The transcriptomes in citrus in response to the HLB bacterial infection have been well documented in four previous reports [5, 10–12], but the information regarding the interactions between the differentially expressed genes is lacking. Through the combination of transcriptome comparative study and gene coexpression network analysis, we have provided for the first time a systems view of how the citrus host plant exerts a genome-wide response to the HLB bacterial infection.
First, we have constructed an HLB response network involving 3,507 Probesets with 56,287 interactions. Using the transcriptome datasets and orthology-based or experimentally verified protein-protein interaction datasets, gene-gene interactions or interactomes have been constructed in the model plants including Arabidopsis and rice (for example, [16, 45–52]) and occasionally in non-model plants such as soybean  and barley . However, there was no report on gene-gene interaction networks in citrus prior to our work. We used the Pcc method to construct a gene coexpression network in citrus, with a focus on the HLB response mechanism. The citrus gene coexpression network will be very useful for the citrus researchers to visualize the subnetworks specific for certain biological processes (such as carbohydrate metabolic process or signaling not analyzed in detail in this report), or to search some potential gene-gene interactions for certain genes or a group of genes in the future. The Citrus Gene Interaction Networks (CitGIN) database has been constructed and made available to the research community to query through the Internet (http://xt.cric.cn/cgr/CitGIN.php).
Second, our analysis of the defense subnetwork has shown that many defense hubs and hormone hubs are intertwined or overlapped. Although the roles of hormone and defense response genes have been discussed in the four previous reports [5, 10–12], our network analysis further indicates that hormone response is interconnected to defense response in citrus when challenged by the HLB bacteria. This may lead to the development of integrating hormone and disease response pathways as a potentially more effective genetic means to improve the citrus resistance to HLB.
Third, our comparative studies of transcriptomes have led to the identification of subsets of commonly up-regulated and stage-specific HLB responsive genes. In contrast to those four GeneChip reports where various statistical methods and fold-change cutoffs were used, we used the same procedure for the analysis of all of the transcriptome datasets. Furthermore, by mapping the subset of commonly up-regulated genes into the HLB response network, we have found that the genes belonging to the categories of carbohydrate metabolic process, transport and hormone response are positioned as the large hubs in the HLB response core subnetwork. This indicates that these three processes constitute a core subnetwork for the citrus host response to the HLB bacterial infection. In addition, we propose that transport is a key component in this HLB response core subnetwork.
Fourth, using PP2 gene as an example of applying the HLB response network, our subnetwork analysis provides an intriguing possibility for the zinc transporter or zinc binding proteins to act with PP2 protein in response to the HLB bacterial infection. PP2 proteins belong to a large gene family in higher plants. However, they have not been assigned a specific biological process, and thus their biological function remains unknown. They are predicted to bind carbohydrates and have been implicated a role in the formation of sieve plug or replacement phloem [5, 10, 12, 44]. Some of the PP2 genes from other organisms such as melon, cucumber and Arabidopsis are specifically or preferentially expressed in companion cells but their protein products are translocated in sieve elements . This indicates a role for PP2 proteins not only in intracellular signaling but also in long distance intercellular communication . Recent evidence show that PP2-type proteins might be involved in aphid-mediated virus transmission in cucumber  and overexpression of PP2-A1 in Arabidopsis increased resistance to a phloem-feeding insect , indicating a possibly more active role for PP2 family proteins in defense response in plants. In citrus, the Probeset Cit.35955.1.S1_at, which is closely related to Arabidopsis PP2-B8, was dramatically up-regulated at late stage and very late stages. The most surprising feature of the PP2-B8 subnetwork is that the 20 Probesets, which are the first degree neighbors of Cit.35955.1.S1_at, are interconnected frequently between each other. This indicates that these genes might be regulated by the precise coordination of various signaling pathways through transcription factors, chromatin modification or remodeling proteins or other factors. Furthermore, seven of the 20 interacting Probesets encode proteins involved in transport, consistent with our proposal that transport is a key component in the HLB response core subnetwork. In addition, three of the seven transporters are predicted to transport zinc, and the PP2 subnetwork also contains four Probesets which represent the genes encoding zinc-binding proteins. Intriguingly, HLB disease symptom was initially thought to be related to zinc deficiency and the zinc transport system is required for virulence in other organisms [1, 2], and therefore the PP2 subnetwork analysis indicates that zinc transporters or zinc binding proteins may have a potentially important role for citrus to respond to the HLB bacterial infection. Taken together, our analysis using the HLB response network can lead to an intriguing but testable hypothesis (which cannot be done without the network analysis) regarding the role of PP2 proteins and zinc transport system or zinc binding proteins in citrus HLB defense response.
It should be noted that there are some potential limitations in our network study. The first one is GO enrichment analysis. The agriGO web tool, which is based on the hypergeometric method and used in this work, does not take into account the local dependency of GO terms. Using the four algorithms provided in the topGO R package which are proposed to eliminate local dependencies, we have found that four of the six hormone GO terms determined to be overrepresented by agriGO are also overrepresented, while the two other hormones (ABA and ethylene) have their child GO terms being truly overrepresented. Therefore, different algorithms or statistical methods in GO enrichment analysis will probably lead to some differences in terms of the overrepresented GO terms for the nodes in the HLB response network.
The second limitation is due to the small sample size. Computational prediction of gene-gene interactions usually requires large sample size; however relatively small number of samples were recently used to construct gene coexpression networks specific to certain aspects of biology (such as [50, 52, 57]). In our analysis, we used the transcriptome datasets described in four previous reports [5, 10–12]. Among these, only one study is available for early stage and late stage respectively, while there are four studies for very late stage (Table 1). We have found that only a small number (222) of significantly regulated Probesets can be identified for early stage, while almost 600 and 2,000–4,000 differentially expressed Probesets can be found for late and very late stages respectively. The variation in the number of differentially expressed genes at different stages could be caused by the difference in experimental conditions given that different ages and varieties of trees and different sources of inoculants were used in different years in those four reports. However, this variation might lead to some sort of bias towards the very late stage genes. To minimize the possibility that the interactions we have detected were the result of random events due to the small sample size, we have selected a high Pcc cutoff value which has led us to believe that the interactions are more likely statistically significant rather than by random and that the topology of the HLB response network is quite similar to most biological networks. Furthermore, the cross-validation result shows a high degree of preservation of gene coexpression patterns, suggesting that the HLB response network is at least moderately robust and biologically relevant. Therefore, despite some limitations due to the small sample size and the experimental variations, the network reported here should be quite useful for the citrus research community and have provided some novel insights into the citrus HLB defense mechanisms. When larger scale transcriptome datasets become available in the future, similar network analysis will provide a comprehensive picture of the gene networks in citrus.
The most daunting challenge in the citrus postgenomic era remains how to identify the best candidate genes for functional dissection of the HLB response mechanism and for genetic modification with an ultimate goal of improving the HLB resistance in citrus. Genetic variations of HLB susceptibility  clearly shows the potential towards dissection of genetic mechanisms of HLB resistance, but understanding the inheritance patterns and subsequently cloning the disease genes requires a long term effort because of long juvenile phase and complex reproductive biology for citrus. Recent developments have shed some lights into the identification of key hub genes as candidate regulatory genes. For example, a seed germination study found that 22–50% of the Arabidopsis hub genes identified from the seed germination network actually have physiological functions in the control of seed germination . Therefore, the hub genes identified in this report may potentially be the first batch of candidates for the functional test in HLB resistance in citrus.