Bioinformatics. 2025 Jul 21:btaf415. doi: 10.1093/bioinformatics/btaf415. Online ahead of print.

ABSTRACT

MOTIVATION: Solenoid proteins, a subset of tandem repeat proteins, have structurally distinct, modular, and elongated architectures that differentiate them from globular proteins. These proteins play essential roles in diverse biological processes, including protein binding, enzymatic catalysis, ice binding, and nucleic acid interactions. Despite their biological significance and increasing commercial applications-such as in therapeutic engineered variants like DARPins and designed PPR proteins-accurate identification and annotation of solenoid structures remain challenging. Given that solenoid structures are more conserved than their sequences, recent advances in protein structure prediction suggest that structure-based solenoid detection methods are preferable to sequence-based ones.

RESULTS: We introduce SOLeNNoID, a deep-learning-based pipeline for predicting solenoid residues in protein structures. Our method employs a convolutional neural network (CNN) architecture to analyze protein distance matrices, enabling accurate identification of solenoid-containing regions. SOLeNNoID covers all three solenoid subclasses: α-, α/β-, and β-solenoids. Comparative evaluation against existing structure-based methods demonstrates the superior performance of our approach. Applying SOLeNNoID to the entire Protein Data Bank (PDB) led to a 71% increase in detected solenoid-containing entries compared to the gold-standard RepeatsDB database, significantly expanding the known solenoid protein repertoire.

AVAILABILITY AND IMPLEMENTATION: SOLeNNoID is implemented in Python and available on github at https://github.com/gnik2018/SOLeNNoID. The source code and pre-trained models are accessible under a free-software license. Training data are available on Zenodo at https://zenodo.org/records/14927497. Contact: James W Murray j.w.murray@imperial.ac.uk.

SUPPLEMENTARY INFORMATION: Available online.

PMID:40689530 | DOI:10.1093/bioinformatics/btaf415

Ann Surg. 2025 Jul 21. doi: 10.1097/SLA.0000000000006841. Online ahead of print.

ABSTRACT

OBJECTIVE: To propose a deep learning (DL) system for the preoperative diagnosis of follicular-like thyroid neoplasms (FNs) using routine ultrasound images.

SUMMARY BACKGROUND DATA: Preoperative diagnosis of malignancy in nodules suspicious for an FN remains challenging. Ultrasound, fine-needle aspiration cytology, and intraoperative frozen section pathology cannot unambiguously distinguish between benign and malignant FNs, leading to unnecessary biopsies and operations in benign nodules.

METHODS: This multicenter, retrospective study included 3634 patients who underwent ultrasound and received a definite diagnosis of FN from 11 centers, comprising thyroid follicular adenoma (n=1748), follicular carcinoma (n=299), and follicular variant of papillary thyroid carcinoma (n=1587). Four DL models including Inception-v3, ResNet50, Inception-ResNet-v2, and DenseNet161 were constructed on a training set (n=2587, 6178 images) and were verified on an internal validation set (n=648, 1633 images) and an external validation set (n=399, 847 images). The diagnostic efficacy of the DL models was evaluated against the ACR TI-RADS regarding the area under the curve (AUC), sensitivity, specificity, and unnecessary biopsy rate.

RESULTS: When externally validated, the four DL models yielded robust and comparable performance, with AUCs of 82.2%-85.2%, sensitivities of 69.6%-76.0%, and specificities of 84.1%-89.2%, which outperformed the ACR TI-RADS. Compared to ACR TI-RADS, the DL models showed a higher biopsy rate of malignancy (71.6% -79.9% vs 37.7%, P<0.001) and a significantly lower unnecessary FNAB rate (8.5% -12.8% vs 40.7%, P<0.001).

CONCLUSION: This study provides a noninvasive DL tool for accurate preoperative diagnosis of FNs, showing better performance than ACR TI-RADS and reducing unnecessary invasive interventions.

PMID:40689491 | DOI:10.1097/SLA.0000000000006841

Environ Sci Ecotechnol. 2025 Jul 5;26:100592. doi: 10.1016/j.ese.2025.100592. eCollection 2025 Jul.

ABSTRACT

Lake ecosystems, vital freshwater resources, are increasingly threatened by pollution from riverine inputs, making the management of these loads critical for preventing ecological degradation. Predicting the combined effects of multiple rivers on lake water quality is a significant challenge; traditional mechanistic models are computationally intensive and data-dependent, while conventional machine learning methods often fail to capture the system's multifaceted nature. This complexity creates a critical need for an integrated predictive tool for effective environmental management. Here we show a multi-task deep neural network (MTDNN) that can accurately and simultaneously predict four key water quality indicators-permanganate index, total phosphorus, total nitrogen, and algal density-at multiple locations within a complex lake system using data from its inflowing rivers. Our model, applied to Dianchi Lake in China, improves predictive precision by up to 56.3 % compared to established mechanistic and single-task deep learning models. Furthermore, the model pinpoints the specific contributions of each river and identifies water temperature and wastewater effluent as dominant, site-specific drivers of pollution. Scenario-based forecasting demonstrates that using reclaimed water for lake replenishment is a viable strategy that does not cause deterioration. This MTDNN framework offers a powerful and transferable tool for data-driven lake management, enabling targeted interventions and sustainable water resource protection.

PMID:40689412 | PMC:PMC12274680 | DOI:10.1016/j.ese.2025.100592

Curr Res Food Sci. 2025 Jul 3;11:101133. doi: 10.1016/j.crfs.2025.101133. eCollection 2025.

ABSTRACT

Conventional egg freshness assessment methods based on the Haugh unit are destructive and time-consuming. Accordingly, this study investigated the use of hyperspectral imaging (450-1100 nm) for nondestructive egg freshness evaluation. Spectral data were preprocessed using standard normal variates to minimize spectral variability, followed by wavelength selection - a crucial step for improving model predictability. Particularly, distance correlation, a statistically robust yet rarely explored method in hyperspectral wavelength selection, was employed to identify informative wavelengths. The selected wavelengths were incorporated into various regression models, namely convolutional neural network, gradient boosting trees, multiple linear regression, partial least squares regression, and support vector regression models. We observed that the convolutional neural network model incorporating the distance correlation method demonstrated the best performance (correlation coefficient of 0.9056 and root mean square error of 4.4152), outperforming the other models using commonly applied wavelength selection methods. Pseudocolor maps of egg freshness were generated based on the best obtained model.

PMID:40689294 | PMC:PMC12270930 | DOI:10.1016/j.crfs.2025.101133

Curr Res Food Sci. 2025 Jun 26;11:101127. doi: 10.1016/j.crfs.2025.101127. eCollection 2025.

ABSTRACT

This review systematically examines the current landscape of data resources and computational models for predicting food flavor and texture. Taste is the most well-defined sensory component, and molecular classification is aligned with the five basic tastes: sweet, sour, bitter, salty, and umami. Odor prediction, while similar in premise, faces greater challenges due to the vast and diverse range of detectable odors and a lack of standardized olfactory metrics. Machine learning models, including graph neural networks and deep learning methods, have shown promise in identifying taste and odor compounds. Texture prediction has seen comparatively less research interest but may prove to be impactful in food quality control pipelines, although more work is needed in creating robust food texture datasets. The review highlights the growing availability of specialized databases which support the development and benchmarking of predictive models. Despite recent advancements, gaps remain in mapping sensory spaces and incorporating receptor-level data. Future directions include creating more extensive and high-quality datasets, improving model explainability, and exploring innovative applications in food design, fragrance, pharmaceuticals, and environmental monitoring. This work provides a comprehensive resource for researchers aiming to advance the field of flavor and texture prediction.

PMID:40689288 | PMC:PMC12274706 | DOI:10.1016/j.crfs.2025.101127

Disaster Med Public Health Prep. 2025 Jul 21;19:e197. doi: 10.1017/dmp.2025.10062.

ABSTRACT

OBJECTIVES: Large-scale crises, including wars and pandemics, have repeatedly shaped human history, and their simultaneous occurrence presents profound challenges to societies. Understanding the dynamics of epidemic spread during warfare is essential for developing effective containment strategies in complex conflict zones. While research has explored epidemic models in various settings, the impact of warfare on epidemic dynamics remains underexplored.

METHODS: We proposed a novel mathematical model that integrates the epidemiological SIR (susceptible-infected-recovered) model with the war dynamics Lanchester model to explore the dual influence of war and pandemic on a population's mortality. Moreover, we consider a dual-use military and civil health care system that aims to reduce the overall mortality rate, which can use different administration policies such as prioritizing soldiers over civilians. Using an agent-based simulation to generate in silico data, we trained a deep reinforcement learning model based on the deep Q-network algorithm for health care administration policy and conducted an intensive investigation on its performance.

RESULTS: Our results show that a pandemic during war conduces chaotic dynamics where the health care system should either prioritize war-injured soldiers or pandemic-infected civilians based on the immediate amount of mortality from each option, ignoring long-term objectives.

CONCLUSIONS: Our findings highlight the importance of integrating conflict-related factors into epidemic modeling to enhance preparedness and response strategies in conflict-affected areas.

PMID:40686043 | DOI:10.1017/dmp.2025.10062

Analyst. 2025 Jul 21. doi: 10.1039/d4an01426j. Online ahead of print.

ABSTRACT

Molecularly imprinted photonic crystal hydrogel (MIPCH) serves as a highly effective platform for the sensitive and selective detection of various analyte molecules. In this study, we present a smartphone-based inverse opal MIPCH (IOMIPCH) sensor designed for the sensitive and selective detection of bisphenol A (BPA) in water samples. The sensor is prepared by photopolymerizing the hydrogel precursor solution within the voids of a polystyrene (PS) photonic crystal (PC) opal film. This is followed by the etching of BPA molecules and the removal of PS spheres, forming an inverse opal structure with binding sites for the analyte BPA. The sensor displays a vibrant structural colour that experiences a redshift upon rebinding of the BPA molecules. The structural colour change provides a visually observable indication of the sensor response. The IOMIPCH-BPA sensor demonstrates a low limit of detection (LoD) of 0.69 fM and a rapid response time of 4 minutes with the ability to selectively detect BPA even in complex sample matrices. Additionally, it is reusable and maintains its performance for up to one month. We used the sensor response images to train a deep learning-based regression model on the smartphone, enabling quantitative predictions of BPA concentration. This integration creates an accurate, portable smart sensor platform capable of real-time BPA sensing.

PMID:40685994 | DOI:10.1039/d4an01426j

Brain Behav. 2025 Jul;15(7):e70689. doi: 10.1002/brb3.70689.

ABSTRACT

PURPOSE: MRI-based whole-brain manual segmentation methods are considered the gold standard for brain volumetric analysis, but are time-consuming and prone to human error. Automated segmentation tools like FreeSurfer can identify differences in brain volumes between healthy and non-healthy individuals. Deep-learning-based segmentation tools, such as FastSurfer, offer faster processing times, but further validation is needed, particularly in pediatric cases. This study aims to compare FastSurfer with FreeSurfer in a pediatric cohort and compare the volume estimates with previously published reference values.

METHODS: A multicenter cohort of 448 subjects aged 4-18 years from three centers was used to compare FastSurfer with FreeSurfer. Validation metrics, including the Dice Similarity Coefficient (DSC), relative volume differences (RVD), and intraclass correlation coefficient (ICC), were computed. Hemispheric asymmetries were assessed by calculating a hemispheric asymmetry index.

FINDINGS: The segmentation methods demonstrated high agreement, with a mean DSC across subjects and regions of interest of 0.90 (95% CI: 0.79; 0.95), RVD of 0.3% (95% CI: -7.6%; 7.4%), and ICC of 0.87 (95% CI: 0.52; 0.94). After a visual inspection, which led to the exclusion of 12 subjects with segmentation errors, growth charts for relative volume estimates of 15 anatomical brain regions were generated, revealing varying growth patterns across ages. A potential clinical application is illustrated by plotting a patient's data on these growth charts, showing a specific atrophy pattern.

CONCLUSION: To our knowledge, this is the first study investigating the use of FastSurfer in volumetric analysis of a pediatric population. Our findings suggest that FastSurfer is a reliable segmentation tool for pediatric data and is particularly promising for clinical practice due to its high accuracy despite rapid processing times. The morphometric data, growth charts, and code are publicly accessible.

PMID:40685764 | DOI:10.1002/brb3.70689

Adv Mater. 2025 Jul 20:e03926. doi: 10.1002/adma.202503926. Online ahead of print.

ABSTRACT

Iontronic pressure sensors are widely used in human motion monitoring and human-machine interactions owing to their high sensitivity, wide measurement range, and excellent resolution. However, conventional dielectric layer designs often involve complex fabrication processes, high costs, and limited performances. This paper proposes a novel sensor structure, the dual-dielectric-layer iontronic pressure sensor (DLIPS), which integrates high- and low-permittivity layers. Validated using silkworm cocoon ion gel and open-cell polyurethane foam as dielectrics, the DLIPS exhibited ultrahigh sensitivity (72548.7 kPa-1), a wide working pressure range (0.001-420 kPa), an exceptionally low detection limit (0.832 Pa), and remarkable durability exceeding 5000 cycles. By leveraging the distinct responses of the capacitance and resistance to pressure and temperature, the sensor can simultaneously measure both parameters. A deep learning regression model is integrated to decouple the mixed temperature and pressure signals, enabling accurate identification. Owing to its ultrahigh sensitivity and capability to detect minute pressure fluctuations, the DLIPS exhibited strong potential for skin-mounted silent speech recognition systems, achieving a recognition accuracy of up to 98.5%. Furthermore, the DLIPS provides a cost-effective and scalable approach for fabricating ultrahigh-sensitivity pressure sensors, underscoring its versatility in wearable technology applications.

PMID:40685692 | DOI:10.1002/adma.202503926

Small Methods. 2025 Jul 20:e00219. doi: 10.1002/smtd.202500219. Online ahead of print.

ABSTRACT

Human skin functions not only as a barrier but also as a sensitive interface responding to environmental stimuli. Skin-attachable conductive materials are gaining increasingly attention in the field of on-skin electronics for applications in health monitoring and human-machine interfaces (HMIs). However, achieving both low modulus and low hysteresis in skin-attachable conductive materials remains challenging. Herein, a skin-like transparent, low-hysteresis, and highly conductive ionogel for human motion monitoring and deep-learning assisted human-machine interface is introduced. The ionogel demonstrates low modulus (5.08 kPa), superior transparency (>92% in the visible range), low hysteresis (<3%), super adhesive and outstanding conductivity (up to 0.86 S/m at 20 °C). The ionogel based sensors demonstrate outstanding sensing sensitivity for human motion monitoring and biopotential detecting. In addition, the ionogel can be integrated in an HMI for handwriting recognition. The 1D-ResNet algorithm is developed for handwriting recognition, achieving an accuracy of 98.13%. It is believe that the ionogels with both low modulus and low hysteresis have great potential for wearable electronics in healthcare monitoring and HMIs in the future.

PMID:40685679 | DOI:10.1002/smtd.202500219

Bioinformatics. 2025 Jul 21:btaf413. doi: 10.1093/bioinformatics/btaf413. Online ahead of print.

ABSTRACT

MOTIVATION: Survival analysis is a branch of statistics that is crucial in medicine for modeling the time to critical events such as death or relapse, in order to improve treatment strategies and patient outcomes. Selecting survival models often involves a trade-off between performance and interpretability; deep learning models offer high performance but lack the transparency of more traditional approaches. This poses a significant issue in medicine, where practitioners are reluctant to use black-box models for critical patient decisions.

RESULTS: We introduce CoxKAN, a Cox proportional hazards Kolmogorov-Arnold Network for interpretable, high-performance survival analysis. Kolmogorov-Arnold Networks (KANs) were recently proposed as an interpretable and accurate alternative to multi-layer perceptrons (MLPs). We evaluated CoxKAN on four synthetic and nine real datasets, including five cohorts with clinical data and four with genomics biomarkers. In synthetic experiments, CoxKAN accurately recovered interpretable hazard function formulae and excelled in automatic feature selection. Evaluations on real datasets showed that CoxKAN consistently outperformed the traditional Cox proportional hazards model (by up to 4% in C-index) and matched or surpassed the performance of deep learning-based models. Importantly, CoxKAN revealed complex interactions between predictor variables and uncovered symbolic formulae, which are key capabilities that other survival analysis methods lack, to provide clear insights into the impact of key biomarkers on patient risk.

AVAILABILITY AND IMPLEMENTATION: CoxKAN is available at GitHub and Zenodo.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40685627 | DOI:10.1093/bioinformatics/btaf413

J Alzheimers Dis. 2025 Jul 20:13872877251359874. doi: 10.1177/13872877251359874. Online ahead of print.

ABSTRACT

BackgroundHandwriting and speech are served as reliable signatures for detecting cognitive decline, playing a pivotal role in the early diagnosing Alzheimer's disease (AD) and mild cognitive impairment (MCI). However, current unimodal approaches for diagnosing AD and MCI have demonstrated constraints in classification accuracy, potentially overlooking the synergistic value of combining handwriting and speech data.ObjectivePresenting an innovative multi-modal screening classification model, that harnesses handwriting and speech analysis to enhance MCI detection, aiming to overcome the constraints of single-modality approaches by integrating data from both modalities, thereby improving diagnostic accuracy.MethodsProposing a multimodal classification model based on gated recurrent unit (GRU) and attention mechanism, treating handwriting and speech data as sequence inputs. The model was constructed and tested on a dataset of 41 participants, including 20 MCI patients and 21 cognitively normal (CN) individuals. To mitigate the risk of overfitting due to the small sample size, we employed a 10-fold cross-validation strategy to ensure the robustness of the results.ResultsOur multimodal classification model achieved an accuracy of 95.2% for MCI versus CN individuals, which shows a significant improvement compared to the results of single-modality. This result indicates the effectiveness of the cross-fusion model in enhancing classification performance, offering a promising approach for the early diagnosis of neurodegenerative diseases.ConclusionsThe proposed GRU_CA effectively improves early MCI detection by fusing handwriting and speech data, outperforming a single modality. It shows strong potential for deployment in primary healthcare settings and establishes a foundation for future research on more complex diagnostic tasks, including CN, MCI, and AD classification, as well as longitudinal studies.

PMID:40685620 | DOI:10.1177/13872877251359874

Sci Rep. 2025 Jul 20;15(1):26304. doi: 10.1038/s41598-025-12526-1.

ABSTRACT

Mild cognitive impairment (MCI) and dementia pose significant health challenges in aging societies, emphasizing the need for accessible, cost-effective, and noninvasive diagnostic tools. Electroencephalography (EEG) is a promising biomarker, but traditional systems are limited by size, cost, and the need for skilled technicians. This study proposes a deep-learning-based approach using data from a portable EEG device to distinguish healthy volunteers (HVs) from patients with dementia-related conditions. We analyzed EEG data from 233 participants, including 119 HVs and 114 patients, and transformed the signals into frequency-domain features using a short-time Fourier transform. A customized transformer-based model was trained and evaluated using 10-fold cross-validation and a holdout dataset. In the cross-validation, the model achieved an area under the curve (AUC) of 0.872 and a balanced accuracy (bACC) of 80.8% in distinguishing HVs from patients. Subgroup analyses were conducted for HVs versus patients stratified by dementia severity and by clinical diagnosis, yielding AUCs ranging from 0.812 to 0.898 and bACCs from 74.9 to 86.4%. Comparable results were obtained in the holdout dataset. These findings suggest that portable EEG data combined with deep learning may serve as a practical tool for the early detection and classification of dementia-related conditions.

PMID:40685486 | DOI:10.1038/s41598-025-12526-1

Sci Rep. 2025 Jul 20;15(1):26361. doi: 10.1038/s41598-025-12533-2.

ABSTRACT

In industrial settings, bearing health directly affects equipment stability, making accurate and efficient fault diagnosis critical for operational safety. Recently, Transformer models have been widely adopted in bearing fault diagnosis due to their strong global modeling capabilities. However, they still face significant challenges under strong noise and limited data. To address this, this paper proposes an end-to-end Vision Transformer with time-frequency fusion and dual attention across spatial and channel dimensions. The model adopts a dual-branch design: the time-domain branch incorporates spatial and channel attention to capture both local and global features, while the frequency-domain branch uses FFT to extract spectral information and fuses it with temporal features for efficient multi-scale modeling. To further enhance sensitivity to local patterns and periodic variations, a cross-scale convolution module and a periodic feedforward network are introduced. Experiments on the CWRU and PU datasets demonstrate that the proposed model achieves 99.42% and 98.14% accuracy, respectively, under noisy and data-scarce conditions. The results confirm superior noise robustness and diagnostic performance over recent state-of-the-art methods, highlighting its practical potential for real-world industrial applications.

PMID:40685451 | DOI:10.1038/s41598-025-12533-2

Sci Rep. 2025 Jul 20;15(1):26355. doi: 10.1038/s41598-025-11510-z.

ABSTRACT

Contemporary supply chain networks in the context of the era of Industry 4.0 are becoming more erratic and complex, and have an influx of structured and unstructured data. Conventional practices of supply chain management (SCM) cannot overcome real-time uncertainties, and it is time to orient the SCM toward AI-guided predictive modeling. This research contains a suggestion of a deep learning (DL) framework that combines Self-Organizing Maps (SOMs), Principal Component Analysis (PCA), and Artificial Neural Networks (ANNs) to predict more accurately the supply chain shipping timing and delivery risk. Applying the DataCo Smart Supply Chain dataset, the offered SOM+ANN model proved much more accurate than conventional Machine Learning (ML) procedures, e.g., Random Forest (RF), XGBoost, or Decision Tree (DT), to address the tasks of predicting the shipping time and categorizing the risk of late delivery. The R2 was found to be 0.92, Root Mean Squared Error (RMSE) 0.936, and Mean Absolute Error (MAE) 0.8459 of the SOM+ANN model estimated the shipping duration. In the classification, the accuracy was 96% and the F1-score was 96.22%. Furthermore, the research also employed another dataset, making it more accurate, better generalized, and robust. The proposed model achieved 89.65% accuracy in dataset 2. The model's outcomes are also interpretable and assisted by SHAP (Shapley Additive exPlanations). The interpretability methods enabled end users to comprehend how the model makes classification decisions. The proposed framework promotes SCM operations, resilience, and decision-making by incorporating transparent AI methodologies.

PMID:40685434 | DOI:10.1038/s41598-025-11510-z

Sci Rep. 2025 Jul 20;15(1):26308. doi: 10.1038/s41598-025-10464-6.

ABSTRACT

Rainfall-induced landslides often result in significant human and property losses, and reliable rainfall thresholds can effectively mitigate the hazards associated with them. However, constructing reliable rainfall thresholds in mountainous areas with sparse rain gauge stations is challenging. This study aims to establish reliable empirical rainfall thresholds for the landslide early warning systems (LEWSs) in the study area, utilizing radar-derived rainfall data processed by deep learning. Firstly, the accuracy of radar-derived rainfall data was verified based on the data with rain gauge measurements. Subsequently, utilizing frequency theory and Bayesian probability analysis methods, in conjunction with the collected landslide data and radar-derived rainfall data, various exceedance probability thresholds for rainfall-induced landslides were determined. Furthermore, the influence of cumulative effective antecedent rainfall on the initiation of landslides was investigated. The proposed threshold equations and the effect of antecedent rainfall on landslides are intended to aid in enhancing the LEWSs for this region. The findings provide valuable insights for managing rainfall-induced landslides, and can be applied to other areas with sparse rainfall data, offering a scientific basis for improved landslide prediction and risk management.

PMID:40685424 | DOI:10.1038/s41598-025-10464-6

Sci Rep. 2025 Jul 20;15(1):26311. doi: 10.1038/s41598-025-09774-6.

ABSTRACT

Unlike traditional reflection-only reconfigurable intelligent surfaces (RISs), simultaneously transmitting and reflecting RISs (STAR-RISs) introduce an innovative technology. They expand the coverage area from half-space to full-space. This advancement provides new degrees of freedom (DoF) for controlling and optimizing signal propagation. This research explores the performance of the STAR-RIS in 6G wireless networks, comparing nearly passive STAR-RIS (NP-STAR) and active STAR (ASTAR) against nearly passive RIS (NP-RIS) and active RIS (ARIS) benchmarks. Through extensive simulations and deep learning (DL)-based optimization, we assess achievable data rates and spectral energy efficiency (SEE) across various system configurations, training dataset sizes, and user locations. Results show that in high-interference environments, NP-STAR configurations can sometimes exceed ASTAR implementations due to their ability to mitigate interference amplification. In addition, the deterioration in spectral energy efficiency (SEE) of active implementations compared to nearly passive ones confirms their greater consumption of energy. The exploration of STAR-RIS technology in 6G networks reveals key research gaps in prior works. These include limited studies on performance under real-world conditions, insufficient understanding of energy-spectral efficiency trade-offs, and the need for reliable DL optimization across scenarios. Practical deployment issues like power management and interference control also lack adequate research. Addressing these gaps is crucial for advancing STAR-RIS technology in this paper. This study emphasizes the importance of managing transmit power levels at base stations to control interference, with active setups particularly excelling in optimal channel conditions. Furthermore, DL approaches can effectively approximate genie-aided performance bounds with adequate training data, especially in complex channel scenarios. These insights provide practical guidance for deploying STAR-RIS technology in demanding wireless networks.

PMID:40685393 | DOI:10.1038/s41598-025-09774-6

BMC Bioinformatics. 2025 Jul 20;26(1):185. doi: 10.1186/s12859-025-06141-0.

ABSTRACT

BACKGROUND: Identification of drug target interactions (DTI) is an important part of the drug discovery process. Since prediction of DTI using laboratory tests is time consuming and laborious, automated tools using computational intelligence (CI) techniques become essential. The prediction of DTI is a challenging process due to the absence of known drug-target relationship and no experimentally verified negative samples. The datasets with limited or unbalanced data, do not perform well. The models that use heterogeneous networks, non-linear fusion techniques, and heuristic similarity selection may need a lot of computational power and experience to implement and fine-tune. The latest developments in machine learning (ML) and deep learning (DL) models can be employed for effective DTI prediction process.

RESULTS: To that end, this study develops a novel DTI Prediction model, namely, DTIP-WINDGRU Drug-Target Interaction Prediction with Wind-Enhanced GRU. The major aim is to determine the DTIs in both labelled and unlabelled samples accurately compared to traditional wet lab experiments. To accomplish this, the proposed DTIP-WINDGRU model primarily performs pre-processing and class labelling. In addition, drug-to-drug (D-D) and target-to-target (T-T) interactions are employed to initialize the weights of the GRU model and are employed for the, DTI prediction process. Finally, the Wind Driven Optimization (WDO) algorithm is utilized to optimally choose the hyperparameters involved in the GRU model.

CONCLUSIONS: For ensuring the effectual prediction results of the DTIP-WINDGRU model, a widespread experimentation process was carried out using four datasets. This comprehensive comparative study highlighted the better performance of the DTIP-WINDGRU model over existing techniques.

PMID:40685357 | DOI:10.1186/s12859-025-06141-0

Br J Anaesth. 2025 Jul 19:S0007-0912(25)00378-2. doi: 10.1016/j.bja.2025.06.016. Online ahead of print.

ABSTRACT

Selection bias has been identified in hypotension prediction models, but its impact on an algorithm's ability to learn relevant information from the arterial waveform remains unclear. The recent study by Yang and colleagues sheds considerable light on this by training and evaluating a deep learning prediction model with biased and unbiased data selections. Unbiased training data allowed an algorithm to learn modestly more than just current blood pressure and the bias significantly distorted and inflated the positive predictive value. We discuss these findings and offer suggestions for further developing effective hypotension prediction algorithms.

PMID:40685291 | DOI:10.1016/j.bja.2025.06.016

Eur J Radiol. 2025 Jul 16;191:112320. doi: 10.1016/j.ejrad.2025.112320. Online ahead of print.

ABSTRACT

BACKGROUND: Coronary Artery Disease-Reporting and Data System (CAD-RADS), a standardized reporting system of stenosis severity from coronary computed tomography angiography (CCTA), is performed manually by expert radiologists, being time-consuming and prone to interobserver variability. While deep learning methods automating CAD-RADS scoring have been proposed, radiomics-based machine-learning approaches are lacking, despite their improved interpretability. This study aims to introduce a novel radiomics-based machine-learning approach for automating CAD-RADS scoring from CCTA images with multiplanar reconstruction.

METHODS: This retrospective monocentric study included 251 patients (male 70 %; mean age 60.5 ± 12.7) who underwent CCTA in 2016-2018 for clinical evaluation of CAD. Images were automatically segmented, and radiomic features were extracted. Clinical characteristics were collected. The image dataset was partitioned into training and test sets (90 %-10 %). The training phase encompassed feature scaling and selection, data balancing and model training within a 5-fold cross-validation. A cascade pipeline was implemented for both 6-class CAD-RADS scoring and 4-class therapy-oriented classification (0-1, 2, 3-4, 5), through consecutive sub-tasks. For each classification task the cascade pipeline was applied to develop clinical, radiomic, and combined models.

RESULTS: The radiomic, combined and clinical models yielded AUC = 0.88 [0.86-0.88], AUC = 0.90 [0.88-0.90], and AUC = 0.66 [0.66-0.67] for the CAD-RADS scoring, and AUC = 0.93 [0.91-0.93], AUC = 0.97 [0.96-0.97], and AUC = 79 [0.78-0.79] for the therapy-oriented classification. The radiomic and combined models significantly outperformed (DeLong p-value < 0.05) the clinical one in class 1 and 2 (CAD-RADS cascade) and class 2 (therapy-oriented cascade).

CONCLUSIONS: This study represents the first CAD-RADS classification radiomic model, guaranteeing higher explainability and providing a promising support system in coronary artery stenosis assessment.

PMID:40684709 | DOI:10.1016/j.ejrad.2025.112320