REVIEW ARTICLE Ten years of predictions … and counting Domenico Cozzetto 1 , Adele Di Matteo 1 and Anna Tramontano 2 1 Department of Biochemical Sciences, University of Rome ‘La Sapienza’, Italy 2 Istituto Pasteur – Fondazione Cenci Bolognetti, University of Rome ‘La Sapienza’, Italy In 2004, the Critical Assessments of Techniques for Pro- tein Structure Prediction (CASP), celebrates its tenth anniversary. The initiative, notwithstanding its relatively long tradition, remains lively and challenging. It is organized by John Moult (Center for Advanced Research in Biotechnology, Rockville, MD, USA), Krzysztof Fidelis (Lawrence Livermore National Labor- atory, Livermore, CA, USA), Tim Hubbard (Sanger Institute, Hinxton, UK), Burkhard Rost (Columbia Uni- versity, New York, NY, USA) and Anna Tramontano (University of Rome, Italy) with the invaluable help of Andriy Kryshtafovych (Lawrence Livermore National Laboratory) and Volker Eyrich (Columbia University). The goals of the experiments are: to evaluate the accu- racy of current methods for protein structure prediction; to identify bottlenecks and to indicate the directions where efforts can best be focused. The scheme is simple: the organizers collect sequences of ‘targets’ i.e. of pro- teins, the structure of which are likely to be solved within a few weeks. These sequences are made available to the community of computational biologists who attempt to predict their three-dimensional structures as well as other relevant biological properties, e.g. domain boundaries, long range inter-residue contacts, disordered regions and, when not previously known, function. Once the experimental structures of the targets are available, they are compared with the collected predictions using a large variety of numerical measures, and the data generated are stored in a database in the Livermore Laboratory Prediction Center. Experts in the field of protein struc- ture prediction are asked to critically evaluate the results and highlight progress and bottlenecks in the field. In 2004, the community selected Alfonso Valencia (Centro Nacional de Biotecnologia, Madrid, Spain), Roland Dunbrack (Fox Chase Cancer Center, Philadel- phia, PA, USA) and B K Lee (National Institutes of Health, Bethesda, MD, USA). The process, lasting from spring to winter of each even-numbered year, is conclu- ded by a meeting where the community convenes to dis- cuss the results. This year, for the first time, the meeting was held in Europe, in Gaeta on 4–8 December. During its ten-year history, CASP has been instru- mental in convincing both the computational and experi- mental communities that the prediction of the structure of proteins non-evolutionarily related to proteins of known structure is not completely out of reach. Indeed, fold recognition methods (i.e. methods that try to iden- tify which of the known topologies is the most likely for an unknown protein); effective techniques for predicting secondary structure and, more recently, methods able to assemble fragments of proteins of known structure to construct the structure of proteins the architecture of Keywords automatic prediction servers; CASP; model evaluation; protein structure prediction Correspondence A. Tramontano, University of Rome ‘La Sapienza’, Department of Biochemical Sciences, 5 Piazzale Aldo Moro, Rome 00185, Italy E-mail: anna.tramontano@uniroma1.it (Received 15 December 2004, accepted 24 December 2004) doi:10.1111/j.1742-4658.2005.04549.x The CASP experiment has been run every other year since 1994. Its objec- tive is to subject the available structure prediction methods to a blind test. This is a short report of the highlights of its last edition. ‘Men who wish to know about the world must learn about it in its partic- ular details’ (Heraclitus of Ephesus, 535–475 bc). FEBS Journal 272 (2005) 881–882 ª 2005 FEBS 881 which is completely novel, have all been fostered and popularized by CASP – a fairly major contribution to molecular biology in the postgenomic era. Indeed, the contribution of CASP has not only been to evaluate the quality of the approaches and promote cross-fertilization between them, but also to validate which of the tools are sufficiently mature and reliable to become part of the standard suite of methods that experimental biologists can use regularly. This year, among the participants selected by the assessors to des- cribe their strategy and results there were ‘the usual sus- pects’, namely, David Baker (Washington University, USA), Jeff Skolnick (Univeristy of Buffalo, NY, USA), David Jones (UCL, UK), Kevin Karplus (UC Santa Cruz, USA), Krzyzstof Ginalski, Janusz Bujnicki and Andrzej Kolinski (Warsaw University, Poland), but also new participants such as Mayuko Takeda-Shitaka (Kitasato University, Japan), Kentaro Tomii (National Institute of Advanced Industrial Science and Technol- ogy, Tokyo, Japan), Yaoqi Zhou (University of Buf- falo), Ming Li (University of Waterloo, Canada). The results and a description of the methods can be found at the CASP website (http://predictioncenter.llnl.gov/ casp6). It is fair to say here that, thanks to the insights and efforts of these groups, as well as to the hard work of many others, the problem of predicting the overall topology of many proteins is clearly within reach, and this is certainly good news for many experimentalists. Figure 1 shows one example where the experimentally determined structure of the first domain of target 272, a hypothetical protein from Thermus thermophilus that is a protein with no detectable sequence similarity with any known structure, is compared with the model pro- duced by the group of David Baker. There is a clear cor- relation between the quality of a model and its range of application. Even in the most difficult cases, these mod- els are usually sufficient to understand the general prop- erties of the molecule thus identifying solvent-exposed regions, flexible parts and, in some cases, to reveal unex- pected evolutionary relationships useful for function assignment. However, for other applications such as drug design or the prediction of substrate specificity, the level of detail required is much higher. Furthermore, methods usually produce alternative models and, in the most difficult cases, distinguishing which one is closer to the real structure represents a serious bottleneck. There is a consensus in the field that this latter task is easier when a model very close to the native structure is pre- sent in the ensemble of models. It is not surprising, therefore, that discussions at the meeting focused on how to push the field towards devoting more effort to the refinement of the models, to the extent that the com- munity is discussing ways to set a required minimum quality of a model below which it would not be consid- ered at all in the assessment. This can be, for example, the quality of the best model obtained by automatic pre- diction servers, some of which have obtained results comparable to those of the best research groups. The appearance of the above mentioned fragment- based methods for predicting the structure of proteins with a new fold had the undesirable effect of discour- aging ab initio methods for protein structure predic- tions, which could not compete with the quality achieved by the heuristic methods. However, at least one example of successful ab initio prediction by the group of Harold Scheraga (Cornell University, Ithaca, NY, USA) was reported in this meeting for target T0215, a 48 residue-long protein. This, together with the possibility of using ab initio energy-based methods for more accurate refinement of the modelled struc- tures should, we hope, revive the interest of the ‘fold- ers’ to the initiative. We are convinced that models obtained by combining energy-based and knowledge- based methods will finally set the foundations for a solution to the protein folding problem. Acknowledgements The Sixth Edition of the CASP meeting was sponsored by National Institutes of Health, NLM and NIGMS, Istituto Pasteur – Fondazione Cenci Bolognetti, Euro- pean Molecular Biology Organization (EMBO), Bio- Sapiens Network of Excellence funded by the European Commission FP6 Programme, contract number LHSG-CT-203-503265, Lawrence Livermore National Laboratory, Italtech Solutions and IBM. Fig. 1. Comparison between the experimental structure (left) and a model (right) of the CASP target T272 domain 1. The structure was solved by A. Ebihara, M. Yao, S. Yokoyama, and S. Kuramitsu (RIKEN Genomic Sciences Center, Yokohama, Japan) (PDB code: 1WJ9), and the model was submitted by D. Baker (Washington University, USA). Ten years of predictions … and counting D. Cozzetto et al. 882 FEBS Journal 272 (2005) 881–882 ª 2005 FEBS . REVIEW ARTICLE Ten years of predictions … and counting Domenico Cozzetto 1 , Adele Di Matteo 1 and Anna Tramontano 2 1 Department of Biochemical Sciences,. methods able to assemble fragments of proteins of known structure to construct the structure of proteins the architecture of Keywords automatic prediction