Work

Developing Deep Learning-based Models for Antibody-Antigen Complex Structure Prediction

Year: 2024
Type: dissertation
Abstract: Antibodies can specifically bind to antigens, making them highly effective for therapeutic applications. Understanding the structure of the antibody-antigen complex is essential to designing novel antibodies that target specific antigens. Despite significant advancements in experimental techniques in antibody-antigen complex structure determination, some of these methods remain labor-intensive and time-consuming, which make them unsuitable for high-throughput antibody screening. Therefore, there is a clear need for computational prediction of antibody-antigen complex structures. However, current end-to-end deep learning methods face challenges in predicting antibody-antigen complex structures due to the lack of coevolution information. Consequently, this underscores the necessity for a more effective antibody-antigen complex structure prediction tool. This thesis presents new methods for the computational prediction of antibody-antigen complex structures. First, we benchmark the performance of antibody-antigen complex structure prediction in traditional docking methods and established deep learning techniques, which have been successful in general protein modeling. Second, using insights from our benchmarking results, we develop an antibody-antigen complex structure prediction method by fine-tuning RoseTTAFold2. Lastly, we introduce a new deep-learning algorithm called Community Maturation (ComMat), an AlphaFold-based architecture employing the spirit of genetic algorithm. This method was successfully applied to the antibody H3 loop modeling task, which is crucial for high-resolution antibody-antigen complex structure prediction. The techniques presented in this thesis can serve as a starting point for accurate antibody-antigen complex structure prediction.
Author: 김유빈
Language: English
Open Access status: green
OpenAlex ID: W7046291912
Cites: 43
Cited by:
Related to: 10