TCRLens: structure-aware equivariant graph learning for TCR-pMHC-I recognition and immunogenic epitope discovery

Area of Science:

• Computational immunology
• Structural bioinformatics
• Machine learning in biology

Background:

• Predicting T-cell receptor (TCR) and peptide-MHC class I (pMHC-I) interactions is crucial for understanding immune responses but is challenging due to structural diversity and limited data.
• Existing methods often struggle with data sparsity and class imbalance, particularly in identifying weak-affinity interactions.
• Accurate modeling is essential for applications like epitope discovery and vaccine design.

Purpose of the Study:

• To develop a structure-aware deep learning framework, TCRLens, for accurate prediction of TCR-pMHC-I interactions.
• To address data sparsity and class imbalance issues using generative data augmentation.
• To evaluate TCRLens performance across multiple prediction tasks and assess its cross-species generalization capabilities.

Main Methods:

• Developed TCRLens, a deep learning framework utilizing multi-scale graph representations and an equivariant graph neural network (EGNN) to model residue-level interactions at the TCR-pMHC-I interface.
• Incorporated a variational autoencoder-generative adversarial network (VAE-GAN) to generate structurally plausible weak-affinity interaction samples, mitigating data sparsity and class imbalance.
• Evaluated TCRLens on peptide-MHC binding, peptide-TCR recognition, and full-complex TCR-pMHC-I interaction prediction tasks using curated datasets from ATLAS and TCR3d.

Main Results:

• TCRLens significantly outperformed state-of-the-art sequence-based, motif-based, and structure-aware methods, including NetMHCpan 4.2 and CapsNet-MHC, across all evaluated prediction tasks.
• Demonstrated robust cross-species generalization, achieving high predictive performance in swine and chicken MHC-I systems.
• Highlighted the effectiveness of geometry-aware representation learning and generative data augmentation in capturing immunological specificity.

Conclusions:

• TCRLens provides a powerful, unified, and extensible platform for modeling TCR-pMHC-I interactions.
• The framework's structure-aware approach and generative data augmentation are key to its superior performance.
• TCRLens has significant potential for advancing epitope discovery and structure-guided vaccine design in both human and veterinary immunology.