Prediction of protein structural classes by decreasing nearest neighbor error rate

Prediction of protein structural classes has been proven to be significant in the field of bioinformatics. A good computational prediction technique may improve the prediction accuracy. In this paper, a new predictor from proteins´ primary sequences is proposed to determine protein structural classes. Firstly, a feature vector which serially fuses pseudo amino acid composition and pseudo position-specific scoring matrix is constructed. Secondly, the classifier based on nearest neighbor error rate is employed and then a heuristic algorithm is proposed to decrease the error rate. Finally, leave-one-out cross-validation is adopted to evaluate our approach on 4 benchmark datasets (Z277, Z498, C204 and W1189). The experimental results demonstrated that our method achieves satisfactory performance in comparison with other existing methods.