Refinement of Antibody Structure Prediction with Deep Learing

Prediction of the 3D tertiary structure of the most variable region of antibody (H3 loop) based on its amino-acid sequence is an important unsolved problem in bioinformatics. Existing methods rarely provide high quality results and, on average, their prediction accuracy remains extremely low. One of the problems is that existing solutions do not use deep learning methods, which have proved to outperform classic algorithms and have become the standard solution in areas such as computer vision, natural language processing, etc. In this thesis, a method for predicting the structure of the H3 loop using only deep learning was proposed. Several methods for encoding amino acid sequences, endowing them with structural and semantic information, as well as enabling them to overcome the problem of small amounts of data, were considered. Finally, several different deep neural network architectures, which have shown outstanding results in natural language processing, were applied in this thesis. The best model obtained was tested on a previously unknown data set and showed results similar to those of existing methods.