Protein-sequence-based comparative analysis to infer biological function is important and familiar to most biologists. Sequence-profile methods such as PSI-BLAST 1 or HMMER 2 are often used to detect distant homologs, and resources such as Prosite 3, BLOCKS 4 and PFAM 5 are representative resources resulting from protein classification based on sequence patterns. Protein structure also plays a crucial role in a full understanding of protein function as it is more conserved than sequence and hence exposes relationships not possible from sequence alone. Many protein domains have less than 10% sequence identity, and yet possess a similar fold and possibly related function.

One of the early insights gained from comparative genomics was domain accretion 6. From prokaryotes to eukaryotes, the number of domains increases. But in higher eukaryotes, different combinations of domains are often observed in the same and different protein families. From a structural point of view domains are discreet compact folding units. PIR 7 classifies proteins into either a homeomorphic superfamily (proteins containing similar domains in the same order) or a homology domain superfamily (proteins from different homeomorphic superfamilies sharing a common ancestral domain). This modular nature of proteins necessitates a new approach to proteome annotation - a structural-domain-based approach.

There already exist a number of automated or semi-automated complete genome annotation systems. For example, GeneQuiz 8 and PEDANT 9 are two pipelines that are comprehensive and highly automated (Table 1). Similarly, there are several sites that provide protein structure annotations for various genomes. Superfamily 10 uses a set of hidden Markov model (HHM) profiles based on SCOP superfamily members. MatDB, based on PEDANT analysis of Arabidopsis thaliana, provides structural annotations using SCOP domain position specific scoring matrix (PSSM) profiles. The National Center for Biotechnology Information (NCBI) maintains a Conserved Domain Database (CDD) that uses PFAM and SMART 11 domain PSSMs to detect possible structural homologs. The 3D-Genomics database 12 uses SCOP domain PSSMs from 3D-PSSM 13. Gene3D uses the CATH domain classification to annotate genes and genomes 14.

We have developed an automated integrative genome annotation pipeline (iGAP) initially to annotate the proteins of A. thaliana and later all proteomes based on a comprehensive fold library (Figure 1). In addition to the domains from SCOP, we have included domains parsed using the protein domain parser (PDP) 15, full-length Protein Data Bank (PDB) chains and chains not classified by SCOP, but associated with SCOP using combinatorial extension (CE), a structural-similarity search algorithm 16. The result is a comprehensive fold library (FOLDLIB) from which comparative and fold recognition models of three-dimensional structure are derived. As a step beyond PSI-BLAST or PFAM profiles, we have used 123D+ 1718, which not only performs target-template profile-profile alignment, but also uses secondary structure and contact capacity potential information for protein fold recognition. Further, the annotation pipeline provides a graded reliability index of functional prediction reliability ranging from A to E based on extensive benchmarking of selectivity versus sensitivity (N.N.A., I.N.S and P.E.B., unpublished work). Here we describe iGAP and the initial results on the analysis of A. thaliana, the first proteome processed, using a combination of web interface and SQL queries (Figure 2). Comparisons are made to other annotation schemes used to process Arabidopsis and to other proteomes processed with iGAP. The iGAP is systematically being applied to more than 1,000 proteomes, completely or partially sequenced and publicly available at NCBI 19, to develop a comparative proteomic resource.

Automated annotation pipelines are crucial to organize the deluge of genomic information. Table 1 compares features of iGAP with those of GeneQuiz and PEDANT, two established genome annotation methodologies. GeneQuiz focuses on homolog and function assignment through sequence similarity search; PEDANT is a comprehensive analysis pipeline with emphasis on gene prediction, secondary and tertiary structure assignment; iGAP puts much more emphasis on fold recognition, threading and, to be released in the near future, homology modeling. Table 2 compares the proteins of A. thaliana (PAT) database to established databases of protein annotations. They differ in both coverage and focus. Again, each of the resources has clear strengths in a number of areas, but PAT stands out in terms of the amount of structural information it provides. Whereas other resources are limited to what is present in PDB or SCOP, PAT provides additional domains from PDP, and genetic domains from Astral. Moreover, an important feature of iGAP is the benchmarking used to establish the reliability measures. Such quality assurance is critical to the future development of these resources if they are to be used in a meaningful way by experimentalists.

Table 3a indicates the coverage of the Arabidopsis proteome provided by each methodology and associated resource. It is clear that InterPro and iGAP represent two approaches that provide very high coverage of the Arabidopsis proteome, based on sequence and structural information respectively. A combination of InterProScan and iGAP is under active development to integrate sequence- and structure-based annotation. Interestingly, only 14% of the Arabidopsis Information Resource (TAIR) GO annotation is based on nonelectronic annotation. This makes an even stronger argument for the integration of sequence- and structure-based annotation, to reduce the possibility of error propagation in electronic annotation. Table 3b highlights some specific examples of results achieved by PAT over other means. Whether these results are meaningful depends on the user's perspective. For one user, a few additional predictions with 90% certainty could be a distraction. To another, they might, in connection with additional experimental evidence, prove valuable. A future challenge to those of us providing such resources is to minimize the pain and maximize the gain for the different types of user. Again quality assurance and user interface design will prove important. While we have made efforts to classify the reliability of our predictions, they are still predictions and should be used, where possible, with associated experimental proof.

With regard to iGAP specifically, we first looked at the overall coverage of the Arabidopsis proteome using iGAP (Figure 3). We were able to assign nearly 70% of the Arabidopsis proteome to folds which had a reliability index C (90% confidence) or better. This compares to 56% of Arabidopsis proteins in the NCBI nonredundant (NR) protein database having an assigned function. While fold assignment does not necessarily translate into functional assignment, it provides a useful indicator.

Second, PAT provides annotations not reported by other databases. Some examples are listed in Table 4. For example, the AP2-domain is a DNA-binding transcription factor that controls flower and seed development 20 in Arabidopsis. The structure of the AP2 domain is found in the PDB (1gcc) 21. Standard BLAST using the 1gcc sequence provides 140 hits at p < 0.1 (a very weak threshold). In PAT, there are 143 hits of A or B reliability (> 99% confidence) plus 12 of reliability C (> 90% < 99% confidence). Another putative protein (GI number 15228210, locus id At3g47660) has a previously undetected domain at the amino terminus which resembles the structure of the pleckstrin homology (PH) domain from phospholipase C delta (PDB 1mai) (C prediction). PH domains are commonly found in signaling proteins 22. Additional domains found in this protein (also documented by TAIR as InterPro domains) include FYVE/PHD zinc finger and an RCC1 like domain (a regulator of chromosome condensation), with A and B reliabilities respectively. TAIR also reported a sugar transporter signature for this protein from Prosite. While the exact function of the protein remains to be determined experimentally, the new finding of a putative PH domain could offer clues to its potential mechanism for signaling and intracellular targeting.

Third, we surveyed a set of Arabidopsis proteins that have known protein structures (confidence level A, Table 4a). For most of these structures, PAT identifies a number of additional Arabidopsis proteins predicted to contain the same domain. For example, the ubiquitin-conjugating enzyme, which is important in protein degradation, identifies 6 unknown proteins out of 12, with 'C' or above confidence, which contain similar domains. In contrast, no additional proteins were found to have TBP-like (TATA binding protein-like) domains.

Recent structures not found in FOLDLIB or SCOP (release 1.55) were examined to see how well they were predicted by iGAP (Table 4b). For PDB structures 1gp4 and 1gp6 (putative leucoanthocyanidin dioxygenase, NCBI NR database 17 October 2001 release), 123D was able to correctly predict the fold to be similar to 1hig (clavaminate synthase-like SCOP superfamily). WU-BLAST only gave a number of low-probability (E reliability) predictions.

Similarly, PDB entry 1e6b (putative glutathione-S-transferase, NCBI NR database 17 October 2001) is a protein with an amino-terminal thioredoxin-like domain and a contiguous glutathione-S-transferase carboxy-terminal domain. Both WU-BLAST and 123D correctly recognized the template structure 1fw1 (glutathione transferase z/maleylacetoacetate isomerase). Both WU-BLAST and 123D predicted the whole protein to be thioredoxin-like with a reliability index of A. However, WU-BLAST made two additional predictions, both correct. The 'pseudo SCOP entry by PAT' is a novel domain parsed by PDP, which at the time was not in SCOP release 1.55. (It is classified as a separate domain in SCOP 1.59.) This was recognized by WU-BLAST. Additionally, WU-BLAST also recognized the amino-terminal thioredoxin-like domain with correct boundaries.

Finally, the SCOP classification of protein structures by fold (Figure 4a) and by family (Figure 4b) provides a convenient way to catalog the relative occurrences of structures in A. thaliana. With respect to folds, the membrane all-alpha fold, alpha-alpha superhelix and protein kinase-like (PK-like) fold ranked highest. The TIM barrel and Rossman folds, and seven-bladed beta-propeller folds are also among the top folds. PK-like proteins have the second highest occurrence at the superfamily level (data not shown). Not surprisingly, serine/threonine kinases and tyrosine kinases are among the most abundant families.

The PAT database was initially developed as a joint development of academia and industry to serve the Arabidopsis and plant proteomics community through the provision of structure and functional assignment to all identified proteins in the Arabidopsis genome. The underlying technology, specifically iGAP and the associated reliability criteria, is well suited for application to other proteomes and this processing is ongoing to provide a comparative proteomics resource. With more of a focus on comparative proteomics, the resource is being expanded in an effort we refer to as the Encyclopedia of Life (EOL). Details on EOL can be found at 23.

The iGAP components are shown in Figure 1, which illustrates how primary protein sequence and structure data are processed by the system. Details are given below.