Data sources used to predict protein functional interactions
We used the following six classes of data to predict protein FIs (Table ): , human physical PPIs catalogued in IntAct [45
], HPRD [46
], and BioGrid [47
]; 2, human PPIs projected from fly, worm and yeast in IntAct [45
] based on Ensembl Compara [48
]; 3, human gene co-expression derived from DNA microarray studies (two data sets [49
]); 4, shared GO biological process annotations [51
]; 5, protein domain-domain interactions from PFam [52
]; and 6, PPIs extracted from the biomedical literature by the text-mining engine GeneWays [53
Data sources used to predict protein functional interactions
Table lists these data sources, the numbers of proteins and interactions, and estimated coverage of the human genome expressed as their coverage of the SwissProt protein database.
The coverage ranges from 7% (Worm PPIs) to 70% (GO biological process sharing). It is notable that the coverage of human physical PPIs from three public protein interaction databases (IntAct, HPRD, and BioGrid) is close to 50%. Many interactions from IntAct were catalogued from co-immunoprecipitation experiments combined with mass spectrometry, and contain multiple proteins in a single interaction record. An odds ratio analysis showed that human PPIs based on all interaction records are much less correlated to FIs (see below) extracted from Reactome pathways than interactions containing four or fewer interactors: 13.91 ± 0.52 versus 36.98 ± 9.17 (P-value = 2.8 × 10-5 based on t-test). Therefore, we selected interactions that contain only four or fewer interactors from the IntAct database. We also tried to use GO molecular functional annotations as one of the data sources. The odds ratio of this data set was 2.99 ± 0.02, much smaller than the GO biological process data set (11.85 ± 0.20). Our results show that this data set contributed little to the prediction. One reason for this may be that the GO molecular functional categories are usually broad and the purpose of our NBC is to predict if two proteins may be involved in the same specific reactions (see below).
Construction and training of a functional interaction classifier
Our goal was to create a network of protein functional relationships that reflect functionally significant molecular events in cellular pathways. The majority of PPIs in interaction databases are catalogued as physical interactions, and there is rarely direct evidence in the interaction databases that these interactions are involved in biochemical events that occur in the living cell. Other protein pairwise relationships have similar issues. To integrate pairwise relationships into a pathway context, we built a scoring system based on the NBC algorithm, a simple machine learning technique [54
], to score the probability that a protein pairwise relationship reflects a functional pathway event.
For our NBC, we used nine features as listed under 'Data source' in Table : 1, whether there is a reported PPI between the human proteins; 2, whether there is a reported PPI between the fly (Drosophila melanogaster) orthologs of the two human proteins; 3, whether there is a reported PPI between the worm (Caenorhabditis elegans) orthologs of the two human proteins; 4, whether there is a reported PPI between the yeast (Saccharomyces cerevesiae) orthologs of the two human proteins; 5, whether there is a domain-domain interaction between the human proteins; 6 and 7, whether the genes encoding the two proteins are co-expressed in expression microarrays based on two independent DNA array data sets; 8, whether the GO biological process annotations for human proteins are shared; and 9, whether there is a text-mined interaction between the human proteins.
An NBC must be trained using positive and negative training data sets in order to determine the proper weighting of different combinations of features. We developed training sets from the curated information in Reactome, relying in part on an independent analysis that reported Reactome as a highly accurate data set for PPI prediction [55
An issue in using PPIs and other pairwise relationships in a pathway context is that the data models used by pathway databases are much richer than a simple binary relationship. A pathway database describes pathways in terms of proteins, small molecules and cellular compartments that are related by biochemical reactions that have inputs, outputs, catalysts, cofactors and other regulatory molecules. To develop the training sets from Reactome pathways for NBCs, we established a relationship called 'functional interaction' using the following definition: a functional interaction is one in which two proteins are involved in the same biochemical reaction as an input, catalyst, activator, or inhibitor, or as two members of the same protein complex.
It is important to note that in Reactome a 'reaction' is a general term used to describe any discrete event in a biological process, including biochemical reactions, binding interactions, macromolecule complex assembly, transport reactions, conformational changes, and post-translational modifications [23
]. We treat two members of the same protein complex as functionally interacting with each other because the activity of the complex as a whole is presumably functionally dependent on the presence of all of its subunits.
Based on the above definition, we extracted 74,869 FIs from Reactome, and used these FIs to create a positive training set for the NBC. After filtering out FIs that did not have at least one feature derived from the data sources in Table , the positive data set comprised 45,079 FIs.
Creating a good negative training set is more difficult than creating a positive set due to the incompleteness of our knowledge of protein interactions [56
]: just because two proteins are not known to interact does not mean that this does not in fact occur. Research groups have addressed this problem using a variety of approaches, including choosing protein pairs from different disjunct cell compartments [57
], or random pairs from all proteins [58
]. For our NBC training, we followed the method in Zhang et al.
] using random pairs selected from proteins in the filtered Reactome FI set.
Choosing an appropriate prior probability or ratio between the positive and negative data sets is important for NBC training. We calculated the prior probability based on the total number of proteins in the filtered FIs from Reactome pathways, which was 5.7 × 10-3. To check the effect of ratio between the sizes of the positive and negative data sets, we test the NBC performance using a ratio of either 10 or 100. NBCs trained with these two ratios yielded similar true and false positive rates, which indicated that our NBC is robust against the size of the negative data set.
The performance of machine learning classifier systems can be evaluated by cross-validation, or more stringently by using an independent data set. We used FIs extracted from pathways in other human curated pathway databases as a testing data set to evaluate the performance of our trained NBC. Figure shows a receiver operating characteristic curve that relates true positive rates to false positive rates across a range of thresholds using this testing data set. We chose a threshold score of 0.50, which trades off a high specificity of 99.8% against a low sensitivity of 20%. The low sensitivity may result, in part, from high false negative rates existing in some of the data sets we used for NBC, especially in PPIs [59
Figure 2 Receiver operating characteristic curve for NBC trained with protein pairs extracted from Reactome pathways as the positive data set, and random pairs as the negative data set. This curve was created using an independent test data set generated from pathways (more ...)
At the threshold score (0.50), a protein pair must have multiple types of FI evidence in order to be scored as a true FI (Table S1 in Additional file 1
). While most (97%) of the predicted FIs have at least one PPI feature (Figure S1 in Additional file 1
), there are no predictions supported solely by human PPI data, and fewer than 3% are supported solely by PPIs in human plus other species. This greatly reduces the weight given to raw human PPI features: the 44,819 human PPIs that went in to the classifier as features resulted in fewer than 15,000 predicted FIs, representing the removal of 68% of the raw PPIs. Most (75%) of the predicted FIs are derived from GO biological process term sharing and protein domain interactions in addition to PPIs.
As a check on the classifier's ability to enrich for FIs, we compared the sharing of GO cellular component annotations (which includes compartments such as 'nucleoplasm') among raw human PPIs to the sharing of these annotations among predicted FIs. Since GO cellular component annotations were not used as a feature during NBC training, we reasoned that this assessment should be independent. Among raw PPIs, 62.9% share GO cellular component terms annotated for both proteins involved in the interaction. In contrast, 96.2% of the predicted FIs share this type of GO term (P-value < 2.2 × 10-16), suggesting a substantial enrichment in true FIs relative to an interaction set derived from raw features alone.
Merging the NBC with pathway data to create an extended FI network
To construct an extended FI network with high protein and gene coverage, we merged FIs predicted from our trained NBC with annotated FIs extracted from five pathway databases. The five pathway databases used were Reactome [23
], Panther [60
], CellMap [61
], NCI Pathway Interaction Database [62
], and KEGG [63
] (Table ).
Pathway data sources in the functional interaction network
To further increase the coverage of our network, we imported interactions between human transcription factors and their targets from the TRED database [64
]. TRED has two parts: one contains highly reliable, human curated data from published literature and the other is uncurated and comprises predictions based on several computational algorithms. For our purposes, we used the human curated part only to ensure the reliability of our FI network, and treat these interactions as a part of the pathway FIs in this report.
The extended FI network contains 10,956 proteins (9,393 SwissProt accession numbers, splice isoforms not counted) and 209,988 FIs (Table ). It covers 46% of SwissProt proteins.
Protein identifiers and functional interactions in the extended FI network
The average connection degree (that is, the number of interacting partners per protein) of the extended network is 38, and the maximum degree is 593 for protein P32121 (ARRB2, Beta-arrestin-2). Most proteins in this network are interconnected: 10,645 proteins are interconnected in the largest connected graph component. The remaining 311 proteins reside in 124 connected graph components of size 7 or smaller.
The FI network shows scale-free properties (data not shown) as do other biological networks [65
]. GO slim annotation enrichment analysis results (not shown) show that our network is enriched in proteins involved in signal transduction, cell cycle and the central dogma. This reflects the ascertainment bias of using Reactome as the training set, as these pathways reflect high priorities for Reactome curation.
Assessing the utility of functional interactions in the network
GBM is the most common type of brain tumor in humans and also has the highest fatality rate. Recently, two data sets from two independent high throughput screens for somatic mutations involved in GBM have been released [12
]. In this section, we demonstrate that the interactions from our network can be used to automatically extend a hand-curated GBM pathway developed to support the analysis of one of these data sets [14
]; the extended GBM pathway captures more observed somatic mutation events and can be used to generate testable biological hypotheses.
In preparation for analysis of The Cancer Genome Atlas (TCGA) somatic mutation data set [14
], a team of bioinformaticians, molecular biologists and clinical oncologists based at Memorial Sloan Kettering Cancer Center and Dana-Farber Cancer Institute developed a human-curated map of the molecular pathways involved in GBM (Figures S7 and S8 in [14
]; the original Cytoscape file can be downloaded from [69
]). Our network captures the majority of proteins and interactions in this map: 96% of proteins (70 of 73) and 69% of interactions (129 of 187).
The TCGA GBM screen captured 341 mutated genes, including both point mutations and copy number variations (CNVs). Of these genes, 38 (11%) are part of the original hand-curated GBM pathway, and 237 (70%) are in the FI network. Of these genes in the FI network, 36 are in the original GBM pathway (15%), and in addition, 108 directly interact with at least one of the curated GBM pathway genes, for a total of 42% of the somatic mutations. This degree of interaction between somatically mutated genes with the GBM pathway is far greater than would be expected by chance (P-value = 1.3 × 10-23 by the hypergeometric test), suggesting that the FI network provides an effective way to enrich the hand-curated GBM pathway for additional genes involved in the disease.
We then added these potential proteins and interactions to the GBM pathway map to extend it. In order to do so, we chose proteins that were found to have one or more somatic mutations in the GBM screen, and had direct interactions with one or more of the proteins in the hand-curated GBM pathway. In this way we were able to extend the hand-curated pathway from 73 proteins and 187 interactions to 181 proteins and 768 interactions. A total of 581 FIs were added between pathway components and new mutated protein interactions (an increase of 148% for proteins and 311% for FIs). Figure shows the original hand-curated map after extending it with predicted and curated FIs from the FI network involving mutated genes. Interactions derived from curated pathways are represented as solid lines (with arrows for FIs involved in catalysis and activation, and with a 'T' bar for those involved in inhibition), while those predicted from the NBC are shown as dotted lines. Many mutated proteins interact with more than one pathway component. For the purposes of readability, Figure shows only proteins that interact with one pathway component. A larger diagram showing the fully extended map is available in Figure S2 in Additional file 1
Figure 3 Overlay of predicted functional interactions onto a human curated GBM pathway from the TCGA data set. Many genes can interact with multiple pathway genes. In this diagram, only genes interacting with one pathway gene are shown to minimize diagram clutter. (more ...)
A total of 23 of the FIs added to the GBM pathway in Figure were predicted by the NBC. To validate the accuracy of these predicted FIs, we searched the published literature for evidence supporting that two genes in the predicted FIs are indeed functionally related. Table lists the literature references that support these interactions. Out of 23 FIs, a total of 18 (78%) are supported by literature evidence for a functionally significant event. One FI (ROS1-EGFR) has no literature evidence supporting it, and the remaining four are confirmed physical interactions but have no evidence of functional significance. These results suggest that the predicted FIs are sufficiently reliable to be safely integrated into known pathways for systematic analysis.
Literature references for predicted FIs added to human curated GBM pathway from the TCGA GBM data set
A detailed examination of the extended GBM pathway can lead to hypotheses that connect the observed sequence alteration in the TCGA data set to known biological pathways. For example, NUP50 is required for degradation of CDKN1B protein [70
]. Copy number deletion in NUP50
, which occurs in three TCGA GBM samples, may inhibit the degradation of CDKN1B and impact the cell cycle process. For another example, tenascin-C (TNC) protein is a ligand for epidermal growth factor receptor (EGFR) [71
]. Three re-sequenced GBM samples have found TNC mutations, which may disturb the RTK/RAS signaling pathway via its interaction with EGFR.
It needs to be pointed out that the directionality of the interaction should be taken into account when using the FI network to frame hypotheses. For example, two of the pathway FIs around TP53, BAX-TP53 and GTSE1-TP53 were originally extracted from the KEGG human p53 signaling pathway [72
]. The BAX
genes are transcriptionally upregulated by TP53 protein. Though it is not annotated in the original KEGG database, there is evidence showing that GTSE1 protein can regulate TP53 protein's activity and localization [73
]. However, there is no evidence to suggest that the P53 pathway is affected by BAX protein, a protein involved in apoptosis [74
]. Hence, mutations in BAX in a particular tumor do not support an etiology involving P53 signaling, but instead might point to events downstream of P53. The same caveat applies to predicted FIs as well.
Clustering of GBM sequence-altered genes in the extended FI network
The previous section described how the FI network can be used to enhance and extract novel hypotheses from a previously created hand-curated disease pathway. In this section, we illustrate how studies of distributions of altered genes in the GBM samples in the FI network can assist in genome-wide functional analysis when a preexisting disease pathway is unavailable.
Both the TCGA [14
] and Parsons et al.
] GBM studies identified recurrent patterns of somatic gene mutations involving multiple classical signaling pathways using a manual process of inspection and correlation to the literature and a variety of pathway databases. Here, we use network community analysis to automatically identify network modules that contain genes and their products that are involved in common processes.
The edge-betweenness algorithm [75
] has been used to find network modules in protein interaction networks [76
]. We applied this algorithm to search for FI network modules for sequence-altered genes identified in the two GBM data sets. Starting with the TCGA data set, we collected 341 mutated and CNV genes from 91 GBM samples that have been re-sequenced in that study. A total of 237 of these genes (70%) were in the FI network. Of these, 168 have mutual FIs and are interconnected. We built a subnetwork around these 168 genes, applied the edge-betweenness network clustering to it, and obtained 17 network modules, 6 of which were greater than size 4 (Figure ).
Figure 4 Edge-betweenness network clustering results for the altered genes from the TCGA data set. Gene nodes in different clusters are displayed in different colors. GO cellular component annotation for clusters 0 and 1 are labeled in the diagram to show the (more ...)
The sizes of the first two modules (modules 0 and 1) are 63 and 50, respectively. The distribution study showed that 76 out of 91 GBM samples have altered genes in both module 0 and module 1 (84%, P
-value < 1.0 × 10-4
from permutation test). As a cross-validation test, we projected 22 samples from the discovery screen in the Parsons data set, which provided both somatic mutation and CNV data, onto these network modules. The result showed that 68% (15 out of 22) have altered genes in both module 0 and module 1 from the TCGA data set (P
-value < 1.0 × 10-4
). We also did a reciprocal test by applying the edge-betweenness clustering algorithm to a subnetwork composed by altered genes from the Parsons data set, and checking sample distributions from both GBM data sets in the network modules. The results are similar to our results in the TCGA data set: 77% (P
-value = 0.0002) of GBM samples in the Parsons data set, and 71% (P
-value < 1.0 × 10-4
) in the TCGA data set have altered genes in two corresponding modules (Figure S3 in Additional file 1
To see what biological features these two modules may connote, we annotated these two modules using pathways and GO terms. GO cellular annotation enrichment assay indicated that module 0 mainly corresponds to proteins present in the cytoplasm and plasma membrane, while module 1 mainly involves gene products present in the nucleus. Many pathways can be assigned to these two modules, but it is clear that module 0 is mainly related to signaling transduction pathways while module 1 is related to the cell cycle, DNA repair and pathways involved in chromosome maintenance (Table S2 in Additional file 1
). The fact that most of the GBM samples have altered genes in both modules implies that these two major modules are acting cooperatively in establishing and/or maintaining the GBM phenotype, and suggests that the development of GBM cancers involve malfunctions in both signaling transduction and cell-cycle regulation.
Our FI network is composed of a combination of curated FIs and predicted FIs. To determine whether the distribution of altered genes is robust, we checked the above results against FI network modules composed of FIs derived from curated FIs only. The results are similar to those obtained using the integrated FI network except that network modules 0 and 1 are smaller than the modules built with both predicted and pathway FIs (results not shown). Figure shows that many mutated genes are brought into modules 0 and 1 based on predicted FIs only, which are shown with dashed lines.
To further explore the distribution of mutations among the network modules, we performed a hierarchical clustering of the TCGA GBM samples based on the occurrence of altered genes in the modules (Figure ). From this clustering, we obtain five sample clusters of size 61, 13, 6, 9, and 2, respectively. Three types of samples were used in the original TCGA screening (rightmost column of Figure ): recurrent samples (15, blue), secondary samples (4, red), and primary samples (72, green). Sample cluster 0, which has a signature of mutations in both network modules 0 and 1, is enriched in primary tumor samples (P-value = 0.055 from Fisher test). In contrast, sample cluster 1, which has additional mutations involving network modules 8, 3, 9, 7 and others, is enriched in samples from tumor recurrences and metastases (P-value = 0.026). Indeed, all but one of the four metastatic samples can be found in sample cluster 1 (P-value = 0.0086).
Figure 5 Hierarchical clustering of GBM samples in the TCGA data set based on altered gene occurrences in the network modules identified by the edge-betweenness algorithm. The rows are samples, while the columns are 17 network modules. In the central heat map, (more ...)
In the original TCGA paper [14
], seven of the recurrent or metastatic samples were labeled as 'hyper-mutated' because of their much higher rate of somatic mutation. We found that except for one sample (TCGA-02-0099) located in sample cluster 0, all of the other six samples are in cluster 1 (P
-value = 1.7 × 10-5
). These results illustrate how the mutated network modules can be used to differentiate cancer samples.
Defining a GBM core cancer network
It is expected that multiple false positive ('passenger') genes exist in the set of sequence-altered genes identified from the GBM samples. It is also expected that true positive ('driver') GBM-related genes should occur more often in GBM samples than by chance. We plotted the percentage of altered genes versus samples for both GBM data sets (Figure ), and compared this distribution against what would be expected by random assignment of genes to samples. There are two phases in the distribution of altered genes across samples. In the first phase, involving gene alterations occurring between two to five samples, there is sharing of fewer altered genes than would be expected by chance. In the second phase, involving genes altered independently in six or more samples, there are more altered genes shared among the samples than would be expected by chance. This result can be explained if there exist a minimum number of driver genes that must be mutated in order to produce GBM, and that this 'GBM core' tends to be recurrently mutated in independent samples. Figure also shows that the average shortest path among shared genes from GBM samples decreases versus sample numbers in contrast to random samples, which implies that GBM candidate genes tend to be closer in the FI network than by chance (see below).
Figure 6 Plots of altered genes versus samples. The horizontal axis is the sample numbers, and the left vertical axis is the percentage of altered genes occurring in samples related to total altered genes. The right vertical axis is the average shortest path among (more ...)
To visualize sequence-altered genes and further define the core set of genes in the GBM samples, we collected genes altered in at least two samples to reduce the number of false positive GBM candidate genes, performed hierarchical clustering among them to identify a set of highly interconnected candidates, and then selected and built subnetworks containing >70% of altered genes (Figure ) by adding the minimum number of linker genes to form a fully connected subnetwork.
Figure 7 Subnetworks for GBM clusters. (a) The TCGA cluster. (b) The Parsons cluster. Shared GBM candidate genes are shown in yellow, non-shared candidate genes in aqua, and linker genes used to connect cancer genes in red. The node size is proportional to the (more ...)
In the TCGA data set, 164 altered genes occurred in two or more GBM samples, 98 (60%, P-value = 3.2 × 10-7) of which were in the FI network. Of these, 71 are in the GBM subnetwork (72%, P-value < 0.001 from permutation test). An average shortest distance calculation (Table ) shows that genes in this cluster are linked together much more tightly than would be expected by chance: 2.29 for subnetwork genes versus 3.83 for a similarly sized random set of genes treated in the same way as the cancer subnetwork. In the Parsons data set, 111 genes occur in two or more GBM samples, 65 (59%, P-value = 8.4 × 10-5) of which are in the FI network. Of these, 46 are in the GBM cancer cluster (71%, P-value < 0.001 from permutation test). Similar to the TCGA data set, the average shortest path among these genes is shorter than by chance (2.76 versus 3.82, P-value < 0.001).
Clustering of GBM-related altered genes occurring in two or more samples in the FI network
In the average shortest path calculation, a potentially confounding factor in the TCGA data set is that 601 genes pre-selected for sequencing may be more tightly interconnected than average. Indeed this is the case. When we performed the permutation test using these 601 pre-selected genes, we obtained an average shortest path of 2.40, which is shorter than the genome-wide average, but still longer than the length of 2.29 calculated for the subnetwork formed by recurrently mutated genes (P-value = 0.023; connection degrees have been considered in permutation test (see below)). This consideration does not apply to the Parsons set, which used an unbiased resequencing approach.
In summary, results from both GBM data sets indicate that more than 70% of the recurrently mutated genes are more tightly interconnected than expected by chance, and occupy a small corner of the large FI network space.
We found that the average connection degrees in the GBM clusters are higher than the average connection degree in the whole FI network (40 based on the biggest connected graph component using gene names): 87 for the TCGA cluster (P-value = 1.3 × 10-5 from t-test), and 60 for the Parsons cluster (P-value = 0.13). The result that the average shortest path among altered genes in cancer clusters is shorter than by chance may be an ascertainment bias due to the higher connection degrees in the cancer clusters resulting from the intensive study of signal transduction pathways, to which most GBM candidate genes belong. To determine whether the differences in average shortest paths between the cancer clusters and randomly selected genes are due entirely to the difference in degree, we performed an additional permutation test in which the genes picked were stratified by degree in order to match the distribution of the cancer gene sets (Table , Degree-based permutation column). This correction reduced, but did not eliminate, the differences in average shortest path between the cancer gene sets and randomly selected genes, and the differences remained statistically significant: both P-values < 0.001.
Average shortest distance for GBM clusters
The reason why the average shortest paths for the TCGA data set are smaller than the those calculated for the Parsons set for both the cancer cluster and degree-based permutation results is that re-sequenced genes in the TCGA data set number 601 in total, which are pre-selected and believed to be more cancer-related, while the Parsons paper resequenced 20,661 protein coding genes.
Looking at the two GBM cancer subnetworks in more detail, each subnetwork consists of GBM candidate genes ('cancer genes') plus the minimum number of interacting genes necessary to interconnect them ('linker genes', shown in red in Figures and ). The TCGA subnetwork contains 77 genes, 5 of them linkers, while the Parsons subnetwork contains 62 genes, 14 of them linkers. Since many of our FIs were extracted from human curated pathways, it is easy to superimpose pathways back to these subnetworks to see what pathways are involved in these cancer genes. Many pathways are statistically significantly hit by these genes. Figure shows four pathways: focal adhesion, signaling by platelet-derived growth factor (PDGF), p53, and cell cycle (false discovery rate <8.0 × 10-4). As illustrated in Figure , there are extensive overlaps among, and cross-talk between, these pathways.
Subnetworks with pathways annotated for GBM clusters. Many pathways are hit by GBM candidate genes. Only four of them are labeled for two GBM clusters in this diagram to simplify the diagram. Colors and symbols are as in Figure 6.
It is surprising that the two GBM studies identified a relatively small number of GBM candidate genes in common: just 15 out of a total of 260 unique genes are shared (6%) for all GBM candidate genes (13 out of a total of 150 unique genes (9%) for altered genes in the FI network). Intriguingly, 12 out of 13 shared genes in the FI network are present in the two cancer subnetworks we built independently for the two data sets (Figures and , shared genes are in yellow; P-value = 0.0014 based on permutation test), suggesting that the GBM cancer clusters capture the common candidate genes. Interestingly, with the exception of COL3A1, all the shared genes directly interact with each other, suggesting that they form the core of a GBM pathway, and that non-shared cancer genes are extensions of the core network.
To further narrow down the list of candidate genes to those that are likely to be drivers, we used the results shown in Figure to investigate candidate genes altered in eight or more GBM samples in the TCGA data set, and five or more samples in the Parsons data set. In the TCGA data set, a total of 20 genes are altered in 8 or more GBM samples. Of these 20 genes, 13 are in our FI network, and 10 are displayed in Figure : CDK2A, CDK2B, CDK4, EGFR, MDM2, NF1, PTEN, RB1, PIK3R1, and TP53. In the Parsons data set, 14 genes occur in 5 or more samples, 10 are in the FI network, and 9 are displayed in Figure : CDKN2A, CDKN2B, EGFR, IFNA1, IFNA2, IFNA8, IFNE1, PTEN, and TP53. Out of these genes, five are shared between these two data sets: CDKN2A, CDKN2B, EGFR, PTEN, and TP53 (P-value = 5.8 × 10-13). The fact that these genes are altered in multiple samples and shared in two studies further indicates the existence of a GBM core network.
Application of the FI network to other cancer types
In this section, we test whether the network patterns found in the GBM samples can also be found in other cancer types. We applied our FI network to the genome-wide data sets for breast [10
], colorectal [10
] and pancreatic cancers [11
], and found similar patterns of clustering of mutations into network modules.
In the breast and colorectal data sets, 18,191 genes based on 20,857 transcripts were re-sequenced for 11 breast and 11 colorectal tumors during the discovery phase. Mutated genes found in the discovery phase are further validated with 24 additional samples of the same tumor type. For network module analysis, we collected mutated genes from the 11 discovery phase samples from both breast and colorectal cancers, built interaction subnetworks based on the FI network, clustered the subnetworks using edge-betweenness, and compared the resulting network modules to the distribution of patient samples.
For breast cancer samples, 1,023 genes were mutated in 11 tumor samples. Of these, 524 genes are in our FI network (51%). The edge-betweenness algorithm generated 33 clusters, of which 12 were of size 5 or greater. The sizes of the first three modules are 55 (module 0), 49 (module 1), and 26 (module 2) (Figure S4 in Additional file 1
). The sample distribution analysis showed that all 11 breast samples had mutated genes in modules 0 and 1, and 9 out 11 breast tumor samples had mutated genes in modules 0, 1, and 2. A GO cellular component annotation enrichment analysis showed that genes in module 0 are mainly from the extracellular region, cytoplasm and the plasma membrane, while genes in module 1 are found mainly in the nucleoplasm, nucleus and nucleolus (Table S3 in Additional file 1
). Pathway annotation shows that genes in module 0 are involved in integrin and focal adhesion pathways, while genes in modules 1 and 2 are involved in transcription, cell cycle and DNA repair (Table S3 in Additional file 1
For colorectal cancers, 766 genes were mutated in 11 tumor samples. Of these, 410 genes are in our FI network (54%). We detected 31 modules from the edge-betweenness algorithm, 8 of which were of size 5 or greater. Nine out of 11 colorectal cancer samples had mutated genes involving three network modules, modules 0, 1 and 4 (Figure S5 in Additional file 1
). A GO cellular component enrichment analysis (Table S4 in Additional file 1
) demonstrated that module 0 is enriched for plasma membrane and cytosol, while module 1 is enriched in GO terms describing the extracellular region, and module 4 is enriched in terms for the nucleoplasm and nucleus. Compared with breast cancer, module 0 in breast cancer has been split into two modules in colorectal carcinoma. Pathway annotation showed that genes in module 0 are highly enriched in components of the endothelin and Nerve Growth Factor (NGF) pathways, in contrast to breast cancer and GBM, which show no such enrichment. Genes in module 1 are involved in focal adhesion and integrin pathways, while genes in module 4 are involved in DNA repair (Table S4 in Additional file 1
), all of which are also enriched in breast cancer and GBM.
As with the two GBM data sets, we performed a hierarchical clustering analysis on genes mutated in at least two or more samples. In this analysis, we used samples from both the discovery and validation phases to increase the sample numbers. Table S5 in Additional file 1
lists the numbers of genes mutated in two or more samples, mutated genes in the whole FI network, and mutated genes in the clusters from hierarchical clustering; these numbers indicate that most of the mutated genes are clustered together in the FI network, as was previously observed with GBM candidate genes. Also similar to the results from the GBM analysis, the majority of cancer candidate genes in both breast and colorectal cancers are closer than would be expected by chance (Table S6 in Additional file 1
). Figures S6 and S7 in Additional file 1
show these clusters with their pathway annotations. As with GBM, many cancer-related pathways are significantly enriched in these clusters.
We performed a similar network module and hierarchical clustering analysis on a pancreatic data set [11
]. Our results show that similar network patterns can be found in pancreatic cancers (Tables S5, S6, and S7, and Figures S8 and S9 in Additional file 1
). We noted that 75% (P
-value = 0.38) of pancreatic samples (18 out of 24 samples) have mutated genes in module 0 (nucleus and cytosol) and module 1 (extracellular region), and 71% (P
-value = 0.24) of samples (17 samples) have mutated genes in module 0 and module 2 (plasma and integral to membrane) (Figure S8 in Additional file 1
As in the GBM analysis, we used permutation testing to assess whether more of the breast, colorectal, or pancreatic cancer samples had mutations involving these modules than would be expected by chance. However, due to the small number of samples available and the relatively large number of mutations per specimen (32 to 49 altered genes per specimen in the breast, colorectal and pancreatic sets, versus 22 per specimen in GBM), we failed to demonstrate statistically significant overrepresentation of specimens with mutated genes in these modules.
In conclusion, we found that most samples from breast, colorectal or pancreatic cancers have sequence-altered genes involving multiple modules localized in different cellular localizations, and that most cancer candidate genes in the FI network are closer than would be expected by chance. The results are broadly similar to those seen with GBM, but there are many intriguing differences as well in the identity of the pathways involved.
Software tools for accessing the functional interactions
To make our FI data sets available to general researchers, we developed a web-based application. This application was built upon the current Reactome SkyPainter using Google Web Toolkit (GWT) and is available at [79
]. Figure is a screenshot of the application when first launched. At the top of the application is a Google-map style view of all reactions manually curated for Reactome; at the bottom is a hierarchical tree of all pathways, organized by biological process.
Front page of the web application for predicted functional interactions.
To find FIs, users can upload lists of UniProt accession numbers or gene names for proteins or genes of interest. The application will then highlight pathways involving FIs with the proteins or genes of interest in both the reaction map and the pathway hierarchy. When the user selects a reaction that has been 'hit', a dialog appears that provides detailed information on the relationship between the reaction, its components, and FIs involving those components (Figure ). The user can also select a hit pathway in the hierarchical tree to bring up a pathway diagram to show pathway components and its FIs (Figure ).
Views of predicted functional interactions. (a) FIs in a reaction diagram. (b) FIs in a pathway diagram.
The entire set of FIs can be downloaded as a MySQL database dump (Additional file 2
) or in tab-delimited files (Additional file 3
) from this web application. There is no restriction on their use or redistribution beyond citing the source of the information.