epiTCR-KDA: knowledge distillation model on dihedral angles for TCR-peptide prediction

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Predicting T-cell receptor (TCR) and antigen binding is vital for immunotherapy development.
• Current predictors often fail on new data due to reliance on sequence data, neglecting structural features.
• Improved generalizability requires integrating structural information into prediction models.

Purpose of the Study:

• To develop a novel predictor, epiTCR-KDA, for TCR-peptide binding.
• To enhance prediction generalizability by incorporating structural information.
• To establish a more effective pipeline for antigen-based immunotherapy.

Main Methods:

• Developed epiTCR-KDA, a Knowledge Distillation model on Dihedral Angles (KDA).
• Utilized dihedral angles of TCR and peptide residues as structural descriptors.
• Integrated structural data into a knowledge distillation framework.

Main Results:

• Achieved an Area Under the Curve (AUC) of 1.00 for seen data and 0.91 for unseen data.
• Consistently outperformed other predictors on public datasets with a median AUC of 0.93.
• Identified cosine similarity of dihedral angle vectors as key to stable performance.

Conclusions:

• epiTCR-KDA offers a significant advancement in predicting TCR-peptide interactions.
• The model demonstrates superior generalizability by leveraging structural features.
• This work provides a highly effective pipeline for antigen-based immunotherapy.