A novel variant effect prediction model based on protein representation with deep learning architecture

When identifying single amino acid variants in the human genome that may cause a detrimental effect on the phenotype, the evolutionary conservation and structural features of the protein sequence are generally examined. Using such properties of proteins, variant effect prediction can be made with computational approaches such as statistical methods and machine learning algorithms. In this thesis, we tried to predict the pathogenic or benign effects of ABL1 protein-specific variants on the phenotype using deep learning and natural language processing methods, which are called innovative variant effect prediction methods. ABL1 variant data obtained from UniProt, ClinVar, EVS and dbSNP databases were used in the developed prediction models. These data were divided into training, validation and test data sets according to six different scenarios determined within the scope of the thesis study. The ABL1 protein sequence used as input in the models was partitioned according to different partitioning methods such as standard partitioning and regional partitioning. These obtained variant sequences were sized according to the embedding representation method. The performance of the models was evaluated according to accuracy, precision, recall, F-measure, specificity and ROC-AUC metrics. When the performances of the developed prediction models are compared according to the scenarios, it is seen that the highest prediction performance is achieved with Convolution Layer modeling (AUC: 0.86). Accordingly, benign variants can be classified with 93% accuracy (selectivity: 0.93) by using the deep learning model with the most successful Convolution Layer only for the classification of benign variants. In the thesis study, a successful model was developed to predict the harmful or harmless effects of protein variants, specifically the ABL1 protein, on the phenotype.