Protein fold identification using machine learning methods on contact maps

Proteins can be classified among the four structural classes of All-Alpha, All-Beta, Alpha+Beta and Alpha/Beta which are further subdivided into 27 folds. Protein fold classification problem is cited in the literature as a challenging unbalanced classification problem with the accuracy results being as low as 51.1% on the bench mark data set of Ding et al. and highest accuracy at 60.5% using 2500 features. We represent the proteins as 11-length feature vectors and adopt Synthetic Minority Over-sampling Technique (SMOTE) based boosting approach to balance the data to address the 27-way fold classification problem. We build C4.5 decision tree classifier in combination with SMOTE boosting algorithm using the novel contact map features and show that the prediction accuracy is enhanced to 64%. An additional advantage of our approach is the reduced dimensionality of the feature vector which is 11 whereas literature uses more than 100 features on average. Further, we propose an algorithm ExtractPatterns that extracts (non) rectangular 2D regions of contacts from the off-diagonal region in linear time. A feature vector of length 11 is formed using this study constituting diagonal and off-diagonal statistical features.