Analysis of protein metal-binding sites using deep neural networks

Proteins fold by forming strong bonds with the metal ions in their environment and reach their three-dimensional structure. The three-dimensional structure of proteins shows which critical function it performs in the cell. Prediction of protein metal binding sites using protein sequence is important for predicting protein structure, functions, and drug discovery. Computational estimates using machine learning methods based on data from amino acid sequences are widely used in various bioinformatics fields. In this thesis, three different deep learning architectures are proposed for the prediction of metal binding status of Histidine (HIS) and Cysteine (CYS) amino acids in protein sequences. These architectures are built on convolutional neural network (CNN), long-short term memory (LSM) and gated recurrent unit (GRU) models, respectively. These architectures are developed using Keras with Tensorflow backend. Since these models cannot work directly on sequence data, digitization techniques based on PAM scoring matrix, protein compositions and binary representation methods have been applied to feed the relevant models. Developed architectures and protein sequence digitization methods have been tested on benchmark data set consisting of 2727 proteins. The results obtained were compared with the results obtained with Naïve Bayes, Support vector machines, Adaboost and Bagging machine learning methods. It seems that the best results for prediction of protein metal binding site are obtained with CNN architecture. This result shows that better performance was obtained with the same dataset than other studies in the literature. Using these results, the geometry of the metal binding sites was evaluated in order to decide which residues are involved in the coordination of a metal ion.