A Biologically Informed Vision-Guided Framework for Interpretable T Cell Receptor-Epitope Binding Prediction.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of T-cell receptor (TCR)-epitope binding is crucial for cancer immunotherapy.
• Current predictive models face challenges in generalization and incorporating key physicochemical properties.

Purpose of the Study:

• To propose DAISY, a biologically informed, vision-guided deep learning framework for robust and interpretable TCR-epitope binding prediction.
• To improve generalization to unseen epitopes and integrate physicochemical properties.

Main Methods:

• DAISY integrates hierarchical physicochemical features using a Condition-Adaptive Fusion module.
• It models residue-level spatial interactions and global biochemical context.
• Score-CAM visualizations provide interpretability by localizing interaction-relevant residues.

Main Results:

• DAISY consistently outperforms state-of-the-art models across four generalization scenarios.
• It achieved an 11% improvement in ROC-AUC and a 16% improvement in PR-AUC in the Unseen-Pair setting.
• Predictions correlate with T-cell clonal expansion, functional TCR identification, and patient survival forecasting.

Conclusions:

• DAISY offers a powerful tool for translational immunology and immune modeling.
• It provides a scalable paradigm for next-generation immune modeling.
• The framework enhances the prediction of TCR-epitope interactions for immunotherapy applications.