Protein Folding: Symbolic Refinement Competes with Neural Networks

Abstract

The Chou-Fasman Algorithm and its associated theory are non-learning methods to predict the secondary structure for proteins from the string of amino acids forming the protein. However, the overall accuracy of the predictions is not high and the predictions for the important structures are particularly low. Therefore, a range of methods have been used to improve the prediction of the secondary structure. We have applied our symbolic knowledge refinement system KRUST to the Chou-Fasman Theory to improve the accuracy of its predictions. We compare our results with several other approaches neural network, probabilistic, inductive learning and case-based. Symbolic refinement is particularly suitable since it makes use of an existing, but not very effective, predictive theory and the learned result takes the form of a similar theory, with the same representation. Testing has revealed that our symbolic refinement approach yields competitive predictions to the other methods, together with the advantage that the learned result retains human comprehensibility.