Drug discovery faces formidable challenges including high technology, high costs, substantial risks, and prolonged development timelines, necessitating disruptive technologies capable of systematically improving efficiency, enhancing predictive accuracy, and reducing failure rates. Artificial intelligence (AI) agent-an emerging intelligent paradigm powered by large language models-holds significant potential to transform the entire drug development pipeline. Their core capability lies in performing autonomous reasoning, planning, and tool utilization directed at complex scientific objectives, thereby integrating and orchestrating multiple research stages and transitioning AI from a mere "tool" to an "active collaborator". Through knowledge integration and hypothesis generation, AI agents can identify underexplored therapeutic targets and novel mecha-nisms of action. In parallel, they can automate complex tasks such as molecular design, optimization, and synthesis planning, and further close the loop between virtual design and physical experimentation by interfacing with automated experimental platforms. Moreover, AI agents are evolving toward higher-level paradigms, including the development of integrated drug design platforms and general-purpose biomedical agents. This review systematically summarizes the core architectures of AI agents, highlights their applica-tions across key stages of drug development, and discusses current limitations along with future directions, providing a reference for researches in related fields.
To address the core limitations of conventional insulin therapy, including delayed glycemic control and the frequent risk of hypoglycemia, the development of glucose-responsive insulin delivery systems capable of dynamically sensing blood glucose levels and releasing insulin on demand has emerged as a pivotal strategy. Based on their underlying sensing mechanisms, these systems are generally classified into three categories: those utilizing glucose oxidase, glucose-binding molecules, and phenylboronic acid. From the perspective of administration routes, injectable and transdermal delivery are the two primary approaches for glucose-responsive insulin. Injectable glucose-responsive insulin delivery systems are highly compatible with existing clinical practices, primarily relying on glucose-responsive carriers to regulate the insulin release rate and achieve stable and efficient bioavailability. Transdermal glucose-responsive insulin delivery systems utilize glucose-responsive microneedle arrays to penetrate the skin stratum corneum and precisely control the rate of insulin release, allowing for sufficient insulin delivery under almost painless conditions. This review systematically summarizes recent advances in both injectable and transdermal glucose-responsive insulin delivery systems, with a focus on carrier design strategies, glucose-responsive release mechanisms, and evolutionary pathways of preparation techniques. It also highlights the contributions of these systems toward improved glucose-responsiveness, therapeutic safety, biocompatibility, and patient adherence. Furthermore, challenges and future prospects for clinical translation are discussed. This overview is expected to provide valuable insights for further research and development in this field.
Objectives: To develop a machine-learning model that integrates routine clinical parameters with tumor mutational burden (TMB) and to evaluate its performance in predicting responses to programmed death-1 (PD-1)/programmed death-ligand 1(PD-L1) inhibitors across various cancer types.
Methods: We conducted a retrospective study of 146 patients with advanced solid tumors who were treated with PD-1/PD-L1 inhibitors. The cohort was randomly divided into a training set (n=116) and a validation set (n=30) at a 4:1 ratio. Using the PyTorch framework, we constructed a neural network model (designated NNT9) incorporating age, sex, body mass index (BMI), TMB, history of systemic therapy, neutrophil-to-lymphocyte ratio (NLR), and other routine blood parameters. The model employed a multilayer perceptron architecture. Hyperparameters were automatically optimized using AutoGluon, and the model was refined via 5-fold cross-validation. SHapley Additive exPlanations (SHAP) was used to perform feature importance analysis on the optimal model in the training set. Predictive performance was compared against TMB alone using metrics including the area under the receiver operating characteristic curve (AUC), accuracy, F1 score, sensitivity, and specificity. Confusion matrices were generated, and the association between model-predicted response groups and progress free survive (PFS) was analyzed.
Results: NNT9 was identified as the optimal model, and the history of systemic therapy, TMB, platelet count, and BMI were the four most important predictive features. NNT9 achieved AUCs of 0.949 and 0.851 in the training and validation sets, respectively, outperforming TMB alone (AUCs: 0.747 and 0.720). In the validation set, NNT9 also demonstrated superior sensitivity (0.571), accuracy (0.867), F1 score (0.667), positive predictive value (0.800), and negative predictive value (0.880). The confusion matrix revealed that NNT9 misclassified only half as many patients as TMB alone in the validation set. Kaplan-Meier analysis showed that patients predicted to be responders by NNT9 had significantly longer PFS than non-responders in both training and validation sets (both P<0.01).
Conclusions: The NNT9 model, which integrates readily available clinical parameters with TMB, represents an accurate and clinically feasible tool for predicting immunotherapy benefit in a pan-cancer cohort, and shows promise for clinical translation.
Objectives: To investigate the relationship between hematocrit (HCT) and low resting oxygen saturation in Tibetan plateau population.
Methods: This retrospec-tive cohort study included 3075 Tibetan adult inpatients who underwent elective non-cardiac, non-thoracic surgery at Tibet Autonomous Region People's Hospital between January 2023 and October 2024. Multivariate logistic regression was used to assess the association between HCT and low resting oxygen saturation (SpO2<88%). Restricted cubic splines were employed to model non-linear relationships, and piecewise logistic regression, in combination with maximum likelihood estimation, was used to identify the HCT threshold.
Results: Multivariate logistic regression analysis showed that elevated HCT was an independent risk factor for low oxygen saturation in the patients. For every 5-percentage-point increase in HCT, the risk of low oxygen saturation increased by 12% (P<0.01). The restricted cubic spline model revealed a significant U-shaped association between HCT and the risk of low oxygen saturation (P<0.01). The HCT threshold for the total population was 42.8%. When HCT was below 42.8%, each 5% increase in HCT was associated with a 22% reduction in the risk of low oxygen saturation in the patients (OR=0.780, 95%CI: 0.619-0.982, P<0.05). When HCT was 42.8% or above, the risk increased by 27.2% (OR=1.272, 95%CI: 1.132-1.429, P<0.01). Stratified analysis indicated that the HCT threshold was 40.4% for female patients and 46.6% for male patients.
Conclusions: There is a U-shaped association between HCT and low resting oxygen saturation in the Tibetan plateau population. An HCT level exceeding the population-specific threshold suggests that the adaptive erythropoietic response to high-altitude hypoxia may transition from being compensatory to potentially detrimental.
Objectives: To systematically evaluate the performance of generative artificial intelligence (GenAI) models, DeepSeek-V3 and the Qwen3 series, in the differential diagnosis of weight loss.
Methods: A search was conducted in the PubMed database for all case reports published in the American Journal of Case Reports between January 1, 2012 and June 2, 2025, containing the term "weight loss" in the title or abstract. Two senior general practitioners independently reviewed each case to determine whether it met predefined diagnostic criteria for weight loss (emaciation). Cases that did not meet these criteria, had incomplete information, or involved clearly defined specialty-specific diagnoses and treatments were excluded. The remaining cases were then compiled into standardized clinical case summaries. These summaries were presented to DeepSeek-V3 and the Qwen3 series models (Qwen3-235B-A22B, Qwen3-30B-A3B, and Qwen3-32B) to generate ranked lists of the top 10 differential diagnoses. The models were not specifically fine-tuned for this task. Sensitivity, precision, and F1-score were used to evaluate performance. Intergroup comparisons were performed using McNemar's test and Cochran's Q test.
Results: A total of 87 case were analyzed. DeepSeek-V3 demonstrated better performance than Qwen3-235B-A22B in sensitivity, precision, and F1-score, especially at the Top5 level (P=0.043). Among the Qwen3 series models, Qwen3-235B-A22B showed the best performance in sensitivity, precision, and F1-score for the Top1 diagnosis, but the differences among the three Qwen3 models across all diagnostic levels were not statistically significant (all P>0.05).
Conclusions: Domestic GenAI models exhibit a characteristic of "breadth over precision" in the differential diagnosis of weight loss, with DeepSeek-V3 performing better at key diagnostic levels. Although the sensitivity and precision for the top-ranked diagnosis require improvement, these models have the potential to serve as effective clinical decision support tools, broadening the diagnostic perspectives of general practitioners.
Spinal cord injury (SCI) is a major cause of motor disability. Epidural electrical stimulation (EES) has emerged as a promising neuromodulation technique and has been extensively investigated in recent years for promoting functional recovery after SCI. This case report describes a patient who sustained a thoracic SCI four months ago due to a fall from height. Following surgical intervention and conventional rehabilitation, the patient's lower limb muscle strength remained at grade C on the American Spinal Injury Association (ASIA) impairment scale. After implantation of a closed-loop spinal neural interface for EES, the patient underwent a regimen of continuous stimulation with spatially targeted configurations and multimodal rehabilitation training. This intervention led to progressive recovery, including the achievement of voluntary single-joint movements, independent standing, and assisted walking. At the 16-week follow-up, the patient's ASIA grade improved from C to D. Improvements were also noted in sensory and autonomic functions in addition to motor recovery. The overall rehabilitation outcomes were substan-tially better than those typically reported for chronic-phase SCI patients in the literature. Importantly, no implant-related infections or neurological complications were observed. These results indicate that the combination of closed-loop spinal neural interface-enabled EES and structured rehabilitation holds significant potential as a therapeutic strategy.
Objectives: To evaluate the feasibility and effectiveness of replacing manual auscultation with artificial intelligence (AI)-assisted cardiac auscultation within the dual-indicator screening strategy for early detection of congenital heart disease (CHD) in neonates in a real-world clinical setting.
Methods: Using data from the provincial CHD treatment network led by the Children's Hospital of Zhejiang University School of Medicine, we retrospectively enrolled 41 320 neonates born between July 2020 and March 2023. All neonates underwent pulse oximetry (POX), AI-assisted auscultation, and manual auscultation. The traditional screening strategy was defined as "POX+manual auscultation," and the AI-assisted strategy as "POX+AI-assisted auscultation". A positive screening result was defined as a positive finding in either POX or the corresponding auscultation method (manual or AI). Echocardiography served as the gold standard for CHD diagnosis. True positive, false positive, true negative, and false negative results were determined, and the missed-diagnosis rate, sensitivity, specificity, Youden index, positive predictive value (PPV), negative predictive value (NPV), and diagnostic accuracy were calculated. To assess the consistency of screening performance across institutions, analyses were also conducted separately in the four hospitals with the highest screening volumes.
Results: A total of 354 neonates were diagnosed with CHD by cardiac ultrasound. Compared with the traditional strategy, the intelligent strategy significantly reduced the missed-diagnosis rate (67.23% vs. 34.75%, P<0.01), increased sensitivity (32.77% vs. 65.25%, P<0.01), NPV (99.41% vs. 99.67%, P<0.01) and Youden index (30.04% vs. 55.15%, P<0.01). However, specificity, PPV, and diagnostic accuracy were lower (all P<0.01). Across the four high-volume institutions, the intelligent strategy consistently showed a significant reduction in missed-diagnosis rate and increased sensitivity, NPV and Youden index, along with decreased specificity, PPV and diagnostic accuracy, indicating robust and reproducible performance across diverse clinical settings.
Conclusions: In multicenter real-world practice, the intelligent screening strategy significantly reduces the missed-diagnosis rate of CHD and demonstrates stable screening performance across different institutions, suggesting AI-assisted auscultation is a feasible and clinically valuable alternative to manual auscultation in neonatal CHD screening.
Elbow orthoses offer a straightforward mechanical approach for rehabili-tation following elbow joint injuries. The convergence of artificial intelligence with these orthoses has led to the development of elbow rehabilitation robots, designed to address the personalized rehabilitation requirements of patients with diverse injury profiles. A typical elbow rehabilitation robot system comprises four core components: rigid mechanical structure, an actuation system, bionic sensors, and integrated software. The rigid structure, analogous to the human skeletal system, includes linkage mechanisms and gear transmission assemblies. The actuation system, mimicking the function of muscles and ligaments, generates and modulates the necessary forces and torques for movement, employing actuators such as pneumatic, elastic and cable-driven types. Bionic sensors, serving as the robot's perceptual interface, encompass photoelectric encoders, force/torque sensors, electromyo-graphic (EMG) signal sensors, and temperature sensors. The software system, encompassing control algorithms and machine learning models, functions as the "neural center," enables intelligent decision-making and motion control. The core technological achievement lies in the seamless integration of hardware and software to enable precise tracking of elbow joint kinematics and adaptive modulation of assistive forces based on real-time human-robot interaction. This integration supports multiple training modalities, including passive, assistive, active, and resistive modes and enables safe, personalized, and intelligent rehabilitation support across different recovery phases. By harnessing technologies like bio-inspired design, precise impedance control, EMG-based assistance, and virtual reality-integrated task training, these robotic systems improve training comfort, assistance accuracy, and patient adherence. This review outlines the current state of elbow rehabilitation robotics, details the key system components and primary training modalities, discusses clinical needs and future development trends, and aims to offer insights for the further refinement of rehabilitation robotic systems.
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：