Designing T cell receptors with deep learning

Abstract

Immunotherapy has been proved as a promising therapeutic alternative to treat chronic viral diseases and cancer. Given the key role of T cells in protecting our body against viruses and altered tumor cells, immunotherapy based on T cells has been applied with success in cases our body presents a deficient response. Recently advances have allowed the engineering of the T cells to redirect their response to a specific target cell. However, T cell immunotherapy is still a complex challenge, especially because of the possibility to cause cross-reaction against normal cells. Thus, a huge effort has been placed into T cell engineering focused on a genetic edition of T cell receptors (TCR), which are the main component of the specific antigen recognition. The rationale behind TCR engineering is to design a TCR that directs a modified T cell to recognize infected or altered cells expressing a specific antigen on its surface through the major histocompatibility complex (MHC). The increasing number of available 3D structures of TCRs in complex with the antigenic peptide and the MHC (TCR:pMHC complex) combined with advances in computational resources and algorithms have boosted the understanding of the mechanism of antigen recognition by TCR, especially its variable CDR3 portion. The comprehension regarding TCR:pMHC made it possible to modify the TCR to optimize its affinity and specificity to a given antigen. However, the available structural information is still limited compared to the vast diversity of TCR repertoire, which difficult the generalization of predictions. In addition, the increase in TCR affinity for the pMHC can be detrimental to the recognition and produce cross-reactivity with self-antigens. On the other hand, developments in sequencing techniques made possible a more complete view of the TCR repertoire and their target antigen. The TCR:pMHC sequence information curated in several public datasets allows the application of data-driven approaches such as machine and deep learning to understand the TCR:pMHC recognition mechanism and generate fine-tuned TCR for a specific antigen. Thus, in this project we propose to use a sequence-to-sequence deep learning method, called Transformer, to generate specific TCR CDR3 sequences from a target antigen sequence. The transformer is a deep learning approach that is based on an encoder and decoder architecture utilizing multiple attention mechanisms. This architecture is applied by the Google team to learn and perform challenging tasks such as automatic text translation. (AU)