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

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Targeted immunotherapy relies on T cell receptors (TCRs) binding to specific epitopes.
• Designing TCRs for novel epitopes is difficult due to limited data and lack of known cognate TCRs.

Purpose of the Study:

• To generate high-quality TCRs for novel epitopes where no cognate TCRs are known.
• To explore the application of in-context learning for TCR generation in under-explored areas.

Main Methods:

• In-context training (ICT) to generate TCRs using a target epitope and known cognate TCRs.
• Self-contemplation prompting (SCP) to generate TCR contexts for novel epitopes lacking known binders.
• Refined prompt selection based on binding affinity and authenticity metrics.

Main Results:

• Aligning training and inference distributions via ICT is crucial for effective context utilization.
• Providing cognate TCRs as context significantly enhances TCR generation for novel epitopes.
• SCP-synthesized context TCRs achieve performance comparable to or exceeding ground-truth contexts.

Conclusions:

• In-context learning, particularly with ICT and SCP, offers a powerful approach for TCR design against novel epitopes.
• This method addresses a critical gap in computational TCR design, paving the way for advanced immunotherapies.