Return to search

Enhancing protein interaction prediction using deep learning and protein language models

Proteins are large macromolecules that play critical roles in many cellular activities in living organisms. These include catalyzing metabolic reactions, mediating signal transduction, DNA replication, responding to stimuli, and transporting molecules, to name a few. Proteins perform their functions by interacting with other proteins and molecules. As a result, determining the nature of such interactions is critically important in many areas of biology and medicine. The primary structure of a protein refers to its specific sequence of amino acids, while the tertiary structure refers to its unique 3D shape, and the quaternary structure refers to the interaction of multiple protein subunits to form a larger, more complex structure. While the number of experimentally determined tertiary and quaternary structures are limited, databases of protein sequences continue to grow at an unprecedented rate, providing a wealth of information for training and improving sequence-based models.

Recent developments in the sequence-based model using machine learning and deep learning has shown significant progress toward solving protein-related problems. Specifically, attention-based transformer models, a recent breakthrough in Natural Language Processing (NLP), has shown that large models trained on unlabeled data are able to learn powerful representations of protein sequences and can lead to significant improvements in understanding protein folding, function, and interactions, as well as in drug discovery and protein engineering.

The research in this thesis has pursued two objectives using sequence-based modeling. The first is to use deep learning techniques based on NLP to address an important problem in cellular immune system studies, namely, predicting Major Histocompatibility Complex (MHC)-Peptide binding. The second is to improve the performance of the Cluspro docking server, a well-known protein-protein docking tool, in three ways: (i) integrating Cluspro with AlphaFold2, a well-known accurate protein structure predictor, for enhanced protein model docking, (ii) predicting distance maps to improve docking accuracy, and (iii) using regression techniques to rank protein clusters for better results.

Identiferoai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/46642
Date30 August 2023
CreatorsHashemi, Nasser
ContributorsVakili, Pirooz, Kozakov, Dmytro
Source SetsBoston University
Languageen_US
Detected LanguageEnglish
TypeThesis/Dissertation

Page generated in 0.0015 seconds