Development of artificial intelligence systems for identification and classification of gene sequences

There are billions of virus species worldwide, and as the smallest parasitic entities, viruses pose a significant threat. Given the vast diversity and rapid evolution of viruses, there is a critical need for the rapid and accurate classification of viral species and their potential hosts to better understand transmission dynamics and facilitate the development of targeted treatments. In this context, the PhyVirus dataset, which consists of pathogenic Single-Stranded RNA viruses and contains different viral species and hosts, is analyzed in this study. The thesis consists of three main chapters and each chapter approaches the classification of genetic sequences from a different perspective. In the first section, viral families and hosts are classified using the K-Mer encoding method along with machine learning (ML) and deep learning (DL) algorithms, such as Random Forest, Gradient Boosting, Extra Trees, and Fully Connected Deep Neural Network (FCDNN). Prediction of virus families with FCDNN method with %99,60 success rate is one of the important results of the study. In host prediction, the highest success rate of %81,53 was obtained with the ExtraTrees classifier. The impact of different K-Mer word lengths on the classification of viral families and hosts was evaluated, and evolutionary relationships, genetic similarities, and host relatedness were examined based on classification results and literature review. In the second section, classification of genetic sequences was performed using graphical and image-based encoding techniques (FCGR, DNAWalk, and Grayscale Transformation). These techniques were analyzed with a CNN model (InceptionV3) and an accuracy rate of %99,89 was achieved with the Grayscale Transform method. In the DNAWalk coding method, the genetic sequence trajectory images were classified with an accuracy of %99,14. In the FCGR coding method, the highest accuracy of %99,85 was obtained with word lengths between 3 and 8. These methods allowed for more accurate classification of viral families and hosts. Upon reviewing the existing literature, no other study was found that comprehensively analyzes the effects of different encoding methods on classification performance using a single dataset. This thesis aims to fill a significant gap in the field and make a meaningful contribution to the literature. Gene sequences are made ready for analysis through various biological and technical processes. However, errors that may occur at any stage of these processes may cause missing data in gene sequences. Missing data prediction, which is frequently discussed in the literature, is usually based on existing methods that require the data to be aligned. In the third part of the thesis, missing data prediction methods are discussed and a new approach for the KNN-Imputation method is developed. The different lengths of the gene sequences in the PhyVirus dataset prevented the direct application of existing missing data prediction methods. This issue was resolved by the newly developed KNN-Imputation approach, which provided a unique contribution to the study. This thesis aims to develop innovative approaches for encoding, classifying, and imputing missing data in genetic sequences and to demonstrate how these methods can be applied in bioinformatics research. The results obtained aim to be an important reference for scientific studies in this field by providing new methodological contributions to viral genome analysis and classification processes.