Conditional generation of real antigen-specific T cell receptor sequences.

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• Designing functional T cell receptors (TCRs) for specific targets is difficult due to complex cross-reactivity rules and limited data.
• Existing methods struggle with generating novel TCR sequences against arbitrary antigens.

Purpose of the Study:

• To develop a computational framework, TCR-TRANSLATE, for generating antigen-specific TCR sequences.
• To evaluate and optimize sequence-to-sequence models for TCR design.
• To experimentally validate a computationally designed TCR against a therapeutically relevant target.

Main Methods:

• Utilized a sequence-to-sequence framework (TCR-TRANSLATE) adapting machine translation techniques.
• Evaluated 12 model variants based on BART and T5 architectures.
• Experimentally validated a designed TCR against the Wilms' tumour antigen.

Main Results:

• The flagship model, TCRT5, demonstrated superior performance on computational benchmarks, ranking functional sequences higher.
• A computationally designed TCR was successfully validated against the Wilms' tumour antigen, a non-viral epitope.
• The designed TCR exhibited cross-reactivity with pathogen-derived peptides, indicating specificity limitations.

Conclusions:

• TCR-TRANSLATE provides a foundation for computational TCR design, enabling the generation of functional TCRs against novel epitopes.
• This work highlights the potential for accelerating personalized immunotherapy by reducing the search space for therapeutic targets.
• Further research is needed to address TCR specificity and improve data availability for enhanced computational design.