Prediction of protein secondary structure using the 3D-1D compatibility algorithm.

A new method for the prediction of protein secondary structure is proposed, which relies totally on the global aspect of a protein. The prediction scheme is as follows. A structural library is first scanned with a query sequence by the 3D-1D compatibility method developed before. All the structures examined are sorted with the compatibility score and the top 50 in the list are picked out. Then, all the known secondary structures of the 50 proteins are globally aligned against the query sequence, according to the 3D-1D alignments. Prediction of either alpha helix, beta strand or coil is made by taking the majority among the observations at each residue site. Besides 325 proteins in the structural library, 77 proteins were selected from the latest release of the Brookhaven Protein Data Bank, and they were divided into three data sets. Data set 1 was used as a training set for which several adjustable parameters in the method were optimized. Then, the final form of the method was applied to a testing set (data set 2) which contained proteins of chain length < or = 400 residues. The average prediction accuracy was as high as 69% in the three-state assessment of alpha, beta and coil. On the other hand, data set 3 contains only those proteins of length > 400 residues, for which the present method would not work properly because of the size effect inherent in the 3D-1D compatibility method. The proteins in data set 3 were, therefore, subdivided into constituent domains (data set 4) before being fed into the prediction program. The prediction accuracy for data set 4 was 66% on average, a few percent lower than that for data set 2. Possible causes for this discrepancy are discussed.