J Imaging. 2025 Mar 19;11(3):88. doi: 10.3390/jimaging11030088.

ABSTRACT

The objective of this study is to propose an advanced image enhancement strategy to address the challenge of reducing radiation doses in pediatric renal scintigraphy. Data from a public dynamic renal scintigraphy database were used. Based on noisier images, four denoising neural networks (DnCNN, UDnCNN, DUDnCNN, and AttnGAN) were evaluated. To evaluate the quality of the noise reduction, with minimal detail loss, the kidney signal-to-noise ratio (SNR) and multiscale structural similarity (MS-SSIM) were used. Although all the networks reduced noise, UDnCNN achieved the best balance between SNR and MS-SSIM, leading to the most notable improvements in image quality. In clinical practice, 100% of the acquired data are summed to produce the final image. To simulate the dose reduction, we summed only 50%, simulating a proportional decrease in radiation. The proposed deep-learning approach for image enhancement ensured that half of all the frames acquired may yield results that are comparable to those of the complete dataset, suggesting that it is feasible to reduce patients' exposure to radiation. This study demonstrates that the neural networks evaluated can markedly improve the renal scintigraphic image quality, facilitating high-quality imaging with lower radiation doses, which will benefit the pediatric population considerably.

PMID:40137200 | DOI:10.3390/jimaging11030088

J Imaging. 2025 Mar 18;11(3):85. doi: 10.3390/jimaging11030085.

ABSTRACT

Semantic segmentation in deep learning is a crucial area of research within computer vision, aimed at assigning specific labels to each pixel in an image. The segmentation of crops, plants, and weeds has significantly advanced the application of deep learning in precision agriculture, leading to the development of sophisticated architectures based on convolutional neural networks (CNNs). This study proposes a segmentation algorithm for identifying plants and weeds using broadband multispectral images. In the first part of this algorithm, we utilize the PIF-Net model for feature extraction and fusion. The resulting feature map is then employed to enhance an optimized U-Net model for semantic segmentation within a broadband system. Our investigation focuses specifically on scenes from the CAVIAR dataset of multispectral images. The proposed algorithm has enabled us to effectively capture complex details while regulating the learning process, achieving an impressive overall accuracy of 98.2%. The results demonstrate that our approach to semantic segmentation and the differentiation between plants and weeds yields accurate and compelling outcomes.

PMID:40137197 | DOI:10.3390/jimaging11030085

J Imaging. 2025 Mar 18;11(3):84. doi: 10.3390/jimaging11030084.

ABSTRACT

Background: This study aims to objectively evaluate the overall quality of colonoscopies using a specially trained deep learning-based semantic segmentation neural network. This represents a modern and valuable approach for the analysis of colonoscopy frames. Methods: We collected thousands of colonoscopy frames extracted from a set of video colonoscopy files. A color-based image processing method was used to extract color features from specific regions of each colonoscopy frame, namely, the intestinal mucosa, residues, artifacts, and lumen. With these features, we automatically annotated all the colonoscopy frames and then selected the best of them to train a semantic segmentation network. This trained network was used to classify the four region types in a different set of test colonoscopy frames and extract pixel statistics that are relevant to quality evaluation. The test colonoscopies were also evaluated by colonoscopy experts using the Boston scale. Results: The deep learning semantic segmentation method obtained good results, in terms of classifying the four key regions in colonoscopy frames, and produced pixel statistics that are efficient in terms of objective quality assessment. The Spearman correlation results were as follows: BBPS vs. pixel scores: 0.69; BBPS vs. mucosa pixel percentage: 0.63; BBPS vs. residue pixel percentage: -0.47; BBPS vs. Artifact Pixel Percentage: -0.65. The agreement analysis using Cohen's Kappa yielded a value of 0.28. The colonoscopy evaluation based on the extracted pixel statistics showed a fair level of compatibility with the experts' evaluations. Conclusions: Our proposed deep learning semantic segmentation approach is shown to be a promising tool for evaluating the overall quality of colonoscopies and goes beyond the Boston Bowel Preparation Scale in terms of assessing colonoscopy quality. In particular, while the Boston scale focuses solely on the amount of residual content, our method can identify and quantify the percentage of colonic mucosa, residues, and artifacts, providing a more comprehensive and objective evaluation.

PMID:40137196 | DOI:10.3390/jimaging11030084

J Imaging. 2025 Mar 3;11(3):76. doi: 10.3390/jimaging11030076.

ABSTRACT

COVID-19 can cause acute infectious diseases of the respiratory system, and may probably lead to heart damage, which will seriously threaten human health. Electrocardiograms (ECGs) have the advantages of being low cost, non-invasive, and radiation free, and is widely used for evaluating heart health status. In this work, a lightweight deep learning network named GM-CBAM-ResNet is proposed for diagnosing COVID-19 based on ECG images. GM-CBAM-ResNet is constructed by replacing the convolution module with the Ghost module (GM) and adding the convolutional block attention module (CBAM) in the residual module of ResNet. To reveal the superiority of GM-CBAM-ResNet, the other three methods (ResNet, GM-ResNet, and CBAM-ResNet) are also analyzed from the following aspects: model performance, complexity, and interpretability. The model performance is evaluated by using the open 'ECG Images dataset of Cardiac and COVID-19 Patients'. The complexity is reflected by comparing the number of model parameters. The interpretability is analyzed by utilizing Gradient-weighted Class Activation Mapping (Grad-CAM). Parameter statistics indicate that, on the basis of ResNet19, the number of model parameters of GM-CBAM-ResNet19 is reduced by 45.4%. Experimental results show that, under less model complexity, GM-CBAM-ResNet19 improves the diagnostic accuracy by approximately 5% in comparison with ResNet19. Additionally, the interpretability analysis shows that CBAM can suppress the interference of grid backgrounds and ensure higher diagnostic accuracy under lower model complexity. This work provides a lightweight solution for the rapid and accurate diagnosing of COVD-19 based on ECG images, which holds significant practical deployment value.

PMID:40137188 | DOI:10.3390/jimaging11030076

J Imaging. 2025 Feb 26;11(3):72. doi: 10.3390/jimaging11030072.

ABSTRACT

In an era where security concerns are ever-increasing, the need for advanced technology to detect visible and concealed weapons has become critical. This paper introduces a novel two-stage method for concealed handgun detection, leveraging thermal imaging and deep learning, offering a potential real-world solution for law enforcement and surveillance applications. The approach first detects potential firearms at the frame level and subsequently verifies their association with a detected person, significantly reducing false positives and false negatives. Alarms are triggered only under specific conditions to ensure accurate and reliable detection, with precautionary alerts raised if no person is detected but a firearm is identified. Key contributions include a lightweight algorithm optimized for low-end embedded devices, making it suitable for wearable and mobile applications, and the creation of a tailored thermal dataset for controlled concealment scenarios. The system is implemented on a chest-worn Android smartphone with a miniature thermal camera, enabling hands-free operation. Experimental results validate the method's effectiveness, achieving an mAP@50-95 of 64.52% on our dataset, improving state-of-the-art methods. By reducing false negatives and improving reliability, this study offers a scalable, practical solution for security applications.

PMID:40137184 | DOI:10.3390/jimaging11030072

Metabolites. 2025 Mar 3;15(3):174. doi: 10.3390/metabo15030174.

ABSTRACT

Background/Objectives: Metabolomics has recently emerged as a key tool in the biological sciences, offering insights into metabolic pathways and processes. Over the last decade, network-based machine learning approaches have gained significant popularity and application across various fields. While several studies have utilized metabolomics profiles for sample classification, many network-based machine learning approaches remain unexplored for metabolomic-based classification tasks. This study aims to compare the performance of various network-based machine learning approaches, including recently developed methods, in metabolomics-based classification. Methods: A standard data preprocessing procedure was applied to 17 metabolomic datasets, and Bayesian neural network (BNN), convolutional neural network (CNN), feedforward neural network (FNN), Kolmogorov-Arnold network (KAN), and spiking neural network (SNN) were evaluated on each dataset. The datasets varied widely in size, mass spectrometry method, and response variable. Results: With respect to AUC on test data, BNN, CNN, FNN, KAN, and SNN were the top-performing models in 4, 1, 5, 3, and 4 of the 17 datasets, respectively. Regarding F1-score, the top-performing models were BNN (3 datasets), CNN (3 datasets), FNN (4 datasets), KAN (4 datasets), and SNN (3 datasets). For accuracy, BNN, CNN, FNN, KAN, and SNN performed best in 4, 1, 4, 4, and 4 datasets, respectively. Conclusions: No network-based modeling approach consistently outperformed others across the metrics of AUC, F1-score, or accuracy. Our results indicate that while no single network-based modeling approach is superior for metabolomics-based classification tasks, BNN, KAN, and SNN may be underappreciated and underutilized relative to the more commonly used CNN and FNN.

PMID:40137139 | DOI:10.3390/metabo15030174

Biomimetics (Basel). 2025 Mar 20;10(3):191. doi: 10.3390/biomimetics10030191.

ABSTRACT

Chemical oxygen demand (COD) is a critical parameter employed to assess the level of organic pollution in water. Accurate COD detection is essential for effective environmental monitoring and water quality assessment. Ultraviolet-visible (UV-Vis) spectroscopy has become a widely applied method for COD detection due to its convenience and the absence of the need for chemical reagents. This non-destructive and reagent-free approach offers a rapid and reliable means of analyzing water. Recently, deep learning has emerged as a powerful tool for automating the process of spectral feature extraction and improving COD prediction accuracy. In this paper, we propose a novel multi-scale one-dimensional convolutional neural network (MS-1D-CNN) fusion model designed specifically for spectral feature extraction and COD prediction. The architecture of the proposed model involves inputting raw UV-Vis spectra into three parallel sub-1D-CNNs, which independently process the data. The outputs from the final convolution and pooling layers of each sub-CNN are then fused into a single layer, capturing a rich set of spectral features. This fused output is subsequently passed through a Flatten layer followed by fully connected layers to predict the COD value. Experimental results demonstrate the effectiveness of the proposed method, as it was compared with three traditional methods and three deep learning methods on the same dataset. The MS-1D-CNN model showed a significant improvement in the accuracy of COD prediction, highlighting its potential for more reliable and efficient water quality monitoring.

PMID:40136845 | DOI:10.3390/biomimetics10030191

PLoS One. 2025 Mar 25;20(3):e0320219. doi: 10.1371/journal.pone.0320219. eCollection 2025.

ABSTRACT

Neuroinflammation is a key feature of human neurodisease including neuropathy and neurodegenerative disease and is driven by the activation microglia, immune cells of the nervous system. During activation microglia release pro-inflammatory cytokines as well as reactive oxygen species (ROS) that can drive local neuronal and glial damage. Phytocannabinoids are an important class of naturally occurring compounds found in the cannabis plant (Cannabis sativa) that interact with the body's endocannabinoid receptor system. Cannabidiol (CBD) is a prototype phytocannabinoid with anti-inflammatory properties observed in cells and animal models. We measured ROS in human microglia (HMC3) cells using CellROX, a fluorescent dynamic ROS indicator. We tested the effect of CBD on ROS level in the presence of three known immune activators: lipopolysaccharide (LPS), amyloid beta (Aβ42), and human immunodeficiency virus (HIV) glycoprotein (GP120). Confocal microscopy images within microglia were coupled to a deep learning model using a convolutional neural network (CNN) to predict ROS responses. Our study demonstrates a deep learning platform that can be used in the assessment of CBD effect in immune cells using ROS image measure.

PMID:40131976 | DOI:10.1371/journal.pone.0320219

PLoS One. 2025 Mar 25;20(3):e0319858. doi: 10.1371/journal.pone.0319858. eCollection 2025.

ABSTRACT

This study aims to explore the integration of the Faster R-CNN (Region-based Convolutional Neural Network) algorithm from deep learning into the MobileNet v2 architecture, within the context of enterprises aiming for carbon neutrality in their development process. The experiment develops a marine oil condition monitoring and classification model based on the fusion of MobileNet v2 and Faster R-CNN algorithms. This model utilizes the MobileNet v2 network to extract rich feature information from input images and combines the Faster R-CNN algorithm to rapidly and accurately generate candidate regions for oil condition monitoring, followed by detailed feature fusion and classification of these regions. The performance of the model is evaluated through experimental assessments. The results demonstrate that the average loss value of the proposed model is approximately 0.45. Moreover, the recognition accuracy of the model for oil condition on the training and testing sets reaches 90.51% and 93.08%, respectively, while the accuracy of other algorithms remains below 90%. Thus, the model constructed in this study exhibits excellent performance in terms of loss value and recognition accuracy, providing reliable technical support for offshore oil monitoring and contributing to the promotion of sustainable utilization and conservation of marine resources.

PMID:40131882 | DOI:10.1371/journal.pone.0319858

IEEE Trans Pattern Anal Mach Intell. 2025 Mar 25;PP. doi: 10.1109/TPAMI.2025.3554560. Online ahead of print.

ABSTRACT

Recent advancements in model pruning have focused on developing new algorithms and improving upon benchmarks. However, the practical application of these algorithms across various models and platforms remains a significant challenge. To address this challenge, we propose ONNXPruner, a versatile pruning adapter designed for the ONNX format models. ONNXPruner streamlines the adaptation process across diverse deep learning frameworks and hardware platforms. A novel aspect of ONNXPruner is its use of node association trees, which automatically adapt to various model architectures. These trees clarify the structural relationships between nodes, guiding the pruning process, particularly highlighting the impact on interconnected nodes. Furthermore, we introduce a tree-level evaluation method. By leveraging node association trees, this method allows for a comprehensive analysis beyond traditional single-node evaluations, enhancing pruning performance without the need for extra operations. Experiments across multiple models and datasets confirm ONNXPruner's strong adaptability and increased efficacy. Our work aims to advance the practical application of model pruning.

PMID:40131753 | DOI:10.1109/TPAMI.2025.3554560

IEEE Trans Cybern. 2025 Mar 25;PP. doi: 10.1109/TCYB.2025.3549583. Online ahead of print.

ABSTRACT

The neural activity patterns of localized brain regions are crucial for recognizing brain intentions. However, existing electroencephalogram (EEG) decoding models, especially those based on deep learning, predominantly focus on global spatial features, neglecting valuable local information, potentially leading to suboptimal performance. Therefore, this study proposed a dynamic hierarchical convolutional attention network (DH-CAN) that comprehensively learned discriminative information from both global and local spatial domains, as well as from time-frequency domains in EEG signals. Specifically, a multiscale convolutional block was designed to dynamically capture time-frequency information. The channels of EEG signals were mapped to different brain regions based on motor imagery neural activity patterns. The spatial features, both global and local, were then hierarchically extracted to fully exploit the discriminative information. Furthermore, regional connectivity was established using a graph attention network, incorporating it into the local spatial features. Particularly, this study shared network parameters between symmetrical brain regions to better capture asymmetrical motor imagery patterns. Finally, the learned multilevel features were integrated through a high-level fusion layer. Extensive experimental results on two datasets demonstrated that the proposed model performed excellently across multiple evaluation metrics, exceeding existing benchmark methods. These findings suggested that the proposed model offered a novel perspective for EEG decoding research.

PMID:40131750 | DOI:10.1109/TCYB.2025.3549583

IEEE J Biomed Health Inform. 2025 Mar 25;PP. doi: 10.1109/JBHI.2025.3554495. Online ahead of print.

ABSTRACT

Training deep learning models for photoplethysmography(PPG)-based cuff-less blood pressure estimation often requires a substantial amount of labeled data collected through sophisticated medical instruments, posing significant challenges in practical applications. To address this issue, we propose Physiological Knowledge-Aware Contrastive Learning (PhysCL), a novel approach designed to reduce the dependence on labeled PPG data while improving blood pressure estimation accuracy. Specifically, PhysCL tackles the semantic consistency problem in contrastive learning by introducing a knowledge-aware augmentation bank, which generates positive physiological signal pairs using knowledge-based constraints during the contrastive pair generation. Additionally, we propose a contrastive feature reconstruction method to enhance feature diversity and prevent model collapse through feature re-sampling and re-weighting. We evaluate PhysCL on data from 106 subjects across the MIMIC III, MIMIC IV, and UQVS datasets under cross-dataset validation settings, comparing it against state-of-the-art contrastive learning methods and blood pressure estimation models. PhysCL achieves an average mean absolute error of 9.5/5.9 mmHg (systolic/diastolic) across the three datasets, using only 2% labeled data combined with 98% unlabeled data for pre-training and 5 samples for personalization, which represents a 6.2%/4.3% improvement, respectively, over the current best supervised methods. The ablation study provides further convincing evidence that the unlabeled data can be utilized to improve the existing cuff-less blood pressure estimation models and shed light on unsupervised contrastive learning for physiological signals.

PMID:40131744 | DOI:10.1109/JBHI.2025.3554495

Chirurgie (Heidelb). 2025 Mar 25. doi: 10.1007/s00104-025-02266-3. Online ahead of print.

ABSTRACT

BACKGROUND: The aim of the present study is to discuss the benefits of intraoperative frozen sections (FS) for the surgical management of endocrine tumors.

METHODS: A systematic search of the literature of the last ten years on FS in the field of oncologic endocrine surgery was carried out and a discussion based on the available evidence and experience of the authors is provided.

RESULTS: A group of publications focused on the role of intraoperative FS in thyroid surgery in identifying the malignant potential of thyroid nodules. The detection of lymph node metastasis and extrathyroidal growth in differentiated thyroid cancer (DTC) were also two other topical groups as well as the diagnosis of lymph node involvement based on stromal desmoplasia in medullary thyroid cancer (MTC). A further group investigated the possibilities of deep learning to overcome technical problems and another investigated the cost-benefit analyses. There is no relevant literature on the role of FS in the surgical treatment of parathyroid and adrenal cancers.

DISCUSSION: The synthesis of the available evidence suggests that FS investigations of the thyroid glands should be restricted to Bethesda V nodules. The technical limitations in the exclusion of vascular and capsular invasion make the FS unsuitable for follicular neoplasms and oncocytic lesions. The Delphi lymph node seems to be suitable for investigation using FS and when positive represents an indication for lymphadenectomy in cN0 patients. Larger studies are necessary in the future to confirm if the absence of desmoplasia with an intact tumor capsule can reliably justify omitting lymph node resection in MTC, independent of the calcitonin level. The costs and benefits depend on the individual context so that generalization is difficult. Deep learning models could generally improve the performance of FS analysis in the future.

CONCLUSION: In thyroid surgery awareness of the technical limitations of FS is crucial for correct implementation and thus to optimize its performance. A preoperative fine needle biopsy and surgical experience help in selecting the nodules that can benefit from FS. Deep image learning could help to overcome current problems in the future. In adrenal and parathyroid oncologic surgery FS do not play a relevant role.

PMID:40131405 | DOI:10.1007/s00104-025-02266-3

Brief Bioinform. 2025 Mar 4;26(2):bbaf065. doi: 10.1093/bib/bbaf065.

ABSTRACT

One of the main goals of metagenomic studies is to describe the taxonomic diversity of microbial communities. A crucial step in metagenomic analysis is metagenomic binning, which involves the (supervised) classification or (unsupervised) clustering of metagenomic sequences. Various machine learning models have been applied to address this task. In this review, the contributions of artificial neural networks (ANN) in the context of metagenomic binning are detailed, addressing both supervised, unsupervised, and semi-supervised approaches. 34 ANN-based binning tools are systematically compared, detailing their architectures, input features, datasets, advantages, disadvantages, and other relevant aspects. The findings reveal that deep learning approaches, such as convolutional neural networks and autoencoders, achieve higher accuracy and scalability than traditional methods. Gaps in benchmarking practices are highlighted, and future directions are proposed, including standardized datasets and optimization of architectures, for third-generation sequencing. This review provides support to researchers in identifying trends and selecting suitable tools for the metagenomic binning problem.

PMID:40131312 | DOI:10.1093/bib/bbaf065

Brief Bioinform. 2025 Mar 4;26(2):bbaf131. doi: 10.1093/bib/bbaf131.

ABSTRACT

DNA methylation plays a crucial role in human diseases pathogenesis. Substantial experimental evidence from clinical and biological studies has confirmed numerous methylation-disease associations, which provide valuable prior knowledge for advancing precision medicine through biomarker discovery and disease subtyping. To systematically mine reliable methylation prior knowledge from known DNA methylation-disease associations and develop robust computational methods for precision medicine applications, we propose MethPriorGCN. By integrating layer attention mechanisms and feature weighting mechanisms, MethPriorGCN not only identified reliable methylation digital biomarkers but also achieved superior disease subtype classification accuracy.

PMID:40131311 | DOI:10.1093/bib/bbaf131

Brief Bioinform. 2025 Mar 4;26(2):bbaf128. doi: 10.1093/bib/bbaf128.

ABSTRACT

Single-cell sequencing technology has enabled researchers to study cellular heterogeneity at the cell level. To facilitate the downstream analysis, clustering single-cell data into subgroups is essential. However, the high dimensionality, sparsity, and dropout events of the data make the clustering challenging. Currently, many deep learning methods have been proposed. Nevertheless, they either fail to fully utilize pairwise distances information between similar cells, or do not adequately capture their feature correlations. They cannot also effectively handle high-dimensional sparse data. Therefore, they are not suitable for high-fidelity clustering, leading to difficulties in analyzing the clear cell types required for downstream analysis. The proposed scSAMAC method integrates contrastive learning and negative binomial losses into a variational autoencoder, extracting features via contrastive unit similarity while preserving the intrinsic characteristics. This enhances the robustness and generalization during the clustering. In the contrastive learning, it constructs a mask module by adopting a negative sample generation method with gene feature saliency adjustment, which selects features more influential in the clustering phase and simulates data missing events. Additionally, it develops a novel loss, which consists of a soft k-means loss, a Wasserstein distance, and a contrastive loss. This fully utilizes data information and improves clustering performance. Furthermore, a multi-head attention mechanism module is applied to the latent variables at each layer of autoencoder to enhance feature correlation, integration, and information repair. Experimental results demonstrate that scSAMAC outperforms several state-of-the-art clustering methods.

PMID:40131310 | DOI:10.1093/bib/bbaf128

Chaos. 2025 Mar 1;35(3):033156. doi: 10.1063/5.0224768.

ABSTRACT

Detection of chaos in dynamical signals is an important and popular research area. Traditionally, the chaotic behavior is evaluated by calculating the Largest Lyapunov Exponent (LLE). However, calculating the LLE is sometimes difficult and requires specific data. Moreover, it introduces some subjective assumptions and is sometimes called a "manual" method. Therefore, there are many attempts to provide alternative ways to assess the dynamical signal as chaotic or deterministic. Some of them use deep learning methods. In this paper, we present a novel method of signal classification that is based on the assumption that it is easier to learn deterministic behavior than a chaotic one. We show that based on this assumption, it is possible to calculate the "amount of chaos" in the signal with the help of a simple LSTM (Long Short-Term Memory) neural network. The main advantage of this method is that-contrary to other deep learning-based methods-it does not require prior data to train the network as the results of the training process for a signal being classified are taken into account as the result of this evaluation. We confirm the method's validity using the publicly available dataset of chaotic and deterministic signals.

PMID:40131283 | DOI:10.1063/5.0224768

Expert Rev Anti Infect Ther. 2025 Mar 25. doi: 10.1080/14787210.2025.2484284. Online ahead of print.

ABSTRACT

INTRODUCTION: Traditional microbiological diagnostics face challenges in pathogen identification speed and antimicrobial resistance (AMR) evaluation. Artificial intelligence (AI) offers transformative solutions, necessitating a comprehensive review of its applications, advancements, and integration challenges in clinical microbiology.

AREAS COVERED: This review examines AI-driven methodologies, including machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs), for enhancing pathogen detection, AMR prediction, and diagnostic imaging. Applications in virology (e.g. COVID-19 RT-PCR optimization), parasitology (e.g. malaria detection), and bacteriology (e.g. automated colony counting) are analyzed. A literature search was conducted using PubMed, Scopus, and Web of Science (2018-2024), prioritizing peer-reviewed studies on AI's diagnostic accuracy, workflow efficiency, and clinical validation.

EXPERT OPINION: AI significantly improves diagnostic precision and operational efficiency but requires robust validation to address data heterogeneity, model interpretability, and ethical concerns. Future success hinges on interdisciplinary collaboration to develop standardized, equitable AI tools tailored for global healthcare settings. Advancing explainable AI and federated learning frameworks will be critical for bridging current implementation gaps and maximizing AI's potential in combating infectious diseases.

PMID:40131188 | DOI:10.1080/14787210.2025.2484284

Rev Sci Instrum. 2025 Mar 1;96(3):035117. doi: 10.1063/5.0231737.

ABSTRACT

According to a publication by the World Health Organization, depression is projected to emerge as the leading mental health issue. In the domain of affective computing, deep learning techniques are frequently employed to represent facial dynamics using both local and global perspectives for the purpose of automatic depression detection (ADD). Yet, current models overlook the crucial interplay between local and global dynamics in discerning the significant features essential for ADD. Addressing this oversight, a novel hybrid computational architecture, named LOGLFormer, has been introduced. This architecture integrates CNN-derived local attributes and transformer-sourced global patterns tailored for ADD. Within LOGLFormer, the design philosophies of ResNet and ViT inspire the CNN and transformer branches, respectively. The synergy of these branches encompasses local convolutional mechanisms, self-attention strategies, and multilayer perceptron entities. Furthermore, the intricacies arising from disparities in CNN and transformer feature sets are reconciled through the specially devised feature alignment module. Rigorous comparative analysis underscores the distinctive efficacy of the LOGLFormer in recognizing depression, notably outperforming several state-of-the-art techniques on two dedicated depression databases: AVEC2013 and AVEC2014. Code will be available at https://github.com/helang818/LOGLFormer.

PMID:40130984 | DOI:10.1063/5.0231737

Protein Sci. 2025 Apr;34(4):e70105. doi: 10.1002/pro.70105.

ABSTRACT

AlphaFold2 (AF2), a deep-learning-based model that predicts protein structures from their amino acid sequences, has recently been used to predict multiple protein conformations. In some cases, AF2 has successfully predicted both dominant and alternative conformations of fold-switching proteins, which remodel their secondary and/or tertiary structures in response to cellular stimuli. Whether AF2 has learned enough protein folding principles to reliably predict alternative conformations outside of its training set is unclear. Previous work suggests that AF2 predicted these alternative conformations by memorizing them during training. Here, we use CFold-an implementation of the AF2 network trained on a more limited subset of experimentally determined protein structures-to directly test how well the AF2 architecture predicts alternative conformations of fold switchers outside of its training set. We tested CFold on eight fold switchers from six protein families. These proteins-whose secondary structures switch between α-helix and β-sheet and/or whose hydrogen bonding networks are reconfigured dramatically-had not been tested previously, and only one of their alternative conformations was in CFold's training set. Successful CFold predictions would indicate that the AF2 architecture can predict disparate alternative conformations of fold-switched conformations outside of its training set, while unsuccessful predictions would suggest that AF2 predictions of these alternative conformations likely arise from association with structures learned during training. Despite sampling 1300-4300 structures/protein with various sequence sampling techniques, CFold predicted only one alternative structure outside of its training set accurately and with high confidence while also generating experimentally inconsistent structures with higher confidence. Though these results indicate that AF2's current success in predicting alternative conformations of fold switchers stems largely from its training data, results from a sequence pruning technique suggest developments that could lead to a more reliable generative model in the future.

PMID:40130805 | DOI:10.1002/pro.70105