Dates: May 18 to 22, 1998

² Glaxo Wellcome Medicine Research Centre

³ Université Claude Bernard - Lyon

⁴ Imperial Cancer Research Fund - London

Comparative molecular modelling (for a review see Bajorath, et al., 1993) allows expanding the number of protein sequences for which we have structural information. Such models can be effectively used to design mutagenesis experiments and to support drug design projects (Peitsch & Guex, 1997). Over the last years we have developed an automated comparative protein modelling server (SWISS-MODEL) and a sequence to structure workbench (Swiss-PdbViewer). These tools, taken together, are a very powerful protein modelling environment which features numerous algorithms for both model construction and structure analysis (Guex & Peitsch, 1998).

The number of protein sequence for which structural information can be derived by computational methods has further increased in the last few years. Indeed, many proteins with undetectable sequence similarity adopting the same 3D fold have been documented. Conventional comparative modelling methods cannot be applied to such cases as the sequence alignment necessary for the modelling process cannot be generated by current algorithms. Fold recognition methods, however, attempt to evaluate the compatibility of a new sequence with a library of 3D folds. These methods are very promising and may prove a useful tool in enhanced genome analysis.

More recently, we have applied the same technology to larger sets of protein sequences, all derived from complete microbial genomes. Our most recent large-scale comparative protein modelling experiment - using the proteins identified from several complete microbial genomes - demonstrated that approximately 15% of the protein sequences of these organisms could be modelled successfully (spring 1997). This number increased by 5 to 10 % when we used the SWISS-PROT database. This effect finds its reasons in the bias of the database content towards well studied and over represented protein families (i.e. serine proteases, dehydrogenases etc…). Another interesting observation was made when we compared these numbers between autumn 1996 and spring 1997. Indeed, the number of proteins for which we could derive a model increased by more than 3 %. This is due to the growth of the Protein Data Bank, and clearly demonstrates that a concerted experimental protein structure elucidation effort has a very significant influence on the global knowledge of protein structure.

Since some time, crystallographers, genome scientists and protein modellers are discussing a major "post-genome" project, which should have partners around the world in both academic and industrial settings. This project is aimed at the experimental elucidation of a large number of protein structures. The proteins to study by either NMR or X-ray crystallography ideally comprise at least one member of each protein family. The selection process will be based on advanced sequence analysis methods and a reliable classification of proteins into families and fold classes.

This structural genomics project will obviously have a major impact on the number of sequences for which a protein model can be derived, and thereby gradually allow the gathering of structural information on every protein family. A wild extrapolation of the above experiment would tend to show we could be able to derive molecular models for "every protein" by 2020. This will have a major effect on the efficiency and time necessary to discover new drugs and thereby improve the quality of healthcare.

3DCrunch will be the first such large-scale computational project initiated in structural biology. The project will yield an unprecedented amount of useful structural information, which we will make available to the scientific community.

• From a scientific point of view, we will learn a lot about protein modelling, as the very large number of structures computed will allow the evaluation of the methods involved using a statistical approach. Thereby, new methods fixing these problems can be developed. Indeed, protein modelling methods are often evaluated using a small number of discrete cases. Large numbers of available models will be very useful to test hypothesis prior to implementation. It will also be a reservoir of models to compare with upcoming experimental structures. Not only can the modelling methods be evaluated through bona fide predictions, but the models can be used for molecular replacement as it has often been the case in the past. Several models made by SWISS-MODEL were used to solve experimental structures, avoiding painstaking isomorphous replacement experiments.
• From a technical point of view, it will allow us to predict more accurately what computational power is required to compute the structure of all proteins of the human genome, provided the structure of one protein of each family has been elucidated experimentally. It will thus help identify the computational challenges, that lay ahead in structural biology.

The very large number of protein models generated by the project will be stored in a database: the Model Database (MODB). This new database will replace today's SWISS-MODEL Repository which presently holds close to 4000 protein models, mainly of sequences of bacterial origin (Peitsch et al., 1997; Peitsch and Guex, 1997).

Each completed model will be fully annotated and stored as a SwissPdbViewer project file. This enables the user to get a complete modelling environment with all the template structures superposed onto the final model. Users can thus further refine the model from within the Swiss-PdbViewer and use the direct SWISS-MODEL access to submit Optimise mode requests.

All results will be accessible to the scientific community via the SWISS-MODEL section of the ExPASy Web Server.