Improved prediction methods for understanding the TCR-peptide-MHC interaction

Abstract

The last decades have seen a rapid increase in our understanding of the immune system, but there are still many unsolved problems. Solving some of these could be invaluable for future advances in drug development and cancer immunotherapy. This thesis introduces methods for understanding an important interaction in the adaptive immune system. One of the key events in the adaptive immune system is the interaction between T-cell expressed receptors (TCRs) and peptides bound to major histocompatibility complexes (pMHCs). If the TCR recognizes a pMHC, the T-cell is activated and the peptide driving this activation is called a T-cell epitope. Predicting T-cell epitopes has been a long standing challenge within the field of immunoinformatics. There are two strategies to solve the problem. One is to use the protein sequences, and the other is to use the structures. Data on protein structures is usually quite limited, so developing reliable tools that use just the sequences is of great interest to the field. A commonly used measure for identifying T-cell epitopes is the pMHC binding strength, as this quantity can be used to limit the number of potential peptide candidates. In the first project of this thesis, we develop an improved method for predicting such peptide-MHC binding strengths by training a neural network on an extended dataset of peptide binding affinities. Further, we show that the updated methods have superior performance when used for detecting T-cell epitopes. However, not all MHC presented peptides are immunogenic. So in order to truly understand what makes a peptide immunogenic we need to understand the interaction between TCRs and pMHCs. One way to do this is to build structural models of the TCR-pMHC complex and use these structures to predict the TCR-pMHC binding strength. In the second project, we develop an automated tool for building such structural models of the TCR-pMHC complex using only the amino acid sequence as input. The tool utilizes comparative modeling techniques and generates accurate models within minutes. In the third project, we investigate the TCR recognition of pMHCs using an experimental technique which measures the relative binding affinity between TCRs and pMHC variants. The relative binding affinities can be translated into TCR motifs, named TCR fingerprints, and these can be used to identify which peptides can be cross-recognized by the TCR. Structural modeling is used in this project to investigate how the TCR recognition is affected by conformational changes in the peptides. In the fourth and last project, we present preliminary results on improving structural models of TCRs by using state-of-the-art machine learning techniques to generate the peptide-binding loops. Collectively, the four projects of the thesis provide improved methods for predicting T-cell epitopes and for structural modeling of the TCR-pMHC complex. We hope that these methods can increase our understanding of T-cell immunogenicity and serve as a foundation for developing improved methods for rational T-cell epitope predictions.