An approach to infer putative disease-specific mechanisms using neighboring gene networks

Abstract

Motivation

The ultimate goal of any experiment is to understand the biological phenomena underlying the condition investigated. This process often results in genes network through which a certain biological mechanism is explained. Such networks have been proven to be extremely useful, for the prediction of mechanisms of action of drugs or the responses of an organism to a specific impact (e.g. a disease, a treatment, etc.). Here, we introduce an approach able to build a network that captures the putative mechanisms at play in the given condition, by using datasets from multiple experiments studying the same phenotype. This method takes advantage of known interactions extracted from multiple sources such as protein–protein interactions and curated biological pathways. Based on such prior knowledge, we overcome the drawbacks of snap-shot data by considering the possible effects of each gene on its neighbors.

Results

We show the effectiveness of this approach in three different case studies and validate the results in two ways considering the identified genes and interactions between them. We compare our findings with the results of two widely-used methods in the same category as well as the classical approach of selecting differentially expressed (DE) genes in an investigated condition. The results show that ‘neighbor-net’ analysis is able to report biological mechanisms that are significantly relevant to the given diseases in all the three case studies, and performs better compared to all reference methods using both validation approaches.

Availability and Implementation

The proposed method is implemented as in R and will be available an a Bioconductor package.

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

The ultimate goal of any experiment is to understand the biological phenomenon underlying the condition investigated. Understanding how genes interact with each other is the key to understand the onset and evolution of a disease, for instance. One of the drawbacks of the existing curated pathways or gene interactions databases is that each interaction is extracted from literature or experimentally validated by independent studies. However, most such interactions were found in specific tissues and/or phenotype, and not all studies employed the same tissue and/or phenotype.

Therefore, these independently identified interactions in the databases may not exist in actual phenotype or tissue studied in a subsequent experiment. Furthermore, new phenomena may be involved in the tissue or phenotype currently being studied. Utilizing only existing pathways from pathways databases or literature, limits one’s ability to discover new phenomena and new interactions. In order to overcome such limitations, some existing methods try to build the regulatory networks based on the correlation or the co-expression existing in the given datasets (Jiang et al., 2008; Langfelder and Horvath, 2008; Rhodes et al., 2007; Zou and Conzen, 2005). The networks resulting from such methods are specific to the condition under study, but the interactions identified are only based on the genes’ expression level. This limitation can produce many false positives as well as false negatives, because an interaction between two genes is not necessarily reflected in the correlation between their expression levels. The interactions between genes can involve an indirect relation between them via their protein products or their transcription factors, and sometimes interactions take place on different time scales. Therefore, there is a need for computational algorithms able to construct network of active interactions by analyzing data in more sophisticated ways, by combining gene expression data with existing pathway information, as well as with data from protein–protein interactions databases.

However, the integration of high throughput datasets such as gene expression data with prior information about gene–gene interactions to find the networks specific to a phenotype is still an open challenge (Komurov et al., 2010). Such methods identify the network of interactions that is most relevant to a given phenotype based on the retrieved prior knowledge, referred to as ‘active network’. This network is also known as ‘network hotspot’ or ‘responsive subnetwork’ (Mitra et al., 2013). The active network, as part of a global interaction network, explains the sudden changes in the genes activity or the characteristic of the phenotype in a given disease. This network is identified based on the given data and can be considered the putative mechanism involved in the given phenotype. The advantage of identifying active network is that it is specific to the condition studied, as opposed to existing curated pathways that can describe more generic knowledge, not necessarily applicable to the given condition.

The discovery of active network can lead to the identification of signature network that is associated to a given disease, rather than a set of gene biomarkers. This can lead to better understanding of the disease, diagnosis and more accurate treatment. Biomarker networks can also achieve more predictive power to classify different diseases such as diabetes or different types of cancer (Mitra et al., 2013). Furthermore, disease-specific networks can also be used to predict drug target mechanisms and to find the response of patients to the drugs.

Based on a comprehensive review published by Mitra et al. (2013) the existing approaches aiming to identify active network using prior knowledge of interactions, can be divided in three main categories, as follows. The first category is the ‘significant-area-search’ methods. In this category the genes and interactions between them are scored based on the input data, and the algorithm tries to find the group of genes and interactions with the highest score. The very first methods in this category are jActiveModules (Ideker et al., 2002) implemented in Cytoscape (Shannon et al., 2003), and Gene Network Enrichment Analysis (GNEA) (Liu et al., 2007). The second category includes ‘diffusion-flow’ and ‘network-propagation’ methods. The methods in this category attempt to find the flow between genes with maximum scores in the existing networks. They identify subset of genes and interactions that accumulate the highest score flows. The most widely used methods in this category are NetWalker (Komurov et al., 2010, 2012), HotNet (Vandin et al., 2011) and Physical Module Networks (Novershtern et al., 2011). The third category includes ‘cluster-based’ methods. These methods use biclustering algorithms to find the interactions that are active in the given conditions. The better known methods in this category are SAMBA (Tanay et al., 2004) and SANDY (Luscombe et al., 2004).

Here we proposed a ‘network-propagation’ algorithm that identifies the maximum flow between genes through their immediately connected genes. This method uses multiple steady state gene expression data that are collected from the same phenotype. The use of multiple datasets allows the proposed approach to capture changes of gene expression that might not be captured in any single dataset due for instance to the snapshot nature of gene expression data. Gene–gene interactions are obtained from protein–protein interaction networks describing the relations between proteins and also from experimentally curated signaling pathways. This method uses the neighborhood of each gene to identify the propagation of disruption that flows through the system. The neighborhood of each gene includes the genes that are directly connected to it based on the known interaction networks.

We apply the method on multiple datasets from experiments studying colorectal cancer, renal cancer and prostate cancer. We assess the results in two ways: first, we assess the enrichment of known biological pathways in the constructed network. This validation process is similar to a gene set analysis approach introduced in Liu et al. (2007). In this reference (Liu et al., 2007), the number of common genes between the associated gene sets to the given phenotype and the identified network is used to validate the results. Similarly, we also consider here the number of interactions overlapping between the constructed networks and known signaling pathways that are associated to the disease investigated. Constructed networks that are significantly enriched in these interactions are considered better than those that are less enriched in such interactions. Second, we obtain a list of genes that are associated to the disease studied from DisGeNET (Bauer-Mehren et al., 2011; Piñero et al., 2015) and calculate the enrichment of each constructed network in such disease-associated genes. DisGeNET integrates information from multiple public data sources and literature (15 different resources) to identify gene–disease associations (Piñero et al., 2015). Networks that are significantly enriched in disease-associated genes are considered better than those that are less enriched in such genes. We compare our result with the results of two widely used methods: NetWalker and HotNet, which also attempt to identify the relevant biological mechanisms on a global network. We also compare our results with the classical approach of considering just the union of differentially expressed (DE) genes in all considered studies. The results of these comparisons show that the proposed method performs better in constructing networks that are significantly relevant to the given diseases based on the resulting genes and interactions.

2 Materials and methods

The approach presented here aims at finding networks of genes that capture the mechanisms that could explain the phenotype. The algorithm requires two types of data. The first type is gene expression data that includes measured gene expression in the control samples versus the investigated disease. The proposed approach uses multiple datasets relevant to the same condition. This is important because each dataset is only a snapshot of the system captured at a given moment in time. The changes in expression levels of the genes will be better estimated by looking at multiple datasets, which represent different snapshots of the same condition, and therefore provide much more information. The second type of data is the prior information about gene–gene interactions that can be collected from different resources describing any known interactions between genes, such as protein–protein interaction, gene regulatory networks and curated pathways. In addition, adding existing information about known interactions allows us to estimate the potential effect of a condition on groups of interacting genes that can ultimately constitute putative models of the mechanisms in action in the given condition. The integration of these two types of data allows us to overcome some of the limitations of existing methods.

The method starts with several existing datasets available for that disease. A list of differentially expressed (DE) genes is calculated as the union of all sets of DE genes from each such individual dataset.

The proposed approach then builds a ‘neighbor network’ for each gene. The neighbor network associated to each gene includes the gene itself, the genes immediately connected to it and the interactions connecting them together based on known interactions from protein–protein interaction databases such as a HPRD (Peri et al., 2003, 2004), as well as from pathway databases such as KEGG (Kanehisa et al., 2002, 2004). This is done such that, even if none of the multiple datasets included in the analysis captures the effect of the gene of interest, by looking at its immediate neighborhood we can still detect changes that propagate from that gene. The neighbor network is constructed exclusively from annotation databases, independently of the DE genes. Even though there is no upper limit on the size of the neighbor network constructed, most such networks (98%) will contain fewer than 50 genes (see Fig. S6 in Supplementary Material). Neighbor networks including fewer than or equal to 2 genes are eliminated from our analysis. In the next step, we calculate the enrichment of each neighbor network based on the number of DE genes they contain. The hypergeometric P-value for each neighbor network is calculated based on the formula below:

$$p_o(x)=1-\sum _{i=0}^{x-1}\frac{\left(\begin{array}{c}M\\ i\end{array}\right)·\left(\begin{array}{c}N-M\\ K-i\end{array}\right)}{\left(\begin{array}{c}N\\ K\end{array}\right)}$$ (1)

where N is the total number of genes, K is the total number of DE genes and M is the number of genes in each neighbor network. This P-value represents the probability of obtaining a number of DE genes in the neighbor network that is equal or higher than the number observed in the analysis, just by chance. Such P-values are computed for all neighbor networks and are corrected for multiple comparisons with false discovery rate (FDR) method. The significant neighbor networks, with FDR-corrected P-values lower than threshold (10%), are identified and combined together to build the final constructed network. The genes and interactions in the constructed network are the integration of all the genes and interactions extracted from all the identified significant neighbor networks. This constructed network resulted from the analysis can be considered as the active network that has the potential to capture the mechanisms involved in the given disease. A summary of the method, referred to as ‘neighbor-net analysis’, is shown in Figure 1.

Fig. 1.

An overview of neighbor-net analysis. (a) The global network combining all the known gene–gene interactions. The colors show three sample neighbor networks for genes a, b and c. (b) The neighbor network for each gene is extracted from the global known interactions. A global list of differentially expressed (DE) genes is obtained by constructing the union of all genes found to be DE in at least one of the given datasets (based on their calculated log fold changes and P-values adjusted for multiple comparisons). The Fisher’s exact test is performed on all extracted neighbor networks based on the number of DE genes they have. (c) The significant neighbor networks (FDR-corrected P-value lower than 10%) are identified. (d) The constructed network which is built by integrating all significant neighbor networks (shown in red)

3 Results

We present the results of our analysis on three different diseases (colorectal cancer, renal cancer and prostate cancer). The method requires a set of datasets studying the same disease under the same conditions. We selected three diseases with three or more datasets available in GEO associated to each of the three cancer types. The results of colorectal cancer are included in the main manuscript and the results of renal cancer and prostate cancer are included in the Supplementary Material. These results are compared with the results obtained with three other approaches: HotNet (Vandin et al., 2011), NetWalker (Komurov et al., 2010, 2012) and the classical ORA (Khatri and Drăghici, 2005). The detailed results of each disease are shown in separate sections. The gene expression data in each case is obtained from Gene Expression Omnibus (GEO) (Barrett et al., 2013). Using GEO2R (Sayers, 2011), we compute a fold change and a P-value for each gene representing the significance of the observed change in the expression levels between normal and disease samples. The information about gene interactions is obtained from the Kyoto Encyclopedia of Genes and Genomes (KEGG) (Release 72.0) (Kanehisa and Goto, 2000; Kanehisa et al., 2002, 2004) and Human Protein database Reference (HPRD) (Release 9) (Peri et al., 2003, 2004). We chose to illustrate the proposed approach using KEGG because it is well recognized and widely used (almost 7800 citations) but this approach is completely independent of the pathway database used and can be used equally well with Reactome (Croft et al., 2014; Joshi-Tope et al., 2005; Nishimura, 2001). Similarly, the approach can be used with any other protein–protein interaction database such as BioGRID (Stark et al., 2006), BIND (Bader et al., 2001), etc.

We collapsed all 173 available KEGG signaling pathways into one network with 5052 nodes and 27 811 interactions. The protein–protein interaction (PPI) network downloaded from HPRD includes 9672 nodes and 39 233 interactions. There are 3638 common genes and 3278 common interactions between KEGG and HPRD. We used gene Entrez IDs to map the genes to the known networks. Combing these two networks results a global graph with 11 086 nodes and 62 934 interactions. The neighbor networks are built for each gene and the associated P-values are computed to represent the significance of their enrichment in the list of DE genes. The constructed network for each case study is the integration of all significant neighbor networks.

There is no universally accepted validation method to assess the accuracy of the constructed networks. Similar to a previous study (Liu et al., 2007), we evaluate the results by performing pathway enrichment analysis based on the edges overlapping between the constructed network and pathways to identify what known biological mechanisms are captured. Such analysis allows us to validate the parts of the constructed network that are obtained from existing pathways known to be involved in the given disease. The hypergeometric P-value for each pathway is computed based on the number of common edges it has with the constructed network. The ranked list of significant pathways for each case study will show which pathways are more enriched in the constructed network. For each disease, we consider the pathway that was created in order to explain that particular disease as the ‘target pathway’ (e.g. Colorectal cancer pathway is the target pathway for colorectal cancer). We validate the constructed network using the rank of target pathway in the reported list of pathways sorted by P-values corrected for multiple comparisons. The constructed network is more relevant to the investigated disease if the rank of the target pathway is lower and the corresponding P-value is more significant (Tarca et al., 2012). Note that other truly impacted pathways may be present in this list for legitimate reasons.

In addition, we also evaluate the results by using the DisGeNET database (Bauer-Mehren et al., 2011; Piñero et al., 2015) to determine how many genes in the network are known to be related to the given condition. A P-value representing the significance of the number of identified genes known to be associated to the investigated disease in the constructed network is computed. A lower P-value will mean that the method identifies genes that are more relevant to the disease and therefore, that the constructed network is more likely to describe the mechanism involved in that disease. Both evaluation approaches are summarized in Figure 2. The results described in the following sections show that the proposed method performs better in both evaluation processes in comparison to all reference methods.

Fig. 2.

An overview of the two evaluation processes for the constructed network. (a) The constructed network which is built by integrating all significant neighbor networks (shown in red). (b) We start the evaluation process by looking at the enrichment of each KEGG pathway in edges also present in the constructed network. The red edges in each pathway represent the edges overlapping between that pathway and the constructed network. The significance of enrichment of each pathway is calculated based on the number of such edges. (c) Pathways are ranked using their enrichment P-values. (d) The second evaluation process calculates an enrichment P-value for each constructed network. This P-value characterizes the enrichment of each constructed network in genes that are known to be associated to the investigated disease based on DisGeNET. Lower P-values represent a significant enrichment of the constructed network in nodes known to be associated with the disease

3.1 Colorectal cancer

We analyze five gene expression datasets studying colorectal cancer from GEO (Barrett et al., 2013) (GSE4173, GSE9348, GSE21510, GSE32323 and GSE8671). A global list of differentially expressed (DE) genes is obtained by selecting the genes with an absolute value of log2 fold change higher than 1.5 and adjusted P-value lower than 0.01 in at least one dataset. The union of DE genes includes 2968 genes out of 19 852 total genes in the five experiments. We perform our analysis on selected DE genes. Based on the calculated hypergeometric P-values for all neighbor networks, 20 of them are significantly enriched in the given list of DE genes. The constructed network is a global graph that integrates all the significant neighbor networks. It includes 144 genes and 251 interactions and is shown in Figure 3. This network can be seen as most likely to include the mechanisms involved in colorectal cancer.

Fig. 3.

The active network that describes the putative mechanisms involved in colorectal cancer. The five subnetworks shown above include 144 nodes and 251 edges. The 130 red edges represent the interactions that exist in significantly enriched KEGG pathways. The 17 edges shown in blue are present in KEGG pathways that are not significantly enriched in edges from this network. The 70 genes shown in green are known to be associated to colorectal cancer based on the DisGeNET database

The edges overlapping between each KEGG signaling pathway and the constructed network are identified. The probability of having the observed number of these edges just by chance is calculated for every KEGG pathway. The list of KEGG pathways that are significantly enriched in edges from the constructed network is shown in Table 1. This table shows that the constructed network includes a significant number of edges from the target pathway. In fact, the Colorectal cancer pathway is ranked 3rd and it has a significant number of edges in common with the constructed network. The two other pathways that are ranked higher than the Colorectal cancer pathway are the Hippo signaling pathway and the Wnt signaling pathway. There is a extensive evidence that both Hippo (Barry and Camargo, 2013; Cai et al., 2010; Pan, 2010), as well as Wnt (Bienz and Clevers, 2000; Clevers, 2006; Reya and Clevers, 2005) are very important in colorectal cancer.

Table 1.

A list of significant pathways (FDR-corrected P-value < 0.05)

Significant pathway	FDR-correctedP-value	References
Hippo signaling pathway	2.1e−104	Barry and Camargo (2013); Cai et al. (2010); Pan (2010)
Wnt signaling pathway	5.5e−22	Reya and Clevers (2005); Bienz and Clevers (2000); Clevers (2006)
Colorectal cancer pathway	3.7e−17
Thyroid cancer	1.8e−15
Endometrial cancer	2.4e−12
Arrhythmogenic right ventricular cardiomyopathy (ARVC)	6.4e−08
Calcium signaling pathway	5.7e−07	Lamprecht and Lipkin (2003)
Melanogenesis	7.05e−06
Pathways in cancer	2.8e−03
Prostate cancer pathway	4.3e−03

Note: These pathways are significantly enriched in the network resulted from neighbor-net analysis. The bold pathway is the target pathway in colorectal cancer. The third column shows the references explaining the association of the respective pathways to colorectal cancer.

Figure S1 in the Supplementary Material shows the edges from the network built by the proposed algorithm, as they appear in the context of existing KEGG pathways. Some of the gene interactions (edges) in this figure, appear multiple times in various significant KEGG pathways.

Interestingly, these edges from the network built by the proposed neighbor-net analysis describe a well-known mechanism known to be involved in colorectal cancer. The β-catenin protein is a very well-known protein that has important impacts on developing colorectal cancer (Morin et al., 1997). It is produced by gene ‘CTNNB1’. This gene is one of the parent nodes in the green network present in 9 out of the 10 significant pathways shown in Figure S1 in the Supplementary Material. Notably, this gene is not identified as a differentially expressed gene by classical approaches in any of the datasets so the classical approach of focusing on differentially express genes would not be able to identify this mechanism. The interactions between ‘CTNNB1’ and its downstream genes, ‘LEF1’ and ‘TCF7L1’, are part of the network built by the proposed approach, network that is present in most of the significant KEGG pathways. These genes are immediately connected to other genes such as ‘MYC’, ‘CCND1’ and ‘BIRC5’ that have important roles in the evolution of colorectal cancer through a number of cell functions (e.g. proliferation, apoptosis) (Lifschitz-Mercer et al., 2001; Morin et al., 1997).

We also compare our result with the results produced by NetWalker and HotNet. Both methods are widely used to construct networks of genes that are meant to describe the active modules in a given phenotype. NetWalker is built based on an algorithm introduced in Komurov et al. (2012, 2010). It accepts as input a list of all genes in the analysis but it requires the selection of a specific group as ‘seeds’. We selected as seeds the genes that are differentially expressed (fold change higher than 1.5 and adjusted P-value lower than 0.01) in at least one of the datasets. NetWalker uses multiple resources such as KEGG, REACTOME interactions and literature based gene regulatory networks, as prior knowledge. The output of this method is a network in which nodes represent genes, and edges represent interactions between them. This network is claimed to explain the mechanisms involved in the investigated disease. The result of NetWalker for colorectal cancer datasets is a network that includes 901 genes and 3028 interactions. The P-value representing the significance of the number of edges overlapping between this network and Colorectal cancer pathway is 0.99 (see Table 2). This P-value shows that the network constructed by NetWalker is not overlapping in any significant way with the KEGG pathway that describes the phenomena involved in this type of cancer.

Table 2.

The ranks and P-values of the target pathway (Colorectal cancer pathway) in neighbor-net analysis and three other methods

Colorectal cancer
Method	rank	FDR-corrected P-value
Neighbor-net analysis	3	3.7e-17
NetWalker	22	0.99
HotNet	95	1.0
ORA	96	0.37

Note: The P-values represent the significance of the enrichment of the Colorectal cancer pathway in the identified active network. The comparisons show that neighbor-net analysis reports the target pathway more significant and highly ranked.

We also compare the results with HotNet (Vandin et al., 2011) that also constructs a network from known protein–protein interaction (PPI) network by considering the degree (number of links) of each gene together with a gene’s score that shows the significance of change in its expression level. We use the genes’ negative log of P-values as their associated scores. The minimum P-value for each gene in five datasets is used to compute the scores. HotNet requires a threshold for selecting the important networks. The authors provide an algorithm that suggests five different thresholds for the given data. We use the minimum threshold suggested by the algorithm. The constructed network includes 215 genes and 175 interactions. The P-value representing the enrichment of the target pathway in the constructed network is 1.0 (see Table 2). In other words, the network constructed by HotNet does not overlap in any significant way with the colorectal cancer pathway from KEGG, suggesting that this constructed network does not capture many of the phenomena considered to be central to the colorectal cancer development by the KEGG’s authors.

We also compare the enrichment of the target pathway in the network constructed from the proposed method with the results of over representation analysis (ORA) which is a classical pathway analysis method (Khatri and Drăghici, 2005; Mitrea et al., 2013). ORA takes into consideration the number of DE genes observed in each pathway and calculates the probability of observing this number just by chance. The analysis is performed on the union of DE genes in all datasets considered. The rank and the P-value of the target pathway representing its enrichment in the list of DE genes is calculated and shown in Table 2.

We also evaluate the results by assessing the number of nodes in each constructed network that are known to be associated to the investigated disease. We compare the genes in the identified active network with the genes known to be associated to colorectal cancer, obtained from the DisGeNET database. The number of associated genes to colorectal cancer from this database is 2277. Based on the extracted associated genes 70 genes out of 144 genes reported by the neighbor-net analysis are known to be associated to colorectal cancer (48%). The percentage of the genes associated to the colorectal cancer in the network constructed by neighbor-net analysis is higher compared to all the reference methods. This means that the neighbor-net analysis is able to identify a higher proportion of genes related to colorectal cancer in comparison to other methods (see Table 3). Also, the computed P-value that represents the significance of enrichment of such genes in the constructed network in neighbor-net analysis is also highly significant (8.4e−15).

Table 3.

The statistical analysis of the results from neighbor-net analysis and all the methods compared

Colorectal cancer
Method	#selected genes	#colorectal cancer genes	%	P-value
Neighbor-net analysis	144	70	48%	8.4e−15
NetWalker	901	283	31%	5.5e−18
HotNet	215	46	21%	0.23
Selected DE genes	2968	552	18%	5.8e−27

Note: The columns show: the number of genes in the identified active network reported by each method, the number of associated genes to colorectal cancer based on information obtained from DisGeNET (Piñero et al., 2015), the percentages of the genes known to be associated to colorectal cancer in the total number of identified genes in each constructed networks, and the corresponding P-values for the enrichment in each method. The P-value of observing the given number of genes that are associated to colorectal cancer in the constructed network is highly significant in the neighbor-net analysis. The percentage of the associated genes in the constructed network in also higher compared to all three existing methods.

The P-values of observing the number of associated genes to colorectal cancer, just by chance, as well as the percentages of such genes in the lists of genes reported by NetWalker, HotNet and selected DE genes are shown in Table 3. The lower P-values represent more significant enrichment of the genes associated to colorectal cancer in the constructed network. The P-values show that the selected DE genes and the list of genes resulted from NetWalker are also significantly enriched in the genes associated to colorectal cancer. However, the percentage of such genes in the total list of identified genes by neighbor-net analysis is higher compared to all other methods. The statistical analysis shows that neighbor-net analysis performs better than the existing methods in both evaluation approaches. Essentially, the neighbor-net analysis is able to find more associated genes to colorectal cancer compared to NetWalker, HotNet and the enrichment approach, as well as more interactions relevant to colorectal cancer based on number of overlaps with the Colorectal cancer pathway.

4 Discussion

A reasonable question might be posed regarding the degree to which the results obtained by the proposed approach depends on the source of annotation, for instance, on the particular database used for protein–protein interactions. In order to investigate this, we also performed the neighbor-net analysis by using BioGRID (Stark et al., 2006) instead of HPRD, as the protein–protein interactions resource. The results of colorectal cancer study (shown in Table 4) demonstrate that the proposed method is still able to find the relevant networks, which include known mechanisms involved in the given disease in two of the three cases. The results of renal cancer, and prostate cancer are shown in Table S7 in the Supplementary Material. We also compared the constructed networks by neighbor-net analysis using two BioGRID and HPRD and the results are shown in Table S8 in the Supplementary Material.

Table 4.

Results of neighbor-net analysis using BioGRID database as protein–protein interactions resource

Ranks and P-values of target pathways using interactions obtained from BioGRID database
Disease	Rank of target pathway		FDR-corrected P-value of target pathway
Colorectal cancer	5		2.7e−16
Number of disease-associated genes using interactions obtained from BioGRID database
Disease	#selected genes	#disease genes	%	P-value
Colorectal cancer	46	22	47%	3.51e-07

Note: The ranks and the P-values of target pathway in colorectal cancer as well as the enrichment P-value of identified genes in obtained disease-associated genes are shown above. The significant P-values for the target pathways and high enrichment of constructed network in known disease-associated genes in this case study determine that the proposed method is not significantly dependent on one specific database and is able to identify the known mechanisms involved in the given disease by using different resources.

Another reasonable question is whether the method might be biased towards heavily studied genes, like known cancer drivers, for which many more protein interactions are known because of study and annotation bias. In order to investigate this, we also applied the proposed approach on the same datasets, but after we excluded the genes with high connectivity in the list of differentially expressed genes for each case study. The results of both evaluation approaches are included in Table S9 in the Supplementary Material. The results show that the proposed method is not simply reporting the highly studied genes that are differentially expressed, but rather it is also able to find the mechanisms involving the less studied genes. These putative mechanisms are still significantly enriched in truly relevant pathways, as well as in disease genes from DisGeNET.

It is also important to investigate whether the proposed approach has the tendency to produce false positives, i.e. construct networks claimed to be representing putative mechanisms that are in fact not related to the data analyzed. In order to show that this is not the case, we applied the proposed approach on a number of randomly generated datasets. From the graph of 11 086 genes and 62 934 interactions constructed from KEGG and HPRD, we selected 1000 genes to be ‘differentially expressed’. The proposed approach was applied to this set of 1000 DE genes to construct neighbor networks and calculate their significance at the 10% level. This whole process was repeated 1000 times. In 974 cases of these 1000 simulations (97.4%), no neighbor network was reported as significant at this significance level. In 24 cases (2.4%) there was only one neighbor network reported as significant. In one case (0.1%), there were two significant networks and in another one case (0.1%) there were three networks reported as significant. These illustrate that the proposed approach produces substantially fewer false positives than usually accepted (10% false positives are normal for a significance level of 10%).

Another potential question is related to the use of KEGG pathway both in the construction of the neighbor networks, as well as in the validation. Does this introduce any bias in our validation? The answer is no, the validation is not biased. The neighbor networks are constructed based on all KEGG pathways. There are 173 such pathways. Out of these 173 pathways, there is only one target pathway for each of the diseases included in the paper. Because we used KEGG in the construction of our initial network, it is expected that some of the edges from our result network will appear in KEGG as well. However, there is no reason to expect that these common edges will be found precisely on the pathway describing the studied disease, in each case. However, this is not the only reason to trust the results. Our validation is based on two other essential facts: (i) that the analysis was able to construct a network describing the putative mechanism in the first place; and (ii) that the constructed network is significantly enriched in genes known to be involved in the given disease according to DisGeNET. The first point is important because we have shown that for random data, no significant networks are found and no final network is constructed at all in the vast majority of cases. For random data, it happened only once in 1000 trials that 3 significant neighbor networks were found. In contrast, there were 20, 69 and 23 such significant neighbor networks for colorectal, renal and prostate cancer, respectively. The second point is crucial because no information from DisGeNET was used in the construction of the networks, nor in any other way. However, for each disease, the network constructed by the proposed approach was enriched in genes know to be implicated in that disease according to DisGeNET at extraordinary levels of significance (${10}^{-14}$–${10}^{-41}$).

5 Conclusion

Inferring the active network involved in the investigated disease is one of the most important goals in system biology. Given the huge number of publicly available datasets, disease-specific networks can be constructed by using multiple datasets from the same condition to capture multiple states of the genes in the given phenotype. We take advantage of known information about genes interactions available in multiple databases to consider the possible disease-specific effects of genes on the genes immediately connected to them by any kind of known interaction. The networks constructed include interactions from KEGG and HPRD (and therefore it is reasonable to believe they are true). We also have shown that these networks are very significantly enriched in genes known to be involved in the respective diseases according to DisGeNET. Hence, we think it is reasonable to consider them as networks describing the putative mechanisms involved in the given phenotypes. Furthermore, the results obtained from 12 datasets involving three diseases constructed networks were shown to be able to include important genes even though they may not be differentially expressed and therefore, they would not be found by a classical approach based on DE genes.

Funding

This research was supported in part by the following grants: NIH R01 DK089167, R42 GM087013 and NSF DBI-0965741, and by the Robert J. Sokol Endowment in Systems Biology in Reproduction. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any of the funding agencies.

Conflict of Interest: none declared.