STAG-LLM: Predicting TCR-pHLA binding with protein language models and computationally generated 3D structures

Area of Science:

• Immunology
• Computational Biology
• Machine Learning

Background:

• T cell receptor (TCR) and peptide-HLA (pHLA) binding is vital for adaptive immunity.
• Accurate binding specificity prediction aids personalized immunotherapy design.
• Current methods primarily use amino acid sequences, neglecting structural information.

Purpose of the Study:

• To develop a multimodal machine learning (ML) model for TCR-pHLA binding specificity prediction.
• To integrate 3D structural data with sequence data for enhanced prediction accuracy.
• To address challenges associated with using computationally generated 3D structures in ML pipelines.

Main Methods:

• Developed STAG-LLM, a multimodal ML model combining a protein language model and geometric deep learning.
• Utilized computationally generated 3D protein structures alongside amino acid sequences.
• Incorporated strategies to manage inference costs, limited training data, and structural noise.

Main Results:

• STAG-LLM demonstrated superior performance in predicting TCR-pHLA binding specificity compared to existing methods.
• The model achieved high accuracy even with smaller training datasets.
• In vitro alanine scanning experiments showed correlation with model attention weights, validating predictions.

Conclusions:

• STAG-LLM shows significant potential for structure-based TCR-pHLA binding prediction.
• The model provides a foundation for advancing immunological and proteomic studies using modeled 3D structures.
• The utility of STAG-LLM is expected to grow with advancements in protein structure and language models.