With the advancement of artificial intelligence, molecular design based on generative models offers novel approaches to accelerate drug discovery. However, existing molecular generation methods suffer from inadequate representational capability in geometric structure and discrepancies between topological and geometric structure representations. These challenges result in generating chemically implausible and structurally unstable molecules. Furthermore, existing methods neglect the crucial properties of both quantum and drug-likeness in drug design. To address these challenges, we propose a novel Geometric Structure-Aware Diffusion Model for molecular generation and optimization tasks, named MolGD. First, we designed a Geometric Structure-Aware Network (GSAN) to directly predict structurally stable molecules from noisy inputs. Within GSAN, a Molecular Graph Attention Network (MGAT) is designed to incorporate geometric information during the topological message-passing process. Then, atomic spatial positions are updated by a Geometric Reconstruction Network (GRN) for enabling integrated modeling of molecular structures. Second, MolGD integrates quantum attributes as conditional constraints for precise quantum property regulation. These conditional constraints can guide MolGD to generate molecules with specific quantum properties. Finally, for drug-likeness property optimization, MolGD integrates self-optimization strategies (MolGD-RL) to guide the model toward generating high drug-likeness and easily synthesisable molecules. Experimental results on the quantum chemistry data set QM9 and the molecular conformation data set GEOM-Drugs demonstrate that the MolGD model outperforms existing molecular generation methods in terms of the effectiveness and stability of generated molecules, the generation of specific quantum properties, and high drug-likeness optimization. This validates its efficacy in molecular generation and optimization tasks, offering novel insights for intelligent molecular design.

Efficiently predicting drug synergy is crucial for developing personalized cancer combination therapy regimens. However, existing methods primarily focus on single-scale structural information and fail to explicitly model the interactions between multi-scale structural information from cell lines and drug pairs, limiting their ability to capture the complex molecular mechanisms of synergistic effects. To tackle these challenges, we propose DeepSTFSynergy, a multi-scale structural information fusion framework for personalized drug combination prediction. DeepSTFSynergy introduces three parallel attention-based subnetworks that comprehensively extract the interaction features of drugs at atomic, sub-structural and global structural scales to adaptively capture multi-scale molecular interactions for predicting synergy. Meanwhile, we design a novel cell line-specific cross-modal fusion mechanism that employs gating units to dynamically identify the contributions of three-scale molecular interaction information to synergistic effects in specific cell lines, thereby filtering out non-critical information and efficiently fusing the features of drugs and cell lines. Comprehensive experiments across real-world benchmark datasets reveal that DeepSTFSynergy exhibits superior performance over current leading approaches in both regression and classification tasks. Case studies also illustrate that the novel drug combinations predicted by DeepSTFSynergy align with previous studies. Moreover, visualization analysis reveals the model’s capability to identify atomic structures and substructures associated with synergy. By quantifying their relative contributions in combination therapy, it provides an interpretable perspective for understanding synergistic mechanisms and assisting personalized treatment decision-making.

Historical landslide inventory serves as a critical tool for analyzing landslide activity patterns and evaluating the long-term geological impacts of triggering events, including earthquakes, extreme weather events, and large-scale infrastructure projects. Although various methods—including visual interpretation, heuristic approaches, machine learning, and deep learning models—have been employed for landslide detection, efficient techniques for historical landslide mapping remain understudied. As a result, comprehensive historical landslide inventories continue to be scarce worldwide. In this study, we developed an advanced landslide detection model using a Swin Transformer architecture integrated with a Pyramid Segmentation Attention mechanism. Subsequently, we applied a network fine-tuning method to achieve cross-domain adaptation, enabling the reconstruction of a decadal-scale landslide inventory across the Wenchuan earthquake-affected region efficiently. Experimental results from the Wenchuan earthquake area demonstrate the proposed approach’s superior temporal transfer mapping performance compared to state-of-the-art models. The proposed historical map also exhibits high accuracy and completeness, offering significant value for analyzing landslide spatiotemporal activity and long-term regional stability. Findings reveal that landslides stabilized overall between 2008 and 2021, with key influences including altitude, slope, and aspect. The results lay the groundwork for regional stability analysis and eco-environment recovery, enabling informed decisions in urban planning and infrastructure investments.

Abstract Electronic health records (EHRs) integrate structured time-series data and clinical records, providing a comprehensive foundation for clinical risk prediction tasks such as in-hospital mortality (IHM) and phenotyping (PHE). However, existing studies still face several critical limitations, including insufficient modeling of irregularly sampled time-series data, lack of explicit cross-modal semantic alignment, and degraded performance under severe class imbalance, particularly for minority classes in both binary and multi-label settings. To address these challenges, we propose a task-aware multimodal risk prediction framework, called UniRiskP, which jointly models temporal irregularity, cross-modal alignment, and imbalance-aware learning. Specifically, UniRiskP incorporates a Unified Time-Driven Embedding (UTDE) module to explicitly model irregular sampling patterns and missing values in structured time-series data by leveraging inter-event time intervals. In parallel, semantic representations are extracted from radiology reports using ClinicalBERT. A maximum mean discrepancy (MMD)-based alignment strategy is employed to project heterogeneous modalities into a shared representation space, and a Dual-level Alignment-based Fusion (DeAF) module enables deep cross-modal interaction. To handle task-specific imbalance, UniRiskP adopts an optimal transport-based reweighting strategy for IHM and a label-aware threshold optimization mechanism for PHE. Experiments on the MIMIC-IV dataset demonstrate that UniRiskP consistently outperforms existing methods in terms of AUROC, AUPRC, F1, and Recall. Ablation studies further verify the effectiveness of irregular temporal modeling, cross-modal alignment, and imbalance-aware strategies. These results highlight the importance of jointly addressing temporal irregularity and semantic alignment in multimodal clinical risk prediction. These findings suggest that UniRiskP has the potential to support reliable ICU risk assessment from heterogeneous EHR data and will facilitate clinical decision-making in real-world settings.

Incremental few-shot relation classification aims to train a model on minimal training instances to dynamically learn novel relations while retaining prior knowledge of base relations. Current approaches still suffer from two persistent challenges: novel-class overfitting exacerbated by prototype drift, and severe catastrophic forgetting of base-class knowledge caused by inadequate discriminability among similar classes. In this paper, we design a Distance-Weighted merged prototypical network with Calibration (DWproto-C) to address these issues. DWproto-C architecturally integrates three modules: (1) a Distance-Weighted (DW) merged prototype network that provides a simpler and more effective solution to the incompatible feature embedding problem in prototypical networks through exclusive use of scalar distance measures; (2) a Base-Attentive Novel Prototype Calibration (BANPC) module designed to mitigate prototype drift in novel class embeddings induced by distributional bias in support samples, thereby obtaining more accurate novel prototypes; (3) a Hardest Negative Mining (HNM) module designed to mitigate catastrophic forgetting caused by similar-class confusion, which explicitly enforces the distinction between the ground truth class and its most confusable counterpart for each query sample. Our model demonstrates superior performance over current state-of-the-art (SOTA) methods across both the FewRel 1.0 and 2.0 benchmarks. Particularly noteworthy is the significant improvement in novel relation classification accuracy on FewRel 2.0, achieving a 11.75

Lossless compression has made significant advancements in Genomics Data (GD) storage, sharing and management. Current learning-based methods are non-evolvable with problems of low-level compression modeling, limited adaptability, and user-unfriendly interface. To this end, we propose AgentGC, the first evolutionary Agent-based GD Compressor, consisting of 3 layers with multi-agent named Leader and Worker. Specifically, the 1) User layer provides a user-friendly interface via Leader combined with LLM; 2) Cognitive layer, driven by the Leader, integrates LLM to consider joint optimization of algorithm-dataset-system, addressing the issues of low-level modeling and limited adaptability; and 3) Compression layer, headed by Worker, performs compression decompression via a automated multi-knowledge learning-based compression framework. On top of AgentGC, we design 3 modes to support diverse scenarios: CP for compression-ratio priority, TP for throughput priority, and BM for balanced mode. Compared with 14 baselines on 9 datasets, the average compression ratios gains are 16.66

MicroRNAs (miRNAs) are key regulators of anticancer drug response, including sensitivity and resistance. Accurate prediction of miRNA-drug response associations is important for understanding response mechanisms and guiding targeted therapy. Existing prediction methods are difficult to fully capture complex high-order biological interactions and are sensitive to data sparsity and class imbalance. In this paper, we present an miRNA-drug response association prediction method called mDADGAN by constructing a diffusion-based generative adversarial network. First, we curated a large-scale and biologically informative data set from four public data resources by aggregating functional associations to facilitate model training. Second, we built a miRNA-lncRNA-drug interaction network by integrating lncRNA into the miRNA-drug network to address data sparsity and employed heterogeneous graph convolution to extract latent features from the interactions among miRNAs, lncRNAs, and drugs. Third, we applied diffusion-based perturbation to denoise and enhance the robustness of feature representations. Finally, we performed cost-sensitive learning to optimize the objective function in an adversarial training process to address class imbalance. The experimental results on four data sets demonstrated that, compared with the existing six methods, mDADGAN achieved higher values for AUC, AUPR, ACC, Recall, and F1-score. The case study demonstrates that integrating lncRNA information effectively enhances miRNA-drug response association prediction. mDADGAN can not only accurately identify known associations in the data sets but also predict previously unreported candidate associations supported by published evidence. Overall, our method mDADGAN has the ability to identify potential miRNA-drug response associations and can provide guidance for future biological experiments to enhance experimental efficiency and reduce experimental costs.

As various types of data grow explosively, large-scale data storage, backup, and transmission become challenging, which motivates many researchers to propose efficient universal compression algorithms for multi-source data. In recent years, due to the emergence of hardware acceleration devices such as GPUs, TPUs, DPUs, and FPGAs, the performance bottleneck of neural networks (NN) has been overcome, making NN-based compression algorithms increasingly practical and popular. However, the research survey for the NN-based universal lossless compressors has not been conducted yet, and there is also a lack of unified evaluation metrics. To address the above problems, in this paper, we present a holistic survey as well as benchmark evaluations. Specifically, i) we thoroughly investigate NN-based lossless universal compression algorithms toward multi-source data and classify them into 3 types: static pre-training, adaptive, and semi-adaptive. ii) We unify 19 evaluation metrics to comprehensively assess the compression effect, resource consumption, and model performance of compressors. iii) We conduct experiments more than 4600 CPU/GPU hours to evaluate 17 state-of-the-art compressors on 28 real-world datasets across data types of text, images, videos, audio, etc. iv) We also summarize the strengths and drawbacks of NN-based lossless data compressors and discuss promising research directions. We summarize the results as the NN-based Lossless Compressors Benchmark (NNLCB, See fahaihi.github.io/NNLCB website), which will be updated and maintained continuously in the future.

Acne vulgaris, a common dermatological condition affecting the pilosebaceous follicular unit, arises from a complex interplay of factors including hyperseborrhea, Propionibacterium acnes colonization, hyperkeratosis, and inflammation. Existing pharmacodynamic evaluation techniques do not facilitate real-time and non-invasive assessment of acne and its severity within living organism. For this study, the aim was to establish a murine model of acne vulgaris, elucidate its pathological processes, and quantitatively assess the severity of the condition using non-invasive imaging techniques. We employed optical coherence tomography (OCT) in conjunction with microscopy to scrutinize the complete life cycle of acne lesions. microscopy provided insights into the micro-anatomical characteristics of murine skin lesions, while OCT allowed for the assessment of skin structural changes. Our findings indicate that the epidermis undergoes hyperplasia concurrent with acne development, and the subsequent reparative phase is marked by scab formation. Integrating deep learning algorithms enabled the precise quantification of epidermal and scab thickness variations. This novel approach provides a unique perspective for acne assessment, potentially guiding more precise clinical drug selection and improving therapeutic outcomes.

To address the issue of insufficient resolution in remote sensing images due to limitations in sensors and transmission, this paper proposes a multi-scale feature fusion model, MSFANet, based on the Swin Transformer architecture for remote sensing image super-resolution reconstruction. The model comprises three main modules: shallow feature extraction, deep feature extraction, and high-quality image reconstruction. The deep feature extraction module innovatively introduces three core components: Feature Refinement Augmentation (FRA), Local Structure Optimization (LSO), and Residual Fusion Network (RFN), which effectively extract and adaptively aggregate multi-scale information from local to global levels. Experiments conducted on three public remote sensing datasets (RSSCN7, AID, and WHU-RS19) demonstrate that MSFANet outperforms state-of-the-art models (including HSENet and TransENet) across five evaluation metrics in ×2, ×3, and ×4 super-resolution tasks. Furthermore, MSFANet achieves superior reconstruction quality with reduced computational overhead, striking an optimal balance between efficiency and performance. This positions MSFANet as an effective solution for remote sensing image super-resolution applications.

In Chinese speech recognition, end-to-end speech recognition models usually use Chinese characters as direct output and perform poorly compared with other language models. The main reason for this phenomenon is that the relationship between Chinese text and pronunciation is more complex. Inspired by the learning process of Chinese beginners, who first master initials, finals, and pinyin before learning characters, we propose the Syllable-Character Collaborative Model (SCCM), which incorporates these phonetic elements into the training process. Additionally, we design a Pinyin-Ensemble module that employs an ensemble learning approach to reduce pinyin recognition errors, which in turn leads to a reduction in text recognition errors. Experiments on AISHELL-1 show that our approach not only reduces pinyin and character error rates compared to a prior end-to-end method using pinyin as auxiliary information, but also achieves a 45.7% relative reduction in Character Error Rate (CER) over the AISHELL-1 baseline.

Few-shot object detection (FSOD) based on fine-tuning is essential for analyzing optical remote sensing images. However, existing methods mainly focus on natural images and overlook the scale variations in remote sensing images, leading to feature confusion among foreground instances of different classes. Additionally, since only a subset of instances are labeled in FSOD training data, the model might mistakenly treat unlabeled instances as background, leading to confusion between foreground features and background features, particularly those of novel classes. The preceding phenomenon indicates that severe feature confusion in remote sensing FSOD hampers the ability of the model to accurately classify and localize instances. To address these issues, this paper proposes a two-stage FSOD framework based on transfer learning via pseudo-sample generation and feature enhancement (PSGFE), including pseudo-sample generation module (PSGM) and feature enhancement module (FEM). The former reduces the feature confusion between foreground and background by generating pseudo-samples for unannotated background areas. The latter dynamically captures and enhances multi-scale features on the region of interest (ROI), and extracts unique core information for each class to eliminate the feature confusion among foreground instances of different classes. Our method has been validated on the optical remote sensing datasets DIOR and RSOD. It demonstrates superior performance compared to existing methods.

Federated learning is effective for Internet of Things data privacy and non-independent and identically distributed issues but not device heterogeneity. Neural Architecture Search can alleviate this by constructing multiple model structures to optimize federated learning performance across diverse edge devices. However, existing methods, whether lightweight networks or client grouping, face a tradeoff between scaling to larger federations and utilizing more powerful structures. We decompose residual network blocks, reformulating them as a Neural Architecture Search task. Furthermore, we propose a method for reinterpreting any sequential architecture into a supernet and developed a training pipeline tailored to this reinterpretated architecture, mitigating this frustrating tradeoff. We conduct pretraining on ImageNet1K and federated training on the CIFAR-100, CIFAR-10, and CINIC-10 datasets under both the ring-based federated learning and FedAvg framework. In less constrained environments, our method maintains performance comparable to another top-two method, which varies across experimental settings, while maintaining a margin of at least 1% Top-1 accuracy over the third-best method. Under balanced settings, our method outperforms the second-best approach by more than 1%, and this advantage increases to over 5% as the task difficulty further rises. Under the most challenging setting, our method outperformed AdaptiveFL, a state-of-the-art dynamic network method for federated learning, by 18.3% on CIFAR-100 with 100 clients under a ResNet backbone.

Learning-based lossless compressors play a crucial role in large-scale genomic database backup, storage, transmission, and management. However, their 1) inadequate compression ratio, 2) low compression & decompression throughput, and 3) poor compression robustness limit their widespread adoption and application in both industry and academia. To solve those challenges, we propose a novel Parallel Multi-Knowledge Learning-based Compressor (PMKLC) with four crucial designs: 1) We propose an automated multi-knowledge learning-based compression framework as compressors' backbone to enhance compression ratio and robustness; 2) we design a GPU-accelerated (s,k)-mer encoder to optimize compression throughput and computing resource usage; 3) we introduce data block partitioning and Step-wise Model Passing (SMP) mechanisms for parallel acceleration; 4) We design two compression modes PMKLC-S and PMKLC-M to meet the complex application scenarios, where the former runs on a resource-constrained single GPU and the latter is multi-GPU accelerated. We benchmark PMKLC-S/M and 14 baselines (7 traditional and 7 leaning-based) on 15 real-world datasets with different species and data sizes. Compared to baselines on the testing datasets, PMKLC-S/M achieve the average compression ratio improvement up to 73.609% and 73.480%, the average throughput improvement up to 3.036X and 10.710X, respectively. Besides, PMKLC-S/M also achieve the best robustness and competitive memory cost, indicating its greater stability against datasets with different probability distribution perturbations, and its strong ability to run on memory-constrained devices. Overall, PMKLC is a balanced compression solution that optimizes compression ratio, throughput, robustness, and resource consumption. PMKLC and linkages of datasets are available at https://github.com/dingyanfeng/PMKLC.

BACKGROUND:Polypharmacy's ability to circumvent acquired resistance to single drug makes it a critical strategy for treating complex diseases. However, it inevitably carries risks of drug-drug interactions (DDIs) that may alter pharmacological activities and potentially lead to severe adverse events or mortality. Computational assessment of drug combination has emerged as an effective approach to support clinical decision-making. Current risk identification methods focus on mining historical interaction patterns to uncover underlying mechanisms, yet face challenges from data sparsity. While data augmentation strategy can mitigate such problem, conventional approaches often introduce noise that obscures core pharmacological mechanisms, undermining safety evaluation. RESULTS:This study proposes a Multi-Mechanism Disentangled Drug-drug Interaction assessment framework integrated contrastive learning, MMDDI, which includes two key components: (1) biologically-informed multi-view generation that creates high-quality augmented views, effectively addressing semantic distortion during data augmentation; (2) Mechanism-aware disentanglement that incorporates mutual information constraints to isolate interaction mechanisms from coupling of multi-modal and heterogeneous data, eliminating quantification bias. Contrastive learning integrates labeled and unlabeled data to enhance robustness against sparse observations. CONCLUSIONS:Comprehensive evaluations demonstrate that MMDDI with hit@4 of 0.86 outperforms the compared baselines, with ablation studies validating the critical contributions of multi-view contrastive and mechanism disentanglement. MMDDI continues to demonstrate excellent performance in cold-start scenarios, achieving accuracy of 0.94 and recall of 0.95. Clinically, MMDDI enables interpretable causal analysis of drug interaction pathways through its mechanism-aware representations, providing operability for optimizing therapeutic regimens.

Identifying disease-associated metabolites could provide critical clues for the diagnosis and treatment of diseases. Although computational approaches have been proposed to predict disease-associated metabolites by training models using positive and negative samples, few efforts have paid attention to optimize the reliability of negative samples, which could possibly improve the prediction accuracy of model. In this work, we propose a novel method called SMDPG to leverage optimized negative sampling and sparse graph convolutional network to predict metabolite-disease associations. In SMDPG, we first build a metabolite-disease bipartite graph based on less similar metabolites and diseases and propose a negative sampling method to select reliable negative metabolite-disease samples from the bipartite graph to reduce the effect of noisy samples. Then a homogeneous metabolite-disease pair graph is constructed based on the selected negative samples and known metabolite-disease associations. Further, an edge sparseness operation is designed to simplify the connection of the homogeneous graph. Finally, the simplified homogeneous graph is fed into the graph convolutional network to predict metabolite-disease associations. Experimental results show that SMDPG can more accurately predicts metabolite-disease associations as compared to the existing methods. Moreover, the study case demonstrates that SMDPG is an effective framework for identifying potential metabolite-disease associations.

Existing drone image processing algorithms for small target detection in Unmanned Aerial Vehicle (UAV) aerial images struggle with challenges like missed detection of small objects, information loss from downsampling, loss of low-dimensional features, and information drop of contextual features. In order to alleviate the four problems just mentioned, we propose a self-adaptive small target detection method, NSC-YOLOv8, based on the YOLOv8 model. First, we introduce a small target detection head that enhances the model’s ability to fuse shallow and deep features, effectively handling low-pixel targets. Second, a Non-lossy Downsampling Block (NDB) is introduced into the backbone, which optimizes the detection accuracy of small targets in large scenes through dimensional transformation. In addition, we introduce a Self-Adaptive Embedding Block (SAEB) based on low-dimensional information, which enhances the comprehensive performance of the model by expanding the local sensing field to enhance the focus on important contextual information. Finally, we design a Content-Aware Resampling Block (CARB), which is able to enhance the model’s ability to recognize small targets by resampling low-dimensional features. Experiments on the VisDrone2019-DET dataset show that NSC-YOLOv8s improves target detection accuracy over YOLOv8s, with an 11.7% increase in mAP@0.5. Additionally, removing a large detection head and adjusting the bottom-up layers reduces NSC-YOLOv8s’ parameters by 1.2 M compared to YOLOv8s. Therefore, NSC-YOLOv8 shows better performance in small target detection for UAV imagery.

Plant resistance proteins are evolved during growth and development to cope with complex environmental changes and infection of pathogens. Predicting plant resistance proteins is of great significance for further exploring plant disease resistance mechanism against viruses. In this paper, we propose a method for predicting plant resistance protein by dual-extracting features. The dual-extracted features are composed of the features extracted by modeling self-attention neural network and detecting sequence structure information respectively to obtain 2381-dimensional protein sequence features. We utilize the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to eliminate redundant features from the extracted 2381-dimensional features to form 53 key features. These 53 key features are inputted into the Lightweight Gradient Boosting Machine (LightGBM) model to predict plant resistance proteins. Experimental results of five-fold cross-validation on real datasets demonstrate that our proposed prediction method outperforms existing methods overall in accuracy, sensitivity, specificity, Matthews correlation coefficient, F1 score, and area under the curve (AUC) in the case of slightly imbalanced datasets. This research work will aid in filtrating plant resistance genes and proteins, and promote disease-resistant breeding for plants.

Automatically finding novel pathways plays an important role in the initial designs of metabolic pathways in synthetic biology and metabolic engineering. Although path-finding methods have been successfully applied in identifying valuable synthetic pathways, few efforts have been made in fusing atom group tracking into building stoichiometry model to search metabolic pathways from arbitrary start compound via Mixed Integer Linear Programming (MILP). We propose a novel method called AFP to find metabolic pathways by incorporating atom group tracking into reaction stoichiometry via MILP. AFP tracks the movements of atom groups in the reaction stoichiometry to construct MILP model to search the pathways containing atom groups exchange in the reactions and adapts the MILP model to provide the options of searching pathways from an arbitrary or given compound to the target compound. Combining atom group tracking with reaction stoichiometry to build MILP model for pathfinding may promote the search of well-designed alternative pathways at the stoichiometric modeling level. The experimental comparisons to the known pathways show that our proposed method AFP is more effective to recover the known pathways than other existing methods and is capable of discovering biochemically feasible pathways producing the metabolites of interest.

Spectral clustering with graph learning usually performs eigen-decomposition on the adaptive graph to obtain embedded representation for clustering. In terms of adaptive graph learning, the embedded representation is usually treated as the principal component of the graph to help improve graph structure. However, most adaptive graph learning methods only use a single graph layer. Therefore, the extraction power of embedded representation is restricted to single graph layer and insufficient to explore the intrinsic information. To break through this limitation, this article proposes a stacked network to realize spectral clustering with adaptive graph learning (SCnet-AGL). Specifically, the network allows the development of latent embedded representation underlying the multiple graph layers to reveal the intrinsic information. Meanwhile, we have designed an adaptive graph learning scheme to exploit the latent embedded representation for graph learning. With the advantage of the network, an augmented graph is obtained by incorporating the representation information for graph learning layer by layer. Finally, an efficient algorithm with feedback training scheme is proposed for network training. Experiments on real datasets demonstrate the effectiveness of the proposed network, and show that it is feasible to develop latent embedded representation to improve clustering performance.