The past and present periods of large-scale genome sequencing have brought an enormous wealth of protein sequences that makes managing, navigating and mining the data an area of research in its own right.

Protein evolution suggests domains as a convenient unit of classification. Mutations, insertions and deletions create sequence diversity inside a domain family. On a higher level, recombination events combine domains in different architectures to give the single- or multi-domain proteins we observe.

Various invaluable tools exist to reduce the diversity of proteins into a reduced set of protein domain families. Semi-automated methods such as Pfam (1), PROSITE (2) or SMART (3) extrapolate the information gained from known members of protein domain families by matching sequences to libraries of hidden Markov models (HMMs), profiles or patterns. Integrative projects such as InterPro (4) combine various primary sources to yield a summary view on protein sequences. Fully automated methods such as ProDom (5) or DOMO (6) apply algorithms to achieve a classification based on first principles.

Previously, we have introduced the Automatic Domain Decomposition Algorithm (ADDA) (7), a method for clustering very large sets of protein sequences. ADDA first splits sequences into domains and then organizes these domains into protein domain families. The classification is constructed in an entirely automated fashion from first principles and thus is not biased by human curation, but only limited by the applied algorithms.

We have applied ADDA to the set of all known protein sequences that are available in the major public databases. Using all sequences for clustering has the advantage of drawing the boundaries between protein domain families in a globally consistent manner. This is in contrast to scanning methods such as Pfam and SMART, where novel or hypothetical sequences are scanned against a library of HMMs or profiles. Matches at the borderline of significance can be due to a newly discovered remote relative or to spurious similarity. In such events, domain families have to be assigned case by case.

Here, we describe a database with a web interface that allows scientists to download and browse the results. The web interface lets a scientist explore the context of a protein sequence in the protein universe: its immediate neighbours as determined by pre-computed sequence similarity searches, and its remote homologues as determined by its domain composition. Alternatively, a scientist can browse the domain families to hunt for domain families of interest.

The database contains more than 1.5 million sequences from UniProt/Swiss-Prot, UniProt/TrEMBL (8), Ensembl (9), NCBI genomes (10) and other sources of protein sequences. The clustering yields 2.7 million domains, which are grouped into 123000 families. Of these, 40000 families have more than five members. The database is built in an entirely automated fashion and is updated regularly.

The domain and domain family definitions result from an automated clustering procedure applied to the set of all protein sequences in the major public databases. The process starts by removing sequences with >40% sequence identity to any other sequence from the input set (11). The remaining representative sequences are then aligned in an all-versus-all manner using the BLAST (12) program.

Representative sequences are split into domains by the ADDA algorithm. Domains are defined so that a minimum number of alignments are intersected by domain boundaries and these alignments cover domains as much as possible.

After splitting protein sequences, the resultant domains are grouped into families of related sequences by a single-linkage clustering algorithm. Domains joined by alignments are grouped into a family if their domain boundaries are consistent. In addition, domains are compared using sensitive profile–profile comparison. The merging process stops when the sequence similarity between domains drops below a threshold.

Finally, domain boundaries are mapped from the representative sequences onto all sequences in the database.

Quality control monitors two aspects of the clustering by comparing ADDA domains to curated databases of domain families like Pfam and SCOP (13). The correspondence of domain boundaries is checked by computing the relative overlap between ADDA and reference domains. ADDA tends to split conservatively; ADDA domains are, on average, larger than reference domains.

The correspondence of domain families is measured by matching each ADDA family to the best matching reference family and counting the relative frequency of other reference families in the ADDA family (selectivity) and the relative frequency of reference domains assigned to different ADDA families (sensitivity). On average, an ADDA domain family unifies 93% of the members of a Pfam family while containing only 5% contamination.

Charts and summary statistics are available on the web server.

In multi-domain proteins, domains of different protein families co-occur. Based on the observed architecture of protein sequences, domain families can be divided into two groups: mobile modules and associated families. Mobile modules are promiscuous and recombine with several other domain families. Associated families either always occur in single-domain proteins or are always associated with the same domain family. In the present release, there are 9252 mobile modules and 49455 associated families (Figure 1A). While the latter tend to be specific to a single kingdom (Archaea, Bacteria and Eukaryota), mobile modules have a larger taxonomic range (Figure 1B).

Figure 1 Overview of domain families in ADDA. The number of families is given in the last row of each category label. (A) Mobile modules, domain families that co-occur with a variety of different domain families, constitute only a fraction of all domain families. (more ...)

A domain family is declared to be structurally covered if one of its domains can be mapped onto a structure in the PDB database of protein structures (14). For each ADDA domain, we register the sequence overlap with domains from curated domain databases (Pfam, SCOP and InterPro). An ADDA domain family is classified as novel if <5% of its domains overlap with domains from the curated domain databases.

ADDA contains 3828 novel mobile modules that are not known to curated domain databases and for which there is no structural information available (Figure 1C). Novel domain families tend to have fewer members (<200) than well-known domain families. The number of novel domains in associated families is even larger comprising 40505 domain families.

The complete protein domain classification can be downloaded from http://ekhidna.biocenter.helsinki.fi:8080/downloads/adda. A web interface is available for browsing and querying at http://ekhidna.biocenter.helsinki.fi:9801/sqgraph/pairsdb.

The interface allows the user to query for protein sequences by identifier, accession number or sequence similarity. The sequence view shows the decomposition of a protein sequence into domains. Links allow the user to browse similar sequences in the direct neighbourhood of a query sequence (multiple alignments pre-computed using BLAST and PSI-BLAST) and to switch to the domain families to get all related sequences beyond the immediate neighbourhood.

Domain families of interest can be retrieved by custom queries. Attributes available for querying are the size of the family, its taxonomic spread, its structural coverage, the number of associated domains (querying for mobile modules), the overlap with other domain classifications (querying for novel domain families) and others. The domain family view includes a summary overview over the protein family and links to other domain classifications. The sequences of all domains in a domain family can be downloaded for local analysis. If structures are known in the family, the domain boundaries are mapped onto the structures for visualization with RASMOL (15).

In the browsing section, the user finds links to precompiled domain sets of interest, e.g. all exclusively eukaryotic mobile domains or a list of domain families without known structures (structural genomics targets). In addition, a genome browser allows access to all or a selection of domain families occurring in a genome.

The web server contains outgoing links to external databases for sequences and domains. For example, ADDA is linked to by the Dali domain database (16) and vice versa.

Links and attributes can be queried in numerous ways. One application of the database interface is to hunt for novel domain families. For example, typing ‘sapiens’ into the genome browser lists 28791 domain families in human protein sequences. The result set can be restricted to exclude domain families that have domains from Archaea or Bacteria and/or are not novel and not a mobile module giving 7933 domain families. Modifying the query to include only domains with more than 20 members produces 108 novel domain families that occur in human protein sequences, are specific to eukaryotes and have at least 20 members.

Our aim is to push the functional annotation of proteins as far as possible using only automated methods. Defining domain families is the first step. We are currently testing methods to split domain families into groups of orthologous proteins and automated methods to define functionally important residues in a family.