Exploring the potential of structure-based deep learning approaches for T cell receptor design

Area of Science:

• Computational biology
• Immunoinformatics
• Protein engineering

Background:

• Deep learning has advanced protein modeling and design, enabling novel protein creation.
• Engineering T cell receptors (TCRs) is a promising strategy for cancer immunotherapy.
• Existing physics-based methods face challenges in designing TCRs due to interface complexities.

Purpose of the Study:

• To explore the potential of deep learning methods (ProteinMPNN, ESM-IF1) for designing fixed-backbone TCRs.
• To evaluate TCR designs for binding target antigenic peptides presented by MHC.
• To compare deep learning approaches against classical physics-based methods.

Main Methods:

• Utilized structure-based deep learning protein design tools: ProteinMPNN and ESM-IF1.
• Designed fixed-backbone TCRs targeting specific peptide-MHC complexes.
• Employed a comprehensive suite of sequence- and structure-based evaluation metrics.

Main Results:

• Deep learning methods demonstrated potential in designing TCRs for specific peptide-MHC binding.
• These AI-driven approaches showed benefits compared to traditional physics-based design.
• Identified areas for improvement in current deep learning-based TCR design.

Conclusions:

• Structure-based deep learning holds significant promise for advancing TCR-based immunotherapeutic design.
• Further refinement of AI methods is needed to overcome current limitations in TCR engineering.
• This study provides a foundation for developing next-generation TCR-targeting cancer therapies.