Residue Depth Prediction in protein kinases using machine learning

Activation of protein kinases by somatic mutation or chromosomal alteration is a usual mechanism in tumorigenesis. In this way, the suppression of activated protein kinases using of targeted small molecule drugs or antibody-based strategies is an effective cancer therapy approach. Therefore, understanding the protein structure and sensitive residues like Serine, Threonine and Tyrosine is an important task. Given a protein structure, it is relatively easy to generally distinguish a buried or exposed residue. However, recognition of burial status or burial degree of each residue using only the sequence information is a challenging task in biology and bioinformatics. In recent years, support vector machines (SVMs) as a powerful classification method gain more popularity among computational biologists. In this work, we introduce an efficient approach that extract useful information from the sequence and uses SVM to identify the depth level of residue using only the protein sequence. We systematically investigated a sequence encoding schemes that include 434 features. The model was built using 489 structures from protein data bank and its performance were evaluated by 135 protein kinases. The depth levels for all kinase residues were predicted using the proposed method by dividing the depth range in to three bins as follow: [0-3)Å, [3-10)Å and ≥ 10 Å. The prediction accuracies of the Serine, Threonine and Tyrosine residues were 66%, 32% and 88% for the first, second and third depth level respectively.