Background: Diabetic retinopathy (DR) is a leading cause of vision loss in working-age adults. Ferroptosis, an iron-dependent form of programmed cell death, contributes to DR pathogenesis. Baicalein, a natural flavonoid, exhibits antioxidant and anti-inflammatory effects and is a potential ferroptosis inhibitor. Methods: We applied a network pharmacology approach to explore Baicalein's therapeutic potential in DR. Differentially expressed genes (DEGs) were identified from GSE102485, and ferroptosis-related genes (FRGs) were obtained from FerrDb. Baicalein-related targets were retrieved from TCMSP and PharmMapper. Protein-protein interaction (PPI) networks were constructed using STRING and Cytoscape. GO and KEGG pathway enrichment analyses were performed. Single-cell RNA sequencing (scRNA-seq) data from GSE178121 were analyzed to determine core target expression across retinal cell types. Molecular docking and molecular dynamics simulations assessed Baicalein's binding affinity and stability with key targets. Results: A total of 4,279 DEGs and 120 ferroptosis-related DEGs were identified, with 21 overlapping targets among DR, ferroptosis, and Baicalein; eleven core targets were selected based on network topology. Enrichment analysis revealed involvement in critical DR-related pathways, including HIF-1A, TNF, PI3K-Akt, and MAPK signaling. scRNA-seq highlighted broad HIF1A expression across retinal cells, particularly in bipolar cells and rods. Molecular docking showed favorable binding with PPARG and ALB, and dynamics simulations indicated stable interactions, suggesting Baicalein can effectively modulate ferroptosis pathways. Conclusions: Integrating network pharmacology and single-cell transcriptomics, this study identifies Baicalein as a promising candidate for DR therapy by targeting ferroptosis. These findings support further preclinical and clinical investigation to protect diverse retinal cell populations from ferroptotic damage.
Inertial measurement units (IMUs) are low-cost, wearable sensors that can estimate body segment orientation by tracking relative sensor orientations. This review aimed to synthesize and evaluate studies investigating the accuracy and reliability of IMUs in measuring shoulder kinematics for clinical application in patients with musculoskeletal injuries. Shoulder kinematics were chosen due to their importance in assessing upper extremity function, performing overhead activities, and the increasing demand for objective, accessible motion-tracking tools in clinical settings. Studies within PubMed/MEDLINE, Scopus, Cochrane Central Register of Controlled Trials, IEEE Xplore, and Google Scholar were screened for eligibility. They were selected based on the following inclusion criteria: (1) application of inertial sensors to assess motions, (2) sensors used accelerometers and gyroscopes or similarly functioning technologies, (3) sensors applied to shoulders, (4) studies published from 2011 to 2024, (5) studies written in English, (6) studies found in peer reviewed, original research articles, (7) studies with full text available. Of 1900 articles identified in our initial literature search, 49 were included. Articles were excluded based on these criteria: (1) Reviews, systematic reviews, or meta-analyses, (2) Studies without ethical approval, (3) Animal or cadaveric studies, (4) Studies prior to 2011. A data extraction was included with key findings of each article. The Consensus-based Standards for the selection of health Measurement Instruments (COSMIN) quality assessment tool was used to assess each article's risk of bias. We compared outcome metrics across studies quantifying IMU accuracy and reliability, including root mean square error (RMSE) and intraclass correlation coefficient (ICC), respectively. IMU-based shoulder kinematics exhibited a wide-range of RMSEs (<1° - 12°) and ICCs (0.32 - 0.98) depending on the motion and number of sensors used. Overall, there was a tolerable RMSE (between 5-10°; mean = 7.10 ± 3.97) and good ICC (>0.75; mean = 0.810 ± 0.145) across studies for 6.77 IMUs on average. The goal of this review was to assess the current IMU use in upper extremities, identify factors preventing clinical use, and inform future research. More IMU-based clinical studies are needed to understand shoulder pathology motor deficits. Additional validation studies are needed to demonstrate IMU efficacy when paired with other technologies.
Human telomerase reverse transcriptase (hTERT) plays a key role in cancer cell immortalization and represents an important therapeutic target for anticancer drug discovery. In this study, a computational pipeline combining genetic algorithms (GAs) and machine learning (ML) was developed to design and screen garlic-derived bioactive compounds as potential telomerase inhibitors. Garlic phytochemicals were used as the initial chemical space, which was iteratively evolved through mutation and fragment expansion to generate novel ligand candidates. A multi-parameter fitness function incorporating Lipinski, Veber, and Ghose drug-likeness rules, quantitative estimate of drug-likeness (QED), hydrogen bond donor/acceptor balance, and aromatic ring constraints was used to guide optimization. In addition, a RandomForest-based classifier was applied to pre-screen compounds for predicted telomerase activity prior to molecular docking. The shortlisted ligands were evaluated using CB-Dock2, and further assessed for pharmacokinetic and toxicity properties using SwissADME, including solubility, lipophilicity, and bioavailability. Out of 125 generated ligands, 14 met both drug-likeness and predicted activity criteria and progressed through the full pipeline. The highest binding affinity observed was -10.5 kcal/mol; however, this top-scoring compound was excluded due to ADMET rule violations. The remaining candidates exhibited favorable physicochemical properties, acceptable solubility, balanced lipophilicity, and good predicted oral bioavailability. Overall, the results demonstrate that genetic algorithms can efficiently generate structurally diverse and pharmacologically relevant scaffolds, and that integrating an activity-based machine learning filter prior to docking improves screening efficiency by prioritizing biologically meaningful candidates. Compared with traditional GA-based docking workflows, this integrated strategy provides a more selective and cost-effective approach for early-stage ligand discovery.
This research offers an in-depth analysis of electro-osmotically induced, immiscible bio-convective flow of urine incorporated with five different nanoparticles (NPs) and actively motile microorganisms within a bio-reactive microchannel, subjected to electromagnetic field effects. The model accounts for critical multiphysical effects including Joule heating, electromagnetic radiation, Hall and ion-slip currents, and interfacial nanolayer (NL) interactions. The results reveal that an increase in interfacial NL thickness amplifies the urine velocity. An artificial neural network (ANN)-based model is further implemented to predict SFC with remarkable precision, achieving a minimal error of 0.01% and demonstrating excellent agreement with analytical outcomes. This research looks at how urine moves through very tiny channels inside the body when electric and magnetic forces are applied. The urine in this study contains very small particles called nanoparticles (NPs), made from materials, such as gold, silver, carbon, and titanium. These particles can improve the way heat and substances move within fluids. The study also includes living microorganisms that can move on their own, making the flow more realistic and related to natural biological processes. Mathematical equations were used to describe how the urine flows and how factors, such as electricity, magnetism, heat, and rotation affect its movement. A computer-based learning method, known as artificial intelligence (AI), was used to check and predict the results. The AI model was highly accurate and matched well with the mathematical results. The study found that a thin coating around NPs helps the urine move faster, while stronger magnetic or rotational forces slow it down. Changes in heat and chemical reactions also affect how microorganisms spread in the fluid. These findings can help in designing new medical technologies, such as lab-on-a-chip devices for urine testing or targeted drug delivery systems. In simple terms, this work improves our understanding of how body fluids containing NPs behave under electric and magnetic fields and may support the development of faster and more efficient biomedical diagnostic tools.
EEG signal reliability in biomedical applications may be affected by ocular artifacts resulting from blinks and eye movements. Existing methods often struggle to remove these artifacts effectively. In order to overcome this constraint, we suggest a new framework that integrating Skewness-based Discrete Wavelet Transform (SDWT) with the Swarm Decomposition (SwD) algorithm. The SDWT separates EEG signals into artifact and non-artifact segments. The artifact segment undergoes SwD processing, which decomposes it into hidden components. The EOG (electrooculographic) component is then analyzed using energy and skewness metrics. The level corresponding to the artifact identified by its highest energy and skewness is discarded, and the remaining elements are recombined with the non-artifact segment. To automate EOG identification, the K-means clustering algorithm is employed. We evaluate the filtering performance of the proposed method using metric: Mean absolute error in the power spectrum (MAE-PS), in power spectrum analysis across the delta(δ), theta (θ), alpha (α), and beta (β) frequency bands. Testing was conducted on two semi-simulated and one real-time datasets: Publicly available EEG signals contaminated by eye-movement and eye-blink artifact. Wearable EEG recordings from epileptic and healthy persons performing physical activities (climbing stairs, sitting, running, and walking). Results demonstrate that our method outperforms existing techniques, achieving: Lower MAE-PS in power spectrum analysis across α,β, and θ bands, indicating better signal preservation. This approach shows promise for enhancing EEG reliability in both clinical and mobile settings.
This paper introduces a mathematical model for the growth of transactive response DNA binding protein of 43 kDa (TDP-43) inclusion bodies in neuron soma. The parameter representing the accumulated neurotoxicity caused by misfolded TDP-43 oligomers is also introduced. The model's equations enable the numerical calculation of the concentrations of TDP-43 monomers, dimers, free oligomers, and oligomers deposited in inclusion bodies. By simulating the deposition of free oligomers into inclusion bodies, the model predicts the size of TDP-43 inclusion bodies. An approximate solution to the model equations is derived for the scenario where protein degradation machinery is dysfunctional, leading to infinite half-lives for TDP-43 dimers, monomers, and both free and deposited oligomers. This solution, valid at large times, predicts that the radius of the inclusion body increases proportionally to the cube root of time, whereas the accumulated neurotoxicity increases linearly with time. To the best of the author's knowledge, this study is the first to model the relationship between the size of TDP-43 inclusion bodies and time, and the first to introduce the concept of accumulated neurotoxicity caused by misfolded TDP-43 oligomers. Sensitivity analysis of the approximate solution indicates that the inclusion body radius and accumulated neurotoxicity become independent of the kinetic constants at large timescales. Unlike the case of infinite half-lives, the numerical solution for physiologically relevant (finite) half-lives demonstrates that the long-term behavior of the inclusion body radius and accumulated neurotoxicity remains dependent on the kinetic constants, converging to distinct curves over time.
Despite their classification as "non-lethal," Kinetic Energy Non-Lethal Projectiles (KENLPs) still causing fatal injuries, necessitating rigorous biomechanical assessment. Ethical and technical limits of Post Mortem Human Subjects (PMHS) and experimental testing have elevated Finite Element Human Body Models (HBMs) for blunt trauma research. This study develops a two-stage numerical framework for cranial and thoracic injury prediction using LS-DYNA. Firstly, Hybrid III (H3) sub-models are validated against the Ballistics Load Sensing Headform (BLSH) envelope and NATO STANREC 4744 (AEP-99) thorax guidelines using validated KENLPs. Building upon these validations, the second stage proposes the Simplified Head (SH-FEM), featuring a dual-layer scalp and skull architecture, and the Simplified Thorax (STh-FEM)-a three-layer construct comprising muscle, a lung slab, and a central skeletal structure preserving dominant load paths while reducing computational cost. Simulation results indicate peak forces scale nonlinearly from 0.82 to 16.03 kN across 20-80 m⋅s⁻¹, with neck coupling reducing peaks by 20-30%. A velocity inflection at 40 m⋅s⁻¹ marks sub-concussive-to-injurious transitions: <33 m⋅s⁻¹ yields <2.1 kN (insignificant risk), while >55 m⋅s⁻¹ exceeds 7.5 kN (fracture/coma). For thoracic impacts across Cases A-E, VCmax-based AIS ≥ 2 risks vary by model and projectile. In Case E, STh-FEM predicted 46% risk versus H3 at 91%; in Case C, values were 10% and 52%, respectively. Furthermore, STh-FEM overpredicted rigid PVC projectile forces (98% in Case C) but matched the deformable SIR-X projectile (∼8 kN peaks). These simplified models demonstrate controlled, reproducible responses, confirming their feasibility as performant alternatives for rapid KENLPs design screening and safety assessment.
Pedestrian injury risks in car-to-pedestrian collisions are strongly influenced by anthropometric characteristics, yet existing human body models rarely represent small-stature Chinese female pedestrians. This study presents the Tianjin University of Science and Technology Injury Bionic Model (TUST IBMs F05-P), developed to represent a 5th percentile Chinese female pedestrian. Detailed anatomical structures were reconstructed directly from medical imaging data without geometric scaling, preserving subject-specific anatomical geometry, and the model was meshed predominantly with hexahedral elements. A representative walking posture was defined, and the model was evaluated through a certification procedure conducted according to the Euro NCAP CP540 pedestrian human body model certification protocol. Simulation results showed that key biomechanical indicators, including Head Impact Time (HIT), contact forces, and kinematic trajectories, predominantly fell within the response corridors specified in CP540. Quantitative assessment using the CORA (CORrelation and Analysis) method defined in ISO/TS 18571:2024 yielded an overall score of 0.84, indicating a high level of correlation with the CP540 reference corridors. The certification results indicate that the TUST IBMs F05-P produces stable and reproducible responses under the tested impact conditions. By providing an anatomically realistic representation of a small-stature Chinese female pedestrian, this model addresses the lack of population-specific pedestrian models and offers a validated basis for pedestrian injury analysis and vehicle front-end safety evaluation.
Pulmonary hypertension (PH) is a progressive cardiopulmonary disorder with high mortality, necessitating non-invasive methods for early detection and treatment evaluation. In this paper, this study proposes a novel machine learning model for non-invasively identifying the therapeutic effects of Baicalin in PH using routine hematological indicators. The core innovation is an enhanced Bat Algorithm (BA) variant, termed RGBA, which integrates an Elite-based Random Walk Strategy (ERWS) and an Elite Guided Strategy (EGS) to achieve a superior balance between exploration and exploitation. And the RGBA demonstrated significantly improved global search capability in IEEE CEC 2014 benchmark tests, outperforming several state-of-the-art meta-heuristic algorithms. Subsequently, a binary version of RGBA (bRGBA) was developed and combined with a Kernel Extreme Learning Machine (KELM) classifier within a wrapper-based feature selection framework, forming the bRGBA-KELM model. Applied to a dedicated PH dataset from murine models, bRGBA-KELM achieved a prediction accuracy of 97.43% via ten-fold cross-validation, outperforming nine comparable hybrid models. Critically, it identified four key blood biomarkers-Red Blood Cell count (RBC), Hemoglobin (HGB), Mean Corpuscular Volume (MCV), and Hematocrit (HCT)-that are mechanistically linked to PH pathogenesis and modulated by Baicalin treatment. In conclusion, the proposed RGBA offers a robust optimization tool, while the bRGBA-KELM model provides a clinically viable, non-invasive technical reference for early PH prediction and therapeutic assessment.
Cancer chemotherapy scheduling presents a significant optimization challenge: it aims to minimize the tumor burden while adhering to toxicity and pharmacokinetic constraints. This study employs a bang-bang optimal control framework applied to a nonlinear cancer chemotherapy model with state constraints. The model incorporates pharmacodynamic parameters and cumulative toxicity limits and is numerically solved via a high-resolution discretization approach in the AMPL modeling environment with IPOPT. The proposed method yields a final tumor size of 9.9466×103, demonstrating a 32.8% improvement over previous optimization techniques. We also investigated the role of time-dependent tumor reduction constraints and performed a sensitivity analysis on key biological parameters, such as the tumor growth rate, drug responsiveness, and biochemical clearance. The proposed framework, through its integration of parameter sensitivity analysis and constrained optimal control, provides a basis for adaptive and patient-specific chemotherapy scheduling that can dynamically adjust to individual tumor and pharmacokinetic profiles. These findings highlight the potential of optimal control methods to inform personalized chemotherapy regimens and suggest directions for clinical translation. However, further validation using real patient data is necessary to confirm the robustness and applicability of the proposed approach.
Mitral valve disease, particularly mitral regurgitation (MR), is among the most common valvular heart diseases, often requiring prosthetic valve replacement to restore normal hemodynamics and preserve myocardial function. Despite advances in prosthetic technologies, predicting patient-informed biomechanical performance remains challenging due to complex blood-structure interactions. This study aimed to develop a patient-informed fluid-structure interaction (FSI) model to evaluate the biomechanical and hemodynamic performance of a mechanical mitral valve prosthesis. A patient-informed FSI model of the left ventricle (LV) and mitral valve prosthesis was developed and validated using echocardiographic data and numerical simulations. The model analyzed the behavior of a St. Jude mechanical mitral valve implanted in a 46-year-old male patient and compared it with a healthy LV model. Hemodynamic and mechanical parameters, including flow velocity, shear stress, pressure gradient, and von Mises stress, were quantified. The prosthetic valve restored global flow direction, stroke volume, and ventricular wall displacement, consistent with echocardiographic data. In the healthy LV, peak inflow velocities reached 0.6-0.7 m/s during early diastole with transient shear stress (∼0.08 s). The prosthetic valve produced higher inflow velocities (up to 6.5 m/s) and prolonged shear stress (∼0.08-0.18 s), indicating potential hemolysis risk. Pressure gradients peaked at ∼13 mmHg, and stress analysis showed preserved central motion with localized stress near the prosthesis. Patient-informed FSI modeling enables accurate, noninvasive assessment of post-operative hemodynamics and mechanical performance, identifying risks such as elevated shear stress and hemolysis while supporting optimized prosthetic design and personalized surgical planning.
Sevoflurane dose-dependently suppresses respiratory activity, yet its molecular mechanisms remain incompletely understood. In this study, we integrated transcriptomic data from the Gene Expression Omnibus and GeneCards to identify differentially expressed genes associated with sevoflurane anesthesia and respiratory depression. A protein-protein interaction network was constructed, and hub genes were screened using the MCC, Betweenness, Closeness, and MCODE algorithms. Functional associations of these hub genes were further explored using the GeneMANIA platform. Potential therapeutic compounds were predicted through the Connectivity Map (CMap) database. The effects of sevoflurane on ICAM1 expression and the intervention of PD-98059 were experimentally validated in SH-SY5Y cells. Ultimately, nine hub genes were identified, including MMP9, CXCL1, IL1B, NF-κB1, CXCL8, IL6, CCL2, ICAM1, and VCAM1. These genes were mainly enriched in pathways related to inflammatory responses, immune regulation, and chemotaxis. Increased infiltration of dendritic cells, MHC class I molecules, and neutrophils was observed in the sevoflurane-treated group. PD-98059 was predicted as a potential therapeutic candidate and was confirmed to reverse sevoflurane-induced upregulation of ICAM1. Overall, this study identifies key genes and inflammatory pathways associated with sevoflurane-induced respiratory suppression, providing new insights into its molecular mechanisms and potential therapeutic strategies.
Although braided stents are widely adopted for intracranial aneurysm treatment, current designs still struggle with simultaneously satisfying the conflicting demands of mechanical properties including radial pressure, bending moment and foreshortening. This study proposed a multi-objective optimization framework based on an improved response surface model (IRSM) for the personalized design of braided stents, to enhance their adaptability to specific lesions. The IRSM was constructed for the aforementioned mechanical properties and integrated with design of experiments to accurately characterize intrinsic relationships between design parameters and mechanical properties. Based on 45 training samples, the IRSM demonstrated high accuracy in modeling the intrinsic relationships between design parameters and mechanical properties. Assessed using leave-one-out cross-validation, it achieved a coefficient of determination R2 > 0.9580 and a mean squared error MSE < 0.0410. Then, multi-objective optimization was implemented to obtain the Pareto-optimal solutions by the non-dominated sorting genetic algorithm II. Finite element simulations were performed separately for both the commercial stent pipeline and the selected Pareto-optimal solution. The comparative results revealed that this selected Pareto-optimal solution effectively increases radial pressure and reduces foreshortening, while inevitably leading to an increase in bending moment, thereby highlighting the importance of conducting coordinated trade-offs and optimizations of multiple objectives based on specific clinical scenarios in personalized stent design.
This study employs patient-specific computational fluid dynamics coupled with a Lagrangian discrete-phase model to quantify the influence of His angle and anastomotic caliber on intragastric hemodynamics after stomach-partitioning gastrojejunostomy (SPGJ). Three His angles (3°, 7°, 11°) and three stoma widths (narrow, intermediate, wide) were systematically analyzed. Within the physiological range, His angle altered global velocity and pressure by <5%, yet extreme values prolonged mean particle residence up to 1.9-fold, indicating a moderate angle (∼7°) minimizes stasis without elevating pressure loss. Stoma size proved decisive: very narrow or very wide orifices produced low-velocity seepage, high stagnation, and either excessive or moderate pressure drops, whereas an intermediate aperture yielded the lowest Δp and the shortest residence time. The results suggest that optimizing SPGJ requires maintaining a moderate His angle and designing an intermediate-sized anastomosis to maximize emptying efficiency while limiting tumor irritation and reflux risk. Clinically, these findings provide actionable guidance for intraoperative configuration and stoma sizing to balance functional patency with complication avoidance.
Malignant diseases remain one of the leading causes of death globally. Drug synergy has emerged as an effective approach for treating malignancy, offering improved therapeutic outcomes. Although techniques like clinical trials and high-throughput drug screening are commonly used to discover promising synergistic drug pairs, but they are time consuming and expensive. With the evolution of artificial intelligence, deep learning models are increasingly being applied to identify synergistic drug combinations. These models rely heavily on large-scale datasets, where size of dataset and hyperparameter selection play a pivotal role in the performance of model. However, determining the optimal hyperparameters for drug synergy models is a complex and time-intensive task, typically involving multiple iterative experiments. As a result, there is growing attention on optimizing hyperparameters for these models. This study emphasizes the role of hyperparameter optimization algorithms (HOAs) in evaluating how different optimization strategies and hyperparameter choices influence model effectiveness. By optimizing the hyperparameters, we achieved good accuracy in predicting drug synergy. Our findings highlight that the effectiveness of hyperparameter optimization is highly task- and dataset-dependent.
Electroencephalogram (EEG) reflect changes in the electrophysiological activity of the brain and can be used in the diagnosis of Alzheimer's disease (AD). Each EEG channel provides real-time information about the brain, while different EEG channels contain different information about the brain. Using all EEG channel data for AD diagnosis contains redundancy data, leading to increased computation and reduced data analysis accuracy. In this paper, a diagnostic method for AD based on Particle Swarm Optimization (PSO) EEG channel selection and Gated Recurrent Unit (GRU) is proposed. Using EEG energy as the fitness function and PSO to select EEG channels, the redundant information in EEG data is reduced and the accuracy of EEG data analysis is improved. GRU is a special kind of recurrent neural network (RNN) structure. It uses EEG data extracted by the principal component analysis (PCA) feature based on singular value decomposition (SVD) as input to the model. And it has a good advantage in analyzing the time series of EEG. The results show that the classification accuracy of the method in this paper reaches 0.9487, which is higher than the performance of other proposed methods. Compared to the results of using all EEG channel data analysis, the classification accuracy of this method was improved by 0.0757. It shows that the method proposed in this paper can improve the classification accuracy of EEG in AD classification tasks and can be applied to related classification tasks.
We analyzed transcriptomic alterations, immune infiltration, and candidate drugs in moderate-to-severe ulcerative colitis (UC) using integrative bioinformatics and in vitro validation. Thirty persistently dysregulated genes were identified and externally validated, with neutrophils emerging as a key immune component correlated with HSD11B2, SPP1, and FAM55A. Drug prediction and molecular docking highlighted Dasatinib and Acetalax as candidate compounds. In LPS-stimulated Caco-2 cells, both drugs improved viability and modulated inflammation-related gene and protein expression. These findings provide hypothesis-generating evidence that Dasatinib and Acetalax may have therapeutic relevance in moderate-to-severe UC.
Machine learning techniques have recently shown significant promise in electroencephalograph (EEG)-based depression recognition. However, existing methods often rely on simple feature concatenation to fuse information from multiple perspectives, failing to adequately exploit the complementarity of heterogeneous features. To this end, we propose a novel affective computing framework for identifying depression that integrates nonlinear analysis with adaptive feature coupling. The framework first uses different entropy measures to effectively characterize the intricate and chaotic dynamics present in EEG signals. Then, a new fusion strategy based on weighted average is proposed to adaptively aggregate complementary information among multi-view features. More importantly, this strategy can mitigate the influence of redundant features. Experimental results on publicly available datasets show that our methodology can dramatically improve the accuracy of depression recognition. Meanwhile, visualization analysis reveals that compared with healthy controls, patients with depression exhibit lower EEG entropy values, reflecting reduced complexity in their brain activity. Due to the good performance of the framework, this study provides important insights into the usefulness of nonlinear analysis and adaptive feature fusion in EEG decoding tasks.
Accurate prediction of drug-target interactions (DTI) is critical for accelerating drug discovery. In this study, we propose deep contrastive learning (DeepCL), a novel deep contrastive learning framework that significantly improves DTI prediction accuracy, particularly in low-coverage scenarios. Unlike traditional methods, DeepCL introduces a generalized sigmoid activation function to resolve numerical underflow issues and employs a margin-based contrastive loss to enforce better separation between interacting and non-interacting pairs. Extensive experiments on three benchmark datasets (Davis, BindingDB, and BIOSNAP) demonstrate that DeepCL outperforms state-of-the-art methods in both standard and zero-shot settings, achieving superior AUPR and AUROC scores. Methodologically, DeepCL constructs a dual-pathway architecture: it leverages the ESM-2 protein language models to capture rich contextual protein features and Morgan fingerprints for precise molecular structural representation. These heterogeneous features are aligned in a shared latent space via modality-specific projectors. DeepCL contributes a robust, scalable, and numerically stable solution for DTI prediction.
As the central hub connecting the upper limb and torso, the shoulder has complex spatial kinematics driven by multi-joint coordination. To realize structured modeling and quantitative evaluation of natural shoulder motion, this study proposes a kinematic modeling approach based on an open-chain spatial hybrid linkage mechanism, validated by the Vicon motion capture system. A 7-degree-of-freedom spatial linkage model incorporating SC, AC, GH and ST joints is constructed, with defined degrees of freedom and coupling relationships. Static equilibrium models in three planes are established to assess joint torque distribution. Vicon-based 3D trajectory data of anatomical markers are analyzed for motion consistency, verifying the model's biomechanical plausibility and reconstruction accuracy, which lays a theoretical foundation for shoulder rehabilitation robots and intelligent exoskeletons.