Digital signal processing in predicting secondary structures of proteins

Abstract

Traditionally protein secondary structure prediction methods work with aggregate knowledge gleaned over a training set of proteins, or with some knowledge acquired from the experts about how to assign secondary structural elements to each amino acid. We are proposing here a methodology that is primarily targeted for any given query protein rather being trained over a predetermined training set. For some query proteins our prediction accuracies are predictably higher than most other methods, while for other proteins they may not be so, but we would at least know that even before running the algorithms. Our method is based on homology-modeling. When a significantly homologous protein (to the query) with known structure is available in the database our prediction accuracy could be even 90% or above. Our objective is to improve the accuracy of the predictions for the so called "easy" proteins (where sufficiently similar homologues with known structures are available), rather than improving the bottom-line of the structure prediction problem, or the average prediction accuracy over many query proteins. We use digital signal processing (DSP) technique that is of global nature in assigning structural elements to the respective residues. This is the key to our success. We have tried some variation of the proposed core methodology and the experimental results are presented in this article.