Detection of remote homology in proteins by machine learning algorithms

The subject of this thesis is to develop a machine learning algorithm application that accurately performs remote homologous protein detection, which is an important problem in the field of bioinformatics. The discovery of remote homolog proteins is important because it is beneficial to discover the structure of unknown proteins. In the thesis, the problem of different lengths of protein sequences is solved by using natural language processing methods such as the bag of words model. The performances were measured by applying motifs of different lengths as protein features. A new application in this thesis provides a solution to the unbalanced data problem. This application, which is a KNN method with k-split with various distance methods, is a competitive study. Remote homologous proteins are a difficult problem to solve because they rely on small sequence similarities. In the thesis, another new application that trains with a new deep neural network that balances TF-IDF feature vectors calculated over n-grams with smoothing operations is carried out. The new application demonstrates the power of deep learning algorithms. The new application achieves better performance and overcomes the unbalanced data set.