While general-purpose large language models (LLMs) demonstrate remarkable capabilities, their clinical application demands rigorous adaptation to ensure safety and accuracy. This review presents a comprehensive framework for transforming LLMs into trustworthy medical specialists. We detail three core knowledge-injection strategies—(1) static embedding to internalize foundational biomedical knowledge; (2) behavioral alignment to enforce clinical safety and verifiable diagnostic logic; and (3) dynamic injection, such as retrieval-augmented generation, for real-time evidence grounding—together with multimodal integration as a complementary perception-injection paradigm extending the input space beyond text to imaging, biosignals, and tabular data. Building on these strategies, we further explore the evolution toward agentic AI systems that orchestrate them for autonomous, collaborative clinical decision-making. Finally, we discuss critical challenges, including model calibration, resource constraints, standardized reporting, and robust safety protocols. Combining these complementary strategies is essential for developing deployable, domain-specialized clinical AI systems.
Recent therapeutic advances have resulted in US Food and Drug Administration (FDA) approval of the first pharmacological agents—resmetirom and semaglutide—for advanced metabolic dysfunction-associated steatohepatitis (MASH) without cirrhosis, reshaping the therapeutic landscape of the disease. Within this evolving framework, incretin-based pharmacotherapies—including glucagon-like peptide-1 (GLP-1) receptor agonists and their dual and triple combinations with glucose-dependent insulinotropic peptide (GIP) and/or glucagon receptor agonists—have emerged as promising options, particularly for individuals with coexisting obesity or type 2 diabetes. These agents exert pleiotropic effects across multiple organs, modulating glucose and lipid metabolism, while providing cardiovascular and renal benefits. Despite these advances, key challenges remain regarding treatment duration, tolerability, and interindividual variability in therapeutic response. This review summarizes the emerging role of GLP-1 mono, dual, and triple agonists in MASH-related fibrosis, focusing on their mechanisms of action and evidence from phase 2 and phase 3 clinical trials with histology-assessed hepatic endpoints.
Bacteria-based living therapeutics exhibit outstanding potential for tumor immunotherapy when combined with functional materials or rewired with rationally designed gene circuits. To stimulate T cells precisely and effectively, herein, we report the surface-engineered bacteria as the living therapeutic for tumor immunotherapy. Escherichia coli cells are genetically programmed to display surface-anchored B7H2 to provide immunological co-stimulation signal of T cells and Histag to coordinate with ZnS quantum dots. In tumors, B7H2 induces the co-stimulatory ICOS-B7H2 signal at the molecular level and triggers the stimulation of cytotoxic T cells, the polarization of helper T cells, the release of cytokines, and long term antitumor immunological memory. Meanwhile, the coordinated ZnS quantum dots image tumors and synergize with B7H2 to enhance immune effects by activating the cGAS/STING pathway. With functional domains displaying on the bacterial surface, we establish an efficient living therapeutic approach for tumor immunotherapy.
Respiratory diseases cause significant morbidity, yet diagnosis remains labor intensive and dependent on physician expertise. Here, we present LungGPT, a unified multimodal system trained on 147 million tokens of domain-specific electronic health records from 125,917 participants. LungGPT comprises two modules: LungGPT-Dx for respiratory disease diagnosis and early warning of critical illness, and LungGPT-Ex for interpretable diagnostic reasoning and treatment recommendations. In large-scale evaluations, LungGPT-Dx achieves a macro-average area under the curve (AUC) of 0.852 (95% confidence interval [CI]: 0.839-0.865) across 22 respiratory diseases, with disease-specific AUCs exceeding 0.900 for lung cancer and pulmonary tuberculosis. Crucially, the model further improves early warning of critical illness by incorporating chain-of-thought (CoT) reasoning into textual data and integrating computed tomography (CT) imaging features. LungGPT-Ex generates high-quality, interpretable reasoning that outperforms specialized clinical models and matches advanced general-purpose models such as GPT-4o and DeepSeek-R1 in correctness, completeness, and truthfulness. By bridging precision diagnostics and rapid decision-making, LungGPT provides a standardized framework to enhance clinical workflows and improve patient outcomes in respiratory healthcare.
We evaluate the efficacy and immunogenicity of HIV clade A/E A244 envelope (Env) immunogens with the 23 amino acids of variable region 1 deleted (ΔV1) or retained (wild-type [WT]) in macaques. Only the ΔV1 regimen significantly reduces the risk of mucosal acquisition of clade C simian/human immunodeficiency virus (SHIV)1157(QNE)Y173H versus controls, providing 81% efficacy and leaving 10 of 12 immunized animals uninfected. ΔV1 vaccination induces higher systemic antibody-dependent cellular cytotoxicity (ADCC) targeting helical-V2 and anti-inflammatory myeloid cells, which, together with IL-17+NKp44+ innate lymphoid cells (ILCs) and systemic PD-1+ helper T cells, correlated with reduced infection risk. By contrast, WT immunization induces higher IL-15, CCR2+pDC, and gp70/V1V2-biased responses, which, along with mucosal IFN-γ+NKG2A-NKp44- ILCs, were associated with increased susceptibility. Ex vivo, ΔV1 gp120 reduced CCR5 expression on CD4+ T cells relative to WT gp120, consistent with the anti-inflammatory mucosal response by ΔV1-vaccine regimens in vivo. Thus, V1 deletion promotes an anti-inflammatory mucosal landscape less permissive to HIV seeding and dissemination following virus exposure.
Enfortumab vedotin is a Nectin-4-directed antibody-drug conjugate designed to deliver the microtubule-disrupting agent monomethyl auristatin E (MMAE) to tumor cells. Using preclinical models of urothelial cancer (UC), we expand the understanding of the multifaceted mechanism of action for enfortumab vedotin that includes direct cytotoxicity on Nectin-4-positive tumor cells, indirect bystander effect on neighboring Nectin-4-negative tumor cells, and MMAE-mediated induction of immunogenic cell death (ICD) and associated increase in activated immune cells in the tumor microenvironment. Importantly, vaccination with enfortumab vedotin-treated tumor cells results in protection against tumor rechallenge in mice, consistent with antitumor immunity. MMAE-mediated ICD induction modulates the tumor microenvironment in a complementary manner to immune checkpoint inhibition. Accordingly, enfortumab vedotin plus PD-1 inhibitor shows enhanced antitumor activity in vivo. These preclinical findings provide mechanistic hypotheses that may be relevant to the improved clinical outcomes observed for enfortumab vedotin plus pembrolizumab relative to chemotherapy.
Although immunotherapy has transformed the treatment landscape for many types of cancer, its therapeutic efficacy in glioblastoma (GBM) is limited by insufficient antigen presentation and the immunogenic cell exclusion in the tumor microenvironment. Here, we develop a candidate-based CRISPR activation (CRISPRa) functional screen to identify regulators of conventional dendritic cell (cDC)-fate specification. We determine that the transcription factors Zfp366/Znf366, Pu.1, Irf8, and Batf3 (ZPIB) are sufficient to convert GBM cells into cDC-like cells. ZPIB-mediated reprogramming results in global transcriptional and epigenetic remodeling in glioma cells. Single-cell RNA sequencing (scRNA-seq) profiling also reveals efficient and dynamic reprogramming of GBM cells to cDCs in vivo. Moreover, reprogrammed tumor cells remodel the microenvironment and elicit systemic tumor-eradicating and durable antitumor immunity in multiple mouse GBM models. Antitumor immunity elicited by ZPIB-DCs is synergistic with immune checkpoint inhibitors. Finally, we evaluate the clinical applicability of this approach by generating ZPIB-DCs from GBM patients within a humanized model. Our study represents a cellular reprogramming therapeutic strategy with broad implications for clinical immunotherapy.
c-Met overexpression promotes tumor progression in many cancers, yet approved inhibitors benefit only patients with MET mutation, leaving most c-Met-overexpressing patients without effective therapy. Inspired by antibody-drug conjugates (ADCs), composed of an antibody linked to cytotoxic agents, a c-Met-targeting drug conjugate enabling MET-independent cytotoxicity offers a strategy to address this gap. Here, we developed an aptamer-drug conjugate (ApDC), integrating SL1, a c-Met-targeting aptamer—short oligonucleotide with high target affinity—with monomethyl auristatin E (MMAE), via cathepsin B-sensitive linker. This ApDC selectively binds c-Met-overexpressing cells, undergoes receptor-mediated internalization, and releases MMAE to induce apoptosis. It achieves efficient tumor accumulation, sustained payload retention, rapid systemic clearance, and robust anti-tumor efficacy across multiple c-Met overexpressing tumor models. With maintained surface receptor expression, rapid tumor accumulation and active payload release, it outperforms a benchmark c-Met-targeting ADC. Combining precise targeting, potent efficacy and favorable safety, this ApDC represents a promising strategy for c-Met-targeted cancer therapy.
The infant gut resistome is established early in life and is shaped by perinatal exposures, yet the mechanisms underlying its modulation remain unclear. We combined shotgun metagenomics of fecal samples from 57 one-month-old infants and paired milk samples from 50 mothers in the MAMI cohort to investigate the influence of maternal secretor status on early-life resistome development. Longitudinal follow-up at 6 and 12 months, and also further validation in the independent Lifelines NEXT (LLNEXT) cohort, support our findings. Cesarean section (C-section) was associated with increased antibiotic resistance gene (ARG) diversity, whereas exclusive breastfeeding reduced ARG abundance and diversity. Maternal secretor status further modified resistome composition among exclusively breastfed infants. Human milk oligosaccharide profiling identified specific glycans underlying these associations, with 2'-fucosyllactose and 6'-sialyllactose showing negative correlations with distinct ARG classes. These findings identify human milk composition as a key determinant of early-life resistome assembly and a potential target for modulating antimicrobial resistance.
Acute myeloid leukemia (AML) persistence and relapse are sustained by leukemia-propagating cells, yet the molecular programs supporting their expansion during disease evolution remain incompletely understood. Using serial patient-derived xenotransplantation, we establish a longitudinal model in which leukemia-initiating capacity progressively increases. Integrated single-cell transcriptomics and multi-omics profiling reveal a predominantly non-genetic trajectory that follows a conserved pattern across models and is associated with coordinated changes across epigenetic, transcriptional, and proteomic layers. Ribosome profiling and rRNA 2'-O-methylation analyses further support a stage-specific increase in translational activity with ribosome remodeling in advanced xenografts. A pharmacological screen of 3,247 compounds uncovers a limited set of vulnerabilities that consistently emerge during disease progression, including CRBN-dependent degradation of GSPT1 (CC-885) and IAP antagonism (AZD5582). In vivo validation shows that both agents markedly reduce leukemic burden, impair leukemia propagation, and enhance cytarabine activity in patient-derived xenograft (PDX) models. Together, these findings show that leukemic propagation is driven by a non-genetic remodeling program, providing a framework to prioritize and test stage-specific therapeutic strategies in AML.
High metabolic heterogeneity and plasticity of triple-negative breast cancer (TNBC) contribute to therapy resistance, necessitating identification of therapeutic vulnerabilities. Here, we identify non-canonical functions of the extracellular matrix (ECM) remodeler, lysyl oxidase (LOX), in regulating glucose metabolism and mitochondrial homeostasis and show that inhibiting LOX generates targetable vulnerability to ferroptosis. Mechanistically, LOX interacts with PARKIN and its upstream kinase PINK1, which we identified as a substrate of LOX. LOX-mediated PINK1 oxidation suppresses PARKIN phosphorylation, stabilizing hypoxia-inducible factor 1-alpha (HIF-1α) and increasing glycolysis. Concomitantly, LOX inhibits PARKIN-mediated mitophagy and maintains mitochondria-ER contacts through VDAC1 stabilization, while the LOX-HSP90 complex promotes mitochondrial Ca2+ transport and ATP production. Inhibiting LOX suppresses glycolysis, disrupts mitochondrial dynamics, reduces OXPHOS and GPX4/FSP1, and induces compensatory DHODH activity. Our “one-two punch” approach combining LOX inhibition with clinical DHODH inhibitor suppresses tumor growth in vivo in chemo-free setting. Notably, LOX protein correlates with HIF-1α/GLUT1/GPX4 in TNBC patient tumors, supporting its clinical relevance.
Parkinson's disease (PD) motor symptoms are linked to excessive beta band (13-35 Hz) oscillations and basal ganglia-cortical (BGC) synchrony, yet their dynamic inter-relation remains unclear. Based on computational modeling, we hypothesize that transient high-beta (20-35 Hz) BGC coupling induces subcortical low-beta (13-20 Hz) amplification. We recorded intraoperative neural signals from the globus pallidus internus and externus (GPi and GPe, respectively) and motor cortex (M1) in 23 PD patients during deep brain stimulation implantation. High-beta bursts in M1 and GPe were classified as either temporally synchronized or isolated. Synchronous M1-GPe high-beta bursts led by GPe were followed by GPi low-beta amplification, which were not evident with either isolated M1 high-beta bursts or M1-led synchronous bursts. Dopaminergic medication attenuated this GPi low-beta amplification and improved symptoms. These findings suggest that excessive M1 high-beta bursts following GPe high-beta activity trigger a cascade establishing low-beta propagation to GPi, which may contribute to PD motor dysfunction.
Cyclin-dependent kinase 4/6 (CDK4/6) inhibitors have become a standard of care for estrogen receptor-positive breast cancer and are being developed for other malignancies. However, resistance to these drugs readily develops, limiting their impact on patient survival. Mechanisms of resistance to CDK4/6 inhibition involve multiple changes in gene expression. We investigated the process of tumor cell adaptation to CDK4/6 inhibitors and the impact of selective inhibitors of CDK8/19 Mediator kinases—broad-spectrum regulators of transcriptional reprogramming—on such adaptations. Adaptive non-genetic resistance to CDK4/6 inhibitors develops rapidly, but the addition of CDK8/19 inhibitors prevents the development of this resistance in different tumor models, in vitro and in vivo. RNA sequencing (RNA-seq) analysis reveals that combining CDK4/6 and CDK8/19 inhibitors suppresses many of the adaptation-associated changes in gene expression, including those previously associated with CDK4/6 inhibitor resistance. The findings suggest that CDK8/19 inhibition may greatly extend the therapeutic benefit of CDK4/6 inhibitors.
The maternal immune system must simultaneously defend against pathogens with tolerance of the antigenically distinct fetus. Despite the central importance of immune modulation during pregnancy, epitope-level analyses of the maternal antibody repertoire remain lacking. Here, we performed high-throughput, multi-isotype longitudinal serology of 119 women from preconception through pregnancy, combining phage display of linear peptides with whole protein-based antigen assays to profile antibody responses to viral and self-antigens. We observed a broad expansion of the maternal IgG repertoire and a relative increase in IgA titers, both tracking with pregnancy progression. These changes may augment transplacental and mucosal immunity in the fetus and newborn, respectively. Notably, pregnancies ending in miscarriage exhibited pronounced IgA titer spikes against both viral and self-antigens unrelated to active infection (sensitivity 14/19, 74%, specificity 95%), indicating an underlying immune perturbation may precede pregnancy loss.
Adoptive T cell therapy for solid tumors is limited by autologous manufacturing complexity and, in allogeneic settings, risks including graft-versus-host disease (GvHD), HLA restriction, and donor variability. We develop a scalable, feeder-free platform to differentiate gene-engineered hematopoietic stem and progenitor cells (HSPCs) into allogeneic, NY-ESO-1-specific cytotoxic T (AlloESO-T) cells. Product phenotype, function, tumor homing, and safety are assessed against solid tumor models and benchmarked to peripheral blood mononuclear cell (PBMC)-derived TCR-engineered T cells. AlloESO-T cells display a uniform cytotoxic phenotype, with dual tumor targeting through a transgenic TCR and natural killer receptors. Relative to PBMC-derived counterparts, AlloESO-T cells show superior cytotoxicity, selective solid-tumor homing, durable killing persistence, and resilience to immune evasion. They also maintain low GvHD and cytokine release syndrome risk, while retaining stable hypoimmunogenic features. These findings establish HSPC-derived AlloESO-T cells as an off-the-shelf, mono-specific cytotoxic T cell therapy with scalable manufacturing, enhanced efficacy, and improved safety, which support broad applicability of AlloESO-T cells across solid tumors.
Prospective evidence linking greenness and cardiovascular disease (CVD) in rapidly urbanizing developing countries remains limited. Here, among 159,590 adults aged ≥40 years from the nationwide China Cardiometabolic Disease and Cancer Cohort with a median follow-up of 10.1 years, we examine the association between residential greenness, measured by satellite-derived normalized difference vegetation index (NDVI) within 500 m of residence, and incident CVD, and evaluate its joint effects with cardiovascular health as defined by Life’s Essential 8. Individuals in the highest quartiles of contemporaneous, one-year, and cumulative NDVI consistently show lower CVD risk compared with those in the lowest quartiles, although associations vary across subpopulations. Notably, individuals with high cardiovascular health scores living in low-NDVI areas exhibit similar CVD risk to those residing in high-NDVI areas. These findings highlight the complementary importance of both green infrastructure and healthy lifestyles in reducing CVD risk in rapidly urbanizing regions of China.
Myotonic dystrophy type 1 (DM1) is a severe neuromuscular disorder caused by CTG repeat expansions in the DMPK gene, leading to the formation of toxic RNA foci that sequester essential splicing regulators MBNL1/2. Beyond muscle impairment, DM1 affects also the brain, leading to significant cognitive deficits, behavioral abnormalities, and intellectual disabilities. This study evaluates the therapeutic potential of the lipid-conjugated antimiR-23b, X82108, designed to promote MBNL1/2 upregulation through inhibition of miR-23b. Systemic administration of X82108 in mice and non-human primates efficiently crosses the blood-brain barrier, increasing MBNL1 in the brain. In DMSXL transgenic mice, treatment increases Mbnl1/2, reduces toxic DMPK, and restores normal splicing patterns across all brain regions. These molecular improvements correlate with improved behavioral outcomes, including reduced impulsivity and normalized exploratory activity. Collectively, the findings highlight X82108 as a promising systemic therapy for DM1, targeting not only muscular features as we have previously shown but also DM1-related CNS alterations.
Antibody-drug conjugates (ADCs) have transformed the treatment of HER2-positive breast cancer, yet resistance remains poorly understood. Using imaging mass cytometry, we profiled 157 regions of interest comprising 912,360 single cells from 47 HER2-positive/hormone receptor-negative breast cancers treated with SHR-A1811 in the FASCINATE-N trial. Spatial proteomic analyses identified two determinants of ADC response: elevated tumor-cell H3K27ac expression was associated with improved ADC efficacy, whereas collagen-positive fibroblasts mediated resistance. Combining ADC with the histone deacetylase inhibitor chidamide or the collagen-modulating agent losartan produced synergistic antitumor effects in preclinical models. These biomarkers and therapeutic vulnerabilities were independently validated in patients with advanced HER2-positive disease receiving trastuzumab deruxtecan. Moreover, based on these spatial features, we developed a clinically applicable ADC barrier prediction model that can be implemented using multiplex immunofluorescence. Taken together, our findings reveal actionable spatial determinants of ADC efficacy and suggest potential combination therapeutic strategies.
Tumor immunotherapies enhance CD8+ T cell function, yet heterogeneous responses in "cold" and "hot" tumors remain a challenge. Although biomechanical cues modulate T cell cytotoxicity, strategies to harness these forces for broad antitumor potentiation remain elusive. Here, integrating pan-cancer single-cell RNA sequencing data, we identify cofilin 1 (CFL1) as a determinant of immunotherapy response. CFL1 overexpression synergizes with CD8+ T cell-targeted immunotherapy across tumor types, driving intratumoral T cell expansion while rendering tumors physically vulnerable. Mechanistically, CFL1 hyperactivation induces F-actin bundling and actin rod accumulation, elevating cytoskeletal tension to facilitate immunological synapse formation. Conversely, CFL1 inactivation through phosphorylation or our newly identified lactylation correlates with poor outcomes. Combination therapy using adeno-associated virus delivering constitutively active CFL1 and PD-1 blockade achieves near-complete tumor eradication. Together, our findings position CFL1 as a biomechanical checkpoint governing tumor vulnerability to CD8+ T cells through immunological synapse licensing, offering a strategy to overcome immunotherapy resistance.
Exercise is an integral therapy for many cardiometabolic diseases, including obesity, type 2 diabetes, and hypertension. Despite its broad health benefits, the circulating factors that mediate exercise adaptations in humans remain incompletely defined, particularly across different exercise intensities. Here, we conducted a multi-cohort human exercise intervention incorporating sprint-interval exercise (SIE) and moderate-intensity exercise (MIE) to analyze intensity-dependent regulation of interorgan crosstalk. We found that exercise intensity distinctly influenced the plasma proteome and metabolome in untrained and trained participants. By integrating multi-organ gene and protein expression datasets with in vitro and in vivo tissue sampling, we mapped regulated proteins to their predicted tissues of origin and destination. Muscle fibers and adipocytes were particularly sensitive to exercise intensity and observed to undergo broad secretory and transcriptomic changes. Moreover, we leveraged a large-scale plasma-phenome database to identify intensity-dependent proteins associated with cardiometabolic health and disease, highlighting how exercise intensity differentially shapes interorgan communication and organismal health.
