In this research highlight, we discuss key studies published in 2025 that advance our understanding of TCR biology, spanning mechanistic insights into activation, and emerging AI-based computational approaches able to model TCR-antigen recognition with translational potential. T cells are central to adaptive immunity, recognizing and eliminating infected or malignant cells through the T-cell receptor (TCR). TCRs are highly diverse heterodimers generated through V (D) J recombination, with additional variability introduced in the CDR3 regions. For the TCR αβ repertoire, this process produces around 108 unique receptors in each individual, 1 thus allowing recognition of peptide antigens presented by major histocompatibility complex (MHC) molecules. 2 Despite this diversity, general principles of antigen recognition are beginning to emerge, including conserved docking geometries between TCRs and peptide–MHC (pMHC) complexes. 3 Beyond structural complexity, TCRs function as mechanosensors, where productive activation depends on the formation of force-dependent catch bonds. 4 Antigen recognition is thus governed by an interplay of structural, biochemical, and mechanical constraints. However, a comprehensive understanding of TCR specificity remains elusive. Given the scale and complexity of the repertoire, identifying antigen specificity still relies largely on labor-intensive experimental approaches. Recent advances in artificial intelligence (AI) are beginning to transform this field, providing new tools to interrogate TCR-antigen recognition, and accelerate the development of targeted immunotherapies5 (Figure 1). In this research highlight, we discuss key studies published in 2025 that advance our understanding of TCR biology, spanning mechanistic insights into activation and emerging AI-based computational approaches with translational potential. Because TCR recognition is governed by dynamic three-dimensional interactions, structural modeling of TCR–pMHC complexes is essential for understanding specificity. Historically, this has been limited by the scarcity of experimental structures. However, recent advances in structure prediction algorithms like AlphaFold36 and RFDiffusion7 are overcoming this bottleneck, enabling systematic interrogation of antigen recognition at atomic resolution. These approaches are complementary: AlphaFold3 predicts the structure of unseen protein complexes from amino acid sequences, whereas RFDiffusion generates entirely novel protein backbones designed to adopt specific geometry or bind defined molecular surfaces. Together, these models expand the scope of structural immunology, shifting the field from predicting molecular interactions to actively designing them. A compelling demonstration of the potential of these AI-based approaches comes from three complementary studies, collectively addressing the de novo design of pMHC-targeting proteins. Using a generative design pipeline based on RFDiffusion, Liu et al. 8 engineered compact binding proteins that adopt TCR-like docking orientations, engaging pMHCs with a strong focus on peptide residue hot spots rather than the surrounding MHC surface. This peptide-centric mode of recognition enables high specificity while minimizing cross-reactivity, a key limitation of engineered high-affinity TCRs. 9 The designed binders achieved high affinity, as measured with surface plasmon resonance, and selectively recognized a broad range of targets across different HLA alleles. These included viral epitopes (SARS-CoV-2, Yellow Fever virus, HIV), tumor-associated antigens (MAGE-A3, MART-1, WT1) and neoantigens, discriminating between closely related peptides presented by the same HLA molecule. Importantly, when incorporated into chimeric antigen receptor (CAR) constructs and expressed in Jurkat cells, these binders mediated antigen-specific activation and cytotoxic responses, with minimal cross-reactivity against unintended peptide targets. Complementary studies by Householder et al. 10 and Johansen et al. 11 extended this strategy across diverse pMHC targets, including NY-ESO-1, establishing that de novo design of functional antigen binders using AI is now feasible across diverse systems. This translational progress should not be conflated with a deepened mechanistic understanding of antigen recognition. A fundamental limitation cuts across all three studies: they rely on black-box models optimized for structural plausibility that are not explicitly trained to capture functional avidity, binding affinity, or the downstream signaling events required for T-cell activation. This structure–function gap constrains these approaches to engineering rather than explaining the biological principles that govern TCR–pMHC recognition. One effort to address this limitation is HERMES, 12 a machine learning framework designed to link structural features of TCR–pMHC interactions with functional outcome. Instead of focusing solely on predicted structures, HERMES models the energetic landscape of the interface, learning how amino acid composition and context-dependent residue interactions influence binding affinity and T-cell activation. This approach was validated on well-characterized systems, including the 1G4 TCR binding to the NY-ESO-1 peptide and the A6 TCR specific for the HTLV-1 Tax viral peptide, as well as additional TCR targeting CMV, self and tumor-associated antigens. Predicted interaction energies correlated strongly with measured binding affinities (Kd) and functional responses (EC50), derived from dose response curves of 4-1BB+ CD8+ T cells. Furthermore, HERMES also enables de novo peptide design by systematically exploring mutations that preserve function while minimizing cross-reactivity, a key requirement for immunotherapy development. Together, these results demonstrate that structure-based models can begin to capture features of TCR–pMHC interactions directly linked to functional activation, representing a step toward more interpretable and mechanistically grounded AI-driven modeling. Despite their ability to link structural features to binding affinity and functional readouts, these approaches remain limited by their reliance on static representations of TCR–pMHC interactions, failing to capture the conformational flexibility known to be central to TCR recognition. This is particularly relevant given that binding affinity, while often used as a proxy for TCR functionality, is an incomplete predictor of T-cell activation, as TCRs operate as mechanosensory receptors. In this context, in a 2025 paper, Qin et al. 13 demonstrated that binding affinity alone is not sufficient to drive functional responses but represents only one component of a more complex process. Using the 2C TCR binding to the tumor-associated SIYRYYGL peptide (R4) or the weaker variant (L4) presented by the 2Kb MHC-I complex, they engineered higher affinity TCRs (m33, m67) by modifying CDR3α motifs and tested the relationship between binding affinity and T-cell reactivity from a mechanistic perspective. Functional assays measuring IL-2 production showed that increasing 3D affinity enhances sensitivity but reduces specificity: while the engineered TCRs (m33, m67) respond strongly to the cognate R4 peptide, they also become responsive to the weaker variant L4, unlike the native 2C TCR, thereby impairing antigen discrimination. Force-dependent analyses revealed that applied force induces additional stabilizing contacts at the TCR–pMHC interface, including interactions involving CDR3 regions, enabling high-affinity TCRs to form catch bonds with weak ligands. Steered molecular dynamic simulations further showed that TCR specificity depends on a defined force regime (~9–12 pN), in which force drives conformational rearrangements of the pMHC that promote CD8 recruitment and stabilization of the ternary complex, supporting productive catch bond formation and T-cell activation in the 2C TCR. Importantly, excessively strong interactions, as observed for m67, reduce the ability of the complex to undergo these force-induced rearrangements, limiting effective CD8 cooperation while still stabilizing interactions with non-cognate ligands, thereby promoting cross-reactivity. These findings highlight a fundamental constraint in TCR engineering and support emerging strategies that focus on tuning mechanosensitivity, rather than simply increasing binding affinity, to achieve functional specificity. 14 Together, these studies demonstrate that TCR-antigen recognition emerges from the interplay of structural, energetic and mechano-dynamical constraints. Accurately modeling this complexity therefore requires not only integrating these features but also access to high-quality functional data that capture their combined effects. Advances remain fundamentally limited by the scarcity of such data as generating functional datasets is labor-intensive, and clonal-level annotations remain rare. To address this gap, Messemaker et al. 15 developed a high-throughput system to generate and functionally test large libraries of TCRs in a controlled cellular context. Using a scalable approach to assemble full-length TCRαβ sequences, the authors combined pooled oligonucleotides with defined V gene segments, enabling accurate and cost-efficient construction of thousands of TCRs in parallel. The TCR libraries are then expressed in TCR-null CD8+ Jurkat T cells, and TCR reactivity is evaluated by co-culture with antigen-presenting cells and measurement of activation markers such as CD69, allowing identification of signaling-competent TCRs. This system provides a scalable framework to generate reliable functional data while enabling systematic identification of false-positive TCR-antigen pairs. The authors applied this method to test the functional reactivity of TCR–pMHC pairs reported in VDJdb. 16 Focusing on two well-characterized repertoires, the HLA-A*02: 01-restricted GLCTLVAML (GLC) epitope from EBV and the YLQPRTFLL (YLQ) epitope from SARS-CoV-2, they demonstrated that only around 40% of YLQ-specific and 70% of GLC-specific TCRs annotated in VDJdb elicited a functional response upon antigen presentation. Importantly, this has direct implication for computational models of TCR specificity and guiding immunotherapy design, as restricting analyses to experimentally validated functional TCRs significantly improves predictive performance. More broadly, this work demonstrates that generating high-quality functional data at scale is both feasible and necessary to build reliable, biologically grounded models of TCR recognition. These recent advances in AI-based approaches have driven substantial progress in modeling and engineering TCR-antigen interactions, with clear translational potential. Much of the current effort remains focused on advancing predictive and generative capabilities rather than addressing the fundamental principles that govern TCR specificity and activation. Realizing the full potential of these tools will require integrating structural, energetic and mechanical features with high-quality functional data to move beyond correlation and toward genuine biological understanding, providing a foundation for mechanistically grounded models of immune recognition. This work was supported by NHMRC Ideas grant APP2028765 (FL). MB acknowledges a UNSW international postgraduate scholarship. Martina Bonomi: Conceptualization; writing – original draft; writing – review and editing. Tom Donaldson: Writing – review and editing. Fabio Luciani: Conceptualization; writing – review and editing; supervision; resources. The authors declare no conflicts of interest.
Bonomi et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: