Disrupting explicit encoding paradigms: property-interactive transformers decode T-cell receptor specificity beyond...

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• T-cell receptors (TCRs) are crucial for immune surveillance, recognizing specific peptide antigens.
• Existing deep learning models for TCR-peptide binding prediction often learn dataset biases and neglect biochemical and spatial binding properties.

Purpose of the Study:

• To develop a novel deep learning model that accurately predicts TCR-peptide binding by incorporating spatial and environmental information.
• To overcome the limitations of current models that overestimate results due to dataset biases.

Main Methods:

• Developed TCRoss, a transformer-based model utilizing cross-mapped amino acid properties to implicitly simulate spatial structure.
• Incorporated environmental information into the training dataset to mitigate learning biases.
• Validated model predictions using wet-lab T-cell activation assays and biophysical analysis.

Main Results:

• TCRoss effectively captures TCR-peptide binding interactions by simulating spatial properties through cross-mapped amino acid interactions.
• Including environmental data improved model performance and reduced dataset biases.
• TCRoss demonstrated superior performance compared to existing models in both known and novel peptide scenarios.
• Wet-lab and biophysical validation confirmed the model's predictive accuracy and biological relevance of high-attention residue pairs.

Conclusions:

• The TCRoss model offers a significant advancement in predicting TCR-peptide binding by integrating spatial and environmental data.
• This approach enhances the accuracy of immune response prediction and provides a more reliable tool for immunological research.
• The findings highlight the importance of considering biochemical and spatial factors in deep learning for biological systems.