Topological analysis of protein structures with statistical learning

Abstract

The study of Protein structure-function relationship has been of key focus in computational biology. A novel method of protein data analysis involves the use of Persistent Homology Analysis (PHA) as a tool for protein classification. The main method of feature engineering from intervals generated is using systematic approach of binning to characterise topological features. These features are then applied into 3 types of statistical learning methods: SVM, Tree-based methods and Neural Networks. Protein classification tasks used include: classification of hemoglobin molecules in relaxed and taut form (task 1) or the identification of all alpha, all beta and alpha-beta protein domains carried out on 450 and 900 proteins samples (task 2 and 3 respectively). The used of modified tree-based approach showed surprisingly stable results that attained the highest overall accuracy of 93.3% and 87.8% for task 2 and 3 respectively.