Protein yapı tahmini için derin öğrenme modellerinin geliştirilmesi

The three-dimensional structure of a protein provides important clues about the function of that protein. Although there have been many studies on protein structure prediction, the problem has still not been solved completely. As it is very difficult to predict the three-dimensional structure of a protein directly, predictions of structural properties of proteins such as secondary structure, solvent accessibility, and torsion angles are carried out first, which are later used as inputs to more elaborate structure estimation tasks. In this thesis, novel deep learning models have been developed by using convolutional neural networks (CNN), graph convolutional networks (GCN) and long-short-term memory (LSTM) recurrent neural networks to predict secondary structure, solvent accessibility and torsion angles of proteins. A rich feature set formed by using PSI-BLAST, HHBlits, physicochemical properties, structural profile matrices, AA index values, and graphs representing the relationship between amino acids were used as inputs to the models. In the first study, a deep learning model was developed by using CNN and GCN layers for secondary structure prediction. In the second study, LSTM layers were added to the first model, which was extended to make solvent accessibility and torsion angle predictions as well using the multi-task learning approach. In both studies, graphs were generated using neighborhood relations between amino acids. In the last study, a novel U-net-based model was designed for secondary structure prediction using CNN, GCN, and LSTM layers. The graph matrices used as input to GCN layers were obtained by using protein contact map prediction. All models were trained, optimized and tested on benchmark data sets. Improvements were obtained in accuracy as compared to the state-of-the-art