Multi-positive contrastive learning-based cross-attention model for T cell receptor-antigen binding prediction.

Area of Science:

• Immunology
• Computational Biology
• Artificial Intelligence

Background:

• T cells are vital for adaptive immunity, recognizing infected or cancerous cells via T cell receptors (TCRs) binding to peptide-MHC (pMHC) complexes.
• Experimental identification of immunogenic TCR-pMHC pairs is resource-intensive and challenging due to TCR diversity.
• Accurate prediction of TCR-pMHC binding is essential for developing effective immunotherapies.

Purpose of the Study:

• To develop a computational tool for predicting peptide-MHC (pMHC) and T cell receptor (TCR) binding.
• To improve the accuracy and interpretability of TCR-pMHC binding predictions.
• To leverage artificial intelligence for advancing TCR-based immunotherapies.

Main Methods:

• Proposing the Contrastive Cross-attention model for TCR (ConTCR) for pMHC-TCR binding prediction.
• Utilizing pretrained encoders for high-level feature embeddings of pMHC and TCR sequences.
• Employing multi-modal cross-attention and contrastive learning for enhanced feature extraction and model pretraining.

Main Results:

• ConTCR effectively captures critical information on TCR-pMHC interactions, with attention scores providing interpretability.
• The model demonstrates strong generalization for predicting binding specificity across diverse TCR repertoires and unseen epitopes.
• Achieved high AUC-ROC scores on both independent non-zero-shot (0.849, 0.950) and zero-shot (0.830, 0.938) test sets.

Conclusions:

• The ConTCR framework offers a promising solution for accurate pMHC-TCR binding prediction and model interpretability.
• ConTCR achieves higher precision than existing advanced models, showing its robustness.
• This AI tool holds significant promise for advancing TCR-based immunotherapies.