A comparative and exploratory analysis of computational methods for TCR structural prediction and antigen-specific TCR discovery.

Reliable structural prediction of T-cell receptors (TCRs) is essential for dissecting antigen recognition and accelerating TCR-based therapeutic development, yet the performance of emerging computational structure prediction models for this task requires further systematic evaluation under practical usage conditions. Here, we assembled a manually curated dataset of TCR crystal structures and compared five state-of-the-art predictors-AlphaFold2 (v2.3.1), TCRmodel2, AlphaFold3, ESMFold, and tFold-TCR-in both single-chain and paired-chain modes. Using pLDDT and model confidence scores (pTM/ipTM), we defined optimized quality thresholds for assessing model reliability. Our analyses revealed a pronounced context dependence in model performance: AlphaFold3 achieved the highest accuracy for paired-chain TCR predictions (excluding the α-FRs domain), while tFold-TCR excelled in single-chain modeling (excluding Vα, CDR1α, CDR1β, and CDR2β domains), indicating that their performance varies significantly depending on the prediction mode and specific structural domains. We further showed that, under our evaluation conditions, structure-guided clustering of predicted CDR3β loops showed improved sensitivity and achieved higher conformational consistency for certain antigen-specific TCR groups compared with conventional sequence-based methods. Applying this framework, we successfully identified two functional SARS-CoV-2-specific TCRs using a TPS (TPSGTWLTY)-reactive TCR template. Our study establishes a practical comparative framework and highlights the translational potential of structure-guided computational workflows for antigen-specific TCR discovery.