PDB COMMUNITY FOCUS

Christine Orengo is a Professor of Bioinformatics in the Structural and Molecular Biology Research Department of University College London (UCL). She studied chemical physics at Bristol University and was awarded a Ph.D. in enzyme kinetics at UCL in 1984. Following a brief spell in industry, she worked as a research fellow at the National Institute for Medical Research (NIMR) in London before moving back to UCL in 1992 to pursue further postdoctoral studies. She was awarded a Senior Research Fellowship by the Medical Research Council in 1995 and was appointed Chair in Bioinformatics in 2002. Together with Janet Thornton, she established the CATH domain structure classification in 1993 which led to the discovery of some highly populated fold groups in nature–the
so-called superfolds.

Her current research interests are in structural, functional, and comparative genomics. Computational analyses exploit the CATH database of structural families and the more recently established sister resource for domain and protein families in completed genomes, Gene3D. She collaborates with a number of experimental groups involved in studying pain, cancer and host-viral interactions. She also participates in several European networks for genome annotation (Biosapiens), grid technologies (EMBRACE), and systems biology (ENFIN) and is a member of the NIH-funded PSI Midwest Center for Structural Genomics (MCSG) headed by Andrzej Joachimiak. She was one of the founding researchers of the bioinformatics based Inpharmatica company. She has authored over 150 papers, book chapters and reviews and is on the editorial board of FEBS, BMC Structural Biology, PEDS, and the Journal of Structural and Functional Genomics. She is on the advisory board of the Swiss Institute of Bioinformatics and the Marie Nostrum Supercomputer Centre in Barcelona.

Dr. Christine Orengo, University College London

Q. In 1997, you and your colleagues established CATH¹–a system that is used to classify protein domain structures. How are researchers using CATH today? What types of research and discoveries does it enable? Has its usage changed in the past ten years?

A: In the early 1990s, there were over three thousand structures deposited in the PDB and Janet Thornton realized that we could get some very useful insights into protein folding and evolution by grouping these into fold groups and evolutionary families. I was fortunate to join her group at that time and we set about doing this classification with the benefit of a very sensitive structure comparison algorithm developed by Willie Taylor and myself, at NIMR. We designed a hierarchical classification which grouped proteins according to their basic secondary structure composition (Class), 3D shape (Architecture), folding arrangement (Topology), and finally evolutionary ancestry (Homology). Although we largely use automated approaches, identifying domain boundaries in multi-domain proteins, and recognizing homologues are difficult and very time consuming, as they need manual validation, which is why we only have ~80% of the PDB classified to date. We have just introduced some sophisticated new protocols that we think will help us to increase this percentage over the next year.

Despite this slight lag with the PDB, CATH is widely used and currently receives about a million web page hits per month from sites all over the world. We have put considerable effort into the design of the resource, trying to present the information in an intuitive and easily accessible form, and I believe this is reflected in its high usage. SCOP², a related resource, is also very widely used but because we exploit slightly different criteria to classify folds and provide additional information on superfamilies (e.g. multiple structure alignments), the two resources are somewhat complimentary. I think CATH is particularly useful for teaching. Perhaps the other distinctive feature of CATH is that we have developed our own structure comparison methods and provide a service (CATHEDRAL web server)³ for scanning new structures against representative domains. This is very popular with structural biologists as it can be used to recognize novel folds or classify new structures into existing superfamilies. The CATH fold library is also exploited by computational biologists developing methods to predict whether a sequence is likely to adopt one of the known structures.

We have now extended CATH to include all sequences in the genomes that can be predicted to belong to a CATH superfamily (CATH-Gene3D)⁴ and this has allowed us to increase the functional annotations associated with each superfamily hugely. Biologists are increasingly using CATH and Gene3D to obtain structural and functional annotations for their proteins and this has been facilitated by further dissemination of the information through the DAS annotation systems set up by the Biosapiens network (www.biosapiens.info).

Perhaps one of the most interesting phenomena revealed by classifying structures is the incredible bias in the populations of the fold groups and evolutionary superfamilies. In 1994, Janet Thornton and I reported the existence of the superfolds, a set of 10 folds which were highly over-represented in CATH⁵. This trend still exists and the integration of sequence data through Gene3D has shown that it is not an artifact of sampling but a genuine reflection of the dominance of certain folds in nature. The bias is also apparent at the evolutionary superfamily level. For instance, the 100 largest superfamilies in CATH account for nearly half the domain sequences of predicted structures in completed genomes.

As CATH has become more highly populated, it has been used to study and characterize the structural mechanisms involved in the evolution of proteins and their functions; in particular, the extent to which structural embellishments to the domain core can modify the geometry of active sites or influence surface features mediating different protein-protein interactions. The integration of genome sequences in CATH-Gene3D has illuminated functional diversity across superfamilies, and recent changes in the usage of CATH reflects biologists’ interests in performing comparative genome analyses with this extensive functional data. For example, a comparison of CATH superfamilies, universal to bacteria, revealed that the expansion of metabolic and regulatory superfamilies with genome size is balanced, allowing maximum enrichment of the metabolic repertoire within the constraints of maintaining a small genome for fast replication.⁶

Q. Do you think that we are close to having representatives of every possible fold? Have the structural genomics projects had an impact?

A. I think this depends on one’s definition of a fold. The huge structural diversity apparent in some of the largest CATH superfamilies has challenged my belief in a rigid hierarchical classification whereby relatives in each evolutionary superfamily adopt the same fold. For example, there is great structural diversity in many of the 100 most highly populated superfamilies, and there are clear examples of relatives with different folds. Whilst these relatives share 40-50% of residues in the cores of their structures, these cores can be embellished so differently that many structural biologists would say that the domains belong to different fold groups. That said, for the remaining ~2000 superfamilies, relatives can be characterized within a single fold group and so I feel that the topology or fold group level in CATH is still valuable.

The structural genomics initiatives, particularly the PSI initiative in the States which has the goal of solving novel folds and aims to determine structures for all large protein families, are helping both to increase the numbers of known folds in the PDB and also to address the question of whether the hundreds of thousands of apparently novel superfamilies in the genomes are truly novel, adopting folds that are distinct from anything seen before. These initiatives have been very successful in increasing the numbers of new folds deposited in the PDB each year. For example, over the last two years a large proportion of the novel folds in the PDB have come from the four major centers associated with this initiative. Interestingly, although PSI deliberately targets superfamilies thought to be unrelated to any known superfamilies in SCOP or CATH, only about 30% turn out to be new superfamilies with distinct folds once their structures are solved. The remainder have been found to be distant relatives of known fold groups and families.

As to whether we have representatives of every possible fold, our analysis of genome data using sensitive threading algorithms like David Jones’s GenThreader⁷ suggests that within each organism about 80% of sequences can now be assigned to one of ~1100 CATH folds. Thus I would say that we do have fold representatives for most of the major superfamilies in nature. However, nearly half of these predicted structures belong to the 100 very structurally-diverse superfamiles and so it is possible their folds may be slightly different to those already characterized.

Sequences which can’t be assigned a fold in CATH tend to belong to very small superfamilies which are species-specific. The number of these superfamilies is growing enormously as the metagenomics initiatives continue. For example, sampling of bacterial proteins from different environments like the Sargasso sea, diverse soils and even the human gut, suggests the existence of hundreds of thousands of very small families and orphan sequences for which we have no structural data at present. Although some of these may be genuinely new superfamilies with folds never seen before, it is more likely that a significant proportion will be found to be distant relatives of structurally characterized families. Some divergence in the structure of these remote relatives would be likely as the different environmental contexts would probably result in the evolution of different functions, and this is frequently mediated by changes in the structure.

The problem in estimating the number of folds that remain to be determined lies in the currently rather subjective approaches used for defining fold similarity. If we assume two domains have similar folds–if they superpose with an RMSD less than 5Å, (normalized for the number of equivalent residues)–we actually find that there are nearly three times as many folds in CATH than represented by the ~1100 fold groups. Practically all of this increase is due to the structural diversity occurring across the 100 largest superfamilies. Since we know that a significant proportion of sequences from each organism are typically assigned to these very large superfamilies, as increasing numbers of structures are solved from different species, the total number of folds will grow, simply from an expansion of these very large superfamilies. In addition, since the thousands of new families arising from the metagenomics will either have novel folds or very likely be distant relatives of these very large superfamilies, and therefore with slightly different folds, over the next decade we could certainly see hundreds more structures which are rather different from any known folds, especially if the structural genomics initiatives continue to be funded.

However, as I mentioned before, whether we view these as completely new folds depends on our definition of fold. We no longer refer to the T level in CATH as the topology or fold group level but rather the ‘topological motif’ or ‘fold motif’ level. In other words, structures grouped at this level share a large central structural motif or core ‘fold motif’ comprising about 40-50% of the domain’s residues. I believe that the majority of ‘fold motifs’ in nature have now been characterized, with the structural genomics contributing significantly to this repertoire of fold motifs over the last decade. Structures remaining to be solved are highly likely to have core motifs similar to one of the ~1100 fold motifs characterized in CATH or SCOP, but these cores may be structurally decorated in ways not seen yet. By improving the characterization of these fold motifs and understanding the manner in which they can be structurally embellished, we hope to improve the structural annotation and modeling of all the sequence relatives in the genomes.

Regardless of the definition of a fold, we are interested in discovering all the different ways in which proteins fold into their 3D dimensional structure and interact with ligands. Gene3D was established to structurally annotate the genomes and integrate functional data from all the sequences. Using these data, we can better understand the structure-function relationship with respect to protein-protein and protein-ligand interactions. Although the ways in which proteins bind ATP could be limitless, they are likely to be very similar in proteins with the same fold. Therefore, by targeting predicted new folds and diverse functional subfamilies the structural genomics initiatives should deepen our understanding of protein folding and protein-ligand interactions and move us further towards a structure-function model for all proteins.

Q. Predicting the function of a given protein is a great challenge. How do Gene3d and the PDB archive play into this type of research?

A. Clearly the value of structures solved by the structural genomics initiatives increases once functions become known for them or as methods for predicting function from structure improve. To facilitate this, we have been increasing the amount of functional information stored in Gene3D. Fortunately, there are now many excellent public resources providing functional information. Those captured in Gene3D include GO⁸, COGs,⁹ FunCat,¹⁰ and EC¹¹ amongst others. We carefully inherit this functional information between relatives using various bioinformatics protocols. Some approaches exploit simple pair-wise sequence identity between relatives whilst others use more sophisticated methods (e.g. HMM-HMM comparisons) to allow safe inheritance between more distant relatives sharing common functions. Knowledge of the pathways or biological processes that a protein participates in is also useful for understanding its functional role, and so we have incorporated information on protein interactions in Gene3D (e.g. from KEGG,¹² Reactome,¹³ and IntAct¹⁴) and developed a suite of bioinformatics tools for predicting interactions between proteins, too. Integrating data in Gene3D in this way allows us to draw together as much collated information on genes as possible both to enhance biomedical research, as well as our model of protein evolution.

The recently created PSI Structural Genomics Knowledgebase (kb.psi-structuralgenomics.org/KB) will help enormously in extending the functional information available for each structure. With the aim of integrating and presenting functional information from a wide range of public resources, this will significantly enhance structural studies on how proteins function. In addition, other initiatives such as the EU-funded Biosapiens network for structural and functional annotation of genomes will also play a part in providing functional annotations for PDB structures. A recent analysis performed for the Midwest Center for Structural Genomics showed that by using Gene3D, some functional information could be gleaned for a large proportion of sequences targeted for structure determination. Some of this is rather general information and may not be that useful at present, except in directing further experiments (e.g. mutation experiments) but a reasonable proportion is detailed enough to allow some mechanistic rationale to be derived from the solved structure.

Furthermore, since recent aims of the PSI structural genomics initiatives include targeting additional relatives from the most highly populated CATH superfamilies, relatives can be targeted which are predicted to be functionally diverse from those with close homologues of known structures. Expanding the repertoire of structures for different functional subfamilies within these superfamilies will increase our understanding of structure-function relationships and ultimately improve function prediction methods. Recent analyses of structures of unknown function solved by the Midwest consortium using the ProFunc resource developed by the Thornton group, showed that some functional information could be predicted for a large proportion of the structures. This success rate is likely to increase as structural genomics initiatives deliberately target sequences with known functions and the resulting increase in coverage of structure-function space improves our function prediction algorithms.

Q. With Richard C. Garratt, you've recently published a great educational tool called The Protein Chart.¹⁵ What was the inspiration for this "periodic table" of proteins? How do you think it will be used?

A. Richard and I really enjoyed developing this chart and we had two excellent CATH researchers in my group, Alison Cuff and Ian Sillitoe, who made the whole project possible. The idea for a protein chart originally arose from the structure modeling kit that Richard had designed for Wiley which is a wonderful teaching tool for explaining how structures are built from their component secondary structures. It’s really a Lego toolkit for proteins! Wiley wanted a protein chart showing examples of representative structures that students could try to build with the kit. We were very excited by the project and inspired to produce a design based on the ideal ‘periodic chart’ of protein structures proposed by Willie Taylor a few years ago.¹⁶ This shows simple representations of all the types of architectures or 3D protein shapes that should be seen in nature given the rules drawn up over the last three decades for protein folding and packing. We thought Willie’s chart was a wonderful way of representing our current knowledge of protein architectures and imagining what shapes and folds remained to be discovered.

A section of the b-proteins shown in The Protein Chart (www.wiley.com)
(Reprinted with permission)

So we designed a protein chart, arranged like a periodic table, but showing representatives of all the domain architectures or shapes currently deposited in the PDB and classified in CATH. There are over 30 different architectures in CATH which are regular enough for the 2D image of the structure to provide meaningful information, and for each of these, the chart shows the ranges of sizes observed. The chart also contains information on the proportion of genome sequences that are predicted to adopt each type of shape, and also the types of functions exhibited in the different fold groups. There are also illustrations of common supersecondary motifs and oligomeric proteins, and so we think it will be a very useful tool for undergraduate teaching and also for structural biology researchers. I would imagine that computational biologists developing structure prediction methods will also find it a useful way of learning about the different shapes and folds they are trying to predict. With the progress of the structural genomics initiatives, especially the PSI, in solving novel structures, we can expect the chart to evolve and expand over the next decade and it will be a useful visual aid for monitoring our knowledge of the structural universe.

1. C.A. Orengo, A.D. Michie, S. Jones, D.T. Jones, M.B. Swindells, and J.M. Thornton (1997) CATH–a hierarchic classification of protein domain structures. Structure. 5: 1093-1108.
2. L. Conte, A. Bart, T. Hubbard, S. Brenner, A. Murzin, and C. Chothia (2000) SCOP: a structural classification of proteins database. Nucleic Acids Res. 28(1): 257-259.
3. O.C. Redfern, A. Harrison, T. Dallman, F.M. Pearl, and C.A. Orengo (2007) CATHEDRAL: a fast and effective algorithm to predict folds and domain boundaries from multidomain protein structures. PLoS Comput Biol. 3(11): e232.
4. C. Yeats, J. Lees, A. Reid, P. Kellam, N. Martin, X. Liu, and C. Orengo (2008) Gene3D: comprehensive structural and functional annotation of genomes. Nucleic Acids Res. 36 (Database issue): D414-8.
5. J. Ranea, D. Buchan, J. Thornton, & C. Orengo (2005) Microeconomic principles explain an optimal genome size in bacteria Genetics. 21: 21-25.
6. C.A. Orengo, D.T. Jones, and J.M. Thornton (1994) Protein superfamilies and domain superfolds. Nature. 372(6507): 631-4.
7. D.T. Jones (1999) GenTHREADER: an efficient and reliable protein fold recognition method for genomic sequences. J Mol Biol. 287(4): 797-815.
8. The Gene Ontology Consortium (2000) Gene Ontology: tool for the unification of biology. Nature Genetics. 25: 25-29.
9. D.L. Wheeler, T. Barrett, D.A. Benson, S.H. Bryant, et al. (2008) Database resources of the National Center for Biotechnology Information. Nucleic Acids Res. 36(Database issue): D13-21.
10. A. Ruepp, A. Zollner, D. Maier, K. Albermann, et al. (2004) The FunCat, a functional annotation scheme for systematic classification of proteins from whole genomes. Nucleic Acids Res. 32(18): 5539-45.
11. Enzyme Nomenclature, Enzyme Classification. www.chem.qmw.ac.uk/iubmb/enzyme.
12. M. Kanehisa and S. Goto (2000) KEGG: Kyoto encyclopedia of genes and genomes. Nucleic Acids Res. 28(1): 27-30.
13. G. Joshi-Tope, M. Gillespie, I. Vastrik, P. D'Eustachio, et al. (2005) Reactome: a knowledgebase of biological pathways. Nucleic Acids Res. 33(Database issue): D428-32.
14. S. Kerrien, Y. Alam-Faruque, B. Aranda, I. Bancarz, et al. (2007) IntAct–open source resource for molecular inter action data. Nucleic Acids Res. 35(Database issue): D561-5.
15. R.C. Garratt and C. Orengo (2008) The Protein Chart Weinheim: Wiley-VCH.
16. W.R. Taylor (2002) A 'periodic table' for protein structures. Nature 416: 657-60.