge. FAC, BIND, TRANS, and PPI for the inferences with the corresponding prior knowledge source. ALL for the inference that exploited all four prior knowledge sources.

Figure 9

Predicted hub TKL1. The labels on the edges tell which inference and prior knowledge predicted this interaction. LASSO for the LASSO inference without prior knowledge. FAC, BIND, TRANS, and PPI for the inferences with the corresponding prior knowledge source. ALL for the inference that exploited all four prior knowledge sources.