Computational identification of antigen-specific T cell groups through generative epitope modeling

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• Single-cell T cell receptor (TCR) sequencing offers deep insights into T cell diversity and clonal expansion.
• Determining the specific antigen recognized by individual T cells is difficult and often requires expensive functional assays.

Purpose of the Study:

• To introduce EpitopeGen, a novel transformer-based computational model.
• To enable population-level prediction of T cell antigen specificity directly from TCR sequences.
• To facilitate scalable analysis of T cell populations in various disease contexts.

Main Methods:

• EpitopeGen utilizes a transformer architecture for predicting epitope sequences from TCR sequences.
• A semi-supervised learning strategy was employed, involving a search of over 70 billion TCR-epitope pairs.
• High-confidence binding predictions were integrated as pseudo-labels with biologically informed quality control.

Main Results:

• Generated epitopes demonstrated high predicted binding probabilities, sequence diversity, and favorable biochemical and biophysical properties.
• In cancer datasets, EpitopeGen identified tumor-specific CD8+ T cells with cytotoxic signatures.
• In COVID-19 patients, the model distinguished T cell responses to different viral proteins across disease severities.

Conclusions:

• EpitopeGen provides a scalable computational framework for inferring T cell antigen specificity.
• The model complements experimental validation methods for T cell epitope discovery.
• It enables deeper investigation into disease-associated T cell populations and their functional roles.