General prediction of T cell receptor antigen specificity from sequence using AlphaFold 3

Technical Summary: General Prediction of T Cell Receptor Antigen Specificity from Sequence Using AlphaFold 3

Problem Statement
The Major Histocompatibility Complex (MHC):peptide:T Cell Receptor (TCR) complex represents the most diverse trimolecular interface in nature and serves as the fundamental trigger for adaptive immunity. While high-throughput sequencing has made TCR repertoire data widely available, determining the specific antigen targets (epitopes) for these receptors remains a significant bottleneck. Existing in silico approaches have advanced but suffer from a critical limitation: they generally fail to predict specificity against "unseen" epitopes. This inability to generalize restricts current methods to a very small fraction of practical cases, hindering the broad decoding of immune responses.

Methodology
This study leverages AlphaFold 3 (AF3) to address the structural prediction of MHC:peptide:TCR triads. The authors applied AF3 to a large-scale dataset comprising over 9,000 TCRs mapped to more than 1,000 distinct epitopes, restricted by over 70 different MHC class I and II alleles.

The core methodology involves two primary steps:

1. Structural Prediction: Utilizing AF3 to generate high-fidelity 3D structures of the MHC:peptide:TCR complexes.
2. Feature Extraction and Classification: Analyzing the predicted structures to identify specific structural features that distinguish cognate (true) triads from non-cognate controls. These features were then used to train a predictive model capable of generalizing to new data.

Key Contributions

• Unprecedented Structural Accuracy: The paper demonstrates that AlphaFold 3 can predict the structures of MHC:peptide:TCR triads with a level of accuracy previously unattained for such diverse interfaces.
• Generalization to Unseen Epitopes: Unlike previous methods, the resulting model successfully predicts specificity against epitopes that were not seen during the AF3 structure generation or the feature selection process.
• Scalable Framework: The approach establishes a workflow that moves beyond limited, case-specific predictions to a generalized prediction system based on sequence data.

Results
The model was evaluated on validation triads that were entirely unseen by both the AF3 structure prediction step and the feature selection process. The results indicate robust performance, achieving median Area Under the Curve (AUC) values ranging from 0.81 to 0.92. These metrics suggest that the structural features derived from AF3 predictions are highly effective at distinguishing true antigen-specific interactions from background noise, even in the absence of prior exposure to the specific epitopes.

Significance
The paper claims that these findings demonstrate the feasibility of generalizing TCR specificity prediction directly from sequence data. By overcoming the "unseen epitope" barrier, this work offers the potential to greatly accelerate the decoding of immune responses. The ability to predict specificity for novel antigens using structure-based features represents a shift from limited applicability to a broader, more versatile tool for understanding adaptive immunity.