Effective framework for protein structure prediction

This paper presents a computational system to predict protein structure using N–grams and a wrapper feature selection framework (the N–gram is a subsequence composed of N characters, extracted from a larger sequence). N–gram features are extracted from a dataset consisting of 277 domains: 70 all–α domains, 61 all–β domains, 81 α/β domains and 65 α + β domains. A wrapper feature selection system, GA–SVM, is applied to obtain an optimised feature set. Using the optimised 3070–feature subset, a classifier model is trained and tested in the Support Vector Machine (SVM) learning system. This model achieves an overall accuracy of 88.09%, evaluated by a 10–fold cross–validation test. This value is 4.7% higher than the one using the initial 6,414 features. Experimental results also illustrate that employing a feature subset selection, by using the proposed GA–SVM wrapper approach, has enhanced classification accuracy in comparison to other GA–based wrapper approaches and existing protein sequence encoding methods.