Research Radar — 2026-06-10
Methods & AI
Computational
Diminishing returns: scaling up training dataset size for single-cell transcriptomic AI models
Nature Methods Published 2026-06-09 research article DOI: 10.1038/s41592-026-03120-y
single-cell foundation model transcriptomics pretraining scaling deep learning computational biology benchmarking
Summary: Two companion papers in Nature Methods systematically evaluate whether scaling up pretraining dataset size improves single-cell transcriptomic foundation model performance. Single-cell foundation models — large transformer-based models pretrained on tens of millions of single-cell transcriptomes — have become increasingly central to single-cell analysis, with the implicit assumption that more pretraining data yields better representations. DenAdel, Hughes, et al. rigorously test this assumption by training models across a range of dataset sizes (from millions to tens of millions of cells), while varying dataset diversity and composition. They find that performance gains from increasing dataset size plateau rapidly: beyond a certain threshold (roughly 5–10 million cells for current architectures), additional data provides negligible improvement in downstream tasks such as cell-type annotation, batch integration, and gene program discovery. The companion study reinforces this finding and further shows that dataset diversity — the inclusion of cells from diverse tissues, species, and experimental protocols — matters more than raw cell count for model generalization. Together, these studies challenge the prevailing "bigger is better" paradigm and suggest that the next frontier for single-cell AI is not simply collecting more data, but developing architectures and training strategies that can extract richer representations from existing datasets.
Why it matters: These findings have immediate practical implications for the single-cell genomics community. If diminishing returns set in well below the scale of the largest existing datasets (which contain hundreds of millions of cells), then simply aggregating more data is not the path to better models. This redirects research effort toward architectural innovations, better pretraining objectives, and strategies for incorporating biological prior knowledge. For the many labs building or fine-tuning single-cell foundation models, these results provide evidence-based guidance on how much data is "enough" — potentially saving enormous computational resources. The finding that diversity matters more than size also argues for more thoughtful dataset curation rather than indiscriminate data aggregation.
Why for Yiru: For TME single-cell analysis, these results suggest that a well-curated dataset of diverse TME samples (across cancer types, treatment conditions, and spatial contexts) may be more valuable than simply pooling all available single-cell data. A TME-specific foundation model trained on a carefully selected, diverse set of tumour-immune-stromal samples could outperform a larger model trained on all of GEO. The benchmarking framework from these studies also provides a template for evaluating how well foundation models capture TME-specific biology — cell states, cell-cell communication programs, and spatial organization patterns — which are often underrepresented in general-purpose models.
Immune BioGraphy: A tale of graphical approaches in systems and virtual immunology
Cell Systems Published 2026-06-05 perspective DOI:
graph machine learning systems immunology virtual cell knowledge graph multimodal immune system computational biology network biology
Summary: Presents a comprehensive perspective on how graph-based machine learning provides a unifying framework for modeling the multiscale complexity of the immune system. The immune system spans molecular (receptor-ligand binding, signaling cascades), cellular (cell-type identification, state transitions), tissue (cell-cell communication, spatial organization), and organismal (immune responses, disease outcomes) scales. Current computational approaches tend to operate at a single scale, making it difficult to trace how molecular perturbations propagate to systemic outcomes. Graph ML naturally bridges these scales: molecules, cells, and tissues can all be represented as nodes in heterogeneous graphs, with edges encoding physical interactions, regulatory relationships, or spatial proximity. The authors review three key graph ML paradigms — graph neural networks for learning on structured biological data, knowledge graphs for integrating diverse biomedical databases and literature, and graph-based language models for reasoning over biological text corpora. They argue that combining these with multimodal data integration (transcriptomics, proteomics, imaging) will enable the construction of "virtual immune cells" — predictive computational models that simulate immune responses to perturbations, from cytokine stimulation to checkpoint blockade.
Why it matters: The immune system is inherently a graph-like system: cells communicate through receptor-ligand interactions, signal through kinase cascades, and organize into spatial networks within tissues. Graph ML is perhaps the most natural computational framework for modeling such systems, yet it remains underutilized compared to sequence- and image-based deep learning. This perspective makes a compelling case that graph-based approaches are not just an alternative method but the right abstraction for immunology. The vision of virtual immune cells — models that can predict how a specific perturbation (a drug, a mutation, a cytokine) will alter immune function — represents a grand challenge that could transform drug development, personalized medicine, and basic immunology research.
Why for Yiru: This perspective is directly relevant to computational TME research. The TME is a spatially organized multicellular system that is naturally represented as a graph: nodes are cells (tumour, immune, stromal) and edges represent physical interactions (ligand-receptor), cytokine-mediated communication, or metabolic coupling. Graph ML could model how specific TME features — checkpoint expression patterns, immune infiltration gradients, metabolic gradients — collectively determine immunotherapy response. Knowledge graphs integrating TME literature, drug-target databases, and clinical trial outcomes could identify novel therapeutic targets. The framework also provides a conceptual bridge between Yiru's computational methods work and translational TME biology — graph representations can encode both molecular mechanisms and clinical phenotypes.
Clonal lineage tracing of innate immune cells in human cancer with Mitotrek
Cancer Cell Published 2026-06-04 research article DOI:
clonal lineage tracing mitochondrial DNA myeloid cells tumour microenvironment single-cell multi-omics epigenetics cancer immunology
Summary: Introduces Mitotrek, a computational method that leverages mitochondrial DNA (mtDNA) mutations as endogenous clonal barcodes to trace the lineage origins of innate immune cells in human tumours. A fundamental question in tumour immunology is whether tumour-associated myeloid cells — macrophages, dendritic cells, monocytes — are recruited from the circulation and subsequently educated by the TME, or whether they arise from tissue-resident precursors with pre-existing functional biases. Answering this requires clonal lineage tracing, but traditional methods (genetic barcoding, viral lineage tracing) cannot be applied to human tumours. Mitotrek solves this by exploiting the fact that mtDNA accumulates somatic mutations at a high rate, and these mutations are stably inherited during cell division, creating naturally occurring clonal barcodes. The authors combine single-cell multi-omics (scRNA-seq + scATAC-seq + mtDNA genotyping) to simultaneously measure transcriptomic state, chromatin accessibility, and clonal identity in individual tumour-infiltrating myeloid cells from human cancer samples. Using Mitotrek, they discover that tumour myeloid populations predominantly arise from circulating monocytes that already carry epigenetic biases toward specific functional states before entering the tumour — suggesting that tumour immunity is shaped by pre-existing immune programming rather than solely by local TME education.
Why it matters: This study challenges a central dogma in tumour immunology — that tumour-associated macrophages and other myeloid cells are primarily shaped by local TME signals after infiltration. The finding that circulating monocytes carry pre-existing epigenetic biases suggests that the systemic immune state of the patient — influenced by factors such as chronic inflammation, prior therapies, and systemic disease — is a major determinant of TME composition. This has therapeutic implications: interventions that reprogram the systemic myeloid compartment (rather than only targeting the TME locally) may be more effective. Mitotrek itself is a methodological advance that makes clonal lineage tracing feasible in any human tissue sample with single-cell multi-omics data, opening the door to lineage-resolved analyses across cancer types and other diseases.
Why for Yiru: Mitotrek is directly applicable to Yiru's TME research. Understanding the clonal origins of TME immune cells could reveal which cell states are targetable by systemic vs. local therapies. Computationally, mtDNA-based lineage tracing could be integrated with spatial transcriptomics to trace how clonally related myeloid cells distribute within the TME, and with ligand-receptor analysis to understand whether clonally related cells engage in similar communication patterns. The method could also be applied to T cell and NK cell populations in the TME, potentially revealing whether exhausted T cell clones arise from pre-existing exhausted precursors or from naive cells that become exhausted within the tumour.
Deep learning of functional perturbations from condensate morphology
Cell Published 2026-06-04 research article DOI:
biomolecular condensates deep learning nucleolus phase separation morphology computer vision ribosome biogenesis TOP1
Summary: Develops a deep learning framework that quantitatively decodes the morphology of biomolecular condensates — membrane-less organelles formed by liquid-liquid phase separation — to infer underlying biochemical states and functional perturbations. Biomolecular condensates such as the nucleolus, nuclear speckles, and stress granules are increasingly recognized as key organizers of cellular biochemistry, but it has been difficult to connect their visually observable morphological features (size, shape, number, internal texture) to the specific molecular processes occurring within them. The authors train convolutional neural networks on high-content microscopy images of condensates under diverse genetic and pharmacological perturbations, enabling the model to classify the perturbation type from condensate morphology alone. Applying this framework to the nucleolus, they discover a previously unrecognized role for topoisomerase I (TOP1) in ribosome biogenesis and nucleolar organization — a finding that emerged from the model's ability to detect subtle morphological changes invisible to human observers. The method is generalizable across condensate types, including nucleoli, nuclear speckles, and viral replication factories.
Why it matters: This work represents an elegant integration of deep learning with cell biology, demonstrating that AI can extract biological insight from image data that human experts cannot perceive. More broadly, it shows that condensate morphology is not merely a phenotype but a rich, quantitative readout of cellular state — a "morphological phenotype" that can be systematically linked to molecular mechanisms. The discovery of TOP1's role in ribosome biogenesis illustrates the power of this approach for uncovering unexpected biology. As condensate biology matures, methods for quantitatively connecting condensate morphology to function will be essential for drug discovery (many diseases involve condensate dysregulation) and for understanding how cells organize their internal biochemistry.
Why for Yiru: Biomolecular condensates are increasingly implicated in TME biology — stress granules in hypoxic tumour regions, signaling condensates at immune synapses, and transcriptional condensates at oncogenic super-enhancers. A deep learning approach for decoding condensate morphology could be applied to TME-relevant condensates, potentially revealing how TME conditions (hypoxia, nutrient stress, cytokine signaling) alter condensate function. The computer vision framework could also be extended to other cellular morphological features in the TME — for example, decoding how macrophage morphology relates to polarization state, or how tumour cell shape relates to invasive potential.
Spatially resolved single-cell analyses of human meningioma identify novel cell states influencing tumor microenvironment and progression
Nature Genetics Published 2026-06-09 research article DOI: 10.1038/s41588-026-02615-w
meningioma single-cell multi-omics tumour microenvironment spatial transcriptomics brain tumour tumour heterogeneity
Summary: Presents a comprehensive single-cell multi-omic atlas of human meningioma comprising more than seven million cells profiled across tumour grades and subtypes. Meningiomas are the most common primary brain tumours, and while typically benign, a subset are aggressive and resistant to treatment. The authors integrate single-cell RNA-seq, single-cell ATAC-seq, and spatial transcriptomics to map the cellular landscape of meningiomas at unprecedented resolution, revealing previously unrecognized cell states within both the malignant and stromal compartments. Key findings include: identification of a tumour cell state characterized by a wound-healing transcriptional program associated with aggressive clinical behaviour; discovery of an immunosuppressive macrophage population enriched in high-grade tumours that expresses multiple checkpoint ligands; and spatial mapping showing that these macrophages co-localize with exhausted T cells in immune-excluded regions. The study also identifies chromatin accessibility differences between tumour grades that suggest epigenetic mechanisms underlying malignant progression.
Why it matters: Meningiomas have been relatively understudied compared to gliomas, despite being more common. This atlas provides a foundational resource for the meningioma field and demonstrates the power of multi-omic single-cell approaches for understanding tumour biology in a clinically relevant context. The identification of an immunosuppressive macrophage population that co-localizes with exhausted T cells in immune-excluded niches mirrors findings in other tumour types and suggests that immunotherapy strategies effective in other cancers might be applicable to aggressive meningiomas. The chromatin accessibility data also provide a resource for identifying regulatory elements that could be targeted therapeutically.
Why for Yiru: The multi-omic approach used in this study — integrating scRNA-seq, scATAC-seq, and spatial transcriptomics — is exactly the type of analysis Yiru might apply to other TME contexts. The finding of spatial co-localization between immunosuppressive macrophages and exhausted T cells is a recurring TME pattern that could be systematically studied across cancer types using similar multi-omic approaches. The computational methods for integrating these data modalities, identifying cell states, and mapping them spatially are transferable to any TME study, and the meningioma atlas provides a useful comparative reference for brain tumour immunology.
Biomedical discoveries
Biomedicine
Widespread transcriptional memory shapes heritable states and functional heterogeneity in cancer and stem cells
Cell Systems Published 2026-06-04 research article DOI:
transcriptional memory epigenetics cancer heterogeneity stem cells lineage tracing single-cell heritability
Summary: Using lineage-resolved single-cell transcriptomics, the authors demonstrate that transcriptional memory — the heritable propagation of gene expression states across cell divisions independent of genetic mutations or ongoing extracellular signals — is widespread in both cancer and stem cell systems, defining previously hidden layers of functional heterogeneity. While it is well established that genetic heterogeneity contributes to cancer evolution and drug resistance, the role of non-genetic, heritable transcriptional states has been difficult to study because standard single-cell profiling cannot distinguish between states that are actively maintained by the environment and those that are epigenetically "remembered" through cell division. The authors solve this by combining single-cell transcriptomics with clonal lineage tracing, enabling them to track gene expression changes along cell division trees and distinguish heritable (memory) from non-heritable (environmental) variation. They identify hundreds of "memory genes" whose expression is stably propagated across multiple cell generations, including genes involved in drug resistance, proliferation, and differentiation. These memory genes are partially conserved between cancer and stem cell systems, predict cancer-relevant phenotypes such as drug sensitivity and metastatic potential, and point to specific epigenetic mechanisms — including DNA methylation and histone modifications — that stabilize memory states.
Why it matters: This study adds an important dimension to our understanding of cancer heterogeneity. The standard model of cancer evolution focuses on genetic mutations and clonal selection, but this work shows that heritable transcriptional states — essentially "epigenetic memory" — create an additional layer of functional diversity that can contribute to drug resistance, metastasis, and tumour recurrence. The finding that memory genes are partially conserved between cancer and normal stem cells suggests that tumours hijack normal developmental memory mechanisms for malignant purposes. Importantly, transcriptional memory is potentially reversible through epigenetic therapies, unlike genetic mutations — making memory states attractive therapeutic targets.
Why for Yiru: Transcriptional memory may explain why some TME cell states — such as T cell exhaustion, macrophage polarization, or CAF activation — persist even when the initiating signals are removed. For example, exhausted T cells might "remember" their exhausted state through epigenetic mechanisms, explaining why checkpoint blockade alone is often insufficient for durable responses. Computational methods for identifying memory genes from single-cell lineage tracing data, and for predicting which memory states are therapeutically reversible, represent an opportunity for computational TME research. Integrating transcriptional memory analysis with spatial data could reveal whether memory states are spatially organized within the TME.
Single-cell spatial pharmacobiology identifies conserved stromal barriers to therapeutic antibody delivery in human solid tumors
Nature Biotechnology Published 2026-06-03 research article DOI: 10.1038/s41587-026-03152-x
spatial proteomics antibody delivery tumour microenvironment stromal barrier therapeutic antibody drug delivery pharmacobiology
Summary: Presents single-cell spatial pharmacobiology (SSP), a method that combines in vivo imaging of fluorescently labeled therapeutic antibodies with high-plex spatial proteomics (CODEX) to map, at single-cell resolution, how stromal barriers in the tumour microenvironment impede antibody-based drug delivery in human solid tumours. A major limitation of antibody-based cancer therapies — including checkpoint inhibitors, antibody-drug conjugates, and bispecific antibodies — is that only a small fraction of the administered dose actually reaches tumour cells, with the rest sequestered by stromal components or cleared systemically. The authors apply SSP to clinical samples from head and neck and pancreatic cancer patients treated with fluorescently labeled cetuximab (anti-EGFR) in phase 1 trials. They find that antibody penetration is highly heterogeneous within individual tumours, with penetration depth inversely correlated with the density of cancer-associated fibroblasts (CAFs) and extracellular matrix (ECM) components. Strikingly, the stromal barriers to antibody delivery are conserved across different tumour types and different antibody targets, suggesting a universal mechanism of antibody exclusion. The companion single-cell spatial proteomics data reveal that regions of poor antibody penetration are enriched for specific CAF subtypes and ECM proteins that could be targeted to improve drug delivery.
Why it matters: Poor drug delivery is a major, underappreciated cause of therapeutic failure in solid tumours. Many drugs that show potent activity in vitro or in preclinical models fail in the clinic not because the target is wrong, but because the drug never reaches the target at sufficient concentrations. SSP provides the first systematic, single-cell resolution map of where antibodies go (and don't go) in human tumours, and identifies the specific cellular and molecular barriers responsible. The finding that these barriers are conserved across tumour types and antibody targets suggests that strategies to overcome stromal exclusion — such as CAF depletion, ECM degradation, or antibody engineering for better penetration — could broadly improve the efficacy of antibody-based cancer therapies.
Why for Yiru: This study is directly relevant to TME research and has clear translational implications. The SSP method could be extended to study how other TME-modifying therapies — checkpoint inhibitors, cytokine therapies, cell therapies — distribute within tumours. Computationally, the rich spatial proteomics datasets generated by SSP are ideal for developing predictive models of drug distribution based on local TME composition. One could imagine a computational tool that, given a patient's TME composition map, predicts which regions will be accessible to antibody-based therapies and suggests stromal targeting strategies to improve delivery. The CAF-ECM axis identified here as the primary barrier is also a key determinant of immune exclusion, linking drug delivery to immunotherapy resistance.
Inflammatory cytokines induce new cancer dependencies that sensitize tumours to immune checkpoint blockade
Nature Genetics Published 2026-06-09 research article DOI: 10.1038/s41588-026-02614-x
CRISPR screen inflammatory cytokine interferon cancer dependency immune checkpoint blockade GPI transamidase FITM2 immunotherapy
Summary: Conducts genome-scale CRISPR loss-of-function screens in the presence of inflammatory cytokines — particularly interferon-gamma (IFN-γ) and tumor necrosis factor (TNF) — to systematically map genetic vulnerabilities that are specifically induced by the inflammatory tumour microenvironment. While CRISPR screens have identified many cancer dependencies, they are typically performed under standard culture conditions that lack the inflammatory signals present in the TME. The authors hypothesized that inflammatory cytokines, which are abundant in immunologically "hot" tumours, might create new dependencies not present under basal conditions — genes that become essential only when cancer cells are under cytokine stress. The screen identifies the glycosylphosphatidylinositol (GPI) transamidase complex and the lipid phosphatase FITM2 as two interferon-specific tumour dependencies. GPI transamidase is required for anchoring proteins to the cell surface, and its disruption selectively kills IFN-γ-exposed cancer cells by preventing surface expression of proteins needed to cope with inflammatory stress. FITM2, an ER-resident lipid phosphatase, becomes essential under IFN-γ due to its role in maintaining ER homeostasis during the secretory stress induced by interferon signaling. Critically, genetic or pharmacologic disruption of either target sensitizes tumours to immune checkpoint blockade in vivo, suggesting a therapeutic strategy: combine checkpoint inhibitors with drugs targeting interferon-induced dependencies.
Why it matters: This study introduces the concept of "context-dependent cancer dependencies" — genes that are not essential under standard conditions but become essential in the inflammatory TME. This has profound implications for cancer drug target discovery: traditional CRISPR screens performed in standard culture may miss the very targets that matter most in vivo. The identification of GPI transamidase and FITM2 as interferon-specific dependencies provides two novel, druggable targets for combination with immunotherapy. More broadly, the study provides a framework for discovering context-specific dependencies induced by other TME signals — hypoxia, nutrient deprivation, mechanical stress — which could dramatically expand the repertoire of cancer drug targets.
Why for Yiru: The concept of cytokine-induced dependencies is directly applicable to TME research. Different TME contexts — immunologically hot vs. cold tumours, different cytokine milieus — may create different sets of dependencies that can be therapeutically exploited. Computational methods could be developed to predict, from single-cell or spatial transcriptomics data, which tumours have the cytokine profile that induces specific dependencies, enabling patient stratification for combination therapies. The experimental framework — performing CRISPR screens under TME-relevant conditions — could be extended to include other TME components such as specific immune cell co-cultures, ECM conditions, or hypoxia.
Plasma signals of lung tumor promotion for molecular cancer prevention
Cell Published 2026-06-04 research article DOI:
lung cancer cancer prevention plasma proteomics biomarker IL-1β tumour promotion alveolar transitional state
Summary: Identifies a 14-protein plasma proteomic signature that predicts which individuals benefit from anti-IL-1β-based lung cancer risk reduction, and mechanistically demonstrates that diverse tumour-promoting factors — smoking, air pollution, chronic inflammation — converge on the induction of an alveolar epithelial transitional cell state that underlies lung tumorigenesis. Building on the landmark CANTOS trial which showed that anti-IL-1β (canakinumab) reduces lung cancer incidence, the authors develop a plasma biomarker to identify individuals most likely to benefit from such preventive therapy. Using proteomic profiling of over 10,000 individuals from multiple cohorts, they identify proteins reflecting inflammation, epithelial damage, and immune activation that together define a "tumour promotion" signature. Single-cell analyses of human lung tissue show that this signature reflects the presence of alveolar type II cells in a transitional, damage-associated state that is primed for malignant transformation. The 14-protein panel stratifies individuals by lung cancer risk and, critically, identifies those most likely to benefit from anti-inflammatory prevention — enabling precision cancer prevention.
Why it matters: Cancer prevention has lagged far behind cancer treatment, partly because we lack tools to identify who should receive preventive interventions and to measure whether prevention is working. This study addresses both gaps: the 14-protein signature provides a blood-based test to identify high-risk individuals, and the identification of the alveolar transitional state as the cellular target of tumour promotion provides a mechanistic framework for developing and monitoring preventive therapies. The convergence of diverse carcinogenic exposures on a common transitional cell state also explains why anti-inflammatory approaches like IL-1β blockade can prevent cancers caused by different exposures — they all funnel through the same biology.
Why for Yiru: The concept of a "transitional cell state" that precedes malignant transformation is relevant to understanding the pre-malignant TME — the tissue environment that permits or promotes cancer initiation. Computational methods could be developed to detect transitional cell states in single-cell data from at-risk tissues and to model how TME signals (inflammation, immune surveillance, stromal remodeling) influence the transition from normal to pre-malignant to malignant states. The plasma proteomic signature approach could be extended to other cancer types, potentially identifying blood-based biomarkers that reflect pre-malignant TME changes across tissues.
Pathogenesis of diffuse large B cell lymphoma proteogenotypes
Cancer Cell Published 2026-06-04 research article DOI:
DLBCL lymphoma proteogenomics tumour microenvironment MYC BCR signaling T cell exhaustion proteomics
Summary: Performs integrated transcriptomic and deep proteomic profiling of diffuse large B cell lymphoma (DLBCL), revealing seven proteogenotypes that capture molecular and TME features across established genetic subtypes, including a high-risk subgroup defined by convergent MYC and B cell receptor activity with an exhausted T cell microenvironment. DLBCL is the most common aggressive lymphoma, and while genetic classifications (ABC, GCB, etc.) have improved biological understanding, they do not fully capture the functional proteomic landscape or the TME context that influences treatment response. The authors profile over 10,000 proteins across DLBCL tumours using mass spectrometry-based proteomics, integrated with transcriptomics and clinical outcome data. The proteogenomic analysis identifies seven proteogenotypes (PG1–PG7) that stratify patients independently of genetic subtypes. PG4 is a particularly high-risk group characterized by enhanced MYC transcriptional activity, hyperactive B cell receptor signaling, and an immune microenvironment enriched for exhausted CD8+ T cells — suggesting that these tumours both drive their own proliferation and suppress anti-tumour immunity. Pharmacologic targeting of PG4-associated kinases shows selective toxicity in PG4 cell lines, identifying potential precision therapy approaches.
Why it matters: This study demonstrates that the proteome provides biological and clinical information beyond what can be inferred from the genome or transcriptome. The identification of PG4 as a high-risk group with co-occurring MYC activity, BCR signaling, and T cell exhaustion is a compelling example of how tumour-intrinsic and TME features converge to drive aggressive disease. The finding that PG4 tumours may be selectively vulnerable to kinase inhibitors opens a path to precision therapy for a currently poor-prognosis group. More broadly, the proteogenomic framework is applicable to other cancer types where protein-level information could refine molecular classifications.
Why for Yiru: The integration of tumour-intrinsic proteomics with TME characterization in this study is directly relevant to TME research across cancers. The co-occurrence of MYC/BCR activity with T cell exhaustion in PG4 suggests a mechanistic link between tumour cell signaling and immune suppression that could exist in other cancer types. Computational methods for integrating proteomic and transcriptomic data to identify such tumour-TME coupling could be developed and applied to solid tumour TMEs. The PG4 kinase dependency finding also illustrates how proteomic data can identify druggable vulnerabilities that are not apparent from genomic data alone.
Cross-disciplinary watchlist
Other Fields
Simultaneous two- and three-photon multiplane imaging across cortical layers in freely moving mice
Nature Methods Published 2026-06-09 research article DOI: 10.1038/s41592-026-03125-7
multiphoton microscopy neuroscience in vivo imaging cortical layers freely moving neural activity calcium imaging
Summary: Presents a lightweight head-mounted multiplane microscope that combines two-photon and three-photon excitation delivered through multiple optical fibers to enable near-simultaneous imaging of neuronal activity from five vertically separated planes spanning cortical layers 2/3 through 5 in freely moving mice performing complex behavioral tasks. Current head-mounted microscopes for freely moving animals are limited to single-plane imaging, which means researchers can only monitor neurons in one cortical layer at a time — a severe limitation given that cortical computation involves coordinated activity across layers. The authors solve this by using separate optical fibers for each imaging plane, with two-photon excitation for superficial layers (where scattering is minimal) and three-photon excitation for deep layers (where longer wavelengths provide better penetration). The system weighs only a few grams, allowing mice to perform natural behaviors including social interactions, navigation, and operant tasks while recording from over 1,800 neurons simultaneously across cortical depths, sampled over weeks.
Why it matters: Recording neural activity across cortical layers during natural behavior has been a longstanding goal in systems neuroscience. Cortical circuits process information through feedforward and feedback connections between layers, and understanding these computations requires simultaneous access to all layers. This technology finally makes that possible in freely behaving animals, opening the door to studying how sensory, motor, and cognitive processes are implemented across cortical microcircuits in naturalistic contexts. The combination of two- and three-photon excitation is an elegant solution to the depth-range trade-off that has limited previous approaches.
Why for Yiru: While primarily a neuroscience tool, the multiplane imaging concept could inspire analogous approaches for imaging the TME across depth. Solid tumours have depth-dependent variations in oxygenation, immune infiltration, and drug penetration, and methods for simultaneous multi-depth imaging in living tissue could reveal how these gradients interact. The fiber-based multi-excitation approach might be adapted for intravital TME imaging, potentially enabling simultaneous recording from superficial and deep tumour regions during therapy.
Spatially resolved m6A profiling using m6A-ARTR-DBiT
Nature Methods Published 2026-06-09 research article DOI: 10.1038/s41592-026-03123-9
m6A epitranscriptomics spatial transcriptomics RNA modification DBiT tissue section N6-methyladenosine
Summary: Introduces m6A-ARTR-DBiT, a method that combines antibody-based detection of N6-methyladenosine (m6A) — the most abundant internal mRNA modification — with the DBiT spatial barcoding platform to profile the spatial distribution of m6A across the transcriptome in intact tissue sections. m6A plays critical roles in mRNA stability, translation, and localization, and its dysregulation is implicated in cancer, neurological disorders, and development. However, existing methods for m6A profiling (such as MeRIP-seq) require homogenized tissue and lose all spatial information. m6A-ARTR-DBiT solves this by: (1) applying m6A-specific antibodies to intact tissue sections to capture methylated RNA fragments, (2) using microfluidic DBiT barcoding to encode spatial coordinates onto the captured fragments, and (3) sequencing to generate spatially resolved m6A maps. The authors demonstrate the method in mouse brain and human cancer tissues, revealing spatially patterned m6A modification that correlates with local cell-type composition, gene expression programs, and histological features — including m6A gradients at tumour-stroma interfaces.
Why it matters: This represents the first method for spatially resolved epitranscriptomic profiling, adding a new layer of information to the spatial biology toolkit. While spatial transcriptomics and spatial proteomics have revealed how gene expression and protein abundance vary across tissues, m6A-ARTR-DBiT now enables investigation of how post-transcriptional RNA modifications are spatially organized. Given the importance of m6A in regulating oncogene expression, immune signaling, and stress responses, spatially resolved m6A maps could reveal previously invisible regulatory mechanisms in the TME — for example, whether tumour cells at the invasive front have different m6A landscapes than those in the tumour core.
Why for Yiru: Spatial epitranscriptomics is directly relevant to TME biology. RNA modifications like m6A regulate key TME processes: immune checkpoint expression (PD-L1 mRNA stability is m6A-regulated), cytokine production, hypoxia response, and metabolic adaptation. Spatially mapping m6A in tumour sections could reveal how the TME shapes the epitranscriptome — for example, whether hypoxic regions have distinct m6A patterns that promote survival or immune evasion. Computationally, integrating spatial m6A data with spatial transcriptomics and proteomics would require new methods for multi-modal spatial data integration, an area ripe for computational innovation.
SAGE-net: A scalable approach to sequence-to-function predictions from personal genomes
Nature Methods Published 2026-06-08 research article DOI: 10.1038/s41592-026-03124-8
sequence-to-function personal genome gene expression regulatory grammar deep learning genomics variant effect prediction
Summary: Introduces SAGE-net, a scalable deep learning framework for training and evaluating sequence-to-function (S2F) models using personal genomes, with improved prediction of how inter-individual DNA sequence variation affects gene expression. S2F models — typically convolutional or transformer neural networks trained to predict molecular phenotypes (gene expression, chromatin accessibility) from DNA sequence — have shown promise for interpreting non-coding variants, but they generally underperform when applied to personal genomes because they are trained on reference genomes and fail to capture inter-individual regulatory variation. SAGE-net addresses this by training directly on personal genome sequences paired with matched expression data, learning to predict how an individual's specific genetic variants affect their gene expression. The authors show that personal genome training improves expression prediction for held-out individuals, but interestingly, the gains come primarily from identifying predictive variants rather than learning a generalizable cis-regulatory grammar — suggesting that current models have not yet captured the underlying regulatory code. The framework is designed to scale efficiently to large cohorts, making it practical for biobank-scale analyses.
Why it matters: Interpreting non-coding genetic variants — which constitute the vast majority of disease-associated variants from GWAS — remains a fundamental challenge in human genetics. S2F models are one of the most promising approaches, but their failure to generalize across individuals has limited their clinical utility. SAGE-net's finding that personal genome training helps but doesn't capture generalizable regulatory grammar is an important reality check for the field, suggesting that fundamentally new architectures or training paradigms may be needed. The scalable framework itself is a practical contribution that enables large-scale personal genome analysis.
Why for Yiru: Personal genome interpretation is relevant to cancer genomics, where somatic and germline variants in regulatory regions may influence TME composition, immune surveillance, and therapy response. S2F models that can predict how an individual's genetic background affects immune gene expression could help explain why patients with similar tumours respond differently to immunotherapy. The observation that current models fail to learn generalizable regulatory grammar is also relevant to the broader challenge of building predictive models of gene regulation in the TME — where the regulatory environment is shaped by both genetic and microenvironmental factors.
Programmable control of bacterial operons with a single Cas13 RNA effector (SONAR)
Nature Biotechnology Published 2026-06-03 research article DOI: 10.1038/s41587-026-03159-4
Cas13 RNA targeting synthetic biology operon gene regulation CRISPR bacterial engineering SONAR
Summary: Introduces SONAR (Synthetic Operon control with Nucleolytic And Regulatory Cas13), a platform that uses a single engineered Cas13 RNA-targeting effector to achieve programmable, multi-modal control of bacterial operons. CRISPR-Cas13 enzymes naturally target and cleave RNA, and previous work has used them for RNA knockdown in bacteria. SONAR extends this dramatically by showing that a single Cas13 protein can be engineered — through simple changes in the guide RNA (crRNA) design — to achieve three distinct regulatory modes at target operons: (1) transcript degradation (RNA knockdown), (2) translation inhibition without degradation (by targeting the ribosome binding site), and (3) translation activation (by blocking inhibitory RNA structures). The authors demonstrate that these modes can be applied simultaneously to different genes within the same cell using different crRNAs, enabling complex multi-gene regulatory programs. This transforms Cas13 from a simple RNA knockdown tool into a versatile RNA regulatory platform.
Why it matters: SONAR represents a significant advance in the synthetic biology toolkit. The ability to achieve three distinct regulatory modes — degradation, translational repression, and translational activation — with a single protein simply by changing the guide RNA dramatically simplifies the engineering of complex genetic circuits. This is a substantial improvement over previous approaches that required different proteins for activation and repression. For metabolic engineering, SONAR could enable dynamic, multi-gene control of biosynthetic pathways. For basic research, the ability to titrate gene expression up or down at will in bacteria opens new experimental possibilities.
Why for Yiru: While SONAR is a bacterial tool, the concept of multi-modal RNA-level control with a single effector protein could inspire analogous systems in mammalian cells. RNA-level regulation is particularly attractive for TME applications because it enables transient, reversible control without permanent genetic changes. RNA-targeting CRISPR systems could be used to modulate immune checkpoint expression, cytokine production, or metabolic enzymes in TME cells — for example, transiently activating immunostimulatory genes in tumour cells or suppressing immunosuppressive factors in Tregs.
OrthoFinder v3: improved phylogenetic orthology inference with enhanced accuracy and scalability
Nature Methods Published 2026-06-09 research article DOI: 10.1038/s41592-026-03126-6
orthology phylogenetics comparative genomics orthogroup gene evolution bioinformatics tool
Summary: Presents OrthoFinder v3, a major update to the widely used phylogenetic orthology inference software that delivers substantially improved accuracy and scalability for analyzing massive and diverse genomic datasets. Orthology — the identification of genes in different species that evolved from a common ancestral gene — is fundamental to comparative genomics, enabling the transfer of functional annotations between species, the reconstruction of gene family evolutionary histories, and the identification of lineage-specific innovations. OrthoFinder v3 introduces several advances: enhanced phylogenetic delineation of orthogroups that provides a 7% relative improvement in accuracy; a new gene-to-orthogroup assignment method that substantially reduces memory usage without compromising accuracy; and improved scalability that allows analysis of thousands of genomes with diverse evolutionary distances. The authors demonstrate the method on datasets spanning the tree of life, from bacteria to vertebrates, showing consistent improvements over previous versions and competing methods.
Why it matters: OrthoFinder is one of the most widely used tools in comparative genomics, with thousands of citations. This update meaningfully improves both accuracy and scalability at a time when the number of sequenced genomes is exploding — making efficient, accurate orthology inference more important than ever. The improved accuracy has practical implications: more correct ortholog assignments mean fewer errors in functional annotation transfer, evolutionary rate estimation, and gene family analysis. The reduced memory footprint makes the tool accessible to researchers without access to large compute clusters.
Why for Yiru: Comparative genomics and orthology analysis are relevant to understanding the evolutionary origins of TME biology. Many genes involved in immune regulation, tumour suppression, and cell-cell communication have complex evolutionary histories with lineage-specific expansions and losses. Accurate orthology inference can reveal when key TME genes — checkpoint molecules, cytokines, chemokines — evolved, how they diversified across species, and which functional innovations are unique to mammals or primates. This evolutionary perspective can inform the selection of model organisms for TME research and the interpretation of cross-species immunotherapy studies.