Integrative machine learning approach for multi-class SCOP protein fold classification

Abstract

Classification and prediction of protein structure has been a central research theme in structural bioinformatics. Due to the imbalanced distribution of proteins over multi SCOP classification, most discriminative machine learning suffers the well-known ‘False Positives ’ problem when learning over these types of problems. We have devised eKISS, an ensemble machine learning specifically designed to increase the coverage of positive examples when learning under multiclass imbalanced data sets. We have applied eKISS to classify 25 SCOP folds and show that our learning system improved over classical learning methods.