Protein secondary structure prediction using deep learning method

Protein structure prediction has been a central focus of study in Bioinformatics. In the past few decades, many statistical methods, such as complex machine learning, followed by deep learning methods have been applied to estimate structural information of protein. Since protein is a significant part of living-organisms, understanding and assessing protein and its functions becomes crucial. Proteins are made by building block, called amino acid. Although protein structure is largely determined by amino acid sequences, known as primary structure, it is difficult to predict protein structure from those sequences alone. Thus, protein secondary structure prediction from the sequences is an important step for the estimation of protein three- dimensional structure. Many approaches have been employed onto protein secondary structure prediction studies. However, up to present days, none of the available techniques in literature is able to provide a fully accurate result, which makes the study more challenging. By using CB513 dataset, this thesis attempts to provide a comparative study of the use of deep learning approaches, CNN, RNN, LSTM and GRU. In the study, the performance of each approach was analyzed and compared with the similar studies in literature. The models, CNN, RNN, LSTM and GRU, developed for protein secondary structure prediction in this study achieved %82,54, %81,06, %81,10, %81,48 accuracy.