Clin Radiol. 2024 Aug 8:S0009-9260(24)00414-8. doi: 10.1016/j.crad.2024.08.002. Online ahead of print.

ABSTRACT

PURPOSE: The aim of this meta-analysis was to assess the diagnostic performance of deep learning (DL) and ultrasound in breast cancer diagnosis. Additionally, we categorized the included studies into two subgroups: B-mode ultrasound diagnostic subgroup and multimodal ultrasound diagnostic subgroup, and compared the performance differences of DL algorithms in breast cancer diagnosis using only B-mode ultrasound or multimodal ultrasound.

METHODS: We conducted a comprehensive search for relevant studies published from January 01, 2017 to July 31, 2023 in the MEDLINE and EMBASE databases. The quality of the included studies was evaluated using the QUADAS-2 tool and radiomics quality scores (RQS). Meta-analysis was performed using R software. Inter-study heterogeneity was assessed by I^2 values and Q-test P-values, with sources of heterogeneity analyzed through a random effects model based on test results. Summary receiver operating characteristics (SROC) curves were used for meta-analysis across multiple trials, while combined sensitivity, specificity, and AUC were calculated to quantify prediction accuracy. Subgroup analysis and sensitivity analyses were also conducted to identify potential sources of study heterogeneity. Publication bias was assessed using the funnel plot method. (PROSPERO identifier: CRD42024545758).

RESULTS: The 20 studies included a total of 14,955 cases, with 4197 cases used for model testing. Among these cases were 1582 breast cancer patients and 2615 benign or other breast lesions. The combined sensitivity, specificity, and AUC values across all studies were found to be 0.93, 0.90, and 0.732, respectively. In subgroup analysis, the multimodal subgroup demonstrated superior performance with combined sensitivity, specificity, and AUC values of 0.93, 0.88, and 0.787, respectively; whereas the combined sensitivity, specificity, and AUC value for the model B subgroup was at a level of 0.92, 0.91, and 0.642, respectively.

CONCLUSIONS: The integration of DL with ultrasound demonstrates high accuracy in the adjunctive diagnosis of breast cancer, while the fusion of DL and multimodal breast ultrasound exhibits superior diagnostic efficacy compared to B-mode ultrasound alone.

PMID:39217049 | DOI:10.1016/j.crad.2024.08.002

Radiography (Lond). 2024 Aug 30:S1078-8174(24)00212-8. doi: 10.1016/j.radi.2024.08.012. Online ahead of print.

ABSTRACT

INTRODUCTION: Magnetic Resonance Imaging (MRI) has revolutionized the diagnosis and treatment of musculoskeletal disorders. Parallel imaging (PI) and compressed sensing (CS) techniques reduce scan time, but higher acceleration factors decrease image quality. Artificial intelligence has enhanced MRI reconstructions by integrating deep learning algorithms. Therefore, the study aims to review the impact of Artificial intelligence-assisted compressed sensing (AI-CS) and acceleration factors on scan time and image quality in musculoskeletal MRI.

METHODS: Database searches were completed across PubMed, Scopus, CINAHL, Web of Science, Cochrane Library, and Embase to identify relevant articles focusing on the application of AI-CS in musculoskeletal MRI between 2022 and 2024. We utilized the Preferred Reporting Items for Systematic Reviews and Meta-analysis guidelines to extract data from the selected studies.

RESULTS: Nine articles were included for the final review, with a total sample size of 730 participants. Of these, seven articles were rated as high, while two articles were considered to be of moderate quality. MRI examination with AI-CS showed scan time reduction of 18.9-38.8% for lumbar spine, 38-40% for shoulder, 54-75% for knee and 53-63% for ankle.

CONCLUSIONS: AI-CS showed a significant reduction in scan time and improved image quality for 2D and 3D sequences in musculoskeletal MRI compared with PI and CS. Determining the optimal acceleration factor necessary to achieve images with higher image quality compared to traditional PI techniques is required before clinical implementation. Higher acceleration factors currently lead to reduced image scores, although advancements in AI-CS are expected to address the limitation.

IMPLICATIONS OF PRACTICE: AI-CS in MRI improves patient care by shortening scan times, reducing patient discomfort and anxiety, and produces high quality images for accurate diagnosis.

PMID:39217002 | DOI:10.1016/j.radi.2024.08.012

Comput Biol Med. 2024 Aug 30;181:109052. doi: 10.1016/j.compbiomed.2024.109052. Online ahead of print.

ABSTRACT

Metastasis driven by cancer cell migration is the leading cause of cancer-related deaths. It involves significant changes in the organization of the cytoskeleton, which includes the actin microfilaments and the vimentin intermediate filaments. Understanding how these filament change cells from normal to invasive offers insights that can be used to improve cancer diagnosis and therapy. We have developed a computational, transparent, large-scale and imaging-based pipeline, that can distinguish between normal human cells and their isogenically matched, oncogenically transformed, invasive and metastasizing counterparts, based on the spatial organization of actin and vimentin filaments in the cell cytoplasm. Due to the intricacy of these subcellular structures, manual annotation is not trivial to automate. We used established deep learning methods and our new multi-attention channel architecture. To ensure a high level of interpretability of the network, which is crucial for the application area, we developed an interpretable global explainable approach correlating the weighted geometric mean of the total cell images and their local GradCam scores. The methods offer detailed, objective and measurable understanding of how different components of the cytoskeleton contribute to metastasis, insights that can be used for future development of novel diagnostic tools, such as a nanometer level, vimentin filament-based biomarker for digital pathology, and for new treatments that significantly can increase patient survival.

PMID:39216406 | DOI:10.1016/j.compbiomed.2024.109052

Comput Biol Med. 2024 Aug 30;181:109072. doi: 10.1016/j.compbiomed.2024.109072. Online ahead of print.

ABSTRACT

Automated generation of knowledge graphs that accurately capture published information can help with knowledge organization and access, which have the potential to accelerate discovery and innovation. Here, we present an integrated pipeline to construct a large-scale knowledge graph using large language models in an active learning setting. We apply our pipeline to the association of raw food, ingredients, and chemicals, a domain that lacks such knowledge resources. By using an iterative active learning approach of 4120 manually curated premise-hypothesis pairs as training data for ten consecutive cycles, the entailment model extracted 230,848 food-chemical composition relationships from 155,260 scientific papers, with 106,082 (46.0 %) of them never been reported in any published database. To augment the knowledge incorporated in the knowledge graph, we further incorporated information from 5 external databases and ontology sources. We then applied a link prediction model to identify putative food-chemical relationships that were not part of the constructed knowledge graph. Validation of the 443 hypotheses generated by the link prediction model resulted in 355 new food-chemical relationships, while results show that the model score correlates well (R2 = 0.70) with the probability of a novel finding. This work demonstrates how automated learning from literature at scale can accelerate discovery and support practical applications through reproducible, evidence-based capture of latent interactions of diverse entities, such as food and chemicals.

PMID:39216404 | DOI:10.1016/j.compbiomed.2024.109072

Comput Med Imaging Graph. 2024 Aug 28;117:102427. doi: 10.1016/j.compmedimag.2024.102427. Online ahead of print.

ABSTRACT

Automated cervical cancer screening through computer-assisted diagnosis has shown considerable potential to improve screening accessibility and reduce associated costs and errors. However, classification performance on whole slide images (WSIs) remains suboptimal due to patient-specific variations. To improve the precision of the screening, pathologists not only analyze the characteristics of suspected abnormal cells, but also compare them with normal cells. Motivated by this practice, we propose a novel cervical cell comparative learning method that leverages pathologist knowledge to learn the differences between normal and suspected abnormal cells within the same WSI. Our method employs two pre-trained YOLOX models to detect suspected abnormal and normal cells in a given WSI. A self-supervised model then extracts features for the detected cells. Subsequently, a tailored Transformer encoder fuses the cell features to obtain WSI instance embeddings. Finally, attention-based multi-instance learning is applied to achieve classification. The experimental results show an AUC of 0.9319 for our proposed method. Moreover, the method achieved professional pathologist-level performance, indicating its potential for clinical applications.

PMID:39216344 | DOI:10.1016/j.compmedimag.2024.102427

Mol Ecol Resour. 2024 Aug 31:e14015. doi: 10.1111/1755-0998.14015. Online ahead of print.

ABSTRACT

Leveraging past allele frequencies has proven to be key for identifying the impact of natural selection across time. However, this approach suffers from imprecise estimations of the intensity (s) and timing (T) of selection, particularly when ancient samples are scarce in specific epochs. Here, we aimed to bypass the computation of allele frequencies across arbitrarily defined past epochs and refine the estimations of selection parameters by implementing convolutional neural networks (CNNs) algorithms that directly use ancient genotypes sampled across time. Using computer simulations, we first show that genotype-based CNNs consistently outperform an approximate Bayesian computation (ABC) approach based on past allele frequency trajectories, regardless of the selection model assumed and the number of available ancient genotypes. When applying this method to empirical data from modern and ancient Europeans, we replicated the reported increased number of selection events in post-Neolithic Europe, independently of the continental subregion studied. Furthermore, we substantially refined the ABC-based estimations of s and T for a set of positively and negatively selected variants, including iconic cases of positive selection and experimentally validated disease-risk variants. Our CNN predictions support a history of recent positive and negative selection targeting variants associated with host defence against pathogens, aligning with previous work that highlights the significant impact of infectious diseases, such as tuberculosis, in Europe. These findings collectively demonstrate that detecting the footprints of natural selection on ancient genomes is crucial for unravelling the history of severe human diseases.

PMID:39215552 | DOI:10.1111/1755-0998.14015

Cancer Med. 2024 Aug;13(16):e70069. doi: 10.1002/cam4.70069.

ABSTRACT

OBJECTIVE: Breast cancer is one of the leading cancer causes among women worldwide. It can be classified as invasive ductal carcinoma (IDC) or metastatic cancer. Early detection of breast cancer is challenging due to the lack of early warning signs. Generally, a mammogram is recommended by specialists for screening. Existing approaches are not accurate enough for real-time diagnostic applications and thus require better and smarter cancer diagnostic approaches. This study aims to develop a customized machine-learning framework that will give more accurate predictions for IDC and metastasis cancer classification.

METHODS: This work proposes a convolutional neural network (CNN) model for classifying IDC and metastatic breast cancer. The study utilized a large-scale dataset of microscopic histopathological images to automatically perceive a hierarchical manner of learning and understanding.

RESULTS: It is evident that using machine learning techniques significantly (15%-25%) boost the effectiveness of determining cancer vulnerability, malignancy, and demise. The results demonstrate an excellent performance ensuring an average of 95% accuracy in classifying metastatic cells against benign ones and 89% accuracy was obtained in terms of detecting IDC.

CONCLUSIONS: The results suggest that the proposed model improves classification accuracy. Therefore, it could be applied effectively in classifying IDC and metastatic cancer in comparison to other state-of-the-art models.

PMID:39215495 | DOI:10.1002/cam4.70069

BMC Bioinformatics. 2024 Aug 30;25(1):284. doi: 10.1186/s12859-024-05917-0.

ABSTRACT

Post-translational modifications (PTMs) are fundamental to essential biological processes, exerting significant influence over gene expression, protein localization, stability, and genome replication. Sumoylation, a PTM involving the covalent addition of a chemical group to a specific protein sequence, profoundly impacts the functional diversity of proteins. Notably, identifying sumoylation sites has garnered significant attention due to their crucial roles in proteomic functions and their implications in various diseases, including Parkinson's and Alzheimer's. Despite the proposal of several computational models for identifying sumoylation sites, their effectiveness could be improved by the limitations associated with conventional learning methodologies. In this study, we introduce pseudo-position-specific scoring matrix (PsePSSM), a robust computational model designed for accurately predicting sumoylation sites using an optimized deep learning algorithm and efficient feature extraction techniques. Moreover, to streamline computational processes and eliminate irrelevant and noisy features, sequential forward selection using a support vector machine (SFS-SVM) is implemented to identify optimal features. The multi-layer Deep Neural Network (DNN) is a robust classifier, facilitating precise sumoylation site prediction. We meticulously assess the performance of PSSM-Sumo through a tenfold cross-validation approach, employing various statistical metrics such as the Matthews Correlation Coefficient (MCC), accuracy, sensitivity, specificity, and the Area under the ROC Curve (AUC). Comparative analyses reveal that PSSM-Sumo achieves an exceptional average prediction accuracy of 98.71%, surpassing existing models. The robustness and accuracy of the proposed model position it as a promising tool for advancing drug discovery and the diagnosis of diverse diseases linked to sumoylation sites.

PMID:39215231 | DOI:10.1186/s12859-024-05917-0

Commun Biol. 2024 Aug 30;7(1):1062. doi: 10.1038/s42003-024-06714-4.

ABSTRACT

Multiplexed imaging technologies have made it possible to interrogate complex tissue microenvironments at sub-cellular resolution within their native spatial context. However, proper quantification of this complexity requires the ability to easily and accurately segment cells into their sub-cellular compartments. Within the supervised learning paradigm, deep learning-based segmentation methods demonstrating human level performance have emerged. However, limited work has been done in developing such generalist methods within the unsupervised context. Here we present an easy-to-use unsupervised segmentation (UNSEG) method that achieves deep learning level performance without requiring any training data via leveraging a Bayesian-like framework, and nucleus and cell membrane markers. We show that UNSEG is internally consistent and better at generalizing to the complexity of tissue morphology than current deep learning methods, allowing it to unambiguously identify the cytoplasmic compartment of a cell, and localize molecules to their correct sub-cellular compartment. We also introduce a perturbed watershed algorithm for stably and automatically segmenting a cluster of cell nuclei into individual nuclei that increases the accuracy of classical watershed. Finally, we demonstrate the efficacy of UNSEG on a high-quality annotated gastrointestinal tissue dataset we have generated, on publicly available datasets, and in a range of practical scenarios.

PMID:39215205 | DOI:10.1038/s42003-024-06714-4

Sci Rep. 2024 Aug 30;14(1):20250. doi: 10.1038/s41598-024-70989-0.

ABSTRACT

The presence of toxic chemicals in water, including heavy metals like mercury and lead, organic pollutants such as pesticides, and industrial chemicals from runoff and discharges, poses critical public health and environmental risks leading to severe health issues and ecosystem damage; education plays a crucial role in mitigating these effects by enhancing awareness, promoting sustainable practices, and integrating environmental science into curricula to empower individuals to address and advocate for effective solutions to water pollution. However, the educational transformation should be accompanied with a technical process which can be eventually transferred to society to empower environmental education. In this study, carbonaceous material derived from Haematoxylum campechianum (CM-HC) was utilized for removing 3-nitrophenol (3-Nph) from aqueous solutions. The novelty of this research utilizes Haematoxylum campechianum bark and coconut shell, abundant agricultural wastes in Campeche, Mexico, for toxin removal, enhancing the adsorption process through artificial neural networks and genetic algorithms to optimize conditions and maximize the absorption efficiency. CM-HC's surface morphology was analyzed using scanning electron microscopy (SEM/EDS), BET method, X-ray powder diffraction (XRD), and pHpzc. Kinetic models including pseudo-first-order (PFO), pseudo-second-order (PSO), and Elovich were applied to fit the data. Adsorption isotherms were determined at varying pH (3-8), adsorbent dosages (2-10 g/L), and temperatures (300.15-330.15 K), employing Langmuir, Freundlich, Temkin, and Redlich-Peterson models. PSO kinetics demonstrated a good fit (R2 > 0.98) for Ci = 50-100 mg/L, indicating a chemical adsorption mechanism. The Langmuir isotherm model exhibited the best fit, confirming chemical adsorption, with a maximum adsorption capacity (Qm) of 236.156 mg/g at T = 300.15 K, pH = 6, contact time = 3 h, and 2 g/L adsorbent dosage. Lower temperatures favored exothermic adsorption. Artificial neural networks (ANNs) were employed for deep learning, optimizing the predictive model for removal percentage. Correlation heat maps highlighted positive correlations between time, dosage, and removal percentage, emphasizing the impact of initial concentration on efficiency. ANN modeling, incorporating iterative optimization, yielded highly accurate predictions, aligned closely with experimental results. The study showcases the success of deep learning in optimizing adsorption processes, emphasizing the importance of diverse correlation algorithms for comprehensive insights into competitive adsorption dynamics. The 5-14-14-1 deep learning architecture, fine-tuned over 228 epochs, demonstrated strong performance with mean squared error (MSE) values of 4.07, 18.406, and 6.2122 for training, testing, and total datasets, respectively, and high R-squared values. Graphical analysis showed a solid linear correlation between experimental and simulated removal percentages, emphasizing the need to consider more than just testing data for optimization. Experimental validation confirmed a 98.77% removal efficiency, illustrating the effectiveness of combining deep learning with genetic algorithms, and highlighting the necessity of experimental trials to verify computational predictions. It is concluded that the carbonaceous material from Haematoxylum campechianum waste (CM-HC) is an effective, low-cost adsorbent for removing 3-nitrophenol from aqueous solutions, achieving optimal removal at pH 6 and 300.15 K with a maximum adsorption capacity of 236.156 mg/g, following Langmuir model and pseudo-second order kinetics. The validated ANN model offers a reliable tool for practical applications in environmental remediation, advancing both environmental science and educational innovation by integrating artificial neural networks and data science methodologies into student learning experiences.

PMID:39215127 | DOI:10.1038/s41598-024-70989-0

Sci Rep. 2024 Aug 30;14(1):20174. doi: 10.1038/s41598-024-70701-2.

ABSTRACT

With large chess-playing neural network models like AlphaZero contesting the state of the art within the world of computerised chess, two challenges present themselves: the question of how to explain the domain knowledge internalised by such models, and the problem that such models are not made openly available. This work presents the re-implementation of the concept detection methodology applied to AlphaZero, by using large, open-source chess models with comparable performance. We obtain results similar to those achieved when applying this methodology to AlphaZero, while relying solely on open-source resources. We also present a novel explainable AI (XAI) method, which is guaranteed to highlight exhaustively and exclusively the information used by the explained model. This method generates visual explanations tailored to domains characterised by discrete input spaces, as is the case for chess. Our presented method has the desirable property of controlling the information flow between any input vector and the given model, which in turn provides strict guarantees regarding what information is used by the trained model during inference. We demonstrate the viability of our method by applying it to standard 8 × 8 chess, using large open-source chess models.

PMID:39215114 | DOI:10.1038/s41598-024-70701-2

Sci Rep. 2024 Aug 30;14(1):20227. doi: 10.1038/s41598-024-71066-2.

ABSTRACT

This study aimed to develop a deep-learning (DL) based method for three-dimensional (3D) segmentation of the upper urinary tract (UUT), including ureter and renal pelvis, on non-enhanced computed tomography (NECT) scans. A total of 150 NECT scans with normal appearance of the left UUT were chosen for this study. The dataset was divided into training (n = 130) and validation sets (n = 20). The test set contained 29 randomly chosen cases with computed tomography urography (CTU) and NECT scans, all with normal appearance of the left UUT. An experienced radiologist marked out the left renal pelvis and ureter on each scan. Two types of frameworks (entire and sectional) with three types of DL models (basic UNet, UNet3 + and ViT-UNet) were developed, and evaluated. The sectional framework with basic UNet model achieved the highest mean precision (85.5%) and mean recall (71.9%) on the test set compared to all other tested methods. Compared with CTU scans, this method had higher axial UUT recall than CTU (82.5% vs 69.1%, P < 0.01). This method achieved similar or better visualization of UUT than CTU in many cases, however, in some cases, it exhibited a non-ignorable false-positive rate. The proposed DL method demonstrates promising potential in automated 3D UUT segmentation on NECT scans. The proposed DL models could remarkably improve the efficiency of UUT reconstruction, and have the potential to save many patients from invasive examinations such as CTU. DL models could also serve as a valuable complement to CTU.

PMID:39215092 | DOI:10.1038/s41598-024-71066-2

Commun Biol. 2024 Aug 30;7(1):1067. doi: 10.1038/s42003-024-06734-0.

ABSTRACT

The identification of miRNA-disease associations is crucial for early disease prevention and treatment. However, it is still a computational challenge to accurately predict such associations due to improper information encoding. Previous methods characterize miRNA-disease associations only from single levels, causing the loss of multi-level association information. In this study, we propose TriFusion, a powerful and interpretable deep learning framework for miRNA-disease association prediction. It develops a tri-channel architecture to encode the association features of miRNAs and diseases from different levels and designs a feature fusion encoder to smoothly fuse these features. After training and testing, TriFusion outperforms other leading methods and offers strong interpretability through its learned representations. Furthermore, TriFusion is applied to three high-risk sexually associated cancers (ovarian, breast, and prostate cancers) and exhibits remarkable ability in the identification of miRNAs associated with the three diseases.

PMID:39215090 | DOI:10.1038/s42003-024-06734-0

Sci Rep. 2024 Aug 30;14(1):20218. doi: 10.1038/s41598-024-70593-2.

ABSTRACT

In therapeutic diagnostics, early diagnosis and monitoring of heart disease is dependent on fast time-series MRI data processing. Robust encryption techniques are necessary to guarantee patient confidentiality. While deep learning (DL) algorithm have improved medical imaging, privacy and performance are still hard to balance. In this study, a novel approach for analyzing homomorphivally-encrypted (HE) time-series MRI data is introduced: The Multi-Faceted Long Short-Term Memory (MF-LSTM). This method includes privacy protection. The MF-LSTM architecture protects patient's privacy while accurately categorizing and forecasting cardiac disease, with accuracy (97.5%), precision (96.5%), recall (98.3%), and F1-score (97.4%). While segmentation methods help to improve interpretability by identifying important region in encrypted MRI images, Generalized Histogram Equalization (GHE) improves image quality. Extensive testing on selected dataset if encrypted time-series MRI images proves the method's stability and efficacy, outperforming previous approaches. The finding shows that the suggested technique can decode medical image to expose visual representation as well as sequential movement while protecting privacy and providing accurate medical image evaluation.

PMID:39215022 | DOI:10.1038/s41598-024-70593-2

Surg Endosc. 2024 Aug 30. doi: 10.1007/s00464-024-11208-9. Online ahead of print.

ABSTRACT

BACKGROUND: There is an increasing demand for automated surgical skill assessment to solve issues such as subjectivity and bias that accompany manual assessments. This study aimed to verify the feasibility of assessing surgical skills using a surgical phase recognition model.

METHODS: A deep learning-based model that recognizes five surgical phases of laparoscopic sigmoidectomy was constructed, and its ability to distinguish between three skill-level groups-the expert group, with a high Endoscopic Surgical Skill Qualification System (ESSQS) score (26 videos); the intermediate group, with a low ESSQS score (32 videos); and the novice group, with an experience of < 5 colorectal surgeries (27 videos)-was assessed. Furthermore, 1 272 videos were divided into three groups according to the ESSQS score: ESSQS-high, ESSQS-middle, and ESSQS-low groups, and whether they could be distinguished by the score calculated by multiple regression analysis of the parameters from the model was also evaluated.

RESULTS: The time for mobilization of the colon, time for dissection of the mesorectum plus transection of the rectum plus anastomosis, and phase transition counts were significantly shorter or less in the expert group than in the intermediate (p = 0.0094, 0.0028, and < 0.001, respectively) and novice groups (all p < 0.001). Mesorectal excision time was significantly shorter in the expert group than in the novice group (p = 0.0037). The group with higher ESSQS scores also had higher AI scores.

CONCLUSION: This model has the potential to be applied to automated skill assessments.

PMID:39214877 | DOI:10.1007/s00464-024-11208-9

J Chem Inf Model. 2024 Aug 28. doi: 10.1021/acs.jcim.4c01174. Online ahead of print.

ABSTRACT

Binding of partners and mutations highly affects the conformational dynamics of KRAS4B, which is of significance for deeply understanding its function. Gaussian accelerated molecular dynamics (GaMD) simulations followed by deep learning (DL) and principal component analysis (PCA) were carried out to probe the effect of G12C and binding of three partners NF1, RAF1, and SOS1 on the conformation alterations of KRAS4B. DL reveals that G12C and binding of partners result in alterations in the contacts of key structure domains, such as the switch domains SW1 and SW2 together with the loops L4, L5, and P-loop. Binding of NF1, RAF1, and SOS1 constrains the structural fluctuation of SW1, SW2, L4, and L5; on the contrary, G12C leads to the instability of these four structure domains. The analyses of free energy landscapes (FELs) and PCA also show that binding of partners maintains the stability of the conformational states of KRAS4B while G12C induces greater mobility of the switch domains SW1 and SW2, which produces significant impacts on the interactions of GTP with SW1, L4, and L5. Our findings suggest that partner binding and G12C play important roles in the activity and allosteric regulation of KRAS4B, which may theoretically aid in further understanding the function of KRAS4B.

PMID:39197061 | DOI:10.1021/acs.jcim.4c01174

Database (Oxford). 2024 Aug 28;2024:baae079. doi: 10.1093/database/baae079.

ABSTRACT

Biomedical relation extraction from scientific publications is a key task in biomedical natural language processing (NLP) and can facilitate the creation of large knowledge bases, enable more efficient knowledge discovery, and accelerate evidence synthesis. In this paper, building upon our previous effort in the BioCreative VIII BioRED Track, we propose an enhanced end-to-end pipeline approach for biomedical relation extraction (RE) and novelty detection (ND) that effectively leverages existing datasets and integrates state-of-the-art deep learning methods. Our pipeline consists of four tasks performed sequentially: named entity recognition (NER), entity linking (EL), RE, and ND. We trained models using the BioRED benchmark corpus that was the basis of the shared task. We explored several methods for each task and combinations thereof: for NER, we compared a BERT-based sequence labeling model that uses the BIO scheme with a span classification model. For EL, we trained a convolutional neural network model for diseases and chemicals and used an existing tool, PubTator 3.0, for mapping other entity types. For RE and ND, we adapted the BERT-based, sentence-bound PURE model to bidirectional and document-level extraction. We also performed extensive hyperparameter tuning to improve model performance. We obtained our best performance using BERT-based models for NER, RE, and ND, and the hybrid approach for EL. Our enhanced and optimized pipeline showed substantial improvement compared to our shared task submission, NER: 93.53 (+3.09), EL: 83.87 (+9.73), RE: 46.18 (+15.67), and ND: 38.86 (+14.9). While the performances of the NER and EL models are reasonably high, RE and ND tasks remain challenging at the document level. Further enhancements to the dataset could enable more accurate and useful models for practical use. We provide our models and code at https://github.com/janinaj/e2eBioMedRE/. Database URL: https://github.com/janinaj/e2eBioMedRE/.

PMID:39197056 | DOI:10.1093/database/baae079

PLoS One. 2024 Aug 28;19(8):e0309459. doi: 10.1371/journal.pone.0309459. eCollection 2024.

ABSTRACT

The reliable operation of electrical power transmission systems is crucial for ensuring consumer's stable and uninterrupted electricity supply. Faults in electrical power transmission systems can lead to significant disruptions, economic losses, and potential safety hazards. A protective approach is essential for transmission lines to guard against faults caused by natural disturbances, short circuits, and open circuit issues. This study employs an advanced artificial neural network methodology for fault detection and classification, specifically distinguishing between single-phase fault and fault between all three phases and three-phase symmetrical fault. For fault data creation and analysis, we utilized a collection of line currents and voltages for different fault conditions, modelled in the MATLAB environment. Different fault scenarios with varied parameters are simulated to assess the applied method's detection ability. We analyzed the signal data time series analysis based on phase line current and phase line voltage. We employed SMOTE-based data oversampling to balance the dataset. Subsequently, we developed four advanced machine-learning models and one deep-learning model using signal data from line currents and voltage faults. We have proposed an optimized novel glassbox Explainable Boosting (EB) approach for fault detection. The proposed EB method incorporates the strengths of boosting and interpretable tree models. Simulation results affirm the high-efficiency scores of 99% in detecting and categorizing faults on transmission lines compared to traditional fault detection state-of-the-art methods. We conducted hyperparameter optimization and k-fold validations to enhance fault detection performance and validate our approach. We evaluated the computational complexity of fault detection models and augmented it with eXplainable Artificial Intelligence (XAI) analysis to illuminate the decision-making process of the proposed model for fault detection. Our proposed research presents a scalable and adaptable method for advancing smart grid technology, paving the way for more secure and efficient electrical power transmission systems.

PMID:39196913 | DOI:10.1371/journal.pone.0309459

PLoS One. 2024 Aug 28;19(8):e0304476. doi: 10.1371/journal.pone.0304476. eCollection 2024.

ABSTRACT

Domain Generation Algorithms (DGAs) are used by malware to generate pseudorandom domain names to establish communication between infected bots and command and control servers. While DGAs can be detected by machine learning (ML) models with great accuracy, offering DGA detection as a service raises privacy concerns when requiring network administrators to disclose their DNS traffic to the service provider. The main scientific contribution of this paper is to propose the first end-to-end framework for privacy-preserving classification as a service of domain names into DGA (malicious) or non-DGA (benign) domains. Our framework achieves these goals by carefully designed protocols that combine two privacy-enhancing technologies (PETs), namely secure multi-party computation (MPC) and differential privacy (DP). Through MPC, our framework enables an enterprise network administrator to outsource the problem of classifying a DNS (Domain Name System) domain as DGA or non-DGA to an external organization without revealing any information about the domain name. Moreover, the service provider's ML model used for DGA detection is never revealed to the network administrator. Furthermore, by using DP, we also ensure that the classification result cannot be used to learn information about individual entries of the training data. Finally, we leverage post-training float16 quantization of deep learning models in MPC to achieve efficient, secure DGA detection. We demonstrate that by using quantization achieves a significant speed-up, resulting in a 23% to 42% reduction in inference runtime without reducing accuracy using a three party secure computation protocol tolerating one corruption. Previous solutions are not end-to-end private, do not provide differential privacy guarantees for the model's outputs, and assume that model embeddings are publicly known. Our best protocol in terms of accuracy runs in about 0.22s.

PMID:39196905 | DOI:10.1371/journal.pone.0304476

Chem Res Toxicol. 2024 Aug 28. doi: 10.1021/acs.chemrestox.4c00199. Online ahead of print.

ABSTRACT

Cytochromes P450 (P450s or CYPs) are the most important phase I metabolic enzymes in the human body and are responsible for metabolizing ∼75% of the clinically used drugs. P450-mediated metabolism is also closely associated with the formation of toxic metabolites and drug-drug interactions. Therefore, it is of high importance to predict if a compound is the substrate of a given P450 in the early stage of drug development. In this study, we built the multitask learning models to simultaneously predict the substrates of five major drug-metabolizing P450 enzymes, namely, CYP3A4, 2C9, 2C19, 2D6, and 1A2, based on the collected substrate data sets. Compared to the single-task model and conventional machine learning models, the multitask fingerprints and graph neural networks model achieved superior performance with the average AUC values of 90.8% on the test set. Notably, the multitask model demonstrated its good performance on the small amount of substrate data sets such as CYP1A2, 2C9, and 2C19. In addition, the Shapley additive explanation and the attention mechanism were used to reveal specific substructures associated with P450 substrates, which were further confirmed and complemented by the substructure mining tool and the literature.

PMID:39196814 | DOI:10.1021/acs.chemrestox.4c00199