Attention network for predicting T-cell receptor-peptide binding can associate attention with interpretable protein...

Area of Science:

• Immunology
• Computational Biology
• Structural Biology

Background:

• T-cell receptor (TCR) recognition of peptide-major histocompatibility complex (pMHC) is vital for immune responses and disease.
• Experimental determination of TCR-pMHC interactions is costly and time-intensive.
• Existing computational methods lack external validation and integration of advanced neural network features.

Purpose of the Study:

• To develop a novel machine learning model for predicting TCR-pMHC interactions.
• To utilize a modified Transformer architecture with attention mechanisms for sequence-based prediction.
• To provide an interpretable model that links prediction weights to structural properties.

Main Methods:

• Developed a source-target attention neural network based on the Transformer architecture.
• Trained the model using amino acid sequences of TCR complementarity-determining region 3 (CDR3) and peptides.
• Validated the model on benchmark and external datasets, analyzing attention weights for structural insights.

Main Results:

• The model achieved competitive performance in predicting TCR-pMHC interactions on both internal and external datasets.
• Analysis of attention weights revealed statistically significant associations between highly attended residues and structural properties like hydrogen bonds within CDR3.
• Created a novel dataset for TCR-pMHC interaction prediction.

Conclusions:

• The developed machine learning model accurately predicts TCR-pMHC binding from amino acid sequences.
• The model's interpretability offers insights into molecular recognition mechanisms.
• This approach facilitates the design of targeted immunotherapeutics and advances understanding of immune interactions.