The complex metabolic and signaling pathways within the cell are controlled by highly coordinated and intricate protein-protein interactions. Information regarding such protein-protein interactions can be obtained from biochemical and biophysical methods like co-immunoprecipitation [1], yeast two-hybrid [2] and mass spectrometry [3, 4]. To complement these often time-consuming experimental methods, computational methods for predicting functional linkages have been developed. Some methods use protein surface interfaces [5, 6]; some use the ordering and/or proximity of genes in genomes [79]; while others use the co-occurrences of genes in genomes [10, 11].

Recently, computational methods that exploit domain-domain relationships have been introduced and proven to be useful for the prediction of functional linkages in genomic research [1215]. In particular, domain fusion analysis exploits the fact that certain proteins in a given genome consist of fused domains that correspond to single, full-length proteins in other genomes [1315]. The proteins with fused domains in a given genome are likely to directly interact or be involved in the same metabolic and signaling pathways. In their analysis of the M. genitalium genome, Huynen et al. showed that the occurrence of a domain fusion event was highly correlated with function [16].

The query genome is defined as the genome where functional linkages are predicted, while the reference genome is the amalgamation of all other genomes excluding the query. Domain fusion events found in the reference genome predict functional linkages between proteins in the query. To date, most domain fusion analysis have compared complete genomes of relatively small sizes and rely on a BLAST comparison [17] between every protein of the query genome to every protein of the reference [13, 14, 16, 1820]. The analysis has not been applied to larger non-redundant sequence databases such as SWISS-PROT, although the analysis becomes a more powerful prediction tool when more reference genomes are included. One reason for this limitation is that the computation time becomes "prohibitively expensive" [18].

Other groups have already appreciated the use of relational databases for domain fusion analysis [21, 22]. To complement their work, we present a fast computational method that enables domain fusion analysis on partial or complete genomes in a non-redundant sequence database using simple relational algebra operations. Instead of using BLAST comparisons, we leveraged on existing efforts to predict protein domains by Pfam's HMM domain database [23]. Beginning with Pfam's domain layout prediction of each protein in the SWISS-PROT+TrEMBL protein sequence database, we applied successive relational algebra operations using SQL to identify putative functional linkages, especially in H. sapiens and S. cerevisiae. These results are compared with experimentally demonstrated cases and published protein interaction databases. Finally, we discuss various factors that can generate false positives.


The majority of protein sequence and domain databases are built on the relational database architecture. Typically, data is acquired from a database of this type by relational algebra operations in the form of SQL queries. Therefore, a method that can be performed directly using these operations will save unnecessary conversion of data and leverage on the scalability and efficiency of commercial RDBMS software (relational database management systems). Our method for finding domain fusions can be performed entirely using relational algebra operations.

The method is described using relational algebra notation with the following conventions: bold text refers to a table; A.attribute refers to an attribute or column of A; σ(predicate)(A) is the selection operation with the predicate in parenthesis, which selects rows in A that satisfy the predicate; A × B is the cartesian product operation, which creates a permutation of information between A and B; π(A.attribute1, A.attribute2,...)(A) is the projection operation, which extracts specified attributes from A.

As a minimum, the database must have a sequence table (denoted by S) and a domain layout table (denoted by D) with some key attributes (Figure 1a). D stores the domain layout of each of the sequences in S and is linked by the seq_id attribute. Let S query be the table of all protein sequences in the sequence database of the query genome and let S ref be the table of all other sequences in the database comprising the reference genomes. Let D query and D ref be the tables of all domain layouts belonging to the query and reference genomes, respectively. Therefore, S query , S ref , D query , and D ref are defined by the following relations:

Figure 1
figure 1

(a) The Crow's Foot entity relationship diagram ofthe database architecture. (b) The flowchart of the method Protein sequences from SWISS-PROT+TrEMBL were divided into query sequences (belonging to the genome of interest) or reference sequences (everything else). All possible DFTs of the query or reference sequences were then permutated using a single SQL command. The valid DFTs were found by subtracting the query set from the reference set. Finally, using the valid DFTs, all functional linkages were permutated. The number of putative functional linkages is generally large, so it is necessary to filter out false positives.

Let F query and F ref be the table of all possible domain fusion templates (DFTs) in the query and reference genomes, respectively. The idea of DFTs is conceptually similar to Rosetta stone [14] and composite proteins [13]. For example, if a gene has four different domains ABCD, there are six different DFTs: AB, AC, AD, BC, BD, and CD.

Let D q1 and D q2 be copies of D query and let D r1 and D r2 be copies of D ref . F query and F ref can be found by performing a projection and selection operation following a cartesian product between the corresponding domain tables. This operation will enumerate all permutations of DFTs. For example, if gene has three different domains ABC, then there are nine possible permutations of DFTs: AA, AB, AC, BA, BB, BC, CA, CB, and CC. The desired DFTs do not have the same domains (i.e., AA, BB, and CC) and order does not matter (i.e., AB is the same as BA). To remove same domain DFTs, the following clauses are added to the selection predicates: (D q1 .dom≠D q2 .dom) for F query and (D r1 .dom≠D r2 .dom) for F ref . At this stage, it is not necessary to consider the removal of one of the two alternatively ordered DFTs.

Let F put be the table of valid DFTs that can be used in the prediction of functionally linked proteins in the query genome. Therefore, F put can be found by the difference between F ref and F query .

F put = F ref - F query     (7)

Finally, let P put be the table of putative functional linkages in the query genome. P put can be obtained by performing a projection and selection operation following a cartesian product between D q1 , D q2 and F put . This operation will, for each DFT in F put , enumerate all permutations of proteins that contain the first domain in the DFT to proteins that contain the second domain in the DFT. Note that this operation can be more efficiently performed if F put includes only domains found in the query genome.

Remember the alternatively ordered DFTs have not been removed. Therefore, if there is a putative functional linkage between protein A and protein B, there will also be a functional linkage between protein B and protein A in P put . To remove these redundant putative functional linkages, it is easiest to re-insert the all rows in P put into a new table with a database trigger enabled that restricts the row insertion of protein A and protein B, if the row of protein B and protein A exists.

Results and Discussion

Domain fusion analysis of H. sapiens and S. cerevisiae

From our domain fusion analysis on the SWISS-PROT+TrEMBL database, we identified 235 and 189 putative functionally linked protein partners in H. sapiens and S. cerevisiae, respectively (Table 1). While it is difficult to rigorously determine the accuracy of the predictions as not all protein-protein interactions have been mapped in these genomes, we searched protein interaction databases [2426] and scientific abstracts on PubMed for those protein partners that have both gene names mentioned in the same article. The databases and scientific literature clearly indicated a functional linkage (such as gene proximity, a complex formation or common pathway) in 33 and 35 protein partners in H. sapiens and S. cerevisiae, respectively (Table 2). For example, we identified three known pairs of functionally linked proteins and one hypothetical in the H. sapiens and S. cerevisiae (Figure 2). First, a glyoxylate cycle protein [27] in C. elegans and D. melanogaster corresponds to two proteins (MASY_YEAST and ACEA_YEAST) that are known to be involved in glyoxylate cycle of C. albicans [28] (Figure 2a). Second, COXW_YEAST and Q12184 are involved in heme A synthesis in S. cerevisiae [29] (Figure 2b). Third, the levels of O76091 and FHIT_HUMAN mRNA are highly correlated in mouse homologs [30] (Figure 2c). Lastly, the functional linkage of TYSY_HUMAN and DYR_HUMAN is predicted by the domain fusions in many grain genomes including Z. mays, G. max and A. thaliana (Figure 2d). There was a higher percentage of positives in S. cerevisiae largely due to the extensive work in mapping protein-protein interactions in yeast using two-hybrid screens [31, 32], microarrays [33, 34] and mass spectrometry [3, 4].

Figure 2
figure 2

Examples of predicted functional linkages The sequences and domains are identified by their SWISS-PROT+TrEMBL and Pfam id, respectively, and the gene name is enclosed in brackets if applicable. The first three examples are known functional linkages in S. cerevisiae and H. sapiens, while the last one is unknown.

Table 1 Analysis of the S. cerevisiae and H. sapiens sequences in SWISS-PROT+TrEMBL
Table 2 Types of functional linkages from the scientific literature

In H. sapiens sequences, there were 771 DFTs arising from 208 organisms, while in S. cerevisiae there were 1,491 DFTs from 328 organisms. The uneven sequence sampling in the SWISS-PROT+TrEMBL database skews the absolute distribution of organisms, however the distributions relative to each other are interesting. When comparing the relative changes in the distributions, the effect of uneven sampling of organisms in our database is normalized. Specifically, if the probability of finding a DFT is equal in all sequences, then the genomic distribution of the source of the DFTs would be the same as the source of the reference sequences (Figure 3a). The genomic distribution of DFTs for S. cerevisiae and H. sapiens are different to the reference sequences and to each other (Figure 3b,3c). In H. sapiens, the multicellular eukaryotic organisms (such as M. musculus) have advanced in the top ten sources, whereas the single cellular organisms (such as E. coli) have declined. Additionally, while C. elegans is still in the top ten sources of DFTs, it requires 476 sequences to find one DFT compared to 45 for X. laevis (Table 3). Conversely, in S. cerevisiae, single cellular organisms (such as M. tuberculosis) have advanced in the top ten sources and multicellular eukaryotic organisms have declined. Furthermore, distantly related organisms appear to require more sequences per DFT, yet closely related organisms do not require the fewest sequences per DFT (Table 4). One possible explanation is that domain fusions require evolutionary time to accumulate, however, if too much time passes, it may be lost. This suggests that domain fusion events cannot be used to accurately predict phylogeny, since their absence could be the result of too short or too long evolutionary distance.

Figure 3
figure 3

The genomic distribution of sequences in the (a) SWISS-PROT+TrEMBL database compared to DFTs in (b) S. cerevisiae and (c) H. sapiens The genomic distribution of the sequences in the SWISS-PROT+TrEMBL should be similar to the genomic distribution of the sources of DFTs, if the assumption is true that DFTs are equally likely to occur in all species. The figure shows clear differences in the genomic distribution, indicating that the assumption is false. Instead, the source of DFTs of a particular query genome is preferentially found in certain genomes.

Table 3 Top ten sources of DFTs in H. sapiens
Table 4 Top ten sources of DFTs in S. cerevisiae

Effect of our distinct definition of a fusion event

Previous methods for domain fusion analysis [13, 14, 20] are essentially identical to our method, except that our method specifically finds individual "domain" fusions, whereas the previous methods used full-length proteins from one organism, which correspond to a fused full-length protein in another organism. We chose our approach as many proteins consist of multiple domains. For example, consider a fusion protein in the reference organism consisting of domains ABCD, which corresponds to two separate proteins in the query organism, consisting of domains AB and CD. Using our method, the list of reference DFTs would be AB, AC, AD, BC, BD and CD; the list of query DFTs would be AB and CD. Therefore, the valid DFTs that can be used for predicting functional linkages are AC, AD, BC and BD. All four DFTs would predict the same functional linkage between the two query organism proteins. In contrast, previous methods would have only a single fusion event that predicts this functional linkage. Therefore, an additional advantage of our approach is that the number of different DFTs predicting a functional linkage could be used to rank our prediction confidence.

Consideration of splice variants

Splice variants are treated intermediately as separate genes in our method since each variant may interact with different proteins. For example, consider a query gene with two variants: one variant consisting of domains ABC while another consisting of domains AC. If it is found that BD is a valid DFT for functional linkage prediction, then the first splice variant could be involved in a putative functional linkage that the second is not. Finally, the putative functional linkages of the gene would be the union of functional linkages of the splice variants.

Consideration of false positives

Any prediction method could produce false positive results. Here, we consider several sources of false positives, which may be generated by the present method. A false positive can occur when a functional linkage is predicted between two proteins where none exists. One possible source of false positives in domain fusion analysis is the promiscuity or paralogy in domains (for example, BTB, PDZ, SH2 and SH3 domains), which occur at a high frequency in many different protein sequences that do not share similar functions [13, 14, 20]. The removal of promiscuous domains reduces false positives, but the criterion for classifying them is a difficult problem. One criterion relies on finding domains with a Z-score greater than 10 [13, 20], while another on domains that are involved in domain fusions events with more than 25 other domains [14].

In our analysis, instead of removing promiscuous domains altogether, we removed only promiscuous DFTs. For example, the RasGAP domain is involved in 72 functional linkages with the SH3 domain in the H. sapiens genes in the SWISS-PROT+TrEMBL database, however it is only involved in 2 functional linkages with the BTK domain. The DFT of RasGAP and SH3 is more promiscuous than RasGAP and BTK. Since the verification of a predicted functional linkage was performed manually through a literature search, there was a limitation to the number of linkages that could be screened. Therefore, we stringently defined a promiscuous DFT as one involved in 10 or more functional linkage predictions (Figure 4). For a higher tolerance for false positives, it is possible to use a value greater than 10.

Figure 4
figure 4

The percentage of DFTs to the number of functional linkages in S. cerevisiae (blue) and H. sapiens (red) As expected, the general distribution of the S. cerevisiae and H. sapiens are almost identical. Some DFTs such as ig and SH3 in H. sapiens and rrm and pkinase in S. cerevisiae predict over 5,000 functional linkages. These DFTs are promiscuous and should be removed from the analysis because they predict functional linkages between proteins that are likely unrelated. In the analysis we applied a stringent definition of a promiscuous DFT – those being involved in greater than 10 predicted functional linkages.

Another possible source of false positives is the inability to list all the DFTs in the query genome. For example, consider two query genes: one consisting of a domain A while another consisting of a domain B. If it is found that AB is a valid DFT for functional linkage prediction, then the two query genes are perhaps functionally linked. However, if the query genome's DFT list is incomplete, AB may potentially exist and therefore, the two query genes may be falsely predicted as functionally linked. A number of factors can cause this problem including the use of an incomplete query genome, absent or inaccurate profile HMM domains and the erroneous prediction of intron and exon sites.

The domain fusion analysis using relational algebra presented here relies on the prediction of domains from profile HMMs. In contrast, previous approaches to domain fusion analysis often employed heuristic local pairwise sequence alignment (PSA) algorithms such as BLAST [17]. Such algorithms emphasize finding long high scoring local alignments, however, the most strongly conserved residues are commonly distributed across the domain. Therefore, the key drawback of a heuristic PSA-based approach in domain fusion analysis is its relative insensitivity for finding remote homologs and, consequently, domain fusions. Within the last decade, however, the sensitivity of sequence searching techniques has been improved by profile- or motif-based analysis, like the profile HMM, which uses information derived from multiple sequence alignments to construct and search for sequence domains and patterns [3537]. Unlike the heuristic PSA algorithms, a profile or motif can exploit additional information, such as the position and identity of residues that are conserved throughout the domain, as well as variable insertion and deletion probabilities. Therefore, the advantage of the profile HMM is the sensitivity and accurate delineation of domains, however, the key drawback is its reliance on the accurate construction of a profile HMM for all domains. If the profile HMM of a domain is not constructed or carelessly done, it will not find all putative domains and, consequently, domain fusions. Thus, as the quality and quantity of separate domain databases increases such as BLOCKS [36], PROSITE [35], Pfam [23], SMART [38], PRINTS-S [39], ProDom [40], TIGRFAMs [41] and amalgamated domain databases such as InterPro [22], our approach to domain fusion analysis will also become increasingly powerful.


The relational algebra method presented here offers an alternative approach to performing domain fusion analysis that leverages on existing efforts to improve the size and quality of domain and motif databases. We have illustrated the efficacy of the method by identifying many possible functional linkages in H. sapiens and S. cerevisiae sequences in the SWISS-PROT+TrEMBL database. Interestingly, the genomic distribution of the sources of DFTs suggests that DFTs are not likely found either in closely or remotely related organisms, but rather there is a balance between the two extremes that is tilted toward closely related organisms. Finally, future work could expand the method presented here to other genomes of interest.


The analysis was performed on the Oracle RDBMS (version 8) installed on a computer with a dual 750 MHz UltraSPARC-III processor and 4 G of RAM running SunOS 5.8. Sequence information from SWISS-PROT (Release 39) + TrEMBL (Release 17) and domain architecture information from Pfam was migrated to the sequence table and domain layout table of the database, respectively, by Perl and Oracle SQL*loader scripts. To perform the analysis, relational algebra expressions were converted to SQL statements and executed by an Oracle SQL*Plus client connected to the database server. The total computation time for H. sapiens and S. cerevisiae were approximately 4 and 3 hours, respectively.