PREDICTION METHOD ABSTRACT The results submitted to the prediction evaluation were obtained by two different methods: (1) Fold recognition and (2) Assembly of backbones from small overlapping fragments.
The fold recognition technique consists of the following parts (1) a knowledge based energy function (mean force potentials), (2) a technique to align sequences with structures (allowing gaps) and (3) quality assessment of the models obtained in the alignment (z-score).
The energy function consists only of C-beta-C-beta interactions and an approximation for solvent interactions. The energy function and z-score calculation is identical to the functions used in the PROtein Structure Analysis program PROSA-II available from gundi.came.sbg.ac.at, which can be used to detect incorrect folds or faulty parts in a structure (Sippl,M.,J., Proteins, 17, 355, 1993).
The alignment technique used is still largely experimental and unpublished (some information on the strategies used has been described by Sippl,M.J., J.Comput.Aided Mol.Design, 7, 473, 1993 and Sippl,M.J. et al. in The Protein Folding Problem and Tertiary Structure Prediction, K.Merz and S.LeGrand (eds.), Birkhauser, Boston, 1994, pp 353-407).
Each target sequence was combined with a subset of structures obtained from the brookhaven data base. The structure yielding the highest score was defined as the best model and the corresponding alignment was submitted as a prediction (in terms of backbone coordinates and the secondary structure assignments, helix or strand, obtained from the respective X-ray structure).
From our previous studies we found that in roughly one out of four cases the current implementation is able to recognise a related fold in the data base in the absence of significant sequence homology. This, of course, is only possible if a related fold exists in the data base. If no related fold exists the method often yields substructures similar to the target structure, but necessarily incorrect overall folds (in fact the extent of similarity is difficult to assess).
The second method employed is based on the combination of small fragments (Sippl et al. Protein Science, 1, 625-640, 1992). The method yields full backbone coordinates, but uses only local energy terms. It can be used to predict the local structures but not the tertiary fold. In this prediction experiment we used the results in the following way: The backbones were assembled and the secondary structure assignments were calculated from the model. These results were then compared to the secondary structure assignments derived from the best model obtained from the sequence structure alignment technique. The consistency between the two results was used as an indication for the quality (in terms of local backbone geometry) of the predictions.
We submitted the best scoring models for 15 target sequences, although
it was clear that the method has a chance to correctly predict a
structure only if a closely related fold is contained in our data base
of known structures.
Structural Biology home page Asilomar Conference home page