Predicting TCR sequences for unseen antigen epitopes using structural and sequence features.

Area of Science:

• Computational immunology
• Bioinformatics
• Structural biology

Background:

• T-cell receptor (TCR) recognition of antigens is crucial for adaptive immunity.
• Large datasets of TCR-antigen pairs enable computational prediction models.
• Accurate prediction of binding affinities for novel TCR-antigen pairs remains a challenge.

Purpose of the Study:

• To present CATCR, a novel framework for enhanced prediction of epitope and TCR interactions.
• To improve the accuracy of forecasting TCR-epitope binding affinities.

Main Methods:

• Utilized convolutional neural networks for peptide feature extraction from residue contact matrices (OpenFold).
• Employed a transformer for encoding segment-based coded sequences.
• Introduced CATCR-D (discriminator) for binding assessment using structural and sequence features.
• Developed CATCR-G (generative module) for predicting CDR3-β sequences against epitope characteristics.

Main Results:

• CATCR-D achieved an AUROC of 0.89 on unseen epitope-TCR pairs, outperforming benchmarks by 17.4%.
• CATCR-G demonstrated precision, recall, and F1 scores exceeding 95% in bidirectional encoder representations from transformers assessments.
• The framework effectively predicts unseen epitope-TCR interactions.

Conclusions:

• CATCR is an effective tool for predicting novel epitope-TCR interactions.
• Incorporating structural insights significantly advances understanding of TCR-epitope recognition.
• Predicting TCRs for novel epitopes using structural and sequence data holds promise for improving binding predictions.