Overcoming data bottlenecks in T cell receptor specificity prediction with effective machine learning
Year: 2024
Type: dissertation
Abstract: T cells are an integral part of the adaptive immune system. They detect host cells that have been compromised due to infection or mutation, and eliminate them through direct attack or recruitment of other effector immune cell types. This is achieved through T cell receptors (TCRs) expressed on their surface, which allows T cells to bind to peptide-major histocompatibility complexes (pMHCs) in a target-specific manner. Uncovering the rules of TCR-pMHC specificity has the potential for profound and positive impact in biomedicine. However, this remains an unsolved challenge made particularly difficult by the immense numbers of possible TCRs (∼ 10^60) and pMHCs (∼ 10^15 ). While it remains implausible to empirically map out anywhere near a majority of the possible binders, an application of machine learning may help us better understand the binding rules by extrapolating from existing data. Recent advances in the natural language processing (NLP) field have demonstrated the impressive ability of transformer language models to learn from unsupervised objectives using large corpora of unlabelled text. Since TCRs, like other proteins, can naturally be represented as a sequence of amino acids, there has been growing interest in applying language modelling technologies to the TCR domain. However, how to most effectively design unsupervised training objectives to optimise language models for downstream TCR-pMHC specificity prediction remains an open question. The core theme of this thesis is the investigation of contrastive learning as a method of training transformer-based TCR representation models. In this regard, I show that combining unsupervised contrastive learning (autocontrastive learning) with the traditional masked-language modelling (MLM) objective is a highly effective way of pre-training a TCR representation model. In addition to the above, the thesis presents related work that I have conducted during my PhD candidacy around automated TCR data standardisation as well as a statistical framework for calibrating TCR distance metrics to probabilities of TCR co-specificity.
Author: Nagano, Yuta
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7061097464