Osteichthyans, comprising sarcopterygians and actinopterygians, dominate modern vertebrate biodiversity1-3, yet their pre-Devonian fossil record remains scarce and fragmentary4,5. The oldest articulated sarcopterygian6 and stem osteichthyan7 date to the late Silurian, whereas undisputed actinopterygian fossils in articulation appear only in the Middle Devonian8. Here we report an articulated, near-complete osteichthyan from the early Silurian Chongqing Lagerstätte (approximately 436 million years ago)9-11, representing the oldest osteichthyan occurrence including microfossils. This tiny fish exhibits a fusiform, generalized osteichthyan body outline, with plesiomorphic osteichthyan characters, including the lack of lepidotrichia and the presence of serial median dorsal plates, pectoral and dorsal fin spines and an anal fin spine reported previously exclusively in stem chondrichthyans12 and one placoderm13. It also displays features, such as a single dorsal fin and caudal fulcra, seen commonly in actinopterygians. Bayesian inference and the 50% majority rule consensus of the maximum-parsimony analysis place the new fish on the osteichthyan stem, whereas the strict consensus leaves its position unresolved within osteichthyans. This discovery increases Silurian osteichthyan diversity and further populates the osteichthyan stem group. The morphological disparity among early osteichthyans implies a more extensive Silurian to Early Devonian radiation of bony fishes than previous lines of evidence suggested.

Osteichthyans (bony fishes and tetrapods) encompass 98% of modern vertebrate species. However, our understanding of the sequence of character evolution among stem osteichthyans has been substantially limited by the fragmentary nature of known stem osteichthyan fossils1-4. Here we investigate newly discovered articulated head and trunk material of Megamastax amblyodus5, which yields previously unseen morphological details of a Silurian stem osteichthyan. Megamastax-previously interpreted as a lobe-finned fish5-exhibits distinct osteichthyan traits in the dermatocranium, such as resorptive tooth shedding and the presence of extrascapular bones. However, the arrangement of its dorsal aortae is reminiscent of crown-group chondrichthyans. The premaxilla with extensive palatal lamina and the elongated post-hypophyseal region of the braincase recall the condition in maxillate placoderms6-8. Crucially, the discovery of an inner dental arcade of discrete tooth cushions on individual attachment bases aligns Megamastax with the fragmentary genera Lophosteus and Andreolepis2-4, corroborating the previous interpretation of isolated tooth cushions as part of the jaw dentition3,9 and verifying their identity as stem osteichthyans. Phylogenetic analysis places Megamastax within the osteichthyan stem, near the osteichthyan crown-group node, and provides a framework for exploring the sequence of character acquisition along the osteichthyan stem. Together, these new findings help to bridge the morphological gap between stem gnathostomes and modern osteichthyans, offering insights into the sequence of early evolutionary steps that shaped the osteichthyan lineage.

Peri-device leak and device-related thrombus1 remain key challenges of current left atrial appendage occlusion (LAAO)2 owing to the incompatibility between the solid occluder and the left atrial appendage (LAA). Here we propose a personalized and complete LAAO using magnetofluids that is suitable for all types of LAAs. Magnetofluids can be injected into LAAs from cardiac catheters. In the presence of a sufficient magnetic field, magnetofluids can resist high-speed blood flow. Magnetofluids can precipitate into magnetogels in contact with water in the blood within only a few minutes. We further confirmed the long-term resilience and biocompatibility of the magnetogel over 10 months in a pig model in vivo. Neither device-related thrombus nor magnetogel leakage was observed in any pigs. The endocardium formed on the Watchman occluder was rough and incomplete, predisposing to thrombosis. Myocardial injuries were unavoidable due to the barbs of the Watchman occluder. The endocardium formed on our magnetogel was smooth, firm and thrombus-free. No crevice was observed between our magnetogel and the LAA, and no injury was caused to the myocardium. These findings may offer a promising clinical strategy for long-term thrombus-free LAAO.

Convective storms can develop rapidly, creating hazards to local populations through intense precipitation, strong winds and lightning1. The large-scale environment in which thunderstorms develop is often well captured in forecast systems, yet predicting where individual storms will initiate remains a fundamental challenge. It is known that differential heating driven by soil moisture (SM) patterns creates atmospheric circulations that favour convective initiation over drier soils2,3, whereas wind shear between low and mid levels can enhance storm growth4,5. Here we show that the most extreme initiations are especially favoured over SM contrasts by means of an interaction with wind shear. Analysing 2.2 million afternoon events across sub-Saharan Africa, we find 68% more initiations classed as extreme given favourable (versus unfavourable) soil conditions, with greatest vertical storm growth occurring where SM-driven circulations oppose the direction of shear-induced cloud displacement. Developing clouds follow the mid-level wind direction and, where this opposes the low-level flow, rainfall is strongly correlated with locally drier soils. Although such shear conditions are particularly common over tropical north Africa, the effect favours negative SM-precipitation feedbacks globally. The combination of SM heterogeneity and wind shear provides a potentially important source of predictability for where deep convection develops, particularly for the most rapidly developing thunderstorms.

Bacterial transcription initiation is a tightly regulated process that canonically relies on sequence-specific promoter recognition by dedicated sigma (σ) factors, leading to functional DNA engagement by RNA polymerase (RNAP)1. Although the seven σ factors in Escherichia coli have been extensively characterized2, Bacteroidetes species encode dozens of specialized, extracytoplasmic function σ factors (σE) whose precise roles are unknown, pointing to additional layers of regulatory potential3. Here we uncover a mechanism of RNA-guided gene activation involving the coordinated action of σE factor in complex with nuclease-dead Cas12f (dCas12f). We screened a large set of genetically linked dCas12f and σE homologues in E. coli using RNA and chromatin immunoprecipitation experiments, revealing systems that exhibit robust guide RNA enrichment and DNA target binding with a minimal 5'-G target-adjacent motif. Recruitment of σE was dependent on dCas12f and guide RNA, suggesting direct protein-protein interactions, and co-expression experiments demonstrated that the dCas12f-gRNA-σE ternary complex was competent for programmable recruitment of the RNAP holoenzyme. Remarkably, dCas12f-RNA-σE complexes drove potent gene expression in the absence of any requisite promoter motifs, with de novo transcription start sites defined exclusively by the relative distance from the dCas12f-mediated R-loop. Our findings highlight a new paradigm of RNA-guided transcription that embodies natural features reminiscent of CRISPR activation (CRISPRa) technology4,5.

The large volume of abdominal computed tomography (CT) scans1,2 coupled with the shortage of radiologists3-6 have intensified the need for automated medical image analysis tools. Previous state-of-the-art approaches for automated analysis leverage vision-language models (VLMs) that jointly model images and radiology reports7-12. However, current medical VLMs are generally limited to 2D images and short reports. Here to overcome these shortcomings for abdominal CT interpretation, we introduce Merlin, a 3D VLM that learns from volumetric CT scans, electronic health record data and radiology reports. This approach is enabled by a multistage pretraining framework that does not require additional manual annotations. We trained Merlin using a high-quality clinical dataset of paired CT scans (>6 million images from 15,331 CT scans), diagnosis codes (>1.8 million codes) and radiology reports (>6 million tokens). We comprehensively evaluated Merlin on 6 task types and 752 individual tasks that covered diagnostic, prognostic and quality-related tasks. The non-adapted (off-the-shelf) tasks included zero-shot classification of findings (30 findings), phenotype classification (692 phenotypes) and zero-shot cross-modal retrieval (image-to-findings and image-to-impression). The model-adapted tasks included 5-year chronic disease prediction (6 diseases), radiology report generation and 3D semantic segmentation (20 organs). We validated Merlin at scale, with internal testing on 5,137 CT scans and external testing on 44,098 CT scans from 3 independent sites and 2 public datasets. The results demonstrated high generalization across institutions and anatomies. Merlin outperformed 2D VLMs, CT foundation models and off-the-shelf radiology models. We also computed scaling laws and conducted ablation studies to identify optimal training strategies. We release our trained models, code and dataset for 25,494 pairs of abdominal CT scans and radiology reports. Our results demonstrate how Merlin may assist in the interpretation of abdominal CT scans and mitigate the burden on radiologists while simultaneously adding value for future biomarker discovery and disease risk stratification.

Humans and animals can sense the negative states of other individuals and respond with prosocial behaviour to improve their conditions1,2. Although prosocial behaviour is hypothesized to have an evolutionary root in caring for vulnerable newborn offspring1,3, whether the neural substrates underlying parenting may contribute to adult-directed prosocial behaviours remains largely unclear. We show that mice with higher levels of parenting exhibit more prosocial allogrooming toward stressed adults. The medial preoptic area (MPOA), a brain area involved in parenting behaviour, bidirectionally regulates allogrooming toward stressed conspecifics. Allogrooming and parenting behaviours recruit a partially overlapping neuronal ensemble in the MPOA, are both controlled by an MPOA-to-VTA pathway and are associated with dopamine release in the nucleus accumbens. Using activity-dependent labeling, we demonstrate that MPOA neuronal ensembles engaged during parenting behaviours are functionally required for allogrooming. Conversely, MPOA neurons activated during prosocial behaviour are functionally required for pup grooming. Collectively, these findings uncover a neural circuit mechanism of prosocial behaviour and reveal partially shared neural substrates between parenting and prosocial behaviours, suggesting that the neural systems evolved for offspring care may have provided a scaffold for the emergence of broader prosocial support between adults.

The impacts of sea-level rise and other hazards on the coasts of the world are determined by coastal sea-level height and land elevation1. Correct integration of both aspects is fundamental for reliable sea-level rise and coastal hazard impact assessments2,3, but is often not carefully considered or properly performed. Here we show that more than 99% of the evaluated impact assessments handled sea-level and land elevation data inadequately, thereby misjudging sea level relative to coastal elevation. Based on our literature evaluation, 90% of the hazard assessments assume coastal sea levels based on geoid models, rather than using actual sea-level measurements. Our meta-analyses on global scale show that measured coastal sea level is higher than assumed in most hazard assessments (mean offsets [standard deviation] of 0.27 m [0.76 m] and 0.24 m [0.52 m] for two commonly-used geoids). Regionally, predominantly in the Global South, measured mean sea level can be more than 1 m above global geoids, with the largest differences in the Indo-Pacific. Compared with geoid-based assumptions of coastal sea level, the measured values suggest that with a hypothetical 1 m of relative sea-level rise, 31-37% more land and 48-68% more people (increasing estimates to 77-132 million) would fall below sea level. Our results highlight the need for re-evaluation of existing coastal impact assessments and improvement of research community standards, with possible implications for policymakers, climate finance and coastal adaptation.

The co-occurrence of germline and somatic oncogenic alterations is frequently observed in breast cancer, yet their combined influence on tumour evolution and therapy resistance remains poorly defined. Through an integrated clinicogenomic analysis of more than 5,800 patients, we show that germline (g) pathogenic variants dictate the evolutionary trajectory of acquired resistance. We specifically find that gBRCA2-associated tumours are uniquely predisposed to develop acquired RB1 loss-of-function alterations, resulting in poor outcomes on standard-of-care frontline CDK4/6 inhibitor (CDK4/6i) combinations. This vulnerability is driven by a dual mechanism: baseline RB1 hemizygosity (heterozygous loss resulting in a single functional RB1 allele), which lowers the evolutionary barrier to biallelic inactivation, and ongoing homologous recombination deficiency, which promotes acquisition of RB1 loss-of-function alterations under the selective pressure of CDK4/6i. Preclinical models from gBRCA2 carriers showed near-uniform resistance to CDK4/6i, with consistent post-treatment Rb loss. Across multiple independent models and in our clinical data, PARP inhibition consistently outperformed CDK4/6i. Our findings suggest that prioritizing PARP inhibition in gBRCA2 carriers may intercept RB1-loss trajectories and delay resistance. More broadly, we establish a predictive framework for forecasting drug-resistant trajectories based on pre-treatment allelic configuration and mutational signatures.

Ferroptosis, a major mechanism of non-apoptotic programmed cell death, critically regulates the homeostasis and functionality of peripheral CD4+ and CD8+ T cells1-6. Here we demonstrate that in mouse, resistance of T cells to ferroptosis depends critically on the composition of standard rodent diets, and that dietary effects on ferroptosis (DEFs) have a crucial role in regulation of T cell homeostasis and immune responses. DEFs are microbiota-independent and are driven by variations in dietary polyunsaturated and monounsaturated fatty acids (PUFAs and MUFAs) that lead to variations in abundance of lipid species in lymphoid tissues and T cells. Consistently, ferroptosis resistance of human T cells also correlated with plasma lipid profiles across multiple healthy cohorts, exhibiting negative associations with PUFA/MUFA ratios in major lipid classes. DEFs dictate T cell resilience in the absence of the essential lipid peroxide scavenger GPX4 and broadly modulate T cell-dependent humoral immunity and T cell-mediated anti-tumour immunity, including in chimeric antigen receptor T cell therapy. Mechanistically, ACSL4, which preferentially biosynthezises PUFA-containing phospholipids7, is highly expressed in T cells and underpins DEF-mediated regulation of follicular helper T (TFH) cell generation and function. Our findings reveal the physiological significance of lipid metabolism in driving DEFs in immunity and suggest strategies targeting lipid metabolism to enhance vaccine efficacy and T cell-mediated immunotherapy.

How a local infection triggers systemic humoral immunity remains unclear. Here we identify farnesyl pyrophosphate (FPP), a mevalonate pathway metabolic intermediate1, as an endogenous alarmin that enhances IgG antibody responses through keratinocyte-derived IL-6 and CCL20. This signalling axis potentiates the differentiation of T follicular helper cells and migratory dendritic cells2,3. FPP accumulates within keratinocytes after infection or ultraviolet irradiation through the activation of the mevalonate pathway mediated by the unfolded protein response-SREBF pathway, amplifying germinal centre (GC) responses in draining lymph nodes. Mechanistically, accumulated FPP in the cytosol engages transient receptor potential vanilloid 3 (TRPV3) by binding to its intracellular domains, inducing Ca2+ influx that subsequently activates the calmodulin-calcineurin-NFAT and PYK2-RAS-ERK pathways to enhance IL-6 and CCL20 production. This FPP-TRPV3-IL-6/CCL20-GC axis potentiates pathogen-specific antibody production, conferring protection in wild-type but not TRPV3-deficient mice. Single-cell RNA-sequencing analyses of systemic lupus erythematosus (SLE) skin lesions and pathogen-infected mouse skin demonstrate hyperactivation of this signalling axis, particularly in the TRPV3high keratinocyte subset. In mouse models of SLE, the activation of this axis correlates with exacerbated disease pathology. Thus, FPP potentiates systemic humoral immunity through the TRPV3-IL-6/CCL20-GC signalling axis, providing insights for the development of vaccine adjuvants and potential therapeutics for SLE.

Metastasis is the major cause of death for patients with triple-negative breast cancer and other solid malignancies. Metastases arise from cancer cells that disseminate from the original tumour, survive systemic immune surveillance and colonize new organs1. Little is known about how initial disseminated tumour cells (DTCs) overcome anti-tumour immunity after seeding a new organ. Here we use a visible antigen in a model of triple-negative breast cancer with cognate CD8+ T cells to study the mechanisms of immune evasion in early metastatic seeding. Analysis of surviving DTCs revealed glucocorticoid receptor (GR) activation as a key driver of resistance to both CD8+ T cells and natural killer cells. Niche profiling using an optimized labelling tool identified FAS-FASL as a key pan-cytotoxic pathway against DTCs, which is repressed by GR activation. Pharmacological inhibition of GR in combination with immunotherapy reduced metastatic burden and expanded lifespan in mice. Thus, we identified a mechanism of immune evasion that operates specifically in DTCs, illustrating the unique immune-cancer interactions at this stage in the metastatic cascade. Our findings suggest that there are therapeutic opportunities to eliminate DTCs, separately from treatments aimed at primary tumours, and GR inhibition is one promising target.

Interactions between diet and the gut microbiota are fundamental to metabolic health, shaping energy balance and disease susceptibility1-5. However, the underlying mechanisms by which dietary and microbial factors converge to regulate host physiology remain unclear. Here we show that protein availability profoundly modulates the functional landscape of the gut microbiota and promotes remodelling of white adipose tissue (WAT). Specifically, low-protein diets (LPDs) robustly induce signature genes of browning in WAT to a similar extent to that seen in response to classical stimuli, such as cold exposure or β-adrenergic receptor activation6-8. LPD-mediated browning was markedly diminished in germ-free mice, and this defect was rescued by colonization with defined bacterial consortia made up of strains that were isolated and down-selected from the faeces of either LPD-fed mice or healthy human volunteers with 18F-fluorodeoxyglucose positron emission tomography (FDG-PET)-confirmed brown- or beige-fat activity9-12. Microbiota-induced browning was mediated both by bile acids driving the activation of the farnesoid X receptor (FXR) in adipose progenitor cells, and by nrfA-encoding commensal-derived ammonia driving the expression of fibroblast growth factor 21 (FGF21) in hepatocytes. The bile acid-FXR and ammonia-FGF21 axes both have non-redundant, essential roles in promoting WAT browning. These findings highlight a mechanistic link between diet, gut microbial metabolism and adipose tissue remodelling, uncovering microbiota-dependent pathways by which the host responds to dietary cues.

Cell-free DNA in blood originates from fragmented chromatin released by dying cells from both healthy and diseased tissues1,2. These fragments carry rich molecular modalities that can reveal pathological alterations in tissues of origin3-10. Here we develop cf-EpiTracing, a highly sensitive automated platform that profiles histone modifications in cell-free DNA from as little as 50 μl of human plasma. By integrating multimodal chromatin states with machine learning, cf-EpiTracing enables accurate deconvolution of cell types of origin. We generated 2,417 cf-EpiTracing profiles from plasma of 125 healthy individuals and 549 patients with inflammatory bowel disease, colorectal cancer, coronary heart disease or lymphoma. cf-EpiTracing enabled unbiased identification of primary diseased tissues and other organ involvement, stratification of B cell lymphoma subtypes with different genetic and epigenetic underpinnings, and detection of early-stage diseases or lesions. Surveying dynamics of epigenetic signatures uncovered disease transformation from follicular lymphoma to diffuse large B cell lymphoma. Further, cf-EpiTracing revealed genomic translocations and epigenetic alterations in patients with mantle cell lymphoma. Of note, our study leverages holistic epigenetic signatures, independently of knowledge of gene transcription, to accurately report recurrence risk and therapeutic response. Together, these findings establish cf-EpiTracing as an automated, non-invasive, epigenome-centric framework with broad applications in early diagnosis, molecular subtyping and prognostic prediction.

Aerosol forecasting is important for air-quality management, health risk assessment and climate change mitigation1,2. However, it is more complex than weather forecasting, owing to the interactions between aerosol physicochemical processes and atmospheric dynamics, resulting in high uncertainty and computational costs3,4. Here we develop a machine-learning-driven Global Aerosol-Meteorology Forecasting System (AI-GAMFS), which provides reliable 5-day, 3-hourly forecasts of aerosol optical components and surface concentrations. AI-GAMFS combines a vision transformer and U-Net in a backbone network, robustly capturing the complex aerosol-meteorology interactions via global attention and spatiotemporal encoding. Trained on 42 years of aerosol reanalysis data and initialized with Global Earth Observing System Forward Processing (GEOS-FP) analyses, AI-GAMFS delivers operational 5-day forecasts in 1 minute. Evaluation with independent ground-based observations suggests improved performance compared with the Copernicus Atmosphere Monitoring Service5 and regional dust models6-9 in forecasting aerosol optical depth and dust components. Compared with GEOS-FP10, it has a lower root-mean-square error for global aerosol optical depth, with comparable dust forecasting skill and improved surface aerosol component forecasts over the USA and China. Our results provide a step forward in leveraging machine learning to refine aerosol forecasting and may help warn against aerosol pollution events such as dust storms and wildfires.

RNA interference (RNAi) depends on DICER, an essential enzyme that processes RNA precursors into small regulatory RNAs. DICER cleaves RNA precursors according to the 5'-end counting rule, in which RNA length is measured from the 5'-end1-3. Previous work proposed a single 5'-end binding pocket that disfavours guanosine (5'-G), leading to cleavage inaccuracies4. Here we show that 5'-G promotes precise cleavage for many substrates. Using massively parallel dicing assays and cryo-electron microscopy, we identify a conserved guanosine-favoured (G-favoured) binding pocket in DICER, distinct from the previously described uridine-favoured (U-favoured) pocket. Together, these pockets influence the alignment between 21-nucleotide and 22-nucleotide cleavage registers, expanding the mechanism of small-RNA biogenesis in metazoan DICERs. We also find that conflicts between 5'-end binding and RNA-motif recognition can trigger RNA conformational adjustments that preserve accurate cleavage-site selection. In addition, conformational adjustments of the double-stranded RNA-binding domain (dsRBD) and PAZ domain help to align substrates with the catalytic centres for precise double-strand cleavage. These results show that the DICER cleavage mechanism integrates dual 5'-end binding pockets, RNA-motif influence and domain motions, advancing our understanding of microRNA biogenesis.

PIEZOs are mechanically gated ion channels that transduce force into electrochemical signals1. PIEZO1 responds to diverse stimuli including membrane stretch2 and shear stress3, whereas PIEZO2 is generally tuned to detect cellular indentation4,5. The functional specialization of PIEZO2 is proposed to underlie its distinct physiological roles, including mediating the sense of touch6,7. How PIEZO2 achieves this selectivity despite its close structural similarity to PIEZO1 is unclear. Here we combine single-molecule MINFLUX fluorescence nanoscopy with electrophysiology to link the conformational states of PIEZO2 to channel gating in intact cells. We find that PIEZO2 is intrinsically more rigid than PIEZO1, and that disparate mechanical stimuli paradoxically evoke opposite conformational and gating responses in each channel. These unique gating properties arise in part from a connection to the actin cytoskeleton, and we identify filamin-B (FLNB) as a molecular tether that is required for this interaction. This complex alters how force is transmitted to PIEZO2 and confers heightened sensitivity to and selectivity for cellular indentation. PIEZO2 and FLNB are co-expressed in somatosensory neurons and colocalize within tens of nanometres at the end organs of cutaneous mechanosensory afferents. These findings help to explain why PIEZO2 is a specialized mechanosensor and provide a molecular blueprint for understanding how cells decode diverse mechanical stimuli across tissues and organ systems.

All of life encodes information with DNA. Although tools for genome sequencing, synthesis and editing have transformed biological research, we still lack sufficient understanding of the immense complexity encoded by genomes to predict the effects of many classes of genomic changes or to intelligently compose new biological systems. Artificial intelligence models that learn information from genomic sequences across diverse organisms have increasingly advanced prediction and design capabilities1,2. Here we introduce Evo 2, a biological foundation model trained on 9 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life to have a 1 million token context window with single-nucleotide resolution. Evo 2 learns to accurately predict the functional impacts of genetic variation-from noncoding pathogenic mutations to clinically significant BRCA1 variants-without task-specific fine-tuning. Mechanistic interpretability analyses reveal that Evo 2 learns representations associated with biological features, including exon-intron boundaries, transcription factor binding sites, protein structural elements and prophage genomic regions. The generative abilities of Evo 2 produce mitochondrial, prokaryotic and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods. Evo 2 also generates experimentally validated chromatin accessibility patterns when guided by predictive models3,4 and inference-time search. We have made Evo 2 fully open, including model parameters, training code5, inference code and the OpenGenome2 dataset, to accelerate the exploration and design of biological complexity.

Known as the 'ultimate semiconductor', cubic diamond (CD) has gained substantial interest both scientifically and industrially. Its polymorph, hexagonal diamond (HD), is even more intriguing because of its fascinating properties associated with the meteorite impacts1-8. As no solid experimental evidence has been provided to prove its existence, the physical properties of HD remain largely unexplored. Here we report the synthesis of millimetre-sized, phase-pure HD from highly oriented pyrolytic graphite (HOPG) compressed along the c-axis at elevated temperatures. Combining advanced structural characterizations and theoretical simulations, we confirm the identity of HD and clarify the transformation pathway from graphite. Bulk HD exhibits a slightly higher hardness than CD and high thermal stability. These findings resolve the long-standing controversy on the existence of HD as a discrete carbon phase and provide new insight into the graphite-to-diamond phase transition, paving the way for future research and practical use of HD in advanced technological applications.

Influenza virus mRNAs are stable and competent for nuclear export and translation because they receive a 5' cap(1) structure in a process called cap snatching1. During cap snatching, the viral RNA-dependent RNA polymerase (FluPol) binds to host RNA polymerase II (Pol II) and the emerging transcript2,3. The FluPol endonuclease then cleaves a capped RNA fragment that subsequently acts as a primer for the transcription of viral genes4,5. Here we present the cryogenic electron microscopy structure of FluPol bound to a transcribing Pol II in complex with the elongation factor DSIF in the pre-cleavage state. The structure shows that FluPol directly interacts with both Pol II and DSIF, positioning the FluPol endonuclease domain near the RNA exit channel of Pol II. These interactions are important for the endonuclease activity of FluPol and FluPol activity in cells. A second structure, trapped after cap snatching, shows that the cleaved capped RNA rearranges within FluPol, directing the capped RNA 3' end toward the FluPol polymerase active site for viral transcription initiation. Together, our results provide the molecular mechanisms of co-transcriptional cap snatching by FluPol.