Nitric oxide is upregulated in inflammatory tissues but has not been used to directly control the activity of folded proteins. Here we report a protein engineering strategy that enables selective restoration of protein function in nitric oxide-rich environments. Protein activity is temporarily suppressed by site-specific substitution of a catalytically or structurally essential glutamate residue with a synthetic amino acid whose side chain is chemically masked. Exposure to nitric oxide triggers decaging of this residue, regenerating the native glutamate and restoring protein function. Using this approach, we engineer nitric oxide-responsive variants of antibodies, enzymes, cytokines, bacterial toxins and viral capsids. In mouse models, this strategy enables inflammation-localized protein activation, selective viral gene delivery in inflamed tissues and rapid detection of intestinal inflammation using engineered probiotic biosensors. These results establish nitric oxide-triggered chemical reactivation of proteins as a generalizable method for post-translational control of protein function, with potential applications in inflammation-targeted therapeutics, gene delivery and biosensing.
Despite explosive growth in biomedical data generation, driven largely by genomics, and in computational capabilities, the probability that a candidate entering phase I ultimately reaches approval has remained stubbornly low over the past decades. This paradox points to a central bottleneck not in data generation, but in converting biological and clinical data into decisions that govern progression, redesign or termination. Here we argue that drug development should be reframed from a linear pipeline into an iterative learning system driven by continuous data feedback. We outline a data-centric framework in which high-dimensional, multimodal molecular and perturbation data, particularly single-cell and spatial readouts, are used to iteratively refine disease models, therapeutic hypotheses, molecular designs and patient stratification strategies across discovery and clinical stages. Using immunotherapies as a proof-of-concept domain, we propose that single-cell molecular readouts from therapeutic perturbations can both de-risk development and deepen mechanistic understanding of immune responses in humans. Finally, we draw parallels to reinforcement learning, in which human molecular and clinical data provide the feedback signal that updates mechanistic models and guides the design of subsequent interventions. Embracing this paradigm offers a path towards more mechanistically grounded, context-aware therapies with higher translational success.
Intracellular protein dysregulation underlies many cancers, neurodegenerative disorders and infectious diseases, yet a substantial fraction of the human proteome remains inaccessible to conventional small-molecule drugs or extracellular antibodies. Genetically encoded intracellular binders, including nanobodies, DARPins, affibodies and de novo-designed scaffolds, provide modular platforms for selective recognition, visualization and functional modulation of endogenous proteins in living cells. These binders can engage extended protein surfaces and conformational epitopes and are increasingly applied in preclinical models to inhibit oncogenic signalling, degrade pathogenic proteins, rewire cellular pathways and engineer programmable cell therapies. Here we outline the structural diversity, discovery strategies and design principles of intracellular binders and summarize their emerging applications in imaging, biosensing and targeted protein control. We further discuss recent advances in computationally assisted binder design and optimization. Finally, we examine key translational challenges, including cellular stability, delivery technologies and immunogenicity, that will shape the clinical development of intracellular binder-based therapeutics.
Artificial intelligence for medical imaging is required to be accurate and interpretable to clinicians. However, current multimodal biomedical foundation models often prioritize performance over explainability. Here we present ConceptCLIP, an explainable biomedical foundation model that achieves state-of-the-art diagnostic accuracy while delivering human-interpretable explanations across diverse imaging modalities. We curate MedConcept-23M, a large-scale dataset comprising 23 million biomedical image-text-concept triplets. Leveraging this dataset, we pretrain ConceptCLIP via joint image-text and region-concept alignment for precise and interpretable medical image analysis. Across a large-scale benchmark covering 78 datasets in 10 imaging modalities, ConceptCLIP demonstrates superior diagnostic performance while providing human-understandable explanations. In a clinician user study spanning three modalities, the concept-based explanations provided by ConceptCLIP help clinicians verify model predictions and identify potential errors. As an explainable biomedical foundation model, ConceptCLIP represents a critical milestone towards the widespread clinical adoption of AI, thereby advancing trustworthy AI in medicine.
Blood flow abnormalities in occluded or narrowed vessels limit drug delivery and contribute to recurrent vascular disease. Existing endovascular or pharmacological approaches do not directly address impaired local flow. Here we propose a miniature endovascular soft robot for active blood flow regulation and enhanced fluid exchange and drug transport in occluded vessels. The soft robot integrates a magnetic body for navigation with cilia that generate metachronal waves to drive the flow. Navigation and flow regulation are decoupled, allowing positioning at target sites followed by local flow enhancement. We investigate the flow regulation mechanism and optimize the structural parameters of the cilia carpet. We validate flow regulation in different vessel phantoms with biologically relevant conditions and in large animal studies. The system restores flow in occluded branches and improves delivery of therapeutic agents, accelerating clot dissolution with reduced residual obstruction and shorter recanalization time both in vitro and in vivo. This technology allows the local modulation of vascular flow, supporting more effective and controlled endovascular therapies.
Islet transplantation is a promising therapy for treating type 1 diabetes. Thin capsules, although offering numerous advantages over thicker ones in islet transplantation, pose persistent challenges for ensuring long-term cell protection. Here we show a biomimetic approach for cell protection using thin capsules, inspired by the biological development of the zona pellucida. A thin 20-μm hydrogel capsule is spontaneously generated on the cell surface with 100% encapsulation efficiency through aptamer-directed molecular recognition, crosslinking and hardening. The entire encapsulation process takes place droplet-free and stress-free under physiological conditions, without exposure to harsh factors or any loss of cellular viability or function. Immunocompetent diabetic mice receiving encapsulated allogeneic islets achieved normoglycaemia, with most sustaining it for more than 100 days.
Fibrosis is a widespread disease implicated in millions of deaths worldwide and impacting diverse organs, including the heart, lung, kidney, liver and synovium. Characterized by excessive extracellular matrix deposition, fibrosis leads to tissue scarring and dysfunction, ultimately resulting in organ failure. The development of faithful experimental models of fibrosis is challenged by its complexity, limiting the discovery of effective therapeutics. Human in vitro tissue models are emerging as powerful systems that capture the pathophysiology of fibrosis to uncover disease mechanisms, identify biomarkers and facilitate drug discovery. Here we define the hallmarks of fibrosis across organ systems to establish foundational design criteria necessary for modelling human fibrosis in vitro. We provide an overview of state-of-the-art technologies and recent advancements in tissue fibrosis models pointing towards potential future directions. Finally, we offer a practical toolkit for researchers with diverse expertise to implement these considerations into future tissue models.
In vivo genetic engineering of haematopoietic stem and progenitor cells (HSPCs) holds the potential to revolutionize the treatment landscape for numerous diseases. However, despite its transformative potential, it remains hindered by the difficulty in efficiently and specifically targeting quiescent human HSCs while maintaining their long-term functionality. Here, after screening 15 lipid nanoparticles (LNPs), we report an LNP that efficiently delivers reporter mRNA to human HSPCs both in ex vivo and in vivo settings when conjugated with the anti-CD34 antibody (CD34/LNPDP). Using CRISPR/Cas editing cargos, CD34/LNPDP achieves high editing efficiency in human HSPCs ex vivo. Intrafemoral administration of CD34/LNPDP in humanized mice results in efficient editing of the erythroid-specific BCL11A enhancer within human HSPCs, enabling the sustained long-term reactivation of fetal haemoglobin (HbF) expression in erythroid cells. In a humanized neutropaenia model harbouring an ELANE mutation, intrafemoral administration of CD34/LNPDP achieves robust editing, targeting exon 2 of ELANE in human HSPCs, partially restoring neutrophil development impairment under long-term observation. Collectively, CD34-targeted delivery enables in vivo HSPC modification without perturbing haematopoiesis, underscoring its suitability for clinical translation.
Hypothesis generation in biomedicine is constrained by human cognitive limitations in synthesizing insights from fragmented biomedical knowledge and multimodal data sources. Here we introduce XunZi, an AI biologist that integrates logical reasoning and multimodal data fusion to autonomously generate de novo therapeutic target hypotheses with testable mechanisms. XunZi has been trained on 24.4 million publications and 613.6 TB of multisource data spanning 21,008 human genes and 5,850 diseases, and outperforms existing methods in both accuracy and interpretability across diverse disease contexts. In Parkinson's disease (PD), where complex mechanisms and limited targets hamper therapy development, XunZi identifies aberrant activation of CHK2 and IRAK4 kinases across multiple models. Pharmacological or genetic inhibition of Chk2 rescues dopaminergic neuron loss and motor deficits in PD mice. We further demonstrate XunZi's broad versatility in diseases such as non-small-cell lung cancer. XunZi establishes a paradigm-shifting framework to translate fragmented biomedical knowledge and data into actionable therapeutics.
Legislative change must fix the European Union’s disharmonized map of early-stage medical device clinical study requirements, whereby fragmented interpretations across Member States create barriers to innovation and delay patient access to life-saving technologies.
The lack of brain penetrant and biologically stable positron emission tomography reporter systems hampers the development of neurological disease models and the monitoring of gene delivery because existing approaches depend on endogenous receptors that vary unpredictably in pathology. HaloTag, a fully exogenous protein label that forms rapid and irreversible bonds with synthetic ligands, provides a modular platform for engineering reporter probes with defined chemical properties. Here we developed a fluorine-18-labelled small-molecule HaloTag ligand optimized for brain entry and covalent retention at the reporter. The tracer showed specific binding in human cells expressing HaloTag and enabled non-invasive imaging of viral gene transfer to striatal neurons in mice, with clear detection of reporter expressing tissue and rapid clearance from surrounding regions. Optical imaging confirmed viral distribution and reporter expression, and a transgenic model expressing HaloTag fused to a postsynaptic protein demonstrated detection of physiologically expressed intraneuronal targets. This system establishes a modular platform for validating preclinical models and quantifying gene expression in the living brain.
Atmospheric water harvesting (AWH), which captures water from air, either by condensing humid air or by using sorbents that bind and release water vapour, is being explored as a decentralized source of clean water where piped supply is absent, intermittent or unsafe1-5. Here we highlight the role of AWH as a reversible vapour sorption and controlled condensation strategy to regulate humidity and water activity (the effective availability of water) within biomedical devices. AWH can buffer local hydration and generate small liquid volumes for sampling, improving the robustness of wearables, point-of-care assays and respiratory monitoring in environments where conventional humidity control is unavailable.
Diagnosing a whole-slide image is an interactive, multistage process, yet practical agentic systems that navigate fields, adjust magnification and deliver explainable diagnoses remain lacking, largely because the tacit, experience-based viewing behaviour of expert pathologists is absent from model training data. Here we introduce Pathology-CoT, a framework that converts expert viewing chain-of-thought behaviour into scalable agent supervision through three contributions. First, an artificial intelligence session recorder unobtrusively captures routine navigation in standard whole-slide image viewers and converts raw logs into standardized behavioural commands and bounding boxes. Second, a human-in-the-loop review pipeline turns artificial intelligence-drafted rationales into paired 'where to look' and 'why it matters' supervision, enabling sixfold faster labelling. Third, using these data, we built Pathology-o3, a two-stage agent that proposes regions of interest and performs behaviour-guided reasoning. On gastrointestinal lymph node metastasis detection, Pathology-o3 outperformed state-of-the-art vision-language models, showed consistent gains across multiple vision-language model backbones and maintained strong performance on an independent external validation cohort.
Diffusion tensor magnetic resonance imaging of the heart is typically performed at millimetre-scale resolution, yielding only four to five voxels across the ventricular wall, limiting the measurement of local spatial variation in cardiomyocyte organization. Here we present a submillimetre in vivo cardiac diffusion tensor imaging method achieved during free breathing that facilitates voxel-level characterization of myocardial microstructure. We introduce a phenomapping framework that combines voxelwise diffusion magnitude and anisotropy with radial and circumferential gradients of cardiomyocyte helix angle to identify distinct microstructural environments. The approach was developed in healthy volunteers and applied to patients with severe aortic valve stenosis, who had preserved cardiac function and marked myocardial thickening. Comparisons to conventional resolution imaging, together with downsampling analyses and ex vivo and histological validation, show that these voxel-scale features are less optimally detected using standard techniques. Four data-driven microstructural classes, defined by combined diffusion properties and orientation gradients, were observed in both healthy and pressure-overloaded hearts. Despite substantial hypertrophy, pressure overload was associated with preserved cardiomyocyte spatial organization. This framework supports studies of myocardial microstructural remodelling in vivo.
Homotypic targeting is the inherent ability of cells to preferentially interact with cells of the same type, a phenomenon seen in cell adhesion, tissue formation and immune responses. However, its potential remains underexploited. Here we report a strategy to substantially enhance homotypic targeting through extracellular vesicles secreted by cells. By engineering the surface of small extracellular vesicles (sEVs) with lanthanides, we amplify specific cell–sEV interactions by more than 25-fold, enabling the selective capture of sEVs by cells of the same lineage even in the presence of excess off-target sEVs. We term this effect ‘super homotypic targeting’. Super homotypic targeting provides a means to distinguish sEVs of different origins within highly heterogeneous sEV populations and enables two applications: using cells to detect specific sEVs and using sEVs to detect specific cells, specifically demonstrated here in the context of cancer detection from blood samples. Super homotypic targeting could hold potential for diagnostics, immunotherapy, drug delivery, rejuvenation and tissue engineering. Engineering the surface of small extracellular vesicles with lanthanide ions enhances homotypic targeting of cells through competitive coordination with sialic acid residues overexpressed on cancer cell surfaces, allowing the detection of cancer cell-derived small extracellular vesicles and circulating tumour cells.
The ability to selectively edit specific RNA modifications is needed to understand their roles in cellular function and disease. However, existing approaches, particularly catalytically inactive Cas (dCas)-based systems, have limited applicability owing to their reliance on eraser proteins, and their performance can vary depending on the modification type and target context. Here we introduce the RNA Modification-Blocking (RModBlock) strategy, which uses chemically modified antisense oligonucleotides (ASOs) with locked nucleic acid to precisely inhibit modifications at targeted sites. Most RNA modification writers, including those for m5C and pseudouridine, require specific structural contexts. Using representative modifications with and without eraser proteins, m5C and pseudouridine, respectively, we demonstrated that RModBlock ASOs blocked the formation of their structural context, inhibiting modified bases by up to 97% in human cells. RModBlock ASOs also inhibited m6A, despite its writer protein not requiring a strict structural context, suggesting broad applicability. Moreover, this strategy achieves comparable or superior performance when dCas13-eraser-based systems are applicable. Finally, by inhibiting cancer-relevant modifications and through in vivo delivery to the mouse liver, we highlight its therapeutic potential. This showcases the RModBlock strategy as a precise, efficient and versatile approach for manipulating RNA modifications, with broad applicability in basic and translational research.
Two studies highlight the ways in which artificial intelligence can be used to improve ionizable lipid design for lipid nanoparticle-mediated mRNA delivery.
