Pulmonary infections pose a significant global health challenge to human life and health. In patients with chronic pulmonary diseases such as cystic fibrosis and bronchiectasis, structural abnormalities of the airways and impaired mucociliary clearance contribute to recurrent and challenging pulmonary infections. These infections are frequently complicated by antimicrobial resistance, making them difficult to treat with conventional antibiotics. As a result, phage therapy has emerged as a promising alternative for treating resistant pulmonary infections. Recently, the integration of artificial intelligence (AI) has improved the efficiency of phage selection. Nevertheless, the accuracy of predicting phage–bacterial host interactions remains limited, posing a significant obstacle to the clinical translation of phage-based therapies. To address this issue, we propose a deep Siamese network framework for precision phage selection in pulmonary infections. Specifically, we employ an identical model architecture to process both phage and host genomes. Initially, the genomic sequences of both phages and hosts are encoded into feature representations using k-mer segmentation followed by the skip-gram model. Subsequently, convolutional neural networks (CNNs) and Transformers are introduced to extract local and global features, respectively. Finally, the extracted features are fused to predict phage–host interactions. Experimental results on dataset created from the NCBI genome database demonstrate that our proposed method achieves superior performance in the precise identification of phages targeting specific bacterial hosts, thereby supporting its potential application in phage therapy for pulmonary infections.

Artificial intelligence (AI) has shown increasing potential in dental diagnostics, yet its accuracy for binary classification of dental caries across different imaging modalities remains unclear. This study aimed to systematically evaluate the diagnostic performance of AI models using clinical intraoral images and dental radiographs. Following the PRISMA-DTA guidelines, PubMed, Embase, Scopus, Web of Science, and IEEE Xplore were systematically searched for studies published between January 2015 and June 2025. Eligible studies applied AI models for caries diagnosis with extractable sensitivity and specificity. Data on dentition, dataset, analysis unit, caries prevalence in test dataset, and preprocessing methods were extracted. Reporting quality and risk of bias were assessed using CLAIM and QUADAS-2. Pooled estimates were calculated with a bivariate random-effects model, with subgroup analyses by image type and analytical unit. 25 studies met the criteria, and 13 were included in the meta-analysis. Pooled sensitivity, specificity, and the area under the curve (AUC) were 0.86, 0.91 and 0.94, respectively. Intraoral image-based models achieved higher sensitivity (0.88) and AUC (0.95), while radiograph-based models showed higher specificity (0.92). Tooth-level analyses yielded stable, clinically relevant performance (0.87/0.91). High heterogeneity (I² > 90%) was partly explained by image type, model architecture, reference standard variation, and test set caries prevalence. AI models showed good diagnostic accuracy for caries detection across imaging modalities and analytical units. However, given the substantial heterogeneity and limitations in study quality and reference standards, these summary estimates should be interpreted with caution. AI-based systems may serve as complementary decision-support tools in clinical practice, but further standardization, external validation, and high-quality multicenter studies are required before broad clinical implementation.

Accurate lung tumor segmentation using computed tomography (CT) scans is needed for efficient tumor treatment. However, the development of deep learning models is often constrained by strict patient privacy regulations that limit direct data sharing. This work presents a system that enables multi-institutional collaboration while training high-quality lung tumor segmentation models without requiring access to sensitive patient data. The proposed framework features the AuraViT suite, which includes the standard AuraViT—a hybrid model with 136 million parameters that combines a Vision Transformer (ViT) encoder, Atrous Spatial Pyramid Pooling (ASPP), and attention-gated residual connections—and the Lightweight AuraViT (LAURA) family (Small, Tiny, and Mobile). These variants are designed for resource-constrained environments and potential edge deployment scenarios. Training is conducted on publicly available datasets (MSD Lung and NSCLC) in a simulated five-client federated learning setup that emulates collaboration among institutions while ensuring patient privacy. The framework uses a federated learning setup with FedProx, adaptive weighted aggregation, and a dynamic virtual client strategy to handle data and system differences. The framework is further evaluated through ablation studies on model architecture and feature importance. The results show that the standard AuraViT-FL achieves a global mean Dice score of 80.81%, while maintaining performance close to centralized training. Additionally, the LAURA variations show a better trade-off between accuracy and efficiency. Notably, the Mobile variant with ∼5 M parameters reduces model complexity by over 96% while maintaining competitive performance (82.96% Dice on MSD Lung).

Medical artificial intelligence (AI) tools, including clinical language models, vision-language models and multimodal health record models, are used to summarize clinical notes, answer questions and support decisions. Their adaptation to new populations, specialties or care settings often relies on fine-tuning, prompting or retrieval from external knowledge bases. These strategies can scale poorly and risk contextual errors-outputs that appear plausible but miss critical patient or situational information. We envision context switching as an emergent solution. Context switching adjusts model reasoning at inference, without retraining. Generative models can tailor outputs to patient biology, care setting or disease. Multimodal models can switch between notes, laboratory results, imaging and genomics, even when some data are missing or delayed. Agent models can coordinate tools and roles based on task and user context. In each case, context switching enables medical AI to adapt across specialties, populations and geographies. This approach requires advances in data design, model architectures and evaluation frameworks, and establishes a foundation for medical AI that scales to an infinite number of contexts, while remaining reliable and suited to real-world care.

By influencing gene expression and contributing to epigenetic modifications, Ribonucleic Acid (RNA) 5-Hydroxymethylcytosine (5hmC) modification significantly affects cellular pathways. It plays an important role in complex regulatory networks and gene expression. Moreover, 5hmC modifications are linked to a variety of human diseases, including diabetes, cancer, and cardiovascular conditions. However, experimental methods to identify RNA 5hmC modifications, such as chromatography and Polymerase Chain Reaction (PCR) amplification, are costly and time-consuming. So, computational methods are necessary to predict these modifications. In this study, several feature descriptors were analyzed and compared to finalize the best ones. Different deep-learning models were explored to design the proposed model architecture. Neighbourhood analysis was conducted on the dataset to provide insights into a deeper understanding of RNA 5hmC modifications. The proposed model, InTrans-RNA5hmC, is a dual-branch deep learning model that has two branches: the Inception branch and the Transformer branch. Word embeddings having the contextual information and language model embeddings from the RiboNucleic Acid Language Model (RiNALMo) were used as the finalized feature descriptors. InTrans-RNA5hmC outperformed existing SOTA methods, achieving 0.97 sensitivity, 0.985 balanced accuracy, and 0.985 F1 score on the Independent test set.

Making AI accessible for forensic DNA profile analysis.

Beyond the model: data infrastructure as the foundation for autonomous virtual laboratories.

Choosing not to know: The emotional and sociocultural architecture of pension willful ignorance.

ReVGG-R2Net: Optimized recurrent framework for microscopic blood cell segmentation.

Elevating adversarial robustness by contrastive multitasking defence in medical image segmentation.

Scaling up self-supervised learning for improved surgical foundation models.

Accelerating VOC-OBP interaction screening via machine learning and molecular docking: Towards semiochemicals discovery for moth pests.

Weather nowcasting, defined as forecasting weather phenomena on time scales from minutes to several hours, is critical for mitigating the impacts of high-impact events such as flash floods, severe convective storms, and extreme precipitation. Traditional nowcasting approaches based on radar extrapolation and convection-permitting numerical weather prediction (NWP) exhibit fundamental limitations in representing rapid storm evolution, convective initiation, and localized extremes at short lead times. Recent advances in artificial intelligence (AI) and deep learning have enabled a new generation of nowcasting systems that learn complex spatiotemporal relationships directly from high-resolution radar, satellite, lightning, and NWP data. This paper provides a comprehensive review of AI-based precipitation nowcasting, covering data sources, model architectures, and evaluation methodologies. We discuss deterministic and probabilistic approaches, including convolutional recurrent networks, encoder–decoder convolutional neural networks, transformers, generative adversarial networks, diffusion models, and emerging physics-informed and hybrid AI–NWP systems. Opportunities such as improved short-lead forecast skill, multi-sensor fusion, probabilistic decision support, and enhanced forecast equity are examined alongside key challenges related to data quality, class imbalance, generalization, interpretability, and operational deployment. Finally, we highlight current research frontiers and methodological trends, outlining open challenges and promising directions for future AI-driven nowcasting systems at the PhD level and beyond.

Artificial intelligence (AI) techniques are increasingly applied to magnetic resonance imaging (MRI) for detecting temporomandibular joint (TMJ) anomalies; however, their overall diagnostic accuracy and generalizability remain uncertain. To systematically review and meta-analyse the diagnostic performance of AI models for TMJ anomaly detection on MRI and to identify factors influencing model performance. A comprehensive search of PubMed, Scopus, Embase, and Web of Science was conducted for studies published between January 2015 and September 2025. Two reviewers independently screened and extracted data. Eligible studies developed and tested AI, machine learning, or deep learning models on human TMJ MRI and reported quantitative performance metrics. Risk of bias was assessed using the QUADAS-2 tool. Pooled sensitivity and specificity were estimated using a bivariate random-effects model, while pooled accuracy was derived using logit transformation. Heterogeneity (I2) was explored through subgroup analyses by model architecture and validation strategy. Fourteen studies were included in the systematic review, of which six met the criteria for meta-analysis. Across these six studies, 18 models were analyzed for accuracy, 29 for sensitivity, and 24 for specificity. The pooled diagnostic accuracy was 0.487 (95% CI 0.403-0.571), with pooled sensitivity and specificity of 0.399 (95% CI 0.348-0.450) and 0.399 (95% CI 0.343-0.456), respectively, all showing substantial heterogeneity (I2 > 90%). Subgroup analyses indicated that advanced architectures such as ResNet-18, Inception v3, and EfficientNet-b4 achieved higher and more consistent diagnostic performance. Advanced deep learning architectures such as ResNet-18, Inception v3, and EfficientNet-b4 demonstrated superior diagnostic performance for detecting temporomandibular joint anomalies on MRI. These findings highlight the potential of AI-assisted MRI interpretation to improve diagnostic consistency, efficiency, and early detection of TMJ pathology. However, substantial heterogeneity and limited external validation currently limit clinical translation. Standardized multicenter studies and transparent model validation are essential to ensure reliable integration of AI tools into clinical TMJ imaging workflows.

Artificial intelligence (AI) is increasingly adopted in manufacturing for tasks such as automated inspection, predictive maintenance, and condition monitoring. However, the opaque, black-box nature of many AI models remains a major barrier to industrial trust, acceptance, and regulatory compliance. This study investigates how explainable artificial intelligence (XAI) techniques can be used to systematically open and interpret the internal reasoning of AI systems commonly deployed in manufacturing, rather than to optimise or compare model performance. A unified explainability-centred framework is proposed and applied across three representative manufacturing use cases encompassing heterogeneous data modalities and learning paradigms: vision-based classification of casting defects, vision-based localisation of metal surface defects, and unsupervised acoustic anomaly detection for machine condition monitoring. Diverse models are intentionally employed as representative black-box decision-makers to evaluate whether XAI methods can provide consistent, physically meaningful explanations independent of model architecture, task formulation, or supervision strategy. A range of established XAI techniques, including Grad-CAM, Integrated Gradients, Saliency Maps, Occlusion Sensitivity, and SHAP, are applied to expose model attention, feature relevance, and decision drivers across visual and acoustic domains. The results demonstrate that XAI enables alignment between model behaviour and physically interpretable defect and fault mechanisms, supporting transparent, auditable, and human-interpretable decision-making. By positioning explainability as a core operational requirement rather than a post hoc visual aid, this work contributes a cross-modal framework for trustworthy AI in manufacturing, aligned with Industry 5.0 principles, human-in-the-loop oversight, and emerging expectations for transparent and accountable industrial AI systems.

Objective.During speech perception, properties of the acoustic stimulus can be reconstructed from the listener's brain using methods such as electroencephalography (EEG). Most studies employ the amplitude envelope as a target for decoding; however, speech acoustics can be characterised on multiple dimensions, including as spectral descriptors. The current study assesses how robustly an extended acoustic feature set can be decoded from EEG under varying levels of intelligibility and acoustic clarity.Approach.Analysis was conducted using EEG from 38 young adults who heard intelligible and non-intelligible speech that was either unprocessed or spectrally degraded using vocoding. We extracted a set of acoustic features which, alongside the envelope, characterised instantaneous properties of the speech spectrum (e.g. spectral slope) or spectral change over time (e.g. spectral flux). We establish the robustness of feature decoding by employing multiple model architectures and, in the case of linear decoders, by standardising decoding accuracy (Pearson'sr) using randomly permuted surrogate data.Main results. Linear models yielded the highestrrelative to non-linear models. However, the separate decoder architectures produced a similar pattern of results across features and experimental conditions. After convertingrvalues toZ-scores scaled by random data, we observed substantive differences in the noise floor between features. Decoding accuracy significantly varies by spectral degradation and speech intelligibility for some features, but such differences are reduced in the most robustly decoded features. This suggests acoustic feature reconstruction is primarily driven by generalised auditory processing.Significance. Our results demonstrate that linear decoders perform comparably to non-linear decoders in capturing the EEG response to speech acoustic properties beyond the amplitude envelope, with the reconstructive accuracy of some features also associated with understanding and spectral clarity. This sheds light on how sound properties are differentially represented by the brain and shows potential for clinical applications moving forward.

Polar sea ice is undergoing rapid change, with recent record-low extents in both hemispheres, raising the demand for skillful predictions from days to seasons for navigation, ecosystem management, and climate risk assessment. Accurate sea ice prediction is essential for understanding coupled climate processes, supporting safe polar operations, and informing adaptation strategies. Physics-based numerical models remain the backbone of operational forecasting, but their skill is limited by uncertainties in coupled ocean–ice–atmosphere processes, parameterizations, and sparse observations, especially in the marginal ice zone and during melt seasons. Statistical and empirical models can provide useful baselines for low-dimensional indices or short lead times, yet they often struggle to represent high-dimensional, nonlinear interactions and regime shifts. This review synthesizes recent progress of DL for key sea ice prediction targets, including sea ice concentration/extent, thickness, and motion, and organizes methods into (i) sequential architectures (e.g., LSTM/GRU and temporal Transformers) for temporal dependencies, (ii) image-to-image and vision models (e.g., CNN/U-Net, vision Transformers, and diffusion or GAN-based generators) for spatial structures and downscaling, and (iii) spatiotemporal fusion frameworks that jointly model space–time dynamics. We further summarize hybrid strategies that integrate DL with numerical models through post-processing, emulation, and data assimilation, as well as physics-informed learning that embeds conservation laws or dynamical constraints. Despite rapid advances, challenges remain in generalization under non-stationary climate conditions, dataset shift, and physical consistency (e.g., mass/energy conservation), interpretability, and fair evaluation across regions and lead times. We conclude with practical recommendations for future research, including standardized benchmarks, uncertainty-aware probabilistic forecasting, physics-guided training and neural operators for long-range dynamics, and foundation models that leverage self-supervised pretraining on large-scale Earth observation archives.

A more complete map of the pattern of genetic variation among inbred mouse strains is essential for characterizing the genetic architecture of the many available mouse genetic models of important biomedical traits. Although structural variants (SVs) are a major component of genetic variation, they have not been adequately characterized among inbred strains due to methodological limitations. To address this, we generate high-quality long-read sequencing data for 40 inbred strains; and design a pipeline to optimally identify and validate different types of SVs. This generates a database for 40 inbred strains with 573,191 SVs, which include 10,815 duplications and 2,115 inversions, which also has 70 million SNPs and 7.5 million insertions/deletions. Analysis of this SV database identifies an SV that can be one component of a bi-genic model for lymphoma susceptibility in SJL mice, which provides mechanistic insight into the genetic basis for susceptibility to murine (and potentially human) lymphomas.

Residential heating, ventilation, air conditioning (HVAC), and water heating systems account for approximately 51% of total household energy consumption in the United States, representing over 5.5 quadrillion BTUs annually [1]. Despite widespread adoption of connected thermostats and smart water heaters, contemporary residential energy management platforms remain fundamentally constrained by device-centric architectures that lack semantic interoperability, suffer from sparse telemetry collection, and operate without predictive optimization capabilities. These systems function as isolated control points rather than as integrated climate ecosystems capable of responding to building thermal dynamics, occupant behavior patterns, distributed energy resource availability, and grid conditions. This paper introduces a comprehensive reference architecture for Climate Intelligence Systems (CIS) that transcends current limitations through four foundational pillars: cryptographically anchored device identity frameworks, metadata-driven equipment modeling hierarchies, cloud-hosted digital twin simulation environments, and predictive machine learning optimization pipelines [2], [3]. The proposed architecture enables anticipatory comfort management that pre-conditions spaces based on forecast weather patterns and predicted occupancy, orchestrates distributed energy resources including rooftop photovoltaic arrays and battery storage systems, and provides proactive grid-responsive demand flexibility without compromising occupant comfort or safety. We present a complete four-layer architectural model encompassing device/field infrastructure, connectivity/identity frameworks, cloud intelligence platforms, and human-facing experience layers. The architecture is augmented with detailed system interaction diagrams, digital twin synchronization pipelines, and demand response control flows that demonstrate practical implementation patterns. Preliminary deployment insights indicate 18-24% reductions in compressor short-cycling events, 12-15% improvements in thermal prediction accuracy under varying weather conditions, and 35-42% increases in reliable demand response participation compared to rule-based approaches. The resulting framework provides a coherent, cryptographically secure, and operationally scalable climate management ecosystem that addresses fundamental architectural limitations in today's smart home platforms while establishing a foundation for next-generation residential cyber-physical systems capable of supporting both individual household optimization and grid-scale energy orchestration.

Detecting multi-layered financial crime networks remains a major challenge in anti-money laundering (AML) due to high false positives, evolving adversarial behaviors, limited relational modeling, and stringent regulatory requirements for transparency and auditability. This study presents a novel, integrated Graph Neural Network–Explainable Artificial Intelligence (GNN-XAI) framework optimized for multi-layered AML detection, explicitly addressing accuracy, scalability, privacy preservation, and regulatory compliance. A heterogeneous Graph Attention Network architecture models transaction flows, entity relationships, device linkages, and temporal interactions within complex financial ecosystems. Over 400 engineered features capture velocity patterns, behavioral deviations, network centrality, and geopolitical risk. The framework was validated on benchmark datasets (IEEE-CIS Fraud Detection, Kaggle Credit Card Fraud, Elliptic Bitcoin) and large-scale synthetic transaction graphs. Results demonstrate robust performance, achieving an AUC-ROC of 0.874, precision of 89.3%, recall of 82.1%, and F1-score of 0.857, outperforming XGBoost and conventional GCN baselines by 5–6%. Relational features accounted for over 51% of predictive contribution. SHAP, LIME, and attention-based explanations enabled regulator-ready interpretability, supporting compliance, auditability, and supervisory review. Scalability experiments confirmed stable performance on networks exceeding 5.6 million transactions with sub-millisecond inference latency, while federated learning and differential privacy ensured viable privacy-utility trade-offs. The findings demonstrate that GNN-XAI architectures provide a practical, regulation-aligned pathway for next-generation AML systems.