Cancer cells activate the integrated stress response (ISR) to adapt to stress and resist therapy1. ISR signals converge on activating transcription factor 4 (ATF4), which controls cell-intrinsic transcriptional programs that are involved in metabolic adaptation, survival and growth2,3. However, whether the ISR-ATF4 axis influences anti-tumour immune responses remains mostly unknown. Here we show that loss of ATF4 decreases tumour progression considerably in immunocompetent mice, but not in immunocompromised ones, by enhancing T cell-dependent anti-cancer immune responses. An unbiased genetic screen of ATF4-regulated genes identifies lipocalin 2 (LCN2) as the principal ATF4-dependent effector that impairs anti-tumour immunity by favouring infiltration with immunosuppressive interstitial macrophages. Furthermore, we find that LCN2 promotes T cell exclusion and immune evasion in preclinical mouse models, and correlates with decreased T cell infiltration in patients with lung and pancreatic adenocarcinomas. Anti-LCN2 antibodies promote robust anti-tumour T cell responses in mouse models of aggressive solid tumours. Our study shows that the ATF4-LCN2 axis has a cell-extrinsic role in suppressing anti-cancer immunity, and could pave the way for an immunotherapy approach that targets LCN2.

The prevalence of metabolic-dysfunction-associated steatohepatitis (MASH) is rising globally, yet effective treatments remain limited1. Here we found that systemic or hepatocyte-specific ablation of the gene encoding glycoprotein non-metastatic melanoma protein B (Gpnmb)-a top upregulated gene in MASH-protected mice from diet-induced MASH. Notably, MASH progression was driven specifically by the secreted GPNMB ectodomain (G-ECD), rather than full-length GPNMB. Serum G-ECD levels showed a strong positive correlation with MASH severity in human patients. Using an unbiased screen of a cell-surface-displayed transmembrane protein library, we identified related to receptor tyrosine kinase (RYK) as a functional receptor for G-ECD. Hepatocyte-specific Ryk ablation protected mice against MASH and abolished the pathogenic effects of G-ECD. Mechanistically, G-ECD binding to RYK activated ERK1/2 signaling, resulting in transcriptional activation of PPARγ-CD36 and SREBP1C pathways that promote hepatic lipid uptake and lipogenesis. Multiple therapeutic strategies targeting the GPNMB-RYK axis-including vaccination, short hairpin RNA, neutralizing antibody and N-acetylgalactosamine small interfering RNA-effectively prevented and treated MASH in preclinical models. Our findings identify the GPNMB-RYK axis as a new pathogenic ligand-receptor pathway and a promising therapeutic target for MASH.

Telecommunication systems are evolving towards ultrawide bandwidth and low latency, supporting wired and wireless links and their non-blocking interconnection1. However, a long-standing bandwidth mismatch between fibre communication and its wireless counterpart arises from fundamental disparities in signal architectures and hardware constraints2,3, which prevent high-speed and compatible transmission across the two domains. This challenge further complicates unified system design and hinders the realization of high-throughput-density, congestion-free fibre-wireless links under wideband-access scenarios4. Here we present an ultra-wideband (UWB) integrated photonics scheme that facilitates fibre-wireless communication over a shared-bandwidth infrastructure. Built on electro-optic (EO) and optic-electro (OE) conversions featuring 3-dB operational bandwidths exceeding 250 GHz and cross-architecture adaptability, our system demonstrates unprecedented data transmission capabilities in both wired and wireless links. Using the same set of devices and powered by the proposed complex bidirectional gated recurrent unit (complex-biGRU) algorithm, ultrahigh single-lane data rates of 512 Gbps for short-reach fibre and, for the first time to the authors' knowledge, 400-Gbps high-speed wireless transmission have been achieved. Furthermore, high-density access is enabled by an all-optically assisted ultra-broadband wireless scheme. Real-time multichannel 8K video transmission is successfully demonstrated across 86 channels, seamlessly using a spectral range from 138 to 223 GHz. These findings in unified telecommunication development show the potential for the development of high-speed, densified and low-latency communication networks.

Paternal care is rare among mammals and the neural mechanisms governing its emergence are poorly understood1. We leveraged the natural paternal behaviour of African striped mice (Rhabdomys pumilio)2,3, and integrated brain-wide cFos mapping, single-nucleus RNA sequencing, virally mediated gene perturbation and environmental manipulation to dissect the neural basis of natural variation in male parenting. Here we find that socio-environmental conditions drive individual variation in male alloparenting such that postweaning social isolation increases paternal care whereas social living in higher density groups increases infanticide. This natural variation in care corresponds to neural activity in the medial preoptic area and changes in correlated activity across brain regions. Within the medial preoptic area, expression of agouti signalling protein (Agouti) in neurons is increased by group housing and is negatively associated with care, and overexpression of Agouti reduces care and enhances infanticide in previously tolerant mice. Naturalistic manipulations further reveal that Agouti integrates long-term housing conditions rather than food availability or hunger. Our findings reveal that variation in male paternal care reflects context-dependent regulation of conserved hypothalamic and melanocortin signalling mechanisms rather than the presence or absence of paternal capacity.

Organic batteries using abundant and recyclable organic electrode materials provide a sustainable and environmentally friendly alternative to commercial lithium-ion batteries1-5, which rely on resource-limited mineral-derived inorganic electrode materials6-8. However, the practical use of organic batteries has been severely hindered by the intrinsic insulation and dissolution of organic electrode materials9,10. Here we report practical organic batteries using an n-type conducting polymer cathode, poly(benzodifurandione) (PBFDO), which exhibits excellent mixed ionic and electronic transport and low solubility. The PBFDO cathode maintains its n-doped state throughout the electrochemical processes and exhibits stable and reversible redox characteristics, high electrical conductivities and significant lithium-ion diffusion coefficients, without the need for additional conductive additives. Consequently, ultrahigh-mass-loading polymer cathodes, with mass loadings up to 206 mg cm-2, are realized, delivering a high areal capacity of 42 mAh cm-2 and demonstrating robust cycling stability. Furthermore, practical 2.5 Ah lithium-organic pouch cells were fabricated, achieving an impressive energy density of 255 Wh kg-1. Notably, the conducting polymer cathode operates efficiently over a wide temperature range from -70 °C to 80 °C and demonstrates excellent flexibility and safety, marking considerable potential for applications in extreme conditions and wearable electronics.

Cytoplasmic dynein-1, a microtubule (MT)-based motor protein, requires dynactin and a coiled-coil adaptor to form the processive dynein-dynactin-adaptor (DDA) complex1,2. The roles of MTs and dynein regulator lissencephaly-1 (LIS1) in DDA assembly have remained elusive. Here we use cryo-electron microscopy to determine the structural basis of MT- and LIS1-mediated DDA assembly. We show that an adaptor-independent dynein-dynactin complex spontaneously forms on MTs with an intrinsic 2:1 stoichiometry in a highly efficient manner, driven by parallel alignment of dynein tails upon MT binding. Adaptors can wedge into and exchange within the assembled MT-bound dynein-dynactin complex; these processes are enabled by relative rotations between dynein and dynactin and facilitated by the dynein light-intermediate chains that assist the adaptor 'search' mechanism. Although LIS1 is dispensable for efficient DD(A)-MT assembly, its presence expands the conformational landscape of DD(A) assemblies on MTs. Cryo-electron microscopy reveals that LIS1 bridges dynactin p150glued and dynein in both the closed Phi-like and open prepowerstroke states, stabilizing low-MT-affinity intermediates that tether dynein molecules in proximity to MTs and prime them for subsequent DD(A) assembly through alternative pathways. These findings demonstrate the dynamic adaptability of the dynein transport machinery and the coordinated roles of MTs and LIS1 in DDA assembly.

HIV-1 integrase (IN) promotes encapsulation of viral genomic RNA into mature viral cores, and this function is a target for ongoing antiretroviral drug development efforts1-3. Here we determined the cryogenic electron microscopy (cryo-EM) structure of a primate lentiviral IN in a complex with RNA, revealing a linear filament made of IN octamer repeat units, each comprising a pair of asymmetric homotetramers. The assembly is stabilized through IN-RNA interactions involving mainly the IN C-terminal domains and RNA backbone. The spacing and orientation of the IN filament repeat units closely matched those of consecutive capsid (CA) hexamers within the mature CA lattice. Using cryo-EM images of native purified HIV-1 cores, we refined the structure of the IN filament as it propagates along the luminal side of the CA lattice. Each IN tetramer within the filament nestled in a CA hexamer, engaging closely with the major homology regions. Substitutions of residues involved in IN-CA contacts yielded eccentric virions with RNA nucleoids located outside of the cores. Collectively, our results establish the structural basis for the HIV-1 IN-RNA interaction and reveal that IN forms an RNA-binding module on the luminal side of the mature CA lattice.

The central nervous system is surrounded by three interconnected membranes referred to as the meninges, which host a diverse immune network1-3. Within the skull-interfacing dura mater are venous sinuses, large veins that are traditionally viewed as passive blood drains for the brain and skull4,5. However, these structures also constitute an important neuroimmune interface6-8. Here we used intravital microscopy to gain mechanistic insight into this interface and reveal that dural sinuses and their endothelial cells form a highly dynamic surface that continually restructures to regulate blood flow, fluid movement and immune surveillance. We show that sinuses are not passive conduits, but instead undergo RAMP1-dependent constriction and dilation mediated by smooth muscle, resembling arterial behaviour. Moreover, the superior sagittal sinus in mice is bifurcated into upper and lower chambers that contribute to intracranial pressure regulation. Both chambers are lined by specialized, highly fenestrated sinus endothelial cells (SECs) that permit movement of fluids, macromolecules and microorganisms between the sinus lumen and leukocyte-rich perisinus space. To safeguard this permeable interface, SECs dynamically open and close intercellular boundaries in a RAMP2-dependent manner. Transcranial RAMP2 antagonism impaired SEC boundary dynamics and reduced immune cell trafficking along the sinus wall during homeostasis and systemic viral infection. Disruption of SEC dynamics during infection compromised local antiviral immunity and promoted pathogen entry into the meninges. Together, these findings establish dural sinuses as dynamic venous structures that regulate fluid exchange and support immune surveillance and antiviral defence.

The brain displays the richest repertoire of post-transcriptional mechanisms regulating mRNA translation1-11. Among these, alternative splicing has been shown to drive cell-type specificity and, when disrupted, is strongly linked to neurological disorders12-17. However, genome-wide measurements of mRNA translation with isoform sensitivity at single-cell resolution have not been achieved. To address this, we deployed Surveying Ribosomal Targets by APOBEC-Mediated Profiling (Ribo-STAMP) coupled with short-read and long-read single-cell RNA sequencing in the brain18. We generated the first isoform-sensitive single-cell translatomes of the mouse hippocampus at postnatal day 25, discovering cell-type-specific translation of 3,857 alternative transcripts across 1,641 genes and identifying isoforms of the same genes undergoing differential translation within and across 8 different cell types. We defined high and low translational states in CA1 and CA3 neurons, with synaptic and metabolic genes enriched in high states. We found that CA3 exhibited higher basal translation compared with CA1, as confirmed by metabolic labelling of newly synthesized proteins and immunohistochemistry of translational machinery components. This accessible platform will expand our understanding of how cell-type-specific and isoform-specific translation drives brain physiology and disease.

Colloidal perovskite nanocrystal (PeNC) has long been synthesized using the hot-injection method and room-temperature ligand-assisted reprecipitation as the prominent techniques1,2. However, both methods have challenges for industrial-scale production3-5: the hot-injection method requires high temperatures, an inert gas environment and rapid cooling, which raise safety concerns, whereas ligand-assisted reprecipitation can exhibit limited productivity on scale-up. Here we present a cold-injection method based on pseudo-emulsion, enabling scalable synthesis of PeNCs with near-unity photoluminescence quantum yield (PLQY, ~100%) and enhanced stability by injecting precursor solution below 4 °C. In the cold-injection method, PeNCs grow through the assembly of fully coordinated plumbates out of the pseudo-emulsion with the assistance of a demulsifier. We discovered that slow assembly of polybromide plumbates, assisted by cold temperature, is essential for defect suppression, resulting in reproducible, stable and pure-green-emitting PeNCs with near-unity PLQY. Furthermore, this method enables efficient large-scale production, achieving 20-l-scale synthesis with remarkable batch weight while maintaining near-unity PLQY. Our findings represent a substantial advancement in synthesis of high-quality PeNCs, offering potential for broad applications in display and lighting industries.

Cytopathology, often abbreviated as cytology, has a central role in the early detection of cancer, such as cervical, lung and bladder cancers, owing to its speed, simplicity and minimally invasive nature1-9. However, its effectiveness is limited by variability in diagnostic accuracy stemming from subjective visual interpretation10-21. Although many artificial intelligence (AI)-powered systems have been proposed to improve consistency22-26, none have achieved fully autonomous, clinical-grade performance. Existing approaches serve as assistive tools and still rely on human oversight for interpretation and decision-making22-26. Here we present a clinical-grade autonomous cytopathology pipeline that combines high-resolution, real-time optical whole-slide tomography with edge computing to deliver end-to-end automation. The system achieves practical performance in imaging speed, quality and data volume, with localized data compression enabling streamlined storage and accelerated AI-driven analysis. In addition to supporting cell-level classification, the platform enables flow cytometry-like, population-wide morphological profiling for comprehensive interpretation of cellular distributions and patterns. A vision transformer achieved area under the receiver operating characteristic (ROC) curve (AUC) values exceeding 0.99 at the single-cell level for detecting low-grade squamous intraepithelial lesions (LSILs), high-grade squamous intraepithelial lesions (HSILs) and adenocarcinoma. In a multicentre evaluation of 1,124 cervical liquid-based cytology samples across four centres, the AI model achieved slide-level AUC values of 0.86-0.91 for LSIL+ and 0.89-0.97 for HSIL+, with LSIL counts correlating strongly with human papillomavirus positivity and HSIL counts scaling with diagnostic severity. The system enables autonomous triage cytology, offering a foundation for routine, scalable and objective diagnostics.

The severity of malaria varies substantially between individuals, but the mechanisms that underlie these differences remain unclear. Because erythrocytes have a key role in malaria biology, genetic variants associated with the development of these cells could inform the mechanisms that determine disease severity. Here we investigate the mechanistic basis of the association of the variant rs112233623-T with erythrocyte properties, and examine its role in modulating malaria severity. This variant is associated with increased levels of haemoglobin A2, increased erythrocyte size and reduced erythrocyte number1,2. It is found in an erythroid enhancer of CCND3, which encodes cyclin D3-a cell-division activator that enhances the pentose phosphate pathway and thereby helps to counteract reactive oxygen species (ROS)3. We show that rs112233623-T disrupts a binding site for the transcription factor SMAD3, weakens enhancer activity and, in erythrocyte precursors (erythroblasts), is associated with reduced CCND3 expression and inhibition of the G1-S cell-cycle transition, concomitant with a reduction in the number of erythrocytes and an increase in their size. Using population genetic methods, we observe signatures of positive selection for rs112233623-T in the genetic history of Sardinia, a region in which malaria was once prevalent. Furthermore, we show that parasite growth is impaired in cultured Plasmodium falciparum-infected erythrocytes from rs112233623-T carriers, and that this impairment correlates with ROS levels. This mirrors our observations in erythrocytes from individuals who are deficient in the pentose-phosphate-pathway enzyme G6PD-a trait associated with protection against malaria in some settings-and highlights a common ROS-based mechanism of malaria resistance. Our results suggest that a reduction in CCND3 in erythroblasts constitutes a mechanism of resistance to malaria, and could enable therapeutic interventions.

Clostridioides difficile infection (CDI) is the leading cause of healthcare- and antibiotic-associated infection and has a 30% recurrence rate1-5. Previous vaccine strategies against CDI failed to reduce pathogen burden, a prerequisite for preventing C. difficile transmission and recurrence6-11. These vaccines were administered parenterally, which induced a systemic immune response, rather than a mucosal response in the colon, the site of infection. Here we compare protection and colonization burden between mucosal (rectal) and parenteral (intraperitoneal) administration routes of a multivalent, adjuvanted vaccine combining inactivated C. difficile toxins and novel surface antigens. We found that mucosal immunization, but not parenteral, clears C. difficile from the host. Unique correlates of decolonization included faecal IgG responses to vegetative surface antigens and a colonic, T helper type 17 (TH17)-skewed tissue-resident memory T cell response against spore antigen. Importantly, mucosal vaccination protected against morbidity, mortality, tissue damage and recurrence. Our results demarcate notable differences in correlates of protection and pathogen clearance between vaccine administration routes and highlight a mucosal immunization regimen that elicits sterilizing immunity against CDI.

Asgard archaea were pivotal in the origin of complex cellular life1. Heimdallarchaeia (a class within the phylum Asgardarchaeota) are inferred to be the closest relatives of eukaryotes. Limited sampling of these archaea constrains our understanding of their ecology and evolution2,3, including their role in eukaryogenesis. Here we use massive DNA sequencing of marine sediments to obtain 404 Asgardarchaeota metagenome-assembled genomes, including 136 new Heimdallarchaeia and several novel lineages. Analyses of their global distribution revealed they are widespread in marine environments, and many are enriched in variably oxygenated coastal sediments. Detailed metabolic reconstructions and structural predictions suggest that Heimdallarchaeia form metabolic guilds that are distinct from other Asgardarchaeota. These archaea encode hallmark proteins of an aerobic lifestyle, including electron transport chain complex (IV), haem biosynthesis and reactive oxygen species detoxification. Heimdallarchaeia also encode novel clades of respiratory membrane-bound hydrogenases with additional Complex I-like subunits, which potentially increase proton-motive force generation and ATP synthesis. Thus, we propose an updated Heimdallarchaeia-centric model of eukaryogenesis in which hydrogen production and aerobic respiration may have been present in the Asgard-eukaryotic ancestor. This expanded catalogue of Asgard archaeal genomic diversity suggests that bioenergetic factors influenced eukaryogenesis and constitutes a valuable resource for investigations into the origins and evolution of cellular complexity.

The three-dimensional architecture of natural habitats is a key determinant of species biodiversity, harvestable biomass and resilience to disturbance1,2. Indeed, some species, including trees, corals and oysters, alter resource availability and modify biotic and abiotic pressures through their own three-dimensional structures-thereby enhancing their own survival3,4. However, which aspects of the three-dimensional architecture of these ecosystem engineers shape ecosystem dynamics and species survival are rarely examined by empirical studies, leaving much of the broader ecological and conservation impact of ecosystem engineering underexplored4-6. Here we show that oyster reefs have combinations of geometric variables that maximize recruit survival, which is a key factor influencing oyster reef growth and persistence. Using three-dimensional habitat designs that capture the full spectrum of natural oyster reef architectures, as well as a geometric theory7 that links habitat surface area, fractal dimension and height, we show that oyster settlement and survival are greatest at particular combinations of fractal dimension and height that minimize predation. Our study provides a template for understanding optimal three-dimensional habitat configurations for habitat restoration projects that are proliferating globally, without targeting key architectural features of habitat space that maximize restoration success8,9.

Neurodevelopmental disorders that arise from de novo mutations in chromatin-remodelling genes lack targeted treatments. Snijders Blok-Campeau syndrome (SNIBCPS)1, which is caused by pathogenic variants in CHD3, manifests with intellectual disability, autistic-like behaviours and motor deficits2. Whether somatic gene correction can reverse such phenotypes in vivo remains unknown. Here we show that modelling the recurrent CHD3 variant p.R1025W in a humanized mouse model (Chd3hR1025W/+) recapitulates key features of SNIBCPS, including reduced CHD3 protein levels and abnormalities in social communication, cognition and motor coordination. We engineered a TadA-embedded adenine base editor (TeABE) and delivered it brain-wide using a dual adeno-associated virus (AAV) system and achieved efficient on-target A•T-to-G•C correction across multiple cortical and hippocampal regions with minimal bystander activity. This intervention restored CHD3 levels and ameliorated behavioural abnormalities in vivo. Furthermore, intrathecal dual AAV delivery in nonhuman primates resulted in widespread neuronal transduction and efficient TeABE reconstitution, a result that supports its translational feasibility. These findings establish in vivo base editing as a viable therapeutic approach for CHD3-related neurodevelopmental disease. More broadly, they demonstrate that precise single-base correction in the postnatal brain can restore protein dosage and function, thereby offering a framework for the treatment of monogenic neurodevelopmental disorders.

Centromeres ensure accurate chromosome segregation, yet their DNA evolves rapidly across eukaryotes leaving the origins of new centromere architectures unclear1-4. The brewer's yeast Saccharomyces cerevisiae exemplifies this long-standing puzzle. Its centromeres shifted ancestrally from large, repeat-rich, epigenetically specified forms to the compact, genetically defined 'point' centromeres1,5. How this transition occurred has remained unresolved6. Here we identify evolutionarily related 'proto-point' centromeres that provide a resolution to the evolutionary origins of point centromeres. Proto-point centromeres contain a single centromeric nucleosome positioned over an AT-rich core, accompanied by relaxed organization and sequence variability of flanking cis-elements. In two species, these proto-point centromeres lie within retrotransposon-derived repeat clusters, linking ancestral repeat-rich centromeres to genetically encoded ones. Comparative and phylogenetic analyses indicate that proto-point and point centromeres evolved in an ancestor with retrotransposon-rich centromeres. These results identify long-terminal-repeat retrotransposons, specifically Ty5 sequences, as the genetic substrate for point-centromere evolution and provide a mechanistic route by which an epigenetic centromere can become genetically specified. More broadly, they show how selfish elements can be co-opted to perform essential chromosomal functions.

Rare diseases affect more than 300 million people worldwide1-3, yet timely and accurate diagnosis remains an urgent challenge1,3-5. Patients often endure a prolonged 'diagnostic odyssey' exceeding 5 years, marked by repeated referrals, misdiagnoses and unnecessary interventions, leading to delayed treatment and substantial emotional and economic burden4,5. Here we present DeepRare-a multi-agent system for rare disease differential diagnosis decision support6-8 powered by large language models, integrating more than 40 specialized tools and up-to-date knowledge sources. DeepRare processes heterogeneous clinical inputs, including free-text descriptions, structured human phenotype ontology terms and genetic testing results to generate ranked diagnostic hypotheses with transparent reasoning linked to verifiable medical evidence. Evaluated across nine datasets from literature, case reports and clinical centres across Asia, North America and Europe spanning 14 medical specialties, DeepRare demonstrates exceptional performance on 2,919 diseases. In human-phenotype-ontology-based tasks, it achieves an average Recall@1 of 57.18%, outperforming the next best method by 23.79%; in multi-modal tests, it reaches 69.1% compared with Exomiser's 55.9% on 168 cases. Expert review achieved 95.4% agreement on its reasoning chains, confirming their validity and traceability. Our work not only advances rare disease diagnosis but also demonstrates how the latest powerful large-language-model-driven agentic systems can reshape current clinical workflows.

Feed algorithms are widely suspected to influence political attitudes. However, previous evidence from switching off the algorithm on Meta platforms found no political effects1. Here we present results from a 2023 field experiment on Elon Musk's platform X shedding light on this puzzle. We assigned active US-based users randomly to either an algorithmic or a chronological feed for 7 weeks, measuring political attitudes and online behaviour. Switching from a chronological to an algorithmic feed increased engagement and shifted political opinion towards more conservative positions, particularly regarding policy priorities, perceptions of criminal investigations into Donald Trump and views on the war in Ukraine. In contrast, switching from the algorithmic to the chronological feed had no comparable effects. Neither switching the algorithm on nor switching it off significantly affected affective polarization or self-reported partisanship. To investigate the mechanism, we analysed users' feed content and behaviour. We found that the algorithm promotes conservative content and demotes posts by traditional media. Exposure to algorithmic content leads users to follow conservative political activist accounts, which they continue to follow even after switching off the algorithm, helping explain the asymmetry in effects. These results suggest that initial exposure to X's algorithm has persistent effects on users' current political attitudes and account-following behaviour, even in the absence of a detectable effect on partisanship.

G-protein-coupled receptors (GPCRs) are important therapeutic targets and have been targeted mainly through their orthosteric site, where the endogenous agonist binds1. However, allosteric modulation has emerged as a promising and innovative strategy in the realm of GPCR drug discovery1. Here, drawing inspiration from the natural regulation of GPCRs by transmembrane proteins, we have developed GPCR exoframe modulators (GEMs), de novo designed proteins that specifically target the transmembrane domain of GPCRs. Utilizing a hallucination-like design approach, we crafted GEMs with three strategic structural prompts to achieve the desired binding modes. We selected the dopamine D1 receptor as a prototypical model and systematically investigated four GEMs. Structural studies and functional assays revealed that these GEMs bind to the transmembrane domains and function as diverse allosteric modulators, including agonist-positive allosteric modulator, negative allosteric modulator and biased allosteric modulator. The ago-PAM GEM restores the activity of various D1 receptor loss-of-function mutants, suggesting a promising therapeutic target for GPCR-related disorders. Our work introduces GEMs that target the transmembrane domain as potent agents for allosteric GPCR modulation and highlights the potential of deep learning-based approaches in the design of function-oriented membrane proteins.