Creation and characterization of a new, non-redundant fragment data bank

The success achieved for protein structure prediction of loop regions with insertions and deletions by knowledge-based methods depends on the quality of the underlying information, i.e. a fragment data bank as complete as possible is needed. However, the greater the number of proteins contributing to the data base the more redundant information is included, which leads to structurally similar proposals in loop predictions and to longer times for extracting fragments. So it is not only necessary to increase the number of proteins for building the loop data base but also to cluster the resulting fragments according to their structural similarities in order to remove redundancy. Here, a new, non-redundant fragment data bank is described, which is based on all proteins in the Brookhaven Protein Data Bank (release 7/95) with a resolution > or = 2.0 A and which can be updated easily by including new information from structures to be solved in the future. In the clustering process presented, the resulting clusters are optimized in several cycles until self-consistency. In this way all redundant information is removed without loosing any significantly different fragments. Finally the resulting fragment data bank is analysed with respect to its completeness.