BERTrand-peptide:TCR binding prediction using Bidirectional Encoder Representations from Transformers augmented with...

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning in Immunology

Background:

• T-cell receptor (TCR) sequencing has increased peptide:TCR binding data.
• Existing machine learning models predict binding for fixed epitopes but struggle with novel peptides.
• Accurate prediction of peptide:TCR interactions is crucial for understanding immune responses.

Purpose of the Study:

• To develop a peptide:TCR binding prediction model with enhanced cross-peptide generalization capabilities.
• To improve the performance of predictive models for previously unseen peptide sequences.
• To provide a robust computational tool for immunoinformatics research.

Main Methods:

• A dataset of known peptide:TCR binders was curated and augmented with negative decoys from healthy donors' T-cell repertoires.
• Deep learning techniques, adapted from Natural Language Processing, were employed for model training.
• The BERTrand model was trained to predict peptide:TCR binding interactions.

Main Results:

• The BERTrand model achieved a cross-peptide generalization performance of 0.69 AUROC.
• BERTrand demonstrated superior performance compared to existing methods when evaluated on unseen peptide sequences.
• The model effectively generalizes predictions to novel peptide epitopes.

Conclusions:

• The developed BERTrand model offers improved prediction of peptide:TCR interactions, particularly for novel peptides.
• This advancement has significant implications for T-cell epitope discovery and vaccine design.
• The study provides open-access datasets and code for further research in immunoinformatics.