Nature Biomedical Engineering

Accurate polyp segmentation from colonoscopy images is critical for colorectal cancer prevention, yet the generalization of deep learning models under domain shift remains insufficiently explored. We propose Boundary-Explicit Guided Attention U-Net (BEGA-UNet), a boundary-aware segmentation architecture that introduces explicit edge modeling as a structural inductive bias to enhance both segmentation accuracy and cross-domain robustness. The framework integrates three components: an Edge-Guided Module (EGM) with learnable Sobel-initialized operators to capture boundary cues, a Dual-Path Attention (DPA) module that processes channel and spatial attention in parallel, and a Multi-Scale Feature Aggregation (MSFA) module to encode contextual information across multiple receptive fields. Evaluated on the combined Kvasir-SEG and CVC-ClinicDB benchmarks, BEGA-UNet achieves 88.53% Dice and 82.51% IoU, outperforming representative convolutional and transformer-based baselines. More importantly, cross-dataset evaluation demonstrates strong robustness under domain shift, with BEGA-UNet retaining 83.2% of its in-distribution performance--substantially higher than U-Net (64.5%), Attention U-Net (47.5%), and TransUNet (53.1%). In a zero-shot setting on an entirely unseen dataset, the model further maintains 72.6% performance retention. Comprehensive ablation studies indicate that explicit boundary modeling plays a central role in improving generalization, while multi-scale context aggregation further stabilizes performance across domains. Feature distribution analyses support this observation by showing that edge-oriented representations exhibit markedly reduced cross-domain variability compared to appearance-driven features. Overall, BEGA-UNet provides an effective and interpretable solution for robust polyp segmentation, demonstrating that explicit boundary modeling serves as a critical inductive bias for ensuring reliability under clinical domain shifts.

Deep learning models utilizing longitudinal healthcare data have significantly advanced epidemiological research. However, contemporary transformer-based models increasingly rely on computationally intensive pre-training steps that entail processing massive real-world datasets with cost-prohibitive hardware. We introduce the Temporal Encoder with Late Fusion (TELF), a lightweight end-to-end predictive model featuring an encoder-only architecture for processing medical codes, followed by post-encoder concatenation with demographic variables. TELF learns code embeddings on-the-fly, thereby bypassing the resource-intensive pre-training bottleneck. Furthermore, its late-fusion design preserves the integrity of the temporal attention mechanism before integrating static demographic predictors. We evaluated TELF using an administrative claims database across three distinct cohorts: pancreatic cancer (n=53,661), type 2 diabetes (n=78,756), and heart failure (n=72,540). TELF consistently outperformed traditional machine learning baselines, including XGBoost, LightGBM, and logistic regression. Specifically, TELF achieved AUCs of 0.9150, 0.8199, and 0.8721 for pancreatic cancer, type 2 diabetes, and heart failure, respectively, compared with 0.9044, 0.7908, and 0.8535 for XGBoost and 0.9014, 0.7800, and 0.8466 for logistic regression. Beyond predictive superiority, TELF's isolated temporal attention mechanism enables population-level motif mining. By extracting high-attention temporal sequences, we mapped aggregated patient journey pathways, revealing interpretable clinical trajectories preceding disease onset. Collectively, these results demonstrate that TELF provides a resource-efficient and accessible framework for advanced temporal modeling in clinical and epidemiological research.

We developed an integrated platform combining high-throughput automated biofabrication, systematic patient-derived tissue experiments, and specialized artificial intelligence to enable patient-specific computational "digital twins" for thrombosis prediction. Our automated manufacturing platform fabricates 80 fully assembled, patient-specific vessel-on-chips within 10 hours from clinical imaging--a [~]100-fold improvement over manual methods--achieving sub-micron precision through novel two-stage pneumatic motion control and integrated optical feedback. Using these chips, we systematically captured thrombosis across 491 high-fidelity videos spanning 6 patient-derived vascular geometries, 5 distinct anatomical injury sites, and 14 anticoagulant/antiplatelet interventions, establishing a "physical twin" experimental corpus. We trained CLoT (Cascade Learner of Thrombosis), a conditional video diffusion model efficiently adapted via lightweight Low-Rank Adaptation (LoRA) to generate realistic thrombosis videos conditioned on patient-specific geometry, injury location, and drug treatment. Rigorous benchmarking against state-of-the-art commercial models (Sora, Wan, Kling, Seedance, Hailuo, Hunyuan) reveals CLoT achieves 7.38-fold superior temporal biological consistency and 5.3-fold higher spatial morphological fidelity. Prospective validation on unseen patients demonstrates >90% temporal accuracy. This integrated paradigm--combining automated fabrication with domain-specialized generative AI--establishes proof-of-concept for personalized medicine enabled by digital twins trained on human-derived vascular anatomy, enabling pre-treatment antithrombotic evaluation while providing a replicable template for translating tissue engineering into clinical practice.

Adaptive, closed-loop control of cellular behavior is essential for next-generation therapies, yet most current treatments operate in an open-loop manner and lack robustness to patient variability and disease dynamics. Here, we establish a controltheoretic platform for rational engineering of closed-loop cell-based therapies that achieve precise and robust regulation. First, we introduce multi-dimensional nullgram profiling, a high-throughput approach that enables quantitative prediction and design of advanced genetic controllers in human cells across circuit topologies and parameter regimes in a single experiment. To evaluate dynamic therapeutic behavior, we next develop Cyberpatient-in-the-loop, an optogenetic digital twin platform that interfaces engineered mammalian cells with computational disease models, enabling systematic testing of closed-loop performance under realistic perturbations. Finally, we leverage these approaches to implement integral feedback cell therapies that sense inflammatory signals and autonomously regulate cytokine levels in primary immune cell cultures. Together, these results establish a general paradigm for engineering cellbased control systems and provide a foundation for next-generation cell therapies.

Chronic constipation is highly prevalent, and cases refractory to treatment are particularly challenging to manage. High-resolution colonic manometry (HRM) is used to further evaluate these patients to identify cases of intrinsic motor dysfunction (underlying myopathy or neuropathy). However, HRM is invasive and resource-intensive, limiting uptake and clinical utility. This study presents Body Surface Colonic Mapping (BSCM), a non-invasive cutaneous electrical recording technique, as a clinical alternative. Simultaneous recordings from HRM (36-channel) and BSCM (8x8 electrode array) were performed in 10 patients with chronic refractory constipation. Lower gut symptom scores were also tracked patients over the duration of the recording. Motility was assessed during meal and bisacodyl challenges. We optimized BSCM signal processing specifically to detect high-amplitude propagating contractions (HAPCs) evoked by bisacodyl. Analysis included time-frequency quantification of motility indices and blinded visual assessment by domain experts to classify the presence or absence of motor responses. BSCM motility indices showed strong correlation with HRM for both meal (r = 0.86) and bisacodyl (r = 0.69) responses. Expert visual analysis yielded concordant classification between BSCM and HRM in the majority (87.5 {+/-} 9.6%) of cases. Furthermore, BSCM identified distinct, patient-specific symptom-motility associations during the meal response. BSCM accurately detects meal- and stimulant-induced increases in colonic motility with high fidelity to invasive HRM. As a non-invasive method that is easy to apply with minimal resource and time requirements, BSCM is well-positioned for clinical translation as a scalable diagnostic tool to elucidate symptom-motility associations and guide personalized management in refractory chronic constipation.

We built and evaluated a zero-shot LLM pipeline with automated, task-aware prompt optimization to extract radiology and symptom fields for gallstone phenotyping from de-identified EHR text. Across symptomatic, asymptomatic, and control cohorts, it performed reliably on high-signal binary fields and symptom flags but lagged on fine-grained stone burden and complications, establishing a practical baseline and motivating targeted refinements

Efficient and scalable isolation of specific cell populations remains a central bottleneck for genome engineering, pooled screening, and cell therapy manufacturing. Here, we present DASIT (Destabilized-nanobody Antigen Selection and Identification Tool), a protein-based circuit for antigen-specific cell selection. DASIT uses a destabilized nanobody fused to an antibiotic resistance protein. In cells expressing the target antigen, binding of the nanobody fusion to the cognate antigen stabilizes DASIT, thereby coupling the presence of an antigen to a selectable signal. We developed DASIT circuits that enable robust selection of antigen-expressing cells and show that they can be designed to target distinct antigen classes and perform across cell types. Because DASIT operates at the protein level, it supports both stable integration and transient delivery, enabling recyclable selection without permanent genomic integration of resistance markers. We demonstrate scalable, FACS-free enrichment in three challenging applications: multiplexed, logic-gated integration of landing pads in human iPSCs, high-throughput CRISPR screening, and phenotypic selection of in vitro-derived neurons at transplantation scale. By decoupling selection from vector integration, DASIT establishes an automation-compatible architecture for multistep genome engineering, high-throughput library screening, and large-scale cell manufacturing. HighlightsDASIT enriches for antigen-positive cells across multiple selection markers and antigens Intermediate levels of DASIT expression support selection across stable and transient delivery modalities Logic-gated, precision genome engineering of human iPSCs via DASIT selection DASIT enables scalable activity-based selection for high-throughput base editing screens DASIT-selected engineered motor neurons survive grafting into acute spinal cord injury

Restoring communication for people with dysarthria secondary to pontine stroke remains a critical challenge. Intracortical brain-computer interfaces (iBCIs) have demonstrated great potential for speech restoration in people with amyotrophic lateral sclerosis (ALS), with 1-24% word error rates (WERs) on a 125,000-word vocabulary. In pontine stroke, electrocorticography (ECoG) BCIs achieved 25.5% WERs with a smaller 1,024-word vocabulary. Whether intracortical BCI performance improvements extend to people with pontine stroke-induced dysarthria remains unclear. Here, we show that neural activity from a single 64-channel microelectrode array in orofacial motor cortex can predict attempted speech in a person with pontine stroke more accurately than prior ECoG BCI work and comparably to prior iBCI work. We trained a neural network decoder to predict phoneme probabilities from spiking rates and spike-band power as BrainGate2 participant T16 mimed (mouthed without vocalization) sentences from a large vocabulary. A series of language models converted these probabilities into word sequences. This decoding architecture has remained stable more than two years post-implantation, achieving a median 19.6% WER with a 125,000-word vocabulary and a median 10.0% WER with a 1,024-word vocabulary (a 60.8% reduction over prior ECoG studies). This framework also generalized beyond cue repetition, enabling T16 to communicate spontaneously via the iBCI in a question-and-answer setting with a 35.2% WER. These results demonstrate that brain-to-text decoding from a small patch of cortex can outperform ECoG-based systems in individuals with pontine stroke and is comparable to early speech iBCIs in individuals with ALS.

Synthetic data generation is a promising approach for biomedical data sharing and dataset augmentation, yet existing methods lack mechanisms to preserve statistical properties necessary for scientific analysis. To address this, we introduce RLSYN+REG, a reinforcement learning-driven generative model, which encourages that regression models trained on synthetic data reproduce the coefficients and predictions of their real-data counterparts. We evaluate RL-SO_SCPLOWYNC_SCPLOW+RO_SCPLOWEGC_SCPLOW on MIMIC-III and the American Community Survey (ACS) across regression model reproduction, fidelity to real data, and privacy. Synthetic data from RLSO_SCPLOWYNC_SCPLOW+RO_SCPLOWEGC_SCPLOW substantially improves upon that of RLSO_SCPLOWYNC_SCPLOW, raising correlations between real and synthetic regression coefficients from 0.054 to 0.600 on MIMIC-III and from 0.160 to 0.376 on ACS. Predictive performance also improves, reducing the gap between real-data baselines by 81.4% and 97.6% on MIMIC-III and ACS, respectively. These improvements come with negligible cost to fidelity or privacy and are robust to reductions in training data.

Foundation models (FMs) promise to standardise predictive modeling across domains, yet their clinical value for tabular data remains unproven. To test this, we performed a large, fully reproducible benchmark of TabPFN, a leading FM for tabular prediction, against twelve established machine learning (ML) methods across twelve binary clinical tasks. Cohorts spanned 788 - 139,528 patients across diverse outcomes, including survival, metastasis, and disease status. Using standardized preprocessing, bootstrapping, and multiple performance metrics, TabPFN was generally competitive but did not consistently outperform strong ML baselines. It exceeded the best ML model in only 16.7% of tasks, with most area under the receiver operating characteristic (AUROC) differences within {+/-}0.01. TabPFN also incurred higher computational cost, with median runtimes 5.5x longer and practical reliance on GPU acceleration. These findings indicate that, for routine clinical tabular prediction, TabPFN offers limited performance gains relative to optimized ML methods, while introducing significant efficiency trade-offs.

We present a streamlined, solid-phase workflow for Oxford Nanopore sequencing that integrates DNA extraction, purification, and library preparation within a single microfluidic cartridge. By eliminating tube transfers and performing all enzymatic steps directly on captured DNA, the method minimizes sample loss, reduces hands-on time, and simplifies library generation for long-read sequencing. Starting from volumes as small as a single drop of blood, this integrated approach produces high-quality sequencing libraries from cell lines, whole blood, and tissue. The workflow achieves robust recovery of high-molecular-weight DNA and high pore occupancy, enabling rapid, low-complexity sample preparation suitable for clinical, field, and decentralized sequencing applications.

"Black box" deep learning models for medical image interpretation limit clinical trust and analysis of performance degradation. Here, we introduce Concept-Level Embeddings for Auditable Radiology (CLEAR), an auditable foundation model based on clinical concepts. Trained on over 0.87 million image-report pairs from 239,091 patients, CLEAR learns a visual representation and projects chest X-rays into a semantically rich space defined by large language model embeddings, making every prediction traceable to specific radiological observations. External validation on four large, physician-annotated datasets from the United States, Europe, and Asia shows that CLEAR not only achieves state-of-the-art classification performance but also enables novel applications: auditable zero-shot pathology detection, systematic identification of radiological confounders, and the creation of expert-level concept bottleneck models from data-driven concepts. By integrating clinical knowledge directly into its reasoning process, CLEAR offers a framework for robust model auditing, safer deployment, and enhanced physician-AI collaboration, advancing towards trustworthy medical AI.

Selecting an effective treatment relies on accurately anticipating patient's response to alternative interventions. However, forecasting longitudinal clinical trajectories remains difficult because electronic health records contain heterogeneous, irregularly sampled data over extended time periods. These issues are especially relevant for laboratory measurements, which are central for diagnostics, assessment of therapeutic responses, and tracking disease progression in routine clinical practice. However, existing deep learning methods for counterfactual prediction usually assume regularly sampled data, an assumption incompatible with the irregular, heterogeneous data-generation processes of real-world clinical practice. Here we present the Time-Aware G-Transformer, which integrates causal G-computation with time-aware attention to predict counterfactual outcomes on irregular data. By explicitly conditioning on the timing of future observations and encoding measurement patterns, the model captures temporal dynamics that previous methods overlook. Evaluated on synthetic tumor growth data and on 90,753 cancer patient trajectories from an academic medical center, our approach demonstrates superior long-horizon (> 1 day) prediction accuracy and uncertainty calibration compared to state-of-the-art baselines. These results demonstrate that embedding temporal relations directly into the attention mechanism enables robust integration of patient history data for evaluating potential treatment strategies in personalized medicine.

Sleep disturbances are pervasive, debilitating non-motor symptoms of Parkinson's disease (PD), where sleep spindle deficits directly drive cognitive decline and disease progression. Current adaptive deep brain stimulation (aDBS) for PD is largely limited to motor symptom management, with no established technical foundation for sleep spindle-targeted closed-loop modulation. The functional role of the basal ganglia in human sleep spindle regulation remains incompletely characterized, and no robust cross-subject pipeline exists to decode these transient events from clinically implanted DBS electrodes. Here, we developed a connectomics-guided meta-learning framework for cross-subject sleep spindle decoding and anticipatory prediction, using whole-night synchronized basal ganglia local field potential and polysomnography data from 17 PD patients with bilateral DBS implants. Our framework achieved 92.63% accuracy for concurrent spindle decoding and 83.44% accuracy for 2-second-ahead prediction, with optimal signals localized to the limbic subthalamic nucleus and <50 ms total latency meeting real-time closed-loop requirements. This work defines the neuroanatomical substrate of basal ganglia spindle signaling in PD, establishes the cross-subject spindle decoding pipeline for clinical DBS systems, and provides a critical translational foundation for sleep-targeted closed-loop aDBS to mitigate PD non-motor burden.

Bispecific T cell engagers (TCEs) have transformed targeted immunotherapy but remain constrained by ex vivo manufacturing and systemic delivery, limiting their application in autoimmune disorders. Here, we introduce the first in vivo TCE platform, utilizing a T cell-targeted integration-deficient lentiviral vector (T-IDLE) to reprogram endogenous T cells into engineered T cells secreting BCMAxCD3 TCEs (STCEs) directly in vivo. In murine Sjogrens syndrome-associated dry eye model, a single intravenous administration of T-IDLESTCE achieved deep and sustained depletion of plasmablasts and plasma cells, reversing lacrimal gland inflammation and restoring tear production without detectable off-target transduction. Longitudinal studies in cynomolgus monkeys confirmed translational promise, demonstrating potent plasma cell clearance and favorable safety profiles over 12 weeks post-vector dosing. Mechanistic analyses demonstrated attenuation of Th1/Th17-mediated inflammation and controlled cytokine responses, accompanied by negligible vector integration in non-T cell lineages, thereby substantiating the safety profile of the integration-deficient design. This first-in-class in vivo TCE approach represents a manufacturing-free, scalable immunotherapy strategy for autoantibody-mediated diseases.

Engineered T cell therapies have achieved significant clinical success in hematological malignancies but remain largely ineffective in solid tumors. Overcoming this limitation requires strategies that enhance T cell function while avoiding systemic immune toxicities and pathological T cell states. Existing approaches typically rely on constitutive gene overexpression or suppression to augment potency or remodel the tumor microenvironment, but these strategies frequently lead to dysregulated immune activation and dose-limiting toxicity. Here, we present Hybrid Receptors (Hybrid-Rs), a modular receptor platform that integrates features of chimeric antigen receptors (CARs) and SyNthetic Intramembrane Proteolysis Receptors (SNIPRs) to couple antigen-dependent T cell activation with programmable gene regulation. Hybrid-Rs enable precise, context-dependent control of T cell potency, differentiation states, and conditional expression of secreted immunotherapeutic payloads with otherwise prohibitive toxicity. Hybrid-Rs are readily humanized and compatible with precision genome editing in primary human T cells, providing a direct and practical path to clinical translation.

Clinical deployment of foundation models requires decision policies that operate under explicit error budgets, such as a cap on false-positive clinical calls. Strong average accuracy alone does not guarantee safety: errors can concentrate among patients selected for action, leading to harm and inefficient use of healthcare resources. Here we introduce SO_SCPLOWTRATC_SCPLOWCP, a stratified conformal framework that turns foundation model predictions into decision-ready outputs through error-controlled selection and calibrated deferral. SO_SCPLOWTRATC_SCPLOWCP first selects a subset of patients for immediate clinical action while controlling the false discovery rate at a user-specified level. For the remaining patients, it returns prediction sets that achieve target coverage conditional on deferral, supporting confirmatory testing or expert review. When clinical guidelines define relationships among disease states, SO_SCPLOWTRATC_SCPLOWCP incorporates a utility graph to produce clinically coherent prediction sets without sacrificing coverage guarantees. We evaluate SO_SCPLOWTRATC_SCPLOWCP in ophthalmology and neuro-oncology across diagnosis, biomarker prediction, and time-to-event prognosis. Across tasks, SO_SCPLOWTRATC_SCPLOWCP controls the false discovery rate among selected patients and provides valid, selection-conditional coverage for deferred patients. In neuro-oncology, it enables H&E-based diagnosis under a fixed error budget, reducing reliance on reflex molecular assays and lowering laboratory cost and turnaround time. SO_SCPLOWTRATC_SCPLOWCP establishes error-controlled decision policies for safe deployment of medical foundation models.

Pancreatic ductal adenocarcinoma (PDA) remains highly lethal, in part, because its dense fibroinflammatory stroma restricts therapy distribution, including adoptive T cell immunotherapies where direct interactions between T and carcinoma cells are essential for effective therapy. While T cell function must be maintained once effector-target engagement occurs, without inducing co-localization subsequent cytotoxic function steps cannot be undertaken. We therefore developed a strategy to "physically optimize" T cells to more effectively sample complex tumor volumes. Informed by pharmacologic perturbations and mathematical modeling we shifted T cell phenotype through expression of constitutively activated RhoA to increase cortical contractility, activation, migration, and sampling in PDA, while showing decreases in exhaustion markers. In CAR T cells this results in more efficient targeting through decreased sampling time and increased engagement with carcinoma cells, consistent with modeling predictions. This significantly increases T cell infiltration and distribution in PDA, resulting in improved tumor control in vivo, suggesting that this is an effective strategy to overcome stromal constraints, improve tumor engagement, and enhance the therapeutic performance of engineered T cell therapies in solid tumors.

Advancing neural interfaces requires large-scale, high-density recording technologies capable of capturing full-spectrum neural activity across cortical and subcortical regions. Here, we present a scalable approach to integrate neural electrodes with advanced application-specific integrated circuits (ASICs). Specifically, we custom-designed an ASIC with 5,376 simultaneous channels, each sampling at 20 kS/s and enabling >1.3 Gb/s total data streaming throughput. The ASIC incorporates in-pixel amplification, time-division multiplexed ADCs, and on-chip stimulation capabilities, ensuring precise signal acquisition with minimal power consumption while maintaining a low noise level of 5.5 {micro}Vrms. We further developed an interconnect strategy using gold bump bonding, which allows for high-density integration of the flexible probe and rigid chip. We demonstrate the capacity of this platform through the integration with a flexible ECoG array. The resulting device allows for the high-resolution mapping of in vivo field potentials on the cortical surfaces of rat brains, supported by the precise localization of evoked sensory activities. These results prove an effective approach towards highly integrated neural interfaces with applications in brain-computer interfaces, neuroprosthetics, and large-scale functional brain mapping.

Drug-induced QT interval prolongation is a key biomarker of proarrhythmic risk and central to drug cardiac safety evaluation alongside in vitro assays and animal studies. Current preclinical frameworks, however, provide limited insight into how experimental uncertainty and extreme exposures translate into real-world arrhythmic risk, despite both factors critically modulating outcomes. To address this, we used sex-specific machine learning surrogate models trained on 3D cardiac digital twins--mechanistic electrophysiology models of anatomically detailed ventricles that integrate multichannel ion-channel block data. These emulators combine the realism of 3D simulations with high-throughput capability, enabling rapid, ethically unconstrained assessment of proarrhythmic risk. We illustrate the approach using loperamide, safe at therapeutic doses but linked to fatal arrhythmias at extreme exposures. Two analyses were performed: propagating experimental IC50 and Hill coefficient variability to quantify its effect on predicted QT prolongation and arrhythmic probability, and simulating extreme exposures to identify sex-specific arrhythmogenic thresholds. Experimental variability substantially broadened predicted QT prolongation and arrhythmic risk near decision thresholds. Extreme exposure simulations identified arrhythmogenic thresholds of approximately 107-109 x Cmax in female models and 213-286 x Cmax in male models. This framework offers a scalable, physics-based tool for early-stage drug cardiac safety evaluation.