AI-driven computational methods and benchmarking for T-cell antigen identification.

Area of Science:

• Immunoinformatics
• Computational vaccinology
• Personalized medicine

Background:

• T-cell recognition involves a complex interaction between T-cell receptors (TCRs), major histocompatibility complex (MHC) molecules, and peptide antigens (forming peptide-MHC or pMHC complexes).
• Accurate prediction of these interactions is vital for developing effective vaccines and personalized therapies.
• Artificial intelligence (AI) offers powerful tools for predicting these complex molecular bindings.

Purpose of the Study:

• To systematically review and categorize AI-driven methods for T-cell antigen identification, focusing on MHC-I, MHC-II, and TCR-pMHC binding prediction.
• To benchmark the performance of state-of-the-art TCR-pMHC prediction models.
• To identify current limitations and suggest future research directions in computational immunoinformatics.

Main Methods:

• Comprehensive survey of AI-based T-cell antigen identification techniques.
• Standardized benchmarking of 18 leading TCR-pMHC prediction models using diverse training data.
• Evaluation of model performance on out-of-distribution (OOD) datasets with unseen epitope variants.

Main Results:

• A significant generalization gap was observed in current AI predictors when evaluated on OOD unseen epitope variants.
• The overall predictive gain across all benchmarked models was marginal under OOD conditions.
• Current models exhibit a severe and persistent challenge in generalizing to novel epitope variants.

Conclusions:

• Urgent need for improved structural modeling and integration of multi-omics data in AI prediction models.
• Development of generative models for de novo T-cell receptor (TCR) design is crucial.
• Advancing computational methods will accelerate the shift from prediction to rational design in immunoinformatics, enhancing vaccine development and personalized medicine.