LANTERN: TCR-peptide binding prediction via large language model representations.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Predicting T-cell receptor (TCR) and peptide-major histocompatibility complex (pMHC) interactions is crucial for developing targeted immunotherapies.
• Current models face challenges with limited data and poor generalization to new epitopes.

Purpose of the Study:

• To introduce LANTERN, a novel deep learning framework for enhanced TCR-pMHC interaction prediction.
• To improve the generalization capabilities of TCR-pMHC binding prediction models, especially in zero-shot and few-shot learning scenarios.

Main Methods:

• LANTERN utilizes pretrained protein (ESM) and molecular (MolFormer) language models to encode TCR sequences and peptide SMILES strings.
• A Multi-Head Cross-Attention (MHCA) module integrates these representations, focusing on interaction-relevant features.
• The framework combines evolutionary and chemical properties for robust prediction.

Main Results:

• LANTERN demonstrates competitive and robust performance on the TCHard benchmark.
• The model shows improved generalization, particularly in random control and unseen epitope settings.
• LANTERN achieves state-of-the-art or comparable results against existing methods.

Conclusions:

• LANTERN offers a powerful new approach for accurate TCR-pMHC binding prediction.
• The framework has significant potential for applications in personalized immunotherapy and vaccine development.
• The study provides a valuable tool for advancing immunoinformatics research.