The researchers focused on the 4T1 model, a highly aggressive mouse breast cancer cell line that is widely used to study triple-negative breast cancer, the subtype of the disease that lacks the three most common therapeutic targets and that has historically been difficult to treat with immunotherapy. Although checkpoint blockade has improved outcomes for some patients with triple-negative breast cancer when integrated into neoadjuvant treatment, the 4T1 model has been an underused resource for dissecting resistance mechanisms because, until now, no tumor-specific epitopes had been defined within it. Without knowing what peptides the immune system can recognize, experiments designed to test immunotherapy combinations in this model have lacked a molecular anchor.

To fill that gap, the team turned to immunopeptidomics, the mass spectrometry-based analysis of the repertoire of peptides presented on the cell surface by major histocompatibility complex molecules. This technique allows researchers to catalog, in an unbiased way, every fragment of protein that a cell displays to patrolling T cells. When the immunopeptidomes of different 4T1 subclones were profiled, a striking pattern emerged. The parental, less metastatic cells presented a broad array of peptides, whereas subclones selected for enhanced metastatic behavior showed a marked contraction of their presented peptide repertoire, consistent with a downregulation of the antigen processing and presentation machinery.

This loss of antigen presentation is a classic strategy of immune evasion, and the study went further by showing that the defect is not irreversible. When the metastatic subclones were treated with type I interferon, a signaling molecule central to antiviral and antitumor immunity, the machinery of antigen processing and presentation was restored, and the cells once again displayed a richer set of peptides on their surface. The finding suggests that metastatic cells can silence their own visibility to the immune system through a pathway that interferon signaling can counteract, a mechanistic insight with direct implications for how interferon-based approaches might be combined with checkpoint inhibitors.

The most consequential discovery, however, lay in the identity of the peptides themselves. Rather than being dominated by fragments of conventional genes mutated or overexpressed in cancer, the immune responses mounted against metastatic 4T1 cells were directed largely at cryptic peptides derived from endogenous retroviral products, particularly those encoded by murine leukemia virus, or MLV. These are sequences that most tumor immunology pipelines would discard as non-coding or repetitive noise, yet in this model they constitute the immunodominant landscape, meaning the targets that the T cell response preferentially recognizes and attacks.

Interferon treatment did more than simply restore presentation; it reshaped the retroviral target landscape itself. Exposure to type I interferon increased both the abundance and the diversity of MLV-derived peptides in the 4T1 immunopeptidome, effectively broadening the range of viral remnant fragments that tumor cells displayed. This interferon-driven amplification of retroviral antigen display provides a mechanistic bridge between innate antiviral signaling and adaptive tumor recognition, and it helps explain why interferon signaling has often been associated with better responses to immunotherapy in both mouse models and human cancers.

Crucially, the functional experiments demonstrated that these retroviral targets are not merely decorative. MLV-specific T cell responses proved to be critical for the effectiveness of immune checkpoint blockade in the model. When the researchers examined tumor growth and metastasis-free survival, the data showed that T cells recognizing the endogenous retroviral peptides were essential for suppressing tumor outgrowth and prolonging the period during which animals remained free of metastatic disease. In other words, the anti-metastatic immunity that determines outcome in this aggressive breast cancer model is, to a substantial degree, an immune response against the tumor’s own ancient viral passengers.

The implications extend beyond the mouse. Human genomes carry their own burden of endogenous retroviral elements, and human tumors frequently express retroviral proteins that are absent from healthy adult tissues. If the same logic applies in patients, then the retroviral peptidome of human tumors could represent a shared, immunodominant and potentially druggable antigen class, one that does not depend on the private mutations that make personalized neoantigen vaccines so logistically complex. The identification of defined, immunodominant ERV epitopes in the 4T1 model now gives researchers a concrete platform on which to build ERV-targeted immunotherapies, including peptide vaccines and T cell-based approaches, and to test them in a system where the relevant antigens are finally known.

The study also reframes the role of interferon in cancer immunotherapy. Rather than acting only as a general inflammatory amplifier, type I interferon emerges here as a specific regulator of the retroviral antigen repertoire, expanding the visible surface of the tumor to the immune system. The authors suggest that this provides a foundation for translating interferon and ERV-targeting strategies into future clinical trial design, potentially pairing agents that derepress endogenous retroviral expression or enhance interferon signaling with checkpoint blockade, so that the immune system is presented with a richer set of retroviral targets precisely when therapeutic antibodies release the brakes on T cells.

For a field that has struggled to explain why some triple-negative breast cancers respond to immunotherapy while others do not, the message of this work is that the answer may lie partly in the viral archaeology of the genome. Metastatic cells that hide their antigens escape; cells forced by interferon to display their endogenous retroviral fragments become visible and vulnerable. By naming the specific MLV-derived epitopes that dominate the anti-metastatic response in 4T1 tumors, the researchers have converted a widely used but immunologically opaque model into a defined testing ground for retrovirus-directed cancer vaccines, and they have strengthened the case that the ancient viral DNA within us is not silent baggage but an active participant in the fight against cancer.

Subject of Research: Identification of immunodominant endogenous retroviral peptides that drive anti-metastatic T cell immunity and checkpoint blockade response in a triple-negative breast cancer mouse model.

Article Title: Identification of immunodominant ERV peptides critical for anti-metastatic immunity in a model of TNBC

Article References: Faridi, P., Zanker, D., Haynes, N. M., Nickson, J., Spurling, A., Panetta, F., Huang, P., Sung, N., Kang, J.-H., Dolcetti, R., Rajapaksha, H., Robinson, B. W. S., Dick, I., Redwood, A., Creaney, J., Chen, W., Anderson, R. L., Caminschi, I., Purcell, A. W., & Parker, B. S. (2026). Identification of immunodominant ERV peptides critical for anti-metastatic immunity in a model of TNBC. Cancer Immunology, Immunotherapy. https://doi.org/10.1007/s00262-026-04554-1

Image Credits: AI Generated

DOI: 10.1007/s00262-026-04554-1

Keywords: endogenous retrovirus, ERV peptides, triple-negative breast cancer, 4T1 model, immunopeptidomics, murine leukemia virus, type I interferon, immune checkpoint blockade, antigen presentation, metastasis, cancer immunotherapy, tumor immunology

The biological logic behind neoantigen vaccines is compelling. Neoantigens arise from somatic mutations that exist only in tumor cells, which means the immune system has not been trained to tolerate them, and the risk of off-target autoimmunity is theoretically low. Advances in next-generation sequencing, HLA typing, immunopeptidomics and synthetic vaccine platforms have made it feasible to sequence a patient’s tumor, identify candidate mutation-derived antigens and manufacture an individualized vaccine in a matter of weeks. Early clinical studies have now established feasibility, safety and immunogenicity across a striking range of solid tumors, including melanoma, pancreatic ductal adenocarcinoma, hepatocellular carcinoma, non-small cell lung cancer, bladder cancer and renal cell carcinoma, using platforms that span mRNA, peptide, DNA, viral vector and dendritic cell technologies.

The clinical momentum is real. In the phase 2b KEYNOTE-942 trial, the individualized mRNA vaccine mRNA-4157, known as V940, combined with pembrolizumab achieved 18-month recurrence-free survival of 79 percent versus 62 percent with pembrolizumab alone in resected melanoma, a hazard ratio of 0.561, with no grade 4 or 5 adverse events attributed to the vaccine. In pancreatic cancer, one of the hardest immunologic targets in oncology, responders to the personalized RNA vaccine autogene cevumeran showed markedly prolonged recurrence-free survival compared with non-responders, with a median not yet reached versus 13.4 months. A phase 3 trial in resected high-risk melanoma, INTerpath-001, is now enrolling roughly 1,089 patients, a sign that the field is moving from proof of concept toward pivotal testing.

Yet beneath these headline numbers lies an uncomfortable attrition problem that the review dissects stage by stage. Only a small fraction of the mutations identified in a tumor ever become genuinely immunogenic vaccine targets. In the autogene cevumeran pancreatic cancer trial, just 11 percent of the administered vaccine neoantigens, 25 out of 230, elicited measurable T-cell responses. In the NeoVax melanoma study, roughly 60 percent of vaccine neoantigens induced CD4-positive T-cell responses but only about 16 percent triggered CD8-positive cytotoxic T cells. A recent systematic evaluation found that only around 22 percent of predicted neoepitopes are actually presented on HLA molecules. Each false positive dilutes the effective immunogenic content of a vaccine that has limited room for passengers.

The root of the problem is that antigen presentation is a multi-step biological process, involving proteasomal cleavage, TAP-mediated transport and peptide-HLA loading, that most computational pipelines compress into a single binding-affinity prediction. Even when a peptide is presented, whether it provokes a functional T-cell response depends on the three-dimensional docking geometry between the peptide-HLA complex and a cognate T-cell receptor, governed by molecular flexibility and peptide conformation that sequence-based models capture poorly. Compounding matters, the scarcity of paired TCR-peptide-MHC datasets and the staggering diversity of the TCR repertoire make generalization across patients extraordinarily difficult. Manufacturing timelines, typically seven to eight weeks for peptide and viral vector platforms, add a further constraint, since tumors can evolve while the vaccine is being built.

This is where the review sees artificial intelligence earning its place, not as a replacement for experiments but as a way to narrow enormous biological and chemical search spaces. At the discovery stage, machine learning frameworks now integrate genomic, transcriptomic and mass spectrometry-based proteomic data to prioritize candidates with multi-layer evidence of expression and presentation. Tools such as ProGeo-neo combine proteomics with patient-specific genomic profiles to confirm protein-level expression, while ImmuneMirror merges whole-exome and RNA sequencing data, and modular workflows like TIminer and pVACtools automate reproducible end-to-end immunogenomic profiling. The authors also highlight the growing importance of noncanonical antigens, derived from alternative open reading frames, aberrant splicing, intron retention, fusion transcripts, circular RNA products and endogenous retroviral elements, some of which may be shared across patients and could inform semi-shared vaccine strategies.

Presentation prediction is where measurable gains have been most convincing. Pan-allelic tools such as NetMHCpan integrate eluted ligand data with binding affinities, and MHCflurry 2.0 folds antigen-processing features into neural network models, achieving an area under the curve of roughly 0.911 for binding classification in a benchmark of 18 predictors across 32 HLA alleles. Newer architectures such as CapHLA jointly model class I and class II presentation, and protein language model-based approaches like MUNIS improve prioritization of presented epitopes. The review also makes a pointed case for CD4-positive T cells, which clinical studies have repeatedly activated, sometimes more frequently than CD8 responses, arguing that HLA class II prediction, despite its computational challenges from open binding grooves and variable peptide lengths, deserves far more attention in vaccine design.

At the T-cell recognition frontier, deep learning is pushing into structure. AlphaFold-based modeling of peptide-HLA and TCR-peptide-HLA complexes can discriminate true target epitopes from decoys, while tools such as NetTCR-2.0, DeepTCR, TRAP and SageTCR integrate paired TCR sequences and three-dimensional structural features to predict receptor binding. Benchmarking studies, however, consistently show that these models remain strongly data-dependent and often fail to generalize to unseen epitopes, so the authors position them as prioritization and hypothesis-generating tools rather than definitive immunogenicity predictors. Downstream, AI is also reshaping formulation science: the AGILE platform uses deep learning-guided screening to identify ionizable lipids for mRNA delivery, and the LiON framework applies message-passing neural networks to large lipid nanoparticle datasets, narrowing chemical design spaces before any wet-lab testing begins.

To tie these threads together, the authors propose an evidence-weighted, closed-loop decision-gate framework spanning six stages: candidate generation, tumor relevance filtering, HLA presentation evidence, immunogenicity refinement, portfolio-level selection, and manufacturability and clinical delivery, with post-vaccination immune monitoring data fed back to refine future models. They stress that repertoire size should reflect biologically meaningful coverage rather than numerical abundance, since no clinically validated optimal neoantigen count exists. On the regulatory front, the UK’s Medicines and Healthcare products Regulatory Agency has released draft guidance for individualized mRNA cancer immunotherapies, in which AI selection algorithms may qualify as Software as a Medical Device, requiring Good Machine Learning Practice, transparent pipelines, version tracking and barcode-based traceability linking genomic data to the administered dose. The review’s bottom line is measured but optimistic: AI will not conjure effective cancer vaccines on its own, but embedded within rigorous experimental validation, scalable manufacturing and evolving regulatory oversight, it may finally make personalized neoantigen vaccination a reproducible pillar of precision oncology.

Subject of Research: The application of artificial intelligence across the translational development pipeline of personalized neoantigen cancer vaccines, including neoantigen discovery, HLA presentation prediction, T-cell recognition modeling and formulation optimization.

Article Title: Artificial intelligence for translational personalized neoantigen cancer vaccine development

Article References: Artificial intelligence for translational personalized neoantigen cancer vaccine development. (n.d.). https://doi.org/10.1186/s12929-026-01286-3

Image Credits: AI Generated

DOI: 10.1186/s12929-026-01286-3

Keywords: neoantigen cancer vaccine, artificial intelligence, personalized immunotherapy, HLA presentation, T-cell receptor, lipid nanoparticle, mRNA vaccine, precision oncology, immunopeptidomics, machine learning, cancer immunotherapy, translational oncology

Traditional antigen discovery has relied heavily on bulk omics. RNA sequencing of tumour biopsies identifies transcripts that are elevated in tumours relative to normal tissues, and mass spectrometry of immunopeptidomes identifies the peptides actually presented on human leukocyte antigen molecules. These approaches have been enormously productive, yielding the tumour-associated antigens that underpin therapeutic cancer vaccines, bispecific antibodies, and adoptive cell therapies. Yet they share a structural blind spot: they measure abundance, not arrangement. A transcript that is highly expressed in a tumour may also be expressed in a vital healthy tissue, and a peptide that appears abundant in a dissociated tumour sample may, in the intact tissue, be presented only on stromal cells rather than on the malignant compartment itself. Dissociation destroys the spatial relationships that the immune system, operating in intact tissue, would encounter.

Proximity labelling technologies emerged precisely to address this limitation. By fusing an engineered enzyme, such as a promiscuous biotin ligase or a peroxidase, to a protein of interest, researchers can covalently tag molecules that come within a few tens of nanometres of that protein in living cells. The tagged molecules can then be enriched and identified by mass spectrometry, producing a snapshot of the local molecular neighbourhood rather than the global composition of the cell. Applied to tumour biology, proximity labelling offers a way to ask what molecules physically congregate around a given marker protein, information that bulk profiling cannot provide. The microenvironment of a membrane protein, its partners, its neighbours in the plasma membrane, and the proteins trafficked alongside it, constitutes a layer of biological organisation that is invisible to standard transcriptomics and only partially accessible to conventional proteomics.

The challenge that arises once proximity data are generated is computational. A proximity labelling experiment produces a network-like structure: bait proteins, their tagged neighbours, the abundances of those neighbours, and the connections among them across conditions and cell states. Representing this as a simple list discards most of its meaning. Graph-structured data demand graph-aware analysis, and this is where modern graph learning methods become relevant. Graph neural networks and related architectures are designed to learn representations of nodes in a network that incorporate information from their local neighbourhoods, so that the identity of a protein is encoded not only by its own properties but by the company it keeps. In biological settings, this inductive bias is often exactly right: function in cellular systems is relational, and proteins that occupy similar network positions frequently share functional roles even when their sequences are unrelated.

Applying graph learning to proximity data in the tumour context creates an opportunity to formalise an intuition that immunologists have long held informally. An antigen that is safe to target therapeutically is not simply one that is absent from healthy tissue in a bulk measurement; it is one whose presentation is unlikely to occur on healthy cells under the conditions the immune system will actually encounter. Proximity provides a proxy for this contextual specificity. A candidate antigen that is consistently found in the immediate molecular neighbourhood of validated tumour markers, and that participates in the same spatial programmes as known malignant-surface proteins, carries more evidence of tumour association than a candidate identified by expression alone. Graph learning allows this evidence to be aggregated systematically across many candidates and many data modalities, rather than assessed one protein at a time by expert curation.

The therapeutic stakes of this problem are considerable. CAR T cell therapy, which engineers a patient’s T cells to recognise a surface antigen, has produced remarkable outcomes in haematological malignancies where a truly tumour-restricted target such as CD19 exists. Solid tumours have proven far more resistant, and a major reason is the absence of comparable target antigens. The antigens that are available on solid tumours, such as HER2, EGFR, or B7-H3, are frequently shared with essential healthy tissues, and targeting them produces on-target off-tumour toxicity that can be dose-limiting or fatal. The field has responded with a variety of engineering strategies, including logic-gated CARs that require two antigens to be present simultaneously, tunable affinity receptors that respond only to high antigen density, and adaptor systems that allow dosing control. All of these strategies depend on knowing which combinations of antigens are jointly specific to tumours, and this is a question about spatial co-organisation that proximity-guided approaches are well positioned to answer.

There is also a deeper immunological rationale for thinking in terms of proximity. The immune synapse itself is a proximity phenomenon: a T cell commits to killing only after sustained engagement with a target cell, integrating signals from many receptor-ligand interactions across the contact interface. Antigens that cluster together on the tumour surface may be recognised more effectively than antigens presented in isolation, because multivalent engagement strengthens T cell receptor signalling and can overcome the inhibitory signals that tumours deploy. Conversely, an antigen that is spatially segregated from co-stimulatory context may be immunologically silent even if it is abundant. Understanding the spatial organisation of tumour antigens therefore has implications not only for target selection but for predicting the quality of the immune response that targeting will provoke.

The tumour microenvironment adds further layers of complexity that proximity-aware methods are suited to capture. Tumours are not homogeneous masses of malignant cells; they are ecosystems containing fibroblasts, endothelial cells, macrophages, T cells, and extracellular matrix, all arranged in structured architectures that vary between patients and between regions of the same tumour. A candidate antigen expressed by tumour-associated fibroblasts, for example, might be attractive for stromal targeting strategies but inappropriate for direct tumour-cell killing. Distinguishing the malignant compartment from the reactive stromal compartment requires information about which proteins co-localise with which, and dissociated single-cell methods, while powerful, lose the tissue architecture that defines these compartments. Proximity labelling performed in intact systems, combined with graph-based inference, offers a route to recovering some of this architectural information from molecular data.

From a machine learning perspective, the tumour antigen problem illustrates a broader trend in computational biology: the shift from classification of individual entities to inference over relational structures. Early bioinformatics treated each gene or protein as an independent feature, and predictive models were built on expression vectors. The realisation that biological molecules operate in networks led to a family of methods that propagate information across interaction graphs, including network propagation algorithms, random walk approaches, and eventually graph neural networks. Each generation of methods has expanded the kinds of questions that can be asked. Where earlier approaches could ask whether a protein is connected to known disease genes, graph learning can ask whether a protein occupies a network position characteristic of disease-relevant molecules, a subtler and often more robust criterion. In the antigen discovery setting, this means the model can learn what the neighbourhood of a validated tumour antigen looks like and then score uncharacterised candidates by the similarity of their neighbourhoods.

Validation remains the essential counterweight to computational prediction, and the history of antigen discovery offers sobering lessons about candidates that looked compelling in silico but failed in vivo. A predicted proximity antigen must survive several successive tests: confirmation that the protein is genuinely presented on the tumour cell surface by human leukocyte antigen molecules, demonstration that healthy tissues lack comparable presentation, evidence that T cells capable of recognising the presented peptide exist in patients, and finally evidence that engaging those T cells produces tumour killing without tissue damage. Each of these tests is demanding, and attrition between stages is high. Computational methods that incorporate proximity information improve the front end of this pipeline by enriching the candidate list for molecules likely to pass the later stages, which is a meaningful advance even though no computational prediction can substitute for experimental validation.

The timing of this work within the broader technological landscape is notable. Spatial transcriptomics and spatial proteomics have matured rapidly, multiplexed imaging now routinely profiles dozens of proteins in intact tissue sections, and proximity labelling has been adapted to an expanding range of model systems. Meanwhile, immunopeptidomics has improved in sensitivity to the point where thousands of presented peptides can be catalogued from limited clinical material. The convergence of these technologies with graph-based machine learning creates conditions in which antigen discovery can become less dependent on serendipity. Historically, many successful tumour antigens were found through patient-specific approaches, such as isolating tumour-infiltrating lymphocytes and identifying their targets, which is powerful but slow and individualised. A systematic, proximity-informed discovery framework aims to generalise this process, producing catalogues of candidate antigens that can serve the broader population of patients.

There are also implications beyond T cell therapy. Antibody-drug conjugates require surface antigens with sufficient differential expression and internalisation behaviour, and the proximity context of an antigen can inform predictions about its trafficking and its accessibility to circulating antibodies. Bispecific molecules that bridge tumour cells and T cells depend on pairs of antigens whose joint expression pattern is tumour-restricted, and proximity data speak directly to co-localisation. Even vaccine design benefits, since peptide antigens that arise in the context of a tumour-specific molecular programme are more likely to elicit responses that discriminate tumour from self. In each of these therapeutic modalities, the fundamental question is the same: which molecular features distinguish the tumour cell surface in its intact context, and how confidently can that distinction be made for a given patient population.

As the field moves forward, the integration of proximity-guided graph learning into antigen discovery pipelines will likely be judged by a practical standard: whether the candidates it surfaces translate into therapies with wider therapeutic windows than those discovered by expression-based approaches alone. The conceptual contribution, however, may prove equally durable. By treating the tumour cell surface as a structured, relational system rather than a list of abundances, this line of work aligns the computational representation of tumour biology with the way the immune system itself reads that biology, through contact, context, and the molecular neighbourhoods in which every antigen is embedded.

Subject of Research: Proximity-guided graph learning reveals tumour-associated proximity antigens

Article Title: Proximity-guided graph learning reveals tumour-associated proximity antigens

Article References: Scandore, C., Malone, C. F., May, C. K., de Regt, A. K., Guernsey, J., Ma, H., Dephoure, N., Setter, B., Howell, R. A., Johnson, K. R., Farr, C. L., Romero, S., Vignale, L., Vittum, T., Dawson, E., Habtetsion, T., Nardi, F., Woodruff, B., Mathay, M., … Fadeyi, O. O. (2026). Proximity-guided graph learning reveals tumour-associated proximity antigens. Nature. https://doi.org/10.1038/s41586-026-11003-7

Image Credits: AI Generated

DOI: 10.1038/s41586-026-11003-7

Keywords: Proximity-guided, graph, learning, reveals, tumour-associated, proximity, antigens, scientific research