Self-Contemplating In-Context Learning Enhances T Cell Receptor Generation for Novel Epitopes

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• T cell receptor (TCR) design for novel epitopes is crucial for targeted immunotherapy but limited by data scarcity.
• Existing computational methods struggle with generating TCRs for epitopes lacking known cognate TCRs.

Purpose of the Study:

• To develop a computational method for generating high-quality TCRs against novel epitopes.
• To address the under-explored challenge of TCR design for epitopes with no known binding partners.

Main Methods:

• Utilized in-context learning (ICT) by training models with target epitopes and known cognate TCRs.
• Introduced self-contemplation prompting (SCP) to generate TCR contexts for novel epitopes.
• Employed prompt selection based on binding affinity and authenticity metrics for enhanced generation.

Main Results:

• In-context training (ICT) is critical for effectively using context TCRs.
• Providing cognate TCRs significantly improves TCR generation for novel epitopes.
• SCP-generated contexts achieved performance comparable to or exceeding ground-truth contexts, especially with refined prompt selection.

Conclusions:

• In-context learning and self-contemplation prompting are effective strategies for TCR generation against novel epitopes.
• The proposed methods enhance the potential for developing novel TCR-based immunotherapies.
• Computational TCR design for under-explored epitopes is now more accessible and promising.