Protein interaction databases (Proteomics)

Protein interactions occur in many different forms: Proteins interact with and modify each other as elements of signaling chains; they form large molecular machines with more than 100 protein and nonprotein elements, for example, the spliceosome; or, they assemble to provide structural elements of the cell. Proteins interact not only with each other but also with DNA, RNA, and other molecule types. Types of interaction may vary from direct physical contact to very indirect interactions, for example, participation in the same signaling pathway. In many contexts, protein states, in particular, posttranslational modifications, are essential for the function of protein assemblies.

The kind of data produced by experimental technologies to elucidate protein interaction data is similarly diverse, from direct, atom-level structural information by NMR (Pellecchia et al., 2002) or X ray via evidence for direct physical contact, for example, by yeast two hybrid assays (Phizicky et al., 2003) and information on participation in the same complex by tandem affinity purification (Rigaut et al., 1999) to indirect information, for example, colocalization.

With the introduction of technologies for high-throughput interaction screens (Uetz et al., 2000; Ito et al., 2001; Gavin et al., 2002; Ho et al., 2002), the number of interactions published in a single publication now varies over four orders of magnitude, from single interactions to more than 20 000 (Giot et al., 2003).

The biological and medical importance of protein interactions, high-throughput technologies, and the broadness of data types and experimental technologies have led to the creation of a significant number of protein interaction data collections, from simple HTML tables providing the results of a particular experiment to large databases providing complex tools. We subsequently introduce some criteria for the comparison of protein interaction databases, and describe some major, publicly accessible protein interaction databases on the basis of these criteria (Table 1), focusing on databases for experimentally, not computationally, derived datasets.

Confidence information: High-throughput experiments are often considered less reliable than detailed, low-throughput experiments. To support the user in estimating the reliability of individual interactions, the authors of large-scale interaction data sets often provide quality indications, for example, the grouping into reliability classes (Li et al., 2004) or numerical values (Giot et al., 2003). In addition, comparative analysis of different interaction data sets provides quality indications. In the simplest case, this is the number of times an interaction has been observed in independent experiments, but quality assessments are also based, for example, on the comparison against a reference data set considered to be correct (von Mering et al., 2002), shared subcellular location and cellular role annotation of interacting proteins (Sprinzak et al., 2003), comparison to RNA expression profiles, or similar interactions in paralogous sequences (Deane et al., 2002).

Table 1 Contents and features of publicly available protein interaction databases

Table 1 (continued)

Standards: The number of interactions in the yeast Saccharomyces cerevisiae interactome has been estimated to be about 10 000-26 000 (Sprinzak et al., 2003; Grigoriev, 2003). Taking into account all relevant species and the change of protein interactions depending on protein and cellular state, the number of interactions is nearly unlimited. As it is unlikely that any given database will be able to collect all available protein interaction data, collecting data from several, potentially specialized databases is essential to assemble a reasonably complete picture of the currently available protein interaction data in a given domain. In addition to providing a complete picture, such collections also provide the basis for interaction data reliability assessments through comparative analysis. However, such collection may be difficult and labor-intensive due to different data formats and annotations used by different databases. To improve this situation, major interaction data providers, among them BIND (Biomolecular Interaction Network Database), DIP (Database of Interacting Proteins), Hybrigenics, HPRD (Human Protein Reference Database), IntAct, MIPS, and MINT, in the framework of the HUPO Proteomics Standards Initiative (http://psidev.sf.net) (see Article 61, Data standardization and the HUPO proteomics standards initiative, Volume 7) have jointly developed the PSI-MI XML format, a community standard for the representation of protein interaction data. PSI-MI 1.0 provides a basic exchange format for protein interaction data (Hermjakob et al., 2004a); level 2.0 is under development (Orchard et al., 2004) and will provide additional features and an extension to additional molecule types, in particular, RNA and DNA. In addition to a standard format, PSI-MI provides a set of interaction-specific controlled vocabularies to standardize not only the format but also the contents of protein interaction data, for example, to select all data derived by a certain technology.

Visualization: In scientific publications, protein interactions are often visualized as groups of adjacent shapes with textual annotation. While such representations are intuitive, they are difficult to generate automatically. For automatically generated visualization, nearly all tools are based on the abstraction of proteins as nodes and interactions as edges in a graph. This representation allows the application of well-established graph layout algorithms to interaction networks, and provides the basis for additional analysis types, for example, interaction distance, or the identification of clusters of highly interconnected proteins. Although mostly based on this basic abstraction, tools differ significantly in the technical implementation, user interface, and, in particular, in methods to project additional information onto such interaction networks, for example, interaction confidence or Gene Ontology (Harris et al., 2004) terms annotated to the interacting proteins. In addition to the tools provided directly by databases, additional tools for graph-based interaction network analysis are provided by commercial and academic organizations, for example, Cytoscape (Shannon et al., 2003). Figure 1(a-e) shows the display of lsm7 (yeast) interactions in some visualization systems. Default settings have been used as much as possible.

Figure 1 (a) Visualization of lsm7 (yeast) interactions in DIP. (b) Visualization of lsm7 (yeast) interactions in Osprey. (c) Visualization of lsm7 (yeast) interactions in IntAct. (d) Visualization of lsm7 (yeast) interactions in MINT. (e) Visualization of lsm7 (yeast) interactions in BIND

Figure 1 (continued)

Figure 1 (continued)

The basic knowledge about protein interactions is currently increasing at a rapid pace, and a broad array of databases provides large data collections and powerful analysis tools, but the representation of protein interaction facts in near-text topic quality, taking into account interaction details such as protein state and dissociation constants, remains a challenge to the scientific community, both in terms of available tools and available data. Unlike other domains of molecular biology, in particular, DNA and protein sequences as well as macromolecular structures, protein interactions share the problem of data fragmentation in proteomics. Systematic deposition of data in public databases, and exchange of such data between databases, is only emerging. However, the standardization of interaction data through the PSI-MI standard, an increasing collaboration between protein interaction databases, and publicly available, open source analysis tools pave the way to public, user-friendly, and well-accessible protein interaction data resources, reflecting the huge biological significance of protein interactions.