Supervised learning of protein thermal stability using sequence mining and distribution statistics of network centrality

Motivation It is expected that the difference in the thermal stability of mesophilic and thermophilic proteins arises, in part at least, from the differences in their molecular structures and amino acid compositions. Existing machine learning approaches for supervised classification of proteins rely on the features derived from the structural networks and the amino acid sequences. However, the network features used leave out several important network centrality values, the statistic used is a simple average and the sequence features used are hand-picked leading to an accuracy of 90%. Results We show that discriminating sub-sequences of the amino acid sequences can significantly improve classification accuracy compared to the existing approaches of counting amino acids, di-peptide or even tri-peptide bonds. We identify notions of network centrality, specifically that depends on the distances between Cα atoms, that appears to correlate better with thermal stability compared to the existing network features. We also show how to generate better statistics from the node- and edge-wise centrality values that more accurately captures the variations in their values for different types of proteins. These improved feature selection techniques make it possible to classify between thermophilic and mesophilic proteins with 96% accuracy and 99% area under ROC. Availability The dataset and source code used are available at https://github.com/ankits0207/Protein_Classification_BIO699 Contact [email protected] online.