Highlight of 2025: From binding prediction to molecular design—computational advances in TCR‐pMHC prediction and targeting

Abstract

In this article, we highlight recent progress in the field of TCR‐peptide–MHC prediction and outline remaining gaps in prediction power. We also touch on exciting advancements in molecular design that allow for precise targeting of peptide–MHC complexes.

T cell immunity hinges on a series of molecular recognition events: intracellular proteins or those captured from the extracellular environment are degraded into peptides and loaded into major histocompatibility complex (MHC) molecules [also known as human leukocyte antigens (HLA) in humans]. The resulting peptide–MHC (pMHC) complexes are displayed on the cell surface for detection by a T cell pool bearing diverse T cell antigen receptors (TCRs).

The system is tuned for generating highly specific and diverse responses to an individual pathogen or threat. Specificity and diversity are determined in part by our polymorphic HLA genes, where allelic variations of only a few bases greatly affect peptide antigen presentation. This means that the repertoire of HLA molecules and the peptide antigens they present varies substantially, even among consanguineous individuals that are not perfectly HLA‐matched. Additionally, the process of T cell development, which involves recombination of TCR gene segments and the addition of nucleotides in the CDR3 region (reviewed elsewhere ¹ , ² ) is geared toward generating a pool of T cells expressing diverse TCRs that are unique to a given individual.

The evolution of this exquisite antigen recognition system is presumed to allow humans and other vertebrates to discriminate self from nonself, and to eliminate pathogens with as little collateral damage as possible. However, the complexities of this system inherently make investigating specific TCR‐pMHC interactions difficult. Being able to predict peptide antigen immunogenicity in the context of specific MHC molecules would potentially transform vaccine design, cancer therapy, and our understanding of the etiology of complex inflammatory diseases where T cells are central to pathogenesis. ³ , ⁴ A central aim of immunoinformatics since the 1980s has been to identify patterns in TCR–pMHC interactions. ⁵ Over the past few years, deep learning has substantially improved pMHC binding prediction, with tools such as NetMHCpan ³ and MHCflurry ⁶ becoming standard components of epitope discovery pipelines, with continued progress on pan‐allele generalization. ⁷ Yet the field has long struggled with several interrelated challenges: predicting not just binding but functional immunogenicity, unifying the fragmented peptide–MHC (e.g. NetMHCpan, MHCflurry) and TCR–pMHC prediction tasks (e.g. TCRdist, NetTCR‐2.2, ERGO‐II, pMTnet), generalizing to unseen peptides and alleles, and most ambitiously, transitioning from passive prediction to active molecular design. Here, we highlight recent studies from 2025 that trace a trajectory from improved epitope prediction through unified frameworks and structure‐based modeling to de novo protein engineering.

DEEP LEARNING FOR EPITOPE PREDICTION BEYOND BINDING

A persistent limitation of HLA‐peptide binding predictors has been their poor correlation with downstream immunogenicity: many predicted binders never elicit a T cell response. ⁸ Wohlwend et al. addressed this with MUNIS, a deep learning model trained on a curated dataset of over 650 000 unique HLA class I ligands drawn from mass spectrometry immunopeptidomics and binding assays. ⁹ MUNIS outperformed established tools including NetMHCpan 4.1 and MHCflurry 2.0 for peptide presentation prediction across diverse alleles. Critically, MUNIS also predicted CD8⁺ T cell epitope immunodominance hierarchies, a task on which earlier tools have shown limited accuracy. The authors validated MUNIS experimentally by screening the Epstein–Barr virus proteome, identifying 12 novel immunogenic epitopes absent from the Immune Epitope Database. MUNIS achieved prediction accuracy comparable to in vitro HLA‐peptide stability assays, suggesting that deep learning can meaningfully reduce the experimental burden in epitope discovery for vaccine design.

UNIFYING THE TCR–pMHC PREDICTION AXIS

Most computational tools treat peptide–MHC binding and TCR specificity prediction as separate problems, an artificial fragmentation that ignores the biological reality that these interactions are coupled. Zhao et al. proposed UniPMT, a unified multitask framework that uses heterogeneous graph neural networks (GNNs) to simultaneously learn peptide–MHC (P–M), peptide–TCR (P–T) and peptide–MHC–TCR (P–M–T) binding relationships. ¹⁰ Peptides, MHC pseudosequences and TCR CDR3β sequences are represented as nodes in a graph, with their interactions as edges, and a GraphSAGE architecture learns robust node embeddings across all three tasks. UniPMT achieved state‐of‐the‐art performance, with up to 15% improvement in precision–recall area under the curve on the P–M–T prediction task compared to single‐task baselines. Importantly, the model correlated with T cell clonal expansion in experimentally validated neoantigen datasets, demonstrating translational relevance for personalized cancer immunotherapy. A notable limitation, acknowledged by the authors, is that UniPMT relies only on the TCR β‐chain CDR3, ignoring α‐chain and non‐CDR3 contributions now accessible through paired single‐cell datasets. Still, UniPMT exemplifies a broader shift toward integrative models that exploit the interdependence of immune recognition events.

STRUCTURE‐BASED PREDICTION AND DE NOVO PEPTIDE DESIGN

While the models above rely on sequence‐level features, Visani et al. took a fundamentally different approach by leveraging three‐dimensional protein structure. ¹¹ Their method uses HERMES, a structure‐based, SE(3)‐equivariant neural network trained on the general protein universe to predict amino acid preferences from local 3D structural environments. Remarkably, despite no direct training on TCR–pMHC data, HERMES accurately predicted both TCR–pMHC binding affinities (Spearman correlation up to 0.72 with experimental data) and functional T cell activities across diverse viral and cancer epitopes. This zero‐shot generalization highlights the power of learning general biophysical principles rather than memorizing specific interaction patterns. The authors also used HERMES for de novo design of immunogenic peptides, designing sequences differing by up to five residues from native peptides, which still proved immunogenic at rates approaching 50% across three therapeutically relevant TCR–pMHC systems (NY‐ESO‐1, EBV and MAGE). This work bridges the gap between prediction and design, demonstrating that structure‐based models can both forecast and engineer immune recognition.

DE NOVO DESIGN OF TCR‐MIMIC BINDER PROTEINS

One major development in 2025 was the simultaneous publication of three studies demonstrating de novo computational design of small, stable proteins that recognize specific pMHC class I complexes with high specificity and nanomolar affinity. ¹² , ¹³ , ¹⁴ All three groups used RFdiffusion, a generative diffusion model for protein backbone design, coupled with ProteinMPNN for sequence design and AlphaFold2 for validation. Liu et al. designed binders for 11 diverse pMHC‐I targets across four HLA alleles (A01:01, A02:01, A03:01 and C07:02), targeting viral, tumor‐associated and neoantigen peptides. ¹² Eight of these conferred peptide‐specific T cell activation when incorporated into chimeric antigen receptors (CARs). Johansen et al. designed minibinders targeting the cancer‐associated NY‐ESO‐1 pMHC, confirmed their structure by cryo‐electron microscopy and demonstrated that CAR‐T cells bearing these minibinders killed NY‐ESO‐1‐positive melanoma cells. Molecular dynamics simulation indicated that the selected binder was more rigid than alternatives, an argument the authors used to support reduced cross‐reactivity. ¹³ Householder et al. solved a 2.05 Å crystal structure of their α‐helical TCR mimic bound to NY‐ESO‐1/HLA‐A*02 (KD ≈ 9.5 nM), revealing a rigid, peptide‐focused docking mode distinct from natural TCRs or antibodies. ¹⁴ Together, these papers demonstrate that generative AI can produce binders with therapeutic‐grade specificity and affinity, potentially bypassing the laborious identification of natural TCRs from patient material and enabling off‐the‐shelf, computationally designed immunotherapeutics.

THE GENERALIZATION CHALLENGE

These advances are exciting, but the field still has lingering limitations. A preprint by Culka et al. provides a sobering benchmark: using a proprietary cancer‐patient dataset profiling TCR binding to novel peptides, they showed that current machine learning models, including state‐of‐the‐art deep learning approaches, perform well on epitopes seen during training but fail to generalize to truly unseen peptides. ¹⁵ This is not a minor caveat since for clinical applications such as personalized neoantigen‐based immunotherapy, the ability to predict TCR binding to novel tumor‐specific peptides is what is needed. The same generalization gap is arguably even more acute in chronic inflammatory and autoimmune diseases, including psoriasis, atopic dermatitis and inflammatory bowel disease, where T cells clearly drive pathology, but where the cognate antigen identity remains largely elusive. Robust TCR–pMHC prediction would be transformative in these settings. Interestingly, the Culka et al. study also found that structure‐based methods using classical energy functions outperformed ML approaches on unseen peptides, while underperforming on known ones. This suggests that hybrid strategies, combining data‐driven learning with biophysical priors, as HERMES begins to do, may represent the most promising path forward for robust generalization.

OUTLOOK

Computational immunology in 2025 is moving from retrospective classification toward prospective molecular design. A limitation cutting across all the articles highlighted is their near‐exclusive focus on CD8⁺ T cells and HLA class I. MHC class II prediction remains harder, but recent work shows the problem is tractable, ¹⁶ and bringing CD4⁺ TCR–pMHC‐II modeling to parity with the CD8⁺ side is arguably the next frontier. A second cross‐cutting limitation is that most current models operate on static structures: incorporating molecular dynamics will be important, since off‐target recognition by TCR mimics and minibinder‐CARs is a real safety concern. The opportunities for early‐career researchers are considerable and include (1) closing the generalization gap, (2) integrating MHC‐II and CD4⁺ T cell biology and (3) applying de novo design to engineer TCR‐mimic therapeutics against patient‐specific neoantigens. ¹⁷ As these approaches converge, the design of vaccines and T cell therapies should become increasingly rational and precise.

AUTHOR CONTRIBUTIONS

Kilian J. Maire: Conceptualization; investigation; writing – original draft; writing – review and editing. Jonathan M. Coquet: Writing – review and editing.

CONFLICT OF INTEREST

The authors declare no conflicts of interest.