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

## TL;DR

This paper reviews AI methods for identifying T-cell antigens and finds that current models struggle with predicting new epitope variants, highlighting the need for better computational approaches.

## Contribution

The paper provides a comprehensive survey and benchmarking of AI-driven methods for T-cell antigen identification, revealing significant generalization gaps in current models.

## Key findings

- Current TCR–pMHC prediction models show a significant generalization gap when tested on out-of-distribution epitope variants.
- Enhanced structural modeling and integration of multi-omics data are urgently needed to improve prediction accuracy.
- Generative models for de novo TCR design could help overcome current limitations in antigen prediction.

## Abstract

The rise of mRNA vaccines highlights the pivotal role of T-cell antigen identification in modern vaccinology and personalized medicine. T-cell recognition relies on the sophisticated ternary interaction between the T-cell receptor (TCR), the major histocompatibility complex (MHC) molecule, and the peptide antigen, which forms the peptide–MHC (pMHC) complex. Computational methods, particularly artificial intelligence (AI), are indispensable for accurately predicting these complex bindings. This review systematically surveys the rapidly evolving AI-driven landscape for T-cell antigen identification, providing a comprehensive categorization of methods for MHC-I, MHC-II, and the highly complex TCR–pMHC binding prediction, alongside foundational data resources. Crucially, we conduct a rigorous, standardized benchmarking of 18 state-of-the-art TCR–pMHC prediction models across diverse training data sources. Our evaluation on two distinct and challenging out-of-distribution (OOD) unseen epitope variant datasets reveals a significant and concerning generalization gap in current predictors. Notably, the overall absolute predictive gain remains marginal across all models under OOD conditions. This result underscores a severe and persistent generalization challenge when faced with novel epitope variants. To address these limitations, we emphasize the urgent need for enhanced structural modeling, the integration of multi-omics data, and the development of generative models for de novo TCR design. By advancing these computational frontiers, our community can accelerate the transition from prediction to rational design in immunoinformatics.