Research Radar — 2026-06-17
Methods & AI
Computational
FlowBench — separating planning, fault recovery and interpretation in agentic bioinformatics
bioRxiv Published 2026-06-16 preprint DOI: 10.64898/2026.06.12.731844
agentic AI large language model bioinformatics benchmark fault recovery planning workflow evaluation
Summary: Introduces FlowBench, a benchmark that decomposes agentic bioinformatics performance into four independently evaluated dimensions: planning (generating valid workflows from biological intent), fault recovery (handling errors and invalid data), biological interpretation (data-grounded reasoning), and end-to-end output fidelity. The authors built FlowAgent, a modular, provider-agnostic framework that can swap backbone models across providers and selectively disable components, and used it to evaluate 23 models from three major providers. Three key findings emerge. First, generating a valid workflow plan from a named toolchain is largely solved, but inferring an appropriate toolchain from biological intent alone compresses all models into a narrow 44–57% pass-rate band regardless of model tier — revealing that biological reasoning about which tools to use is the bottleneck, not plan execution. Second, ablation experiments show that the dependency-structured plan and a completeness-reflection step drive performance, while adding a same-context validator-driven retry makes structural quality worse. Third, fault recovery and data-grounded interpretation remain unsolved across all models: models frequently propose fixes that force a clean exit while leaving the underlying data invalid, and their data-grounded interpretation consistently lags behind internal-knowledge recall. Critically, safety does not emerge from capability — reasoning-tier models were among the least reliable at recognizing unrecoverable faults, proposing plausible-sounding but incorrect fixes. The study concludes that once planning saturates, agent architecture and refusal calibration, not model scale, are the productive frontier for improving agentic bioinformatics.
Why it matters: Agentic LLM systems are being deployed in bioinformatics at a pace that far outstrips our understanding of their failure modes, and single-metric evaluations have created a false sense of progress by conflating capabilities that fail independently. FlowBench's decomposition of performance into planning, fault recovery, interpretation, and output fidelity provides the first rigorous framework for understanding where these systems actually succeed and fail. The finding that all models — including reasoning-tier frontier models — converge to a narrow 44–57% pass-rate band on tool selection from biological intent is a sobering reality check. It means that the hardest problem in agentic bioinformatics is not executing workflows but knowing which workflow to execute — a task that requires genuine biological understanding. The fault recovery findings are equally important: models that confidently produce incorrect fixes while reporting success represent a safety hazard that will only grow as these systems are deployed in clinical and translational settings. FlowBench provides the evaluation infrastructure needed to make agentic bioinformatics an engineering discipline rather than a collection of impressive demos.
Why for Yiru: FlowBench's decomposition of agentic performance into planning, fault recovery, and interpretation maps directly onto the challenges of using AI agents for TME research. Planning — knowing which computational tools to apply to a biological question — is the daily work of computational biology. The finding that all models struggle with tool selection from biological intent suggests that AI agents need structured domain knowledge about TME analysis workflows (e.g., which tools to use for cell-type deconvolution vs. spatial niche analysis vs. immune repertoire characterization). The fault recovery findings are practically important: TME datasets are noisy and heterogeneous, and an agent that confidently produces incorrect results while reporting success could lead to erroneous biological conclusions. FlowBench's modular evaluation framework could be adapted to create a TME-specific agent benchmark, testing whether agents can correctly select and execute workflows for common TME analysis tasks like identifying differentially expressed immune genes, characterizing spatial immune niches, or integrating multi-omics data.
scIsoAgent enables autonomous isoform-resolved characterization and sequence-informed interpretation of long-read single-cell transcriptomes
bioRxiv Published 2026-06-16 preprint DOI: 10.64898/2026.06.11.731519
agentic AI large language model long-read sequencing single-cell transcriptomics isoform analysis alternative splicing workflow automation
Summary: Presents scIsoAgent, an autonomous LLM-powered scientific agent specifically designed for long-read single-cell RNA-seq analysis, addressing the gap between general-purpose biomedical AI agents and the specialized needs of isoform-resolved transcriptomics. Long-read single-cell RNA sequencing uniquely links cellular identity to full-length transcript isoforms and sequence-level features (splice junctions, UTRs, RNA modifications), but realizing its biological value requires reproducible workflows that connect specialized long-read analysis tools with biological interpretation. Existing LLM-based biomedical agents support general omics analysis but are not designed for the isoform-resolved, sequence-informed workflows characteristic of long-read single-cell data. scIsoAgent turns heterogeneous long-read single-cell inputs into traceable, reproducible analysis pipelines by orchestrating specialized tools for isoform quantification, differential transcript usage analysis, and sequence-level feature characterization. The agent incorporates domain-specific knowledge about long-read data characteristics (higher error rates, read length distributions, isoform complexity) that general-purpose agents lack. The system produces structured outputs that connect isoform-level findings to biological interpretation, including identification of isoform switches, novel transcript discovery, and sequence-informed functional annotation. Benchmarking demonstrates that scIsoAgent outperforms general-purpose biomedical agents on long-read-specific tasks while maintaining the traceability and reproducibility essential for scientific workflows.
Why it matters: Alternative isoform usage can alter gene function independently of total gene expression — a protein with an included vs. excluded domain can have opposite functions — yet most single-cell analysis pipelines aggregate all isoforms into gene-level counts, discarding this critical layer of regulation. Long-read single-cell sequencing solves this by capturing full-length transcripts, but the analysis is substantially more complex than short-read scRNA-seq, requiring specialized tools for isoform quantification, splice junction analysis, and sequence-level interpretation. This complexity has been a barrier to adoption. scIsoAgent addresses this by providing an autonomous agent that handles the specialized workflows, making long-read single-cell analysis accessible to researchers without deep bioinformatics expertise in this area. The agentic approach is particularly well-suited to this problem because long-read analysis involves many tool-specific decisions (which aligner, which isoform quantifier, which differential transcript usage method) that depend on the biological question and data characteristics — exactly the kind of context-dependent tool selection that AI agents can automate.
Why for Yiru: Isoform-level regulation is directly relevant to TME biology. Many immune checkpoint molecules, cytokine receptors, and signaling proteins have functionally distinct isoforms — for example, soluble vs. membrane-bound variants of PD-L1, or alternative splice forms of CD44 that affect immune cell trafficking. Current TME single-cell studies overwhelmingly use gene-level quantification and miss this layer of regulation entirely. scIsoAgent could be applied to long-read single-cell data from tumour and immune cells to discover isoform switches associated with immune evasion, T cell exhaustion, or therapy resistance — changes that would be invisible at the gene level. The traceable, reproducible workflow design also aligns with best practices for computational TME research, where methodological transparency is essential for cross-study comparisons. The domain-specific knowledge integration approach (teaching the agent about long-read data characteristics) is a template for building TME-specific AI agents that understand the particular challenges of tumour transcriptomics, such as tumour purity effects, stromal contamination, and the need for careful batch correction.
HiCFoundation — a generalizable foundation model for chromatin architecture, single-cell and multiomics analysis across species
Nature Methods Published 2026-06-15 research article DOI: 10.1038/s41592-026-03097-8
foundation model 3D genome Hi-C chromatin architecture self-supervised learning multiomics cross-species
Summary: Reports HiCFoundation, a foundation model pretrained on a large corpus of Hi-C data for comprehensive 3D genome and epigenomics analysis. The three-dimensional organization of chromatin in the nucleus — measured by Hi-C and related technologies — regulates gene expression by bringing enhancers into proximity with promoters and organizing the genome into topologically associating domains (TADs) and A/B compartments. However, extracting functional insights from Hi-C data has been challenging because existing methods are often task-specific, species-specific, or require substantial retraining for each new analysis. HiCFoundation addresses this through large-scale self-supervised pretraining on Hi-C data from multiple species and cell types, learning a general representation of chromatin architecture that captures both universal principles of genome folding and context-specific features. After pretraining, the model can be fine-tuned on downstream tasks including single-cell Hi-C analysis, multiomics integration (linking chromatin structure to gene expression, histone modifications, and DNA methylation), and cross-species chromatin comparison. The model provides a unified, efficient, generalizable, and interpretable paradigm for 3D genome analysis, substantially reducing the data and computational requirements for task-specific models. A companion Research Briefing article discusses the regulatory implications and how HiCFoundation supports integrative analysis linking genome architecture to downstream regulatory function.
Why it matters: Foundation models — large models pretrained on broad data that can be fine-tuned for specific tasks — have transformed natural language processing and protein structure prediction, but their application to 3D genome biology has lagged due to the unique challenges of chromatin contact data (sparse matrices, distance-dependent signal decay, variability across resolutions). HiCFoundation demonstrates that the foundation model paradigm can be successfully applied to 3D genome organization, with the key advantage that a single pretrained model can support diverse downstream tasks without task-specific architecture design. The cross-species generalizability is particularly valuable — it means insights about chromatin architecture can be transferred from well-studied model organisms to less-characterized species or to patient samples where training data is scarce. The multiomics integration capability also addresses a practical bottleneck: connecting 3D genome structure to functional genomics data (RNA-seq, ChIP-seq, ATAC-seq) typically requires complex multi-step pipelines that HiCFoundation streamlines into a unified framework.
Why for Yiru: 3D genome organization is increasingly recognized as a determinant of gene regulation in cancer, including in the TME. Chromatin structural variations — TAD disruptions, enhancer hijacking, compartment switching — can activate oncogenes or silence tumour suppressors, and these structural changes may influence how tumour cells and immune cells respond to therapy. HiCFoundation's ability to integrate Hi-C data with gene expression and epigenomic data in a unified framework could be applied to cancer samples to identify structural variants that drive immune evasion or therapy resistance. The cross-species generalizability is also relevant for comparing chromatin architecture between mouse tumour models and human tumours — a common need in preclinical TME research. The foundation model approach (pretrain once, fine-tune for many tasks) provides a template for building similar models for other TME-relevant data types, such as spatial transcriptomics or imaging mass cytometry, where task-specific models currently dominate.
PhenoBIC — operator-free single-cell spatial phenotyping in multiplex imaging data using deep learning of cell staining patterns
bioRxiv Published 2026-06-16 preprint DOI: 10.64898/2026.06.11.731702
multiplex imaging spatial phenotyping deep learning cell classification tissue microenvironment image analysis computational pathology
Summary: Introduces PhenoBIC (Biomarker Imprint of a Cell), a pre-trained deep learning model for image classification of multiplexed biomarker signals to achieve operator-free single-cell spatial phenotyping. Multiplex imaging technologies that simultaneously measure 30–60 protein markers in tissue sections have revolutionized spatial biology, but their analysis remains a bottleneck: most workflows require manual gating to define cell phenotypes, which is slow, subjective, and operator-dependent. PhenoBIC addresses this by training a deep learning model to classify cells directly from the multiplexed biomarker staining patterns — the "biomarker imprint" — without manual threshold setting. The model achieves an F1-score of approximately 0.88, outperforming manual gating and other machine learning-based approaches for cell marker expression classification. Validation spans multiple biomarkers, tissue sampling strategies (whole-slide and tissue microarray), and cancer types. Critically, PhenoBIC generalizes across different antibody panels and tissue contexts without retraining because it learns the visual patterns of positive and negative staining rather than panel-specific thresholds. The operator-free workflow dramatically reduces the time from image acquisition to biological insight while eliminating the inter-operator variability that has plagued multiplex imaging studies.
Why it matters: Multiplex imaging is one of the most powerful tools in spatial biology, but the analysis bottleneck — manual gating of 30–60 markers across millions of cells — has limited its throughput, reproducibility, and clinical translation. PhenoBIC's operator-free phenotyping is not just a convenience; it fundamentally changes what is possible with multiplex imaging by enabling truly large-scale, reproducible spatial profiling. The cross-panel generalization is particularly important because it means PhenoBIC can be applied to existing multiplex imaging datasets without requiring panel-specific training — a crucial feature given the diversity of antibody panels used across studies. The improvement over manual gating (F1 ~0.88 vs. manual gating) is substantial enough to change biological conclusions in cases where manual gating misclassifies cells with intermediate or ambiguous marker expression. This work is part of a necessary transition in spatial biology from artisanal, operator-dependent analysis to automated, reproducible computational pipelines.
Why for Yiru: Spatial phenotyping of the TME — identifying which immune cells are present, where they are located, and what functional states they occupy — is central to understanding tumour-immune interactions. PhenoBIC directly addresses the most time-consuming step in this workflow. In TME studies using multiplex imaging (e.g., CODEX, MIBI, CyCIF), manually gating 40+ markers to define cell types like exhausted CD8+ T cells, immunosuppressive macrophages, or activated fibroblasts can take weeks and introduce operator bias. PhenoBIC could reduce this to minutes while improving reproducibility. The cross-panel generalization is practically valuable: TME researchers often use different antibody panels for different studies or different tumour types, and a model that works across panels without retraining would enable meta-analysis of published multiplex imaging data. The operator-free approach also supports clinical translation, where reproducible, automated analysis is a regulatory requirement.
oxo-flow — compiled, memory-safe bioinformatics workflow orchestration in Rust
bioRxiv Published 2026-06-15 preprint DOI: 10.64898/2026.06.11.731578
workflow engine bioinformatics Rust compiled language performance memory safety software engineering
Summary: Introduces oxo-flow, a bioinformatics workflow engine written in Rust that compiles to a single native binary, addressing the startup latency and runtime safety limitations of existing workflow engines. The dominant bioinformatics workflow engines — Snakemake, Nextflow, CWL, and WDL — run on interpreted or just-in-time compiled language runtimes (Python, Java/Groovy, JavaScript), incurring hundreds of milliseconds of startup latency and providing no compile-time safety guarantees. On an Apple M5 processor, oxo-flow parses, validates, and dry-runs a production-scale workflow in approximately 22 milliseconds — before Snakemake or Nextflow have finished loading their runtime environments. Peak memory usage is 16 megabytes, representing roughly a 100x reduction compared to existing engines. The Rust implementation provides memory safety guarantees at compile time, eliminating entire classes of runtime errors (null pointer dereferences, data races, buffer overflows) that can cause silent workflow failures in production. The engine supports the same workflow description language semantics as existing engines, maintaining compatibility with the large ecosystem of existing bioinformatics workflows. The single-binary deployment model simplifies installation and containerization, reducing the dependency complexity that plagues bioinformatics workflow environments.
Why it matters: Workflow engines are the invisible infrastructure of computational biology — virtually every large-scale genomic, transcriptomic, or proteomic analysis runs through Snakemake, Nextflow, or CWL. But these engines carry significant overhead: hundreds of megabytes of memory, multi-second startup times, and complex dependency chains that make reproducibility challenging. For large-scale production environments processing thousands of samples, these overheads compound into hours of wasted compute time and memory. oxo-flow's 100x improvements in startup time and memory usage are not incremental — they change the economics of running bioinformatics workflows at scale, particularly in cloud environments where faster startup enables more efficient autoscaling and lower costs. The memory safety guarantees from Rust are also practically important: silent workflow failures from memory errors in production bioinformatics pipelines can produce incorrect scientific results that propagate through downstream analyses before being detected — or never detected at all.
Why for Yiru: TME computational research increasingly involves large-scale workflows processing hundreds of tumour samples across multiple omics modalities — workflows that strain existing engines. oxo-flow's performance characteristics (22ms startup, 16MB memory) would be particularly valuable for iterative TME analysis pipelines where workflows are repeatedly executed during parameter exploration, or for deploying TME analysis in clinical settings where fast turnaround is essential. The single-binary deployment model also addresses a real pain point in computational TME research: sharing analysis pipelines with collaborators who may not have the exact same computing environment. A single binary that bundles the entire workflow engine could significantly improve the reproducibility of TME computational analyses across labs. The memory safety guarantees are relevant for clinical-grade TME analysis pipelines where computational errors have patient consequences.
Super Learner Ensemble Modeling of CPTAC Proteomic Data for Survival Prediction in Head and Neck Squamous Cell Carcinoma
bioRxiv Published 2026-06-16 preprint DOI: 10.64898/2026.06.11.731237
survival analysis proteomics ensemble learning head and neck cancer CPTAC machine learning prognosis
Summary: Evaluates the Super Learner (SL) ensemble algorithm for survival prediction in head and neck squamous cell carcinoma (HNSCC) using proteomic features from the CPTAC cohort. Survival analysis in cancer has traditionally relied on Cox proportional hazards models, with some exploration of black-box machine learning methods. The Super Learner algorithm addresses the model selection dilemma by combining diverse candidate algorithms into a weighted ensemble that performs comparably to the best individual candidate method — without requiring the analyst to guess which method will work best. This study applies SL to proteomic features and clinical covariates from 96 CPTAC HNSCC samples, using three candidate algorithms (Cox LASSO, Cox Ridge, and Random Survival Forest) with model optimization via Uno's time-dependent Concordance Index. Models were tested at 1-year and 3-year time horizons using bootstrap resampling. The ensemble approach identifies proteomic features associated with survival that would be missed by single-method approaches, demonstrating the value of proteomics beyond traditional clinical and transcriptomic prognostic models. The framework provides a template for applying ensemble survival modeling to proteomic data across cancer types.
Why it matters: Most cancer survival models use gene expression data, but proteins are the functional effectors of cellular phenotypes and drug targets — there is often poor correlation between mRNA and protein levels. The CPTAC (Clinical Proteomic Tumor Analysis Consortium) has generated proteomic data for multiple cancer types, but methods for extracting prognostic information from these data lag behind transcriptomic approaches. The Super Learner framework addresses a fundamental challenge in survival modeling: different algorithms capture different aspects of the data (linear effects, non-linear interactions, regularization), and the best algorithm varies by dataset and clinical question. By combining multiple algorithms into an ensemble, SL provides robust performance without requiring the analyst to commit to a single method. The application to HNSCC proteomics demonstrates that protein-level features add prognostic information beyond what is available from clinical variables alone.
Why for Yiru: Proteomic characterization of the TME — measuring the actual protein-level expression of immune checkpoints, cytokines, and signalling molecules — provides information that transcriptomics cannot capture, particularly for secreted and cell-surface proteins that are regulated post-transcriptionally. The Super Learner ensemble framework could be applied to TME proteomic data to build prognostic models that integrate protein-level features of both tumour cells and immune infiltrates. The framework's ability to combine different model types is well-suited to the heterogeneity of TME data, where some features may have linear effects (e.g., CD8+ T cell density) while others have complex non-linear relationships with outcome (e.g., macrophage polarization states). The CPTAC proteomic data used in this study are publicly available, and similar proteomic cohorts exist for other cancer types, providing opportunities to extend this ensemble approach to TME-focused prognostic modeling.
Biomedical discoveries
Biomedicine
Profilin-1 Deficiency Activates STING to Drive T Cell-Mediated Anti-Tumor Immunity in Breast Cancer
bioRxiv Published 2026-06-10 preprint DOI: 10.64898/2026.06.05.730362
STING cGAS anti-tumor immunity breast cancer Profilin-1 cytosolic DNA CD8+ T cell genomic instability
Summary: Reveals that depletion of the actin-binding protein Profilin-1 (Pfn1) in breast cancer cells triggers genomic instability leading to cytosolic DNA accumulation, cGAS-STING pathway activation, and CD8+ T cell-mediated anti-tumor immunity. Profilin-1 is a key regulator of actin dynamics, and its dysregulation in cancer has well-characterized effects on tumour cell-intrinsic processes including migration and invasion. However, whether Pfn1 modulation influences immune surveillance has been unknown. Using an inducible CRISPR/Cas9 knockout model, the authors demonstrate that triggering Pfn1 depletion in breast cancer cells leads to features of genomic instability including polyploidy, micronuclei, and DNA damage, with intrinsic defects in both homologous recombination and non-homologous end-joining DNA repair. Pfn1-deficient cells exhibit nuclear envelope abnormalities and accumulate cytosolic DNA, which activates the nucleic acid-sensing cGAS-STING pathway. This innate immune activation triggers a type I interferon response and production of T cell-recruiting chemokines, leading to increased CD8+ T cell infiltration and tumour control in immunocompetent mice. The anti-tumor effect requires both STING signaling in tumour cells and an intact adaptive immune system — it is lost in STING-knockout tumours and in immunodeficient hosts. The study establishes a direct mechanistic link between actin cytoskeleton dysregulation, genomic instability, innate immune sensing, and adaptive anti-tumor immunity.
Why it matters: The cGAS-STING pathway has emerged as a central node connecting genomic instability to anti-tumor immunity, and STING agonists are in active clinical development. However, most research has focused on exogenous STING activation (agonist drugs, radiation-induced DNA damage) rather than endogenous mechanisms that tumours use to regulate STING signaling. This study reveals that Profilin-1 — an actin-binding protein with no obvious connection to innate immunity — is a critical regulator of the STING pathway in cancer cells. The implication is that tumours may actively suppress STING signaling by maintaining Pfn1 expression, and that Pfn1 loss or inhibition could be exploited therapeutically to "unmask" tumours to the immune system. The finding that cytoskeletal protein dysregulation triggers genomic instability and STING activation also suggests that other cytoskeletal perturbations — including those caused by microtubule-targeting chemotherapies like taxanes — may have underappreciated immunostimulatory effects mediated through this pathway. More broadly, this work exemplifies how tumour cell-intrinsic processes traditionally studied in isolation (actin dynamics) can have profound immunological consequences.
Why for Yiru: The TME is shaped by both tumour-intrinsic and immune-extrinsic factors, and this study demonstrates a direct molecular connection between a tumour cell cytoskeletal protein and the activation of anti-tumor T cell responses. The Pfn1-STING axis could be interrogated in TME datasets to determine whether Pfn1 expression levels correlate with STING pathway activation, T cell infiltration, or immunotherapy response across breast cancer subtypes. The finding that Pfn1 loss causes nuclear envelope abnormalities and cytosolic DNA accumulation raises the question of whether other cytoskeletal perturbations in the TME — such as mechanical stress from dense extracellular matrix or cytoskeletal changes during epithelial-mesenchymal transition — similarly activate STING signaling and contribute to immune surveillance. For computational TME analysis, this work highlights the importance of considering tumour cell-intrinsic features (cytoskeletal gene expression, DNA damage signatures, STING pathway activation) alongside traditional immune infiltration metrics when modeling immunotherapy response.
Cycling persister clones with elevated NR2F1-mediated cholesterol biosynthesis cause chemotherapy resistance
bioRxiv Published 2026-06-10 preprint DOI: 10.64898/2026.06.06.730520
drug resistance chemotherapy persister cells cholesterol biosynthesis NR2F1 organoid tongue cancer clonal dynamics
Summary: Uses a human tongue cancer organoid library with time-lapse imaging to track individual cancer cell clones during and after chemotherapy exposure, identifying a subpopulation of "cycling persisters" (CPs) that continue to proliferate under drug treatment and drive tumour relapse. While the concept of drug-tolerant persister cells — a subpopulation that survives drug exposure through non-genetic mechanisms — is well established, most persister studies have focused on non-cycling, quiescent cells. This study identifies CPs that maintain proliferation under chemotherapy and form larger colonies than non-persister clones during treatment. By directly sampling thousands of CP and non-CP clones from 3D organoid cultures based on colony size differences, the authors performed molecular characterization revealing that CPs are driven by elevated NR2F1-mediated cholesterol biosynthesis. Tumour-intrinsic interferon signaling and cholesterol metabolism distinguish CPs from non-CPs, and pharmacological inhibition of cholesterol biosynthesis sensitizes CPs to chemotherapy. The study challenges the prevailing model of persisters as dormant cells and instead identifies an active, proliferative persister state sustained by metabolic reprogramming — specifically, cholesterol biosynthesis — that represents a therapeutically targetable vulnerability.
Why it matters: Chemotherapy resistance remains the leading cause of cancer treatment failure, and understanding the cellular mechanisms that enable tumour cells to survive drug exposure is essential for developing strategies to prevent relapse. The identification of cycling persisters — cells that continue to divide under chemotherapy rather than entering quiescence — fundamentally revises the persister cell concept. Quiescent persisters have been notoriously difficult to target because most cancer drugs target proliferating cells. Cycling persisters, by contrast, maintain proliferation and are potentially vulnerable to therapies that target the metabolic pathways sustaining their growth, such as cholesterol biosynthesis. The NR2F1-cholesterol biosynthesis axis identified here represents a druggable vulnerability — statins and other cholesterol-lowering drugs are already clinically available and could potentially be repurposed to prevent chemotherapy resistance. The organoid time-lapse imaging approach — directly observing clonal dynamics during and after treatment — is also methodologically important, as it captures the dynamic nature of resistance that is missed by endpoint analyses.
Why for Yiru: The concept of cycling persisters has direct relevance to the TME. Chemotherapy not only kills tumour cells but also reshapes the TME by releasing damage-associated molecular patterns (DAMPs), altering cytokine gradients, and affecting immune cell viability. Cycling persisters that survive chemotherapy may interact differently with the immune system than the original tumour population — their elevated cholesterol metabolism, for example, could affect lipid-mediated immune signaling or the formation of lipid-laden immunosuppressive myeloid cells. The cholesterol biosynthesis dependency is also interesting from a TME perspective: cholesterol metabolism is increasingly recognized as a regulator of immune cell function, with cholesterol accumulation in CD8+ T cells promoting exhaustion and cholesterol in tumour cells affecting membrane fluidity and signaling. The organoid time-lapse approach could be adapted to study how persister clones interact with immune cells in co-culture systems, revealing whether persisters have altered immunogenicity or immune evasion capacity.
Single-nucleus multiomics unveils malignant cellular states, regulatory architectures and microenvironmental reorganization across the G-CIMP epigenomic transition in IDH-mutant glioma
bioRxiv Published 2026-06-09 preprint DOI: 10.64898/2026.06.05.730437
glioma IDH-mutant G-CIMP single-nucleus multiomics epigenomics tumour microenvironment cellular states
Summary: Integrates single-nucleus RNA-seq and ATAC-seq across 18 tumour specimens from 10 patients to resolve the cellular states, regulatory architectures, and microenvironmental reorganization that characterize the G-CIMP (glioma CpG island methylator phenotype) epigenomic transition in IDH-mutant gliomas. IDH-mutant gliomas are stratified by G-CIMP into High (GCH) and Low (GCL) subtypes that exhibit markedly divergent clinical outcomes, with GCL representing a more aggressive disease state. However, the cellular and regulatory determinants of this transition have remained poorly defined. The authors identify six malignant cellular states whose differential enrichment across G-CIMP strata delineates the GCL epigenomic transition. GCL tumours are enriched for independently prognostic Mesenchymal and Mitotic Proliferative states, driven by convergent E2F, MYC, MEF2, and NFI-family regulatory networks confirmed across chromatin, histone, and transcriptomic modalities. Pseudotime trajectory inference reveals a multifurcating developmental model of malignant state transitions. Critically, the study also characterizes how the tumour microenvironment reorganizes across the G-CIMP transition, with GCL tumours showing altered immune composition, including changes in myeloid and T cell compartments that accompany the shift toward more aggressive malignant states. The multi-modal regulatory architecture provides a comprehensive resource for understanding the molecular basis of glioma progression and nominates regulatory networks as potential therapeutic targets.
Why it matters: IDH-mutant gliomas are the most common gliomas in adults and are ultimately fatal despite their initially indolent course. The G-CIMP classification has been clinically useful for prognosis, but the biological mechanisms driving the transition from low-grade GCH to high-grade GCL have been poorly understood. This study provides the most comprehensive molecular characterization of the G-CIMP transition to date by simultaneously profiling the transcriptome and chromatin accessibility at single-cell resolution. The identification of convergent regulatory networks (E2F, MYC, MEF2, NFI) driving the aggressive mesenchymal and proliferative states provides a molecular framework for understanding glioma progression and nominates specific transcription factor networks as therapeutic targets. The characterization of TME reorganization across the G-CIMP transition is equally important — it suggests that the epigenomic evolution of tumour cells actively remodels the immune microenvironment, potentially creating new therapeutic vulnerabilities or resistance mechanisms during disease progression.
Why for Yiru: This study demonstrates how tumour cell-intrinsic epigenomic evolution (the G-CIMP transition) is accompanied by coordinated changes in the TME, including immune cell composition and activation states. The single-nucleus multiomics approach — simultaneously measuring gene expression and chromatin accessibility in the same cells — is directly applicable to TME research, where understanding how chromatin-level regulation controls immune cell states and tumour-immune interactions is a frontier. The regulatory network analysis framework (linking transcription factor activity to cellular states across chromatin, histone, and transcriptomic modalities) could be applied to TME studies to identify the regulatory drivers of immune evasion, T cell exhaustion, or macrophage polarization. The finding that G-CIMP transition is associated with TME reorganization also underscores the importance of longitudinal TME profiling during tumour evolution — the immune contexture of a low-grade glioma is fundamentally different from that of its high-grade descendant, with implications for immunotherapy timing.
Glioblastoma-derived extracellular vesicles released after radiation promote cognitive impairment through NFκB-mediated microglial activation
bioRxiv Published 2026-06-11 preprint DOI: 10.64898/2026.06.09.730969
glioblastoma extracellular vesicles radiation therapy cognitive impairment NFκB microglia neuroinflammation
Summary: Investigates whether glioblastoma-derived extracellular vesicles (EVs) released after radiation treatment contribute to the cognitive impairment commonly experienced by GBM survivors, identifying an NFκB-mediated microglial activation mechanism. Cognitive decline is a frequent and devastating sequela in glioblastoma patients who survive their disease, yet its mechanisms remain poorly understood — it has been attributed to both the tumour itself and to treatment effects (radiation, chemotherapy), but the molecular mediators are unclear. EVs are established mediators of intercellular signaling within the tumour microenvironment. The authors show that GBM-derived EVs released specifically after radiation treatment (RT-EVs) — but not EVs from unirradiated tumours — cause cognitive deficits and neuroinflammatory responses when administered to mice in vivo. In vitro, RT-EVs activate the NFκB pathway in microglia, inducing the release of neurotoxic hydrogen peroxide (H₂O₂). Critically, NFκB p50 knockdown in microglia abolishes the H₂O₂ release triggered by RT-EVs, demonstrating mechanistic dependence on NFκB signaling. The study identifies a causal chain: radiation → GBM EVs with altered cargo → microglial NFκB activation → neurotoxic H₂O₂ release → cognitive impairment. This mechanism suggests that targeting EV release, EV cargo, or microglial NFκB signaling could mitigate radiation-induced cognitive decline in GBM patients without compromising the anti-tumour effects of radiation.
Why it matters: As GBM therapies improve and patients live longer, treatment-related morbidities — particularly cognitive decline — are becoming a central clinical challenge. Radiation therapy is essential for GBM local control, but its contribution to cognitive impairment has been difficult to separate from tumour effects and has lacked a mechanistic explanation. This study provides that mechanism: radiation fundamentally alters the cargo of tumour-derived EVs, transforming them from neutral or pro-tumour signals into neurotoxic agents that activate microglia. The therapeutic implications are immediately actionable: pharmacological blockade of EV biogenesis, neutralization of specific EV cargo components, or inhibition of microglial NFκB signaling could all be tested as neuroprotective strategies during radiation therapy. The study also has broader implications for understanding how cancer therapies reshape intercellular communication in the TME — radiation, chemotherapy, and targeted therapy may all alter EV cargo in ways that have unintended consequences for normal tissue function.
Why for Yiru: The concept that treatment reshapes tumour-derived extracellular vesicles, with functional consequences for host tissue, extends beyond the brain. In the TME, radiation, chemotherapy, and immunotherapy all alter the secretome of tumour cells, including EV cargo. These treatment-induced changes could affect immune cell function in the TME — for example, radiation-induced EVs might carry different immunomodulatory cargo (DAMPs, cytokines, checkpoint ligands) than EVs from untreated tumours, potentially enhancing or suppressing anti-tumour immunity. The NFκB-microglia axis identified here has parallels in the TME, where tumour-derived EVs can activate NFκB in tumour-associated macrophages to drive immunosuppressive phenotypes. The experimental paradigm — comparing EVs from treated vs. untreated tumours — could be applied to study how different TME-directed therapies alter EV-mediated communication between tumour cells, immune cells, and stromal cells.
SHERLOC — An interpretable deep learning model for longitudinal circulating tumor DNA data in survival analysis
bioRxiv Published 2026-06-09 preprint DOI: 10.64898/2026.06.04.730097
liquid biopsy ctDNA deep learning survival analysis longitudinal data interpretability cancer monitoring
Summary: Introduces SHERLOC, a deep learning framework specifically designed for survival analysis using longitudinal on-treatment circulating tumor DNA (ctDNA) data. Serial ctDNA measurements offer a noninvasive window into treatment response and emerging resistance, but clinical ctDNA data present substantial methodological challenges: high-dimensional short longitudinal sequences of variant allele frequencies (VAFs) with irregular sampling intervals, combined with typically small sample sizes in clinical trials. SHERLOC addresses these challenges through three integrated components: (1) shared temporal representations of gene-level VAF trajectories that capture how individual mutations evolve during treatment, (2) feature-specific temporal trajectories of panel-level ctDNA biomarkers (e.g., mean VAF, number of detectable mutations), and (3) survival-aware genomic representations pre-trained on a large pan-cancer tissue-biopsy dataset (MSK-CHORD) that encode prior knowledge about the prognostic significance of mutations in specific genes. These components are integrated within an interpretable Cox proportional hazards framework, meaning the contribution of each feature to risk prediction can be inspected. Benchmarked against diverse statistical, ensemble, and deep learning approaches, SHERLOC demonstrates superior performance in predicting survival from longitudinal ctDNA data, particularly in the clinically challenging setting of small cohort sizes.
Why it matters: Liquid biopsy using ctDNA has transformed cancer monitoring, but the analytical methods for extracting prognostic information from longitudinal ctDNA data have not kept pace with the technology. Most clinical ctDNA analyses reduce rich longitudinal trajectories to single time-point metrics (e.g., "ctDNA clearance at week 4"), discarding the temporal dynamics that contain prognostic information. SHERLOC's ability to model the full trajectory of individual mutations over time — capturing not just whether ctDNA goes down but how it goes down (rate, pattern, which mutations persist) — represents a substantial methodological advance. The pre-training on MSK-CHORD (a large pan-cancer genomic dataset with survival outcomes) is a clever way to address the small-sample problem that plagues ctDNA studies: it transfers knowledge about which mutations are prognostically meaningful from tissue-based cohorts to liquid biopsy analysis. The interpretability is clinically essential — clinicians need to understand why a model predicts poor prognosis, not just that it does, to inform treatment decisions.
Why for Yiru: Longitudinal monitoring of the TME — how the immune infiltrate, cytokine milieu, and tumour cell states evolve during treatment — is the next frontier in immuno-oncology, and the analytical challenges mirror those of ctDNA: short, irregularly sampled time series with many features and small cohorts. SHERLOC's framework for modeling temporal trajectories of multiple features simultaneously could be adapted to longitudinal TME data from serial biopsies or blood-based immune monitoring. The pre-training strategy (transferring knowledge from large cross-sectional cohorts to longitudinal analysis) is also applicable: large TME profiling cohorts (TCGA, CPTAC) could be used to pre-train models of which TME features are prognostically important, with that knowledge transferred to the analysis of smaller longitudinal TME studies. The interpretable Cox framework is particularly important for TME research, where understanding which immune features drive prognosis is as important as predicting prognosis itself.
Spatial Compartmentalization of TCR Repertoires Between Primary Melanomas and Sentinel Lymph Nodes Reveals Distinct Clonal Architectures and Shared Antigen Recognition
bioRxiv Published 2026-06-09 preprint DOI: 10.64898/2026.06.05.730356
TCR repertoire melanoma sentinel lymph node T cell clonal architecture antigen recognition spatial immunology
Summary: Performs TCR β-chain sequencing on paired primary melanoma tumours and sentinel lymph nodes from 24 treatment-naive patients to quantify how T cell receptor repertoires are organized across these functionally linked but anatomically distinct sites of anti-tumour immunity. Primary tumours exhibited markedly reduced TCR diversity and pronounced clonal dominance compared with matched lymph nodes, consistent with selective expansion of tumour-reactive T cells at the tumour site. Tumour-associated clonotypes displayed significantly longer CDR3 sequences driven by increased non-templated nucleotide insertions — a feature associated with neoantigen recognition — suggesting that tumour-infiltrating T cells are enriched for clones recognizing mutation-derived antigens. Despite the distinct clonal architectures between compartments, the study identified shared clonotypes present in both tumour and lymph node, indicating that the same T cell clones can patrol both sites. The spatial organization of TCR repertoires reveals principles of how anti-tumour T cell responses are compartmentalized: the tumour enriches for oligoclonal, neoantigen-reactive populations through local antigen-driven selection, while the lymph node maintains a more diverse repertoire serving as a reservoir of T cell clones that can be recruited to the tumour. The balance between compartment-specific expansion and cross-compartment sharing has implications for understanding both natural anti-tumour immunity and the mechanism of action of immunotherapies like checkpoint blockade.
Why it matters: The spatial organization of anti-tumour T cell responses — how T cell clones are distributed between the tumour and lymphoid organs — is fundamental to understanding how the immune system controls (or fails to control) cancer. Most TCR repertoire studies profile either the tumour or peripheral blood in isolation, missing the spatial dimension. This paired analysis of primary tumour and draining lymph node from the same patients provides the most detailed view to date of how T cell repertoires are compartmentalized in human cancer. The finding that tumours are oligoclonal and enriched for clones with features of neoantigen reactivity validates the central model of cancer immunoediting: tumours selectively expand T cells that recognize tumour-specific antigens. The identification of shared clones between compartments suggests that the lymph node serves as a reservoir that can replenish tumour-infiltrating T cells — a concept with direct therapeutic relevance. Checkpoint blockade is thought to act in part by expanding T cell clones in lymph nodes that then traffic to tumours; the cross-compartment sharing observed here provides the clonal infrastructure for this mechanism.
Why for Yiru: TCR repertoire analysis is an increasingly important tool for understanding anti-tumour immunity in the TME, but most studies analyze the tumour in isolation. This study demonstrates that understanding the spatial organization of T cell clones — which clones are in the tumour vs. lymph node, which are shared — provides insights that tumour-only analysis cannot. For TME research, this suggests that analyzing TCR repertoires in both tumour and adjacent normal tissue or draining lymph nodes could reveal which T cell clones are tumour-selective vs. broadly reactive, and whether the TME actively excludes certain clones while enriching others. The CDR3 length feature as a marker of neoantigen reactivity is methodologically useful — it could be applied to TCR data from TME studies to estimate the neoantigen reactivity of tumour-infiltrating T cells without requiring knowledge of the specific neoantigens. The cross-compartment clonal sharing analysis also provides a framework for tracking how immunotherapy redistributes T cell clones between compartments.
Cross-disciplinary watchlist
Other Fields
B cell clones combine overlapping mechanisms to diversify during infection
Nature Immunology Published 2026-06-16 research article DOI: 10.1038/s41590-026-02571-x
B cell clonal diversification class-switch recombination somatic hypermutation malaria infection antibody
Summary: Demonstrates that during malaria parasite infection, individual activated B cell clones simultaneously combine multiple diversification mechanisms — early class-switch recombination, clonal expansion, effector fate bifurcation, and somatic hypermutation — to rapidly generate diverse effector and memory B cell populations in vivo. B cells can produce antibodies of different isotypes (IgM, IgG, IgE, IgA) with different affinities for antigen, and can differentiate into antibody-secreting plasma cells or memory B cells. The traditional model has treated these diversification processes as sequential and largely independent: first clonal expansion, then class switching, then affinity maturation through somatic hypermutation. This study challenges that model by showing that individual B cell clones commonly deploy all of these mechanisms concurrently during an active immune response. The overlapping deployment of diversification mechanisms allows a single B cell clone to simultaneously generate short-lived plasma cells producing low-affinity IgM for immediate protection, germinal center B cells undergoing affinity maturation for long-term high-affinity responses, and memory B cells poised for future encounters — all from the same founding clone and within the same temporal window. This parallel diversification strategy maximizes the breadth and speed of the antibody response during infection.
Why it matters: Understanding how B cells generate antibody diversity is fundamental to vaccine design, where the goal is to induce broad, high-affinity, long-lasting antibody responses. The finding that B cell clones deploy multiple diversification mechanisms simultaneously — rather than sequentially — has important implications for vaccine strategies. It suggests that vaccines designed to maximize only one aspect of the B cell response (e.g., high-affinity neutralizing antibodies) may be suboptimal because they ignore the parallel generation of other protective B cell populations (e.g., memory B cells, broadly reactive low-affinity clones). The malaria infection model is particularly relevant because malaria has been a notoriously difficult vaccine target — natural infection generates antibody responses that are often short-lived and strain-specific. Understanding how B cells naturally diversify during malaria infection could inform vaccine designs that recapitulate the most protective aspects of this diversification. More broadly, the concurrent deployment model means that interventions targeting one diversification mechanism (e.g., boosting somatic hypermutation) may have unintended effects on others (e.g., reducing class switching) because they compete for the same cellular resources.
Why for Yiru: B cells and tertiary lymphoid structures (TLS) are increasingly recognized as important components of the TME, with B cell infiltration and TLS formation associated with favourable responses to immunotherapy in multiple cancer types. Understanding how B cells diversify their antibody responses is directly relevant to anti-tumour B cell immunity: tumour-infiltrating B cells may generate antibodies against tumour antigens, and the mechanisms of diversification described here — class switching, affinity maturation, memory formation — determine the quality and durability of that antibody response. The parallel diversification model suggests that effective anti-tumour B cell responses may require simultaneous engagement of multiple diversification pathways, which has implications for designing B cell-targeted immunotherapies or cancer vaccines. The malaria model used here provides a well-characterized system for studying B cell diversification under strong immune stimulation, and the analytical approaches for tracking clonal diversification could be adapted to study B cell responses in TLS within tumours.
Functional and dysfunctional T regulatory cell states in human tissues in RA and other autoimmune arthritic diseases
Nature Immunology Published 2026-06-16 research article DOI: 10.1038/s41590-026-02540-4
regulatory T cell autoimmune disease rheumatoid arthritis synovial tissue Treg dysfunction tissue immunology microenvironment
Summary: Identifies two predominant tissue-resident T regulatory cell (Treg) subsets in synovial joints from patients with rheumatoid arthritis and other autoimmune arthritic diseases: a suppressive CD25hi CXCR6-positive subset and a dysfunctional CD25lo AREG-positive subset. Tregs are essential for maintaining immune tolerance and preventing autoimmunity, and their dysfunction is thought to contribute to autoimmune disease pathogenesis. However, whether Treg dysfunction in autoimmunity reflects intrinsic Treg defects or microenvironment-driven functional modulation has been unclear. By profiling Tregs directly from inflamed synovial tissue, the authors show that these two subsets are shaped by distinct local microenvironmental cues: the suppressive subset is maintained by macrophages, while the dysfunctional subset is induced by fibroblasts. The dysfunctional CD25lo AREG-positive Tregs retain FoxP3 expression (identifying them as bona fide Tregs) but have lost suppressive function and instead produce amphiregulin (AREG), an epidermal growth factor family member that can promote tissue remodeling and potentially contribute to the hyperplastic synovial tissue characteristic of rheumatoid arthritis. The study demonstrates that Treg dysfunction in human autoimmune disease is not necessarily due to intrinsic Treg defects but can be actively imposed by the local tissue microenvironment — specifically, by fibroblast-derived signals that reprogram Tregs from a suppressive to a tissue-remodeling phenotype. This microenvironment-driven model of Treg dysfunction has therapeutic implications: rather than trying to boost Treg numbers or function systemically, strategies that target the local signals driving Treg dysfunction (e.g., fibroblast-Treg interactions) may be more effective.
Why it matters: Treg-based therapies — adoptive transfer of expanded Tregs or low-dose IL-2 to boost endogenous Tregs — have shown promise in autoimmunity and transplantation but have had limited and variable efficacy. This study provides a potential explanation: simply increasing Treg numbers may not restore tolerance if the local tissue microenvironment converts those Tregs into a dysfunctional state. The identification of specific microenvironmental drivers — macrophages maintaining suppressive Tregs vs. fibroblasts inducing dysfunctional Tregs — provides cellular targets for therapeutic intervention. Rather than systemic Treg boosting, local modulation of the synovial fibroblast-Treg interaction could restore Treg function in the joints without systemic immunosuppression. The finding that dysfunctional Tregs produce AREG, a tissue-remodeling factor, is particularly interesting: it suggests that "dysfunctional" Tregs are not simply passive but may actively contribute to tissue pathology. This framework — microenvironment-driven immune cell functional reprogramming — likely extends beyond Tregs in RA to other immune cells in other inflammatory and autoimmune conditions.
Why for Yiru: The concept that the local tissue microenvironment can reprogram immune cells from a functional to a dysfunctional state is directly relevant to the TME. Tumour-infiltrating Tregs are typically considered immunosuppressive and detrimental to anti-tumour immunity, but this study raises the possibility that the TME may also contain functionally heterogeneous Treg populations — some truly suppressive, others reprogrammed to alternative functions by tumour-derived signals. The macrophage-vs-fibroblast axis identified here (macrophages support suppressive Tregs, fibroblasts induce dysfunctional Tregs) has a direct parallel in the TME, where tumour-associated macrophages and cancer-associated fibroblasts are key stromal populations that shape immune cell function. Profiling Treg heterogeneity and function in the TME across different tumour types and spatial contexts could reveal whether similar microenvironment-driven Treg reprogramming occurs in cancer and whether it affects immunotherapy response. The AREG connection to tissue remodeling is also relevant — AREG is implicated in tumour-promoting tissue remodeling in several cancer types.
mRNA-based influenza vaccine expands the B cell response breadth in humans
Nature Immunology Published 2026-06-15 research article DOI: 10.1038/s41590-026-02569-5
mRNA vaccine influenza B cell antibody response germinal center memory B cell vaccine comparison
Summary: Compares the humoral immune responses of individuals who received conventional inactivated influenza vaccines to those who received an mRNA-based quadrivalent influenza vaccine, revealing that the mRNA platform induces qualitatively superior B cell responses. Seasonal influenza vaccines — traditionally inactivated virus or recombinant protein formulations — provide variable protection that is often modest, particularly in elderly populations and against drifted strains. The mRNA vaccine platform that proved transformative for COVID-19 is now being applied to influenza. This study directly compares the B cell response in humans receiving either an inactivated influenza vaccine or an mRNA-based quadrivalent influenza vaccine encoding haemagglutinin from four seasonal strains. The mRNA vaccine induced prolonged germinal center responses — the specialized microanatomical structures in lymph nodes where B cells undergo affinity maturation and class switching. This prolonged germinal center activity translated into quantitatively more antibodies and, critically, increased numbers of memory B cells. The expansion of the memory B cell compartment is particularly important because memory B cells are the substrate for rapid, high-quality antibody responses upon subsequent exposure to the virus. The mRNA vaccine also induced antibodies with broader reactivity against variant influenza strains, suggesting that the mRNA platform may overcome one of the major limitations of current influenza vaccines: strain-specificity that requires annual reformulation. The study provides mechanistic evidence that mRNA vaccines' superior immunogenicity — well-established for COVID-19 — extends to influenza through enhanced germinal center biology and memory B cell generation.
Why it matters: Influenza remains a major global health burden, causing 290,000–650,000 deaths annually, and current vaccines provide incomplete and variable protection. The demonstration that mRNA-based influenza vaccines induce prolonged germinal center responses, more antibodies, and more memory B cells compared to conventional inactivated vaccines in humans is a significant step toward next-generation influenza vaccines. The expansion of the memory B cell compartment is the key mechanistic insight: conventional inactivated vaccines primarily boost existing antibody titers without substantially expanding the memory B cell pool, which may explain why protection wanes rapidly. mRNA vaccines, by driving sustained germinal center reactions, generate new memory B cells that can respond to future exposures — including potentially to drifted strains not exactly matching the vaccine. The broader reactivity against variant strains addresses the fundamental limitation of current strain-matched influenza vaccines and raises the possibility of more universal influenza vaccine approaches using the mRNA platform. The study also reinforces the general principle that mRNA vaccines are not just a delivery platform but actively shape the quality of the immune response through sustained antigen expression and innate immune stimulation.
Why for Yiru: The mechanistic insights from this vaccine comparison — prolonged germinal center responses driving memory B cell expansion and antibody breadth — are directly relevant to cancer vaccine development. Therapeutic cancer vaccines face similar challenges to influenza vaccines: they need to generate durable, high-quality immune responses against antigens that may vary (neoantigens can be lost through immune editing, analogous to antigenic drift). The finding that mRNA vaccines expand the memory B cell compartment is particularly important for cancer vaccines, where durable immune memory is essential for preventing late relapse. The prolonged germinal center response induced by mRNA vaccines may also be beneficial for generating high-affinity antibodies against tumour antigens, which are often self-antigens with low inherent immunogenicity. The comparison framework — directly benchmarking mRNA vs. conventional vaccine platforms in humans with detailed immune profiling — provides a template for evaluating different cancer vaccine platforms in early-phase clinical trials.
Recurrent COPA mutation drives R-spondin-independent Wnt activation in intestinal tumors
Nature Genetics Published 2026-06-12 research article DOI: 10.1038/s41588-026-02616-9
Wnt signaling COPA mutation intestinal tumor small intestine cancer genetics R-spondin oncogenesis
Summary: Identifies recurrent in-frame deletions in exon 13 of COPA — a gene encoding a subunit of the COPI coatomer complex involved in Golgi-to-ER retrograde transport — as a driver of small intestinal tumorigenesis through activation of Wnt signaling via a mechanism that is independent of R-spondin but dependent on Wnt ligands. Wnt/β-catenin signaling is a central oncogenic pathway in colorectal and small intestinal cancers, typically activated by loss-of-function mutations in APC or gain-of-function mutations in CTNNB1 (β-catenin), or by R-spondin fusions that potentiate Wnt receptor activity. COPA mutations represent a novel mechanism of Wnt pathway activation: the in-frame deletions in COPA exon 13 cause aberrant COPI complex function, which in turn leads to increased secretion of Wnt ligands through a mechanism involving disrupted intracellular protein trafficking. The Wnt ligands then act in an autocrine or paracrine manner to activate Wnt signaling in the tumour cells. This R-spondin-independent, Wnt ligand-dependent mechanism is distinct from the APC/CTNNB1/R-spondin mechanisms that dominate colorectal cancer and represents a new oncogenic route to Wnt activation specifically in the small intestine. The study expands the genetic landscape of intestinal tumorigenesis and identifies COPA as a potential therapeutic target or biomarker for small intestinal tumours, which are rare but often aggressive and poorly understood compared to colorectal cancers.
Why it matters: Wnt pathway activation is nearly universal in intestinal cancers, but the specific genetic mechanism of activation matters for both prognosis and therapy. APC-mutant colorectal cancers respond differently to Wnt pathway inhibitors than R-spondin-fusion tumours, and this study adds a third mechanism — COPA mutation-driven Wnt ligand secretion — that may require yet different therapeutic strategies. The COPI coatomer connection to Wnt signaling is mechanistically novel: it links a fundamental intracellular trafficking complex to oncogenic signaling in a way not previously appreciated. This discovery also highlights the value of studying rare tumour types (small intestinal cancer) for uncovering new cancer mechanisms that may have broader relevance. The Wnt ligand-dependent nature of the activation suggests that these tumours may be sensitive to Wnt ligand inhibitors (e.g., PORCN inhibitors that block Wnt secretion) rather than downstream pathway inhibitors, providing a potential targeted therapy for what is currently a therapeutically neglected cancer type.
Why for Yiru: Wnt signaling is not only a driver of tumour cell proliferation but also a key regulator of immune cell function in the TME. Wnt ligands produced by tumour cells can suppress dendritic cell function, exclude T cells from the tumour, and promote immunosuppressive macrophage polarization — contributing to the "immune-excluded" phenotype characteristic of Wnt-activated tumours. The COPA mutation mechanism — increased Wnt ligand secretion due to trafficking defects — would be expected to create a Wnt-rich TME that actively suppresses anti-tumour immunity. From a TME perspective, COPA-mutant tumours might represent a subset of Wnt-activated cancers with particularly high Wnt ligand levels, making them both dependent on Wnt signaling for tumour cell-intrinsic growth and actively immunosuppressive through paracrine Wnt effects on immune cells. This dual role makes Wnt ligand inhibitors particularly attractive for COPA-mutant tumours, as they could simultaneously target tumour cell growth and relieve immune suppression — a combination that could synergize with immunotherapy.
Activity-dependent adaptive deep brain stimulation improves gait in Parkinson's disease
Nature Medicine Published 2026-06-15 research article DOI: 10.1038/s41591-026-04432-4
Parkinson's disease deep brain stimulation adaptive therapy gait neural decoding subthalamic nucleus neuromodulation
Summary: Demonstrates that real-time decoding of ongoing locomotor activities from subthalamic nucleus neural dynamics can steer activity-dependent adaptive deep brain stimulation (aDBS) to improve gait deficits in Parkinson's disease. Parkinson's patients experience a spectrum of locomotor deficits — including shuffling gait, freezing of gait, and postural instability — that vary with daily activities and the fluctuating physiology of the disease. Conventional DBS delivers continuous, fixed-parameter stimulation optimized for cardinal motor symptoms (tremor, rigidity, bradykinesia) but is activity-agnostic, meaning the same stimulation is delivered whether the patient is sitting, walking on smooth ground, or navigating stairs — despite different neural control requirements for each activity. The authors identify physiological principles that enable real-time decoding of different locomotor activities (standing, walking, turning, stair climbing) from the neural dynamics of the subthalamic nucleus. This decoded activity information is used to adapt DBS parameters in real time: different stimulation patterns are delivered for different activities, optimized for the specific locomotor demands. The adaptive approach improved locomotor deficits while preserving or enhancing the benefits for cardinal motor symptoms, effectively expanding the therapeutic range of DBS beyond what fixed-parameter stimulation can achieve. A companion paper provides additional clinical characterization and demonstrates the feasibility of long-term aDBS use.
Why it matters: Deep brain stimulation has been a transformative therapy for Parkinson's disease, but its benefits for gait and balance — the symptoms that most affect quality of life and independence — have been limited and variable. Gait is a complex behaviour requiring dynamic coordination of posture, rhythm, and steering, and the neural control signals likely differ from those governing simpler movements like wrist flexion that are used to optimize DBS parameters. The concept of activity-dependent stimulation — delivering different stimulation patterns for different activities — is a logical extension of the closed-loop DBS paradigm that has shown promise for tremor and rigidity. Demonstrating that neural activity in the subthalamic nucleus contains sufficient information to decode different locomotor states in real time, and that using this information to adapt stimulation improves gait, is a significant technical and clinical advance. The approach moves DBS from a "one-size-fits-all" therapy to a personalized, context-aware intervention that adapts to the patient's moment-to-moment needs — analogous to how a healthy basal ganglia system modulates its output based on behavioural context.
Why for Yiru: The adaptive stimulation paradigm — decoding physiological state from neural signals and adapting therapy in real time — has conceptual parallels to adaptive immunotherapy, where treatment could be modulated based on real-time monitoring of the TME. Just as DBS parameters are adjusted based on decoded locomotor state, immunotherapy dosing, combination, or timing could be adjusted based on biomarkers of TME state (e.g., T cell activation, cytokine levels, tumour antigen presentation) measured through liquid biopsies or implantable sensors. The neural decoding framework — extracting behaviourally-relevant signals from complex neural population activity — is analogous to extracting clinically-relevant signals from complex immune monitoring data. The demonstration that activity-specific stimulation outperforms continuous stimulation also provides a general principle: "more" is not always better, and context-appropriate intervention can achieve outcomes that fixed interventions cannot.
A progeria syndrome links DNA hypermethylation to age-related pathology
Nature Genetics Published 2026-06-12 research article DOI: 10.1038/s41588-026-02633-8
epigenetics DNA methylation aging progeria DNMT3A stem cell age-related disease
Summary: Describes Heyn-Sproul-Jackson syndrome, a newly characterized epigenetically driven accelerated aging syndrome caused by germline DNMT3A gain-of-function mutations, establishing a causal link between DNA hypermethylation and tissue aging. Declining tissue function and regenerative capacity underlie many chronic diseases of aging, but establishing whether age-associated epigenetic changes — particularly gains in DNA methylation at specific genomic loci — are causal drivers or passive correlates of aging has been a central unanswered question. The identification of a human progeria syndrome driven by DNMT3A gain-of-function provides a natural experiment to answer this question. DNMT3A is a de novo DNA methyltransferase, and the mutations identified in this syndrome increase its catalytic activity, leading to accelerated, ectopic DNA methylation that recapitulates the age-related methylation patterns seen in normal aging — but on a dramatically compressed timescale. Patients exhibit multilineage stem cell dysfunction affecting haematopoietic, mesenchymal, and other tissue stem cell compartments, phenocopying key aspects of physiological aging including immune dysfunction, bone marrow failure, and tissue atrophy. Mouse models carrying the homologous Dnmt3a mutations recapitulate the human phenotype, demonstrating that the accelerated DNA methylation is sufficient to drive the aging-like pathology. The study establishes that DNA hypermethylation is not merely a biomarker of aging but a causal driver of age-related stem cell decline and tissue dysfunction — a finding with profound implications for understanding and potentially intervening in the biology of aging.
Why it matters: The question of whether epigenetic changes cause aging or merely reflect it has been one of the most debated topics in aging research. Epigenetic clocks can accurately predict chronological age, and age-related DNA methylation changes are remarkably consistent across individuals and tissues, but correlation is not causation. This study provides the strongest evidence to date that DNA hypermethylation — specifically, DNMT3A-driven gains in methylation — is causally sufficient to drive tissue aging and stem cell dysfunction. The identification of a human syndrome that compresses decades of epigenetic aging into childhood provides both mechanistic insight and a model system for studying epigenetic aging. The demonstration that the same mutations cause aging-like phenotypes in mice opens the door to preclinical testing of interventions that target DNMT3A activity or reverse pathological DNA methylation. The connection between DNMT3A and aging also has implications for clonal haematopoiesis: DNMT3A loss-of-function mutations are the most common mutations in age-related clonal haematopoiesis (CHIP), and this study shows that the opposite — gain-of-function — causes an even more dramatic aging phenotype, highlighting the exquisite sensitivity of stem cell function to DNMT3A dosage.
Why for Yiru: Epigenetic aging is increasingly recognized as relevant to cancer biology, including in the TME. The age-related decline in immune function (immunosenescence) is thought to contribute to increased cancer incidence and reduced immunotherapy efficacy in elderly patients, and epigenetic changes in immune cells may be a key mechanism. DNMT3A is particularly relevant: it is the most commonly mutated gene in CHIP, and CHIP is associated with altered immune function and potentially with immunotherapy outcomes. The demonstration that DNMT3A gain-of-function drives stem cell dysfunction through hypermethylation suggests that DNMT3A activity must be precisely calibrated for normal stem cell function — too little (CHIP) or too much (this progeria syndrome) both cause pathology. In the TME, age-related epigenetic changes in tumour-infiltrating immune cells could affect their function, and DNMT3A activity in both tumour cells and immune cells may influence the TME through methylation-dependent gene regulation. Understanding how DNMT3A-driven methylation affects immune cell function could inform strategies to rejuvenate aged immune cells for more effective immunotherapy in elderly cancer patients.