ConsensusPathDB--a database for integrating human functional interaction networks.

ConsensusPathDB is a database system for the integration of human functional interactions. Current knowledge of these interactions is dispersed in more than 200 databases, each having a specific focus and data format. ConsensusPathDB currently integrates the content of 12 different interaction databases with heterogeneous foci comprising a total of 26,133 distinct physical entities and 74,289 distinct functional interactions (protein-protein interactions, biochemical reactions, gene regulatory interactions), and covering 1738 pathways. We describe the database schema and the methods used for data integration. Furthermore, we describe the functionality of the ConsensusPathDB web interface, where users can search and visualize interaction networks, upload, modify and expand networks in BioPAX, SBML or PSI-MI format, or carry out over-representation analysis with uploaded identifier lists with respect to substructures derived from the integrated interaction network. The ConsensusPathDB database is available at: http://cpdb.molgen.mpg.de.

INTRODUCTION

Functional interactions between cellular entities like genes, proteins, metabolites, etc. are the key drivers of cellular functions. Different experimental methods like chromatin immunoprecipitation (1) and two-hybrid assays (2), among others, have generated large amounts of interaction data for many organisms, usually stored in interaction databases. In the past few years, the analysis of interaction networks has become crucial to understand biological processes and their dysfunctions in human diseases. For example, reaction networks build the basis of computational models in systems biology. Analyses combining expression and interaction data have recently been used to reveal previously unknown disease mechanisms (3,4). Thus, collecting comprehensive human interaction data is the key to gain new insights into cell biology.

While for several model organisms like Saccharomyces cerevisiae (5) and Caenorhabditis elegans (6), such comprehensive functional interaction networks are available, the larger part of the human interactome remains undiscovered (7). Even worse, the existing knowledge on human functional interactions is dispersed in over 200 interaction databases, each of which has a specific data format, focus and bias (8). Most integration efforts with respect to interaction data so far have focused on merging homogeneous interaction networks. For example, APID (9), MiMI (10) and UniHI (11) integrate protein–protein interaction networks from multiple sources. However, the integration of heterogeneous interactions remains a challenge. Such integration is highly relevant because the resulting network reflects multiple functional aspects of the nodes at the same time (like regulatory relations, physical interactions, catalyzed reactions), and thus constitutes a more complete picture of the living system.

We have developed ConsensusPathDB, a database for integrating human molecular interaction networks, in order to address such a comprehensive integration of interaction data. The integrated content comprises different types of functional interactions that interconnect diverse types of cellular entities. In order to gain an immediate critical number of interactions, we have focused primarily on the integration of existing database resources although our schema has also been used for additional manual upload of experimental interactions. Currently, the database contains human functional interactions, including gene regulations, physical (protein–protein and protein-compound) interactions and biochemical (signaling and metabolic) reactions, obtained by integrating such data from 12 publicly accessible databases (referred to as source databases): Reactome (12), KEGG (13) (metabolic reactions only), HumanCyc (14), PID (http://pid.nci.nih.gov), BioCarta (http://www.biocarta.com), NetPath (http://www.netpath.org), IntAct (15) (data from small-scale experiments only), DIP (16), MINT (17), HPRD (18), BioGRID (19) and SPIKE (20). In this article, we describe the methods used for data integration, the database schema, as well as the main functions of the web interface.

RESULTS

Mapping of functional interactions

In order to assess the content overlap of the source databases and to reduce redundancy, we have applied a method to merge identical physical entities and identify similar interactions. The method is straightforward and efficient for the integration of networks from any single species. Simple physical entities of the same type (genes, proteins, transcripts, metabolites) are compared on the basis of common database identifiers like UniProt (21), Ensembl (22), Entrez (23), ChEBI (24), etc. Since different databases tend to annotate physical entities with different identifier types (e.g. some databases annotate proteins with UniProt identifiers, others with Ensembl identifiers), we first translated the annotations to a uniform identifier type, which is a UniProt entry name in case of proteins, Ensembl gene ID in case of genes and transcripts, and KEGG/ChEBI ID in case of metabolites. Protein complexes are compared according to their individual protein composition. Simple physical entities with the same identifier, and complexes with the same composition, are merged in ConsensusPathDB. Information provided by the according source databases for the merged entities is stored in a complementary manner.

Functional interactions of physical entities are also compared with each other. Here, we distinguish between primary and secondary interaction participants. Primary participants are substrates and products in case of biochemical reactions, interactors in case of physical interactions and target genes in case of gene regulation. All other participants, e.g. enzymes and interaction modifiers, are secondary participants. If the primary participants of two or more interactions match, these interactions are considered similar. Two similar interactions may have different stoichiometry, modification and/or localization of the participants. To allow for flexibility, similar interactions are marked as such in the database, but the decision whether they should be considered identical despite mismatching details is left to the user and depends on his specific problem. Moreover, ConsensusPathDB does not provide any additional quality control filters. All interactions provided by the different database sources are treated in the same way. The results of our mapping method applied on the data from the source databases mentioned above are summarized in Table 1.

Table 1.

Database content and pairwise database overlaps in terms of matching physical entities and similar interactions

	Reactome	Kegg	Humancyc	Pid	Biocarta	Netpath	Intact	Dip	Mint	Hprd	Biogrid	Spike
Database overlaps in terms of matching physical entities
Reactome	6831	1037	893	560	796	241	1067	352	1217	2041	1704	1506
Kegg	1037	3271	1482	59	383	15	225	23	329	1297	540	508
Humancyc	893	1482	3892	253	688	110	553	138	701	1656	1065	947
Pid	560	59	253	3614	764	354	799	371	886	1357	1263	1027
Biocarta	796	383	688	764	3387	302	892	392	1011	1695	1477	1346
Netpath	241	15	110	354	302	739	392	173	450	615	568	479
Intact	1067	225	553	799	892	392	4138	576	2660	3546	3016	3237
Dip	352	23	138	371	392	173	576	964	637	865	828	707
Mint	1217	329	701	886	1011	450	2660	637	5849	4684	3939	3915
Hprd	2041	1297	1656	1357	1695	615	3546	865	4684	10165	7185	5973
Biogrid	1704	540	1065	1263	1477	568	3016	828	3939	7185	8696	5248
Spike	1506	508	947	1027	1346	479	3237	707	3915	5973	5248	7012
Database overlaps in terms of similar interactions
Reactome	4129	262	122	101	80	34	97	31	51	304	212	118
Kegg	262	1655	213	0	4	0	0	0	0	0	0	0
Humancyc	122	213	1322	0	2	0	1	2	2	7	4	2
Pid	101	0	0	3510	264	96	63	46	73	333	243	186
Biocarta	80	4	2	264	2221	67	51	35	43	139	114	173
Netpath	34	0	0	96	67	1915	57	34	123	821	510	224
Intact	97	0	1	63	51	57	6880	312	2714	3216	1583	4080
Dip	31	0	2	46	35	34	312	1218	389	821	656	418
Mint	51	0	2	73	43	123	2714	389	13187	7211	4428	5676
Hprd	304	0	7	333	139	821	3216	821	7211	37955	19463	11351
Biogrid	212	0	4	243	114	510	1583	656	4428	19463	28038	10303
Spike	118	0	2	186	173	224	4080	418	5676	11351	10303	22232

Current versions of the integrated databases are: Reactome 25, KEGG 47.0, HumanCyc 12.1, PID 2008_06_10, BioCarta 2008_01_08, NetPath downloaded on 6.7.2008, IntAct 2008-06-27, DIP 2008-01-13, MINT 2008-05-19, HPRD I_090107, BioGRID 2.0.42 and SPIKE downloaded on 6.7.2008.

Biological pathways in ConsensusPathDB are represented as sets of interactions, whose compositions are adopted from the source databases. This means that individual interactions rather than entire pathways from different databases are compared with each other. This was necessary because the concept of pathway is defined very differently in the respective source databases and the pathway boundaries are rather unclear. For example, KEGG's Glycolysis/gluconeogenesis pathway contains 31 reactions whereas Reactome's Glycolysis contains 10 reactions.

Database schema and content

Interaction data in ConsensusPathDB currently originates from 12 interaction databases and comprises physical interactions, biochemical reactions and gene regulations. The data are gathered in different formats: database dumps, standard interaction file formats like BioPAX (25) or PSI-MI (26), database-specific flat or XML files, or is retrieved through web services. All interaction data are translated to the schema of our relational database and are integrated using the method described above. The database has a graph-like architecture: its three main classes are ‘Interaction’, ‘Physical entity’ and ‘Edge’. The first two classes store information about different interaction types and physical entities, respectively, and edges connect both, indicating that a specific entity participates in a specific interaction. Edges are carriers of information on the particular role, stoichiometry, state and compartment of the physical entity in the particular interaction, if such information is available from the source databases. Several other classes in our database contain information about the source databases, external identifiers of entities or interactions, literature references, cellular compartments, biological pathways, etc. Importantly, the source of physical entities and interactions is always recorded, which allows linking to the original data in the source database.

The database content is updated automatically on a regular basis with the latest releases of the source databases.

Web interface

ConsensusPathDB is accessible via the Internet at http://cpdb.molgen.mpg.de. The main functions of the web interface are described here and summarized graphically in Figure 1. The web interface gives information about the database and its current content summarized in overlap tables (Table 1). It is obvious that the content of existing human interaction resources is partially overlapping but, more importantly, complementary to a large extent. The web interface contains documentation in form of a regularly updated tutorial intended to guide the user through the different functions of the web interface and provide further details (Supplementary Material 1).

Figure 1.

Graphical summary of the main functions of the ConsensusPathDB web interface. The search for interactions of specific entities or pathways, search for interaction paths, overrepresentation analysis and model upload induce specific interaction subnetworks that can be visualized in the dynamic visualization environment and exported in BioPAX format for further computational analysis, or as images. Different colors of nodes and edges refer to different types of interactions or entities and different source databases, respectively. The background image shows a screenshot of the ConsensusPathDB start page.

Search functions

The user can search for interactions of specific physical entities or pathways by name or database identifiers. Rules according to which interactions with the same primary participants but different compartment/modification/stoichiometry information are to be merged can be specified here. Interactions of interest are merged with their similar counterparts according to these rules (Figure 2), and are displayed as network graphs in the visualization environment of the ConsensusPathDB web interface. In these graphs, two classes of nodes exist: physical entity nodes and interaction event nodes. Node colors encode the specific object type (protein, metabolite, etc.; physical interaction, biochemical reaction, etc.). Edges connect interactions with physical entities and indicate which physical entities participate in which interactions. Different edge styles encode the roles of the entities in the interactions, and edge colors refer to the source of this annotation. The network graphs are automatically generated and dynamical. For example, interactions can be removed from the graph, or new ones can be added by expanding a specific physical entity. Details about nodes, like alternative names and external identifiers, are shown in tool-tips. Physical entities can be easily located in large graphs by searching them by name. Network graphs can be exported as image or as a computer-readable file (currently, in BioPAX level 2 format). In the latter case, networks extracted from ConsensusPathDB can be used as input to various software programs for further analysis, e.g. for modeling and simulation studies.

Figure 2.

Illustration of the mapping procedure. Two biochemical reactions that are identical according to their primary participants and the user-specified mapping criteria are mapped. The reaction A + B -> C + D, catalyzed by enzyme E1, originates from database X. A similar reaction, catalyzed by enzymes E1 and E2, originates from database Y. Since both reactions are identical according to the user-specified rules, the interactions are merged and visualized reflecting the source annotations.

Apart from the search for interactions of single physical entities, the user can search for shortest paths of interactions connecting two distinct physical entities in the overall interaction network stored in the database. If a path between the entities of interest exists, it can be further constrained by forbidding certain intermediates. Interaction paths can be visualized in the visualization environment.

Overrepresentation analysis

Using the web interface, over-representation analysis can be carried out with gene sets and functional modules derived from two methods. The first method incorporates pathway definitions as given by the source databases. The second method is based on functional modules defined by proximity measures using the integrated network structure. It includes a node (for example a gene) as the module centre and its neighbours within a user-specified interaction proximity radius. For example, modules with radius 1 include the central gene and all its direct neighbors (genes that appear in a functional interaction together with the central one), and modules with radius 2 additionally include the neighbours of the neighbours of the central gene. Moreover, interconnectivity of module members can be specified by the user with the clustering index. For each predefined module, a P-value is calculated based on the hypergeometric distribution. The P-value reflects the significance of the observed overlap between the input gene list and the module's members as compared to random expectations. A small P-value indicates that more of the module's members are present in the input list than expected by chance. If, for example, the input list contains differentially expressed genes from a case–control study, overrepresentation analysis may point to pathways and functional sub-networks that are dysregulated in the disease state. Overrepresented modules are shown in downloadable lists sorted by the significance of over-representation (P-value) and modules of interest can be visualized in order to see the specific relations between their members.

Import, export and expansion of networks

Apart from the possibility to export interaction networks from the visualization environment, the user can upload an interaction network file in any of three different formats: BioPAX (level 2), PSI-MI (level 2.5) and SBML (level 2) (27). If the physical entities from the uploaded file are annotated with external identifiers, these entities and their interactions are mapped to the interaction network in the database and matching information is indicated. Thereby, the model can be validated against existing interaction knowledge stored in the database, and extended in the context of the database content.

DISCUSSION

ConsensusPathDB is a database that integrates interaction data from heterogeneous resources for human functional interactions in order to create a more complete and less biased picture of the cellular interactions. It can be used in many ways—for example, to retrieve network topologies necessary for mathematical simulations, to interpret gene lists, to assess the distribution of interaction knowledge across interaction databases, or to carry out topological analysis with the integrated human interaction network, which, according to our analysis, provides quite different results compared to topological analysis of the separate interaction databases (results will be published elsewhere).

Although we apply our best efforts to collect and integrate available interaction data, more data sources remain to be integrated. Importantly, human gene regulatory interactions are currently weakly represented in our database (283 interactions) because on the one hand, such data are rare compared to other interaction types (e.g. protein–protein interactions), and, on the other hand, access to the majority of existing gene regulatory data is mostly limited by license constraints (28, 29). Apart from physical interactions, biochemical reactions and gene regulatory interactions, other relations exist between cellular entities that will be integrated into the database, for example, genetic and epigenetic relationships or relations with respect to experimental co-regulation patterns or to more general co-occurrences. Some of these relations are present e.g. in the STRING database (30) and can be integrated into ConsensusPathDB. Although ConsensusPathDB is currently focused on Homo sapiens, the integration of data from other species is an ongoing issue since it will reveal conserved and species-specific cellular processes on the interaction level.

Interaction maps are of great importance in many areas of life sciences, for example in systems biology and molecular medicine. However, more work remains to be done in order to assemble a complete map of the human functional interactome. ConsensusPathDB marks a first step towards achieving this goal by collecting, integrating and interpreting heterogeneous interaction knowledge.

AVAILABILITY

ConsensusPathDB is available freely to academic users via http://cpdb.molgen.mpg.de. Data in form of flat files are available upon request (please contact kamburov@molgen.mpg.de).

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

FUNDING

This work was supported by the EMBRACE and CARCINOGENOMICS projects that are funded by the European Commission within its 6th Framework Programme under the thematic area ‘Life Sciences, Genomics and Biotechnology for Health’ (LSHG-CT-2004-512092 and LSHB-CT-2006-037712); 7th Framework Programme project APO-SYS (HEALTH-F4-2007-200767); German Federal Ministry of Education and Research within the NGFN-2 program (SMP-Protein, FKZ01GR0472); Max Planck Society within its International Research School program (IMPRS-CBSC). Funding for open access charge: European Commission.

Conflict of interest statement. None declared.