Yapısal özellikler ve makine öğrenme metodaları kullanarak PDZ domain etkileşimlerini ve sınıfını tahmin etme

PDZ domains (PSD-95/Discs-large/ZO-1 homology) are one of the most abundant and evolutionary conserved domain families through uni- and multi-cellular organisms. As its abundance and high evolutinary conservation indicates, PDZ domains mediate a number of distinct functions in the cell including vesicular sorting, neuronal synaptic plasticity, development and neural guidance. Therefore, malfunction of the PDZ domain causes several crucial diseases such as Usher's syndrome, epilepsy, schizophrenia and types of cancer. PDZ domains are 80-100 residues long and consist of two α-helices (αA- αB) and six β-sheets (βA to βF). The canonical interaction is the most common interaction type of PDZ domain where the PDZ domain binds the C-terminal of the target protein via the binding cavity. PDZ domains are categorized into three classes according to the motif of their binding partners as Class I, Class II and Class III. Altough, PDZ domains prefer to bind a particular class of peptides there are cases where the PDZ domain can interact with both Class I and Class II peptides, classified as Class I-II. This study focuses on building prediction models for PDZ domain mediated interactions and PDZ domain classification by using their structural features. By utilizing the properties of the PDZ domains and their ligands, those have the available interaction experimental data, an interaction prediction and classification model was built via machine learning approaches. One of the most robust machine learning approach, support vector machine (SVM) algorithm, was selected to train the models. The interaction prediction and the classification models performances were evaluated by cross-fold validation test and validation of human proteome scanning results on experimentally known interactions data. The interaction prediction model and the classification model have area under ROC curve with a number of 0.99 and 0.91, respectively. Moreover, the human proteome scanning results showed that the interaction prediction model was able to predict the known PDZ domain mediated interactions correctly with a TP rate of 17 %. Additionally, the general knowledge of classification of PDZ domains were supported by our results. These models could be utilized by future experimental studies to narrow the search space of the novel binding partners of PDZ domains as well as drug discovery studies that target PDZ domain containing proteins.