Prediction of the secondary structures of proteins by using PREDICT, a nearest neighbor method on pattern space

We introduce a novel method for predicting the secondary structure of proteins, PREDICT (PRofile Enumeration DICTionary), in which the nearest-neighbor method is applied to a pattern space. For a given protein sequence, PSI-BLAST is used to generate a profile that defines patterns for amino acid residues and their local sequence environments. By applying the PSI-BLAST to protein sequences with known secondary structures, we construct pattern databases. The secondary structure of a query residue of a protein with unknown structure can be determined by comparing the query pattern with those in the pattern databases and selecting the patterns close to the query pattern. We have tested the PREDICT on the CB513 set (a set of 513 non-homologous proteins) in three different ways. The first test was based on a pattern database derived from 7777 proteins in the Protein Data Bank (PDB), including those homologous to proteins in the CB513 set and gave an average Q(3) score of 78.8% per chain. In the second test, in order to carry out a more stringent benchmark test on the CB513 set, we removed from the 7777 proteins all proteins homologous to the CB513 set, leaving 4330 proteins. Pattern databases were constructed based on these proteins, and the average Q(3) score was 74.6%. In the third test, we selected one query protein among the CB513 set and built pattern databases by using the remaining 512 proteins. This procedure was repeated for each of the 513 proteins, and the average Q(3) score was 73.1%. Finally, we participated in the CASP5 (group ID: 531) where we employed the first-layer database based on the 7777 proteins and the second-layer database based on the CB513 set. The PREDICT gave quite promising results with an average Q(3) (Sov) score of 78.1 (77.4) % on 55 CASP5 targets.