Neural network based protein structure prediction

Feed-forward BPN architecture with a window size 13 has been used for secondary structure prediction. A novel approach has been devised where amino acids have been grouped into 12 classes with each class using a 12-bit unary-format representation. This avoids over-fitting and makes system less bulky. The system has been tested on a cross-validation set of 70 proteins comprising insulin and insulin-like structures. The validation results show Q_3total accuracy of 68%. The Q₃ accuracy´s for Helix, Strand and Coil prediction were 98%, 95% and 5%, respectively. The extraction of insulin-like polypeptide p (whose structure is not yet known) from Momordica Charantia is being aimed at to develop oral insulin of this institute. The NN model developed in this research has been used to predict secondary structure of this case protein. The results were benchmarked against the predictions made from the public domain servers.