VitTCR: A deep learning method for peptide recognition prediction.

Area of Science:

• Computational biology
• Immunoinformatics
• Machine learning in immunology

Background:

• T cell receptors (TCRs) recognize peptide antigens presented by MHC molecules.
• Accurate prediction of TCR-peptide interactions is vital for designing effective cancer immunotherapies and vaccines.
• Existing computational models have limitations in predicting these complex interactions.

Purpose of the Study:

• To develop and evaluate VitTCR, a novel predictive model for TCR-peptide interactions.
• To leverage the vision transformer (ViT) architecture for enhanced prediction accuracy.
• To assess the biological relevance and potential applications of the model in immunotherapy and vaccine development.

Main Methods:

• Development of VitTCR, a model utilizing the vision transformer (ViT) architecture.
• Conversion of TCR-peptide interaction data into numerical AtchleyMaps using Atchley factors.
• Integration of a positional bias weight matrix (PBWM) for improved accuracy.
• Benchmarking VitTCR against existing computational models.

Main Results:

• VitTCR achieved an AUROC of 0.6485 and AUPR of 0.6295 in predicting TCR-peptide interactions.
• Model performance was comparable to other benchmarked models.
• Incorporating PBWM slightly enhanced prediction accuracy.
• Model predictions showed statistically significant correlations with immunological factors like T cell clonal expansion and activation.

Conclusions:

• VitTCR is a valuable computational tool for predicting TCR-peptide interactions.
• The model offers insights relevant to the development of cancer immunotherapies and vaccines.
• Further comparative studies are recommended to explore VitTCR's effectiveness across diverse contexts.