HeteroTCR: A heterogeneous graph neural network-based method for predicting peptide-TCR interaction.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• T-cell receptor (TCR) and peptide interactions are crucial in immunology and clinical applications.
• Existing models struggle with predicting binding for novel antigens or limited TCR repertoires.
• Unsupervised clustering models (UCMs) cannot directly predict peptide-TCR binding.

Purpose of the Study:

• To develop an accurate supervised predictive model (SPM) for peptide-TCR binding probabilities.
• To address the limitations of current SPMs in identifying novel antigens and diverse TCR repertoires.
• To introduce HeteroTCR, a novel SPM leveraging Heterogeneous Graph Neural Network (GNN) for enhanced prediction.

Main Methods:

• Proposed HeteroTCR, a novel SPM based on Heterogeneous Graph Neural Network (GNN).
• HeteroTCR integrates within-type (TCR-TCR, peptide-peptide) similarity and between-type (peptide-TCR) interaction data.
• Model performance evaluated on independent datasets and validated using single-cell data.

Main Results:

• HeteroTCR demonstrated superior performance compared to state-of-the-art models on independent datasets.
• Ablation studies confirmed the critical role of the Heterogeneous GNN module in capturing binding features.
• Validation with single-cell datasets showed predicted binding probabilities correlate with observed binding fractions.

Conclusions:

• HeteroTCR accurately predicts peptide-TCR binding probabilities, outperforming existing methods.
• The Heterogeneous GNN approach effectively captures key features of the peptide-TCR binding process.
• HeteroTCR offers a robust and reliable tool for analyzing immune system interactions, particularly for novel antigens.