The ProFAT server
The input of a ProFAT analysis is a protein sequence and a keyword list that describes the cellular process or putative function relevant for the protein under analysis.
ProFAT workflow
The workflow of ProFAT can be divided into 3 parts (Figure 1): a domain search or domain prediction, whereby identified conserved domains are used to split the input query for further processing with 2) the Annotation Engine and 3) the Threading Engine.
Domain search and prediction
Initial domain searches (using RPS-BLAST against the CD-database (NCBI)) are carried out using a restrictive E-value cutoff (E <= 1E-04) by default. Identified conserved domains can be selected in the results page for further processing. Figure 2A shows the results of a domain search for the protein Dip13α/APPL1 [GenBank:NP_036228]. In this case, ProFAT identified a central PH-domain and a C-terminal PTB domain in the input query. If the domain search fails to identify conserved domains, the user can perform a domain prediction. In this case, RPS-BLAST is run using relaxed settings (E <= 100). Identified weak domain hits are subsequently submitted to a BLAST-search, whereby the resulting hits are mined for their biological relevance using the user-provided keyword list. Figure 2B shows the results of the domain prediction for the N-terminal 280 amino acids of Dip13α/APPL1. If both approaches fail to detect conserved domains, the input sequence can be split into fragments of sizes between 150 and 300 amino acids.
At this stage, selected domains can be in parallel submitted to the Annotation Engine, the Threading Engine, as well as to an HMMerThread run for a keyword-independent domain prediction.
Annotation engine
The Annotation Engine sends the selected conserved domains and/or regions of the input query to a PSI-BLAST search against the non-redundant database. Associated information including GO-annotation, GenBank features, as well as associated publication abstracts of identified hits are subsequently text-mined for the occurrence of user-provided keywords. Figure 3 shows a typical output from the Annotation Engine, where full-length Dip13α/APPL1 was used as the query sequence with a keyword list tailored for 'Endocytosis' [see Additional file 1]. Results from the Annotation Engine for each domain are represented by red bars – when keywords of the user-provided list are found in information associated with identified database entries – and by blue bars – when none of the keywords have been detected (top bars in Figure 3A). Results from the Threading Engine are represented in a similar fashion (bottom bars in Figure 3A). The output of both core modules is highly interactive. Mouse-over of the BAR-domain region of Dip13α/APPL1 (N-terminal red bar or region 1–280 in the associated table of the Annotation Engine) results in a graphical representation of the individual alignments (Figure 3B). By clicking on the region bar or the number in the column 'Keyword Hits' in the associated table the user gets access to the post-processed alignments from the PSI-BLAST search (Figure 3C). The user can individually access the information associated with a database entry (GenBank Features, PubMed Abstracts, Gene Ontology annotation, as well as sequence). Identified keywords are highlighted in bold. The linked number in the column 'Total Hits' links to the complete PSI-BLAST results, where each sequence is annotated with its associated information. PSI-BLAST leads to the raw PSI-BLAST results and GO leads to a tabular listing of the frequency of GO keywords associated with all identified hits.
Using Dip13α/APPL1 as a query ProFAT identified sequence similarity between the first 280 amino acids and BAR-domain containing proteins. The presence of a BAR domain in the N-terminus of Dip13α/APPL1 has been previously reported [3, 18].
Threading engine
The Threading Engine performs a Threader run with the selected regions/conserved domains of the input query and subsequently mines information associated with identified structures for keywords from the user-provided keyword list. Information used for text-mining of the Threading Engine includes protein databank (PDB-) features, PDB keywords, compound information, as well as literature from PubMed abstracts. The core output of the Threading Engine is similar to the Annotation Engine described above. By clicking on the number in the column 'Keyword Hits', the user retrieves the alignments provided by Threader 3.5, with individually accessible information concerning the database entry found (Figure 4A). As an example, we show the Threading Engine results from the PH-domain of Dip13α/APPL1, which for instance identified the PH-domain of the Rac-GEF Tiam1 ([PDB:1FOE]). The link 'Threader' in the table links to tabular output of the top hits from the Threader search (Figure 4B).
HMMerThread
One of the limitations of Threader 3.5 is that its sensitivity drops if the protein region submitted does not correspond in length to the sequence of the crystal structure in the threading database. Threader was for instance not able to detect the BAR-domain in the N-terminal region (amino acids 1 – 280) of Dip13α/APPL1 with a significant score, since the region encompassing the BAR-domain is smaller than the un-annotated N-terminus of the protein. To circumvent this problem we combine an HMMer-based domain search on the input query with a subsequent fold recognition run. Only domains with an associated 3D-structure are considered for further structure prediction (see Figure 1B). When an HMMer search was combined with a threading run for Dip13α/APPL1, two structures of BAR domains were detected as top hits in the threading run (Figure 5). HMMer detected a BAR-domain with an E-value of 0.88 in the N-terminus of Dip13α/APPL1 (Figure 5A, amino acids 4 – 224). When this region was selected for further processing using Threader, it identified the BAR-domains of Amphiphysin and Arfaptin2 with nearly 90% certainty (Figure 5B and 5C). HMMerThread could therefore confirm the findings of the Annotation Engine.
Gene Ontology tree mapping
One limitation of ProFAT is that if the keyword list does not correspond to the actual biological background of the protein input query, results may be misleading. To avoid this, ProFAT maps the GO-annotation of identified hits onto the GO-tree, whereby the number of hits in a certain branch are shown next to the biological processes, molecular functions and cellular compartments. When the user does not find any significant hit with the keyword list used, the ProFAT search can be repeated with a selection of keywords based on the biological function most relevant to the input query, as defined by the associated GO-terms.
Design of keyword lists
Text-mining for the selection of biologically relevant hits in ProFAT is performed using keywords from a user-provided list. The results from a ProFAT search are therefore directly influenced by the keywords a user provides for the ProFAT search. While the stemming algorithm [14] used here takes care of differential suffixes of words, users should still follow a few rules in order to obtain optimal results: 1) the user should try to fully describe the process of interest in the keyword list. A CH domain, for example, has been annotated for actin-binding proteins, but is also found in microtubule-associated or cytoskeleton interacting proteins. Assuming that a protein query has been implicated in actin binding, interesting results could therefore be missed, in the case where only the keywords 'actin binding' were present in the keyword list. This is mainly due to firstly, that the actin-binding domain could show remote similarity to a domain which was initially annotated as a microtubule-interacting domain and secondly, because annotations, whether they are manual or automatic, can be inaccurate; 2) the user should try to avoid common words that are found in any GenBank record, like 'RNA' or 'protein' or also names of organisms. Other common words found in protein names are for instance 'alpha', 'beta' or 'delta', which should also be avoided; 3) in case the user is uncertain about the exact wording of keywords that describe a certain process, we would recommend to use commonly used wordings as are for instance found in functional annotation databases such as Gene Ontology or the Panther database [19]; 4) if the user already has an idea concerning the identity of a weakly conserved domain found in the protein query, it is recommended to include the name of the domain in the keyword list, as the Annotation and Threading Engines will then also specifically show those hits that contain similarity to this conserved domain.
Validation of the ProFAT server
In order to evaluate the performance of the ProFAT server, we chose three domain families from the Superfamily database [20], namely the PABP, PLAT and HNF-1α families. All members of these three domain families, including predicted hits that show only weak sequence conservation, were submitted to automated ProFAT searches using the Annotation Engine, as well as HMMerThread. The Annotation Engine was executed using domain-specific, as well as unspecific keyword lists [see Additional file 1]. Proteins that were correctly predicted using HMMerThread searches or the Annotation Engine, respectively, were scored (Figure 6). The correct prediction of superfamily association with HMMerThread was strongly dependent on the domain (Figure 6A and [see Additional file 2]). While HMMerThread correctly identified 92% of all PLAT family members, it detected only 38% of PABP domains and did not find significant scoring for any HNF-1α domain. The failure of HMMerThread to detect any HNF-1α also did not change upon increase of HMMerThread extensions and increasing the depth of HMMerThread hits to 75. A correct domain prediction using fold recognition techniques therefore seems to rely heavily on the domain under analysis, which has been reported before [21]. However, HMMer itself already identified nearly all domains correctly (93% of all 238 PABP family members, 94% of 438 PLAT domains and 92% of 48 HNF-1α domains). RPS-BLAST on the other hand performed worse in terms of domain prediction, with only 60% of correctly predicted PABP domains and 74% correctly predicted PLAT domains. An HNF-1α domain was however detected for all 48 superfamily members by RPS-BLAST, even for more divergent members. The Annotation Engine showed an overall good performance on the accurate assignment of superfamilies [see also Additional file 2]. It correctly scored for around 90% of all superfamily members for all three domains in their respective keyword lists (96% of HNF-1α members, 89% of PABP members and 90% of PLAT family members, respectively). Interestingly, the Annotation Engine detected almost all weakly conserved members of the PLAT and HNF-1α superfamilies, while it only detected a single predicted protein from the PABP superfamily [see Additional file 2], which suggests that the sequence conservation of predicted PABP-members seems to be too low for detection by PSI-BLAST. A detailed analysis of keyword hits of ProFAT's Annotation Engine in all keyword lists is shown in Figure 6B. We observed only a minor false-positive assignment of superfamily members in unrelated keyword lists. Members of the transcription factor family HNF-1α naturally scored with a similar rate in the keyword list 'Transcription' compared to the keyword list which was designed for 'HNF-1α ' specifically. 17% of the members of the PLAT superfamily, which is a domain characteristic for membrane- and lipid-associated proteins, showed also significant scoring in the keyword list for 'Endocytosis', which can be explained by the fact that association with membrane or lipids plays an essential role in intracellular transport. The term 'membrane binding' was furthermore also present in the keyword list designed for 'Endocytosis' [see Additional file 1]. Finally, the PABP domain is found in the C-terminal region of poly(A)-binding protein, 9% of which seem to score significantly in the keyword list 'Cell Cycle'. No co-occurring keywords could in this case account for cross-scoring of PABP-family members in 'Cell Cycle'. However, translational control via poly(A)-binding proteins has also been implicated in the regulation of cell cycle, especially in oocyte maturation (see for instance [22–25]), which could explain the observed hit frequency of PABP superfamily members in 'Cell Cycle'. These data suggest that ProFAT is indeed able to mine functional annotation of proteins in a highly specific manner. When combining the results of the Annotation Engine and HMMerThread on the three domain families, ProFAT could detect 96% of all HNF-1α-, 90% of PABP- and 98% of all PLAT family members (Figure 6A, 'ProFAT combined'). Combination of the data from the Annotation Engine and HMMerThread therefore provides overall better identification than any of the methods alone.
Identification of novel and weak domain hits using ProFAT
Identification of a CH domain in Hook proteins and the microtubule-associated protein KPL2
Hook3 is a member of the Hook family of proteins involved in intracellular trafficking and associates with its N-terminus to the cytoskeleton [26]. We were interested whether ProFAT would detect sequence similarity with any other protein known to bind to microtubules in the N-terminal part of Hook3 (residues 1 to 153). The Annotation Engine identified two Fimbrin-like proteins from A. thaliana as potential weak homologues with E-values of 0.96 and 1.1, respectively ([see Additional file 3B]; for figure legends for Additional files 3 to 8 [see Additional file 9]); for accession numbers of proteins used for ProFAT searches and for construction of multiple sequence alignments [see Additional file 10]. The Threading Engine did not pick up significant or biologically relevant hits. We therefore submitted Hook3 (1–153) to the HMMerThread module, which identified the Calponin Homology (CH) domains from T-Fimbrin ([PDB:1AOA]) and from the APC-binding protein EB1 ([PDB:1PA7], [see Additional file 3C]). We then aligned the three human members of the Hook family to representatives of the CH domain family. As is shown in Figure 7A, all except for one of the essential residues conserved in CH domains are also present in the three human Hook-proteins. The structure of the N-terminus of mouse Hook1 ([PDB:1WIX], Ohashi, et al., unpublished) was recently added to the PDB-database. We compared the structure of the Hook1 CH domain to the CH domain of EB1 using the DALI-server, which gave a Z-score of 8. When 1WIX was used to search the structure database for structural neighbors, it identified the CH domain of Calponin alpha as the first hit with a Z-score of 9.1. The N-terminal domain of Hook1 can therefore be considered to be significantly similar in fold to CH domains. ProFAT was therefore able to correctly identify the N-terminal domain of Hook3 as a CH domain.
KPL2 is an essential component of the central pair complex in ciliated cells. The orthologue from rat was characterized as a gene that is specifically expressed in ciliated cells [27]. The orthologue in Sus scorfa was recently linked to an autosomal recessive disease in pigs that leads to immotile short-tail sperm [28]. The orthologue of KPL2 in Chlamydomonas reinhardtii, Cpc1, was identified as a component of the central pair complex, which is a large protein complex that regulates the activity of axonemal dynein [29]. The central pair complex consists of 2 central microtubules that associate with a large number of additional factors [30], some of which link the two central microtubules. Central pair complex (CPC-) associated proteins also extrude from this structure and thus help in the assembly of a cylindrical cage of filaments surrounding the microtubules. At open positions in this cage, some CPC-associated proteins interact with external radial spokes and thereby transmit signals that regulate dynein activity for coordinated movement of flagella. Mutations in Cpc1 disrupt the assembly of the central pair complex and alter flagellar beat frequency in Chlamydomonas [29]. Biochemical analysis showed that when Cpc1 is deleted, a large portion of the central pair complex is missing.
Rat KPL2 was predicted to have a N-terminal CH domain, with which it could interact with the cytoskeleton or the central microtubule pair [27]. This domain however is undetectable by RPS-BLAST and comes with an insignificant E-value in SMART analysis. We were interested in whether ProFAT would detect a CH domain in human KPL2. The domain search of ProFAT detected a domain of unknown function DUF1042 in the N-terminal part of the protein, which was selected for further processing [see Additional file 4A]. HMMer, on the other hand, detected the presence of a CH domain between amino acids 1 – 105 in the sequence, which was sent to Threader [see Additional file 4A]. The Annotation Engine of ProFAT identified among other CH domain – containing sequences, the proteins Mal3 from S. pombe and the microtubule-associated protein EB1 from Arabidopsis [see Additional file 4B]. Along the same lines, HMMerThread detected the presence of a CH domain with 83% confidence [see Additional file 4C]. The alignment of 3 KPL2 orthologues with representatives from the CH domain family reveals good conservation of KPL2 to CH domain family members (Figure 7A). These results suggest that the domain DUF1042 is essentially a member of the CH domain family.
Identification of a SAM domain in the C-terminus of EPS8 family members
Eps8 proteins are downstream targets of the Epidermal Growth Factor (EGF) pathway. Members of this protein family are implicated in EGF-mediated signal transduction, though their exact role is so far unknown. It has been shown that Eps8 coordinates EGF-receptor signaling via regulation of small GTPases. A C-terminal effector region in Eps8, for instance regulates activation of Rac, which leads to actin cytoskeleton remodeling [31]. Eps8 family proteins are predicted to have a SAM domain in the C-terminus of the protein [31]. Domain searches using RPS-BLAST and/or SMART fail to identify this domain, even at permissive E-values. We were interested in whether ProFAT could detect the SAM domain in those proteins. The domain search of ProFAT identified an EPS8/PTB domain in the N-terminus of EPS8L3, as well as a SH3 domain in the C-terminal part, but failed to recognize the SAM domain. HMMer on the other hand detected a SAM_1 domain with an E-value of 2 in the C-terminus of the protein, which was selected for further processing [see Additional file 5A]. ProFAT's Annotation Engine detected SAM-domain containing proteins, as, for instance, a sequence from chicken and the kinase suppressor of Ras from Drosophila simulans [see Additional file 5B]. The HMMerThread pipeline predicted a SAM_1 domain in the C-terminus of EPS8L3 with a certainty of over 90% [see Additional file 5C]. The multiple sequence alignment of Eps8 and Eps8-like proteins 2 and 3 with representatives of the SAM domain family, as well as the structural representative of the SAM_PNT domain, which is a subfamily of the SAM domain, shows a conserved pattern of hydrophobic, aromatic and charged amino acids (Figure 7B). These results suggest that the C-termini of Eps8 and Eps8 like proteins contain a SAM domain, as was proposed previously [31].
Identification of an RRM domain in PARN proteins and an uncharacterized protein family
The poly(A)-specific ribonuclease PARN is a 3' exonuclease which is involved in the destruction of cellular mRNAs [32]. Members of the PARN family contain a split CAF1 domain, which has ribonuclease catalytic activity. In the center of the CAF1 domain, RPS-BLAST predicts a PARN_R3H domain, which is predicted to bind single- or double-stranded RNAs. RPS-BLAST also predicts a weakly conserved RRM domain C-terminal of the CAF1 domain with an E-value of 1.8. We were interested as to whether ProFAT could detect the weakly conserved RRM domain in human PARN. The domain search of ProFAT correctly predicts the CAF1 and PARN_RH3 domains and the HMMer module of HMMerThread predicts the presence of an RRM_1 domain adjacent to the CAF1 domain [see Additional file 6A]. We selected the RRM_1 module for further processing with HMMerThread, as well as the C-terminal part of PARN for analysis using the Annotation and Threading Engines of ProFAT. The Annotation Engine identified, as an example, the Bruno-like RNA binding protein 5 from chicken [see Additional file 6B]. HMMerThread identified RRM motifs from several crystallized proteins with a confidence of nearly 90% [see Additional file 6C]. The crystal structure of the region containing the RRM domain of a PARN family member has been determined (Nagata T., et al., 2004, unpublished; [PDB:1WHV]). Using the DALI server, we searched for similar structures to 1WHV. The closest hit is the structure of the central RRM of human La protein ([PDB:1S79], [33]), which is detected with a significant score of 7.5. We next performed a multiple sequence alignment of PARN family members to representatives of the RRM domain (Figure 7C) and observe a high level of conservation between these two domains. ProFAT was therefore able to detect the weakly conserved RRM domain in PARN family members.
The uncharacterized human protein LOC84060 has not been associated with any biological function. Domain searches using standard parameters did not reveal any conserved domains for this protein. However, when increasing the E-value in the RPS-BLAST search, an RRM domain is found with an E-value of 4.6. Assuming that this protein would be involved in RNA metabolism or regulation, we submitted the protein sequence of LOC84060 to the ProFAT server. The domain search pipeline of ProFAT did not find any conserved domain, while HMMer identified the presence of an RRM_1 domain in this protein [see Additional file 7A]. We selected the RRM_1 domain for processing with HMMerThread and submitted the protein sequence of LOC84060 to ProFAT's Annotation and Threading Engine. For more accurate results, we invoked the option of splitting the input sequence using 150 amino acids. ProFAT's Annotation Engine identified among others the RRM domain in the poly (A)-binding protein PABPC from human [see Additional file 7B]. HMMerThread found the RRM_1 domain of splicing factor U2AF as significantly similar [see Additional file 7C]. Next we aligned LOC84060 to representatives of the RRM domain (Figure 7C). The multiple sequence alignment reveals that LOC84060 shares all except for two residues that are conserved in this domain family. Based on this data, we suggest that LOC84060 is a RRM domain containing, RNA-binding protein.
Identification of an acetyltransferase domain in the unknown human protein LOC79969
No functional information is so far available for the uncharacterized human protein LOC79969. Domain searches using RPS-BLAST or SMART predict the presence of a domain of unknown function, DUF738. As there was no hint on the biological context this protein could be associated with, we performed only an HMMerThread search with the protein sequence of LOC79969. HMMer detected a weakly conserved acetyltransferase domain within the DUF738 region [see Additional file 8A]. We selected the predicted Acetyltransf_1 domain for further processing using the threading pipeline of HMMerThread, which identified the 3-dimensional structures of several acetyltransferases with a confidence of nearly 90% [see Additional file 8B]. We next aligned members of the LOC79969 family to representatives of the Acetyltransf_1 domain family (Figure 7D). LOC79969 seems to be most closely related to the GNAT subfamily of acetyltransferases. Interestingly, the proposed catalytic Tyrosine residue at the C-terminus of the Acetyltranferase domain (reviewed in [34]) is mutated to a Leucine in human and fly LOC79969 and a Methionine in C. elegans. A conserved Tyrosine is however located 4 residues C-terminal to the proposed catalytic site. As our data suggest that LOC79969 adopts a GNAT-like fold, it will have to be tested experimentally, whether the Acetyltransf_1 domain is catalytically active.
Applications of ProFAT
ProFAT finds its utility in several applications: 1) the ProFAT server should be used when standard similarity search programs fail to predict the function of a so-far uncharacterized protein that can be associated with a certain cellular process/molecular function. In this case, ProFAT would be used as an aid for post-filtering of complex Threading and PSI-BLAST outputs; 2) the user might be interested in whether a conserved domain shows remote sequence similarity or is structurally related to proteins from a specific cellular process/molecular function and can therefore use ProFAT to specifically search for weakly related sequences or structures that are found in the biological context of interest; 3) the domain prediction pipeline is applicable to regions of proteins with no obvious conserved domain. In this case, the combination of RPS-BLAST and a subsequent BLAST-search of weak domain hits with a text-mining step can strengthen evidence from subtle sequence similarity with additional biologically relevant evidence; 4) finally, HMMerThread presents itself as a very powerful pipeline for accurate prediction of weakly conserved domains by looking for remote sequence similarity with conserved domain hits in combination with a subsequent threading step. HMMerThread in addition has the advantage of not relying on the user-provided keyword list and can be applied to proteins, which cannot be associated with any biological function. This module can therefore be used as a means of predicting weakly conserved domains with high accuracy.
