Comparative Analysis of TCR and TCR-pMHC Complex Structure Prediction Tools

Area of Science:

• Computational biology
• Structural bioinformatics
• Immunoinformatics

Background:

• AI advancements like AlphaFold have accelerated computational structure prediction for T cell receptors (TCRs) and TCR-peptide-major histocompatibility (TCR-pMHC) complexes.
• Standardized benchmarks are lacking to assess the accuracy and limitations of these rapidly developing prediction tools.

Purpose of the Study:

• To systematically evaluate the performance of various computational tools for predicting isolated TCR structures and TCR-pMHC complex structures.
• To identify the strengths and weaknesses of different prediction methods, including homology-based, general AI, and TCR-specific approaches.

Main Methods:

• Evaluated seven tools for TCR structure prediction and six tools for TCR-pMHC complex structure prediction.
• Utilized a post-training dataset of 40 αβ TCRs and 27 TCR-pMHC complexes (Class I and II).
• Assessed model accuracy at global, local, and interface levels using diverse metrics.

Main Results:

• AlphaFold2, AlphaFold3, and tFold-TCR demonstrated high overall accuracy for TCR structure prediction.
• TCRmodel2 and AlphaFold2 showed good performance in TCR-pMHC structure prediction.
• TCR-specific AlphaFold2 tools had lower accuracy in framework regions; all tools struggled with CDR3 loops, docking, and certain MHC interfaces.

Conclusions:

• No single tool excels in all aspects; tool selection depends on specific prediction needs.
• Emphasizes the need for multiple evaluation metrics and highlights areas for improvement in TCR and TCR-pMHC structure prediction methodologies.