Neusoft (China)

Accurate prediction of small interfering RNA (siRNA) silencing efficiency is critical for accelerating nucleic acid drug development. Over the past two decades, the field has transitioned from empirical sequence-derived heuristics to data-driven methods such as deep learning and graph neural networks. In this review, we systematically discuss current research from three integrated perspectives: data resources, design rules, and computational algorithms. Firstly, we summarize strengths, biases, and gaps in existing datasets, including low-throughput saturation scanning data, high-throughput panels, and specialized repositories for chemically modified siRNAs. Next, we outline the evolution of siRNA design principles, from classical guidelines such as GC content windows, terminal thermodynamic asymmetry, and positional base preferences, to more sophisticated thermodynamic-structural features incorporating target mRNA accessibility, and finally, to advanced clinical GalNAc-siRNA strategies involving seed destabilization and stereochemically purified backbones. Regarding computational algorithms, the progression is categorized into three stages: early linear regression models, traditional machine learning approaches, and modern deep learning frameworks, alongside various specialized algorithms tailored for chemically modified siRNAs. Current studies demonstrate steadily improving prediction accuracy, yet significant challenges persist, including insufficient data coverage for disease-relevant targets and hard-to-transfect cells, limited paired measurements of bare versus modified sequences, and inadequate integration of off-target effects and immunogenicity metrics. Moving forward, there is an urgent need to establish highquality, dose- and time-resolved datasets across diverse tissues, embed biophysical priors into interpretable deep-learning architectures, and adopt multi-task modeling approaches to optimize efficiency, safety, and delivery simultaneously. Collectively, these advancements promise to propel siRNA design beyond liver-centric applications towards precision therapeutics targeting multiple tissues and diseases.

Vessels encapsulating tumor clusters (VETC) serve as a crucial adverse prognostic indicator in hepatocellular carcinoma (HCC). This study aimed to develop and validate a Delta radiomics-based nomogram model on dynamic contrast-enhanced CT (DCE-CT) to predict VETC status and patient prognosis in HCC. A cohort of 222 patients from two centers with HCC undergoing DCE-CT scans and CD34 immunochemical staining was enrolled. Each liver lesion was segmented on intratumoral and peritumoral regions in the arterial phase (AP) and portal vein phase (PP) CT images. A total of 10,128 (1,688*6) radiomics features, including absolute and relative delta radiomics features, were extracted. Using four machine-learning algorithms, the features were trained and optimized (training set), and validated (internal and external test sets) to classify VETC patterns. Multivariable logistic regression incorporating signature scores and clinical predictors generated the nomogram. Model performance was evaluated through area under the curves (AUC) analysis, calibration curves, and decision curve analysis (DCA). The Kaplan-Meier survival analysis was used to assess recurrence-free survival (RFS) in the VETC+ and VETC- patients. The logistic regression-based nomogram incorporating three radiomic signatures and two clinical factors showed powerful predictive ability in internal and external test sets with AUCs of 0.854 and 0.803, respectively. The calibration curves, DCA showed favorable predictive performance of the nomogram. Patients classified as high-risk by the nomogram exhibited significantly shorter RFS compared to low-risk counterparts (P < 0.001). The developed nomogram demonstrated clinical translatability in preoperative VETC prediction and recurrence risk stratification, providing a potential imaging biomarker for guiding personalized therapeutic strategies in HCC management.

The online vectorized map construction technique employs a neural network model to forecast the vectorized representation of a specific region around automobiles, using data obtained from sensors mounted on automobiles. Due to advances in end-to-end object detection with transformers framework, the research on query-based online mapping has attracted substantial attention. However, the fixed number of queries and the random initialization of query embeddings constrain the model's performance. Moreover, the transformer architecture for object detection is based on the assumption that queries are identically distributed and independent, a premise that is not entirely applicable to map point queries which possess established subordinate relationships with map element instances. To address these issues, we initially incorporate a simplified transformer layer that utilizes semantic priors in bird's-eye view features for query initialization. The queries are then sent to a transformer-based map decoder for optimization and combined with a dynamic query management mechanism to eliminate low-confidence queries, hence maintaining computational efficiency. Furthermore, to guarantee that point queries within each instance preserve a consistent representation and avoid feature confusion among map element instances, we proposed an instance internal consistency map decoder. We conduct extensive experiments on commonly used map construction datasets to evaluate the proposed method. The experimental results demonstrate that our proposed method achieves state-of-the-art performance on the nuScenes and Argoverse 2 datasets.

Digital subtraction angiography (DSA) devices guide procedures across numerous diseases, performed on more than 100,000 patients daily worldwide. However, these procedures expose patients and healthcare providers to radiation, increasing the risk of health issues. Despite many low-dose DSA imaging methods proposed, none have been prospectively clinically validated. In this study, 46,829 patients (over 5 million DSA images) from 70 centers were used to iterate our previously developed generative artificial intelligence system (named GenDSA-V2). A total of 1,068 patients (533 in intervention arm and 535 in control arm), with suspected cerebral aneurysms (n = 435), lung cancer (n = 417) or advanced liver cancer (n = 216), meeting surgical criteria, were enrolled to validate the GenDSA-V2. The primary outcome was radiation dose, while secondary outcomes included efficiency, operation time and intraoperative complications. Group assignments were blinded to patients, surgeons and investigators, while technicians were aware but not involved in data collection or analysis. The GenDSA-V2 group showed substantially reduced radiation exposure, with an air kerma (AK) of 151.3 ± 125.1 mGy compared to 457.4 ± 407.4 mGy in the standard clinical protocols (SCP) group (mean difference = -306.1 mGy, 95% confidence interval (CI) = -342.3 to -269.9, P < 0.001 for superiority) and a dose-area product (DAP) of 4009.7 ± 2767.9 μGy m2 versus 12531.6 ± 9145.9 μGy m2 (mean difference = -8521.9 μGy m2, 95% CI = -9333.1 to -7710.7, P < 0.001 for superiority). Mean operation time was 33.1 ± 10.8 min in the SCP group and 34.8 ± 11.8 min in the GenDSA-V2 group (mean difference = 1.7 min, 95% CI = 0.3 to 3.1, P < 0.001 for noninferiority). Complication rates were similar (SCP = 8.1%, GenDSA-V2 = 7.5%, mean difference = -0.6%, 95% CI = -3.8% to 2.6%, P < 0.001 for noninferiority). The GenDSA system reduces radiation exposure to both physicians and patients by approximately two-thirds during DSA-guided procedures, demonstrating substantial clinical and translational value. Chinese Clinical Trial Registry: ChiCTR2400084789 .

This paper focuses on the application of large language models (LLM) in intelligent customer service and personalized recommendation in e-commerce. By constructing an intelligent customer service system that integrates retrieval-enhanced generation (RAG) and domain fine-tuning, as well as a personalized recommendation framework based on semantic understanding, the paper systematically explores the mechanism by which LLM enhances the quality of e-commerce services. Experimental results show that, compared with traditional methods, the accuracy of intent recognition in intelligent customer service based on LLM has increased by 23.6%, and user satisfaction has reached 91.2%; the NDCG@10 indicator of personalized recommendation has increased by 18.3%, and the satisfaction of recommendation explanations has increased by 32.5%. The research confirms that LLM can effectively address issues such as semantic understanding deviations and insufficient recommendation adaptability in e-commerce services, providing theoretical and practical references for the construction of a collaborative intelligent e-commerce service system.

Individualized postoperative management of colorectal ovarian metastases demands precision medicine tools, yet current approaches lack consideration of prognostic heterogeneity and targeted therapy benefit guidance and suffer from high costs and long turnaround times of genetic testing. In this retrospective, prospective multicohort study, we developed and validated an interpretable transformer-based transfer learning model to predict patient prognosis, targeted therapy benefits and molecular mutations by integrating digital pathology with RNA data. The performance of the model was assessed with the AUC, accuracy, sensitivity, specificity, PPV and NPV. The model accurately predicted peritoneal recurrence, with AUCs of 0.90, 0.74, and 0.83 across patient cohorts. It also achieved precise prognostic stratification for peritoneal recurrence-free survival in the training (HR=107.22, 95% CI 24.18-475.52; p<0.001), external test (HR=4.97, 95% CI 1.16-21.37; p=0.03), and prospective test (HR=10.53, 95% CI 2.02-54.94; p=0.01) sets. The model revealed a significant association between prognostic stratification and tumor microenvironment heterogeneity (p < 0.05), thereby enhancing its biological interpretability. Further analysis revealed that only patients classified as high risk with BRAF/RAS mutations could benefit from the addition of targeted therapy to adjuvant chemotherapy (HR 0.38, 95% CI 0.18-0.79; p=0.007). Moreover, the model predicted BRAF/RAS mutations with AUCs of 0.96/0.94 in the training set, maintaining cross-cohort generalizability with AUCs of 0.64-0.83. This pathobiology-based deep learning model can robustly detect prognosis and mutation and identify targeted therapy beneficiaries, serving as a potential precision tool in clinical decision-making for the management of colorectal ovarian metastases.

Background Accurate prediction of postoperative pancreatic fistula (POPF) after radical distal pancreatectomy (DP) is critical for optimizing surgical strategies. The objective of this study was to develop and validate a deep learning-based framework for three-dimensional (3D) body composition analysis to forecast POPF in patients with pancreatic cancer (PC). Methods A retrospective analysis was conducted on patients who underwent radical DP at two institutions between 2015 and 2022. A deep learning-based 2D and 3D segmentation model was developed to assess abdominal muscles and fat using preoperative computed tomography (CT) images. Predictive models for POPF were constructed using Gradient Boosting Decision Trees (GBDT) and their performance was evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA) on validation and external test datasets. Results The study comprised 230 patients with PC, with a mean age of 62 years. The incidence of POPF was observed to be 47.4% (Grade B/C). In terms of muscle segmentation, the Dice similarity coefficients (DSCs) for the testing set varied between 91.84 and 98.41% across different regions of the abdominal musculature. For the segmentation of visceral and subcutaneous adipose tissue, the DSCs in the testing set were 97.10 and 98.57%, respectively. The integrated clinical and imaging-based POPF prediction model demonstrated superior performance, achieving an AUC of 0.82, with a sensitivity of 0.81 and a specificity of 0.76 in the external test set. Conclusion The implementation of a deep learning-based 2D and 3D body composition analysis pipeline exhibited high accuracy in predicting POPF following radical DP for PC. However, further validation in larger, multicenter cohorts is required to confirm the generalizability of these findings.

Social cognition develops through a complex interplay between neural maturation and environmental factors, yet the neurobehavioral mechanisms underlying this process remain unclear. Using a naturalistic fMRI paradigm, we investigated the effects of age and parental caregiving on social brain development and Theory of Mind (ToM) in 34 mother-child dyads. The functional maturity of social brain networks was positively associated with age, while mother-child neural synchronization during movie viewing was related to dyadic relationship quality. Crucially, parenting and child factors interactively shaped social cognition outcomes, mediated by ToM abilities. Our findings demonstrate the dynamic interplay of neurocognitive development and interpersonal synchrony in early childhood social cognition, and provide novel evidence for neurodevelopmental plasticity and reciprocal determinism. This integrative approach, bridging brain, behavior, and parenting environment, advances our understanding of the complex mechanisms shaping social cognition. The insights gained can inform personalized interventions promoting social competence, emphasizing the critical importance of nurturing parental relationships in facilitating healthy social development.