Improving prediction of protein secondary structure using physicochemical properties of amino acids

Abstract

Protein Structure Prediction is important in the sense that it helps to extend knowledge about the understanding of protein structures and functions. The knowledge is essential for prediction of secondary structures of unknown proteins required for applications related to drug discovery. A novel technique for protein secondary structure prediction is presented here. In this work, two levels of multi-layer feed forward neural networks are used. In the first level network, sequence profiles from PSI-BLAST and physicochemical properties of amino acids are used for sequence to structure predictions. Confidence values of forming helix, sheet and coil, obtained from the first level network are then used with the second level network for structure to structure predictions. The overall prediction accuracy as obtained through experimentation is in the range of 75.58% to 77.48%. This method is trained and tested with nrDSSP datasets using four folds cross validation. It is also tested on target proteins of Critical Assessment of Protein Structure Prediction Experiment (CASP3) and achieves better results than PSIPRED over some target proteins.