T-cell-receptor biology and TCR-engineered cell therapy directed at tumor neoantigens — the readout side of antigen recognition.
T cells are a key part of the adaptive immune system. Using their surface-bound T cell antigen receptors (TCRs), these cells scan peptides and other antigens presented to them by major histocompatibility complex molecules (MHCs) on the surface of cells, searching for abnormalitie…
Primary tumors and their sentinel lymph nodes are functionally linked sites of anti-tumor immunity, yet how T cell receptor (TCR) repertoires are organized across these compartments remains incompletely understood. We thus performed TCR{beta} sequencing on paired primary melanoma…
IgGM2 is a new all-atom foundation model for adaptive immune receptor design that jointly generates CDR residue identities and full-atom structures. Unlike modular pipelines, it uses a structure-to-design strategy to allow framework geometry to adapt to designed CDRs without separate inverse folding. This unified approach aims to improve accuracy in modeling the coupled variation of sequence, conformation, and binding geometry for antibodies and TCRs.
Researchers benchmarked AlphaFold2, AlphaFold3, and related deep learning methods for modeling antibody-protein, antibody-peptide, and TCR-pMHC recognition. Results indicate that increased sampling and AlphaFold3 generally outperform default AlphaFold2 settings, though predictive accuracy varies significantly by interface class. The authors note that antibody-peptide complexes remain particularly difficult to model accurately and propose model pooling as a potential solution to leverage method complementarity.
Researchers have introduced TCR-FramePose, a local-frame representation method that decomposes TCR-pMHC docking geometry into reach, offset, and orientation coordinates. Applied to 378 crystal structures, the tool captures CDR3 loop geometry and affinity-associated information beyond conventional descriptors, potentially improving the computational prediction of T cell receptor recognition.
PRISM is a metric-learning framework that embeds TCR beta sequences into a peptide-organized latent space to predict pMHC ligands. By using structure-guided synthetic receptors to counter viral bias in training data, it abstains on out-of-distribution inputs via uncertainty modeling. Benchmarks show PRISM matches or exceeds existing models, particularly for rare epitopes.
T cell receptor (TCR)-based immunotherapy holds immense potential for treating cancers, autoimmunity, and infectious diseases, where antigen-specific TCR recognition is crucial for adaptive immune responses. Engineering or de novo generation of the complementarity-determining reg…
Researchers identified neogenes (Ew_NGs) induced by the EWSR1::FLI1 fusion in Ewing sarcoma, which present peptides on HLA-I complexes. CD8+ T cells specific to these public antigens killed EwS cells in an HLA-I restricted manner, a process dependent on both the fusion and the neogene expression. Transduced TCR-T cells replicated this cytotoxicity in vivo without off-target or allogeneic activation, suggesting a viable cell therapy for relapsed/resistant cases.
The replicateFest R package and Shiny app provide a framework for analyzing T cell receptor repertoire data from Functional Expansion of Specific T (FEST) assays, specifically addressing variability from biological and technical replicates. By applying negative binomial modeling to replicate experiments, it identifies clonotypes significantly expanded in antigen-stimulated conditions with adjusted p-values. This tool aims to improve reproducibility in detecting neoantigen-specific T cell responses, a critical step for guiding vaccine development and assessing checkpoint blockade efficacy.
Alaunos Therapeutics has terminated its Phase I/II study of autologous T cells engineered via the Sleeping Beauty system to express neoantigen-reactive TCRs for solid tumors.
Researchers present Loop-Diffusion, an energy-based diffusion model for designing and scoring protein loops. The model demonstrates state-of-the-art performance in recognizing binding-enhancing mutations at TCR-pMHC interfaces, addressing limitations in current ML methods for functional prediction.
The paper introduces SCEPTR, a T cell receptor language model using contrastive learning and masked-language modeling. It achieves state-of-the-art performance in TCR specificity prediction by leveraging data-efficient transfer learning, outperforming existing protein language models.
A new study in Nature elucidates the structural basis for self-discrimination by neoantigen-specific T-cell receptors (TCRs). Understanding these mechanisms is critical for ensuring that neoantigen vaccines do not trigger autoimmune responses against healthy tissue while effectively targeting tumor mutations.
Beijing DCTY Biotech is initiating a Phase 1/2 study of DCTY1102, an autologous TCR T-cell therapy targeting KRAS/NRAS G12D mutations in HLA-A11:01 positive patients. The trial will employ a 3+3 dose-escalation design to determine the maximum tolerated dose and recommended Phase 2 dose, evaluating safety, pharmacokinetics, and preliminary efficacy in advanced malignant tumors.
The National Cancer Institute is recruiting for a Phase 1 study of neoepitope-specific T cells for patients with nine types of blood cancers, including those with TP53 or RAS mutations. The protocol involves autologous T-cell isolation and expansion to target patient-specific neoepitopes, addressing the challenge of treating hard-to-treat hematologic malignancies.
Providence Health & Services is managing an intermediate-size expanded access program for T-cells engineered with TCRs targeting the KRAS-G12D neoantigen in advanced pancreatic and colorectal cancers. The protocol involves autologous T-cell transduction with GMP-grade retroviral vectors followed by lymphodepletion, offering a potential treatment avenue for patients with limited options harboring this specific mutation.
Researchers present CaliPPer, a post-hoc framework that quantifies and improves AI model performance for binding prediction using distance-aware Bayesian recalibration. Applied to TCR and BCR models, CaliPPer significantly reduces mean absolute errors in AUROC/AP/F1 predictions and increases true discovery rates on unseen epitopes by providing per-sample confidence scores.
A new training framework called Counterfactual Invariant Prediction (CIP) addresses shortcut learning in TCR-pMHC binding models by enforcing invariance to non-anchor peptide edits. Evaluated on VDJdb-IEDB benchmarks, CIP reduces the shortcut index by 39.7% compared to baselines, improving model robustness under family-held-out and distance-aware evaluation protocols.
This paper introduces calibrated abstention for TCR-pMHC binding prediction, allowing models to output confidence scores or explicitly abstain when encountering unseen epitopes. Using a dual-encoder architecture and conformal abstention rules, the method reduces error rates by nearly 40% at 80% coverage, offering a principled approach to managing uncertainty in neoantigen discovery.
ImmSET, a sequence-based transformer architecture for TCR-pMHC specificity prediction, is introduced to model interactions among variable-length biological sequences. The authors identify and correct a failure mode in prior approaches that inflated performance metrics, demonstrating that ImmSET's generalization scales consistently with training data under stricter evaluation protocols.
Researchers propose TCR-EML, an explainable model layer architecture for TCR-pMHC prediction that incorporates known biochemical binding mechanisms. This 'explain-by-design' approach aims to improve upon black-box transformer models by providing interpretable rationales for predictions while maintaining competitive accuracy on large-scale datasets.
A new framework for rational multi-modal transformers uses post-hoc explainability to optimize encoder-decoder architectures for TCR-pMHC prediction. By identifying informative sequence inputs and introducing explanation-based early stopping, the method claims state-of-the-art performance alongside improved robustness and mechanistic insight into binding behavior.
The LANTERN framework combines large protein language models (ESM-1b) with chemical representations (MolFormer) to enhance TCR-peptide interaction prediction. It demonstrates superior performance in zero-shot and few-shot scenarios compared to existing models like ChemBERTa, leveraging robust negative sampling to address data scarcity in personalized vaccine development.
Researching the specificity of TCR contributes to the development of immunotherapy and provides new opportunities and strategies for personalized cancer immunotherapy. Therefore, we established a TCR generative specificity detection framework consisting of an antigen selector and…
Researchers propose Quantifying Cross-Attention Interaction (QCAI), a post-hoc method to interpret cross-attention mechanisms in encoder-decoder transformers used for TCR-pMHC binding. The work introduces the TCR-XAI benchmark of 274 experimentally determined structures to address the lack of interpretability in current black-box models.
Researchers developed an integrated in silico-in vitro workflow to generate high-affinity, selective antibodies targeting the KRAS G12D neoantigen presented by HLA-C*08:02. The method combines in silico docking and CDR design with yeast surface-display selection, yielding antibodies with no off-target reactivity that can be reformatted as CARs or bispecific T-cell engagers for therapeutic use.
Research reveals that HLA micropolymorphisms confine neoantigen conformational adaptability, directly guiding T cell receptor selectivity. Specifically, micropolymorphisms in HLA-A*03:02 versus A*03:01 prevent TCR binding by altering the neoantigen's conformational ensemble rather than peptide binding, highlighting a critical mechanistic constraint for vaccine design that must be accounted for in antigen selection pipelines.
A Science publication reports that healthy donor T cell receptors can expand functional neoantigen recognition beyond patient vaccination. This finding suggests potential for off-the-shelf TCR therapies or broader immune monitoring strategies that do not rely solely on patient-specific vaccine responses.
Researchers introduced HERMES, a structure-based, physics-guided machine learning model that predicts TCR interactions with peptides on MHC class I alleles. Trained on the protein universe without direct TCR-pMHC data, HERMES achieved up to 0.72 correlation with experimental data and enables de novo design of immunogenic peptides.
A new attention-aware contrastive learning model, ATMTCR, improves the prediction of TCR-antigen binding specificity. By using Transformer encoders and masked contrastive views on large-scale TCR CDR3 sequences, the model extracts high-order semantic information to outperform existing algorithms in predicting neoantigen reactivity.
Researchers introduced DapPep, a domain-adaptive peptide-agnostic learning framework for universal TCR-antigen binding affinity prediction. Using a lightweight self-attention architecture combined with protein language models, DapPep outperforms existing tools in predicting binding for unseen peptides, addressing a key bottleneck in neoantigen vaccine design for data-scarce settings.
Researchers introduce tcrLM, a lightweight masked language model for predicting TCR-antigen binding specificity. Pretrained on over 100 million TCR CDR3 sequences with virtual adversarial training, tcrLM outperforms existing methods in predicting binding, offering a computationally efficient tool for neoantigen vaccine design.
This study analyzes the contact map dependence of TCR binding repertoires using an affinity-based model. By weighting pairwise amino acid interactions via crystal structure-derived contact maps, the work provides insights into TCR-pMHC binding energy distributions and their influence on T cell recognition probability during negative selection.
PhysicoGPTCR, a dual-conditioned generative protein Transformer, is introduced for designing TCR variable regions. Trained on TCR-peptide-HLA triples with physicochemical descriptors, the model improves binding-competent clone generation and sequence space exploration compared to baselines like GPTCR and VAEs, advancing computer-aided cellular therapy.