Weighted amino acid composition based on amino acid indices for prediction of protein structural classes

Prediction of protein structural classes is one of the most important and challenging tasks in the bioinformatics field. A protein is classified into one of the four main types of protein structural classes; all-¿, all-ß, ¿/ß and ¿+ß. This paper investigates the role of amino acid indices (AAI) combined with traditional amino acid composition (AAC) to create a weighted amino acid composition (WAAC) feature-set to predict the structural class of a protein. There are over 500 amino acid indices that can be used to develop the novel weighted amino acid composition feature-set which has a great potential of increasing accuracy for the prediction of protein structural classes. For evaluation of these indices a high quality 40% homology dataset is used that contains over 7000 protein sequences (the largest of its kind) extracted from proteomic databases. The predictive technique developed is an optimum k-nearest-neighbour classifier, named multiple-k-nearest-neighbour (MKNN). In order to evaluate the classifier a 10- fold cross-validation test procedure is used throughout the study. Over 1 million analyses were carried out, the highest accuracy obtained was from index LEVM780101 at 48.35%, which is 9% higher than traditional AAC and 6.6% higher than that of the best sequence-driven-feature sub-set used in other studies. There is great potential for further improvement as WAAC is a feature-set with the least number of attributes without any feature selection and the numbers of indices that yielded higher accuracies than traditional AAC and other sequence-driven-features are 536 and 435, respectively, out of the 548 amino acid indices analysed in this study.