Structure-based TCR-pMHC binding prediction and generalization to unseen peptides

Area of Science:

• Immunology
• Computational Biology
• Structural Biology

Background:

• T-cell receptor (TCR) and peptide-bound major histocompatibility complex (pMHC) interactions dictate adaptive immunity.
• Advancements in protein structure modeling enable structure-based computational methods for TCR recognition.
• Current graph neural network (GNN) classifiers exhibit limited accuracy for predicting TCR-pMHC binding specificity with unseen peptides.

Purpose of the Study:

• To comprehensively assess factors influencing the generalization performance of GNN-based TCR-pMHC binding specificity classifiers.
• To analyze the sensitivity of predictors to TCR-pMHC interface interaction features and structural uncertainty.
• To improve the generalization capabilities of GNN models for novel peptide recognition.

Main Methods:

• Utilized computationally predicted protein structures for TCR-pMHC binding interface analysis.
• Performed experiments to evaluate the impact of interaction features and structural uncertainty on classifier accuracy.
• Designed GNN classifier architectures incorporating auxiliary training objectives.

Main Results:

• Identified critical factors affecting the generalization performance of GNN-based TCR-pMHC specificity predictors.
• Demonstrated that specific classifier architecture designs and auxiliary training improve prediction accuracy for unseen peptides.
• Highlighted the challenges and limitations of current GNN approaches in generalizing to novel peptide recognition.

Conclusions:

• The generalization performance of GNN-based TCR-pMHC binding specificity classifiers is significantly influenced by interface features and structural uncertainty.
• Architectural modifications and auxiliary training objectives are effective strategies to enhance the ability of GNNs to recognize novel peptides.
• Further research is needed to overcome the limitations of current GNN paradigms for robust TCR-pMHC binding specificity prediction.