Wildfires sweeping through the vast boreal forests of Alaska, Canada, Scandinavia, and Russia could be having a larger impact on the climate than scientists once believed. A new study led by researchers at UC Berkeley suggests these northern fires may release far more carbon into the atmosphere than current estimates indicate. The reason is that these fires do not only burn trees. In many boreal regions, flames can spread downward into thick layers of carbon rich soil beneath the forest floor. These soils, known as peat, contain partially decomposed plant material that has accumulated over hundreds or even thousands of years. As a result, they may overlook slower, less visible fires that smolder deep within peat and organic soils. "Many of the fires that matter most for the climate don't look dramatic from space," said study lead author Johan Eckdahl, a postdoctoral scholar in Berkeley's Energy and Resources Group. "Peatlands and organic soils can smolder for weeks to years, releasing enormous amounts of ancient carbon." Eckdahl and his colleagues combined detailed national forest records with direct field measurements to reconstruct the amount of carbon released by each fire. Using these data, the team created a detailed map of wildfire emissions. Their analysis showed that local conditions such as climate, vegetation, and soil characteristics strongly influence how much carbon is stored in forests and how much is released during a wildfire. When the researchers compared their reconstructed emissions with six widely used global wildfire models, they discovered significant discrepancies. In other areas, especially where fires burned deep into soil layers, emissions were dramatically underestimated. For instance, the models predicted higher emissions in the county of Gävleborg, where intense fires burned through dry forests and were clearly visible from satellites. However, the situation was very different in neighboring Dalarna County. There, lower intensity fires burned quietly into thick layers of organic soil and were less noticeable from space. "Sweden is a very large country, but it's quite small compared to Siberia and Canada," Eckdahl said. "Sweden has a good network of forest roads, but in Siberia, I hear it's a real trek, which is one reason why we're severely missing measurements from that region." Eckdahl is now working with colleagues at UC Berkeley and other institutions as part of the Western Fire & Forest Collaborative to apply similar research methods in forests across the Western United States. Although forests in the western U.S. generally do not contain the same thick peat soils found in northern boreal regions, several other factors still influence wildfire emissions. "By improving our understanding of how this element flows between the land and the atmosphere, we can better anticipate the impact of future fire regimes in a warming world and design smarter strategies to reduce climate risks on society." Common Arthritis Drug Found To Lower Blood Pressure and Boost Heart Health Cosmic Voids Aren't Empty – They're Full of Something Far Stranger Stay informed with ScienceDaily's free email newsletter, updated daily and weekly. Keep up to date with the latest news from ScienceDaily via social networks: Tell us what you think of ScienceDaily -- we welcome both positive and negative comments.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Interactions between mutant cells and their environment have a key role in determining cancer susceptibility1,2,3. However, understanding of how the precancerous microenvironment contributes to early tumorigenesis remains limited. Here we show that newly emerging tumours at their most incipient stages shape their microenvironment in a critical process that determines their survival. Analysis of nascent squamous tumours in the upper gastrointestinal tract of the mouse reveals that the stress response of early tumour cells instructs the underlying mesenchyme to form a supportive ‘precancerous niche', which dictates the long-term outcome of epithelial lesions. Stimulated fibroblasts beneath emerging tumours activate a wound-healing response that triggers a marked remodelling of the underlying extracellular matrix, resulting in the formation of a fibronectin-rich stromal scaffold that promotes tumour growth. Functional heterotypic 3D culture assays and in vivo grafting experiments, combining carcinogen-free healthy epithelium and tumour-derived stroma, demonstrate that the precancerous niche alone is sufficient to confer tumour properties to normal epithelial cells. We propose a model in which both mutations and the stromal response to genetic stress together define the likelihood of early tumours to persist and progress towards more advanced disease stages. Groundbreaking studies in human genomics over the past decade have revealed that our healthy tissues accumulate cancer-associated mutations with age4,5,6,7,8. These observations highlight new levels of complexity in the early pathophysiology of cancer, raising the question of what other factors, beyond cancer mutations, may have a role during early carcinogenesis. Models of early tumours spanning a range of epithelial tissues, including oesophagus, skin and intestine, have started to offer a clearer understanding of what drives tumour initiation1,2,3,9,10,11. Work in this area has shown that tumour formation represents more than the mere accumulation of genetic alterations, highlighting the important role of environmental cues and non-genetic mechanisms in this process3,12,13,14,15. Indeed, mounting evidence indicates that the predisposition of a mutated epithelium to develop tumoral lesions depends on complex interactions between mutant cells and their dynamic surroundings. Coexisting mutant clones could either synergize or compete, contributing to early tumour initiation16,17. Indeed, even after tumours have formed, the presence of neighbouring mutant clones can continue to influence tumorigenesis1. Alternative environmental cues, such as the stiffness of the extracellular matrix (ECM)3,13, as well as direct cell–cell communication between mutant cells and non-mutant cells2,9,10,11,14, have also been shown to affect the expansion of mutant clones, susceptibility to tumour initiation and invasion18,19,20. Despite this, understanding of the mechanisms by which environmental factors determine the formation and long-term persistence of emerging tumours remains limited. Previous studies using an oesophageal early-tumour model demonstrated that not all nascent tumours have the same chance of survival. Most tumours are cleared from the tissue soon after formation by competition with neighbouring mutant clones. Surviving tumours instead persist long term, becoming susceptible to cancer progression1. But a key question remains: how are precancerous tumours able to withstand the competitive mutant environment that surrounds them? Understanding the processes underlying early tumour persistence and the relevance of the microenvironment at pre-neoplastic stages provides a critical opportunity to dissect the mechanisms driving precancer progression, opening new avenues to halt cancer in its tracks. Here we combine single-cell RNA sequencing with lineage tracing and 3D heterotypic cultures to study the unique features of the few nascent tumours that escape the existing protective barriers preventing tumorigenesis. We demonstrate that, during the earliest stages of tumour development, fibroblasts react to the pre-neoplastic epithelium by promoting the formation of a fibrotic precancer niche that, in turn, feeds back on the epithelium favouring early tumour growth and survival. To study the processes that underlie the persistence of pre-neoplastic nascent tumours, we used a well-established, clinically relevant mouse model of upper gastrointestinal tract (including oesophagus and forestomach) tumorigenesis driven by a mutagen found in tobacco smoke (diethyl-nitrosamine (DEN))1,21 (Extended Data Fig. After DEN treatment, the tissue becomes an evolving patchwork of mutant clones competing for space and survival, recapitulating the complex mutational landscape of the normal human ageing oesophagus22. This results in the emergence of pre-neoplastic squamous tumours with the potential to persist long-term (Extended Data Fig. Nascent epithelial tumours, marked by KRT17 (keratin 6A (KRT6A) and keratin 17)1, can be detected in tissue whole-mounts from their most incipient stages, from as early as 10 days after DEN treatment (Extended Data Fig. The emerging tumours are microscopic, containing as few as 10 cells, and are characterized by their distinctive rosette-like structure1,21 (Extended Data Fig. This brief window of formation is followed by a tumour-clearing process, in which more than one-third of the initial tumours are progressively eliminated1. The surviving tumours can persist in the tissue for more than a year, largely as low-grade dysplasia (pre-neoplastic or precancer stages), with sporadic progression to invasive squamous cell carcinomas1 (Extended Data Fig. As a result, only a subset of the original tumours survive long term, enabling us to study the mechanisms that modulate precancerous tumour persistence. To understand what drives early tumour survival, we first set out to compare the phenotypic traits of nascent tumours and those persisting long term (10 days and more than 8 months, respectively, after DEN treatment; Fig. Histological analysis showed that persistent dysplastic tumours (Extended Data Fig. 1h) were characterized by a prominent stromal remodelling (Fig. These nest-like structures were formed by stromal fibroblasts (PDGFRα+) that protruded towards the epithelial compartment, seemingly enclosing early tumours to create a supportive scaffold or a ‘precancerous niche' (Fig. Unlike in persisting tumours, at nascent stages, most epithelial lesions (around 70%; 199 of 296) showed no apparent stromal reorganization (Fig. 1c, d), denoting the existence of two phenotypically different nascent tumour subtypes, referred to here as Niche+ and Niche− (Fig. a, The experimental DEN carcinogen protocol. Wild-type mice were exposed to DEN in the drinking water for 2 months. b,c, Representative confocal images of long-term-persisting tumours 8 months after DEN treatment (b) and nascent tumours 10 days after DEN treatment (c), stained for DAPI (blue), KRT6A (tumour marker; red) and PDGFRα (fibroblast marker; grey). Image settings were adjusted to the upper stromal layer. d, Percentage of Niche+ and Niche− tumours at the indicated time points after DEN administration from three mice per time point; statistical significance was determined by a one-sided chi-squared test. e, Confocal images showing the incorporation of 5-ethynyl-2′-deoxyuridine (Edu; green) in KRT6A+ nascent tumours (red, dashed line) 10 days after DEN treatment. Images were generated omitting the uppermost suprabasal layer. f, Diameter (×100 µm) of Niche+ (red) and Niche– (blue) tumours at the indicated time points after DEN treatment. Data are expressed as mean ± s.e.m. Two-tailed Welch's t-test comparing Niche– versus Niche+ tumours. g, Cartoon illustrating the association between tumour niche remodelling and long-term tumour survival. Next, we assessed the dynamic nature of these two nascent tumour subtypes. We found that the number of Niche+ tumours, despite constituting the minority of all initial tumours, remained constant over time, whereas the number of Niche– tumours decreased markedly (Extended Data Fig. As a result, the tissue became progressively enriched in Niche+ tumours, with most (around 82%; 65 of 79) showing a supportive stromal scaffold by 8 months following DEN treatment (Fig. This enrichment in Niche+ tumours prompted us to explore whether stromal remodelling was associated with nascent tumour persistence. Close analysis revealed that keratinocytes in contact with the niche showed a particularly high proliferative activity (Extended Data Fig. 2e), indicating active epithelial–stromal communication at precancerous stages. These results were further reinforced by observations in the squamous forestomach, where long-term pre-neoplastic tumours also exhibited a remodelled stromal niche (Extended Data Fig. Collectively, our data support a model in which the remodelled stromal scaffold acts as a ‘precancerous niche' promoting tumour growth and survival. These observations link stromal remodelling in nascent tumours with pre-neoplastic tumour progression. The importance of the precancerous niche during early tumorigenesis became evident in 3D heterotypic cultures. These cocultures revealed that signals from the early tumour stroma are sufficient to confer tumour features to epithelial cells that had never been exposed to carcinogens24 (Extended Data Fig. Using reporter mouse lines to track the tissue origin, we found that untreated phenotypically normal epithelium directly exposed to the denuded tumour niche (lacking the epithelial compartment) acquired a tumour-like morphology and became highly proliferative, reaching levels similar to those of early tumours in vivo (Extended Data Fig. Moreover, heterotypic tissue constructs grafted into immune-deficient NOD-SCID-γ mice (NSG; NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ) showed that the pro-survival phenotype conferred by the tumour niche was also observed in vivo. Normal epithelial cells were more likely to engraft long term when exposed to early tumour stromal signals (Extended Data Fig. Overall, these results demonstrate that the early tumour microenvironment promotes epithelial cell growth, favouring precancerous tumour survival and, ultimately, disease progression. Given the key role of the niche in nascent tumour survival, we next explored its cellular composition. Under normal conditions, the squamous upper gastrointestinal tract is characterized by three distinct layers of stromal tissue: the lamina propria, a thin loose connective tissue directly beneath the epithelium; the muscularis mucosae, a layer of smooth muscle cells; and the submucosae, a dense irregular lower stromal compartment25 (Extended Data Fig. In line with observations in other epithelial tissues26,27, immunofluorescence analysis revealed two fibroblast populations that showed distinctive tissue compartmentalization and morphology and different expression levels of the pan-fibroblast marker PDGFRα (Extended Data Fig. Histological analysis of emerging tumours revealed that the main cellular component of the precancerous niche was PDGFRαlow fibroblasts, phenotypically indistinguishable from neighbouring lamina propria fibroblasts (Fig. Further characterization showed that endothelial and immune cells were largely absent from the niche in nascent tumours (10 days after DEN treatment; Fig. The side view shows that the niche is composed of lamina propria (Lp), not submucosae (Sb), fibroblasts. Mm, muscularis mucosae; Ep, epithelium. The white arrowhead points to the nascent tumour niche arising from lamina propria. Blue, DAPI; red, KRT6A; grey, PDGFRα. b, Number of stromal cells in tumour-free DEN tissue, Niche− and Niche+ tumours per unit of surface area, 10 days after DEN withdrawal; n = 27 Niche−, n = 22 Niche+, n = 15 (DEN) areas, from 3 mice; dots represent each area; PDGFRα+ fibroblasts, CD45+ immune cells and CD31+ endothelial cells are shown. Data are expressed as mean ± s.e.m. Statistical significance was assessed by one way Welch's analysis of variance (ANOVA) with multiple comparisons. c, 3D-rendered confocal side views of Niche− and Niche+ tumours 10 days after DEN withdrawal; green, CD45; orange, CD31 (absent); red, KRT6A; grey, PDGFRα. Dashed lines show the basal membrane. Image settings were adjusted to the upper stromal layer. d, Experimental protocol for fibroblast lineage tracing: Col1a2CreER and R26FlConfetti/wt mice received a dose of tamoxifen (TAM) followed by DEN treatment. Samples were collected 6 months after DEN treatment. e, Representative top-down (top) and side views (bottom) of control and tumour tissue from d, Grey, PDGFRα; yellow and red, lineage-traced Confetti+ cells. Dashed lines separate stromal compartments. f, A single channel from e shows a marked difference in PDGFRα expression across the two stromal compartments, both in control and tumour samples; lamina propria, PDGFRαlow (underlying epithelium); submucosae, PDGFRαhigh (deeper stromal layer). Since the role of cancer-associated fibroblasts (CAFs) in tumorigenesis, drug resistance and disease progression is well recognized28,29, we next assessed whether niche-forming fibroblasts exhibited CAF features. Except for the nuclear localization of YAP (active YAP, aYAP; Extended Data Fig. 3e), the expression of CAF markers, such as fibroblast activation protein (FAP) and α-smooth muscle actin (α-SMA), was not detectable in 10-day tumour fibroblasts (Extended Data Fig. We conclude that, although Niche+ fibroblasts in incipient tumours lack a full CAF phenotype, their aYAP status is consistent with a pre-CAF transitional state in the nascent tumour niche30. Another important stromal contributor to tumorigenesis and cancer progression is the immune compartment31. Immune-cell characterization of Niche+ and Niche– tumours, however, did not show significant differences (Fig. Accordingly, the emergence and persistence of Niche+ and Niche− tumours remained unaltered in immune-deficient mice (NSG; Extended Data Fig. These results indicate that immune cells do not discriminate between nascent Niche+ and Niche− tumours, acting as bystanders in early tumour persistence. To better understand the contribution of stromal fibroblast to tumour niche formation, we used an unbiased genetic lineage-tracing approach to target fibroblasts. Sporadic confetti labelling of fibroblasts, across the lamina propria and submucosae compartments, was induced in Col1a2-CreER;R26FlConfetti/WT mice followed by DEN treatment (Fig. Analysis of confetti clones 6 months after DEN treatment showed that fibroblasts in the niche underwent clonal expansion (Fig. Immunostaining further revealed that clones in the niche were formed by lamina propria PDGFRαlow fibroblasts (Fig. a, Microdissection of squamous upper gastrointestinal tract 8 months after DEN treatment for single-cell RNA sequencing. b, Uniform manifold approximation and projection (UMAP) showing cell-type annotation. d, Violin plots showing levels of Pdgfra and Fn1 expression in fibroblast clusters. e, Representative images from 6 mice of the DEN area and nascent tumour stroma 10 days after DEN treatment, showing the accumulation (white arrowhead) of fibronectin (FN1, green) in the niche. f, Heatmap (left) of the top 1,500 differentially expressed genes along the basal keratinocyte pseudotime trajectory. Pseudotime trajectories, top right (blue, committed; yellow, basal; magenta, tumour). g, Representative images from 6 mice, showing tumour 12 markers in nascent tumours 10 days after DEN treatment, showing homogeneous KRT6A (red) and KRT17 (yellow); heterogenous SOX9 and EGR1 (cyan); and AREG and RUNX1 (magenta); DAPI (blue). h, Representative images showing SOX9 distribution in Niche− and Niche+ tumours 10 days after DEN treatment. White arrowheads highlight SOX9+ keratinocytes (cyan), KRT6A (red) and PDGFRα (grey). i, Images of SOX9+ clusters in DEN-treated tumour-free areas 10 days after DEN treatment. Blue, DAPI; cyan, SOX9; grey, PDGFRα. White arrowheads highlight keratinocyte to fibroblast proximity; the white dashed line shows the epithelia to stroma border. j, Left, experimental protocol: Krt14CreER;Sox9flox/flox mice received tamoxifen (TAM) followed by DEN. Tissues were collected 1 month after DEN treatment and compared with DEN-treated uninduced controls. Right, quantification of tumour burden; n = 3 mice per condition; data shown as mean ± s.e.m. Images captured by confocal microscopy. To confirm this, we traced the PDGFRαlow and PDGFRαhigh fibroblast populations separately in Pdgfra-CreER;R26FlConfetti/WT mice (Extended Data Fig. We reasoned that differential PDGFRα expression levels would enable us to control the level of recombination in the lamina propria and submucosae. Paradoxically, the PDGFRαlow fibroblast population showed a markedly higher recombination efficiency, with negligible recombination detected in the lower PDGFRαhigh compartment (Extended Data Fig. 4j–m), potentially owing to different tamoxifen accessibility between stromal layers32. The distinctive recombination efficiency enabled us to trace PDGFRαlow fibroblasts at early tumour stages (6 weeks after DEN treatment) to explore their contribution to the niche-formation process (Extended Data Fig. We observed that the early tumour niche was composed of PDGFRαlow-derived fibroblast clones that expanded locally in the upper stromal compartment. No clonal expansion events were found in the lower stroma (PDGFRαhigh compartment) or spanning across stromal compartments (Extended Data Fig. Overall, these observations demonstrate that local PDGFRαlow fibroblasts in the lamina propria not only maintain, but also contribute to, the formation of the pre-neoplastic tumour niche. Single-cell RNA sequencing (scRNA-seq) was done in individually micro-dissected dysplastic tumours of the squamous upper gastrointestinal tract 8 months after DEN treatment (Fig. The transcriptional profile of pre-neoplastic tumour cells was compared with that of cells from adjacent tumour-free areas in DEN-treated mice (internal control), and with that of cells from healthy untreated control animals (Extended Data Fig. In line with our histological and lineage-tracing characterization (Fig. 3c,d and 4), the scRNA-seq analysis revealed a marked degree of heterogeneity in the expression of the fibroblast marker Pdgfra (Fig. Distinctive Pdgfralow and Pdgfrahigh fibroblast populations were present across conditions (Supplementary Table 5). Pdgfralow fibroblasts expressed higher levels of genes encoding structural (scaffolding matrix) ECM components (such as Fn1, Fbn1, Has1, Has2, Loxl2, Mfap5, Cd248 and Col1a1)33 common in loose connective tissue. Pdgfrahigh fibroblasts instead showed an enrichment of genes encoding ECM fibrillar (underlying matrix) collagens (Col6a3 and Col5a3), vascular support collagens (such as Col4a1/2, Col8a1, Col15a1 and Col13a1) and other basement-membrane components33 (such as Lama1 and Thbs1/2; Fig. Accordingly, immunolabelling of Pdgfra-derived clones (Extended Data Fig. 4j) in the lamina propria and submucosae, respectively, revealed increased fibronectin production in the Pdgfralow fibroblast compartment (Extended Data Fig. 6e), showed a significant upregulation of matrisome genes associated with wound healing/fibrosis9,34 (matrix deposition: Fn1, Fbln5, Tnxb, Cd248, Vim, Plaur, Mfap5; thrombospondins: Thbs1, Thbs3; collagens: Col3a1, Col1a1/2, Col5a3; remodelling: Timp1, Adam9, Loxl1; other factors: Fgfr2, Bmp1, Bmp6, Cx3cr1; Extended Data Fig. The pro-fibrotic nature of these niche fibroblasts was supported by immunolabelling of both nascent and surviving tumours (Fig. 6g), as well as by second-harmonic generation (SHG) imaging, which showed a marked ECM remodelling in the tumour niche (Extended Data Fig. Together, these findings pointed at the activation of a tissue repair response in the stroma of early tumours, in line with the long-standing notion that tumours are “wounds that do not heal”36. We next explored whether niche-forming fibroblasts (C19) in long-term surviving tumours presented a CAF signature. Despite a subtle upregulation of a reduced subset of CAF-associated genes12,28,37,38,39 (Vim, S100a4, Mfap5 and Col1a2; Extended Data Fig. 6l–n), their expression at the protein level (VIM, FSP and FAP) remained largely unaltered or undetectable (Extended Data Fig. Accordingly, fibroblast proliferation did not show significant changes28 (Extended Data Fig. These observations indicate that Niche+ fibroblasts in surviving dysplastic tumours lack a fully established CAF phenotype. However, the presence of nuclear YAP and the profibrotic nature of Niche+ fibroblasts point to a pre-CAF state30 (Extended Data Fig. 3e), transitioning to myCAFs at more advanced stages (invasive squamous cell carcinomas 14 months after treatment; Extended Data Fig. Label-transfer analysis from the fibroblast atlas in ref. 40 indicated that tumour niche fibroblasts probably derive from universal, rather than tissue-specific, fibroblast populations. Pdgfralow fibroblasts, including tumour-enriched cluster 19, partly recapitulated the transcriptional signature of the Pi16+ universal population, whereas Pdgfrahigh fibroblasts aligned with the Col15a1+ universal fibroblast subset (Extended Data Fig. To assess whether stromal genetic alterations drive the tumour niche phenotype, we performed deep-targeted sequencing of 192 cancer-related genes22 (Supplementary Table 9). The results argued against somatic mutations in fibroblasts being responsible for precancerous niche formation. The data revealed that DEN treatment induces gene perturbations, mainly in the epithelial compartment, showing a minimal mutational burden in the tumour stroma that matched the level of untreated or internal DEN control samples (Extended Data Fig. In contrast to nascent stages (Fig. 2b,c), we found that, as tumours progressed, they showed a marked remodelling of the vascular network and an increased immune infiltrate (Extended Data Fig. In line with previous findings, transcriptional analysis revealed notable changes in the immune-cell composition in long-term surviving tumours, indicative of active immune-cell recruitment with progression towards an immunosuppressive microenvironment at later stages41 (Extended Data Fig. Taken together, our data reveal a significant stromal reorganization in surviving tumours. Analysis of lamina propria-derived Pdgfralow fibroblasts was consistent with a notable fibrotic ECM remodelling in the precancerous niche, before the emergence of a fully established CAF phenotype. Next, we explored the epithelial transition from healthy and normal to pre-neoplastic states. Pseudotime analysis revealed two distinctive basal cell trajectories denoting tumour and non-tumour states. These trajectories largely converged in committed and differentiating cells (Fig. Gene score enrichment analysis of tumour-specific gene modules identified by pseudotime analysis revealed further heterogeneity in epithelial tumour states, referred to as Tumour 1 and Tumour 12 (Fig. 8d–h and Supplementary Tables 11–13), that expressed increased levels of the early tumour markers Krt6a and Krt17 (refs. Gene set enrichment analysis revealed a unique signature in Tumour 12 cells (Extended Data Fig. This was reinforced by increased levels of genes encoding transcription factors associated with a tumour stress response (Jun, Fos, Fosb, Runx1, Atf3, Egr1, Egr3 and Myc)41,44,45,46,47,48 (Extended Data Fig. Crucially, validation at the protein level further supported the heterogenous nature of the epithelial cells populating nascent early tumours, with Tumour 12 markers staining only a subset of tumour cells (Fig. Further expression changes in Tumour 12 comprised the upregulation of genes associated with stromal communication, including cell adhesion (Col12a1, Itgav, Itga2, Itgb6, Lama3, Vcl, Cadm1, Icam1 and Runx1 (ref. Increased expression of the cell-adhesion genes Ccn1, Ccn2 (ref. 52 was of particular interest, owing to their recognized role in communication with fibroblasts (Extended Data Fig. Since the data so far had indicated that there was close communication between Tumour 12 epithelial and stromal cells, we reasoned that Tumour 12 cells could be linked to the emergence of the tumour niche, and thereby to tumour persistence and survival. Indeed, SOX9 expression, used as a proxy to mark the Tumour 12 state (alongside KRT6A and/or KRT17 (refs. 1,21)), was expressed mainly in early tumour cells in direct contact with the niche (Extended Data Fig. 8o), indicative of their close interaction. To explore whether SOX9+ cells were associated with the formation of the pre-neoplastic tumour niche, we took advantage of sporadic clusters of KRT6A+ and SOX9+ cells in phenotypically normal (non-tumour) regions of DEN-treated tissue, potentially marking prospective tumour cells before lesion formation. Isolated SOX9+ cell clusters showed signs of fibroblast attraction, presenting fibroblasts in closer proximity and at a higher density than in the surrounding tissue (Fig. Overall, these data establish the Tumour 12 state as a relevant player in early tumour stromal remodelling and niche formation. Accordingly, SOX9 depletion in Krt14-CreER; Sox9flox/flox DEN-treated mice (Fig. 9g) led not only to a significant reduction in early tumour survival (1 month after DEN treatment; Fig. 3j) but also to a reduction in the size of nascent Niche+ tumours compared with that of Niche− tumours (Extended Data Fig. To study epithelial–mesenchymal communication in surviving tumours, we assessed ligand–receptor interactions53 enriched across coexisting cell populations in DEN-treated tissue (Fig. Pro-fibrotic ECM-related pathways were among the top outgoing interactions predicted to preferentially signal from tumour niche fibroblasts to tumour niche keratinocytes. These pathways included laminin, collagen, fibronectin, thrombospondin and tenascin (Fig. Given its well-established role in ECM assembly and association with fibrosis and advanced cancer progression54, we reasoned that fibronectin (FN1) might be a central player modulating ECM interactions across tumour cell compartments. The relevance of FN1 interactions was supported both by the specific upregulation of FN1 receptor genes in tumour niche keratinocytes (receiver cells) and by the increased expression of FN1, at both mRNA and protein level, in tumour niche fibroblasts (sender cells; Fig. Overall, our analyses predicted robust pro-fibrotic and wound healing epithelial–mesenchymal interactions in surviving tumours. Circles represent cells; circle sizes show cell numbers; and the thickness of the connections represents the number of significant interactions. c, Heatmap showing the top 10 signalling predictions. Interactions in Tumour (Tmr) 12 keratinocytes and Tmr fibroblasts are highlighted in yellow. d, Top: schematics of the chemoattractant assay. e, Percentage of fibroblasts crossing the membrane. Data are expressed as mean ± s.e.m. Dots represent replicate cultures from n = 4 mice. Significance assessed by one-way Welch's ANOVA with multiple comparisons (other groups in EDF 10c). f, Schematic representation of epithelioid and fibroblast co-culture treated with gefitinib (GFT) or vehicle (DMSO). g, Representative images from f showing fibroblast interaction with keratinocytes in DMSO and EGFR inhibition (GFT) conditions. Top insets show PDGFRα+ fibroblast (red) heterogeneity: PDGFRαlow fibroblasts assemble adjacent to keratinocytes, whereas PDGFRαhigh fibroblasts position further away in controls. Fibroblast activation at the border is labelled by VIM (white) and FN1 (green). This was inhibited in GFT (bottom). Green, CDH1; nT, nuclear Tomato; KRT, keratinocytes, red; blue, DAPI. h, Schematic representation of keratinocyte–fibroblast interactions under DMSO and GFT conditions. i, Experimental protocol of drug intervention with a fibronectin assembly-inhibiting peptide (FUD) or GFT 20-day regimen with the DEN treatment. j, Tumour burden decreased in the GFT and FUD groups compared with vehicle (VHC) or scrambled (SCR) control, respectively. Tissues were collected 10 days after DEN treatment. Significance was assessed by one-tailed Mann–Whitney test. Images were captured by confocal microscopy. Illustrations in a, d, f, h and i were created in BioRender; Alcolea, M. https://BioRender.com/eghet5p (2026). Next, we explored outgoing signals from tumour niche keratinocytes. Here, EGF was identified as one of the strongest incoming signals for tumour niche fibroblasts (Fig. EGF ligands (including AREG and HBEGF) were enriched in tumour niche keratinocytes (sender cells) at both the mRNA and protein levels (Fig. 8l, m and 10b), with its receptor (EGFR) being markedly expressed in the underlying tumour niche fibroblast population (receiver cells; Extended Data Fig. The upregulation of both AREG (amphiregulin) and FN1 in nascent tumours (10 days after DEN treatment; Fig. 6g and 8m) highlighted the importance of epithelial–mesenchymal communication from nascent tumour stages. Given the well-known role of EGF and FN1 signalling in epithelial–stromal communication during wound healing and tissue damage55,56, we reasoned that they might have a similar role in response to DEN-induced genetic stress. To determine whether tumour niche keratinocytes (Tumour 12) exert their mesenchymal remodelling effect (Fig. 8k,o and 9a–f) through the EGFR pathway (Fig. 4c), we used a chemoattractant assay (Fig. We found that the EGFR ligand AREG positively stimulated fibroblast migration, confirming that keratinocytes in nascent tumours can promote fibroblast chemotaxis and mesenchymal remodelling through paracrine EGF stimulation (Fig. We gained further insights into the dynamic nature of epithelial–mesenchymal communication by coculturing regenerative 3D oesophageal cultures (epithelioids57) with primary fibroblasts (Fig. We found that expanding keratinocytes, which exhibit increased levels of stress markers (including SOX9; Extended Data Fig. 10d), prompted fibroblasts to segregate spatially into two distinct populations, mirroring the in vivo scenario. PDGFRαlow fibroblasts localized immediately adjacent to the growing epithelium, whereas PDGFRαhigh fibroblasts were found in distant areas (Fig. The interaction between expanding epithelial cells and fibroblasts also led to FN1 deposition and vimentin upregulation in the PDGFRαlow population (Fig. Gefitinib-mediated inhibition of EGFR signalling in epithelioids further confirmed the role of EGFR signalling in epithelial–mesenchymal communication in a regenerative or stress context, showing reduced epithelial SOX9 expression, diminished fibroblast segregation or compartmentalization, and hindered fibroblast ECM remodelling (Fig. These data directly link EGFR signalling and SOX9 expression in the epithelium with mesenchymal remodelling. FN1, which is an important component of the fibrotic niche in early tumours (Fig. 6b–g), also represents a critical and well-established regulator of the stromal wound-healing response58. To determine whether the newly formed fibronectin-rich tumour niche has a critical role in promoting early tumour growth and survival, we treated established 3D epithelioids (exhibiting steady-state levels of proliferation)57 with soluble FN1 for 24 hours. Indeed, FN1 promoted epithelial proliferation (Extended Data Fig. These data indicate that, although fibrosis promotes tumour growth, a fibrotic environment alone is not sufficient to drive a pre-neoplastic response. These ex vivo experiments revealed that the EGF–SOX9–FN1 axis governs epithelial–mesenchymal communication and subsequent tissue reorganization in response to epithelial perturbations. Accordingly, in vivo experiments showed that inhibition of either fibronectin fibrillogenesis (using the functional upstream domain (FUD) peptide)59 or EGFR signalling with Gefitinib (GFT) led to a significant reduction in the number of Niche+ tumours (Fig. Taken together, our results demonstrate the central role of the EGF–SOX9–FN1 axis in early tumour niche formation. In response to genetic stress, SOX9+ epithelial cells stimulate fibroblast migration and ECM remodelling through EGF signalling. This, in turn, promotes the formation of a pro-fibrotic, fibronectin-rich tumour scaffold, which favours tumour persistence and progression by perpetuating the pro-tumorigenic phenotype. To validate the relevance of our observations in a human context, we analysed chemo-naive early-stage human oesophageal squamous cell carcinomas (T1a, T1b and Tis; Fig. 5a–d) and residual dysplastic tissues after chemotherapy (ypT0; Fig. a, Schematic representation of the tumour sample origin shown in b–d; tumour resections from human patients performed before chemotherapy (chemo-naive). b, Representative confocal image (from n = 4 patients) of a T1B tumour section showing widespread KRT6A (red), and heterogenous SOX9 (cyan) and AREG (magenta) expression; blue, DAPI. The control sample is normal area identified by a pathologist in the tumour section. c, Representative confocal image (from n = 5 patients) of a T1A tumour section showing widespread KRT17 (yellow), heterogenous SOX9 (cyan) expression and fibroblast attraction to areas marked by PDGFRα (grey) expression. d, Representative confocal image (from n = 1 patient) of a carcinoma in situ (Tis) whole mount showing heterogenous SOX9 (cyan) expression and fibronectin (FN1, green) accumulation underneath (white arrowhead). e, Schematic representation of the tumour sample origin shown in f; resections were performed after chemotherapy. f, Representative confocal image (from n = 1 patient) of a post-treatment, pathological staging T0 (dysplasia) whole mount, showing heterogenous expression of SOX9 (cyan) and AREG (magenta) and fibronectin (FN1, green) accumulation underneath. g, Schematic of proposed model whereby epithelial cells respond to genetic perturbation by activating a stress gene signature (tumour 12), denoted by SOX9 and EFG ligand overexpression. The establishment of this EGF–SOX9–FN1 signalling axis between epithelial and mesenchymal nascent tumour cells results in the formation of an early tumour niche that favours tumour growth and promotes long-term persistence. Immunolabelling showed that patient tumours, unlike adjacent tissue, recapitulate observations in the DEN mouse model. T1a and T1b tumours displayed homogeneous expression of KRT6A and KRT17 (pan-tumour markers in mice) and heterogeneous expression of SOX9 (marking tumour cells associated with stromal remodelling in mice; Tumour 12; Fig. Accordingly, SOX9-expressing cells exhibited high AREG expression levels (Fig. 5b) and were found next to areas with increased fibroblast density (PDGFRα; Fig. Analysis of tissue whole-mounts reinforced the link between SOX9 expression and stromal remodelling, with marked FN1 deposition in the vicinity of SOX9+ tumour cells (Fig. These features were also observed in residual dysplastic tissue after chemotherapy (Fig. 5e,f), which showed marked ECM remodelling (fibronectin deposition) in the proximity of AREG+ or SOX9+ tumour cells. Our observations support the presence of a heterogeneous AREG+ and/or SOX9+ population in early-stage squamous tumours of the human oesophagus. The data further reinforce the association between this population, mesenchymal changes and ECM remodelling in early human oesophageal tumorigenesis, revealing the potential clinical relevance of our study. Studies in the past decade have shown that mutations, conventionally thought to be the sole cause of cancer, can also be found in healthy ageing tissues4,5,6,7,8, where they form part of normal tissue physiology. This has redirected the interest of the cancer community to fill the knowledge gap around the earliest disease stages, and particularly to understand how mutant cells interact with adjacent tissue compartments1,2,3,13,21,22. This study provides mechanistic insights into the processes that determine whether tumours emerging in complex mutant landscapes persist long term or are outcompeted and eliminated from the tissue1. We show that early tumour survival and subsequent progression rely on intricate interactions between nascent tumour cells and their dynamic niche. This work reveals that exposure to mutagens activates a heterogeneous ‘tissue stress' response, whereby incipient epithelial and mesenchymal tumour cells signal and feedback onto each other through the EGF–SOX9–FN1 axis. Tumours failing to activate this communication axis are less likely to persist and grow. In particular, a tumour-specific stress state, defined by high SOX9 expression, promotes the recruitment of fibroblasts to the nascent tumour through EGF signalling. This in turn facilitates the formation of a precancerous niche, rich in fibronectin, that perpetuates a pro-tumorigenic phenotype, favouring tumour growth and persistence. Interfering with fibronectin fibrillogenesis in vivo impaired niche formation, prevented tumour survival and reduced the overall tumour burden. These findings support a self-sustaining process in which the reciprocal communication between niche mesenchymal cells and SOX9high epithelial cells supports tumour survival, favouring disease progression over time (Fig. These data are relevant to early human carcinogenesis. The heterogeneous expression of SOX9, EGFR ligands and associated deposition of FN1 are recapitulated in early stage human oesophageal squamous cell carcinomas, consistent with active epithelial–stromal communication in nascent human tumours. Whether interfering with ECM assembly represents a valid approach to prevent cancer progression in patients, and whether analogous mechanisms operate in other tumour types, requires further investigation. Overall, our data demonstrate the unprecedented capacity of the early tumour niche to perpetuate tumour survival signals beyond intrinsic changes driven by genetic alterations, enabling nascent tumours to persist in highly competitive mutant landscapes. This offers the new perspective that not only mutations, but also the environmental response to genetic stress, defines the likelihood of tumours to progress towards more advanced disease stages. Our findings indicate that strategies targeting tumour cells, as well as supporting neighbouring cells, could open new avenues for cancer prevention and improve long-term outcomes. High-grade dysplasia, squamous cell carcinoma and macroscopically normal, healthy clinical samples, as well as the corresponding clinical information, were collected following research ethics approval and individual informed consent from patients who underwent oesophagectomy for oesophageal cancer. The T1A and T1B stage chemo-naive surgical tumour samples were donated by patients who had undergone surgery at the Clinic for Visceral, Thoracic and Vascular Surgery at TU Dresden or at the Medical Department I of the Carl Gustav Carus University Hospital. Macroscopically normal samples adjacent to the proximal resection margin were sampled from cancer resection specimens. The corresponding formalin-fixed, paraffin-embedded material (tumour and healthy tissue) from a total of ten characterized oesophageal squamous cell carcinomas was selected from the archive of the Institute of Pathology of the University Hospital Carl Gustav Carus (EK 59032007) by the Tumour and Normal Tissue Bank (TNTB) Dresden. Studies presented in the manuscript involving early chemo-naive human oesophageal tumour samples from Dresden were approved by the Ethics Committee of TU Dresden, Germany (ref. Studies presenting chemo-naive or post-chemo human oesophageal tumour samples from Guy's and St Thomas' (London) and Addenbrooke's Hospital (Cambridge, UK), respectively, were approved by the East of Scotland Research Ethics approval committee (REC 18/ES/133). Histological sectioning of the tissue samples and haematoxylin and eosin staining of reference slide series for determining the tumour cell content of the individual patient samples were done at the Institute of Pathology, University Hospital CGC Dresden, TU Dresden. All animal experiments were approved by the local ethical review committees of the University of Cambridge and conducted according to the Home Office project licences PPL70/8866 and PP7037913 of the Cambridge Stem Cell Institute, University of Cambridge. Unless otherwise specified, C57BL/6J mice (Charles River, strain code 632) were used. Other mouse strains used include: cell-cycle reporter line R26Fucci2aR (Fucci2a)60, provided by I. J. Jackson; PdgfraEGFP (007669, Jackson Laboratory); Sox9flox/flox (ref. Further information about the experimental mouse lines can be found in the Supplementary Methods section Experimental mouse lines. Recombination of Col1a2CreERR26FlConfetti/WT mice was induced by a single intraperitoneal tamoxifen injection (3 mg per 20 g body weight). The Col1a2CreERR26FlConfetti/WT mice were induced by a single intraperitoneal tamoxifen injection (0.5 mg or 5.0 mg per 20 g body weight). The K14CreERSox9flox/flox received two subcutaneous tamoxifen injections (5 mg per 20 g body weight) 48 h apart. Tamoxifen was prepared by dissolving in ethanol (less than 10% total volume) and diluting in sunflower-seed oil. All strains were maintained in a C57BL/6 background. All experiments used a mixture of male and female mice with no gender-specific differences observed (unless specified otherwise). For RNA-sequencing experiments, only male animals were used to avoid confounding effects from the oestrous cycle. All animals exposed to the carcinogen and their respective controls were adults between 8 and 14 weeks of age (see the section on chemically induced mutagenesis below). Mice were bred and maintained under specific-pathogen-free conditions at the Gurdon Institute and the Anne McLaren Building, University of Cambridge. All animals were housed at 20–24 °C, 45–65% humidity and a 12 h:12 h light:dark cycle. Mice were treated with DEN (Sigma-Aldrich; N0756) at 40 mg l−1 in Ribena-flavoured water for 24 h, three times a week (Monday, Wednesday and Friday) for 8 weeks1,21. Mice received sweetened water between DEN dosages and normal water after the completion of DEN treatment. Control mice received sweetened water as a vehicle control for the length of the treatment. Animals exposed to DEN were monitored for adverse effects as stated in our Home Office project licences (PPL70/8866 and PP7037913) for regulated procedures on protected animals. In summary, animals were weighed daily on weekdays for the first week, weekly for the next month and then monthly thereafter. Animals were also checked every day for any clinical signs or abnormal behaviour. Any concerning animals were weighed every other day or daily, if necessary, until the weight was stable again. If the weight loss approached 10%, animals were weighed daily until stable and received wet mash or palatable diet. Animals showing 15% weight loss measured for 2 consecutive days were killed immediately. For EdU labelling experiments, mice received 100 µg EdU in PBS (Life Technologies, A10044) intraperitoneally 2 h before tissue collection. In vitro 3D cultures (see above) received media supplemented with 10 µM EdU and were incubated for 2 h at 37 °C and 5% CO2 before fixation. EdU incorporation in tissue whole-mounts (see above) was detected using a Click-iT EdU kit according to the manufacturer's instructions (Invitrogen, C10337). EdU+ cells were quantified using confocal microscopy. Mice were treated with Gefitinib for 20 days at 80 mg per kg body weight (or vehicle control) three times a week to inhibit the EGFR pathway. Treatment started 10 days before the end of DEN treatment and ended 10 days after it. Pharmacological inhibition of FN1 fibrillogenesis was achieved by treating mice with functional upstream domain (FUD) peptide64 intraperitoneally for 20 days at a concentration of 12.5 mg per kg body weight. Treatment started 10 days before the end of DEN treatment and ended 10 days after it. Peptides were synthesized at more than 95% purity (WatsonBio; peptide sequence below). Lyophilized peptides were reconstituted in PBS. The upper gastrointestinal tract (oesophagus and forestomach) from control and DEN-treated mice was dissected at the time points indicated in the main text and/or figure legends. Oesophagi were excised and cut open longitudinally. The muscle layer was then removed and the tissue was flattened under a dissecting microscope using fine forceps. Stomachs were cut open longitudinally and rinsed twice with PBS to remove any food remains. The glandular stomach was excised away and the forestomach kept and flattened for downstream analysis. For epithelial-only and stromal-only whole-mounts, tissues were incubated in 5 mM ethylene-diamine-tetraacetic acid (EDTA) (Life Technologies, 15575020) in PBS for 3 h at 37 °C. After incubation, the epithelium was gently peeled from the stroma using fine forceps. Subsequently, each of these layers was flattened individually. Oesophageal and forestomach whole-mounts (either peeled or unpeeled) were fixed in 4% paraformaldehyde (Alfa Aesar, 043368) in PBS for 30 min at room temperature. Haematoxylin and eosin staining was done in 7-μm paraffin-embedded sections by the Histology Core Service at the Cambridge Stem Cell Institute and imaged using a Zeiss AxioScan Z1 microscope. Histological analysis of murine tumour samples was done by B. Mahler-Araujo at the MRC Metabolic Diseases Unit (MC_UU_00014/5). Under a dissecting microscope in a laminar flow hood, oesophagi were dissected and epithelial–stromal layers isolated as described above in the section ‘Whole-mount preparation'. Thereafter, tissues were rinsed in 1% P/S in PBS three times to remove residual EDTA and flattened. Combinations of epithelium and stroma from different experimental conditions (DEN treated and/or control) were prepared by carefully placing the epithelial layer over the relevant stroma (referred to as ‘tissue recombination' composites). Flattened epithelium–stromal constructs were cultured in six-well plate inserts (ThinCert Greiner Bio-One, 657641). Size-matched polydimethylsiloxane (PDMS) stencil frames were placed around the tissue construct to prevent cell expansion (see the ‘Stencil production' section below). The tissue was allowed to settle for 10 min before adding 2 ml of minimal medium (mFAD) containing one-part DMEM (Fisher Scientific, 41966029) and one-part DMEM/F12 (Fisher Scientific, 11320033) supplemented with 5 μg ml−1 insulin (Sigma-Aldrich, 15500), 5% fetal calf serum (Fisher Scientific, 26140079), 1% P/S and 5 µg ml−1 Apo-Transferrin (Sigma-Aldrich, T2036), as previously described24,65. The 3D heterotypic cultures were maintained in standard humidified cell-culture incubators at 37 °C with 5% CO2 for up to 7 days. At the end point, samples were fixed in 4% PFA in PBS for 30 min at room temperature and stored for downstream confocal analysis. Silicone elastomer (PDMS) was mixed with a curing agent (Avantor VWR; Sylgard 184 Elastomer Kit, 634165S) at a 10:1 ratio and centrifuged at 300g for 10 min to remove the bubbles. The resulting mix was poured on a dish at around 70 mg cm−2 and left on an even surface to polymerize overnight at 37 °C. The next day, the resulting polymer was cut into 2 × 5 mm rectangle-shaped frames, sterilized in 70% ethanol overnight, and treated with 1% pluronic acid (Sigma-Aldrich, P2443-250g) in PBS for 1 h at 37 °C. The frames were then left to air dry before use. Tissue recombination composites (as described above in the section ‘Ex vivo tissue recombination assay') of DEN-treated oesophageal stroma and untreated (control) oesophageal epithelium were prepared for in vivo grafting adapting the strategy described above. Before separating the epithelium from the stroma, all visible tumours were marked with a partial incision using a punch biopsy tool (1 mm diameter; Merck, WHAWB100040). After separating the tissue layers, all the stromal compartments were assessed for peeling efficiency under a fluorescence dissecting microscope (Leica M165 FC), and any remaining epithelium, identified by the dense epithelial nuclei clusters, were excised from the tissue using a 1 mm biopsy punch. For heterotypic tissue constructs, 2 mm biopsies (Selles Medical, instrument BP20F) of tumour or control stroma were excised and a 2 mm healthy untreated epithelium biopsy placed above. Composites were cultured overnight as described above and grafted in the back skin of anaesthetized shaved NSG female mice (two constructs per incision, and two incisions per animal). Longitudinal incisions for grafting were approximately 5 mm in length. The wounds were closed with GLUture glue (Fisher Scientific, NC0632797) and the mice were left to recover. Then, 3–6 months later, the mice were killed and the back skin fixed with 4% PFA in PBS for 30 min at room temperature and stored for downstream confocal analysis. Oesophagi were dissected as described above and cut in half. Tissue was incubated in 0.5 mg ml−1 Dispase (Sigma-Aldrich, D4818) for 10 min at 37 °C while rotating. After incubation, the epithelium was peeled away and the stroma was minced finely and incubated in Trypsin-EDTA (0.25%) (ThermoFisher, 25200056) for 15 min at 37 °C while rotating. The resulting suspension was mixed by pipetting and DMEM supplemented with 10% FBS and 1% P/S was added (1:1 v/v). The suspension was passed through a 70 µm filter (PluriSelect, 43-10070-40) and cells were pelleted by centrifugation at 300g for 5 min at 4 °C. Pellets were resuspended in 0.5% FBS, 1% P/S in DMEM, and seeded on 8.0 µm pore transwell insert (24-well plates; ThinCert Greiner Bio-One, 662638). Primary fibroblasts were cultured for 48 h before fixation in 4% PFA in PBS for 10 min. Membranes were incubated with 1 µg ml−1 DAPI (Sigma-Aldrich, D9542) in PBS for 30 min at room temperature, cut and mounted in 1.52 Rapiclear mounting medium (SUNJin Lab, RC152001) keeping their original orientation, followed by confocal analysis. Further information on quantification can be found in the Supplementary Methods section ‘Analysis of fibroblast migration assay'. Oesophagi were cultured using the 3D epithelioid organ culture approach65. In brief, tissues were dissected, cut into 3 × 5 mm rectangles and placed on a transwell insert with the epithelium side up. The tissue was left to settle for about 5 min. Explants were expanded in complete medium (cFAD) containing mFAD supplemented with 1 × 10−10 M cholera toxin (Sigma-Aldrich, C8052), 10 ng ml−1 EGF (Fisher Scientific PeproTech, AF-100-15), 0.5 μg ml−1 hydrocortisone (Calbiochem, 386698). Tissue explants were removed by aspiration 5 days after culture set-up and maintained in mFAD for 2 weeks to confluence. Soluble fibronectin (Fisher Scientific Corning, 356008) was added to the medium for 24 h at 100 μg ml−1 after diluting it in mFAD with 25 mM HEPES (Fisher Scientific, 15630056). Samples were fixed with 4% PFA in PBS for 30 min at room temperature, after a 2 h EdU chase, and kept for downstream confocal analysis. Epithelioids were set up as described above and maintained in mFAD until keratinocyte migration started. The original tissue was then removed and the explant left overnight before adding freshly isolated oesophageal fibroblasts (as described above). DMSO as vehicle control or 2 µM of Gefitinib prepared in DMSO were added together with the fibroblast suspension. Then, 3 days after the fibroblasts were introduced to culture, samples were fixed with 4% PFA in PBS for 30 min at room temperature and kept for downstream confocal analysis. After fixation, epithelial–stromal composites or human tissue whole-mounts were incubated for 30 min in permeabilization buffer (PB1; 0.5% bovine serum albumin (VWR International, 126575-10), 0.25% fish-skin gelatin (Sigma-Aldrich, G7765), 1% Triton X-100 (Fisher Scientific, 10102913) in PBS), then blocked for 2 h in PB1 containing 10% donkey serum (DS) (Scientific Laboratory Supplies, D9663). Next, tissues were incubated with primary antibodies diluted in 10% DS in PB1 for 3 days at 4 °C followed by four washes of 30 min each with 0.2% Tween-20 (Promega UK, H5151) in PBS. Thereafter, tissues were incubated overnight with secondary antibodies diluted 1:500 in 10% DS in PB1 at room temperature. Unbound antibody was removed by four washes with 0.2% Tween-20 in PBS throughout the next day. Antibody details are provided in Supplementary Table 15. To stain cell nuclei, tissues were incubated with 1 µg ml−1 DAPI in PBS at 4 °C overnight. Afterwards, samples were rinsed three times in PBS and mounted in 1.52 RapiClear mounting media for imaging. Immunolabelling of individual tissue layers (epithelium or stroma) or sections consisted of an incubation for 30 min in permeabilization buffer (PB2; 0.5% bovine serum albumin, 0.25% fish-skin gelatin, 0.5% Triton X-100 in PBS). Tissues were then blocked for 2 h in PB2 containing 10% DS. Next, samples were incubated overnight at room temperature with primary antibodies diluted in 10% DS in PB2 followed by three washes with 0.2% Tween-20 in PBS for 30 min each. Secondary antibodies were diluted 1:500 in 10% DS in PB2 and incubated with tissues overnight at 4 °C, after which unbound antibody was removed by three washes with 0.2% Tween-20 in PBS, and staining continued as above. Thick cryosections of fixed tissues embedded in optimal cutting temperature compound (OCT; Thermo Scientific, 12678646), cut with a thickness of 50 µm onto glass slides, were immunolabelled using the same protocol. Likewise, 7-μm paraffin-embedded sections were immunolabelled according to the protocol described above, after antigen retrieval performed by heating of tissue sections in either 1 mM EDTA buffer (pH 8.0) or 10 mM sodium citrate buffer (pH 6.0) for 10 min at 95 °C. When staining with primary antibodies raised in the same host, one of the antibodies was acquired as preconjugated with a fluorophore or conjugated in house following the manufacturer's instructions (Invitrogen, A20186/A20187). After incubation with the corresponding secondary antibodies, the samples were blocked for 3 h at room temperature with 10% DS in PB with the IgG from the relevant host species (1:500). Afterwards, samples were incubated with conjugated antibodies diluted in PB containing 10% DS and the relevant host IgG (1:500) overnight at room temperature. At this point, staining proceeded as described above. Immunostained samples were analysed by confocal imaging. Confocal images were acquired using either an inverted Leica SP5 microscope with standard laser configuration or a Stellaris 8 FALCON FLIM microscope with a white-light laser using LAS X 4.7.0.28176 or 3.5.5.19976 software. Typical confocal settings used included: bidirectional scanning, a 40× immersion objective lens, an optimal pinhole size (as defined by the software), a scan speed of 400–600 Hz with 2–3× line averaging, optimal Z-step size (as defined by the software) and a resolution of 512 × 512 or 1,024 × 1,024 pixels, unless stated otherwise. Then, sD reconstructions from optical sections and their corresponding image renders were generated using Volocity 5.5.5 (PerkinElmer) and Volocity 7 (Quorum), Zen 3.2 and Arivis 3.5.1. Further information about specific types of image analysis, such as second-harmonic generation imaging, can be found in the Supplementary Methods. Sample preparation methods for libraries can be found in the Supplementary Methods in the section ‘Single-cell and RNA isolation for single-cell RNA-sequencing (scRNA-seq)'. The scRNA-seq libraries were generated using the 10× Genomics Chromium Next GEM Single Cell 3′ Reagent Kit (v.3) and sequenced at the Genomics Core Facility of Cancer Research UK (CRUK), Cambridge Institute. Libraries were generated in two different batches. Information about library batches can be found in Supplementary Table 2. Control samples were included in both batches to provide a reference to assess potential batch effects. The cells for each biological replicate were loaded into a 10× Chromium microfluidics chip channel to generate one library from each. In total, 17 libraries were sequenced on either an Illumina HiSeqx4000 or a NovaSeq6000 system using one SP, two S1 and two S2 flow cells. Note that, given the punch biopsy approach used, DEN samples could contain sporadic tumour cells from tumours not visible under the dissection microscope. The raw scRNA-seq data were processed with CellRanger (v.7.0.1). Reads were aligned to the mouse reference genome (mm10 2020-A), empty droplets were filtered and unique molecular identifiers were counted to generate gene-expression matrices. Doublets were identified using Scrublet66 (v.0.2.3) and removed, along with low-quality cells, on the basis of per-sample quality-control metrics (Supplementary Table 2); cells with more than 15% mitochondrial reads or genes expressed in fewer than three cells were excluded, resulting in 91,347 high-quality cells. Count matrices were processed using a standard Seurat workflow67 (v.5.0.3) up to dimensionality reduction. Data were integrated by tissue of origin (oesophagus or forestomach) using Harmony68 (v.1.2.3), and the integrated embeddings were projected in two dimensions using Seurat's RunUMAP function. Further details on data clustering, annotation, and trajectory and communication analysis can be found in the Supplementary Methods and on the Alcolea lab's GitHub page. Oesophagi from control and DEN-treated animals were dissected as described above. Tissues were flattened with the epithelial side up and visible tumour lesions were marked with a partial incision using a 1 mm diameter punch tool under a dissecting microscope. Tissues were then incubated in 5 mM EDTA for 3 h at 37 °C while rotating. After incubation, the epithelium was removed, the stroma was flattened and tissues were fixed as described above. Immunofluorescent labelling against KRT14 was done as described above. Only tumour stroma footprints negative for KRT14 (lacking epithelial cells) were considered for DNA sequencing to avoid the identification of genetic mutations present in epithelial cells. Tumour stroma matching the criteria was dissected under a fluorescence dissecting microscope (Leica M165 FC) using a 1 mm punch biopsy tool. Punch biopsies of equivalent size were collected from untreated healthy tissues as a control. The DNA from individual biopsies was extracted using the Arcturus PicoPure DNA Extraction Kit (Fisher Scientific, KIT0103) following the manufacturer's instructions. Thereafter, proteinase K was inactivated by incubation at 95 °C for 10 min. Samples were then sheared, libraries prepped using the NEBNext Ultra II Fragmentase System, and index tags applied (Sanger 168 tag set). Then, 500 ng of pooled material was taken forward for hybridization capture and enrichment (SureSelect Target enrichment system, Agilent Technologies) using a previously designed bait panel of 192 genes (Supplementary Table 9), including those commonly mutated in squamous cancers22. After clean-up, libraries were normalized to around 6 nM and submitted to cluster formation for sequencing on a Novaseq6000 (Illumina) to generate 100-base pair paired-end reads. Aligned reads were mapped to the mouse GRCm38 reference genome using BWA-mem (v.0.7.17)69. Depth of coverage was also calculated using SAMtools to exclude reads that were unmapped, not in the primary alignment, failed platform or vendor quality checks, or PCR or optical duplicates. Bedtools (v.2.23.0)71 was then used to calculate the depth of coverage per base across samples. Variant calling was done using the deepSNV R package (also commonly referred to as ShearwaterML; v.1.21.3; https://github.com/gerstung-lab/deepSNV). Variants were annotated using VAGrENT. Mutations called by deepSNV ShearwaterML were filtered using the following criteria: first, positions of called single nucleotide variants (SNVs) have a coverage of at least 100 reads; second, germline variants called from the same individual are omitted from the list of called variants; third, adjustment for false discovery rate and mutations use support from at least one read from both strands for the mutations identified; and finally, pairs of SNVs on adjacent nucleotides in the same sample are merged into a dinucleotide variant if at least 90% of the mapped DNA reads containing at least one of the SNV pairs also contained the other one. DeepSNV ShearwaterML was run with a normal panel of approximately 12,000 reads. The numbers of biological replicates and animals are indicated in the figure legends (n refers to the number of independent replicates per time point and/or condition). A minimum of three independent mice or ex vivo cultures were used in all cases. All experiments were done independently at least three times with similar results, unless otherwise stated. The reproducibility of all key findings was confirmed in independent experiments conducted on different days and using independent biological samples. For image analysis, a minimum of three independent samples were inspected in all cases. The data are expressed as mean values ± s.e.m. All statistical tests were done comparing biological replicates. Differences in tumour burden were assessed by one-tailed unpaired non-parametric Mann–Whitney U-tests. Differences between Niche− and Niche+ tumour distribution in ageing animals was assessed by one-sided chi-squared test. For large datasets, normality was assessed using a Kolmogorov–Smirnov test; for normally distributed data, differences between two groups were assessed by two-tailed Welch's t-tests; for non-normally distributed data, a two-tailed Mann–Whitney U-test was used. Differences between more than two groups were calculated using either one-way Welch's ANOVA, followed by a Dunnett's T3 multiple-comparison test, or Kruskal–Wallis one-way ANOVA, followed by Dunn's multiple-comparison test, for normally distributed or non-normally distributed data, respectively, unless specified otherwise. Exact P-values are indicated in the relevant figures with a precision of up to four decimal places. Statistical tests were conducted in GraphPad Prism (v.10.5.0) with 95% confidence intervals. No statistical method was used to predetermine sample size. The experiments were done without randomization. Blinding was done for tumour count per condition and in vitro sample analyses by confocal microscopy. In cases for which quantification was done in tumours and morphologically normal areas, blinding was not possible owing to differences in physical sample appearance. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Mouse reference genomes GRCm38 and mm10 2020-A were used. The single-cell RNA sequencing data generated in this study have been deposited in the Gene Expression Omnibus (GEO) repository under accession code GSE271962. The DNA sequencing dataset was deposited at the European Nucleotide Archive (ENA) under dataset accession number ERP134942. Source data are provided with this paper. No new algorithms were developed for this paper. Colom, B. et al. Mutant clones in normal epithelium outcompete and eliminate emerging tumours. Yum, M. K. et al. Tracing oncogene-driven remodelling of the intestinal stem cell niche. The extracellular matrix dictates regional competence for tumour initiation. Kakiuchi, N. & Ogawa, S. Clonal expansion in non-cancer tissues. Fowler, J. C. & Jones, P. H. Somatic mutation: what shapes the mutational landscape of normal epithelia? Martincorena, I. et al. High burden and pervasive positive selection of somatic mutations in normal human skin. Martincorena, I. et al. Somatic mutant clones colonize the human esophagus with age. Martincorena, I. Somatic mutation and clonal expansions in human tissues. Injury prevents Ras mutant cell expansion in mosaic skin. van Neerven, S. M. et al. Apc-mutant cells act as supercompetitors in intestinal tumour initiation. Flanagan, D. J. et al. NOTUM from Apc-mutant cells biases clonal competition to initiate cancer. Chen, Y. et al. Epithelial cells activate fibroblasts to promote esophageal cancer development. Mechanics of a multilayer epithelium instruct tumour architecture and function. Tagore, M. et al. GABA regulates electrical activity and tumor initiation in melanoma. Transient loss of Polycomb components induces an epigenetic cancer fate. Marusyk, A. et al. Non-cell-autonomous driving of tumour growth supports sub-clonal heterogeneity. Reeves, M. Q., Kandyba, E., Harris, S., Del Rosario, R. & Balmain, A. Multicolour lineage tracing reveals clonal dynamics of squamous carcinoma evolution from initiation to metastasis. & Campbell, I. G. Co-evolution of tumor cells and their microenvironment. Karras, P., Black, J. R. M., McGranahan, N. & Marine, J.-C. Decoding the interplay between genetic and non-genetic drivers of metastasis. Janiszewska, M. et al. Subclonal cooperation drives metastasis by modulating local and systemic immune microenvironments. Alcolea, M. P. et al. Differentiation imbalance in single oesophageal progenitor cells causes clonal immortalization and field change. Colom, B. et al. Spatial competition shapes the dynamic mutational landscape of normal esophageal epithelium. Chang, J. et al. Genomic alterations driving precancerous to cancerous lesions in esophageal cancer development. Lifting regenerative barriers promotes epithelial cell fate plasticity supporting lineage conversion. Tétreault, M.-P. & Kahrilas, P. GI manifestations with a focus on the esophagus: recent progress in understanding pathogenesis. Huycke, T. R. et al. Patterning and folding of intestinal villi by active mesenchymal dewetting. Manieri, E. et al. Role of PDGFRA+ cells and a CD55+ PDGFRALo fraction in the gastric mesenchymal niche. Sahai, E. et al. A framework for advancing our understanding of cancer-associated fibroblasts. The evolving tumor microenvironment: from cancer initiation to metastatic outgrowth. Calvo, F. et al. Mechanotransduction and YAP-dependent matrix remodelling is required for the generation and maintenance of cancer-associated fibroblasts. Roles of the immune system in cancer: from tumor initiation to metastatic progression. Kumar, R., Mao, Y., Patial, S. & Saini, Y. Induction of whole-body gene deletion via R26-regulated tamoxifen-inducible Cre recombinase activity. Frantz, C., Stewart, K. M. & Weaver, V. M. The extracellular matrix at a glance. Guerrero-Juarez, C. F. et al. Single-cell analysis reveals fibroblast heterogeneity and myeloid-derived adipocyte progenitors in murine skin wounds. Dynamic expression of α6 integrin indicates epidermal cell behaviors. Dvorak, H. F. Tumors: wounds that do not heal. Nurmik, M., Ullmann, P., Rodriguez, F., Haan, S. & Letellier, E. In search of definitions: cancer-associated fibroblasts and their markers. Integrated single-cell transcriptome analysis reveals heterogeneity of esophageal squamous cell carcinoma microenvironment. Davidson, S. et al. Single-cell RNA sequencing reveals a dynamic stromal niche that supports tumor growth. Buechler, M. B. et al. Cross-tissue organization of the fibroblast lineage. Yao, J. et al. Single-cell transcriptomic analysis in a mouse model deciphers cell transition states in the multistep development of esophageal cancer. Garg, M. et al. Laminin-5γ-2 (LAMC2) is highly expressed in anaplastic thyroid carcinoma and is associated with tumor progression, migration, and invasion by modulating signaling of EGFR. Stem cell lineage infidelity drives wound repair and cancer. Larsimont, J.-C. et al. Sox9 controls self-renewal of oncogene targeted cells and links tumor initiation and invasion. The pioneer factor SOX9 competes for epigenetic factors to switch stem cell fates. Zhang, X. et al. Dissecting esophageal squamous-cell carcinoma ecosystem by single-cell transcriptomic analysis. & Beck, B. Dedifferentiation of esophageal progenitors in metaplasia and cancer. Osorio, K. M., Lilja, K. C. & Tumbar, T. Runx1 modulates adult hair follicle stem cell emergence and maintenance from distinct embryonic skin compartments. Canals, F., Colomé, N., Ferrer, C., Plaza-Calonge, Mdel C. & Rodriguez-Manzaneque, J. C. Identification of substrates of the extracellular protease ADAMTS1 by DIGE proteomic analysis. Chen, C. C., Chen, N. & Lau, L. F. The angiogenic factors Cyr61 and connective tissue growth factor induce adhesive signaling in primary human skin fibroblasts. Sakai, K. et al. Thrombospondin-1 promotes fibroblast-mediated collagen gel contraction caused by activation of latent transforming growth factor β-1. Inference and analysis of cell-cell communication using CellChat. Singh, P., Carraher, C. & Schwarzbauer, J. E. Assembly of fibronectin extracellular matrix. Mimura, Y. et al. Epidermal growth factor induces fibronectin expression in human dermal fibroblasts via protein kinase C δ signaling pathway. Aging fibroblasts present reduced epidermal growth factor (EGF) responsiveness due to preferential loss of EGF receptors. Herms, A. et al. Self-sustaining long-term 3D epithelioid cultures reveal drivers of clonal expansion in esophageal epithelium. Arwert, E. N., Hoste, E. & Watt, F. M. Epithelial stem cells, wound healing and cancer. Tomasini-Johansson, B. R. et al. A 49-residue peptide from adhesin F1 of Streptococcus pyogenes inhibits fibronectin matrix assembly. Mort, R. L. et al. Fucci2a: a bicistronic cell cycle reporter that allows Cre mediated tissue specific expression in mice. Kist, R., Schrewe, H., Balling, R. & Scherer, G. Conditional inactivation of Sox9: a mouse model for campomelic dysplasia. Snippert, H. J. et al. Intestinal crypt homeostasis results from neutral competition between symmetrically dividing Lgr5 stem cells. Kang, S. H., Fukaya, M., Yang, J. K., Rothstein, J. D. & Bergles, D. E. NG2+ CNS glial progenitors remain committed to the oligodendrocyte lineage in postnatal life and following neurodegeneration. Sottile, J. et al. Fibronectin-dependent collagen I deposition modulates the cell response to fibronectin. Herms, A. et al. Self-sustaining long-term 3D epithelioid cultures reveal drivers of clonal expansion in esophageal epithelium. Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Li, H. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. Danecek, P. et al. Twelve years of SAMtools and BCFtools. Quinlan, A. R. & Hall, I. M. BEDTools: a flexible suite of utilities for comparing genomic features. Menzies, A. et al. VAGrENT: Variation Annotation Generator. We thank members of the Alcolea lab for comments and suggestions; I. Michalk for coordinating human sample collection and documentation at the Tumour and Normal Tissue Bank Dresden; J. Cordle for coordinating human sample collection and delivery from Guy's and St Thomas' NHS Foundation Trust; the Cancer Aging and Somatic Mutation (CASM) teams for accommodating work related to human samples, especially E. Anderson, L.-A. Gilbey and the rest of the CASM research management team; P. Humphreys and D. Clements for guiding imaging analysis at the Jeffrey Cheah Biomedical Centre (JCBC); N. Lawrence for guiding imaging analysis at the Gurdon Institute Imaging Facility; I. Pshenichaya for technical histology support at the JCBC; B. Mahler-Araujo and J. Warner for histopathology analysis; D. Streichert for technical support with human-sample staining; M. Paramor and V. Murray; K. Kania; I. Mohorianu's team for their contribution to scRNA-seq sample processing, library preparation and data pre-processing; the staff of the University Biomedical Services Gurdon Institute and the Anne McLaren Building technical biomedical assistance; and I. J. Jackson (Fucci2a) and H. Clevers (Rosa26Confetti) for donating mouse lines. This work was supported by grants to M.P.A. This research was funded in whole, or in part, by Wellcome (203151/Z/16/Z, 203151/A/16/Z) and the UKRI Medical Research Council (MC_PC_17230), and core support grant for Cambridge Stem Cell Institute Discovery Research Wellcome Platform Discovery Research Platform for Tissue Scale Biology (226795/Z/22/Z). This work also received funding from the European Research Council (ERC) Executive Agency under HORIZON ERC Synergy Grant Programme (grant agreement 101167202 — ClonEScape — ERC-2024-SyG, to M.P.A.). was funded by the Isaac Newton Trust (21.07(a)), the Medical Research Council (MR/P019013/1) and Worldwide Cancer Research (19-0192 and 23-0063). was supported by a University of Cambridge/Wellcome Junior Interdisciplinary Fellowship (ISSF 11/2/2020) and the Medical Research Council (MR/P019013/1). was funded by an ELBE Postdoctoral Fellowship from the Center for Systems Biology Dresden (CSBD). received funding from the European Union's Horizon 2020 research and innovation programme under Marie Sklodowska-Curie grant agreement 794664 (OESOPHAGEAL FATE). and M.T.B were also supported by the Isaac Newton Trust (research grants 16.24(e)) and the Leverhulme Trust (RPG-2023-136). was funded by the Cancer Research UK Cambridge Centre (CANCTA-2023/100003). acknowledges funding from the Human Frontier Science Program (LT000092/2016-L). was also supported by Cancer Research UK programme grants (C609/A27326 andDRCRPG-Nov24/100003). received funding from the ERC under the European Union's Horizon 2020 research and innovation program (grant agreement 950349) and acknowledges the Center for Nanoscience (CeNS), Munich. acknowledges funding from the Royal Society (E.P. Present address: RhyGaze, Basel, Switzerland These authors contributed equally: G. Skrupskelyte, J. E. Rojo Arias G. Skrupskelyte, J. E. Rojo Arias, H. Ajith, D. Rossetti, M. K. S. Tang, M. T. Bejar, B. D. Simons & M. P. Alcolea G. Skrupskelyte, J. E. Rojo Arias, H. Ajith, D. Rossetti, M. K. S. Tang, M. T. Bejar & M. P. Alcolea Max Planck Institute for Molecular Cell Biology and Genetics, Dresden, Germany Max Planck Institute for the Physics of Complex Systems, Dresden, Germany Center for Systems Biology, Dresden, Germany Wellcome Sanger Institute, Hinxton, UK B. Colom, J. C. Fowler, K. Murai, W. Knight, A. Noorani & P. H. Jones Department of Gastroenterology, Guy's and St. Thomas' Hospital, London, UK Arnold Sommerfeld Center for Theoretical Physics, Ludwigs-Maximilians-Universität Munchen, Munich, Germany Department of Applied Mathematics and Theoretical Physics, Centre for Mathematical Science, University of Cambridge, Cambridge, UK Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar guided the experimental design for single-cell RNA sequencing and, together with H.A. did cell-to-cell communication analysis and advised and supported the single-cell RNA sequencing analysis. analysed lineage-tracing data and provided advice and guidance on single-cell RNA sequencing data analysis. processed DNA sequencing samples and did data analysis. provided insights and technical expertise in targeted DNA sequencing. assisted with human tissue experiments. did histopathological staging of human tumour samples. assisted with in vitro validation experiments. advised on parts of the study, provided expertise regarding epithelial stem cells and tumour biology, and assisted with writing the manuscript. conceived the project, supervised and performed experiments, and wrote the manuscript with input from all authors. Review and editing of the manuscript was done by all authors. For the purpose of open access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. Correspondence to G. Skrupskelyte or M. P. Alcolea. The authors declare no competing interests. Nature thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. a, H&E image of transversal cross-sections from the murine upper gastro-intestinal tract (including oesophagus and forestomach) displaying multiple tumours (black arrowheads) 6 m after DEN treatment. b, Number of tumours per oesophagus (OE) at the indicated time points after DEN treatment, n = 3 mice per time point. A line is drawn across the means of each time point. One way Welch's ANOVA with multiple comparison was used to assess significance, indicated p values show two consecutive timepoint comparison indicated by the black line. c, Cartoon (edges) illustrating tumour clearance over time, with examples of confocal images (middle) of growing persistent tumours stained for DAPI (blue) and KRT6A (red). Tissues were collected at the indicated time points after DEN treatment: 10 d, 2 m and 1 y from 5 mice per timepoint. d-e, Confocal images showing early tumour marker expression in nascent tumours 10 d after DEN treatment from 18 mice. f, H&E image of a transversal cross-section depicting a nascent tumour 10 d after DEN treatment (black triangle), representative image from 4 mice. g, H&E images of tumours transversally cross-sectioned at different time points after DEN treatment at different stages of progression in the squamous upper gastro-intestinal tract. h, Bar plot (left) showing distribution of DEN induced pathology before (<) and after (>) 9 m after DEN treatment. i, Representative whole mount confocal images from control (ctrl) or 10 d post-DEN oesophagi showing localized KRT17 expression (yellow) in tumours, otherwise absent in surrounding or age-matched control tissues. j, Percentage of proliferating (KI67+) fibroblasts in DEN-treated oesophagi (0 d post-DEN) and in normal age-matched controls from PdgfraEGFP mice. n = 17 and 9 areas per condition respectively, each from 3 animals. Statistical significance was assessed by a two-tailed Mann-Whitney U test. Days (d), months (m), year (y). Source data (b,h,j). a, Density of Niche+ and Niche– tumours at the indicated time points after DEN treatment. n = 3 mice per time point; solid lines represent means. One-tailed Mann-Whitney test Niche– versus Niche+ tumours per time-point. b, Confocal images showing EdU incorporation (green) in KRT6A+ nascent tumours (red, dashed line) 10 d after DEN treatment. c, Number of EdU+ basal keratinocytes normalised per basal cell density at the indicated time points after DEN treatment. N, number of normal areas or Niche– or Niche+ tumours; n = 84 (10 d), 20 (2 m) Niche– tumours, n = 35 (10 d), 38 (2 m) Niche+ tumours, and n = 20 (10 d), 14 (2 m) DEN areas from 3 mice per condition. Data expressed as violin plots with mean (solid lines) and quartiles (dashed lines). One-way Welch's ANOVA with multiple comparison was used to assess significance. d, Quantification of basal keratinocyte cells per area in DEN, Niche– or Niche+ early tumours. N is number of DEN or tumour areas from 3 mice. One-way Welch's ANOVA with multiple comparisons was used to assess significance. e, 3D-rendered confocal side-views of a tumour 3 m after DEN treatment and its respective age-matched control showing R26Fucci2aR tissue (mCherry, G1 cells; red mVenus, S/G2/M cells; green). White arrows mark proliferating cells in the basal layer. f, 3D-rendered confocal side-views showing growing Niche+ tumours at the indicated time points after DEN treatment; DAPI (blue); KRT6A (red); PDGFRα (greyscale). g, Representative confocal images of forestomach epithelia 10 d and 3 m after DEN treatment. White arrowheads highlight tumours at indicated time points. Data form n = 3 animals per timepoint. h, Representative confocal images of a transversal cross-section of a forestomach tumour showing the presence of the stromal niche 8 m afterDEN treatment. DAPI (blue); KRT6A (red); KRT17 (yellow); PDGFRα (grey). i, Schematic illustration of ex vivo 3D heterotypic organ cultures. After separating epithelial and stromal layers, control epithelia were combined with stroma of surviving tumours 3 m after DEN withdrawal (dashed lines). j, Schematic representation of the heterotypic construct approach used (left). Images (middle and right) show the emerging tumour like structure in control epithelium, side and top-down views, respectively. Grafted epithelium expressed nuclear tdTomato (nT, red, from nTnG mice). DAPI (blue); β-catenin (CTNNB1, green), PDGFRα (greys). k, Experimental protocol (left) and representative side-views of 3D-rendered confocal images (right) of 3D heterotypic tissue constructs 7 d post-culture and in vivo control or tumour tissue for comparison. Ep, epithelium; WT, wild-type; Ctrl, control. Tumour bearing tissue was collected 3 m after DEN treatment. Dots are data from an individual tumour (Tmr, in vivo), tumour-like structure (Tmr-like, heterotypic culture), and the respective in vivo and in vitro control areas. Dots represent areas assessed across different biological replicates n = 9 control areas, from 3 mice in vivo, 23 control areas from 3 mice in vitro and n = 10 tumours from 6 mice in vivo and 7 tumour like structures from 5 mice in vitro. One-way Welch's ANOVA with multiple comparisons were used to assess significance. m, Schematic representation of the subcutaneous heterotypic grafting approach to NOD SCID gamma (NSG) immunodeficient mice. Grafted epithelium expressed nuclear tdTomato (nT, red, nTnG mice), while accompanying stroma was EGFP (green, from H2B-EGFP mice) to distinguish the origin of cells in the graft. n, Representative confocal images (from m) showing the long-term survival of a graft combining healthy untreated nT epithelium with EGFP tumour or control stroma. o, Quantification of surviving and lost graft constructs from n, 3-6 m post-transplantation expressed as percentage. 19 control and 23 DEN constructs were transplanted across 10 animals, respectively. Statistical significance was determined by a one-sided Chi-squared test. Days (d), months (m), control (ctrl). Source data (a,c,d,l,o). b, Schematic representation of the imaging angle used according to the sample preparation method. c, Image of tissue whole-mounts showing differential PDGFRα (grey) expression between the fibroblast layers. α-SMA smooth muscle layer (green), DAPI (blue). Lamina propria-Lp, Muscularis mucosae- Mm, Submucosae- Sb. d, Image of a tissue section showing reduced PDGFRα expression in upper stromal layers (i.e. lamina propria). Dashed lines (c,d) separate the tissue layers as listed in schematic. e, Confocal images of nascent tumour stroma 10 d after DEN treatment. Images show active YAP (aYAP) expression in PDGFRα labelled fibroblasts (white arrowheads). aYAP (magenta); CD31 and CD45 (green); PDGFRα (grey); DAPI (blue). f, Representative confocal images of Niche+ tumours 10 d after DEN treatment, showing the absence of typical cancer associate fibroblast (CAF) marker in the niche forming fibroblasts. g, Number of CD45+ (immune) cells in epithelia of Niche−/Niche+ tumours and adjacent DEN area normalised to surface area, 10 d post-DEN. n = 27 (Niche−), 22 (Niche+) tumours, from 3 mice each, dots represent a tumour. Statistical significance assessed by one-way Welch's ANOVA with multiple comparisons. h, Experimental DEN carcinogen protocol in wild-type (wt) and immunocompromised NOD SCID gamma (NSG) mice. N = 3 wt and 4 NSG mice at 10 d and n = 3 wt and 3 NSG at 2 m after DEN treatment time point; expressed as mean±s.e.m. One-tailed Mann-Whitney test was used to assess statistical significance. k, Representative confocal images of Niche+ nascent tumours from 10 d post-DEN showing fibroblast (PDGRα, grey) recruitment (white arrow) to a forming epithelial tumour (KRT6A, red) in wt and NSG mice. Scale bar, 10 µm, l, Diameter (µm) of Niche+ and Niche− tumours, n = 126 (tumours in wt from 3 mice) and n = 189 (tumours in NSG from 4 mice) at 10 d post DEN and n = 57 (tumours in wt from 3 mice) and n = 36 tumours in NSG from 3 mice) at 2 m after DEN treatment. Two-tailed Welch's t-test was used to assess statistical significance. Source data (g, i, j, l). a, Experimental protocol of fibroblast lineage tracing in Col1a2CreERRs26FConfetti mice. b, Confocal images of confetti-labelled cells 24 h after induction in different stromal layers. c, d, Efficiency of confetti construct recombination after a 72 h chase in the PDGFRα+ fibroblast population. Total percentage of recombined cells in the PDGFRα+ fibroblast population (c), and split colour percentage within the recombined population (d). N represents number of fields from 3 animals, n = 36; each replicate shown with a different shade of colour. e, Confocal 3D top-down or side-view images of a DEN area. 2d-f. f, Distance to the nearest neighbour fibroblast (fbr) labelled in the same colour in external control (Ctrl), internal control (DEN) and tumours. Pairwise Wilcoxon signed-rank tests were performed, and P-values were adjusted using the Holm-Bonferroni correction. g, Cartoon illustrating fibroblast expansion in the early tumour niche: lineage tracing results revealed that stroma between tumour lobes show increased fibroblast density and decreased distance between the same colour fibroblasts (black arrowheads) if compared to the tissue outside of lobes or control. Results show that labelled cells are closer together on the lobe contour than elsewhere. Data expressed as mean and error bars denote 95% confidence intervals obtained from boot strapping. j, Schematic representation of the confetti cassette and experimental protocol of fibroblast lineage tracing in PdgfraCreERTRs26FConfetti mice. k, Images of labelled fibroblasts (from j) in lamina propria and submucosae; data from 5 animals. l, Percentage of fibroblasts labelled in lamina propria (Lp) and submucosae (Sb) (from j). m, Representative confocal images of tissue from j showing top-down projections and confetti labelled fibroblast; 5 and 3 animals for low and high TAM dose, respectively. Quantified induction efficiency per field of view shown as an average above the images. n, Experimental DEN protocol in TAM induced PdgfraCreERTRs26FConfetti mice. o, Representative confocal images of confetti labelled fibroblasts in early tumours 6w after DEN treatment. Dashed lines mark tumour and stromal layer margins. 0.5 mg/20 g body weight is low; 5 mg/20 g body weight is high tamoxifen dose. PDGFRα (grey), α-SMA (in b grey); CONFETTI (cyan, yellow and red) in (b). Dashed lines in (e,k,o) label lamina propria and submucosae layers in side views. Hours (h), weeks (w), days (d), months (m), tamoxifen (TAM). Scale bars (k, m, o) 50 µm. Source data (c, d, l). a, Schematic visualisation of representative DEN-induced tumours under the dissection microscope, before and after microdissection using punch biopsy tool. b, Image composite depicting sample collection strategy for scRNA-seq. Squamous oesophagus and forestomach (outlined in red) first underwent tumour marking and were then peeled to separate epithelial (Ep) and Stromal (Str) compartments. This was followed by sample biopsying, where tumours and size-matched control biopsies (adjacent tissue as internal control, DEN; untreated tissue as external control, Control) were micro-dissected for scRNA-seq. c, UMAP representing cell cluster distribution as defined by Seurat. d, UMAP cell distribution of cell-types identified using label transfer from ref. 73. e, Heatmap showing expression of representative marker genes used for curated cell type annotation across the 22 clusters. Log-transformed normalised expression levels were averaged by cluster for each gene and scaled across all cells belonging to each group. f, UMAPs showing expression of representative genes for each cell type identified. Colour bars of UMAPs indicate log-transformed normalised expression levels. a, Heatmap showing the log2 fold-change (FC) of differentially expressed genes in Pdgfralow relative to Pdgfrahigh fibroblasts. Dashed line marks fibroblast contour. b, UMAP projection denoting the heterogenous expression of Fn1 (like Pdgfra) in fibroblasts; inset shows fibroblast cluster distribution. c, Violin plots showing the levels of Pdgfra and Fn1 expression in fibroblast clusters i.e., 3, 2, 4, and 19 in control conditions. Fn1 expression levels are inversely proportional to those seen of Pdgfra. Black line across violin plots shows is mean. d, Representative images of stromal whole mount showing fibronectin (FN1, green) expression in lamina propria PDGFRαlow (white arrowhead) and submucosae PDGFRαhigh fibroblasts, labelled in yellow (from induced PdgfraCreERRs26FConfetti mice; EDF 4j; n = 3 mice). e, Stacked bar plot (left) showing fibroblast cluster enrichment across conditions. Significance was assessed by one-sided Chi-squared test, standardised residual (Std Res) values are shown in the heatmap on the right, indicating Cluster 19 (C19) to be the most enriched cluster in tumour conditions f, Volcano plot showing differential gene expression between Tumour and DEN conditions in Cluster 19, Pdgfralow fibroblast cluster enriched in tumours. Non-significant genes (p ≥ 0.05), green and grey. Genes defining the profibrotic nature of cluster 19 in tumours are labelled on the right. g Representative confocal images of a DEN adjacent area and nascent tumours 10 d after DEN treatment or persisting tumour 6 m after DEN treatment showing the accumulation of Fibronectin (FN1, green). h, Second harmonic generation (SHG) image of lamina propria (directly underneath the epithelium) in a control (Ctrl) and tumour areas 3 m post-DEN. Extracellular matrix (ECM) fibres detected by SHG, (cyan); nuclei (red). i, ECM fibre density was scored as SHG signal intensity 3 m after DEN treatment. N is the number of regions or tumours assessed from 6 animals; n = 36 (Ctrl), 20 (DEN), 31 (Tmr). One-way Welch's ANOVA with multiple comparison was used to assess significance. j, Plot depicting the orientation of ECM fibres in control regions and tumours 9 m after DEN treatment. N is the number of fields of view assessed in 3 animals per condition, n = 12 (Ctrl) and 8 (Tmr). k, Representative confocal image of a nascent tumour from 6 animals, showing integrin α6 (CD49f, magenta) sequestering in the early tumour niche. Samples were also labelled for KRT6A (red), PDGFRα (grey) and DAPI (blue). Image settings adjusted to upper stromal layer. l, Venn diagram displaying overlapping genes between known CAF markers (top left, red) and differentially expressed genes (DEGs) in DEN tumour (Tmr) fibroblasts (Cluster 19) relative to external control (Ctrl; top right, yellow) or to internal control cells (DEN; bottom, green). Known subtypes of CAFs: Myofibrotic (my), immune (i), and antigen-presenting (ap) were considered. m,n, Violin plots showing expression of signature genes found to be enriched in Cluster 19 tumour fibroblasts (m) or other canonical CAF markers (n); external control (ctrl, green), internal control (DEN, blue), and tumours (Tmr) (red). o, Representative 3D-rendered side-view images of control tissue and tumours 6-7 m post-DEN. Expression of CAF markers (FAP, VIM, FSP) was not detected in niche fibroblasts. KRT6A (red); CD45 (immune), and CD31 (endothelia) (cyan); Vimentin (VIM) (green); PDGFRα (top left and right) and FSP (bottom), (grey); FAP (green), DAPI, blue. Dashed line outlines tumour area. p, Quantification of Ki67+ EGFP+ fibroblasts from PdgfraEGFP mice in control and Niche+ tumours 6-12 m after DEN treatment. q, Representative bright field (H&E) and confocal (IF) images of invasive carcinoma from 14 m, post-DEN withdrawal tissue. H&E was used to diagnose the pathology. In different panels αSMA, FAP, VIM (green) co-expression with PDGFRα (grey) show CAF emergence in DEN model. Scale bars, 100 µm (black) 50 µm (white); representative data from 5 animals >12 m after DEN treatment. r, Violin plots showing the prediction score from label transfer for Pi16+ and Col15a1+ cross-tissue universal fibroblast population40 in upper GI fibroblast clusters. s, upper GI fibroblast clusters (top left). Remaining UMAPS show Pdgfra expression in UMAP space of upper GI and cross-tissue fibroblasts40. a, Experimental protocol for targeted DNA sequencing of stromal niche 7 m after DEN treatment. b, Substitutions per megabase (Sb/Mb) were calculated and used as indicative of the mutation burden in stroma across conditions. Epithelial tissue (12 m after DEN treatment) was used as benchmark and includes both tumour and tumour-free tissues (DEN+Tmr; from 3 mice). Two-tailed Welch's t-test comparing tumour stroma with internal or external control stroma or with epithelium. c, Stacked bar plot showing scRNA-seq captured cell type distribution across conditions. d, Representative 3D-rendered confocal image of an early tumour (KRT6A, red) 4 m (left) and 7 m (right) post-DEN withdrawal showing the recruitment of immune cells (CD45, green) and blood vessels (CD31, orange) to the persisting tumour niche. e, Heatmap showing expression of representative marker genes across the 14 immune cell types identified in the scRNA-seq data. Log-transformed normalised expression levels were averaged by cluster for each gene and scaled across all cells belonging to each group. f, UMAP of identified immune cell types split by condition. g, Top, stacked bar plot showing immune cell type enrichment across conditions. Significance was assessed by a one-sided Chi-square test. (Bottom) heatmap showing standardised residuals values from Chi-square test. Asterisk (*) marks cell types used in CellChat analysis. h, Heatmaps representing outgoing (left) and incoming (right) interaction strengths as identified by CellChat and rescaled to their row maxima. Cumulative interaction strength is depicted by bars on the right. a, Left (inset), basal keratinocyte cell distribution in the UMAP space annotated by condition (right, UMAP consisting of all cell types). b, UMAP of basal keratinocyte cell distribution representing cells in each condition. Colour bars of UMAPs indicate log-transformed normalised expression levels. d, UMAP showing cell clusters as defined by Seurat. e, Mapped RNA velocity of basal keratinocytes colour coded by condition. Black arrows are velocity vectors that show predicted cellular state change overtime. f, Basal keratinocyte two-dimensional pseudotime (PST) UMAP (top) and trajectories (black line) (bottom) inferred by Monocle 3. Legend shows the range of pseudotime values g, Heatmap (left) of the top 1500 genes identified as differentially expressed between cells in the tumour and basal trajectories. Log-transformed expression levels were scaled in a gene-wise fashion from −4 to 4 (scale) and are shown after hierarchical clustering along trajectories from committed to basal (blue to orange) and from committed to tumour (blue to magenta). The average expression patterns of the genes contained in each cluster are depicted on the right, with two modules (1 and 12) identified as tumour unique based on their preferential upregulation at the end of the tumour trajectory (Δ_tumour > 2, Δ_basal <1). (Blue, committed; Magenta, tumour; Orange, basal). Dashed vertical lines label the highest PST score of committed. Significant genes with FC ≤ 1, red. Genes driving stress and extracellular matrix related processes are labelled. Terms ranked by FDR (false discovery rate); representative genes listed below the bar plot. GO:BP (Gene Ontology Biological Processes); KEGG (Kyoto Encyclopedia of Genes and Genomes). Cartoon (top right) representing Tumour 12 keratinocyte communicates with stroma. m, Representative confocal images from 6 mice showing expression of tumour markers: SOX9 and EGR1 (cyan); AREG and RUNX1, (magenta); KRT6A, red; KRT17 (yellow). DEN images were acquired 10 d after DEN treatment in regions adjacent to tumours. n, Relative frequency (%) (top) of tumours expressing cells positive for SOX9 (Tumour 12 state marker) per 1,000 µm2, tumours 10 d post-DEN, n- number of tumours=119 from 3 biological replicates. o, Representative confocal image from 10 mice demonstrating SOX9+ (cyan) keratinocytes are situated in direct contact to stroma (PDGFRα, grey). a, 3D-rendered confocal images of Niche+ and Niche- tumours showing KRT6A (red) and SOX9 (cyan) 10 d after DEN treatment. b, c Quantification of SOX9+ keratinocytes (krt) (b) and SOX9 intensity (c) in a, normalised to surface area. DEN, equivalent morphologically normal areas from DEN-treated tissues. 3 mice per condition with n = 20 DEN areas, 84 Niche- and 35 Niche+ tumours. Statistical significance assessed by one-way Welch's ANOVA with multiple comparisons. d, 3D-rendered confocal images (basal view, left; side-view, right) of SOX9+ keratinocyte clusters exclusively found in DEN condition. SOX9 (cyan), β catenin (green), PDGFRα (grey), DAPI (blue). e, Quantification of fibroblast proximity to basal layer from d, expressed as the inverse of the distance between basal keratinocytes and underlying fibroblasts. f, Number of fibroblasts directly underneath the basal layer corrected per area. The number of control (n = 27) or SOX9 expressing regions (n = 42) assessed in tissues from 4 mice. g, Strategy to validate SOX9 knock-out in basal keratinocytes (left). SOX9 expression was induced by culturing oesophageal tissue using the 3D epithelioid approach. Established cultures from tamoxifen-induced Krt14CreER R26mTmGSox9flox/flox mice were confocal imaged (right), confirming the presence of areas lacking SOX9 (cyan) expression. h, Top, experimental protocol: Krt14CreERSox9flox/flox mice received a dose of Tamoxifen (TAM) followed by the DEN treatment. Tissues were collected 1 m after DEN treatment (from n = 3 mice) and tumour diameter was compared to DEN-treated un-induced controls (n = 3 mice). Bottom, tumour diameter (µm) of Niche+ and Niche- tumours, from the above. Significance was assessed using two-tailed Welch's t-test. Source data (b,c,e,f,h). a, Scatterplot displaying the dominant senders and receivers as identified by CellChat. Circle size represents ‘communication probabilities'. b, Heatmaps (top) showing the relative importance of each cell group based on the computed network centrality scores for the Fibronectin (FN1) and EGF signalling network, respectively. Violin plots of relevant ligands and receptors across different cell types (bottom). These include expression of fibronectin (Fn1) and fibronectin-binding receptors (Itga3, Itgav, Itgb6), as well as EGF ligands (Areg, Hbegf) and binding receptors (Egfr, Erbb2). c, Migration of fibroblast through cell culture insert membrane (cartoon, top) in low (0.5%) or high (10%) serum (FBS) conditions was quantified (bottom) and expressed as a percentage of the total number of cells. Each dot represents a technical replicate; data 4 biological replicates. One way Welch's ANOVA test with multiple comparison was used to assess significance between groups presented in the main Fig. d, schematic representation of regions imaged in epithelioid cultures from (top) and representative confocal images showing SOX9 (cyan) downregulation in keratinocytes in central epithelioid areas as well as on the border of GFT-treated epithelioids. Representative images of the DMSO (ctrl) condition shown for comparison. Nuclear tdTomato (nT) keratinocytes (krt) (red); CDH1 (green); DAPI (blue). e, Schematic representation of an experiment (left) and representative confocal images (right) of expanding epithelioid cultures exposed to fibroblasts under DMSO (control) or Gefitinib (GFT) conditions. Vimentin, VIM (grey); PDGFRα (red); nuclear Tomato (nT) keratinocytes (krt) (red); DAPI (blue), E-Cadherin, CDH1 (green); Fibronectin, FN1 (green). f, Top, experimental schematics, bottom, quantification of EdU incorporation (2 h chase) in confluent 3D epithelial cultures (Epithelioids) exposed to Fibronectin for 24 h. Data expressed as a percentage of the total number of cells (mean±s.e.m). Each dot represents a technical replicate; data from 3 biological replicates. Two-tailed Welch's t-test comparing EdU incorporation in control (Ctrl) and fibronectin-treated cells. g, Schematic representation of Bleomycin (BLEO) administration to mice (top) and representative confocal images (bottom) of SOX9 (cyan) expression in epithelia of bleomycin treated animals. Illustrations in c,d,e,f,g were created in BioRender; Alcolea, M. https://BioRender.com/nm3lx8y (2026). Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Skrupskelyte, G., Rojo Arias, J.E., Ajith, H. et al. Precancerous niche remodelling dictates nascent tumour persistence. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing: Cancer newsletter — what matters in cancer research, free to your inbox weekly.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. 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. Creating a machine to design functions across the diversity of life would require it to learn a deep, generalist representation of biological complexity. Although this complexity surpasses straightforward human intuition, advances in artificial intelligence offer a universal framework that leverages data and compute at scale to uncover higher-order patterns6,7. We reasoned that training a model with these capabilities would require data spanning the full spectrum of biological diversity to discover emergent properties similar to those found in other fields8. We previously demonstrated that machine learning models trained on prokaryotic genomic sequences can model the function of DNA, RNA and proteins, as well as their interactions that create complex molecular machines1,2. Here we present Evo 2, a biological foundation model trained on a representative snapshot of genomes spanning all domains of life. We extend the sequence modelling paradigm to the scale and complexity of eukaryotic genomes through advances in data curation, model architecture, large-scale pre-training, advanced interpretability methods and inference-time prediction and generation approaches. Emphasizing generalist capabilities over task-specific optimization, Evo 2 represents an important milestone in biological sequence modelling, laying a broad foundation for prediction and design tasks that are relevant to all modalities of the central dogma, that span molecular to genome scale and that generalize across all domains of life. Evo 2 was trained on prokaryotic and eukaryotic genetic sequences, with potential downstream utility for predictive and generative tasks across multiple scales of complexity (Fig. We trained two versions of Evo 2: a smaller version with 7 billion parameters trained on 2.4 trillion tokens (Evo 2 7B), and a larger version with 40 billion parameters trained on 9.3 trillion tokens (Evo 2 40B). This new training dataset, which we call OpenGenome2, was compiled from curated, non-redundant nucleotide sequence data with a total of more than 8.8 trillion nucleotides from bacteria, archaea, eukarya and bacteriophage (Fig. a, Evo 2 models DNA sequence and enables applications across the central dogma, scaling from molecules to genomes and spanning all domains of life. b, Evo 2 was trained on data encompassing trillions of nucleotide sequences from all domains of life. Arabidopsis thaliana, Bacillus subtilis, Bacteroides fragilis, Caenorhabditis elegans, Chlamydomonas reinhardtii, D. melanogaster, E. coli, Gallus gallus, Gorilla gorilla, Haloferax volcanii, Homo sapiens, Mycobacterium tuberculosis, Pan troglodytes, Pseudomonas aeruginosa, S. cerevisiae and Tetrahymena thermophila are highlighted. c, A two-phase training strategy was used to optimize model performance while expanding the context length up to 1 million base pairs to capture wide-ranging biological patterns. d, Novel data augmentation and weighting approaches prioritize functional genetic elements during pretraining and long-sequence composition during midtraining. f, Schematic of the new multi-hybrid StripedHyena 2 architecture, showing the efficient block layout of short explicit (SE), medium regularized (MR) and long implicit (LI) hyena operators. g, Comparison of iteration time at 1,024 GPU, 40B scale between StripedHyena 2, StripedHyena 1 and Transformers, showing improved throughput. h, Validation perplexity of Evo 2 midtraining comparing the model size and context length, showing benefits with scale and increasing context length. i, A modified needle-in-a-haystack task was used to evaluate long context recall ability up to 1 million sequence length, and shows that Evo 2 performs effective recall at 1 million token context. Both Evo 2 7B and 40B are trained in two phases to capture biological length scales from molecular to organismal (Fig. Our first stage of pretraining uses a context length of 8,192 tokens, with data weighting focused on genic windows to learn functional genetic elements, followed by a multi-stage midtraining phase over which we extend the context length of Evo 2 to 1 million tokens to learn the relationships between elements across long genomic distances (Fig. This matches best practice in natural language, in which initial pretraining at shorter context lengths improves both efficiency and overall model quality9,10,11. As in Evo 1, we excluded genomic sequences from viruses that infect eukaryotic hosts from the training data for biosafety purposes. We verified that these data exclusions led to high perplexity on genomic sequences from eukaryotic viruses (Extended Data Fig. 2a), indicating poor language modelling performance in this domain. Evo 2 uses StripedHyena 2, a convolutional multi-hybrid architecture5 that relies on a combination of three different variants of input-dependent convolution operators12 and attention (Fig. 1b), improving training efficiency at scale on both short and long sequences, as well as allowing each layer to model interactions at variable distances. StripedHyena 2 provides substantially higher throughput (at 40 billion parameters, up to 3× speedup at 1 million context length) than highly optimized Transformer6 baselines and previous generation hybrid models based on recurrences or long convolutions, such as StripedHyena 1 (ref. StripedHyena 2 also improves loss scaling on DNA against both Transformers and StripedHyena 1 (Extended Data Fig. 1c), thereby achieving both lower prediction error with the same amount of training data and enabling more efficient use of computational resources. We train up to 1 million base pairs in context length through a multi-stage extension phase, which showed improvements in loss with both model scale and longer context (Fig. With a synthetic long-context evaluation called ‘needle-in-a-haystack', we show that Evo 2 can identify and predict the value of a specific 100 base pair sequence (the needle) hidden within 1 million base pairs of random DNA (the haystack), serving as a synthetic quality check that the model can retrieve information from its full context window, as desired for long-context models (Fig. By learning the likelihood of sequences across vast evolutionary datasets, biological sequence models capture conserved sequence patterns that often reflect functional importance. These constraints allow the models to perform zero-shot prediction without any task-specific fine-tuning or supervision1,14,15,16. Here, likelihood refers to the probability that the model assigns to a given sequence, where mutations that reduce this probability are predicted to be deleterious. Given that Evo 2 learns a likelihood landscape across all three modalities of the central dogma (DNA, RNA and protein) and all three domains of life, we sought to assess whether Evo 2 could perform mutational effect prediction across these modalities and organisms (Fig. a, Evo 2-predicted zero-shot likelihoods can be used to predict the effects of DNA, RNA or protein mutations on molecular function or organismal fitness. b, Effects on Evo 2 prediction of sequence likelihood caused by mutations along gene start sites for various model species across the domains of life. c,d, For different prokaryotic (c) and eukaryotic (d) sequences, the likelihood of different types of mutations in different genomic elements were scored using Evo 2 7B. Scatter represents the median change in likelihood from wild type to mutant sequence per species, coloured by domain (c) or kingdom (d). Shown are the standardized median of delta likelihood values across 5 species, where medians were calculated across approximately 4,100 randomly selected mutation loci. f, DMS assays were used to assess the Spearman correlation of zero-shot likelihoods from models with experimental assays. Notably, Evo 1 and GenSLM were exclusively trained on prokaryotic datasets. g, Schematic of our single-nucleotide resolution exon classifier based on embeddings from Evo 2. h, Single-nucleotide exon classifiers were trained on embeddings from Evo 2, Nucleotide Transformer (NT) and Evo 1, and were evaluated on the basis of their AUROC across eight held-out species. Performance was compared to SegmentNT-30 kb multispecies (asterisks indicate species in SegmentNT training data), ab initio AUGUSTUS, and to baseline nucleotide content and conservation metrics. j, Evo 2 predicts genes as essential or nonessential, as determined by experimental gene essentiality assays across bacterial, archaeal and phage species (shown as overlaid scatter) using mutational likelihood of premature stop codon insertions (as a genetic perturbation). To assess whether Evo 2 captures core biological principles, we first evaluated how single nucleotide variants (SNVs) affect Evo 2 likelihoods in the genomic sequences around the start codons of protein-coding genes. We introduced these mutations at each position in the wild-type sequence and calculated the resulting changes in Evo 2 predicted likelihoods across thousands of such loci (Fig. We observed strong changes in the likelihood for mutations within the start codons in both prokaryotes and eukaryotes. This was followed by a three-base periodicity pattern reflecting the triplet codons, with changes at the wobble positions showing lower impact on likelihood. For both prokaryotic and eukaryotic genomes, we observed a pattern upstream of the coding DNA sequence (CDS) that was consistent with the locations of known consensus sequences associated with translation initiation, namely, the Shine–Dalgarno sequence17 for prokaryotes and the Kozak sequence18 for eukaryotes. We also observed similar patterns for SNVs around stop codons (Extended Data Fig. Next we measured the effect of mutations across a variety of both noncoding and coding sequences (Fig. Across 20 prokaryotic species and 16 eukaryotic species, we observed changes in model likelihoods consistent with known biological constraints. In noncoding regions, deletions in transfer RNAs (tRNAs) and ribosomal RNAs (rRNAs) had much larger effects than deletions in intergenic and other noncoding loci, reflecting the known essential roles of these RNAs. The 40B model exhibited higher sensitivity to deletions in microRNA (miRNA) and small nucleolar RNA (snoRNA) sequences compared with the 7B model. Evo 2 also predicted that less efficiently translated codons had lower likelihoods than more efficient codons (Extended Data Fig. Recognizing that our training data contained genomes with distinct genetic codes, we tested how different premature stop codons impacted species that differ in their stop codon usage (Fig. When ciliate genomes were artificially recoded to the standard genetic code, Evo 2 predicted mutations from the standard stop codons as deleterious, demonstrating that the model relies on sequence context to determine the appropriate genetic code (Extended Data Fig. Although Evo 2 likelihoods reflect the expected importance of different genetic alterations, a key question is whether these likelihoods also correlate with functional effects, which can be empirically measured via deep mutational scanning (DMS) of proteins and noncoding RNAs (ncRNAs). Although state-of-the-art methods for this task tend to leverage both sequence alignments and structural conditioning, general-purpose single-sequence protein language models also learn likelihood distributions that correlate with fitness15. Evo 2 sequence likelihoods correlate with diverse definitions of fitness across nine prokaryotic protein datasets; six eukaryotic protein datasets; and seven datasets of rRNAs, tRNAs and ribozymes (Fig. Evo 2 is competitive with widely used ProGen language models for protein DMS and with RNA language models for ncRNA DMS, although it underperforms state-of-the-art models on protein DMS. Consistent with observed trends for protein language models, the performance of Evo 2 on these fitness prediction benchmarks begins to saturate and can decrease at the largest model scales19,20,21. We also tested the ability of Evo 2 to predict mutation effects in protein sequences from viruses that infect human hosts. We found no correlation between Evo 2 likelihood and viral protein fitness (Extended Data Fig. 2b), consistent with our data exclusions having the intended effect of weakening both language modelling and downstream performance (Extended Data Fig. Evo 2 likelihoods also have modest zero-shot association with human mRNA decay rates (Extended Data Fig. Since Evo 2 learns from eukaryotic genomes, which can be challenging to annotate, we assessed whether its embeddings capture exon–intron architecture. We trained lightweight models on Evo 2 7B base embeddings to develop single-nucleotide resolution classifiers of exon labels (Fig. On eight diverse species held out from classifier training, our best classifier achieved areas under the receiver operating characteristic curve (AUROCs) ranging from 0.91 to 0.99 (Fig. 2h,i), outperforming models trained on embeddings from other genomic language models, Nucleotide Transformer22 and Evo 1 (ref. As a practical baseline, we show that our classifier outperforms ab initio AUGUSTUS23 across all species tested. Evo 2 also outperforms SegmentNT24 on all species outside the SegmentNT training set and on one of the three species in its training set. These results suggest that combining Evo 2 sequence embeddings with supervised approaches can aid the functional annotation of genetic components across diverse species, including non-model organisms. Beyond molecular or gene-level prediction tasks, we previously showed that high likelihood under Evo 1 is associated with whole organism replication fitness in prokaryotes and phage as quantified by gene essentiality experiments1. Using zero-shot likelihoods to score the effects of premature stop codon insertions into bacterial, archaeal and phage genomes, we found that Evo 2 models performed similarly to Evo 1 and better than other zero-shot methods in predicting gene essentiality across diverse species (Fig. On zero-shot prediction of human gene essentiality (Methods), Evo 2 40B (AUROC = 0.66, area under the precision-recall curve (AUPRC) = 0.15) outperformed other genomic language models (AUROC range 0.50–0.59, AUPRC range 0.09–0.12) and performs within the range of four PhyloP conservation scores (AUROC range 0.65–0.71, AUPRC range 0.13–0.21) (Extended Data Fig. 3i), although the overall predictive performance remains modest. Together, these results demonstrate that Evo 2 captures information across biological modalities and domains of life. Notably, the 7B and 40B models expand predictive capabilities without compromising the prokaryotic insights captured by Evo 1. The utility of both zero-shot likelihoods and simple classifiers trained on Evo 2 embeddings for a variety of predictive tasks across prokaryotic and eukaryotic genomes indicates that Evo 2 provides a strong foundation model for downstream applications in computational biology. Variant effect prediction represents a critical challenge in genomics, with direct implications for clinical diagnosis and therapeutic development. Genomic language models have previously struggled in eukaryotic variant effect prediction, lagging considerably behind species-specific models that use multiple sequence alignments16,22,25. Evo 2 can perform accurate zero-shot variant effect prediction for both coding and noncoding DNA by considering the changes in the model's likelihoods after introducing mutations involving single or multiple nucleotides (Fig. a, Overview of zero-shot variant effect prediction using Evo 2. Evo 2 was used to assign likelihood scores to human genetic variants, distinguishing pathogenic and benign variants in both coding and noncoding regions. Shown are the AUROCs and AUPRCs for classifying pathogenic and benign variants from ClinVar, across models. For non-SNV evaluations, a modified version of PhyloP was used (Methods). d, Zero-shot evaluation on splice-altering variants in SpliceVarDB, split by exonic (n = 1,181) and intronic (n = 3,769) scoring. e, Evo 2 and other models were used to evaluate BRCA1 variant effect predictions against BRCA1 saturation mutagenesis data, comparing classification of loss-of-function versus functional and intermediate variants in both coding (n = 2,077 SNVs) and noncoding (n = 1,125 SNVs) regions. f, Evo 2 zero-shot likelihood scores plotted for loss-of-function (LOF) versus functional/intermediate variants (n = 3,893), demonstrating the ability of Evo 2 to separate these classes. P value calculated by two-sided Wilcoxon rank sum test. g, Evo 2 embeddings were extracted and concatenated to train a supervised classifier for BRCA1 variant effect prediction. h, Predictions of the supervised classifier on functional/intermediate variants compared with true loss-of-function variants on the test set (n = 789), with decision scores on the horizontal axis. P value calculated by two-sided Wilcoxon rank sum test. i, Comparison of a supervised classifier trained on Evo 2 embeddings on the BRCA1 test set against zero-shot baselines, highlighting the value of using Evo 2 embeddings to build lightweight supervised models. We used annotations of human clinical and experimentally determined variants to evaluate the ability of Evo 2 to predict biologically important sequence variation. We also contextualize the performance of Evo 2 against a wide range of models, including statistical measures of conservation (for example, PhyloP); unsupervised language models of proteins, RNA and DNA (for example, ESM-1b); supervised splicing prediction models (for example, Pangolin and SpliceAI); and human variant effect prediction models (for example, AlphaMissense, GPN-MSA and CADD). Using the ClinVar database, we compared the ability of Evo 2 against other methods for predicting the pathogenic effects of human genetic variants across diverse variant classes (Supplementary Data 1). For coding region SNVs, the 40B and 7B models performed competitively, ahead of zero-shot methods, including ESM-2, but behind ESM-1b, GPN-MSA and some PhyloP variants (Fig. For non-SNV coding variants (for example, insertions and deletions), both Evo 2 models outperformed all other methods; notably, these non-SNV variants are not possible to score by leading models such as AlphaMissense and GPN-MSA (Fig. For noncoding SNVs, Evo 2 40B ranked first among unsupervised models and only trailed behind supervised models (Fig. For noncoding non-SNVs, Evo 2 40B outperformed all models tested (Fig. Across variants stratified by levels of conservation or distance from splice sites, Evo 2 maintains competitive performance among unsupervised models for noncoding variants and the best performance for coding and noncoding non-SNVs out of all methods tested (Extended Data Fig. To further evaluate performance on splice variants, we used SpliceVarDB, a repository containing experimentally validated splicing effects. On intronic variants, zero-shot prediction with Evo 2 was competitive with supervised models, slightly trailing SpliceAI and CADD but ahead of Pangolin; on exonic variants, Evo 2 trailed specialized supervised models but outperformed all zero-shot models (Fig. We next focused on a dataset measuring functional consequences of variants across both exons and introns of the BRCA1 gene26. Zero-shot prediction with Evo 2 exhibited strong performance on coding SNVs and outperformed all other models on BRCA1 noncoding SNVs (Fig. Evo 2 7B and 40B achieved better performance than other models when coding and noncoding SNVs were evaluated together, suggesting well-calibrated predictions across included variant types (Extended Data Fig. When separately considering BRCA1 noncoding variants near or far from splice sites, Evo 2 40B outperformed all tested models, including supervised splicing predictors (Extended Data Fig. A recently released BRCA2 variant dataset with experimental measurements27 enabled us to extend this analysis to a related gene. Evo 2 surpassed specialized models such as GPN-MSA when predicting coding and noncoding variants together, achieving second-best performance behind CADD, a supervised model (Extended Data Fig. These results indicate that Evo 2 is an effective zero-shot predictor across diverse types of functional human variants. Although zero-shot scoring is particularly valuable when task-specific training data are unavailable, model-derived embeddings can also serve as inputs to supervised classifiers that learn task-specific decision boundaries, thereby enhancing both sensitivity and specificity. To illustrate this capability, we assessed whether a simple ridge regression model trained with Evo 2 embeddings exclusively on BRCA1 variants could surpass zero-shot prediction with Evo 2 (Fig. Given that different layers within large language models capture distinct features, we systematically extracted sequence embeddings from each block of the Evo 2 40B model to identify which layer yielded the most informative features for variant classification (Extended Data Fig. Our supervised model achieved a clear separation between loss-of-function variants and all other variants (Fig. 5e), outperforming zero-shot prediction by Evo 2 40B on the test set (AUROC = 0.95, AUPRC = 0.88) (Fig. These results underscore how Evo 2 embeddings can be harnessed to train models aimed at more specialized tasks, including those with high clinical relevance. Unlike the highly constrained sequences typically found in clinical variant datasets which are biased towards coding, splicing or untranslated region (UTR) variants, other regulatory sequences—particularly those distal to genes—exhibit substantially lower conservation. In this context, we used DART-eval to assess how effectively Evo 2 embeddings and likelihoods capture regulatory function28. On zero-shot tasks in DART-eval, Evo 2 40B (chromatin accessibility quantitative trait loci (caQTL) AUROC = 0.58, DNase I sensitivity quantitative trait loci (dsQTL) AUROC = 0.66) outperforms other unsupervised DNA language models, such as Nucleotide Transformer (caQTL AUROC = 0.52, dsQTL AUROC = 0.61), but trails sequence-to-function models trained on accessibility data, such as ChromBPNet (caQTL AUROC = 0.77, dsQTL AUROC = 0.89) (Extended Data Fig. These results indicate that while multi-species language models trained on sequence alone capture some regulatory information, sequence to function models with task-specific training achieve higher performance in this setting. On human clinical variant prediction, Evo 2 represents a major improvement over previous multi-species DNA language models across different variant types, with leading performance on non-SNVs (insertions, deletions, duplications), and maintains this performance even in the absence of strong site-independent sequence conservation, although it falls behind supervised models for distal regulatory variants. Furthermore, leveraging the representations in a supervised setting illustrates how Evo 2 embeddings can serve as a foundation for downstream prediction tasks. Notably, Evo 2 is not trained on any human genetic variation or functional genomics data. In sum, these findings support the versatility of Evo 2 as a genome-scale language model for both unsupervised and supervised variant effect prediction. Evo 2 learns complex representations of genomic sequences without explicit biological labels or annotations. Contrary to the common critique of large language models as black box systems, recent advances in the field known as mechanistic interpretability have demonstrated that sparse autoencoders (SAEs) can reveal latent dimensions that correspond to semantically meaningful features in natural language29,30,31. Without any prior biological annotations or labels, we trained SAEs on Evo 2 representations (or neuron firing patterns), to decompose the model into sparse, high-dimensional representations in which each latent dimension often exhibits human-interpretable patterns (Fig. a, SAEs were trained on Evo 2 to extract SAE features associated with interpretable biological function that can be used for annotation, discovery and steering of sequence generations. b, Phage-associated feature activates preferentially on RefSeq-annotated prophages (left and top right) in the E. coli K12 MG1655 genome and fires on phage-derived spacer sequences within CRISPR arrays (bottom right). c, Activations of features associated with ORFs, intergenic loci, tRNAs and rRNAs, in a 100-kb region in E. coli K12 MG1655. d, Activations of features associated with α-helices, β-sheets and tRNAs at an E. coli K12 MG1655 locus containing tufB and a tRNA array ending with thrT (left) and the rpoB–rpoC locus (right). AlphaFold 3 (AF3) structure predictions with feature activations overlaid, of EF–Tu in complex with the tRNA (left) and of RpoB and RpoC in complex (right). e, A feature in the human genome with preferential activation immediately after frameshift mutations over other less deleterious mutation types. f, Features with activation on DNA motifs in the human genome that correspond to transcription factor-binding motifs. g, Features associated with exons, introns and their boundaries in the human genome generalize to a segment of the woolly mammoth genome. We trained a Batch-TopK SAE32 on Evo 2 representations from layer 26 (Methods). The SAE was trained on representations from 1 billion tokens evenly split across several complete eukaryotic and prokaryotic genomes (Extended Data Fig. We matched learned SAE latent dimensions, also referred to as features, and known biological concepts by finding features that were enriched in sequence segments containing a particular annotation, a process that we refer to as contrastive feature search (Extended Data Fig. This revealed diverse features that align with known biological concepts. For example, Evo 2 developed internal representations associated with mobile genetic elements. Feature f/19746 is closely associated with prophage regions across prokaryotes (Extended Data Fig. 7b) and activates on annotated prophages in the Escherichia coli genome, including the cryptic prophage CPZ-55 (Fig. This feature also activates on spacer sequences within a CRISPR array, which are integrated during CRISPR adaptation from foreign genetic material such as phage DNA (Fig. 4b), as well as after the last CRISPR direct repeat and on synthetic, scrambled spacer sequences, suggesting that Evo 2 associates CRISPR spacers with phage sequences as opposed to directly memorizing phage genomes (Fig. This feature also activates on other regions that are not annotated as phage by geNomad33 yet contain genes associated with prophages, such as integrases and invertases (Extended Data Fig. Next, we sought to identify concepts associated with canonical biological genomic elements. We identified diverse features corresponding to open reading frames (ORFs), intergenic regions, tRNAs and rRNAs in the E. coli genome (Fig. We further probed for structural signatures at the protein level and identified features linked to protein secondary structures, such as α-helices and β-sheets (Fig. These associations highlight the multimodal nature of genome language modelling, capturing higher-order structural information beyond DNA alone. We extended our analysis to the human genome in search of eukaryotic features. By introducing mutations into thousands of human coding sequences and applying contrastive feature search on a eukaryotic-only SAE, we identified a mutation-sensitive feature (f/24278) that preferentially activates on frameshifts and pre-mature stop mutations (Fig. We also observed other activations on DNA motifs in the promoter regions of human genes (Fig. 4f, left) that closely resemble the known binding sites of human transcription factors (Fig. Across a random sample of human promoter sequences, Evo 2 unsupervised SAE features have significant hits (q < 0.01, Sandelin–Wasserman similarity) to 70% promoter-enriched motifs (Extended Data Fig. For comparison, HOMER36, a specialized motif discovery algorithm, only recalls 35% of the same motifs (Extended Data Fig. We provide a full report on transcription factor motif-associated features in Supplementary Data 2. These results suggest that Evo 2 contains distinct internal representations of noncoding regulatory elements. Finally, we identified features that were closely associated with the exon and intron architecture of the human genome, including features that activate preferentially on coding regions (f/15680), introns (f/28339), the first bases of an exon following an intron (f/1050), and the last base of an exon followed by an intron (f/25666) (Extended Data Fig. The coding region feature also activates on bacterial ORFs, suggesting a learned universal representation of coding sequences (Extended Data Fig. For instance, exon boundary features (f/1050 and f/25666) integrate signals across splice sites that span multiple nucleotides, and the prophage feature (f/19746) identifies mobile genetic elements requiring kilobase-scale context. Although we identified these features on the human genome using an SAE trained only on model organisms (including primates, Mus musculus, Xenopus tropicalis and Drosophila melanogaster), we further observed that these features transferred to a genic region within a portion of the woolly mammoth genome37 (Fig. These results demonstrate that an Evo 2 SAE learns features that transfer across species and suggest utility for genome annotation, although systematic benchmarking against established annotation tools remains necessary. Overall, we demonstrate that Evo 2 latent representations capture a broad spectrum of biologically relevant signals, from prokaryotic mobile genetic elements and eukaryotic regulatory motifs to protein secondary structure and mutational severity. Since conceptual features for natural language can capture abstract concepts, other Evo 2 SAE features could represent more complex biological patterns (Extended Data Fig. We have released the SAE models and a visualization tool to facilitate exploration of Evo 2 features for the scientific community. Beyond its utility on prediction tasks, Evo 2 is also a generative model. We therefore sought to generate DNA sequences from diverse organisms with Evo 2 and assess the quality of designed sequences (Fig. To evaluate the ability of Evo 2 to respond to genomic prompts, we first assessed performance across six diverse species, spanning archaea, prokaryotes and four eukaryotic lineages (fungi, protists, plants and animals). For each species, we selected highly conserved representative genes and prompted Evo 2 with 1,000 base pairs of upstream sequence plus the first 500–1,000 base pairs of the target gene. We found that Evo 2 achieves gene completion with high amino acid sequence recovery, which improved with scale (Fig. Evo 2 40B and 7B also demonstrated improved performance over Evo 1 and maintained high sequence recovery throughout long context training (Fig. a, Evo 2 can generate chromosome- and genome-scale DNA sequences using unconstrained autoregressive generation. The model was prompted with portions of the H. sapiens mitochondrial genome, M. genitalium genome and S. cerevisiae chromosome III to generate DNA sequences with similar lengths to those of the native sequences. c, Predicted rRNA, CDS and tRNA counts in Evo 2-generated mitochondrial sequences using MitoZ compared with the natural H. sapiens mitochondrial genome values. d, Query cover versus sequence identity of generated mitochondrial sequences against nucleotide BLAST hits in the core_nt database with expect threshold of 0.05, coloured by the E-value. e, Visualizations of Evo 2-generated sequences when prompted with a 3-kb sequence from the H. sapiens mitochondrial genome, demonstrating variation that still retains natural synteny patterns of coding sequences. f, AlphaFold 3-predicted structure of multimeric complexes from an Evo 2-generated sequence resembling human mitochondrial DNA. Genes are annotated with Prodigal and coloured on the basis of statistically significant sequence similarity to natural proteins (hmmscan E-value < 0.001). h, The fraction of Prodigal-annotated genes with hmmscan hits between Evo 2 40B and M. genitalium generated by Evo 1. i, Distribution of Prodigal-annotated genes from Evo 2-generated M. genitalium compared with the natural genome. k, AlphaFold 3 structure predictions of example proteins found on Evo 2-generated prokaryotic genomic sequences, with high observed structural similarities to natural proteins while diversifying the sequence com-position. l, The native genome sequence from S. cerevisiae chromosome III and an Evo 2-generated DNA sequence of similar length, which was generated by prompting the model with a 10-kb sequence from S. cerevisiae chromosome III, are visualized alongside predicted homologous yeast gene, exon, promoter and tRNA annotations. Consistent with poor performance for viruses that infect humans on the language modelling task and on function prediction downstreams, Evo 2 also has poor performance on generating proteins from human viruses (Extended Data Fig. Even when directly trying to elicit a viral protein, Evo 2 had essentially random performance in sequence recovery, preventing Evo 2 from unconstrained or accidental generation of human viral proteins. To test the ability of Evo 2 to generate DNA at the scale of entire genomes, we assessed its ability to generate all known components of a natural organelle genome. We prompted Evo 2 7B and 40B with portions of human mitochondrial DNA, generating over 250 unique 16-kb sequences (Methods). When annotated with MitoZ38, we found that the generated sequences have the correct number of CDSs, tRNA genes and rRNA genes expected in human mitochondria (Fig. 5c), with varying degrees of sequence similarity to natural genes (Fig. 5d and Supplementary Table 7) while maintaining proper synteny (Fig. Evo 2-generated mitochondrial sequences contained proteins with predicted multimeric complexes matching those of human mitochondrial proteins (Fig. The codon usage of generated sequences also closely matched that of the human mitochondrial genome (Extended Data Fig. 9d) and Evo 2 generations on average successfully generated one of each type of expected tRNA, without duplicating the two included in the prompt (Extended Data Fig. We next leveraged the million-base-pair context window of Evo 2 to generate DNA sequences at the scale of small prokaryotic genomes. For this task, we focused on M. genitalium, a model minimal genome of length approximately 580 kb (refs. Using a 10.5-kb segment from the M. genitalium reference sequence as the prompt, we generated ten 580-kb sequences that we annotated with Prodigal41. Nearly 70% of generated genes contained significant Pfam hits, a substantial improvement over Evo 1 131k (18%) (Fig. Generated proteins also have structural alignment to natural proteins, though the structure prediction confidences of generated proteins are relatively lower than those of natural genes (Fig. To assess the eukaryotic sequence generation capability of Evo 2, we prompted it with 10.5 kb from Saccharomyces cerevisiae chromosome III (approximately 316 kb in length) to generate 20× 330-kb DNA sequences. These sequences include tRNAs, promoters and genes with intronic structure (Fig. 9i and Methods) The density of tRNA and gene features was below those found in the native yeast genome (Fig. 5l), though the generated genes had similar length distributions to natural proteins (Extended Data Fig. Generated genes also demonstrate varying structural similarity to natural proteins while demonstrating sequence diversity (Extended Data Fig. A common measure of phylogenetic relatedness, the tetranucleotide usage deviation (TUD) of generated S. cerevisiae sequences correlates with that of native S. cerevisiae, and this agreement is higher for Evo 2 40B than for Evo 2 7B (Extended Data Fig. These results demonstrate that Evo 2 can generate DNA sequences that resemble organellar, prokaryotic and eukaryotic genomes on the basis of several in silico metrics. Of note, these evaluation metrics do not guarantee functional or replication-competent genomes, and our genome-scale generations lack important elements, such as some essential genes. Experimentally testing genome-scale designs will also require large-scale, iterative effort42. However, Evo 2 provides a stronger and more versatile foundation model for genome-scale generation than Evo 1 and can also generate eukaryotic sequences. Inference-time guidance of generative models via a separate scoring or reward function has enabled powerful domain-specific conditioning in tasks such as code generation, algorithm design and mathematical reasoning43,44. We therefore sought to demonstrate how Evo 2 can be guided to generate long genomic sequences that would not be sampled naturally by standard autoregressive generation. We focused on designing genomic segments to have artificial chromatin accessibility patterns45. Previous methods have used gradient-based, diffusion-based or black box optimization guided by oracle scoring functions to successfully design regulatory elements in the hundreds of nucleotides46,47,48,49,50,51. Building off of these methods, especially those that leverage black box optimization, we demonstrate that coupling a generative language model with inference-time guidance enables de novo design of multi-kilobase sequences with controllable chromatin accessibility. Although Evo 2 does not explicitly learn chromatin accessibility, we can still leverage inference-time guidance to enable epigenomic conditioning. Models such as Enformer3 and Borzoi4 can predict chromatin accessibility from DNA sequences across cell types from human and mouse. However, Enformer and Borzoi are not generative models and are exclusively trained on natural genomes. We guided Evo 2 to generate DNA sequences for which we can specify the location and length of chromatin-accessible regions, which are often visualized as ‘peaks' along a one-dimensional genome sequence (Fig. We used an ensemble of Enformer and Borzoi to define a scoring function that accepts or rejects generated sequences on the basis of how well their predicted chromatin accessibility matches a desired pattern. Instead of sampling and scoring full, multi-kilobase designs, we conducted a beam search that re-evaluates Enformer and Borzoi after each new 128 bp of sampled sequence and only continues autoregressive generation off of the most promising samples (Fig. We used DNase hypersensitivity predictions in 129 ES-E14 cells from both Enformer and Borzoi, using the natural mouse genome context to inform the predictions (Methods). a, Multi-kilobase sequences were designed to control the locations and lengths of chromatin-accessible regions, which are visualized as peaks indicating the degree of accessibility along a one-dimensional genomic sequence. b, 128-bp DNA chunks from a prompt were autoregressively generated with Evo 2. A beam search algorithm then selects the optimal chunks by scoring how well their Enformer- and Borzoi-predicted chromatin accessibility profiles match a target pattern. c, Design runs are plotted by how successfully they matched the target pattern versus the compute used. AUROC quantifies how well predicted accessibility profiles can distinguish our desired open- versus closed-chromatin positions. Individual design runs are plotted as grey dots and the averages across design runs for each beam search width are plotted as crosses. d, Two different peak patterns were designed with varying total compute budgets, with more compute leading to clearer designed peaks. e, Designs were experimentally tested by synthesizing and assembling the DNA, performing site-specific integration into mouse or human cells, and measuring chromatin accessibility with ATAC-seq. f–h, Control over the position and width of chromatin accessibility peaks enables Morse code messages (‘EVO2' (f), ‘LO' (g) and ‘ARC' (h)) in the epigenome. Enformer and Borzoi predictions are based on the DNase hypersensitivity tracks in 129 ES-E14 cells. Designs were sampled using 30–84 tokens per base pair (Methods). i,j, Integrating the same generated sequence into both HEK293T and K562 cells enables the design of identical patterns across both cell types (i) or of cell-type-specific accessibility profiles (j). k, AUROC quantifies how well experimental accessibility profiles can distinguish our desired open- versus closed-chromatin positions. Five designs were tested in HEK293T and 31 designs in K562 for which we varied the chromatin accessibility along the sequence. l, The paradigm of using an accurate scoring function to guide a capable generative model extends beyond chromatin accessibility design, enabling many complex biological design applications. Sampling 30 or more 128-bp chunks and selecting the top two chunks at each step of the beam search was sufficient to achieve final designs with AUROCs above 0.9 (Fig. In some of our designs, we varied the length and location of accessible regions to write Morse Code messages, where narrow peaks indicate dots, wide peaks indicate dashes, and inaccessible regions indicate spaces. We experimentally tested these predictions by synthesizing and assembling the designed DNA, performing site-specific integration into the genomes of mouse embryonic stem cells (mESCs), and measuring chromatin accessibility with assay for transposase-accessible chromatin using sequencing (ATAC-seq) (Fig. The three Morse code patterns had good agreement between the predicted and experimentally measured chromatin profiles, with the experimental patterns having AUROCs of 0.92–0.95 (Fig. We also observed experimental success (AUROC > 0.89) for simpler peak patterns (Extended Data Fig. Using a capable generative model to propose sequences outperforms simpler proposals and produces designs with favourable properties that emerge without direct optimization. For example, prompting Evo 2 with the native genomic context resulted in natural dinucleotide frequencies (Extended Data Fig. By contrast, when we repeated the same design pipeline except with a uniform or bigram proposal replacing Evo 2, we observed poorer token-matched inference-time scaling and poor consensus among ensemble predictions (Extended Data Fig. Given that agreement among ensembled predictors was, retrospectively, critical for experimental design quality (Extended Data Fig. 10a), we hypothesize that these simpler proposals are prone to generating adversarial samples as observed in other domains, such as protein design52,53. Further analysis of our Morse Code designs revealed that designed peak regions contain significantly higher predicted transcription factor motif density than non-peak regions (two-sided Welch's t-test, P = 3.6 × 10−7) (Extended Data Figs. Unlike unigram or bigram designs, Evo 2 designs were significantly enriched for transcription factors expressed in mESCs (one-sided hypergeometric P = 2.0 × 10−4), with GC and CpG density correlating with chromatin peaks at levels similar to natural sequences (Extended Data Figs. Notably, Evo 2 was not explicitly conditioned to generate motif-rich sequences. These findings suggest that Evo 2 can serve as a powerful generative proposal with sequence outputs containing greater regulatory potential than simpler methods. We then demonstrated the generality of this approach by designing and experimentally testing the chromatin accessibility profiles of 1-4 kb sequences in two human cell lines, HEK293T and K562 (Methods). We tested designs with predicted chromatin accessibility patterns that differ between HEK293T and K562 cells, as well as designs with the same pattern in both cell lines (Fig. We observed strong experimental success rates when we varied the level of chromatin accessibility within a designed sequence, with 33 out of 36 designs (92%) having an AUROC greater than 0.8 (Fig. When designing regions with differential accessibility between 2 cell types, a much more challenging task, we observed that 4 out of 24 of these designs (17%) had greater than twofold differential accessibility, and 1 out of 24 designs (4%) had greater than threefold differential accessibility (Fig. In the four designs with more than twofold differential accessibility, the predicted transcription factor motifs in the design peaks were significantly enriched for K562-expressed transcription factors (one-sided hypergeometric P = 0.0017) (Extended Data Fig. 11i) but not for HEK293T-expressed transcription factors (one-sided hypergeometric P = 0.25) (Extended Data Fig. This design task shows how Evo 2 can be coupled with task-specific supervised models to achieve controllable design of mammalian chromatin architecture. While beam search requires increasing inference-time compute (Fig. 6c) to improve generations, it is also highly flexible, requires no additional training compute and can leverage non-differentiable scoring functions. We note that other application-specific models could also be used to guide Evo 2's generations (Fig. 6l), enabling biological design in any downstream application for which there exists a capable predictive model. Here we report a genomic language model, Evo 2, that achieves generalist prediction and design capabilities across all domains of life. Developing Evo 2 required substantial investment in machine learning research and engineering5, as well as data curation and evaluations. We provide several resources under an open-source license, including the following: (1) parameters for the Evo 2 models; (2) distributed training code; (3) code for multi-GPU inference; and (4) the full OpenGenome2 training dataset (Data availability and Code availability). The Evo 2 40B 1 million (1M)-context model demonstrates best overall performance, though the 7B 1M-context model is competitive and useful for settings requiring lightweight inference. Although we also release an experimental one-billion-parameter short-context model (Supplementary Table 1), this version should be avoided owing to overall weaker performance. We also release a tool for generating and scoring sequences with Evo 2 40B in a simple web interface (at https://arcinstitute.org/tools/evo/evo-designer) and a tool for exploring SAE features alongside genomic annotations (at https://arcinstitute.org/tools/evo/evo-mech-interp). Evo 2 is one of the largest-scale fully open models thus far (including training and inference code, data and parameters), even across other modalities, such as language and vision. As with all new biotechnologies, there are safety, security and ethical considerations. Aligned with the Responsible AI × Biodesign commitments (https://responsiblebiodesign.ai/), we preemptively assessed and mitigated potential concerns prior to open source publication. Fully open-source models enable researchers to interrogate, reproduce, and build upon advances in artificial intelligence. They may also be used in unanticipated ways that could lead to accident or misuse risks54. We collaborated with multidisciplinary experts to reduce risks via data exclusion measures, safety and security evaluations, and population bias evaluations (Methods). By excluding genomic sequences of viruses that infect eukaryotes from our training data, we aimed to ensure our openly shared model did not disseminate the capability to manipulate and design pathogenic human viruses. Task-specific post-training may circumvent this risk mitigation measure and should be approached with caution. Our data exclusions had the intended outcomes of weakening language modelling performance (Extended Data Fig. 2a) and downstream mutational effect prediction (Extended Data Fig. Probing and testing these measures by red teaming meant to directly elicit pathogenic human viral proteins showed generations were effectively random in this domain (Extended Data Fig. We also showed that the population-free design of Evo 2 mitigated ancestry biases in model predictions55 (Extended Data Fig. Few examples of empirical risk assessment of biological foundation models exist; this work represents one of the most comprehensive evaluative efforts thus far that considers both precaution and access. Further research is also needed to expand the suite of available evaluations and risk mitigation approaches. Evo 2 offers a powerful foundation for future work. Combining Evo 2 with additional information such as population-scale genomic variation56,57 or data from sequence-to-function experiments3,58 could enable an even greater breadth of downstream tasks. Whereas our mechanistic interpretability analysis focused primarily on well-annotated features, future work could leverage these approaches for genome mining and the discovery of more complex combinations of biological elements. Although Evo 2 generates more realistic DNA sequences than Evo 1, the current generative evaluations described in this study do not guarantee that the sequences will function in cells. Improving generation with inference-time guidance can notably require computationally intensive sampling. Supervised fine-tuning and reinforcement learning with feedback from biological experiments is likely to improve the efficiency and quality of sequences generated by Evo 2 for complex applications. The Evo series of models lays the groundwork for biological modelling and design that unifies the diverse length scales of biology with a common representation. These capabilities, combined with large-scale DNA manipulation59, may enable programmable design of more complex biological functions. We expect that future work integrating genomic sequence data with additional modalities could produce a model that productively simulates complex phenotypes in health and disease. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2. Raw reads for ATAC-seq analysis of mESC chromatin accessibility designs have been uploaded to the Sequence Read Archive (SRA) under BioProject accession number PRJNA1314301. Raw reads for ATAC-seq analysis of human chromatin accessibility designs have been uploaded to the SRA under BioProject accession number PRJNA1314272. Code and tools for model exploration are available at the following links: top-level code repository: https://github.com/arcinstitute/evo2; pretraining, midtraining and fine-tuning code: https://github.com/zymrael/savanna; inference code: https://github.com/zymrael/vortex; Evo Designer, an interactive user interface for generation and scoring with Evo 2: https://arcinstitute.org/tools/evo/evo-designer; Evo Mech Interp Visualizer, an interactive user interface for exploring SAE features: https://arcinstitute.org/tools/evo/evo-mech-interp; NVIDIA Evo 2 NIM (generation): https://build.nvidia.com/nvidia/evo2-protein-design; NVIDIA Evo 2 NIM (forward): https://build.nvidia.com/arc/evo2-40b; NVIDIA BioNeMo version of Evo 2 code: https://github.com/NVIDIA/bionemo-framework. Nguyen, E. et al. Sequence modeling and design from molecular to genome scale with Evo. Merchant, A. T., King, S. H., Nguyen, E. & Hie, B. L. Semantic design of functional de novo genes from a genomic language model. Effective gene expression prediction from sequence by integrating long-range interactions. Linder, J., Srivastava, D., Yuan, H., Agarwal, V. & Kelley, D. R. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Ku, J. et al. Systems and algorithms for convolutional multi-hybrid language models at scale. Scaling laws for neural language models. Language models are unsupervised multitask learners. Gao, T., Wettig, A., Yen, H. & Chen, D. How to train long-context language models (effectively). 63rd Annual Meeting of the Association for Computational Linguistics 1, 7376–7399 (ACL, 2025). In vivo perturb-seq of cancer and microenvironment cells dissects oncologic drivers and radiotherapy responses in glioblastoma. Poli, M. et al. Hyena hierarchy: towards larger convolutional language models. 40th International Conference on Machine Learning (eds Karuse, A. et al.) 28043–28078 (2023). Poli, M. et al. Mechanistic design and scaling of hybrid architectures. 41st International Conference on Machine Learning 235, 40908–40950 (2024); https://proceedings.mlr.press/v235/poli24a.html. Notin, P. et al. ProteinGym: large-scale benchmarks for protein design and fitness prediction. Benegas, G., Albors, C., Aw, A. J., Ye, C. & Song, Y. S. A DNA language model based on multispecies alignment predicts the effects of genome-wide variants. & Dalgarno, L. The 3′-terminal sequence of Escherichia coli 16S ribosomal RNA: complementarity to nonsense triplets and ribosome binding sites. Kozak, M. The scanning model for translation: an update. A., Weinstein, E. N., Naik, N. & Madani, A. ProGen2: exploring the boundaries of protein language models. Li, F.-Z., Amini, A. P., Yue, Y., Yang, K. K. & Lu, A. X. Feature reuse and scaling: understanding transfer learning with protein language models. Dalla-torre, H. et al. Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Stanke, M., Steinkamp, R., Waack, S. & Morgenstern, B. AUGUSTUS: a web server for gene finding in eukaryotes. de Almeida, B. P. et al. SegmentNT: annotating the genome at single-nucleotide resolution with DNA foundation models. Ji, Y., Zhou, Z., Liu, H. & Davuluri, R. V. DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome. Accurate classification of BRCA1 variants with saturation genome editing. Huang, H. et al. Functional evaluation and clinical classification of BRCA2 variants. Cunningham, H., Ewart, A., Smith, L. R., Huben, R. & Sharkey, L. Sparse autoencoders find highly interpretable features in language models. Towards monosemanticity: decomposing language models with dictionary learning. Scaling monosemanticity: extracting interpretable features from Claude 3 Sonnet. Bussmann, B., Leask, P. & Nanda, N. BatchTopK Sparse Autoencoders. Identification of mobile genetic elements with geNomad. Vorontsov, I. E. et al. HOCOMOCO in 2024: a rebuild of the curated collection of binding models for human and mouse transcription factors. A., Bailey, T. L. & Noble, W. S. Quantifying similarity between motifs. Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Sandoval-Velasco, M. et al. Three-dimensional genome architecture persists in a 52,000-year-old woolly mammoth skin sample. Meng, G., Li, Y., Yang, C. & Liu, S. MitoZ: a toolkit for animal mitochondrial genome assembly, annotation and visualization. Complete chemical synthesis, assembly, and cloning of a Mycoplasma genitalium genome. Karr, J. R. et al. A whole-cell computational model predicts phenotype from genotype. Hyatt, D. et al. Prodigal: prokaryotic gene recognition and translation initiation site identification. Fredens, J. et al. Total synthesis of Escherichia coli with a recoded genome. Li, Y. et al. Competition-level code generation with AlphaCode. Brown, B. et al. Large language monkeys: scaling inference compute with repeated sampling. Allis, C. D. & Jenuwein, T. The molecular hallmarks of epigenetic control. & Noble, W. S. Ledidi: Designing genomic edits that induce functional activity. & Seelig, G. Fast activation maximization for molecular sequence design. Controlling gene expression with deep generative design of regulatory DNA. Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo. DaSilva, L. F. et al. DNA-diffusion: leveraging generative models for controlling chromatin accessibility and gene expression via synthetic regulatory elements. Designing DNA with tunable regulatory activity using score-entropy discrete diffusion. Robust deep learning-based protein sequence design using ProteinMPNN. Language models generalize beyond natural proteins. Bloomfield, D. et al. AI and biosecurity: The need for governance. Pervasive ancestry bias in variant effect predictors. Schubach, M., Maass, T., Nazaretyan, L., Röner, S. & Kircher, M. CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Pampari, A. et al. ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants. Durrant, M. G. et al. Bridge RNAs direct programmable recombination of target and donor DNA. We thank J. Caputo, G. Filsinger, M. Fischbach, J. Kazaks, A. Kundaje, M. Macek Jr, S. Newins, S. Ovchinnikov, R. Park, B. Plosky, A. Woodrow and M. Zhang for helpful discussions and assistance with manuscript preparation; D. Traphagen, A. Gordon, H. Lewis, H. Estela, T. Rvachov, D. Ahn, S. Nah, C. Adams, X. Ren, S. Bak and D. Chang for behind-the-scenes help; and C. Dallago, K. Tretina, J. Israeli, N. Tadimeti, A. Stern, D. Voss, E. Calleja, C. Ye, R. Izzo, M. Bala, S. Alborghetti, V. Sirohi, V. Mehta, P. Bhattacharya, J. Sewall, A. Milesi, D. Toczydlowska, J. Mitchell, T. Moon, V. Balas, A. Aithal, P. Karbasi, C. Xuan, G. Guo, J. Wilber, M. Uhls, M. Harwood, N. Patel, O. Mosafi, R. Haukioja, S. Poulos and S. Bryson for additional support. We thank Ansa Biotechnologies for providing synthesized and assembled DNA for our mESC Morse code experiments. acknowledge funding support from the National Science Foundation Graduate Research Fellowship Program. acknowledges funding support from the Fannie and John Hertz Foundation. acknowledges funding support from the Knight-Hennessy Graduate Scholarship Fund. is supported in part by NHGRI/NIH grant RM1-HG009491 subaward, NIH grant DP5OD036167, by grant 2024-349901 from the Chan Zuckerberg Initiative, and funds from the Brotman–Baty Institute for Precision Medicine. acknowledges funding support from Arc Institute, Yosemite, Rainwater Foundation, Curci Foundation, Rose Hill Innovators Program, V. and N. Khosla, S. Altman, and anonymous gifts to the Hsu laboratory. acknowledges funding support from Arc Institute, the Gates Foundation, Stanford Institute for Human-Centered Artificial Intelligence (HAI) Hoffman-Yee Research Grants, Stanford Center for Digital Health, V. Gupta and R. Tonsing. Present address: OpenAI, San Francisco, CA, USA These authors contributed equally: Garyk Brixi, Matthew G. Durrant, Jerome Ku, Mohsen Naghipourfar, Michael Poli, Gwanggyu Sun These authors jointly supervised this work. Dave P. Burke, Hani Goodarzi, Patrick D. Hsu, Brian L. Hie Arc Institute, Palo Alto, CA, USA Garyk Brixi, Matthew G. Durrant, Jerome Ku, Mohsen Naghipourfar, Gwanggyu Sun, Daniel Chang, Alison Fanton, Gabriel A. Gonzalez, Samuel H. King, David B. Li, Aditi T. Merchant, Chiara Ricci-Tam, Jonathan C. Schmok, Daniel Guo, Michael H. Herschl, Rajesh Ilango, Reshma Mehta, Jeremy Sullivan, Joseph Tey, Patrick Collison, Dave P. Burke, Hani Goodarzi, Patrick D. Hsu & Brian L. Hie Evo 2 Core Team, Palo Alto, CA, USA Garyk Brixi, Matthew G. Durrant, Jerome Ku, Mohsen Naghipourfar, Michael Poli, Gwanggyu Sun, Greg Brockman, Daniel Chang, Alison Fanton, Gabriel A. Gonzalez, Samuel H. King, David B. Li, Aditi T. Merchant, Eric Nguyen, Chiara Ricci-Tam, David W. Romero, Jonathan C. Schmok, Ali Taghibakhshi, Anton Vorontsov, Brandon Yang, Sudarshan Pinglay, Dave P. Burke, Hani Goodarzi, Patrick D. Hsu & Brian L. Hie Garyk Brixi, Michael Poli, Daniel Chang, Samuel H. King, David B. Li, Aditi T. Merchant, Eric Nguyen, Euan A. Ashley, Stephen A. Baccus, Haoyu Dai, Steven Dillmann, Stefano Ermon, Daniel Guo, Madelena Y. Ng, Christopher Ré, Joseph Tey, Ben Viggiano, Tina Hernandez-Boussard & Brian L. Hie Mohsen Naghipourfar, Michael H. Herschl, Amy X. Lu, Mohammad R. K. Mofrad, Kevin Zhu & Patrick D. Hsu Liquid AI, San Francisco, CA, USA Independent Researcher, San Francisco, CA, USA David W. Romero, Ali Taghibakhshi, Anton Vorontsov, Ken Janik, John St. John, Greg Zynda, Anthony B. Costa, Ming-Yu Liu & Kimberly Powell Myra Deng, Liv Gorton, Nam Nguyen, Nicholas K. Wang, Michael T. Pearce, Elana Simon, Daniel Balsam, Eric Ho & Thomas McGrath Columbia University, New York, NY, USA Johns Hopkins University, Baltimore, MD, USA Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar G. Brixi conducted training data composition and loss weighting experiments. performed zero-shot human mRNA decay rate prediction. performed zero-shot human variant effect prediction. conducted the generative epigenomics design runs and scaling analyses. conducted the mESC Morse code experimental validation. designed and implemented the Evo Designer tool. conducted the safety, security and ethics investigation. Correspondence to Patrick D. Hsu or Brian L. Hie. acknowledges outside interest in Stylus Medicine, and is currently an employee of Anthropic PBC. acknowledges outside interest as a co-founder of Radical Numerics. J.P. acknowledges outside interest as an artificial intelligence policy consultant for the Chan Zuckerberg Initiative. acknowledges outside interest in Factory and Google Ventures. acknowledges outside interest as a Google Advisor. acknowledges outside interest as a co-founder of Exai Bio, Vevo Therapeutics and Therna Therapeutics, serves on the board of directors at Exai Bio, and is a scientific advisory board member for Verge Genomics and Deep Forest Biosciences. acknowledges outside interest in Stylus Medicine, Terrain Biosciences, and Monet AI as a co-founder, serves on the board of directors at Stylus Medicine and the scientific advisory board at Amgen, and is a venture partner at Thrive Capital. acknowledges outside interest in Arpelos Biosciences and Genyro, Inc. as a scientific co-founder. The other authors declare no competing interests. Nature thanks Teresa Przytycka and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. (a) Data composition of OpenGenome2; total eukaryotic genomes per kingdom (left), total base pairs per training data subset (middle), and detailed breakdown of other/augmented training data subset (right). (b) Core input-dependent convolution operators in StripedHyena 2, with a diagram showing their composition in the architecture. (c) Scaling ablations on OpenGenome2, showing the loss convergence of multi-hybrids compared to previous generation hybrids and Transformers. (e) A null distribution of needle-in-a-haystack scores by randomly shuffling the needle sequence across a sweep of haystack lengths and needle positions, computing the resulting retrieval score based on a categorical Jacobian analysis (Methods). The distribution of N = 1040 scores is plotted here. To decrease dual use risks, safety filtering was performed on the training data to remove viral sequences that can infect eukaryotic hosts. Evo 2 is less performant on eukaryotic viruses, as intended. (a) Perplexity scores for viral sequences from the USDA Select Agents and Toxins List consistently demonstrate elevated perplexity values compared to non-pathogenic viruses and prokaryotic viruses. Blue violin plots show the distribution of scores, with individual data points overlaid representing 512-bp chunks sampled uniformly at random across viral genomes. (b) Correlation of language model likelihood with experimental deep mutational scanning (DMS) fitness measurements for human viral proteins. Gray bars represent mean correlation coefficients, with individual data points corresponding to DMS datasets from ProteinGym. Results indicate poor predictive capability on viral protein mutational effects for Evo 2 and Evo 1 models. (c) Comparative analysis of protein sequence generation success rates across different model conditions. Bar heights represent percentage amino-acid sequence recovery in the response sequences when prompted with a portion of a viral protein, with error bars showing standard deviation across multiple responses to the same prompt. Models were tested with various prompting proteins (shown on the horizontal axis) with different Evo 2 models (indicated by color). Random sequence generations are included as a control condition. (d) Analysis of ancestry bias for Evo 2 as a variant effect predictor compared to baselines, with protein mutations converted to DNA codons. Baseline performance data is taken from Pathak et al. Most variant effect predictors have ancestry bias, and score non-European ancestry variants as more pathogenic. Evo 2 has similar ancestry bias as other population-free methods, examined by taking both the ratio (heatmap) and mean difference (bar plot) of min-max scaled scores of each population subgroup to the European subgroup. (c) Evo 2 40B predicts lower likelihoods for deletions in miRNA and snoRNA loci compared to Evo 2 7B. The same sequences were analyzed with both models. (d) The translational codon ramp pattern detected across all coding sequences across four species, focusing on the first and last 100 codons for each coding sequence. The local mean tRNA-adaptation index (tAI) was calculated using pre-computed tAI values for each species, and then z-score normalized. (e) The average change in log-likelihood across hundreds of genes and codon positions in each species' genome. Blue lines indicate synonymous codon mutations with a higher tAI than the reference sequence, while red lines indicate synonymous codon mutations with a lower tAI than the reference sequence. (f) Evo 2 predicts stop codons dependent on the sequence context and stop codons present in the genome sequence, responding to artificially altered stop codon code by predicting the mutations as high effect. Showing median z-score standardized median Δlikelihood values for two ciliate genomes across 6 sequence context lengths. (g) Length-adjusted Evo 2 likelihoods of human mRNA sequences showed a negative correlation with their experimentally measured decay rates. Borzoi was included as a supervised sequence-to-expression model by selecting and averaging RNA expression prediction tracks. (h) Zero-shot prokaryotic gene essentiality prediction including the base pretrained models and the final checkpoints extended to 1-million token context for both the 7B and 40B parameter Evo 2 models; compare to Fig. (i) DepMap human gene essentiality classification performance measured by AUROC and AUPRC metrics comparing conservation baselines and language models. (a) ClinVar variants stratified by PhyloP score and variant type (coding or noncoding, SNV or non-SNV). (b) ClinVar variants stratified by proximity to splice sites, where “near” is defined as within 5 bp or 3 bp to donor or acceptor sites, respectively, and “far” is defined as otherwise. Because there were only 22 coding non-SNVs that are near splice sites and all of these variants have the same label (“pathogenic”), we excluded this condition in our evaluations due to the limited sample size and the inapplicability of binary classification metrics like AUROC and AUPRC. (c) Zero-shot noncoding ClinVar evaluations after filtering out all variants with either a SpliceAI score ≥ 0.1 or a Pangolin score ≥ 0.1 (i.e., keeping variants with low SpliceAI and low Pangolin scores). (a) Zero-shot prediction of BRCA1 variant pathogenicity for coding and noncoding SNVs evaluated in aggregate, showing AUROC and AUPRC scores across models. (b) Zero-shot prediction of BRCA1 variant pathogenicity for noncoding SNVs separated according to “intronic” and “splice site” variants based on whether the variant site is located more than 8 bp away from the intron-exon boundary26. (c) Zero-shot prediction of BRCA2 variant pathogenicity for coding (left), noncoding (center), and aggregated (right) SNVs, showing AUROC and AUPRC scores across models. (d) Comparison of AUROCs for the supervised classification of BRCA1 SNVs using embeddings from Evo 2 40B block 20, for pooling functions that use different window sizes around the variant site to average embeddings. AUROCs are averaged from 5-fold cross-validation on the training set, and are from the best-performing ridge regression regularization parameter for each pooling function. AUROCs are averaged from 5-fold cross-validation on the training set, and are from the best combination of pooling function and ridge regression hyperparameter α for each layer. (f) DART-Eval results for zero-shot regulatory DNA tasks. Task 1 evaluates models on their ability to distinguish candidate cis-regulatory elements (cCREs) from shuffled sequences, Task 2 tests their ability to identify transcription factor (TF) motifs by distinguishing true TF binding sites from control sequences, and Task 5 predicts variant effects for chromatin-accessibility QTLs (caQTLs, African) and dynamic sequence QTLs (dsQTLs, Yoruban). Sequence likelihoods were computed under each model and used to measure classification accuracy with mean AUROC. For Task 5, while we did not see strong signal across all models when using the zero-shot log-likelihoods, Evo 2 embeddings were predictive of noncoding variant effects. (a) Composition of the prokaryotic sequences randomly subsampled for SAE training. (b) Composition of the eukaryotic sequences randomly subsampled for SAE training. (c) Feature density histograms for SAEs demonstrate that Layer 26 converged with fewer low-frequency features while the distribution peaks around a firing frequency of 1e-3 representing sparse yet generalizing features. (d) Activation density for all layer 26 SAE features over the E. coli K12 MG1655 genome (left) or for a length-matched segment of human chromosome 17 (right). (e) Mean non-zero activation for all layer 26 SAE features over the E. coli K12 MG1655 genome (left) or for a length-matched segment of human chromosome 17 (right). (f) UMAP embedding of layer 26 SAE feature weights colored by activation density difference between eukaryote and prokaryotic sequence for each feature, with features presented in Fig. (a) Diagram depicting the contrastive feature search strategy used to identify and quantify selected features. (b) Mean activations of the phage feature across phage regions annotated by geNomad and matched-length bacterial non-phage sequences across 100 randomly selected GTDB genomes. AUROC was computed over mean activations for each region. (c) (i) Scrambling spacer sequences does not ablate the phage feature activation pattern on spacer sequences. (ii) Using a constant scrambled CRISPR direct repeat sequence ablates phage feature activation for the first two spacers. (iii) Using different scrambled sequences instead of CRISPR direct repeats ablates the phage feature activation pattern. (iv) Natural sequence activation pattern, as in Fig. (d) Additional examples of sequences not annotated as phage sequences by geNomad which the phage feature activates on. (e) Activations of additional features associated with open reading frames (ORFs), plus strand or minus strand ORFs ((+) ORF and (–) ORF), and intergenic loci in a 100 kb region in E. coli K12 MG1655. (f) Mean activations for prokaryotic organizational features on different annotation types across the E. coli K12 MG1655 genome. AUROC was computed over mean activations for each region. Consecutive intergenic positions were merged into single regions. The ORF associated features were evaluated for their abilities to predict the presence of either plus strand or minus strand ORFs. (g) Mean activations for protein secondary structure features on different secondary structure types across ORFs in the E. coli K12 MG1655 genome. AUROC was computed over mean activations of positions annotated as each secondary structure type per protein. (h) Ablation experiments on SAE features with high F1 scores for biological elements demonstrate increases in average CE and ratio of ablated to original CE. These results suggest that learned features can be causally relevant downstream. (b) F1, precision, and recall scores across mutation types for features shown in (a). (c) Activations for SAE features associated with exons, introns, and their boundaries in the human genome, shown for a 6000 bp region in chromosome 1. (d) F1, precision, and recall scores for each SAE feature shown in (c) to its corresponding genomic element. These scores were calculated at the level of individual bases across 1,000 genes randomly selected from the human genome. (e) Mean activations for each SAE feature shown in (c) on different annotation types across the human genome, and the corresponding AUROC values of the features to their corresponding annotation type. AUROCs were calculated at the level of individual bases across 1,000 genes randomly selected from the human genome. (f) Recall-FDR curve of Evo 2 SAE features compared with HOMER on human H1-2CORE motifs and promoter-enriched motifs. (a) Amino acid sequence recovery for different genes across Evo 2 models when prompted with genomic context of the respective gene. (c) Predicted aligned error (PAE) of the Evo 2-generated mitochondrial complexes from AlphaFold 3. (e) Frequency of each tRNA anticodon across Evo 2-generated mitochondria as annotated by MitoZ. (f) ESMFold predicted local distance difference test (pLDDT) distributions of natural and Evo 2-generated M. genitalium genes called by Prodigal. (g) Distribution of TM scores from Evo 2-generated M. genitalium genes against UniRef50 AlphaFold DB. (h) Predicted AlphaFold 3 structures, TM scores, and sequence identity comparing genes from Evo 2-generated M. genitalium with natural proteins. (i) Distribution of genes, introns, tRNAs, and promoters on an Evo 2-generated S. cerevisiae sequence compared with the natural S. cerevisiae chromosome III (gray line). (j) Distribution of gene lengths and pLDDTs of Evo 2-generated and natural S. cerevisiae genes, annotated by GeneMark-ES. (l) Distribution of different secondary structures in Evo 2-generated against S. cerevisiae chromosome III wildtype genes. (m) Protein structure of genes from Evo 2-generated S. cerevisiae sequence compared with the natural structure and sequence. (n) Tetranucleotide usage deviation (TUD) comparison between natural S. cerevisiae, S. pombe, and Evo 2-generated sequences for whole sequence, CDS, and promoters (upstream of gene start). (a) Simpler patterns were designed in which each peak had uniform width. Good experimental validation of the predicted peaks was observed, especially when the peak width is longer and when all Enformer and Borzoi models agree in their predictions. Accuracy drops to an AUROC of 0.89 with short peaks of 384 bp in the “short wave” (right), especially when the Enformer and Borzoi predictions disagree. 6f-h. (b) Top: Dinucleotide frequencies in randomly proposed sequences after beam search filtering with Enformer and Borzoi still show significant deviation from the baseline mm39 frequency. Middle: Dinucleotide frequencies, generated by a bigram proposal distribution based on the mm39 reference genome, match the baseline mm39 frequency by construction. Bottom: Sequences generated by Evo 2 when prompted with a portion of the mouse genome have natural dinucleotide frequencies, despite this never being directly enforced during inference-time sampling. The bigram proposal appears to have better initial scaling at lower beam search widths, but this plateaus as the beam search width increases. The Evo 2 proposal is the first to reach an AUROC > 0.95 threshold, above which designs tend to have qualitatively clear design success. Individual design runs are plotted as circles and the averages across design runs for each beam search width and each generative model are plotted as bold Xs. (d) Quantified agreement between the single Enformer and four aligned Borzoi prediction tracks using the Intraclass Correlation Coefficient (ICC) from a two-way mixed-effects model (ICC(2, k); Methods). A value of 1 indicates perfect agreement across all five tracks. Despite never directly optimizing for ensemble agreement, sequences generated by Evo 2 have consistently high ICC values of ~0.95 even at the lowest beam search widths. Both uniform and bigram proposals have much lower ensemble agreement, though the ICC tends to improve as the increasing beam search width filters out poor designs. Individual design runs are plotted as circles and the averages across design runs for each beam search width and each generative model are plotted as bold Xs. (e) The best scoring designs combining both AUROC and ICC for the “ARC” Morse code pattern across different generative proposals. Despite high AUROC and ICC values, the uniform and bigram proposals are qualitatively worse in terms of desired pattern agreement and Enformer/Borzoi ensemble agreement than the sequence generated by an Evo 2 proposal. We hypothesize that ensemble disagreement corresponds to greater uncertainty in the Enformer/Borzoi accessibility predictions and is consistent with adversarial inputs. (f) Genomic statistics are plotted for three experimentally validated Morse code designs (from left to right, “ARC”, “EVO2”, and “LO”). Then, these statistics are plotted combining all three designs into a single plot (“Evo 2”) followed by plotting statistics for the same three Morse code designs but generated by a uniform or a bigram proposal. Statistics are separated by regions with a designed peak and without a designed peak. The rightmost column plots the statistics for regions of the Mus musculus genome in experimentally determined DNASE-seq peak regions and in regions without peaks. Additional details on these statistics are provided in Methods. Individual plots are shaded gray if there is a significant difference (P < 0.05) between “peak” and “no peak” conditions under a two-sided Welch's t-test. (h) Genomic tracks visualizing TF motif-related statistics along the “LO” design, where the motifs have been restricted to TFs expressed in mESCs. (i) Distribution of genes plotted by log(1 + TPM) expression values on the x-axis, used to determine a gene expression cutoff for mESCs (log(1 + TPM) > 1). This expression cutoff was used to determine whether TF motifs found in the “LO” design were significantly enriched for TFs expressed in mESCs. (a) Genomic tracks visualizing the sequence statistics plotted in (Extended Data Fig. (b) A histogram distribution of 24 designed sequences [see panels (f) and (g)] where high chromatin accessibility in K562 and low accessibility in HEK293T was specified, plotted according to the K562/HEK293T fold change in mean coverage across all positions in the peak. A fold change of 1 indicates the same mean coverage in both cell types. Four sequences with >2-fold change and one sequence with >3-fold change in mean coverage were observed (where K562 has higher mean coverage), representing a 4-17% success rate. (c) Summary results of designs in human cells that attempt to either maximize accessibility (“K562 on” and “HEK293T on”) or minimize accessibility (“K562 off” and “HEK293T off”) across the full designed sequence (see panels (d) and (e), respectively). All designs that maximize accessibility have mean coverage values > 2.7 and all designs that minimize accessibility have mean coverage values < 1.3. (d–h) Plots showing ATAC-seq coverage of 1-4 kb designs in which the same sequence was integrated into both HEK293T and K562. The design patterns were grouped into five main categories. (d) The first consists of designs that aim to maximize accessibility in both cell lines across an entire 1-kb region. (e) The second consists of designs that aim to minimize accessibility in both cell lines across an entire 1-kb region. (h) The fifth consists of miscellaneous 4-kb designs that specify either two or four peaks in either K562 alone or in both cell lines. For all human cell line experiments, ATAC-seq coverage values are the average across two transfection/nucleofection replicate populations of cells. (i,j) Distribution of genes plotted by log(1 + TPM) expression values on the x-axis, used to determine a gene expression cutoff for K562 (log(1 + TPM) > 1.5) (i) and for HEK293T (log(1 + TPM) > 2.15) (j). This expression cutoff was used to determine whether TF motifs found in the B7, B10, B11, and B12 designs were significantly enriched for TFs expressed in K562 (i) or in HEK293T cells (j). This file contains Methods, supplementary text and appendix. Scores and metadata for human variant effect prediction analysis, related to Fig. Mechanistic interpretability transcription factor motif report. A comprehensive list of promoter-enriched motifs from the HOCOMOCO v.12 CORE database and associated SAE feature hits, related to Fig. DNA sequences for experimental chromatin accessibility designs. Experimentally tested DNA sequences in the Morse code mESC and HEK293T/K562 chromatin accessibility design tasks, related to Fig. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. Brixi, G., Durrant, M.G., Ku, J. et al. Genome modelling and design across all domains of life with Evo 2. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing: Translational Research newsletter — top stories in biotechnology, drug discovery and pharma.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. 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. Sea-level rise (SLR) poses a high risk to vast coastal lowlands around the world, including low-elevated and populous river deltas and coastal plains. According to the Intergovernmental Panel on Climate Change (IPCC) assessment report 6 (AR6) (shared socioeconomic pathway (SSP)1-1.9 to SSP5-8.5)4, global mean sea level (MSL) is projected to rise between 0.28 m and 1.01 m by 2100 compared with that in 1995–2014. Including deep uncertainties in the polar ice-sheet dynamics, these projections could increase by several metres4. This rise is amplified by negative vertical land motion, that is, land subsidence, a natural but increasingly human-accelerated phenomenon in coastal lowlands5,6, which drives higher rates of relative sea-level rise (RSLR)7,8. Consequently, the impacts of RSLR and coastal (and compound) flooding are closely related to coastal elevation relative to sea level. Assessing coastal exposure and vulnerability consequently requires the use of elevation information, commonly provided by digital elevation models (DEMs). Constituting the fundamental base of any such impact and exposure assessment, the quality (mainly vertical accuracy and spatial resolution) of the DEMs is fundamental to the accuracy and reliability of the derivatives and widely addressed in scientific literature (see, for example, ref. Although high-quality elevation information, for example, acquired through airborne lidar, is available in several regions of the Global North (see, for example, ref. 10), the best available elevation data for the vast majority of the coastal areas worldwide is satellite-based. The global availability of satellite-based elevation data enabled unprecedented (global) SLR and coastal hazard studies (see, for example, ref. However, spaceborne DEMs can have vertical errors up to several metres, contain sensing or interpolation artefacts, or are outdated, thereby affecting the quality of coastal hazard assessments, especially in flat, subsiding coastal plains and densely populated river deltas10,12,13. Apart from considering vertical uncertainty of elevation data, using DEMs in coastal hazard assessments requires correctly combining coastal elevation with local sea-level height and the proper conversion to a common vertical reference frame2,3. Through a systematic review evaluating recent SLR impacts and coastal hazard assessment studies, we found that these crucial steps were often not considered or performed incorrectly. Rather than considering actual, local sea-level height, coastal sea level is most often assumed to equal (an often outdated) global geoid (or in some instances even ellipsoid), to which open-access global DEMs are typically referenced when provided. A geoid is an equipotential surface model that approximates MSL based on gravity and the rotation of Earth. As geoid quality depends on gravity observations, uncertainties in global geoid models can range up to several metres in regions that suffer from gravitational data paucity (see, for instance, refs. 14,15), predominantly located in the Global South. Moreover, actual sea-surface height is not just determined by the gravity and rotation of Earth, but also by, for example, ocean currents and large-scale circulation, winds, tides, seawater temperature and salinity. As a result, time-average sea-surface height can deviate strongly (up to several metres) from a geoid, and its difference is the so-called mean dynamic topography (MDT). The widespread omission to properly reference coastal elevation to measured local sea level generally leads to an underrepresentation of actual coastal sea-level height in (global) SLR and coastal hazard assessments and can introduce errors with magnitudes as large as a century of projected SLR for affected regions in the world2,3. By means of meta-analyses, we quantify the implications of the most-frequently encountered omissions or errors (that is, incorrect or absent vertical datum conversion) on estimates of exposed people and coastal area at global and regional scales. To evaluate the broader implications of our findings, we quantified the magnitude of potential underestimation or misjudgement in assessment studies included in the most recent IPCC AR6 reports. To facilitate proper, future (re)assessments of coastal hazard impacts, we converted several state-of-the-art global DEMs to coastal sea-level height and provided them ready for use (see Data availability and Code availability sections). We conclude with concrete recommendations, such as data documentation guidelines and peer-review checklists, to ensure correct vertical datum alignment in future publications and improve community research standards. Herewith, we aim to eliminate future propagation of erroneous methodologies that caused this community-wide blind spot and resulted in widespread underestimations of coastal SLR and hazard impact assessments. We evaluated 385 peer-reviewed, scientific publications (systematically selected through a PRISMA-guided literature search, published between 2009 and 2025, with >53% in the past 5 years) on SLR and/or RSLR and coastal flood exposure, vulnerability and risk, presenting both global and regional hazard assessments. We scrutinized each for the correctness of DEM usage, vertical datum conversion and proper integration of sea-level height and coastal elevation and documented the ubiquity of errors and omissions (see the Methods for details; see also Extended Data Fig. The evaluated coastal hazard assessment literature investigates coastal settings all over the world and across spatial scales. The impact assessments focused on SLR and/or RSLR (14%), storm surge (8%), tsunami (8%), coastal exposure, vulnerability and/or risk more in general (41%), or included combinations of different single-type coastal hazards, exposure, vulnerability and/or risk assessments and technical aspects and methodological advancements (29%). The evaluated literature includes studies from data-sparse coastal lowlands in Africa (14%) and Asia (58%), which are severely affected by RSLR and coastal flooding16, poor data availability and accessibility as well as data inaccuracy due to the comparably poor performance of global Earth gravity models (see, for example, ref. Moreover, the evaluated literature includes high-impact, global hazard assessments (10%), which dictate the contemporary scientific understanding of global coastal hazard impacts. In the bulk of the evaluated publications (73%), documentation of used sea-level height, coastal elevation and vertical datums is either incomplete (13%) or entirely missing (60%) (Fig. Although 27% of the literature correctly documented the vertical datum(s) used, only 1% correctly described and aligned coastal sea-level height information to the land elevation data. For the vast majority of evaluated assessments, sea level and coastal elevation alignment and datum conversion description were absent, and conversion was probably omitted (90.6%). In the remaining studies, the description was often incomplete (making the study irreproducible) or described an incorrectly performed conversion (8.6%). Correct alignment of elevation data from global DEMs to sea level requires a vertical datum conversion in combination with additional sea-level data, either a local sea-level datum (for example, local tide gauge and/or national sea-level-aligned datum) or global ocean surface topography (for example, satellite altimetry and/or buoys combined product such as MDT). We expect that authors aware of the necessity and with the expertise to successfully perform a vertical datum conversion using additional sea-level data do properly document these crucial (and often time-consuming) methodological steps and additional datasets. Therefore, we presume that the absence of any documentation of sea-level information (25%) and/or methodological conversion steps (65%) means that sea-level height data was not included or datum conversion to a sea-level reference was omitted. Repeated evaluation across multiple studies confirmed the validity of this presumption (see, for example, ref. Complete vertical datum documentation means all necessary vertical datum information is provided in the study itself or in the cited references of the data used. Vertical datum conversion is correct if all necessary datum conversion steps required to properly align all data to a common vertical reference are properly described and applied. The correct implementation of a sea-level reference involves the use of sea-level information (for example, MDT or tide gauge data), correctly aligned with all other data used in the assessment in a common vertical reference. A sea-level reference is considered up to date if the latest available sea-level data were used. In 73% of the studies, vertical datum documentation was incomplete or completely absent. In nearly all evaluated assessments, sea-level data and their proper alignment to coastal elevation data were either not documented and datum conversion likely omitted (90.6%; see underlying presumption in main text), or (seemingly) incorrectly performed (8.6%). Only 0.3% of the evaluated studies completely documented, converted and properly adjusted coastal elevation data with sea-level information (shown with green colour). The Sankey diagram for the results of this study was created using SankeyMATIC (https://sankeymatic.com/). The most prominent issue in the evaluated literature (demonstrably present in 25% and presumingly present in 63% more; Figs. 3) was the neglection of datum conversion from geoid (in some cases even ellipsoid) to a sea-level reference, thereby implicitly assuming a geoid height of 0 m to match local sea-level height. Of all literature containing this issue, we encountered only two papers that reflected on the potential discrepancy between geoid and actual sea level8,18. The second most frequent issue (9%) was incomplete, and thereby erroneous, datum conversion and inadequate alignment of vertical datums of datasets involved (for example, land elevation, bathymetry, sea level and heights of coastal infrastructure). This group of literature also contains few studies that pioneered the use of MDT data to create a sea-level reference19,20,21,22, indicating first signs of community awareness on the necessity for correct land–sea-level alignment. Although these studies arguably improved on the bulk of geoid-based studies neglecting sea-level alignment, they suffer from conversion documentation shortcomings and demonstrated20,21,23 incorrect datum conversion (Extended Data Fig. 13) out of the 385 evaluated studies (0.3%) had complete vertical datum and conversion documentation and contained no conversion and alignment errors. Apart from evaluating individual studies, our literature evaluation also revealed widespread and persistent propagation of erroneous workflows, for example, omitting sea-level datum conversion24,25 or introducing methodological datum conversion errors19,22,26,27, to consequently affect follow-up studies that apply the same data and processing approach20,21,23,28 (Supplementary Data 1). Similarly, assessments using modelling frameworks that include elevation data in the coastal zone also suffer from the investigated issues such as incomplete or absent vertical datum documentation (see ref. 30) and/or lack of a sea-level datum (ref. 31, for instance) (Supplementary Data 1). We performed several meta-analyses on global and regional scales to quantify the magnitudes of coastal sea-level height misrepresentation stemming from the most-frequently encountered processing omission or errors for the most widely used geoids (Methods and Extended Data Fig. Impact assessments that neglect datum conversion and consequently assume the geoid surface to represent MSL result in a global average underrepresentation of coastal sea-level height as represented by MDT of respectively 0.27 m (median 0.19 m, standard deviation (s.d.) In regions in which the respective geoids perform more poorly (for example, higher inaccuracy caused by data paucity), the discrepancies can go up to several metres (5.5–7.6 m for EGM96 and 2.8–3.4 m for EGM 2008) (Fig. a–c, The still widely used Earth Gravitational Model 1996 (EGM96) shows large deviations from measured coastal sea level, here indicated by the latest available MDT44, especially profound at the regional scale. d–f, The more recent EGM2008 geoid model shows overall improvement over EGM96 and provides globally a slightly better approximation of local coastal sea level. In data-rich countries in the Global North, the global geoids represent coastal sea level relatively well (for example, Eastern United States, Northern Europe and Western Europe), whereas in the more data-sparse Global South, regions such as Latin America, East Africa and the Indo-Pacific, with Southeast Asia and Oceania as global hotspots, the geoids substantially underrepresent actual sea-surface height, ranging from several decimetres up to several metres locally. The vast majority of the evaluated literature assumed the geoid surface (0 m) to represent contemporary local MSL, thereby introducing the above discrepancy as error into their respective coastal hazard and SLR impact assessments. For visualization purposes, the spatial scale of the data shown was resampled to 1° using bilinear resampling, whereas all statistics are given at 90 m spatial resolution. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. Although the global statistics average out the larger regional and sub-regional discrepancies, these seem to be particularly large for several key regions, most located in the Global South. Largest discrepancies are observed in Southeast Asia (hosting large, populous and low-lying river deltas; Extended Data Figs. 6 and 7) and the Pacific Region (often lowly-elevated atolls), on average amounting to an underrepresentation of coastal sea level of more than 1 m, as was previously already highlighted in local studies on the Mekong and Ayeyarwady deltas2,3. Other areas with large discrepancies are located in Latin America, the west coast of North America, the Caribbean, Africa, the Middle East and the larger Indo-Pacific. Although, on average, the geoid models underrepresent sea-surface height at global and regional scales, locally the discrepancies can also range in the opposite direction (for example, northern Mediterranean coast, Antarctica and some islands in the Atlantic and the Pacific (EGM2008 only)) (Fig. 2), consequently resulting in an overrepresentation of sea-surface height. The lowest discrepancies between MSL and geoid are prevalent in Eastern North America, as well as Northern and Western Europe (Supplementary Data 2), reflecting the stronger performance of geoid models to approach sea-surface height in data-rich regions in the Global North. Continuous advances in global geoid modelling are reflected by new global geoids (for example, GOCO2025s), but these also contain comparably large discrepancies to MDT-determined MSL (Supplementary Fig. Therefore, omitting to include sea-level information and to properly convert from a geoid to a sea-level datum, especially in data-sparse and remote regions in which geoid and sea level do not align well, immediately transfers these discrepancies (Fig. 3) as errors into hazard and SLR impact assessments24,32,33,34 or into the delineation of the low-elevation coastal zone (LECZ)25. Meta-analysis of the most common shortcoming in existing coastal hazard assessments (that is, assuming the geoid to represent local sea-level height), using modern, globally available DEMs (CoastalDEM v.2.1 (ref. The difference given is the coastal sea-level height discrepancy for each DEM between assumed coastal sea level (that is, assuming 0 m elevation of the respective geoid-referenced DEM to equal contemporary MSL) and our assessment in which we correctly align coastal elevation to measured mean sea level using the latest MDT data44. On average, global coastal sea level is about 0.3 m higher than commonly assumed in coastal hazard assessments, whereas in Southeast Asia, the discrepancies are the largest, with measured sea level exceeding previous assumed levels by, on average, 0.9–1.1 m. All statistics are computed at 90 m spatial resolution. Colour outlines and grey-shaded areas indicate the regions and subregions (Supplementary Fig. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. Studies that do include sea-level information but subsequently omit to convert elevation (DEMs) and sea-level data (for example, altimetry-based MDT or mean sea surface (MSS)) to a common datum before combining them, introduce datum offset errors (that is, the respective height difference between the different datums, present in 9% of the studies; see, for instance, refs. The few studies that did incorporate MDT data but suffer20,21 from a specific geoid conversion error (Extended Data Fig. 3) statistically nearly equal our global average coastal sea-level height representation (underrepresentation of 0.02 m), but suffer regionally from discrepancies up to several metres, ranging from −4.1 m to +4.3 m (s.d. A third source of error stems from using an outdated sea-level reference. Tide gauges provide locally specific information, but their ability to adequately reflect recent sea level can be severely restricted when the time series is too short, incomplete or outdated. As tidal datums may have been established decades to centuries ago, for example, the North American Vertical Datum of 1988 (NAVD 88) in 1985 (ref. 36), or the Amsterdams Peil (AP) in the 17th century (later the Normaal AP in 1875)37, their usage is valid as long as sea-level changes since establishment are correctly taken into account. The incorrect use of sea-level height information, for example, from tide gauges or satellite altimetry21,23, without considering past relative sea-level changes since datum establishment and start of the projection period, contradicts the implied actuality of the hazard assessment conducted. Given the high rates of RSLR due to coastal subsidence5,8, which exceed global SLR in many parts of the world, particularly in densely populated Asian deltas and coastal lowlands6,8, assessments of contemporary coastal exposure must account for relative sea-level changes since tide gauge datum establishment or, for MDT/MSS, the average observation period38. Similarly, as coastal elevation itself is also not static, we highlight the urgency of using the most recent and accurate elevation data, ideally corrected for spatial heterogeneous vertical land motion39 and consequent elevation change since data acquisition38. However, a thorough evaluation and inclusion of relative sea-level or elevation change since datum establishment or elevation acquisition practice is far from common practice, as only 22% of the evaluated studies use actual DEMs (that is, the latest available elevation data by the time of initial submission; Supplementary Table 2), whereas only 9% include DEM accuracy assessments (Extended Data Fig. We performed several meta-analyses to evaluate the impacts of the most-frequently witnessed issues in existing coastal hazard assessments and quantified the potential exposure misjudgement present in global relative SLR impact assessments. We used four of the best-performing global DEMs to date as supplied in their data repositories40,41,42,43 and applied a RSLR scenario of 1 m, while omitting proper inclusion of a sea-level datum to mimic the most commonly applied methodology (>90% of all assessments), that is, assuming the global geoid surface to represent contemporary sea-level height (Fig. We compared the results with the same DEMs correctly aligned to measured local MSL by properly applying the most recent MDT data44, showing the discrepancy between the commonly assumed sea-level height (that is, geoid) and measured local mean sea-level height. Our meta-analyses show that worldwide estimates of land area and population below MSL after a 1 m relative SLR are substantially underestimated when MDT data are not included. Proper sea-level referencing using MDT increases the exposed area from 294,500–431,100 km2 to 460,100–670,000 km2 (that is, by 31–37%), and population from 34.0–49.2 million to 77.0–132.2 million people (that is, by 48–68%) across different geoid-based DEMs and various population datasets (Fig. 4 and Extended Data Fig. This adds, respectively, 287,400–470,700 km2 and 55.1–101.6 million people to the 172,700–234,000 km2 area and 21.9–34.5 million people already below MSL so far (with respect to 114,200–158,600 km2 and 10.5–15.4 million people estimated so far below MSL without MDT referencing). In the most affected region, Southeast Asia, the estimated area and population below MSL following 1 m RSLR increased up to 94% and 96% following proper MDT referencing, increasing the exposure numbers to 78,000–99,700 km2 and 24.2–46.9 million people for this region only. Global assessments of the LECZ (first 10 m of coastal elevation) based on geoid elevation underestimate area up to 4% and people up to 8% globally and proper MSL referencing increases the global LECZ to 3.0–4.1 million km2 being inhabited by 0.82–1.07 billion people (Supplementary Data 3). Meta-analysis showing the impact of the most-frequently observed errors in existing coastal hazard assessments, that is, omitting or improper alignment of measured coastal sea level to land elevation, using modern, globally available DEMs (CoastalDEM v.2.1 (ref. Properly referencing coastal elevation from geoid to local MSL increases global population exposure estimates from 44.0–108.4 million to 102.8–132.2 million people below sea level following 1 m RSLR (that is, by 12–67%). Population estimates are based on WorldPop 2020 (ref. 52) and do not account for future population change. The largest population exposure increase, that is, by 13–96%, is observed in Southeast Asia, where 32.1–46.9 million people (of which 5.0–9.3 million currently already reside below sea level) fall below sea level following 1 m RSLR. All statistics are computed for the respective spatial resolution of the DEMs. Colour outlines and grey-shaded areas indicate the regions and subregions used (Supplementary Fig. All data related to this figure are included in Supplementary Data 3. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. 51 (Open Government Licence v.3.0) and icons modified from Font Awesome Free v.7.1.0 (CC-BY 4.0; https://creativecommons.org/licenses/by/4.0/). When we corrected the encountered geoid conversion error present in existing studies applying MDT data (that is, omitting to correct the offset between the EGM96 and DIR-R4 geoids) (Extended Data Fig. 9), our meta-analysis showed that global coastal area below MSL after a 1 m RSLR decreased by 2% (11,400 km2), from 527,000 km2 to 515,700 km2 (of which, respectively, 133,100 km2 and 117,100 km2 are currently already below MSL). However, the estimated global population below MSL following a 1 m RSLR considerably increased by 10–12% (8.0–14.1 million) from 68.6–108.4 million to 76.6–122.4 million people (of which, respectively, 18.2–27.9 million and 16.5–25.2 million are currently already below MSL). This shows that this specific geoid conversion error (Extended Data Fig. Consequently, assessments that do use MDT as sea-level reference but suffer from this error (demonstrated for ref. 23 and presumably more, for example, refs. 19,45) underestimate population exposure in affected regions (Fig. At the local scale, the encountered vertical reference issues particularly affect hazard assessments in large, often populous, low-elevated coastal-deltaic areas (Extended Data Figs. To evaluate the relative impact of vertical reference issues with respect to other (DEM-dependent) uncertainties, we performed a detailed meta-analysis for the Vietnamese Mekong Delta (VMD). The VMD exemplifies many densely populated coastal landscapes exposed to RSLR worldwide, being one of the largest, flattest and low-lying deltas worldwide2. Not including a proper sea-level reference (in our case using MDT data), added 10–60% additional (DEM-specific) error to the relative elevation assessment on top of existing data-inherent inaccuracy, sea-level and elevation change effects38 (Supplementary Table 3). Proper inclusion of MDT data increased the area and population in the VMD exposed to a 1 m RSLR from 1,400–6,000 km2 to 18,400–24,800 km2 (that is, by 72–95%), and from 312,900–2.4 million to 5.4–10.0 million people (that is, by 74–96%), confirming the findings from earlier research on the elevation of Mekong Delta2 (Supplementary Table 4). The specific geoid conversion error encountered for incorrect MDT implementation (Extended Data Fig. 9) is particularly prominent in coastal lowlands of Southeast Asia and when corrected for the VMD, exposed area and population following a 1 m RSLR increases by, respectively, 18% (from 20,900 km2 to 25,500 km2) and 19–23% (from 5.5–7.9 million people to 6.8–10.2 million people). These numbers highlight the large spatial variability and locality-specific impacts of geoid conversion errors that are not evident in regional and global statistics. A total of 46 evaluated studies (categorized as follows: correct implementation of sea-level reference (n = 1); incorrect implementation of sea-level reference (n = 9); absence of sea-level reference (n = 36)) are included in the screened IPCC AR6 reports, most frequently in chapters from Working Group II and particularly (n = 25) in the Cross-Chapter Paper (CCP2) (Supplementary Figs. Our quantitative comparison of selected assessments indicates that datum conversion errors23 (potentially for ref. 9) and sea-level reference omissions (see ref. 46, for example) may have led to underestimated coastal exposure in the IPCC reports. For instance, CCP2-reported estimates of people residing in the LECZ (896 million; about 11% of the global population in 2020)45 are below our estimates following proper MDT referencing (966 million to 1.07 billion; 12.3–13.7% of the 2020 global population) (Supplementary Table 5). Our study reveals fundamental misalignment issues of sea level and coastal elevation throughout a wide body of scientific literature, which introduces errors and creates large uncertainties in the vast majority of coastal hazard and SLR and/or RSLR impact assessments. From all evaluated studies, more than 99% did not use sea-level information, omitted or made errors during sea-level datum conversion and missed crucial datum and processing documentation, rendering the studies irreproducible. In most cases, the encountered methodological issues lead to an underrepresentation of coastal sea-level height, causing existing assessments to underestimate the spatial extent and timing of future RSLR and coastal hazard impacts. This raises concerns about the correctness and reliability of existing assessments and calls for re-evaluation of the workflows and results. At present, we risk that global efforts to improve sea-level measurements and projections to mm accuracy (see, for example, refs. 4,47) are nullified by erroneous sea-level and elevation data implementation in coastal hazard and SLR impact assessments. Our findings reveal a community-wide blind spot, which calls for a systemic change in how we deal with sea-level and (coastal) land elevation data in the global scientific community and beyond. A potential explanation for the encountered vertical datum issues may be the current constellation that leaves complex geodetic transformations, required to correctly use elevation data for coastal hazard assessments, in the hands of non-specialist end users unfamiliar to the required processing steps. One solution to avoid future errors from omitted or wrongly performed datum conversion and sea-level referencing may lie in the hands of the data providers, which could provide readily combined products of digital terrain with sea level to facilitate proper end use, as we do so in this paper (see Data availability section). This would especially make sense for DEMs that are specifically developed to target the coastal zone and facilitate coastal hazard assessments, such as CoastalDEM40,48 and DeltaDTM43. Although we addressed issues on coastal elevation and sea level, we found indications for similar issues in studies using bathymetry data, suggesting further future research and critical reflection in those domains as well. The fact that geoid models are developed and performing relatively well in reflecting local sea-level height in the Global North (for example, the United States and Western Europe) may perhaps explain the overconfidence placed in geoid-model performance by Global-North-based scientists when performing assessments on a global scale or in other regions of the world in which geoid models perform less well. Although we did not further investigate or test this hypothesis, we raise this to call for further scrutiny on the Global North–South transferability of scientific approaches and datasets in future research. A noteworthy example of incompatible North–South transferability occurred with CoastalDEM v.1.1 (ref. 48), of which the neural-network approach used to create the CoastalDEM was Global-North-trained (the United States) and Global-North-validated (Australia), but performed considerably worse elsewhere. The approach, for example, placed half of the entire Mekong Delta already well below present-day sea level, thereby greatly overestimating consequent population exposure to high-water levels in the region (Extended Data Fig. Our results indicate that scientific peer-review has so far been unsuccessful in withholding the investigated errors from publication and propagation through the literature. Journals could introduce dedicated steps in their submission and peer-review procedures, for example, by providing elevation and datum documentation guidelines, requesting author declarations and adopting review checklists to aid referees. Apart from helping to avoid errors by ensuring proper data use and datum conversion and making studies transparent and replicable, these actions will also raise awareness on proper vertical referencing with the wider research community. This key factor will become even more relevant with the emergence of new elevation datasets (for example, high-accuracy measurements of relative sea level and coastal elevation using ICESat and SWOT data), which will also require proper vertical referencing and documentation when applied for further assessment. Moreover, the correct use of sea-level and land elevation data is also critical to the emerging integration of vertical land motion (VLM) into RSLR projections5,39,49. We suggest adopting a proper ‘dynamic elevation' approach, combining multiple elevation datasets50, performing multi-source uncertainty and accuracy evaluations38 and correctly interpreting and integrating VLM observations49, to create reliable, state-of-the-art projections of future RSLR. This study probably reveals only the tip of the iceberg, as the evaluated publications form only a representative selection of the full body of coastal hazard assessments. It is concerning that many of the evaluated studies (see, for example, refs. 22,24) are used to underpin SLR impact and coastal hazard exposure statements in IPCC reports4,47, which, in turn, inform global disaster risk reduction (United Nations Office for Disaster Risk Reduction) efforts, governments and policymakers worldwide on coastal vulnerability, adaptation needs and timelines, as well as provide quantitative input for climate risk rankings and loss and damage discussions. We recommend that future IPCC reports include a specific review step to verify the methodological validity of referenced coastal hazard assessments. Apart from scientific publications, our investigations suggest that much grey literature, such as policy-forming documents, governmental reports and other consultancy-based assessments, especially in the more data-sparse Global South, focusing on coastal exposure and risk, contain similar issues. We did not investigate whether the issues identified in the above-mentioned reports and assessments have led to misinformed decision-making, but this cannot be ruled out. Our findings may have far-reaching implications for existing coastal adaptation, protection and mitigation strategies, especially those using satellite-derived elevation data as information base. This necessitates re-evaluating existing coastal hazard assessments to rule out vertical reference and sea-level datum issues and, if those assessments informed decision-making, potentially updating and expediting implementation timelines of coastal adaptation strategies, as exposure thresholds may be reached much sooner than previously projected. We conducted a systematic and reproducible literature research and evaluation, adhering to standards of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) process. All details of the literature research, selection and evaluation are provided in Supplementary Tables 1 and 2. We used the Scopus search engine and used a search term for titles, abstracts and keywords that covered keywords from low-lying coastal environments, processes and hazards, as well as elevation information and elevation dynamics, impact and/or hazard assessments, projections or evaluations (Supplementary Table 1). The Scopus search engine was used as it constitutes one of the most transparent search engines. We are aware of the limitation that Scopus does not include the entire available (scientific) literature and therefore might introduce a bias into our investigation. Therefore, we performed a second independent literature evaluation on 96 additional studies (90 after excluding publications by the authors; non-systematically collected through other search engines) (Extended Data Fig. All results presented in the main paper are based on the Scopus-based systematically researched literature only. From the initial 35,001 documents that were identified through Scopus, we included only peer-reviewed scientific articles, reviews including original analyses, or data papers. Moreover, we limited our search to the most recent 15 years (2009–2025) to include only the most recent research from the period during which both global DEMs and global MDT data were available and well-established, thereby ensuring a levelled playing field of global data availability across the globe to evaluate datum conversion and sea-level reference methodologies for assessments. Further filters applied refer to the exclusion of certain unrelated subject areas, keywords and limiting the search to the English language. This refined search term reduced the amount of literature to 7,241 documents that were subsequently screened in two consecutive rounds following established inclusion and exclusion criteria (Supplementary Fig. The first screening was conducted based on title, abstract and keywords, and the second screening was conducted based on in-depth analysis of the full article and, when needed, its references and supplementary material. The screening was consistently performed by K.S., whereas a subset of 1,000 articles (first screening) was cross-checked and validated by P.S.J.M., hereby not encountering conflicting decisions. Publications that were categorized as ‘yes' and ‘maybe' in the first screening were included in the second, full-article screening. The inclusion criteria were focus on coastal land and coastal sea level, focus on coastal hazard and impact assessment, focus on coastal lowlands and the use of satellite-borne elevation information. The main exclusion criteria were the absence of elevation data use, use of airborne lidar without combination with satellite-borne elevation data, methodological and technical articles without application, articles referring to elevation data cited in tertiary source or beyond (only primary or secondary sourced studies included) and focus on secondary hazards (cascading effects). The finally included literature, consisting of 385 publications, underwent an in-depth evaluation of their respective research data and methodology, scrutinizing them, and—where needed—also their secondary references, to assess all relevant details related to study area, distance to shore, study focus and/or flood type, number of DEMs, DEMs used, actuality of DEM, DEM type, DEM coverage, acquisition and technique, performance of DEM accuracy assessment, DEM accuracy, spatial resolution, DEM availability, horizontal datum used, vertical datum documentation, vertical datum conversion, vertical datum type, vertical datum used, implementation of sea-level reference, and—where needed—additional comments (Supplementary Table 2). The individual papers were accessed by either the Scopus search engine, ISI Web of Science, Science Direct, Google Scholar or ResearchGate. Only publications that handle all datasets in a consistent vertical reference frame (that is, performing a correct vertical datum conversion) and refer to the latest available sea-level information at their time (that is, using an up-to-date sea-level reference and correct vertical datum conversion) were considered as proper. As vertical datum conversion always requires additional data next to a DEM-based coastal elevation dataset (for example, vertical datum information, offset between geoid/geoids and tidal level, sea-level height data) as well as a conversion procedure (including a specific GIS or coding environment or conversion service), we expect researchers who perform a datum conversion to document the additional datasets and procedures in their papers and/or supplementary information or supplementary data. In case a paper does not provide any documentation of additional datasets or required conversion steps, we presume that no datum conversion was performed. This assumption was confirmed to be correct for numerous papers for which the absence of datum conversion becomes apparent in the results of the study. Finally, we evaluated more than 90 additional publications that fitted the selection criteria, but were not included in the Scopus-searched literature. This subset was separately collected by the authors through additional, non-systematic literature searches, for example, to include relevant IPCC-referenced assessments, and served as an independent dataset to evaluate potential bias present in our systematic Scopus-searched literature dataset (Supplementary Data 1). The additional literature dataset underwent the same scrutinizing evaluation (Supplementary Table 1). We found nearly equivalent findings on the occurrence and percentage distribution for the various vertical datum and sea-level reference issues (Extended Data Fig. 8) as for the systematic review (Fig. 1), which suggests that our Scopus-based systematic review is unbiased and provides representative results for the existing body of literature. To obtain land elevation above continuous local sea level globally and quantify the discrepancies to land elevation with respect to global geoids or datum conversion errors, we processed four of the most recent DEMs at a global scale by converting them from their original vertical reference system to MSL as indicated by the MDT product. The DEMs included were CoastalDEM v.2.1 (original vertical reference system: EGM96, spatial resolution: 90 m × 90 m)40, FABDEM v.1.0 (original vertical reference system: EGM2008, spatial resolution: 30 m × 30 m)41, GLL-DTM v.2 (original vertical reference system: MDT53, spatial resolution: 1° × 1°, that is, about 1,000 m × 1,000 m)42 and DeltaDTM v.1 (original vertical reference system: EGM2008, spatial resolution: 30 m × 30 m)43. The publicly available GLL-DTM v.2 (referenced to CNES-CLS13 MDT53) contains a geoid conversion error (that is, the EGM-DIR R4 geoid was assumed to equal the EGM96) and does not apply the latest available MDT dataset, at the time. Therefore, we obtained the pre-converted GLL-DTM v.2 (referenced to EGM96) from the authors and conducted the vertical datum conversion to MSL44. We used the latest available global MDT HYBRID-CNES-CLS2022 dataset44, which provides sea surface height above geoid (GOCO06s) across the globe, measured by satellite altimetry and combined with gravitational field information, oceanographic data from drifting buoys, high-frequency radar velocities and hydrological profiles. The MDT dataset44 provides spatially continuous information of sea surface height above the GOCO06s geoid at a resolution of 0.125° averaged over a period from 1993 to 2021. Therefore, it provides the latest available information on global MSL and an accurate substitute for local tide gauge information in those regions. Several studies confirm the accuracy of MDT data in the range of cm (see, for example, refs. 54,55), although vertical accuracy decreases up to about 4 cm within 10 km of the coast56. It serves as an open-accessible product that—if properly aligned with elevation information—can be used to adjust elevation with respect to local sea level continuously along the coastlines of the world. The following section documents a proper, consistent and reproducible vertical datum conversion of the four DEMs, providing all required datasets, datum information, processing steps and used software environments and may serve as an example for future studies or existing studies aiming to re-evaluate previous assessments. For all our computations, we used the ArcGIS Pro environment. To reference coastal elevation of the DEMs to MSL as given by MDT, the offsets of the underlying respective vertical reference systems (that is, EGM96, EGM2008 and GOCO06s) were determined (Extended Data Fig. Geoid information was obtained from the openly accessible calculation service of the International Centre for Global Earth Models from GFZ Helmholtz Centre for Geosciences57. Point data on geoid height anomaly to the WGS84 ellipsoid was obtained for the entire globe at a resolution of 0.085° for the EGM96, EGM2008 and GOCO06s geoids, respectively. For each geoid, the height anomaly points were interpolated into a global raster by using multiquadric radial basis functions, which gave the most accurate interpolation results and is also used by gravitational field modelling studies58,59. Subsequently, the global geoid raster files were resampled to a common spatial resolution of 90 m × 90 m (for CoastalDEM v.2.1, FABDEM v.1.0 and DeltaDTM v.1) and 1,000 m × 1,000 m (for GLL-DTM v2) by using bilinear resampling to be comparable with the spatial resolution of the DEMs throughout the entire datum conversion process. Geoid offsets were determined by subtracting the GOCO06s geoid height anomaly raster from the EGM96 and EGM2008 geoid height anomaly rasters. Using MSL as indicated by MDT as a vertical datum for land elevation data requires the extrapolation of MDT data over land. We extracted point values from the MDT HYBRID-CNES-CLS2022 raster using bilinear interpolation at point locations and subsequently extrapolated them over land using an inverse distance weighting algorithm and a smooth neighbourhood type with a smoothing factor of 0.5. The resulting raster dataset was resampled to two spatial resolutions of 90 m × 90 m (for CoastalDEM v.2.1, FABDEM v.1.0 and DeltaDTM v.1) and 1,000 m × 1,000 m (for GLL-DTM v.2) by using bilinear resampling to enable comparability with the spatial resolution of the DEMs and to avoid the introduction of potential artefacts stemming from large differences in spatial resolution in the datum conversion process. The processing of DEMs involved two main processes (Extended Data Fig. Therewith, the vertical references of DEM-derived land elevation and MDT-derived sea surface height are aligned to a common datum (that is, GOCO06s), which is a prerequisite to obtain land elevation above MSL by subtracting the MDT data from the respective DEM (Extended Data Fig. To automate the processing of the vertical datum conversion of the DEMs, we applied two ArcGIS Pro model workflows to convert DEM tiles from the EGM geoids to GOCO06s and subsequently to MDT, while preserving their original spatial resolution and properties (the Python codes are provided online). As the performance and the reliability of MDT extrapolation over land decreases with increasing distance from the sea, we applied a distance threshold of 500 km from the coastline (as defined by Open Street Map60), for which we consider this approach of DEM vertical datum conversion valid. We converted all land elevation information within the distance threshold to MSL, thereby ensuring adequate extrapolation performance and full coverage of vast low-lying coastal plains and river deltas such as the Ganges–Brahmaputra–Meghna Delta. We investigated coastal land elevation and sea-level height globally by extracting point elevation data from the original and vertically converted DEMs and their respective differences, at a 90 m interval along the coastline using bilinear interpolation of values at point locations and excluding no data values and water bodies before rasterization. As some of the evaluated DEMs contain large negative, unrealistic elevation values that are likely artefacts from the source DEM acquisition and post-process steps, we excluded these by applying a minimum elevation threshold of 7 m below MSL, which represents some of the lowest elevations in the coastal lowlands worldwide, such as the Netherlands (see, for example, ref. The elevation statistics (including minimum, maximum, mean, median and standard deviation) were calculated for global (both including and excluding Antarctica), continental and regional scales (using administrative boundaries provided in ref. 51). The global statistics provided in the main text and figures exclude the results for Antarctica, as there are no people living there. To show the impacts of neglected conversion to a sea-level datum, we investigated the impact of 1 m RSLR on area and population when simulated for DEMs with their original vertical reference system and after conversion to MSL, apart from computing area and population at present already below sea level. Similarly, we investigated the discrepancies in area and population within the 10 m LECZ when different DEMs with and without proper vertical datum alignment are used. After no data values and large negative values (that is, ≤7 m below MSL) were excluded and water bodies masked, we reclassified the DEMs applying thresholds of ≤1 m and ≤10 m, respectively. Area in km2 was calculated for global, continental and regional scales (Supplementary Fig. We limit our assessment to relative elevation only, and do not apply a hydrodynamic inundation (for example, bathtub) approach, nor report on flood or inundation extent or impacts in our results. To estimate population currently below MSL with 1 m RSLR and within the LECZ, we use three global population datasets and thereby avoid potential bias in absolute population counts arising from single datasets25,61,62. Uncertainties in and between population data stem from resolution (grid cell size) and quality of input data (for example, census data) and ancillary products as well as models to calculate statistics (ref. 63 and references therein), daily population dynamics and inconsistencies in administrative boundary data61,62. Therefore, we apply a multi-dataset approach and use unconstrained WorldPop data from 2020 at 100 m spatial resolution52 as well as the LandScan Global dataset (800 m spatial resolution) for the years 2020 (to be comparable to the WorldPop 2020 data)64 and the latest available 2023 data65. For the WorldPop dataset, we used data for individual countries as the global dataset provides only aggregated data that would lead to substantial overestimation in population. Population estimates were derived for global, continental and regional scales using zonal statistics. Global population was calculated by computing binary rasters for each DEM (1 if elevation is between 7 m below MSL and ≤0 m, ≤1 m and ≤10 m above MSL, respectively) and aggregating these at the respective (coarser) resolution of the population datasets, using the arithmetic mean. This created the fraction of which each raster cell meets the elevation requirement. Subsequently, the raster was multiplied by the respective population data and spatially summed to estimate population for each extent of interest, implicitly assuming equal distribution of population within a single population data cell. We note that this constitutes an uncertainty factor as people may be distributed disproportionately within a single raster cell, but data resolution restricts further detailing. Another shortcoming is that we do not account for population change in our exposure projections and use static population numbers. Adding a projection of population change in our 1 m RSLR scenarios is not possible, as the projection is spatio-temporally variable. Therefore, the actual number of future exposed population is probably higher, given the predominant projected growth of the human population, particularly in coastal zones of the Global South (see, for instance, refs. All reported coastal elevation values, RSLR impact and LECZ statistics, as well as the respective deviations between estimates for DEMs with and without proper datum conversion, were quantified in absolute value and as discrepancies in percentage (Supplementary Data 2 and 3). To investigate links of the evaluated literature to the latest IPCC reporting cycle (AR6), we screened the AR6 working group (WG)I–III and the Special Report on the Ocean and Cryosphere in a Changing Climate (SROCC) reports against our evaluated literature reference list. Through a reproducible, large-language-model-supported, screening protocol (protocol details are available in the Supplementary Information), we used ChatGPT-5 to screen all the chapter-specific literature lists from the AR6 WGI, II, III reports and WGII cross-chapter papers (using .bib files available on https://www.ipcc.ch/report/ar6/wg1/downloads/, https://www.ipcc.ch/report/ar6/wg2/downloads/, https://www.ipcc.ch/report/ar6/wg3/downloads/) and the SROCC report chapters (using PDFs of individual chapters, available on https://www.ipcc.ch/srocc/download/) for using references from our systematic review (385) and additional literature (96). We found that 46 studies in our systematic review and 29 in our additional literature were included as references in the IPCC reporting (Supplementary Table 5, Supplementary Figs. We then grouped these studies according to their evaluated categories: proper integration of sea-level reference (n = 1 and 2, for systematic review and additional literature, respectively); incorrect integration of sea-level reference (n = 9 and 7, for systematic review and additional literature, respectively); and absence of sea-level reference (n = 36 and 20, for systematic review and additional literature, respectively) (Supplementary Table 5) and performed a quantitative comparison (area and/or population exposure) with several representative and methodologically comparable studies from each category with the results from our meta-analyses, to quantify the magnitude of error in the existing, IPCC-referenced impact assessments as a result of various vertical referencing issues. The original DEMs, MDT and population data used in this study are available in their respective online repositories40,41,42,43,44,52. The processed, global DEMs, converted to local MSL using MDT data and used for the meta-analyses in this study, are available for reuse and are accessible at Zenodo (https://doi.org/10.5281/zenodo.17722669). All computations were performed using ArcGIS Pro (https://www.esri.com/en-us/arcgis/products/arcgis-pro/overview) and layouting was done using QGIS (https://qgis.org/). The results of the in-depth literature evaluation, workflow protocol and all statistics of the coastal impact meta-analyses are available in the Supplementary Data and Supplementary Information. The codes used for this article are available here at Zenodo (https://doi.org/10.5281/zenodo.17953234). Best practices for elevation-based assessments of sea-level rise and coastal flooding exposure. Minderhoud, P. S. J., Coumou, L., Erkens, G., Middelkoop, H. & Stouthamer, E. Mekong delta much lower than previously assumed in sea-level rise impact assessments. Seeger, K. et al. Assessing land elevation in the Ayeyarwady Delta (Myanmar) and its relevance for studying sea level rise and delta flooding. Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC, 2023). Measuring, modelling and projecting coastal land subsidence. Global subsidence of river deltas. Ericson, J. P., Vörösmarty, C. J., Dingman, S. L., Ward, L. G. & Meybeck, M. Effective sea-level rise and deltas: causes of change and human dimension implications. Nicholls, R. J. et al. A global analysis of subsidence, relative sea-level change and coastal flood exposure. Uncertainty and bias in global to regional scale assessments of current and future coastal flood risk. Schumann, G. J.-P. & Bates, P. D. The need for a high-accuracy, open-access global DEM. Neumann, B., Vafeidis, A. T., Zimmermann, J. Future coastal population growth and exposure to sea-level rise and coastal flooding – a global assessment. Causes and consequences of error in digital elevation models. Kulp, S. & Strauss, B. H. Global DEM errors underpredict coastal vulnerability to sea level rise and flooding. & Zeray Öztürk, E. Height biases of SRTM DEM related to EGM96: from a global perspective to regional practice. Local orthometric height based on a combination of GPS-derived ellipsoidal height and geoid model: a review paper. Becker, M. et al. Coastal flooding in Asian megadeltas: recent advances, persistent challenges, and call for actions amidst local and global challenges. Nyoka, C. J., Din, A. H. M., Pa'suya, M. F. & Omar, A. H. Rigorous evaluation of global geopotential models for geoid modelling: a case study in Kenya. Edmonds, D. A., Caldwell, R. L., Brondizio, E. S. & Siani, S. M. O. Coastal flooding will disproportionately impact people on river deltas. Muis, S. et al. A comparison of two global datasets of extreme sea levels and resulting flood exposure. & Pronk, M. New ICESat-2 satellite LiDAR data allow first global lowland DTM suitable for accurate coastal flood risk assessment. New LiDAR-based elevation model shows greatest increase in global coastal exposure to flooding to be caused by early-stage sea-level rise. Projections of global-scale extreme sea levels and resulting episodic coastal flooding over the 21st century. & Vernimmen, R. Global LiDAR land elevation data reveal greatest sea-level rise vulnerability in the tropics. Muis, S., Verlaan, M., Winsemius, H. C., Aerts, J. C. J. H. & Ward, P. J. A global reanalysis of storm surges and extreme sea levels. MacManus, K., Balk, D., Engin, H., McGranahan, G. & Inman, R. Estimating population and urban areas at risk of coastal hazards, 1990–2015: how data choices matter. Zhang, S. et al. Quantitative risk assessment of typhoon storm surge for multi-risk sources. Climate change and urban sprawl: unveiling the escalating flood risks in river deltas with a deep dive into the GBM river delta. Global dataset of extreme sea levels and coastal flood impacts over the 21st century. Bolleter, J., Grace, B., Hooper, P. & Foster, S. Wet-bulb temperature and sea-level rise in the United Arab Emirates – planning responses. Lincke, D. & Hinkel, J. Coastal migration due to 21st century sea-level rise. Ghomsi, F. E. K. et al. Sea level rise and coastal flooding risks in the Gulf of Guinea. Brown, S. et al. What are the implications of sea-level rise for a 1.5, 2 and 3 °C rise in global mean temperatures in the Ganges-Brahmaputra-Meghna and other vulnerable deltas? Gebremichael, E. et al. Assessing land deformation and sea encroachment in the Nile Delta: a radar interferometric and inundation modeling approach. Merkens, J.-L., Lincke, D., Hinkel, J., Brown, S. & Vafeidis, A. T. Regionalisation of population growth projections in coastal exposure analysis. Vousdoukas, M. I. et al. African heritage sites threatened as sea-level rise accelerates. North American Vertical Datum of 1988 (NAVD 88) (NOAA, 2018). Seeger, K. & Minderhoud, P. S. J. Attributing uncertainties in elevation assessments for data-sparse coastal lowlands using global elevation models: a globally applicable approach showcasing the Vietnamese Mekong Delta. Preprint available at Research Square https://doi.org/10.21203/rs.3.rs-7706762/v1 (2025). Minderhoud, P. S. J., Middelkoop, H., Erkens, G. & Stouthamer, E. Groundwater extraction may drown mega-delta: projections of extraction-induced subsidence and elevation of the Mekong delta for the 21th century. Pronk, M. DeltaDTM: a global coastal digital terrain model, v.1, 4TU. Hybrid Mean Dynamic Topography CNES-CLS22-CMEMS2020. Long-term sea-level rise necessitates a commitment to adaptation: a first order assessment. & O'Neill, B. C. Spatially explicit global population scenarios consistent with the shared socioeconomic pathways. Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC, 2021). & Strauss, B. H. CoastalDEM: a global coastal digital elevation model improved from SRTM using a neural network. Minderhoud, P. S. J., Shirzaei, M. & Teatini, P. From InSAR-Derived Subsidence to Relative Sea-Level Rise – A Call for Rigor. Seeger, K., Minderhoud, P. S. J., Hauser, L. T., Teatini, P. & Woillez, M.-N. How to Assess Coastal Flood Risk in Data-Sparse Coastal Lowlands? Accurate Information on Land Elevation Is Key. A Question of Development No. United Nations World Food Programme. World Administrative Boundaries - Countries and Territories. Population Counts: unconstrained Individual Countries 2020 (100 m Resolution). Global High Resolution Population Denominators Project – Funded by The Bill and Melinda Gates Foundation (OPP1134076) (2018). Rio, M.-H., Mulet, S. & Picot, N. Beyond GOCE for the ocean circulation estimate: synergetic use of altimetry, gravimetry, and in situ data provides new insight into geostrophic and Ekman currents. Knudsen, P., Bingham, R., Andersen, O. & Rio, M.-H. A global mean dynamic topography and ocean circulation estimation using a preliminary GOCE gravity model. Hamden, M. H., Din, A. H. M., Wijaya, D. D., Yusoff, M. Y. M. & Pa'suya, M. F. Regional mean sea surface and mean dynamic topography models around Malaysian seas developed from 27 years of along-track multi-mission satellite altimetry data. Birol, F. et al. Understanding uncertainties in the satellite altimeter measurement of coastal sea level: insights from a round-robin analysis. Ince, E. S. et al. ICGEM – 15 years of successful collection and distribution of global gravitational models, associated services, and future plans. Local geoid determination in strip area projects by using polynomials, least-squares collocation and radial basis functions. Using radial basis functions in airborne gravimetry for local geoid improvement. Data Derived from OpenStreetMap for Download – Coastlines. Lichter, M., Vafeidis, A. T., Nicholls, R. J. & Kaiser, G. Exploring data-related uncertainties in analyses of land area and population in the “Low-Elevation Coastal Zone” (LECZ). Uncertainties in measuring populations potentially impacted by sea level rise and coastal flooding. The spatial allocation of population: a review of large-scale gridded population data products and their fitness for use. Rose, A. et al. LandScan Global 2020. Oak Ridge National Laboratory. Lebakula, V. et al. LandScan Silver Edition. Oak Ridge National Laboratory. Reimann, L., Vafeidis, A. T. & Honsel, L. E. Population development as a driver of coastal risk: current trends and future pathways. Tessler, Z. D. et al. Profiling risk and sustainability in coastal deltas of the world. & Strauss, B. H. New elevation data triple estimates of global vulnerability to sea-level rise and coastal flooding. We thank A. Collins, R. Kopp and Z. Siyum for providing extensive feedback and suggestions that helped to shape our assessments further and considerably strengthened the manuscript. We thank A. Bolten, M. van der Meij and L. Steinbuch for supporting the workstation use and server capacities needed to conduct the global processing, and for support during the data publication, and R. Vernimmen for providing the GLL DTM v.2 referenced to EGM96. For this paper, we went deeper than ever before. We thank J. Pernack and D. Brill for sparking and encouraging an everlasting scientific curiosity about Earth. We acknowledge the inspiration and remember H.-J. Here's to the crazy ones and an amazing better half. We thank the brothers Arthur and Luuk Minderhoud for providing P.S.J.M. ample opportunity for nighttime contemplations, which played a key part in shaping this manuscript. We also thank our supportive families and a wonderful grandma. Finally, we acknowledge the inquisitive ‘paper tiger' that grew into a graceful lion. May it bask in the morning sun and sleep soundly tonight. acknowledges the funding from the Dutch Science Foundation (NWO) under the NWO Veni TTW 2022 (Applied and Technical Sciences) call with the project: ‘Drowning Deltas—why deltas sink and what to do about it' (no. This work received funding to support open-access publication by the Dutch Ministry of Agriculture, Fisheries, Food Security and Nature under the Wageningen University & Research Knowledge Base Programme (KB). Soil Geography and Landscape Group, Wageningen University and Research, Wageningen, The Netherlands Department of Civil, Environmental and Architectural Engineering, University of Padova, Padova, Italy Department of Groundwater and Water Security, Deltares Research Institute, Utrecht, The Netherlands Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar investigated the literature review and data processing. investigated the result analysis and interpretation. investigated the IPCC reference screening. handled the project administration. wrote the original draft and reviewed and edited the final paper. Correspondence to Katharina Seeger or Philip S. J. Minderhoud. The authors declare no competing interests. Nature thanks Alexandra Collins, Robert Kopp, Zenebe Siyum and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Use of DEMs based on (A) study focus/flood type and (B) study area. The Sankey diagrams for the results of this study were created using SankeyMATIC (https://sankeymatic.com/). A, Schematic overview of vertical datums. B, Calculation of offsets between vertical reference systems as prerequisite for vertical datum conversion of elevation data. C, Preparation of mean dynamic topography (MDT) data as prerequisite for vertical datum conversion of elevation data. D, Vertical datum conversion of elevation data from their original vertical reference system to the latest earth gravity model GOCO06s and to mean sea level as indicated by MDT. Global coastal elevation according to GLL-DTM v2 above (A–C) mean sea level after refs. 21,42,53 and (D–F) mean sea level following this study (MDT HYBRID-CNES-CLS2022), as well as (G–I) the offset between both. 21,42) contains a geoid conversion error (i.e. the EGM-DIR R4 geoid was assumed to equal EGM96, hence introducing the offset between both geoids (Extended Data Fig. 9) when applying the MDT data. In addition, the GLL-DTM v2 applied an, at the time already, outdated MDT dataset (i.e. CNES-CLS13 MDT53). The results were visualized using QGIS v.3.28.6 and shapefiles from ref. Global coastal elevation according to DeltaDTM v1 above (A–C) geoid EGM2008 and (D–F) mean sea level (MDT HYBRID-CNES-CLS2022), as well as (G–I) the offset between both. For visualization purposes, the spatial scale of the data shown was resampled to 1° using bilinear resampling while all statistics are given at 90 m spatial resolution. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. Meta-analysis showing the impact of the most frequently observed errors in existing coastal hazard assessments, i.e. omitting or improper alignment of measured coastal sea level to land elevation, using modern, globally available digital elevation models (CoastalDEM v2.1 (ref. Due to the widespread omission to tie elevation data to local sea level instead of global geoid models, area impacted by 1 m RSLR is 31% to 37% larger globally. Maximum discrepancies of up to 94% (that is up to 93,800 km2) are observed along the coastlines of Southeast Asia where – if properly assessed – up to 100,300 km2 will fall below sea level following 1 m RSLR. All statistics are computed for the respective spatial resolution of the DEMs. Colour outlines and grey-shaded areas indicate the regions and subregions used (Supplementary Fig. All data related to this figure is included in Supplementary Data 3. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. Impact of 1 m RSLR and LECZ based on GLL-DTM v2 referenced to (A–K) mean sea level after refs. 21,42,53 and (L–V) mean sea level following this study (MDT HYBRID-CNES-CLS2022). The results were visualized using QGIS v.3.28.6 and shapefiles from ref. 51 (Open Government Licence v3.0) and ref. Impact of 1 m RSLR and LECZ based on DeltaDTM v1 referenced to (A–K) EGM2008 and (L–V) mean sea level (MDT HYBRID-CNES-CLS2022). The results were visualized using QGIS v.3.28.6 and shapefiles from ref. 51 (Open Government Licence v3.0) and ref. The Sankey diagram for the results of this study was created using SankeyMATIC (https://sankeymatic.com/). We show this discrepancy to quantify potential residual errors stemming from incomplete vertical datum conversion (erroneously assuming the EGM-DIR R4 geoid of the CNES-CLS13 MDT53 to equal the EGM96 geoid and thereby omitting this geoid-to-geoid datum conversion). This error is present in ref. 21 who based their processing approach on ref. 19 (presumably containing the same conversion error, but this is not irrefutable due to incompleteness of conversion documentation). This error is likely present in other papers (e.g. ref. 45) that follow the described approach of MDT implementation. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. CoastalDEM v1.1 and v2.1 (90 m spatial resolution, respectively) were acquired from Climate Central (https://go.climatecentral.org/coastaldem/). The vertical datum was converted to local MSL by properly applying the most recent MDT data44 as described in the Methods section (see also Extended Data Fig. TopoDEM_v2 was created by interpolation of topographical elevation points referenced to the same local MSL and validated locally using independent geodetic measurements38. TopoDEM_v2_MSL provides a mean delta plain elevation of about 0.77 m, while CoastalDEM_v1.1_MSL documents the mean elevation of the delta plain more than a meter lower, at about −0.65 m. This showcases the neural network to correct the SRTM DEM, which was trained on US coastlines and validated in Australia, to be performing very inadequate in other places of the world, thereby, at least for the Mekong Delta, largely overestimating the amount of people exposed to high water levels as reported in ref. These issues were largely addressed in CoastalDEM 2.1, which was released in 2021 (ref. However, it still suffers from inaccuracies due to sensing artefacts. The elevation profiles of CoastalDEM v1.1 and v2.1 over transect A–A' were binned to 500 m (median) for visualization purpose to match the spatial resolution (500 m) of TopoDEM_v2. The results were visualized using QGIS v.3.28.6 and shapefiles from ref. This file includes Supplementary Figs. 1–11 and Supplementary Tables 1–5 as well as the protocol for screening the evaluated literature against IPCC AR6 WG I–III and SROCC references using ChatGPT-5. Scientific literature included into detailed evaluation following evaluation categories outlined in Supplementary Tables 1 and 2. Further literature and material referred to in those publications were investigated as well but not included in the evaluation process. Coastal elevation, geoid and sea-level statistics. Coastal minimum, maximum, mean and median elevations and standard deviations for CoastalDEM v.2.1, originally referenced to a global gravity model (EGM96) (ref. 40), FABDEM v.1.0, originally referenced to a global gravity model (EGM2008) (ref. 41), GLL-DTM v.2, originally referenced to MSL (MDT after ref. 42), and DeltaDTM v.1, originally referenced to a global gravity model (EGM2008) (ref. 43) and referenced to MSL (MDT after ref. Offsets between both versions were calculated by subtracting the DEM without vertical datum correction from the DEM with the vertical datum correction conducted in this study. All statistics are based on coastal elevations extracted every 90 m from the DEM. Although both versions of the GLL-DTM are referenced to a sea-level dataset, the original version of GLL-DTM v.2 does not use the latest available dataset and suffers from incorrect vertical datum conversion (Supplementary Data 1). Processing details are given in the Methods. The dataset further includes coastal minimum, maximum, mean and median and standard deviation statistics for geoid heights of EGM96 and EGM2008 relative to WGS84, respectively, as well as the offset between EGM96 and EGM-DIR R4, and sea-level heights as indicated by MDT (ref. 44) above geoid (GOCO06s, EGM96, EGM2008, EGM-DIR R4 and GOCO2025s, respectively). Assessments of area and population below present MSL, exposed to 1 m relative SLR (RSLR) impact and in the 10 m LECZ. Area and population below present MSL, falling below sea level following 1 m RSLR and being exposed in the 10 m LECZ for CoastalDEM v.2.1, FABDEM v.1.0, GLL-DTM v.2 and DeltaDTM v.1 in absolute numbers (in km2, population counts) and relative (in %) to total area and population of the areas of interest, as well as their discrepancy from area and population estimates for correctly conducted assessments with vertical datum conversion to actual local MSL. Literature screening against IPCC AR6 WG I–III and SROCC references. Outcomes of the assessment whether the literature evaluated in this study had been cited in the IPCC AR6 Working Group I–III bibliographies and in the IPCC SROCC report. Details on the workflow and evaluation of the IPCC-referenced literature are given in the Methods and Supplementary Information. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Sea level much higher than assumed in most coastal hazard assessments. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. The synthesis of organoarsenicals has traditionally involved multistep procedures and the use of toxic arsenical reagents, impeding their widespread application. Now, a photocatalytic approach enables the direct conversion of arsenic sulfide minerals, such as orpiment, into various organoarsenicals using visible light. This is a preview of subscription content, access via your institution Get Nature+, our best-value online-access subscription Subscribe to this journal Receive 12 print issues and online access Prices may be subject to local taxes which are calculated during checkout Haxton, K. All about arsenic. This article examines arsenic's duality and our evolving understanding of this element. Direct conversion of white phosphorus to versatile phosphorus transfer reagents via oxidative onioation. This work reports that organoarsenicals can convert white phosphorus into useful phosphorus-transfer reagents through oxidative onioation. Tay, W. S. & Pullarkat, S. A. C–As bond formation reactions for the preparation of organoarsenic(III) compounds. This review summarizes the methodologies for the synthesis of organoarsenicals. Ye, J., Rensing, C., Rosen, B. P. & Zhu, Y.-G. Arsenic biomethylation by photosynthetic organisms. This review examines the process of arsenic methylation in photosynthetic organisms. Direct catalytic transformation of white phosphorus into arylphosphines and phosphonium salts. This study presents a photocatalytic method for the efficient synthesis of organophosphorus compounds directly from white phosphorus. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This is a summary of: Wang, Y. et al. Mineral-to-molecule arsenic transfer via photoredox catalysis. Photocatalytic synthesis of organoarsenicals directly from minerals. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Nature Photonics (2026)Cite this article Modern wireless technologies—spanning mobile communications to satellite links—rely on systems operating across disparate microwave bands. Although escalating data demands have driven the evolution from 2G to 6G, each generation has traditionally required dedicated, frequency-specific hardware, complicating multiband integration. This challenge intensifies at higher frequencies (5G and beyond), where conventional approaches incur prohibitive costs and power consumption in wireless terminals. Here we present a scalable and unified platform that supports all-generation (2G to 6G+) parallel wireless systems by combining photonic circuits with electronic metasurfaces. Using a self-synchronized dual-comb technique, we simultaneously generate over 60 reconfigurable microwave frequencies up to 100 GHz, with beamforming enabled by compact, low-power metasurfaces. This architecture facilitates all-generation wireless links with advanced modulation formats. Crucially, we demonstrate the direct drive of the wireless edge by data-centre silicon photonic transceivers, seamlessly merging data centre and wireless networks. Our solution unifies signal generation, processing and beamforming in a compact, cost-effective platform, offering a transformative foundation for future wireless systems. This is a preview of subscription content, access via your institution Access Nature and 54 other Nature Portfolio journals Get Nature+, our best-value online-access subscription cancel any time Subscribe to this journal Receive 12 print issues and online access only $21.58 per issue Buy this article Prices may be subject to local taxes which are calculated during checkout The data that support the plots within this paper and other findings of this study are available via Zenodo (https://doi.org/10.5281/zenodo.18241616)66. All other data used in this study are available from the corresponding authors upon request. The codes that support the findings of this study are available from the corresponding authors upon reasonable request. Dang, S., Amin, O., Shihada, B. & Alouini, M.-S. What should 6G be? Google Scholar Wang, W. et al. On-chip topological beamformer for multi-link terahertz 6G to XG wireless. Google Scholar Sen, P., Siles, J. V., Thawdar, N. & Jornet, J. M. Multi-kilometre and multi-gigabit-per-second sub-terahertz communications for wireless backhaul applications. Google Scholar Willner, A. E. Optical and THz broadband integrated circuits for mode-dependent free-space communications. Optical Fiber Communication Conference (OFC) Tu2A.2 (Optica Publishing Group, 2024). Bekkali, A., Ishimura, S., Tanaka, K., Nishimura, K. & Suzuki, M. Multi-IF-over-fiber system with adaptive frequency transmit diversity for high capacity mobile fronthaul. Google Scholar Saeidi, M. A., Tabassum, H. & Alouini, M.-S. Multi-band wireless networks: architectures, challenges, and comparative analysis. Google Scholar Pozar, D. M. Microwave Engineering: Theory and Techniques (John Wiley & Sons, 2021). Lee, T. H. The Design of CMOS Radio-Frequency Integrated Circuits (Cambridge Univ. Heydari, P. Terahertz integrated circuits and systems for high-speed wireless communications: challenges and design perspectives. IEEE Open J. Solid-State Circuits Soc. Google Scholar Georgiadis, A., Rogier, H., Roselli, L. & Arcioni, P. Microwave and Millimeter Wave Circuits and Systems: Emerging Design, Technologies and Applications (John Wiley & Sons, 2012). Zhang, W. et al. Broadband physical layer cognitive radio with an integrated photonic processor for blind source separation. Larsson, E. G., Edfors, O., Tufvesson, F. & Marzetta, T. L. Massive MIMO for next generation wireless systems. Google Scholar Kutty, S. & Sen, D. Beamforming for millimeter wave communications: an inclusive survey. Google Scholar Di Renzo, M., Haas, H., Ghrayeb, A., Sugiura, S. & Hanzo, L. Spatial modulation for generalized MIMO: challenges, opportunities, and implementation. Google Scholar Wei, M., Zhao, H., Galdi, V., Li, L. & Cui, T. J. Metasurface-enabled smart wireless attacks at the physical layer. Google Scholar Huang, Y. et al. Radiofrequency signal processing with a memristive system-on-a-chip. Saad, W., Bennis, M. & Chen, M. A Vision of 6G wireless systems: applications, trends, technologies, and open research problems. Google Scholar Zhang, H., Di, B., Song, L. & Han, Z. Reconfigurable Intelligent Surface-Empowered 6G (Springer International Publishing, 2021). Marpaung, D., Yao, J. Integrated microwave photonics. Google Scholar Shu, H. et al. Microcomb-driven silicon photonic systems. Google Scholar Nagatsuma, T., Ducournau, G. & Renaud, C. C. Advances in terahertz communications accelerated by photonics. Google Scholar Zhang, C. et al. Clone-comb-enabled high-capacity digital-analogue fronthaul with high-order modulation formats. Google Scholar Tetsumoto, T. et al. Optically referenced 300 GHz millimetre-wave oscillator. Google Scholar Sengupta, K., Nagatsuma, T. & Mittleman, D. M. Terahertz integrated electronic and hybrid electronic–photonic systems. Google Scholar Tong, W. et al. 200-m photonics-aided terahertz wireless transmission of 253-Gbit/s DP-OFDM signals utilizing multidimensional nonlinear equalization. Google Scholar Xiang, C. et al. 3D integration enables ultralow-noise isolator-free lasers in silicon photonics. Google Scholar Yao, X. S. & Maleki, L. Optoelectronic microwave oscillator. Google Scholar Maleki, L. The optoelectronic oscillator. Google Scholar Hao, T. et al. Breaking the limitation of mode building time in an optoelectronic oscillator. Liu, J. et al. Photonic microwave generation in the X- and K-band using integrated soliton microcombs. Google Scholar Chang, L., Liu, S. & Bowers, J. E. Integrated optical frequency comb technologies. Google Scholar Zhou, Z., Nopchinda, D., Darwazeh, I. Clock and carrier synchronized multi-band wireless communications enabled by frequency comb dissemination in radio access networks. Google Scholar Torres-Company, V. & Weiner, A. M. Optical frequency comb technology for ultra-broadband radio-frequency photonics. Laser Photonics Rev. Google Scholar Moreira, R. L. et al. Integrated ultra-low-loss 4-bit tunable delay for broadband phased array antenna applications. IEEE Photonics Technol. Google Scholar Hong, S. et al. Versatile parallel signal processing with a scalable silicon photonic chip. Zhang, X. et al. Microcomb-synchronized optoelectronics. Jin, W. et al. Hertz-linewidth semiconductor lasers using CMOS-ready ultra-high-Q microresonators. Zhang, X. et al. High-coherence parallelization in integrated photonics. Zhang, K. et al. A power-efficient integrated lithium niobate electro-optic comb generator. Cui, T. J., Qi, M. Q., Wan, X., Zhao, J. & Cheng, Q. Coding metamaterials, digital metamaterials and programmable metamaterials. 3rd Generation Partnership Project (3GPP). NR; Base Station (BS) Radio Transmission and Reception. Technical Specification number TS 38.104 (3GPP, 2025); https://www.3gpp.org/ftp/Specs/archive/38_series/38.104/ Dai, L. et al. Reconfigurable intelligent surface-based wireless communications: antenna design, prototyping, and experimental results. Zeng, S. et al. RIS-based IMT-2030 testbed for mmWave multi-stream ultra-massive MIMO communications. IEEE Wireless Commun. Xia, D. X. et al. Adaptive wireless-powered network based on CNN near-field positioning by a dual-band metasurface. Wei, B. et al. Ku/Ka dual-band shared-aperture leaky-wave antenna with fixed-frequency and wide-angle beam scanning based on ridged SIW. Che, D. Digital SNR adaptation of analog radio-over-fiber links carrying up to 1048576-QAM signals. European Conference on Optical Communications (ECOC) 1–4 (IEEE, 2020). Koepfli, S. M. et al. Metamaterial graphene photodetector with bandwidth exceeding 500 gigahertz. Lao, C. et al. Quantum decoherence of dark pulses in optical microresonators. Jin, X. et al. Microresonator-referenced soliton microcombs with zeptosecond-level timing noise. Sun, S. et al. Microcavity Kerr optical frequency division with integrated SiN photonics. et al. Dispersive-wave-agile optical frequency division. Li, W. et al. Intelligent metasurface system for automatic tracking of moving targets and wireless communications based on computer vision. Taghvaee, H. et al. Holographic mmWave metasurface integrating THz sensing for 6G wireless networks. IEEE Wireless Commun. James, A. et al. Scaling comb-driven resonator-based DWDM silicon photonic links to multi-Tb/s in the multi-FSR regime. Rizzo, A. et al. Massively scalable Kerr comb-driven silicon photonic link. Snigirev, V. et al. Ultrafast tunable lasers using lithium niobate integrated photonics. Li, K. et al. An integrated CMOS–silicon photonics transmitter with a 112 gigabaud transmission and picojoule per bit energy efficiency. Daudlin, S. et al. Three-dimensional photonic integration for ultra-low-energy, high-bandwidth interchip data links. Xing, S. et al. Seamless optical cloud computing across edge-metro network for generative AI. Sludds, A. et al. Delocalized photonic deep learning on the internet's edge. Davis, R., Chen, Z., Hamerly, R. & Englund, D. RF-photonic deep learning processor with Shannon-limited data movement. Xu, X. et al. 11 TOPS photonic convolutional accelerator for optical neural networks. Zhou, Z. et al. Communications with guaranteed bandwidth and low latency using frequency-referenced multiplexing. Hu, F. et al. 400-Gbps/λ ultrafast silicon microring modulator for scalable optical compute interconnects. Preprint at http://arxiv.org/abs/2509.01555 (2025). Lyu, Y. et al. Dynamic routing for integrated satellite-terrestrial networks: a constrained multi-agent reinforcement learning approach. Chen, Y. et al. Data for “Multiband Wireless Systems Based on Microwave Integrated Photonics with Metasurfaces”. acknowledges the National Natural Science Foundation of China (grant numbers U25D8009 and 12293052), the Beijing Municipal Natural Science Foundation (grant number Z220008), Shanghai 2025 ‘Science and Technology Innovation Action Plan', Joint Research Project of the Shijiazhuang-Peking University Cooperation Program, He Science Foundation, support from Qiming Photonics for the microcomb fabrication, High-performance Computing Platform of Peking University and Peking Nanofab. acknowledges the Beijing Outstanding Young Scientist Program (grant number JWZ020240102001). acknowledges the National Natural Science Foundation of China (grant number 62322101). Xiangpeng Zhang acknowledges the National Natural Science Foundation of China (grant number 62401020). We thank J. E. Bowers and W. Jin from the University of California, Santa Barbara, as well as D. Pan, P. Cai, N. Zhang and W. Wang at SiFotonics Technologies for their support in the experiments. We thank Y. Dai, D. Ren and H. He from Peking University for his assistance in polishing the English. We thank Z. Hao from Peking University for his assistance with the code. We thank Keysight for loaning the high-speed spectrum analyser. The experiments are supported by Peking University Nano-Optoelectronic Fabrication Center. These authors contributed equally: Yujun Chen, Jiahao Gao, Xuguang Zhang. State Key Laboratory of Photonics and Communications, School of Electronics, Peking University, Beijing, China Yujun Chen, Jiahao Gao, Xuguang Zhang, Zixuan Zhou, Xiangpeng Zhang, Xiaoyu Zhang, Lei Zhang, Lingyang Song, Boya Di & Lin Chang Department of Electrical Engineering & State Key Laboratory of Terahertz and Millimeter Waves, City University of Hong Kong, Hong Kong, China Ke Zhang, Yikun Chen, Chengfei Shang & Cheng Wang Key Laboratory of All Optical Network and Advanced Telecommunication Network, Ministry of Education, Institute of Lightwave Technology, Beijing Jiaotong University, Beijing, China School of Information Science and Technology, ShanghaiTech University, Shanghai, China Frontiers Science Center for Nano-optoelectronics, Peking University, Beijing, China Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar The experiments were conceived by L.C., B.D. and Yujun Chen. The EO comb was designed by K.Z., Yikun Chen, C.S. The metasurfaces were designed by Jiahao Gao, Xiaoyu Zhang, Z.L., Jiafan Gao, L.S. The experiments were performed by Yujun Chen, Xuguang Zhang and Jiahao Gao, with assistance from Z.Z., Xiangpeng Zhang, Xiaoyu Zhang and K.Z. The results were analysed by Yujun Chen, Jiahao Gao, Xuguang Zhang, Z.Z. All authors participated in writing the manuscript. The project was performed under the supervision of L.C., B.D. Correspondence to Lingyang Song, Boya Di or Lin Chang. The authors declare no competing interests. Nature Photonics thanks Armands Ostrovskis and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Notes I–XVIII, Figs. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Reprints and permissions Chen, Y., Gao, J., Zhang, X. et al. Multiband wireless systems based on microwave integrated photonics with metasurfaces. Version of record: 04 March 2026 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Nature Photonics © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Nature Synthesis (2026)Cite this article Installation of a functional group onto pyridines with predictable regiocontrol represents an appealing method for drug discovery programmes in the pharmaceutical industry. However, functionalization at the meta position of pyridines is often non-trivial owing to their inherent electronic properties, which commonly guide the reactions at the ortho and/or para positions. Here we report a method to enable regioselective migration of a peripheral aryl group on pyridines via 1,2-aryl migration of azacylohexadienyl radical intermediates generated by temporary dearomatization of the pyridine rings under photoredox catalysis. This process made it possible to transfer aryl substituents pre-installed at the para or ortho position of pyridines to the meta position by three different permutations: C4-to-C5, C4-to-C3 and C3-to-C2 aryl group transpositions. The present protocol exhibits broad functional group compatibility, offering streamlined access to a library of diverse meta-arylpyridines without the need for de novo synthesis. Pyridines are among the most common structural motifs found in various active pharmaceutical ingredients1,2 (Fig. The installation of functional groups of choice onto pyridines at a late stage offers an opportunity to enrich molecular libraries and thus accelerates the discovery process3,4. While recent advances in chemical synthesis have unlocked a range of non-canonical bond disconnections on pyridines, the synthetic capability of precisely controlling regioselectivity still has room for development5,6,7. The functionalization at the meta position (C3 or C5) of pyridines presents a paramount challenge compared with that at the ortho (C2 or C6) and para (C4) positions, owing to their inherent electronic properties (Fig. Processes involving regioselective metalation and the ensuing functionalization of the metallated species have advanced the state of the art in the meta-functionalization of pyridines8. The promising meta-regioselectivity for the metallation could be dictated by the directing group pre-installed at the suitable position of the pyridine substrates9 or by the precise design of ligands on the transition metal catalysis based on iridium, palladium and nickel complexes10,11,12,13,14,15,16,17. Another approach to execute meta-selective functionalization, which has recently emerged, is the use of temporary dearomatization (Fig. In this process, pyridines are first dearomatized through 1,4-hydroboration18,19,20, 1,4-hydrosilylation21, 1,2-hydride reduction of pyridinium salts22, or dearomative annulation with dimethyl acetylenedicarboxylate and methyl pyruvate23,24,25,26,27,28,29. The resulting dihydropyridine intermediates, which possess nucleophilic enamine moieties, are functionalized through ionic or radical pathways before rearomatization affords meta-functionalized pyridines. McNally demonstrated meta-halogenation of pyridines via a sequence of ring opening of pyridines to the Zincke imine intermediates, their regioselective halogenation, and rearomative ring closure30,31,32. However, meta-arylation has rarely been developed, with only limited examples reported using palladium or copper catalysis15,27, despite the fact that meta-(hetero)aryl pyridines could serve as a core structure of several pharmaceutical drugs such as etoricoxib (Fig. a, Active pharmaceutical ingredients based on pyridine motif1,2. b, Peripheral functionalization of pyridines. c, Temporary dearomatization for the meta-functionalization of pyridines18,19,20,21,22,23,24,25,26,27,28,29,30,31,32. d, Aryl group transposition on pyridines via 1,2-aryl migration. e, Synthesis of meta-arylpyridines via peripheral aryl group transposition (this work). PC, photocatalysts; pin, pinacolato; Tf, trifluoromethanesulfonate; Bn, benzyl; red., reduction; ox. oxidation; e, electron; Ar, aryl. To bridge this synthetic challenge, we conceived a peripheral editing strategy33 that allows transposition of an aryl group at the readily introducible para or ortho position of pyridines under a single-electron transfer (SET) redox catalytic manifold (Fig. For example, the Birch-type dearomatization of 4-arylpyridines I via a sequence of SET reduction and protonation forms 1,4-dihydropyridine intermediates II, which, in turn, undergo SET oxidation34 and deprotonation to form azacyclohexadienyl radicals III4. Radical 1,2-aryl migration35,36 is then envisioned to transpose the C4-aryl group to either C3 or C5 to form III3 or III5. Further aryl group transposition from C3 to C2 on III3 and from C5 to C6 on III5 would be viable to afford III2 and III6, respectively. Nonetheless, reverse aryl group transposition should be possible based on the reversible nature of 1,2-aryl group migration. Finally, SET oxidation and deprotonation of III could furnish new pyridine products IV, potentially bearing the aryl group at a different position, thereby closing the redox-neutral catalytic loop. Despite the feasibility of the underlying redox processes, the central challenge lies in achieving precise and predictable control over aryl group transposition. We hypothesized that this control arises from an equilibrium between azacyclohexadienyl radicals III, which could be modulated by the substituents on the pyridine scaffold. Here, we describe a protocol that transposes a peripheral aryl group on pyridines under visible-light photoredox catalysis (Fig. We demonstrate three different transposition permutations (C4-to-C5, C4-to-C3 and C2-to-C3), paving the way for the synthesis of meta-arylpyridines. We embarked on our investigation using methyl 4-(4-(dimethylcarbamoyl)phenyl)nicotinate (1), which is easily prepared from a readily available nicotinic acid derivative, as the model substrate (Table 1). We found that the treatment of 1 with a cyanoarene-based donor-acceptor photocatalyst, 2,4,6-tris(diphenylamino)-5-fluoroisophthalonitrile (3DPAFIPN, 2 mol%)37, is optimal to promote C4-to-C5 aryl group transposition in the presence of potassium formate (HCO2K, 5 equiv. ), caesium carbonate (Cs2CO3, 2 equiv.) and 1,1,1,3,3,3-hexafluoro-2-propanol (HFIP, 5 equiv.) as the additives in dimethylsulfoxide (DMSO, 0.1 M), under irradiation with indigo light (427 nm, 40 W), affording C5-aryl pyridine 2 in 71% assay yield based on 1H nuclear magnetic resonance (NMR) spectroscopy (67% isolated yield) (entry 1). The reaction in 10-mmol scale operated without detrimental impact on the process efficiency (entry 2). The process without 3DPAFIPN did not proceed at all (entry 3), and likewise, in the absence of HCO2K, the reaction ended up with poor conversion (entry 4). These experiments suggested that the interplay between photocatalysts and HCO2K would be key to promoting the present process (vide infra). Replacement of HCO2K with N,N-diisopropylethylamine (i-Pr2NEt) resulted in a very low product yield of 2 despite full conversion of 1 (entry 5). The solvent screening revealed that the process in dimethylformamide (DMF) was somewhat productive (entry 6), whereas acetonitrile (MeCN) was incompatible (entry 7). With the optimized conditions in hand, we next examined the substrate scope of C4-to-C5 aryl group transposition (Fig. As for the 4-arylnicotinates, the process was compatible with aryl groups having a polar π-electron-withdrawing group such as amino carbonyl (3 and 4), methoxy carbonyl (5 and 6), lactone (7) and cyano (8 and 9) moieties as well as chloro, fluoro and trifluoromethyl aryl groups (10–12). Similarly, the method was amenable to transposing pyridyl and pyrimidyl groups (13–15). Although the transposition of a biphenyl group (16) is viable, the processes with phenyl and 2-naphthyl groups (17 and 18) ended up with a moderate efficiency, and no transposition of a 4-methoxyphenyl group was observed (Supplementary Fig. The synthesis of 5-arylnicotinonitriles (19 and 20) could also be achieved efficiently. As for the trisubstituted pyridine series, 5,6-diarylnicotinate and -nicotinonitrile (21 and 22) as well as 5-aryl-6-methylnicotinate 23 were found to be viable targets. Interestingly, the aryl group transposition in the synthesis of 5-aryl-6-cyclopropylnicotinate 24 results in partial opening of the cyclopropyl ring to afford 5-aryl-6-propylnicotinate 24′, indicating the involvement of dearomatized C4/C6-radical intermediates such as III5, as depicted in Fig. Meanwhile, a butenyl group at C6 could be kept intact during the aryl group transposition for the formation of 25. The substrates bearing a carbohydrate moiety (26) and bezafibrate analogue (27) were well tolerated during the aryl group transposition, showcasing a potential utility of the present catalytic method for late-stage functionalization. aThe reaction conditions: substrates (0.2 mmol), 3DPAFIPN (2 mol%), HCO2K (5 equiv. ), DMSO (4 ml), Ar atmosphere, irradiation with 427-nm light (Kessil lamp, 40 W) with fan cooling. b10-Phenylphenothiazine was used instead of 3DPAFIPN under irradiation with 390-nm light (40 W). cIrradiation with 427-nm light (Kessil lamp, 40 W × 2). d2,4,6-Tris(diphenylamino)-3,5-difluorobenzonitrile was used instead of 3DPAFIPN. eLi2CO3 wase used instead of Cs2CO3. EWG, electron-withdrawing groups. The elementary steps for the C4-to-C5 aryl group transposition from 1 to 2 are depicted in Fig. The process should be initiated by the first SET reduction and protonation of 1 to form dihydropyridine radical A (red-1), followed by the second SET reduction and protonation to form 1,4-dihydropyridine B (red-2) (Fig. This electrochemical-chemical-electrochemical-chemical (ECEC) mechanism could be characterized by the cyclic voltammetry of pyridine 1 in DMSO. In the absence of HFIP, the cyclic voltammogram of pyridine 1 showed the EE mechanism with two reversible reduction waves with E1/2 at –1.72 V and –1.97 V versus standard calomel electrode (SCE) (Supplementary Fig. Conversely, in the presence of HFIP as the proton donor, the cyclic voltammogram of 1 showed one irreversible reduction wave observed at Ep/2 = –1.73 V (Supplementary Fig. 4), indicating that the second SET reduction occurs at a potential equal to or less negative than that of the first SET reduction. Upon the formation of 1,4-dihydropyridine B, the process proceeds to the SET oxidation and aryl transposition phase (Fig. A sequence of SET oxidation and deprotonation of B (ox-1) forms 4-aryl azacyclohexadienyl radical C4, which is followed by C4-to-C5 aryl group transposition via spirocyclic cyclohexadienyl radical D45 to generate 5-aryl azacyclohexadienyl radical C5. Finally, deprotonation and SET oxidation of C5 (ox-2) completes the overall redox-neutral process to afford 2. The cyclic voltammogram of 2 showed an irreversible reduction wave with Ep/2 at –1.95 V versus SCE (Supplementary Fig. 5) in the absence of HFIP and at –1.80 V versus SCE (Supplementary Fig. S6) in the presence of HFIP in DMSO, both of which are more negative than those of 1. a, Elementary steps for aryl group transposition from 1 to 2. b, Transformation of 3DPAFIPN to chromophores (28, 29 and 29′) and their photocatalytic reactivity. c, A proposed reaction mechanism. a, Calculated energy diagram for all the possible aryl group transpositions starting with C4. All calculations were carried out at the D3-M06-2X/cc-pVTZ level with solvation by DMSO estimated using a conductor-like polarized continuum model (CPCM). b, Reactivity of 2- and 6-arylnicotinates 30 and 31. c, Substituent effect to the activation energy towards TS145-R. TS, transition state. Interestingly, monitoring of the reaction progress (Supplementary Fig. 24) indicated that the first hour of the reaction was an unproductive induction period, during which the photocatalyst, 3DPAFIPN, was converted into new chromophores38. We characterized three species, and their structures were identified as the carbazole 28 formed via defluorinative cyclization (45% assay yield based on 1H NMR)39, and the methyl-substituted carbazoles 29 and 29′ formed via decyanomethylation of 3DPAFIPN40 followed by defluorinative cyclization (8% assay yield each) (Fig. The evaluation of their photocatalytic reactivity for the aryl group transposition of 1 revealed that these chromophores were found to show reasonable photocatalytic reactivity without unproductive induction period, and after 7 h irradiation, a large portion of them (>75%) could be recovered (Fig. 3b(ii) and Supplementary Figs. Based on these observations as well as their electrochemical potentials in both ground and excited states (Fig. 3b(ii) and Supplementary Section 4.2.10), the present aryl group transposition is probably facilitated by 28, 29 and 29′ generated in situ from 3DPAFIPN. The Stern–Volmer experiments showed that both pyridine 1 and HCO2K quenched the excited 28 and 29′ at similar rates (Supplementary Figs. 45 and 51), whereas the excited 29 was quenched by 1, but not by HCO2K (Supplementary Fig. The oxidative quenching of the excited 28, 29 and 29′ by pyridine 1 generates the corresponding radical cations 28•+, 29•+ and 29'•+, respectively, which are capable of oxidizing formate (E1/2 = +0.93 V versus SCE)41 (Fig. The resulting formyloxy radical undergoes hydrogen atom transfer (HAT) from another molecule of HCO2K, generating CO2 radical anion [CO2•−] and formic acid in an irreversible fashion41,42. Similarly, the reductive quenching of the excited 28 and 29′ by HCO2K also leads to the formation of CO2•− (Fig. CO2 radical anion is a strong reductant (E1/2 CO2/CO2•− = –2.3 V versus SCE in DMSO)43, capable of initiating dearomatization of 1 to B via red-1 and red-2 processes and/or SET reduction of 28, 29 and 29′ to form the corresponding radical anion 28•− (E1/2 28/28•− = –1.63 V versus SCE), 29•− (E1/2 29/29•− = –2.21 V versus SCE) and 29′•− (E1/2 29′/29′•− = –1.88 V versus SCE), respectively, which also mediate red-1 and red-2 processes (Fig. In the steady state of the photoredox catalytic cycle, the SET reduction and oxidation phases are synergistically catalysed by the 28*/28•−, 29*/29•− and 29′*/29′•− redox couples, while being complementarily repaired by the CO2 radical anion (Fig. To elucidate the origin of the observed C4-to-C5 transposition selectivity against other transposition permutations, we computed all the potential transposition pathways consisting of five azacyclohexadienyl radicals C2–C6 derived from 1 based on density functional theory calculations44,45 (Fig. As for the C4-to-C5 aryl group transposition, the first step is intramolecular radical addition of the 4-aryl azacyclohexadienyl radical C4 to the aryl group from the C5 side, requiring an activation energy (ΔG‡) of 25.2 kcal mol−1 to form a spirocyclic cyclohexadienyl radical D45 via TS145. Ensuing rearomative migration of the aryl group via TS245 affords the 5-aryl azacyclohexadienyl radical C5, which lies only 0.3 kcal mol−1 below C4. Furthermore, the C5-to-C6 transposition of C5 might also be viable to generate 6-aryl azacyclohexadienyl radical C6. Based on the activation energies, C4, C5 and C6 are presumed to exist in equilibrium via reversible aryl group transpositions. Indeed, the reaction of 6-arylnicotinate 30 under the present reaction conditions gave 2 in 41% yield through C6-to-C5 aryl group transposition (Fig. The cyclic voltammogram of 30 showed one irreversible reduction wave at Ep/2 = –1.52 V versus SCE (in the presence of HFIP in DMSO; Supplementary Fig. Taken together with the reduction potentials of C4-aryl 1 (Ep/2 = –1.73 V) and C5-aryl 2 (Ep/2 = –1.80 V) as well as the lowest energy barrier estimated for the C4-to-C5 aryl group transposition, the selective formation of C5-aryl 2 from the equilibrium between C4, C5 and C6 could be validated. Spin density analysis of C4 indicated that it has similar magnitude of magnetic moment of unpaired electron at the C3 and C5 positions. Thus, C4 could potentially induce C4-to-C3 aryl group transposition in addition to the C4-to-C5 pathway. The C4-to-C3 transposition was predicted to proceed via an uphill process, forming the corresponding intermediate D34, which features a quaternary carbon centre. Given that D34 lies 3.5 kcal mol−1 above D45, C4 would be interconverted into C5 more favourably than C3, and thus, the presence of C3 and its further aryl group transposition to C2 might be unlikely in the present process. It turned out that 2-arylnicotinate 31 (Ep/2 = –1.76 V versus SCE) did not undergo aryl group transposition, probably due to the high energy barrier estimated for C2-to-C3 aryl group transposition (Fig. To account for the migratory aptitude of the aryl groups experienced in Fig. 3b, analogous computations using the derivatives of C4-R bearing different C4-aryl groups (R) were conducted. We found that the aryl groups having π-accepting substituents give lower activation energy towards TS145-R, and these activation energies correlated with observed experimental yields (Fig. Given that TS145-R is expected to be a late transition state due to the endergonic nature of the reaction46, the stability of D45-R could affect the activation energy more than that of C4-R. We speculated that π-accepting substituents on the aryl group could stabilize the singly occupied molecular orbital of D45-R, resulting in the lower activation energy. While exploring a different transposition permutation, we found that 2,4-diarylpyridine 32, which has a more electron-rich aryl group at the C2 position than that at the C4, underwent C4-to-C3 aryl group transposition to form 2,3-diarylpyridine 33 as a sole product (Fig. Having a 4-methoxyphenyl group at the C2, electron-deficient aryl groups at the C4 exhibited good migratory aptitude towards the C3 (33–39). A phenyl group could also be transposed efficiently under this modality, providing 40 in 61% yield. Installation of methyl and cyclopropyl groups at the C5 position did not influence the transposition, providing trisubstituted pyridines 41 and 42, respectively, in good yields. As for the C2-aryl groups, the method was found to be compatible with alkoxyphenyl groups (43 and 44), tolyl (45), phenyl (46), halophenyl (47 and 48) and five-membered heteroaryl groups (49–52). We observed that 2,4-diphenylpyridine, which has equivalent electronic properties at the C2 and C4 positions, preferentially undergoes C4-to-C3 transposition to afford 2,3-diphenylpyridine (53). aThe reaction conditions: substrates (0.2 mmol), 3DPAFIPN (2 mol%), HCO2K (5 equiv. ), DMSO (4 ml), Ar atmosphere, irradiation with 427-nm light (Kessil lamp, 40 W) with fan cooling. bKessil lamp, 40 W × 2. c3DPA2FBN was used instead of 3DPAFIPN. d3DPA2FBN (4 mol%) was used instead of 3DPAFIPN. To elucidate the mechanistic origin on the C4-to-C3 selectivity, which transposes the C4 aryl group to the sterically more hindered C3 side, all the possible pathways from azacyclohexadienyl radical E4 derived from pyridine 31 were computed (Fig. Although the unpaired electron of E4 is equally distributed at the N1, C3 and C5 positions based on the spin density analysis, the radical addition to the C4-aryl group from the C3 side to form F34 was calculated to be 2.7 kcal mol−1 more energetically favourable than that from the C5 side to form F45. This is probably due to more extended π-conjugated system in the 3,4-dihydropyridine moiety of F34 than the 4,5-dihydropyridine moiety of F45, resulting in a lower activation energy towards the late transition state TS′134 than that towards TS′145 (Supplementary Fig. The spirocyclic cyclohexadienyl radical F34 could facilitate the aryl group transposition to the C3 position via TS′234 to form E3, which lies 5.7 kcal mol−1 lower in energy than E4. The cyclic voltammetry of 32 and 33 in DMSO in the presence of HFIP indicated that 33 has more negative reduction potential (Ep/2 = –2.1 V versus SCE) than 32 (Ep/2 = –1.82 V versus SCE). These data were consistent with the experimental observation that the aryl group transposition starting from 32 would preferentially accumulate C3-arylpyridine 33. To elucidate the connectivity of other regioisomeric congeners through aryl group transpositions, we conducted the reactions of 2,5-diarylpyridine 54 and 2,6-diarylpyridine 55 (Fig. Neither 54 nor 55 showed transposition reactivity; 54 underwent reduction of the carboxamide moiety to give alcohol 56 in 10% yield, whereas 55 simply underwent the Birch-type dearomatization to form tetrahydropyridine 57 in 15% yield. a, Calculated energy diagram for all the possible aryl group transpositions. All calculations were carried out at the D3-M06-2X/cc-pVTZ level with solvation by DMSO estimated using a CPCM. b, Reactivity of 2,5-diarylpyridine 54 c, Reactivity of 2,6-diarylpyridine 55. Interestingly, we observed a reversal of the transposition pathway when the relative electronics of the C2 and C4 aryl substituents were inverted (Fig. An electron-deficient aryl group at C2 preferentially migrated to the C3 position, affording 3,4-diarylpyridines (58–62) in moderate yields under the slightly modified reaction conditions. These complementary outcomes highlight the decisive role of substituent electronics in governing the course of transposition and demonstrate that the transposition selectivity is tunable. aThe reaction conditions: substrates (0.2 mmol), 3DPAFIPN (2 mol%), K2Sx (10 mol% per S), HCO2K (5 equiv. ), MeOH (20 equiv. ), DMSO (4 ml), Ar atmosphere, irradiation with 427-nm light (Kessil lamp, 40 W) with fan cooling. bThe reaction was conducted using HFIP (5 equiv.) instead of MeOH and in the absence of K2Sx. We have developed a peripheral editing strategy of pyridines under photoredox catalysis that enables the transposition of a preinstalled aryl group at the para- (C4) or ortho- (C2/C6) position to the meta- (C3/C5) position, offering a solution to the long-standing synthetic challenge of selectively introducing an aryl group to the meta position of pyridines. Our mechanistic analyses revealed that the meta-selectivity on the transposition is driven by the intricate interplay of various factors, including the electronic and steric effects of the substituents and their positions as well as the structures of the transient radical intermediates and the reduction potential of the pyridine substrates. Based on the operational simplicity and broad functional group compatibility as well as the mechanistic rationale showcased in this work, we view the present protocol to be useful in synthetic endeavours towards the functionalization of various heteroarenes. To a 25-ml sealed tube with a magnetic stir bar was added the pyridine substrate (0.20 mmol), HCO2K (5 equiv. ), 3DPAFIPN (2 mol%) and DMSO (2–4 ml). The solution was sparged with Ar through a needle for 15 min. To the degassed reaction mixture was added HFIP (5 equiv.). After the reaction mixture was sealed, it was stirred under irradiation with visible light (Kessil lamp, λmax = 427 nm, 40 W) with fan cooling for the time specified in the Supplementary Information. The reaction mixture was then quenched with pH 10 buffer (3 ml), diluted with H2O (30 ml) and extracted with EtOAc (3 × 30 ml). The combined organic layers were washed with water (30 ml) and brine (30 ml), dried over Na2SO4, filtered and concentrated in vacuo to afford a crude residue. The crude residue was then purified by flash column chromatography (silica gel) to afford the desired product. For complete experimental details, including procedures and full characterization of all new compounds, see the Supplementary Information. All data are available in the Supplementary Information. X-ray data are available free of charge from the Cambridge Crystallographic Data Centre under accession number CCDC 2494400 for 3DPAFMeCN and 2494404 for 29. Shearer, J., Castro, J. L., Lawson, A. D. G., MacCoss, M. & Taylor, R. D. Rings in clinical trials and drugs: present and future. Google Scholar Marshall, C. M., Federice, J. G., Bell, C. N., Cox, P. B. & Njardarson, J. T. An update on the nitrogen heterocycle compositions and properties of U.S. FDA-approved pharmaceuticals (2013−2023). Google Scholar Castellino, N. J., Montgomerym, A. P., Danon, J. J. & Kassiou, M. Late-stage functionalization for improving drug-like molecular properties. Google Scholar Börgel, J. & Ritter, T. Late-stage functionalization. Google Scholar Mohite, S. B. et al. Advances in pyridine C−H functionalizations: beyond C2 selectivity. Google Scholar Josephitis, C. M., Nguyen, H. M. H. & McNally, A. Late-stage C−H FUNCTIONALIZATION OF AZINes. Google Scholar Murakami, K., Yamada, S., Kaneda, T. & Itami, K. C-H functionalization of azines. Google Scholar Cao, H., Cheng, Q. & Studer, A. meta-Selective C−H functionalization of pyridines. Google Scholar Balkenhohl, M. & Knochel, P. Regioselective C–H activation of substituted pyridines and other azines using Mg- and Zn-TMP-bases. Google Scholar Zhang, T. et al. A directive Ni catalyst overrides conventional site selectivity in pyridine C–H alkenylation. Google Scholar Yang, L., Uemura, N. & Nakao, Y. meta-Selective C−H borylation of benzamides and pyridines by an iridium−Lewis acid bifunctional catalyst. Google Scholar Salamanca, V., Toledo, A. & Albéniz, A. C. 2,2′-Bipyridin]-6(1H)-one, a truly cooperating ligand in the palladium-mediated C–H activation step: experimental evidence in the direct C-3 arylation of pyridine. Google Scholar Cheng, C. & Hartwig, J. F. Iridium-catalyzed silylation of aryl C−H bonds. Google Scholar Larsen, M. A. & Hartwig, J. F. Iridium-catalyzed C−H borylation of heteroarenes: scope, regioselectivity, application to late-stage functionalization, and mechanism. Google Scholar Ye, M. et al. Ligand-promoted C3-selective arylation of pyridines with Pd catalysts: gram-scale synthesis of (±)-preclamol. Google Scholar Ye, M., Gao, G.-L. & Yu, J.-Q. Ligand-promoted C-3 selective C–H olefination of pyridines with Pd catalysts. Google Scholar Ir-catalyzed highly selective addition of pyridyl C–H bonds to aldehydes promoted by triethylsilane. Google Scholar Liu, Z. et al. Asymmetric C3-allylation of pyridines. Zhou, X.-Y., Zhang, M., Liu, Z., He, J.-H. & Wang, X.-C. C3-selective trifluoromethylthiolation and difluoromethylthiolation of pyridines and pyridine drugs via dihydropyridine intermediates. Liu, Z. et al. Borane-catalyzed C3-alkylation of pyridines with imines, aldehydes, or ketones as electrophiles. Wübbolt, S. & Oestreich, M. Catalytic electrophilic C H silylation of pyridines enabled by temporary dearomatization. Grozavu, A., Hepburn, H. B., Bailey, E. P., Lindsay-Scott, P. J. & Donohoe, T. J. Rhodium catalysed C-3/5 methylation of pyridines using temporary dearomatisation. Cao, H., Cheng, Q. Radical and ionic meta-C–H functionalization of pyridines, quinolines, and isoquinolines. Haring, M., Balanna, K., Cheng, Q., Lammert, J. Formal meta-C–H-fluorination of pyridines and isoquinolines through dearomatized oxazinopyridine intermediates. Xu, P., Wang, Z., Guo, S.-M. & Studer, A. Introduction of the difluoromethyl group at the meta- or para-position of pyridines through regioselectivity switch. Qin, S. et al. Electrochemical meta-C–H sulfonylation of pyridines with nucleophilic sulfinates. Guo, S.-M., Xu, P. & Studer, A. meta-Selective copper-catalyzed C–H arylation of pyridines and isoquinolines through dearomatized intermediates. Bhattacharya, D. & Studer, A. meta-Hydroxylation of pyridines, quinolines, and isoquinolines using dearomatized intermediates. Balanna, K. & Studer, A. meta-Nitration of pyridines and quinolines through oxazino azines. Boyle, B. T., Levy, J. N., de Lescure, L., Paton, R. S. & McNally, A. Halogenation of the 3-position of pyridines through Zincke imine intermediates. Li, S. et al. C3 selective chalcogenation and uorination of pyridine using classic Zincke imine intermediates. Wang, H. & Greaney, M. F. Regiodivergent arylation of pyridines via Zincke intermediates. Ma, C., Lindsley, C. W. & Yu, B. Rational molecular editing: a new paradigm in drug discovery. Li, Y., Wang, D., Zhang, L. & Luo, S. Redox property of enamines. Allen, A. R., Note, E. A. & Stephenson, C. R. Aryl transfer strategies mediated by photoinduced electron transfer. Chen, Z.-M., Zhang, X.-M. & Tu, Y.-Q. Radical aryl migration reactions and synthetic applications. Speckmeier, E., Fischer, T. G. & Zeitler, K. A toolbox approach to construct broadly applicable metal-free catalysts for photoredox chemistry: deliberate tuning of redox potentials and importance of halogens in donor–acceptor cyanoarenes. Ko, H. M., Lee, C. W. & Kwon, M. S. Degradation behavior of donor–acceptor cyanoarenes-based organo photocatalysts. Pinosa, E. et al. Light-induced access to carbazole-1,3-dicarbonitrile: a thermally activated delayed fluorescent (TADF) photocatalyst for cobalt-mediated allylations. Dang, Y. et al. Alkene carboxy-alkylation via CO2•−. Huang, Y. et al. Photoredox activation of formate salts: hydrocarboxylation of alkenes via carboxyl group transfer. ACS Catal. Chmiel, A. F., Williams, O. P., Chernowsky, C. P., Yeung, C. S. & Wickens, Z. K. Non-innocent radical ion intermediates in photoredox catalysis: parallel reduction modes enable coupling of diverse aryl chlorides. Koppenol, W. H. & Rush, J. D. Reduction potential of the carbon dioxide/carbon dioxide radical anion: a comparison with other C1 radicals. Zhao, Y. & Truhlar, D. G. The M06 suite of density functionals for main group thermochemistry, thermochemical kinetics, noncovalent interactions, excited states, and transition elements: two new functionals and systematic testing of four M06-class functionals and 12 other functionals. Grimme, S., Antony, J., Ehrlich, S. & Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H–Pu. Hammond, G. S. A correlation of reaction rates. Download references Financial support was provided by Nanyang Technological University (NTU Singapore), the Singapore Ministry of Education (Academic Research Fund Tier 2: MOE-T2EP10122-0007) and the National Research Foundation, Prime Minister's Office, Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) programme through Decarbonisation Grand Challenge Project SM3: Sustainable Manufacture of Molecules and Materials for S.C. We acknowledge Y. Li (NTU Singapore) for assistance in the X-ray crystallographic analysis. We thank D. A. Pratt (University of Ottawa) for helpful discussion. School of Chemistry, Chemical Engineering and Biotechnology, Nanyang Technological University, Singapore, Republic of Singapore Eugene Yew Kun Tan, Tian-Yu Peng, Taku Wakabayashi & Shunsuke Chiba Cambridge Centre for Advanced Research and Education in Singapore, Singapore, Republic of Singapore Taku Wakabayashi & Shunsuke Chiba Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Conceptualization and methodology, E.Y.K.T., T.P., T.W. ; writing – original draft, E.Y.K.T., T.P., T.W. and S.C.; writing – review and editing, E.Y.K.T., T.P., T.W. and S.C.; funding acquisition, resources and supervision, S.C. Correspondence to Shunsuke Chiba. The authors declare no competing interests. Nature Synthesis thanks the anonymous reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Thomas West, in collaboration with the Nature Synthesis team. Peer reviewer reports are available. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Figs. 1–53, Supplementary Tables 1–4, mechanistic studies, experimental protocols for synthesis of starting materials and products section, crystallographic analysis and computational studies. Crystallographic data for 3DPAFMeCN. Crystallographic data for 29. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. Reprints and permissions Tan, E.Y.K., Peng, TY., Wakabayashi, T. et al. Peripheral aryl group transposition on pyridines using photoredox catalysis. Download citation Received: 10 November 2025 Accepted: 10 February 2026 Published: 04 March 2026 Version of record: 04 March 2026 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Nature Synthesis ISSN 2731-0582 (online) © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Scientists at UCLA Health and UC San Francisco have discovered why certain brain cells are better equipped than others to withstand the buildup of tau, a toxic protein closely linked to Alzheimer's disease and related dementias. The findings point to biological differences that may help explain why some neurons survive longer, and they could open the door to new treatment strategies. Tau is the most common protein known to aggregate in neurodegenerative disorders, yet scientists have long puzzled over why some neurons are more vulnerable than others. This complex labels tau with molecular tags that direct it toward the cell's waste disposal system for breakdown and removal. The results suggest that boosting this natural cleanup pathway could form the basis of new therapies for neurodegenerative diseases, which affect millions of Americans and still lack effective treatments. "We wanted to understand why some neurons are vulnerable to tau accumulation while others are more resilient," said study first author Dr. Avi Samelson, assistant professor of Neurology at UCLA Health, who conducted the research while at UCSF. "By systematically screening nearly every gene in the human genome, we found both expected pathways and completely unexpected ones that control tau levels in neurons." In experiments using neurons derived from human stem cells, the researchers switched off individual genes to see how each one influenced toxic tau clumping. When the team examined brain tissue from people with Alzheimer's disease, they found that neurons with higher levels of CRL5SOCS4 components were more likely to survive despite tau accumulation. The study also uncovered an unexpected link between mitochondrial problems and tau toxicity. When the researchers disrupted these energy-producing structures, cells began producing a specific tau fragment measuring about 25 kilodaltons. This fragment closely matches a biomarker detected in the blood and spinal fluid of Alzheimer's patients, known as NTA-tau. "This tau fragment appears to be generated when cells experience oxidative stress, which is common in aging and neurodegeneration," Samelson said. "What makes this study particularly valuable is that we used human neurons carrying an actual disease-causing mutation," Samelson said. "These cells naturally have differences in tau processing, giving us confidence that the mechanisms we identified are relevant to human disease." Beyond CRL5SOCS4, the large-scale genetic screen revealed additional biological pathways not previously tied to tau regulation. These include a protein modification process known as UFMylation and enzymes that help build membrane anchors within cells. Materials provided by University of California - Los Angeles Health Sciences. Common Arthritis Drug Found To Lower Blood Pressure and Boost Heart Health Cosmic Voids Aren't Empty – They're Full of Something Far Stranger Stay informed with ScienceDaily's free email newsletter, updated daily and weekly. Keep up to date with the latest news from ScienceDaily via social networks: Tell us what you think of ScienceDaily -- we welcome both positive and negative comments.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). Adele L. Marston is in the Centre for Cell Biology, School of Biological Sciences, University of Edinburgh, Edinburgh EH9 3BF, UK. When a cell divides, its chromosomes are copied and then pulled apart so that each daughter cell receives a complete set of genetic material. These sites are among the most rapidly evolving parts of the genome, and have diverged greatly across species. Writing in Nature, Haase et al.1 and Helsen et al.2 reveal how new types of centromere can arise while maintaining the connections required for the genome to be inherited accurately. Prices may be subject to local taxes which are calculated during checkout Read the paper: Ancient co-option of LTR retrotransposons as yeast centromeres Complete ape genomes offer a close-up view of human evolution Mammalian cells repress random DNA that yeast transcribes Largest Silurian fish illuminates the origin of osteichthyan characters Pokémon turns 30 — how the fictional pocket monsters shaped science AI can write genomes — how long until it creates synthetic life? Genome modelling and design across all domains of life with Evo 2 Cell-free chromatin state tracing reveals disease origin and therapy responses Lipid metabolism drives dietary effects on T cell ferroptosis and immunity Clonal-aggregative multicellularity tuned by salinity in a choanoflagellate CLCC1 promotes hepatic neutral lipid flux and nuclear pore complex assembly The Medical Faculty Mannheim of Heidelberg University offers the position of a Full professorship (W3) for “Transfusion Medicine and Immunology” ... Job Title: Associate or Senior Editor Nature Progress Brain Health Location: Shanghai, Beijing or Pune, Hybrid Working Model Application Dead... Shanghai, Beijing or Pune, Hybrid Working Model Job Title: Associate or Senior Editor, Nature Progress Oncology Location: Shanghai, Beijing or Pune, Hybrid Working Model Application Deadlin... Title: Associate or Senior Editor, Communications Sustainability Location: Shanghai, Beijing, Nanjing, Pune or New Delhi Application Deadline: Marc... Shanghai, Beijing, Nanjing, Pune or New Delhi Read the paper: Ancient co-option of LTR retrotransposons as yeast centromeres Complete ape genomes offer a close-up view of human evolution Mammalian cells repress random DNA that yeast transcribes An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.