人工智能

Introduction Assessment format may influence the extent to which student learning approaches are reflected in performance, yet whether constructed-response descriptive assessments (DAs) and selected-response multiple-choice question assessments (MCQAs) differ in their sensitivity to variation in learning approach within formative physiology education has not been directly examined. This study tested whether baseline deep and surface learning approaches were differentially associated with performance in DAs and MCQAs within a longitudinal formative undergraduate medical physiology programme. Methods This three-month longitudinal observational study was conducted at a single medical college in South India. Among 150 invited first-year medical students, 109 completed the baseline Revised Two-Factor Study Process Questionnaire (R-SPQ-2F) and contributed to the study. Eight physiology topics, selected through a modified Delphi process, were taught sequentially and assessed on a rolling basis. For each topic, students completed both a DA and an MCQA in the same sitting, with items matched on construct and revised Bloom's taxonomy level. A linear mixed-effects model was used to test whether the association between learning approach and marks differed by assessment format, with a student-level random intercept to account for repeated observations. A post hoc inter-rater reliability audit was conducted on a randomly selected subset of DA scripts. Sensitivity analyses included a random-slopes model and a Deep-minus-Surface composite parameterisation. Results The association between learning approach and marks differed by assessment format, with significant format-by-deep-learning and format-by-surface-learning interactions (both p < 0.001). Within DAs, higher deep-learning scores were associated with higher marks (&#x3b2; = 0.032, p = 0.022), whereas higher surface-learning scores were associated with lower marks (&#x3b2; = -0.038, p = 0.003). Within MCQAs, neither learning-approach dimension was significantly associated with marks. The negative association between surface learning and DA performance remained robust across model specifications, whereas the positive association between deep learning and DA performance was attenuated in the random-slopes model (p = 0.072). The Deep-minus-Surface sensitivity analysis supported the same overall format-dependent pattern. A post hoc inter-rater reliability audit on 25% of DA scripts yielded strong agreement with an intraclass correlation coefficient (ICC) of 0.87. Conclusions Assessment format moderated the association between learning approach and assessment performance&#xa0;in this cohort: DAs showed clearer differentiation by learning approach than MCQAs, with poorer DA performance at higher surface-learning scores. These findings do not show that MCQAs&#xa0;reward surface learning but suggest that DAs may provide more discriminating information about variation in learning approach than MCQAs within a programmatic assessment framework.

Rain gauge networks are the core infrastructure for hydrological and water resource monitoring, flood control and disaster mitigation early warning, and water resource planning and regulation. The rationality of their layout directly determines the accuracy, representativeness, and economy of regional precipitation data acquisition. Considering that information entropy can accurately characterize the spatial distribution law and information complexity of rainfall, and spatiotemporal deep learning models have strong capabilities in fitting spatiotemporal features, this paper couples mutual information entropy with a spatiotemporal deep learning model and proposes a novel optimal layout method for rain gauge networks. Daily observed rainfall data from 50 ground-based rain gauges in the upper reaches of the Tuojiang River during 2015-2024, as well as the PERSIANN-CCS remote sensing precipitation product for the same period, were used in the study. A CNN-LSTM spatiotemporal deep learning model integrating spatial features and temporal dependence was constructed, coupled with the mutual information entropy index, and the GA-PSO hybrid optimization algorithm was applied for solution. The superiority of the proposed method was verified by comparison with the calculation results of the traditional mutual information entropy-based greedy optimization algorithm. The results show that the hybrid optimization algorithm driven by the spatiotemporal deep learning model coupled with mutual information entropy is significantly superior to the comparison algorithm in terms of the rationality of the station network structure, the ability to characterize spatial rainfall distribution, the control of average relative error, and the improvement of total information entropy. After optimization, the number of rain gauges in the upper reaches of the Tuojiang River can be reduced from 50 to 25. While greatly reducing the number of stations, the optimized network can still relatively accurately reflect the spatiotemporal characteristics of rainfall in the basin, which can provide a theoretical basis and technical support for the optimal layout of basin rain gauge networks and water resource management.

There are several clinical and research applications for determining the amount of brain tissue resected after epilepsy surgery; however, manual segmentation of postoperative magnetic resonance imaging (MRI) is imprecise and time-consuming. In this study, we developed and benchmarked ResectVol DL, a freely available deep learning-based tool that performs this task automatically. To create ResectVol DL, we trained a UNet-like deep learning model using postoperative T1-weighted MRI from epilepsy surgery patients and evaluated it against manual delineations (ground truth). ResectVol DL was also compared with three automated methods (ResectVol 1.1.2, DeepResection, and Auto3DSeg) using Dice similarity coefficient (DSC), Pearson correlation coefficient, and relative volume difference from manual segmentation. To assess false-positive detections and generalizability beyond epilepsy, we additionally processed images from healthy controls (no resection) and brain tumor cases. The final epilepsy cohort comprised 120 patients (57 women, mean age at surgery&#x2009;=&#x2009;31.5&#x2009;&#xb1;&#x2009;15.9&#x2009;[SD]&#x2009;years), split into training (n&#x2009;=&#x2009;72) and test (n&#x2009;=&#x2009;48) sets. An additional 42 images (22 healthy controls and 20 brain tumor cases) were included to test for false positives and generalizability. Segmentation performance differed across methods (Friedman test, p&#x2009;<&#x2009;.001). ResectVol Dl achieved the highest median DSC (.925), significantly outperforming ResectVol 1.1.2, DeepResection, and Auto3DSeg after Bonferroni correction. Volume-based metrics were similar for Auto3DSeg and ResectVol DL (r&#x2009;=&#x2009;.988, relative difference&#x2009;=&#x2009;8.4% vs. r&#x2009;=&#x2009;.985, 8.1%; no significant difference), yet Auto3DSeg produced three false-positive cavities in no-surgery controls (3/22, 95% confidence interval [CI] =&#x2009;3%-35%), whereas none was observed for ResectVol DL and DeepResection (0/22, 95% CI&#x2009;=&#x2009;0%-15%). ResectVol DL provides accurate, fully automated segmentation of postoperative resection cavities, offering a robust and reproducible methodological tool for large-scale postoperative imaging studies in epilepsy surgery. ResectVol DL also provides volumetric information derived from region labeling, which may serve as potential input for predictive models associated with surgery outcome; however, this application has not yet been validated.

With the continuous evolution toward sixth-generation (6G) wireless communication systems, emerging scenarios such as terahertz transmission, integrated sensing and communication (ISAC), and ultra-massive multiple-input multiple-output (MIMO) have significantly increased the complexity, nonlinearity, and uncertainty of wireless propagation environments. The conventional model-driven paradigm, established upon Shannon information theory and precise mathematical modeling, is increasingly constrained by model-mismatch issues in real-world deployments. This paper systematically reviews recent advances in deep learning-enabled physical-layer signal processing. We examine intelligent channel estimation, signal detection, and end-to-end communication systems based on autoencoder architectures. We then analyze key technical challenges-including interpretability, data dependence, computational complexity, privacy and security in distributed learning, and system-level performance-overhead trade-offs-along with state-of-the-art solution strategies such as deep unfolding, transfer learning, model compression, federated learning, and lightweight design. Future evolutionary directions toward AI-native 6G networks, integrated sensing-communication-computing architectures, and intelligent reconfigurable wireless environments are discussed. Furthermore, emerging generative AI techniques, including diffusion models, are identified as a promising direction for addressing data scarcity and enhancing system adaptability. The study demonstrates that hybrid intelligence-integrating model-based prior knowledge with data-driven learning-will become the dominant design philosophy for next-generation intelligent physical-layer systems.

Glioblastoma multiforme (GBM) is the most aggressive primary brain tumor in adults, with a median survival of 14.6 months under standard radiotherapy and temozolomide (TMZ) chemotherapy. The methylation status of the O&#x2076;-methylguanine-DNA methyltransferase (MGMT) promoter is a critical biomarker predicting TMZ response; however, its determination currently requires invasive tissue sampling. Non-invasive prediction of MGMT promoter methylation from multiparametric MRI (mpMRI) through deep learning represents a compelling alternative, yet its clinical feasibility remains unresolved. Using the BraTS 2021 dataset (582 patients, four MRI sequences: FLAIR, T1w, T1wCE, T2w), we conducted a systematic comparative study of unimodal and multimodal deep learning approaches based on VGG-16, exploring 1,380 experimental configurations (unimodal: 192; multimodal: 1,188) across three imaging planes, eight slice counts, and three multimodal fusion strategies (early, intermediate, and late fusion). In the unimodal setting, the best model trained on T2w coronal images (32 slices, no transfer learning) achieved an accuracy of 0.6458 and an AUC of 0.6422 on the validation set, but dropped to 0.5586 and 0.5533 on the independent test set, revealing substantial overfitting attributable to limited dataset size. Strikingly, multimodal fusion consistently failed to outperform the best unimodal model, with all three fusion strategies plateauing at ~0.64 accuracy and ~0.64 AUC on validation data. Transfer learning improved generalization across train/test distributions at the cost of peak performance. These findings suggest, for the tested framework in this study, that MGMT methylation status prediction from mpMRI remains fundamentally constrained by dataset heterogeneity and size, irrespective of modality combination strategy, and that T2w coronal acquisitions could be more interesting in future data collection efforts.

EndoPredict is a gold-standard for HR+/HER2- breast cancer risk stratification but limited by high cost. We propose Multi-modal Clustering-constrained Attention Multiple Instance Learning (MCLAM), a novel deep learning framework, offering a cost-effective, accurate alternative for recurrence risk prediction. We retrospectively analyzed 254 early-stage HR+/HER2- breast cancer patients from multiple centers, supplemented by an external validation cohort of 338 cases. MCLAM innovatively combines histopathological features extracted via ResNet-50, nuclear morphology quantified by HoVer-Net, and clinical variables through a clinical-nuclear feature enhancement module and a multimodal fusion strategy. MCLAM outperformed all comparison models, achieving an impressive area under the curve (AUC) of 0.83 and 0.82 (independent test set), significantly exceeding traditional deep learning baselines (ResNet50 AUC = 0.61) and MIL variants (DTFD_MIL_CN AUC = 0.78). Clinically, it stratified patients into risk groups with significantly better 5- and 10-year disease-free survival and distant disease-free survival in the low-risk group (all P < 0.05), and maintained predictive power in HER2-low populations (all P < 0.05). Attention heatmaps and nuclear feature analysis further enabled biological interpretability of recurrence risk. To our knowledge, this is the first and largest multi-center study validating an EndoPredict-predicting model in a Chinese HR+/HER2- breast cancer cohort, addressing a key population gap. MCLAM provides an accurate, interpretable, and cost-efficient alternative to molecular assays, enabling accessible individualized risk assessment and supporting personalized treatment strategies.

Early detection of Micro-cracks in monolithic zirconia crowns remains a challenge because conventional inspection methods cannot identify subsurface defects that may lead to clinical failure. Therefore, it is of interest to develop and evaluate a deep learning convolutional neural network model for detecting and classifying Micro-cracks in zirconia crowns using high-resolution micro-computed tomography imaging. Hence, sixty zirconia crowns were fabricated and divided into control and experimentally stressed groups, generating 1,440 labeled micro-CT cross-sectional images that were used to train (80%) and test (20%) a ResNet-50 model. The model achieved an overall accuracy of 94.7%, sensitivity of 93.2%, specificity of 96.1% and an AUC of 0.97 for micro-crack detection and classification. Deep learning combined with micro-CT imaging provides a highly accurate and automated approach for identifying Micro-cracks in zirconia crowns, with potential to enhance quality assurance in dental manufacturing workflows.

This work investigated whether deep learning models trained on optic disc-centered fundus images retain sufficient subject-specific information for biometric verification compared with models trained on full-field fundus photographs. A total of 30,836 color fundus photographs from 7,724 eyes of 4,500 subjects were obtained at the Bascom Palmer Eye Institute. Each fundus photograph was processed into three image representations: full-field fundus, optic disc region including 0.5 disc diameters of peripapillary retina, and tightly cropped optic disc only. Images were partitioned at the subject level into training (70%), validation (10%), and test (20%) sets. Separate Siamese convolutional neural network models were trained for each image type using triplet loss to learn subject-discriminative embeddings. Biometric verification was evaluated on the independent test set using exhaustive same-eye image pairing and cosine similarity. All image representations retained measurable subject-specific biometric signal. The full-fundus model achieved the highest performance (AUC, 0.992; EER, 4.4%), followed by the disc-region model (AUC, 0.989; EER, 5.5%) and the disc-only model (AUC, 0.969; EER, 10.5%). Accuracy was 0.968 for full fundus, 0.945 for disc region, and 0.919 for disc-only images. Pairwise comparisons showed significantly worse performance for disc-only images compared with full fundus (P&#x2009;<&#x2009;0.001). Differences between full-fundus and disc-region models were small and not significant for AUC or EER. These findings demonstrate that deep learning models restricted to optic disc-centered fundus images retain meaningful subject-specific information, although performance declines as available retinal context is reduced. Inclusion of a narrow peripapillary rim yields biometric verification performance comparable to full-field fundus images. Although identity cannot be established from a fundus image without a linking key, recognizing that even restricted retinal regions retain subject-specific features highlights the importance of cautious data-sharing practices while supporting continued scientific collaboration.

Caenorhabditis elegans is a premier model organism for aging and neurobiology research, valued for its short lifespan, optical transparency, genetic tractability, and well-mapped nervous system. Non-invasive automated recording of biomarkers is a fundamental goal in modern biology because it preserves natural physiology and eliminates confounds from anesthesia, restraint, or repeated handling in C. elegans . Yet high-magnification imaging of freely moving worms remains a persistent challenge: as magnification increases, the narrowing field of view compounds target loss, motion blur, and focal drift, pushing researchers toward immobilization strategies that compromise physiology, suppress natural behavior, and preclude the continuous longitudinal observation essential for aging and neurobiological studies. Here, we present a real-time tracking workflow for imaging individual worms in a microfluidic platform under controlled culture conditions. The system integrates deep learning head detection, image-based autofocus, and rapid motorized-stage feedback to support stable imaging across multiple magnifications, including neuronal-scale imaging. Hundreds of individually housed worms in separate incubation chambers enable repeated daily imaging of the same animals throughout their lifespan. Built entirely on a commercially available inverted microscope without additional custom hardware, the platform features a modular, user-configurable interface adaptable to diverse microscope setups, specimens, and experimental goals. Fluorescence images from freely moving worms were visually comparable to those from immobilized animals, supporting longitudinal phenotyping in aging and neurobiology studies.

This study aims to develop a deep learning model utilizing both surface and esophageal electrocardiogram (ECG) data to accurately differentiate types of paroxysmal supraventricular tachycardia (PSVT), including slow-fast atrioventricular nodal re-entrant tachycardia (S-F AVNRT), and orthodromic atrioventricular reentrant tachycardia with left-sided (AVRT-L) and right-sided (AVRT-R) accessory pathways. We analyzed 921 ECG cases from 775 patients from four hospitals between 2014 and 2022, segmented into 6261 ten-second ECG segments. A Residual Network (ResNet)-based model was developed. For comparison, Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) classifiers were also constructed using handcrafted time-domain and frequency-domain features. It was thoroughly evaluated using a comprehensive set of metrics. These metrics included accuracy (ACC), the area under the receiver operating characteristic curve (AUC); sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and F1-Score. The diagnostic efficacy of ECG lead configurations was robust. Among surface-only leads, the three-lead combination II+V1+aVF achieved the highest AUC (0.989). The single-lead aVF demonstrated remarkable efficiency (AUC 0.961), approaching the performance of the full 12-lead ECG (0.974). The bipolar esophageal lead (EB) alone achieved an AUC of 0.989, comparable to II+V1+aVF (AUC 0.989). The combination aVF+EB yielded the highest overall AUC of 0.996. ResNet significantly outperformed RF and XGBoost across all lead configurations (P < 0.001). This model effectively distinguishes between PSVT types, surpassing traditional diagnostics in accuracy and reliability. Future research should focus on model optimization and dataset expansion to enhance diagnostic capabilities and interpretability.

Cryogenic electron microscopy (cryo-EM) has become an increasingly important for structure-based drug discovery by enabling characterization of interactions between macromolecules and small-molecule ligands. However, computational interpretation of ligand density remains challenging, particularly when ligand locations are unknown or local map resolution is limited. Existing methods generally require well-resolved macromolecular structures and predefined binding sites, limiting their applicability during early-stage structure determination. To date, no approach has been able to both reliably detect ligand density and subsequently reconstruct ligand atomic structures directly from experimental cryo-EM maps. Here, we present Emap2lig, a two-stage deep learning framework for automated ligand detection and atomic modeling directly from cryo-EM maps. Emap2lig-Find identifies ligand-associated densities and remains effective for maps at resolutions as low as &#x223c;5 &#xc5;. Emap2lig-Build subsequently uses a diffusion-based generative model to build atomic ligand structures. Together, Emap2lig provides a unified framework for ligand discovery and modeling across a broad range of resolutions.

Wireless capsule endoscopy (WCE) enables painless, minimally invasive visualization of the gastrointestinal tract. Still, its diagnostic potential is limited by incomplete mucosal coverage and poor transferability of existing navigation methods across patient anatomies. We propose a transferable, anatomical landmark-guided deep reinforcement learning framework for robust autonomous gastric navigation. Leveraging a lightweight edge-contour-depth fusion module, our policy operates on stable, low-dimensional landmark coordinates rather than high-dimensional video streams. This design effectively bridges the sim-to-real visual gap and ensures robustness across diverse anatomies, enabling low-cost deployment by reducing computational overhead. In simulations across eight patient-derived models, the method achieves >97% coverage within 50 s, significantly outperforming vanilla Proximal Policy Optimization, Soft Actor-Critic, and Deep Q-Network agents by enhancing coverage and minimizing variance. To ensure deployment reliability, a two-stage sim-to-real pipeline supported by an adaptive dynamic programming controller actively mitigates physical disturbances, including actuator latency and peristalsis. Ex vivo experiments across five independent scans demonstrate high coverage stability, achieving a mean coverage of 87% and a 53% reduction in procedure time compared with expert manual control. This study establishes a scalable paradigm for autonomous, high&#x2011;coverage endoscopic navigation, advancing the clinical deployment of intelligent WCE systems for GI diagnostics.

Building detection from drone imagery represents a transformative approach to rural governance by enabling precise spatial data acquisition for critical applications including illegal construction monitoring, disaster assessment, and cadastral mapping. However, automated detection systems face persistent challenges including extreme scale variations in rural buildings, complex background interference from vegetation and shadows leading to boundary ambiguity, and severe scarcity of high-quality annotated datasets that limit model generalization. To overcome these limitations, this study introduces an integrated framework featuring three innovative components: the Multi-scale Hybrid Attention module employs parallel convolutional pathways with channel and spatial attention to dynamically capture multi-scale features while suppressing background noise; the Dynamic Feature Pyramid Network utilizes content-aware routing to adaptively fuse hierarchical features for optimal scale-invariant representation; and the Progressive Contrastive Learning strategy leverages both labeled and unlabeled data through hard sample mining to enhance discriminability under data constraints. Extensive experiments validate the model's efficacy, achieving a mean Intersection over Union (MIoU) of 87.3%, pixel accuracy (PA) of 94.2%, and mean Average Precision (mAP) of 89.6% on the Massachusetts Buildings Dataset, substantially surpassing benchmarks like U-Net (80.1% MIoU), SegNet (78.9% MIoU), and DeepLabV3+ (82.4% MIoU), with ablation studies confirming critical module contributions (e.g., MIoU drops to 81.5% without MHA). The framework demonstrates robust cross-dataset generalization (72.3% MIoU on Chinese rural data) and effective problem resolution, establishing a scalable solution for intelligent rural governance through accurate building extraction. The dataset and code used in this study have been uploaded to the GitHub website: https://github.com/xiexie1234567890/rural_building_detection/tree/main.

Deep learning has advanced electrocardiogram (ECG) analysis but remains difficult to interpret, limiting clinical adoption and electrophysiological insight. We propose a post-hoc explainability method for convolutional neural networks (CNNs) applied to ECG. The method clusters the feature activations of the last three residual blocks of a 1D-ResNet, segmenting each ECG into a sequence of clusters and quantifying assignment entropy as a per-timestamp uncertainty. It is evaluated on PTB-XL super-class classification under 10-fold cross-validation and on CODE-15% age regression. Cluster proportions correlate with predicted labels and align with P/QRS/T/TP landmarks. A random forest trained on cluster proportions reproduces the CNN's predictions at $94.9 \pm 0.5\%$ and matches its accuracy and AUROC on the true labels (88.2 vs. $88.4\%$; 79.6 vs. $81.4\%$). Encoder uncertainty surfaces class-dependent representational stability not visible to Grad-CAM. Time-localized clusters recover physiologically aligned structure from CNN activations and faithfully summarize the encoder's discriminative information. The method provides a post-hoc, architecture-light tool to audit CNN-based ECG models and surface label-quality issues.

The evolution of soft robots into embodied intelligent systems relies fundamentally on precise proprioception. However, a universal solution for capturing continuous deformations during diverse interactions, particularly in spatially confined interventional scenarios, remains lacking. Here, we introduce a deep learning-enabled versatile shape perception method based on a single-ended multimode fiber (MMF). By leveraging the intrinsic integration advantages of optics, our minimalist reflective architecture physically eliminates the dependence on complex demodulation units and distal devices. Furthermore, treating chaotic optical speckle fields as data streams encoding high-dimensional shape information, reconfigurable neural decoders resolve a single physical channel into versatile perception modes tailored to heterogeneous tasks: discrete state confirmation on soft grippers (>99% accuracy), continuous shape tracking on bionic dexterous hands (~5-fold spatial resolution enhancement), and intuitive 3D morphological reconstruction of soft surgical robots (IoU>0.93). Overall, our work establishes a versatile framework for breaking hardware adaptability limits via computation, laying a solid foundation for closed-loop control in digital twins of soft robots.

The proliferation of distributed energy resources (DERs) and the ubiquity of Internet of Things (IoT) devices are driving the integration of mobile edge computing (MEC) into smart grids. This convergence enables real-time data processing for prosumers but introduces a complex cyber-physical coupling: computational offloading decisions directly impact local energy consumption, thereby altering the prosumer's status in the peer-to-peer (P2P) energy market. Conversely, dynamic market prices influence the economic viability of offloading. This paper addresses the joint optimization of computational task offloading and P2P energy trading in an edge-assisted smart grid ecosystem. We formulate the problem as a mixed-integer nonlinear programming (MINLP) model aimed at maximizing long-term system utility, balancing throughput, latency, and economic incentives under strict edge server capacity and community energy neutrality constraints. To tackle the curse of dimensionality and system stochasticity, we propose a hybrid framework combining Deep Q-Networks (DQN) with a constraint-aware heuristic mechanism. The DQN agent learns adaptive offloading policies from high-dimensional states, while a deterministic rule-based layer ensures strict adherence to community energy balance. Simulation results based on real-world solar generation and market data demonstrate that our proposed method outperforms baseline strategies-including local-only execution and greedy heuristics-improving average utility by 12.3% and reducing task delay by 16.5%, while maintaining robust operational feasibility.

Abdominal aortic aneurysms (AAA) pose a significant clinical risk due to their potential for rupture, which is often asymptomatic but can be fatal. Although maximum diameter is commonly used for risk assessment, diameter alone is insufficient as it does not capture the properties of the underlying material of the vessel wall, which play a critical role in determining the risk of rupture. To overcome this limitation, we propose a deep learning-based framework for elasticity imaging of AAAs with 2D ultrasound. Leveraging finite element simulations, we generate a diverse dataset of displacement fields with their corresponding modulus distributions. We train a model with U-Net architecture and normalized mean squared error (NMSE) to infer the spatial modulus distribution from the axial and lateral components of the displacement fields. This model is evaluated across three experimental domains: digital phantom data from 3D COMSOL simulations, physical phantom experiments using biomechanically distinct vessel models, and clinical ultrasound exams from AAA patients. Our simulated results demonstrate that the proposed deep learning model is able to reconstruct modulus distributions, achieving an NMSE score of 1.6%. Similarly, in phantom data, the predicted modular ratio closely matches the expected values, affirming the model's ability to generalize to phantom data. We compare our approach with an iterative method which shows comparable performance but higher computation time per reconstruction. In contrast, once trained offline, the deep learning method can provide quick and accurate estimates of tissue stiffness from ultrasound images, so as to provide near real-time predictions of AAA growth.

Molecular features play critical roles in shaping cellular responses to therapeutic agents, and understanding their influence on drug sensitivity and resistance is essential for explaining heterogeneous treatment outcomes. Integrating multi-omics molecular information can uncover complex cross-modal dependencies, identify potential biomarkers, and enhance drug response prediction. However, the high dimensionality and strong interdependencies of multi-omics data pose substantial modeling challenges, underscoring the need for robust, interpretable computational approaches. This study presents DiffDR, a diffusion-based framework that models multi-omics features and drug representations through an energy-constrained diffusion module. This module encodes batched samples and efficiently propagates information while preventing over-smoothing, enabling the capture of both global and local dependencies without relying on explicit graph structures. To enhance model transparency, DiffDR incorporates an integrated gradient-based interpretability module that quantitatively attributes prediction outcomes to specific omics features. DiffDR demonstrates superior predictive performance compared with several state-of-theart drug response prediction methods. Ablation analysis indicates that the energy-constrained diffusion mechanism substantially improves predictive accuracy, confirming its effectiveness in handling high-dimensional multi-omics data. The findings highlight the value of DiffDR in capturing cross-modal molecular dependencies and providing interpretable insights into drug response mechanisms. Overall, DiffDR represents a robust and interpretable approach for drug response prediction, enabling biologically meaningful mechanistic insights into molecular drivers of drug response.

Currently, multiple unmanned aerial vehicles (UAVs) can cooperatively work as mobile edge computing (MEC) servers in the sky to provide computation services to ground terminals (GTs). Such an MEC-enabled multi-UAV system will greatly benefit the GTs, each of which can offload its tasks on demand to a nearby UAV. In particular, if a GT has to process computation-intensive deep learning tasks in a catastrophic environment, it can partially offload these tasks to UAVs using a scheme like Partial Program Offloading (PPO). This ensures the quick processing of the deep learning tasks while saving computing resources on both the GT and UAV sides. Nevertheless, UAV-GT offloading links are frequently blocked by ground obstacles in complicated environments, and individual UAVs may have limited computation capacity. Moreover, UAVs lack a constant propulsion energy supply to sustain a long mission time. All these factors lead to a degraded Quality of Service (QoS) for GTs in terms of task latency. To address this issue, we propose to jointly optimize the UAV trajectories, computing resource allocation, and the partial offloading of deep learning tasks. The formulated joint optimization problem is challenging to solve optimally, as it is non-convex and involves multiple coupled constraints. We propose utilizing the Successive Convex Approximation (SCA) method alongside a Block Coordinate Descent (BCD) approach to tackle this joint problem. Numerical results demonstrate that the proposed joint optimization scheme significantly outperforms the benchmark solutions.

暂无摘要（点击查看详情）