Predicting drug-disease associations is a crucial step in drug repositioning, especially with computational methods that quickly locate potential drug-disease pairs. Heterogenous network is a common tool for introducing multiple type relation information about drugs and diseases. However, the diversity of relations is ignored in most of existing methods, which makes them difficult to explore type semantic information with structure properties. Therefore, we propose a relation-centric GNN framework to encode critical association patterns. Firstly, we utilize a relation-centric graph, line graph, to represent the context of a drug-disease pair identified as the center node. The prediction problem is modeled to learn the embedding vector of the center node. Secondly, a multi-grained line graph neural network (MGLGNN) is designed to excavate fine-grained features that encapsulate local graph structures. We theoretically define a handful of typical nodes that can be regarded as high-order abstractions of relations in each type. Then, MGLGNN distills the local information and passes it to typical nodes from a global perspective. With learned multi-grained features, the center node automatically captures heterogenous relation semantics and structure patterns. Thirdly, a hierarchical contrastive learning (HCL) mechanism is proposed to ensure the quality of multi-grained features in an unsupervised way. Extensive experiments show the great potential of our model in mining drug-disease associations.

Antibiotic resistance has reduced the effectiveness of traditional antibiotics for public health needs. Antibacterial peptides (ABPs) hold clinical value due to their inherent antimicrobial activity and resistance to degradation. Existing deep learning approaches for predicting the minimum inhibitory concentration (MIC) of ABPs typically rely on a single modality or a single scale, limiting their ability to capture ABP complexity. Moreover, these methods cover only a narrow range of bacterial species. To address these challenges, we propose the geometric graph network (GGN)-ABPMIC model, which integrates multi-scale structural features at the atomic and residue scales with peptide sequence features. These features are aggregated through a GGN enhanced with geometric vector perceptrons. To enrich the feature representation, we also incorporate descriptors that quantify the physicochemical properties of ABPs and fuse them with the GGN-processed features. The model predicts the MIC values for ABPs across datasets spanning 10 bacterial species. For training and robustness, we introduce a multi-stage dynamic-weight hybrid loss that combines mean-squared error (MSE), the Huber loss, and a contrastive learning loss. Across the 10 species, GGN-ABPMIC achieves a mean MSE of 0.221, a mean R$<^>2$ of 0.564, and a mean Pearson correlation coefficient of 0.749, outperforming existing approaches. Additional validate via case studies on supplementary Escherichia coli ABP sequences indicates that the MIC predictions of GGN-ABPMIC are closer to the true values, demonstrating strong predictive performance and generalization for ABP activity. The data and codes for GGN-ABPMIC are available at https://github.com/LiJinYu1231/GGN-ABPMIC/tree/master.

Background Single-cell RNA sequencing (scRNA-seq) allows for the measurement of gene expression at the transcriptomic level with single-cell precision, thereby deepening our comprehension of cellular heterogeneity. However, the high dimensionality and sparsity of scRNA-seq data impede downstream analyses (such as cell clustering and trajectory inference), and learning effective embedded representations of the data has become a key aspect in scRNA-seq data analysis. Result We present scDEBGCL, a novel deep embedding algorithm based on bipartite graph contrastive learning. scDEBGCL leverages Singular Value Decomposition (SVD) for bipartite graph enhancement and integrates graph contrastive learning, graph reconstruction, and data reconstruction to jointly learn low-dimensional embedded representations of cells, which are used for downstream tasks such as cell clustering, trajectory inference, and marker gene identification. Specifically, scDEBGCL first converts the gene expression matrix into a cell-gene bipartite graph and applies SVD to this bipartite graph for graph enhancement. This strategy effectively preserves the global cell-gene interactions and facilitate the learning of global synergistic signals within the data. To further capture the discriminative cellular representations, scDEBGCL performs contrastive learning between the enhanced graph and the original bipartite graph. Then, scDEBGCL integrates the contrastive learning loss, bipartite graph reconstruction loss, and ZINB distribution-based reconstruction loss to jointly optimize and learn the low-dimensional representations of cells for downstream analyses such as cell clustering, cell trajectory inference, and marker gene identification. Conclusions Experiments results demonstrate that scDEBGCL is a useful GCL framework for deep embedding in scRNA-seq data, providing a reliable foundation for various downstream analyses.

Uncovering genes that drive cancer is fundamental to elucidating the mechanisms underlying cancer development and to advancing cancer research. Recent years have witnessed the emergence of cancer driver gene identification from multi-omics data as a key research area, facilitated by the rapid progress of high-throughput molecular technologies. Although numerous algorithms have been proposed for cancer driver gene discovery, the precise identification of these genes continues to pose a challenge owing to the lack of labeled data. This study presents SDMGAE, a Self-supervised Dual Masked Graph AutoEncoder-based method for cancer driver gene identification. This framework integrates two components: a self-supervised graph learning module and a driver gene prediction module. During the self-supervised graph learning phase, nodes and edges of protein–protein interaction (PPI) networks are masked separately to consider both node and structural information. Subsequently, the graph autoencoder is employed to reconstruct the PPI network without using labelled data. In the driver gene prediction stage, we employ the pre-trained graph neural network encoder to obtain the embeddings, which are then processed through the logistic regression to generate prediction outcomes. To evaluate the effectiveness of SDMGAE, we performed benchmarking experiments across 10 distinct types of cancer data. Experimental outcomes reveal that SDMGAE exhibits improved performance in cancer driver gene detection compared with state-of-the-art methods.

The identification of cancer driver genes is crucial for in elucidating the molecular pathogenesis of carcinogenesis and advancing precision oncology interventions. Although progress has been made in integrating multi-omics data, which has enhanced the predictive ability regarding cancer driver genes, current methods still have their limitations. They merely concentrate on local sample pairs within a single view or employ a fixed number of graph convolutional network layers, which are hard to adapt to the constraints of diverse biological networks. In response to these challenges, this paper introduces a multi-view contrastive learning method (MVCL) to distinguish cancer driver gene. The MVCL first constructs four distinct gene relationship networks from distinct dimensions: a Protein-Protein Interaction network, a Gene Ontology network, a pathway co-occurrence network, and a protein sequence similarity network. To accommodate the varying connection densities across different network views, a topology-adaptive encoder is designed. It dynamically adjusts GCN layer numbers based on the radius of the largest connected subgraph in each view. Feature-level and cluster-level contrastive loss functions are also introduced. They ensure consistent gene feature representation from both local and overall view. Experimental results demonstrate that MVCL significantly enhances the area under the ROC curve and the area under the precision-recall curve for identifying driver genes for pan - cancer and specific cancer types compared to existing methods. In general, MVCL shows great potential in the realm of precision tumor therapy and is applicable to predicting biomarkers of diverse complicated diseases.

Single-cell multi-omics sequencing represents an advanced technology capable of simultaneously measuring multiple omics data from the same cell. The joint clustering of single-cell multi-omics sequencing data enables a comprehensive depiction of cell states and uncovers intricate molecular mechanisms, holding immense significance in fields such as oncology, neurology, and developmental biology. However, the disparities in feature spaces across different omics layers and data noise present substantial challenges for achieving accurate clustering. To tackle these challenges, we introduce a novel clustering method for single-cell multi-omics data, termed scMSAC, which is grounded in a denoising subgraph attention autoencoder. The proposed method employs a weighted nearest neighbor graph strategy to ascertain the weights of multi-omics data, subsequently generating a similarity graph that holistically encapsulates intercellular connections through the weighted amalgamation of diverse omics perspectives. The scMSAC model captures the topological features of cells through the subgraph attention autoencoder, constructing relationships among cells. For the omics features extracted by the subgraph attention autoencoder, scMSAC incorporates an SCA (Spatial Channel Attention) mechanism for feature fusion to reduce the differences in feature spaces of different omics and achieve better clustering performance. Comparative experiments with various existing methods demonstrate that scMSAC has excellent clustering performance and performs well in detecting rare cell types and differential expression analysis.

Accurate prediction of enzyme catalytic constants ( k_cat ) is essential for functional annotation and industrial biocatalysis. However, existing computational methods often face trade-offs between high-dimensional feature complexity and the high cost of 3D structural data. This study presents EnzyKPred, a deep multi-modal learning framework designed for high-throughput k_cat prediction. EnzyKPred integrates evolutionary semantic priors from the pre-trained protein language model ESM-2 with a lightweight hybrid encoder-comprising a bidirectional Gated Recurrent Unit (BiGRU) and a 1D Residual Network (ResNet) to capture both long-range dependencies and local structural motifs. For substrate characterization, a Message-Passing Neural Network (MPNN) is employed to encode molecular topological graphs. A key innovation is the adaptive gated fusion mechanism, which dynamically weights heterogeneous features to optimize enzyme-substrate specificity modeling. Benchmarked on the CatPred-DB dataset, EnzyKPred achieves a peak coefficient of determination (R2) of 0.612 on a strictly partitioned test set, outperforming state-of-the-art methods such as CatPred and UniKP. Furthermore, out-of-distribution (OOD) evaluations demonstrate superior generalization robustness even at 40

Convolutional Neural Networks struggle with long-range dependencies modeling in medical image segmentation, and traditional Transformer models rely on Multi-Layer Perceptron (MLP) for channel information mixing, with performance issues as data dimensions increase. These issues prompt a reassessment of the model’s design to enhance segmentation performance and effectively capture long-range dependencies. Consequently, this study presents the Kadformer, a novel network optimized for fine-grained multi-organ segmentation. The Kadformer model adopts an innovative U-shaped network architecture, which enhances the extraction of spatial and channel features in the encoder through the KAN-Enhanced Multi-Dimensional Attention (KMA) mechanism, effectively compensating for information loss during downsampling. We design a Dynamic Path Selection (DPS) strategy to mitigate the feature extraction discrepancies encountered by the linear attention mechanism when processing category-sparse and category-dense images while enhancing feature discrimination through long-range sequential modeling Mamba. Furthermore, we construct the Data Interaction (DAI) module to guide the dual-path encoder’s channel and spatial information filtering and effectively integrate the semantically inconsistent features between the KMA and DPS modules. Our approach achieves more than 30% parameter reduction compared to state-of-the-art methods. In addition, the Kadformer network outperforms existing segmentation methods on six public datasets, demonstrating excellent performance. The code has been made available on GitHub: https://github.com/wxc9927/Kadformer.

Brain imaging genetics generally combines genotype data with brain structure and functional measures to investigate the genetic basis of neurological disorders. Multimodal brain imaging data carry different but complementary information, which can clearly measure the same brain. Multi-task sparse canonical correlation analysis (MTSCCA) is often used to analyze the bi-multivariate correlation between genotype data and multimodal imaging features; however, it lacks the capacity to identify class-specific biomarkers and to deeply capture complex genetic mechanisms. In this paper, a linear discrimination and decomposition method based on MTSCCA (LDMTSCCA) is proposed to solve above problems. Specifically, LDMTSCCA first extracts disease-related genetic information using sparse linear discriminant analysis, and then employs parameter decomposition to learn multi-level expression patterns of genetic loci. To better explore intricate genetic mechanisms, LDMTSCCA jointly incorporates disease states, parameter decomposition, and network connectivity constraints into the association analysis between multilocus genetic information and multimodal neuroimaging phenotypes. The proposed method is compared with the traditional single task, multi-task, multi-view, and deep learning-based canonical correlation analysis methods in Alzheimer's Disease Neuroimaging Initiative, which shows that LDMTSCCA not only achieves the highest canonical correlation coefficients but also identifies multi-level biomarkers.

Introduction: Single-cell RNA sequencing (scRNA-seq) is crucial for unraveling gene expression complexity. However, existing feature selection methods often overlook the biological significance of co-expressed gene regions, leading to the omission of potential biomarkers. Methods: We propose RF-SCGFS, a co-expressed gene region and gene joint selection method based on random forests. The method identifies co-expressed gene regions within homologous cell populations and builds a random forest model using cell type labels generated by the Scalable and Efficient speCtral clUstERing algorithm (Secuer). Feature importance evaluation is applied to select key co-expressed gene regions and genes. Results: Experiments on 13 public scRNA-seq datasets demonstrate that RF-SCGFS outperforms traditional methods with average improvements of 0.15 and 0.19 in normalized mutual information (NMI) and adjusted Rand index (ARI), respectively. When combined with mainstream unsupervised algorithms, RF-SCGFS achieves excellent performance (NMI > 0.91 on Yan and Biase datasets). In the PBMC-ctrl dataset, the method successfully identifies genes associated with immune system processes (GO:0006955, p = 2.02E-37). Discussion: RF-SCGFS addresses key challenges in single-cell analysis by reducing computational burden through efficient feature selection while maintaining biological relevance through unsupervised clustering-guided selection. Conclusion: RF-SCGFS provides an interpretable framework for feature selection in single-cell data, successfully identifying relevant disease genes and revealing the potential value of co-expressed gene regions in analyzing cellular heterogeneity.

G-quadruplexes (G4s) are non-canonical nucleic acid secondary structures that help maintain genomic stability and regulate gene transcription. Although the genome contains a vast number of putative G4-forming sequences (PQSs, sequences with intrinsic in vitro G4-forming potential), only a small fraction fold stably into G4 structures within the complex chromatin environment of living cells. Existing deep learning approaches improve predictive accuracy by incorporating cell line-specific epigenetic data; however, their heavy reliance on costly, large-scale sequencing assays (e.g., ChIP-seq) limits broader application to clinical samples and newly profiled cell lines. To address this challenge, we propose iDualG4, an interpretable dual-channel deep learning framework that uses DNA sequence as the only input at inference time. By leveraging a pretrained Enformer module, iDualG4 infers epigenomic proxy features directly from DNA sequence and integrates them with local sequence features, thereby replacing the need for newly measured cell-specific epigenomic assays during prediction. Evaluations across multiple cell lines, including K562, demonstrate that iDualG4 significantly outperforms existing methods, particularly in handling imbalanced data (achieving an AUPR of 0.981 on K562). Interpretability analysis based on DeepSHAP indicates that iDualG4 provides an in vivo G4 prediction tool combining high precision and interpretability without the need for additional experimental sequencing data, and offers a novel computational framework for elucidating how sequence and the epigenetic environment jointly determine genomic G4 formation.

Spatial multi-omics sequencing enables the simultaneous profiling of transcriptomics, proteomics, and epigenomics at a spatial resolution, offering insights into complex tissue organization and molecular regulation. However, the effective integration of multiple omics modalities in a spatial context remains a major challenge. Here, we present SpaDDM, a spatial multi-omics integration framework based on directional diffusion models (DDMs), which supports spatial pattern identification, cross-omics alignment, and inter-and intracellular signaling flow analysis. SpaDDM employs DDM-based graph networks to learn omics-specific representations by jointly incorporating spatial coordinates and molecular measurements within each modality, followed by an attention mechanism to align features across modalities. We benchmarked SpaDDM on diverse spatial multi-omics datasets, including transcriptomics-epigenomics and transcriptomics-proteomics combinations across multiple tissues and species. SpaDDM consistently outperformed existing methods by more accurately deciphering spatial tissue patterns and effectively reducing the boundary noise between spatial regions. Moreover, the learned low-dimensional coembedded representations of individual cells serve as integral mediators for inferring the signaling flows that underlie spatial patterning. Finally, we demonstrated that SpaDDM alignment of complementary information across multi-omics layers facilitates cross-omics translation and significantly improves the prediction of cell state alignments.

The radiosensitivity of cancer patients determines the efficacy of radiotherapy, and patients with low radiosensitivity cannot benefit from radiotherapy. Therefore, accurately predicting radiosensitivity before treatment is essential for personalized and precise radiotherapy. However, most existing studies rely solely on genomic and clinical features, neglecting the tumor microenvironmental information embedded in histopathological images, which limits prediction accuracy. To address this issue, we propose Resfusion, a deep multimodal fusion framework that integrates patient-level gene expression profiles, clinical records, and histopathological images for tumor radiosensitivity prediction. Specifically, the pre-trained large-scale pathology model is used as an image encoder to extract global representations from whole-slide pathological image. Radiosensitivity-related genes are selected using an autoencoder combined with univariate Cox regression, while clinically relevant variables are manually curated. The three modalities are first concatenated and then refined through a self-attention-based module, which captures inter-feature dependencies within the fused representation and highlights complementary information across modalities. The model was evaluated using five-fold cross-validation on two common tumor datasets suitable for radiotherapy: the Breast Invasive Carcinoma (BRCA) dataset (282 patients in total, with each fold partitioned into 226 training samples and 56 validation samples) and the Head and Neck Squamous Cell Carcinoma (HNSC) dataset (200 patients in total, with each fold partitioned into 161 training samples and 39 validation samples). The average AUC values obtained from the five-fold cross-validation reached 76.83% and 79.49%, respectively. Experimental results demonstrate that the Resfusion model significantly outperforms unimodal methods and existing multimodal fusion methods, verifying its effectiveness in predicting the radiosensitivity of tumor patients.

Hyperspectral image classification (HSIC) plays an important role in the field of remote sensing. However, its performance improvement faces challenges such as spectral variability and scarcity of labeled samples. In recent years, Graph Neural Networks (GNNs) have shown significant advantages in HSIC, but existing methods still have limitations. For example, although the Subgraph Adaptive Convolutional Network (SGDNet) alleviates the problem of subgraph topological imbalance by dynamically adjusting the feature aggregation method, it still fails to effectively capture long-range dependencies after subgraph division, resulting in insufficient global semantic consistency and limiting further improvement in classification performance. To address this issue, this paper proposes a Subgraph Contrastive Learning Network (SCLNet). SCLNet constructs subgraph datasets through seed sampling strategies and optimizes the feature representation within subgraphs using a contrastive learning mechanism, enhancing intra-class similarity and inter-class differences. In addition, SCLNet introduces a global feature buffer pool to integrate feature information between subgraphs, effectively solving the problem of long-range dependency loss caused by subgraph division in SGDNet and achieving classification performance improvement through local refinement and global collaboration. Experimental results show that SCLNet has achieved significant improvements in classification performance on multiple HSI datasets, verifying its effectiveness in solving the problem of long-range dependency loss and providing a new effective solution for HSIC.

Graph transformer networks have received more attention in hyperspectral image (HSI) classification. However, they overlooked the influence of graph connectivity strength in positional encoding and distribution. In order to address the above deficiencies, we proposed the novel graph transformer with structural embedding and training (GTSET) for HSI classification. Specifically, the structural embedding module firstly aimed at extracting effectively local and non-local feature information via patch-based distance encoding and centrality correlation coefficients based on graph connectivity strength, alleviating spectral variability. Secondly, the structural training module aimed at addressing imbalanced structural position distribution of labeled samples by leveraging the topological graph connectivity to determine their structural position distribution and reweighting the influence of labeled samples on the graph transformer training stage, exploring the guiding role of labeled samples in low spatial resolution of HSI. Next, we further refine training weights based on the spectral feature smoothness of labeled samples. Finally, comprehensive experiments on three real-world HSI datasets demonstrate that the GTSET achieves superior performance in HSI classification with limited labeled samples, compared to other popular classification methods. Implementation of GTSET, along with examples, can be found on the GitHub repository: https://github.com/xuchengchao0/GTSET.

The precise identification of promoters is crucial for understanding gene regulation. Deep learning methods have achieved considerable success in promoter prediction, yet most operate at the sequence level with coarse-grained labels. This means they label an entire DNA segment as either a “promoter” or “non-promoter,” which results in a lack of the nucleotide-level resolution in prediction. In this study, we propose EvoSNR-Prom, a model designed for promoter prediction at single-nucleotide resolution. EvoSNR-Prom is built on the Evo foundation model and formulates promoter identification as a token-level sequence labeling problem, analogous to named entity recognition in natural language processing. To address the limited contextual information available in single-nucleotide tokenization, we introduce a lexicon-enhanced embedding strategy that incorporates biologically meaningful DNA lexicons, enriching contextual representations and improving the model’s ability to capture complex sequence motifs. Furthermore, to enhance predictive performance on small size datasets, we integrate a label-aware transfer learning framework to leverage knowledge from well-annotated source species to a target organism. The results across various prokaryotic datasets show that EvoSNR-Prom achieves excellent performance. This work provides a valuable computational framework for the high-precision analysis of gene regulatory elements, contributing to the advancement of promoter prediction at single-nucleotide resolution.

Predicting drug-disease associations (DDAs) plays a crucial role in drug development and disease treatment. However, existing researches predominantly focus on single DDAs prediction task, often overlooking the intricate relationships among different tasks, which can further improve the performance of methods for DDAs prediction. To address this limitation, a multi-task prediction framework, capable of simultaneously predicting drug-disease, drug-protein, and disease-protein associations, is proposed, named MTP-DDA. The framework constructs three distinct graphs to reflect different relationships between biological entities. Then, based on these graphs, two sub-views and one main-view are constructed. For sub-view, corruption strategy is adopted to generate corrupted view, and Graph Convolutional Network (GCN) is employed to extract features from both the original view and its corrupted version, with contrastive learning applied to enhance feature representations. For main-view, GCN and Node2Vec are utilized to extract low-order and high-order node features respectively, and an attention mechanism is utilized for feature fusion. Finally, the node features from three views above are integrated, and the dot product operation is applied to the node features of association pairs to derive association scores, thereby enabling multi-task association prediction. Under 10-fold cross-validation, the proposed framework outperforms current methods on public datasets, demonstrating its effectiveness and robustness.

Aiming to address the challenges of excessive model parameters, high computational complexity, strong complex background interference, and inadequate small-target detection found in insulator defect detection when using UAV remote sensing imagery of transmission lines, we propose a lightweight multi-defect detection model—Insulator Defect Detection-DETR (IDD-DETR). Specifically, we introduce a lightweight multi-starblock feature extractor (LMS-FE) as the backbone network to enhance its feature extraction capacity. Next, in order to enhance small-defect detection performance, a multi-scale feature pyramid (SOEP) is constructed by integrating shallow high-resolution features into the neck network. Additionally, a lightweight multi-branch feature fusion module (LMB-FF) is designed to efficiently fuse spatial and semantic information of small defects, suppressing background interference while optimizing model complexity. Finally, experimental results demonstrate that IDD-DETR achieves a 2.2% improvement in mean average precision (mAP) on the insulator small-defect dataset compared with the baseline algorithm, with model parameters and computation reduced by 44.9% and 47.1%, respectively. It also reaches a detection speed of 61.2 frames per second, satisfying the lightweight and high-precision requirements for edge deployment in transmission line inspection scenarios.

The development of single-cell RNA sequencing (scRNA-seq) technology provides unprecedented opportunities for elucidating cell heterogeneity and gene expression. Identifying and discovering cell types through cell clustering is a crucial step in analyzing scRNA-seq data. However, the high-dimensionality nature and frequent dropout events of the data raise great challenges for cell clustering. Here, we propose a novel contrastive clustering framework called scSCCNIA (Similarity-matrix-based Contrastive Clustering with Neighbor Information Aggregation), for the accurate identification of cell clusters from scRNA-seq data. scSCCNIA adopts a Laplacian filter to conduct neighbor information aggregation, constructs different graph views by using special un-shared parameters Siamese encoders for data augmentation, and learns the latent low-dimensional embedding representations via similarity-matrix-based contrastive learning. Comparative analyses of multiple scRNA-seq datasets from different platforms and with varying cell numbers demonstrate that scSCCNIA outperforms existing methods in terms of cell clustering and marker gene identification. Furthermore, scSCCNIA reveals the heterogeneity and functional specificity of various cell types through Gene Ontology terms and Kyoto Encyclopedia of Genes and Genomes enrichment analyses. Overall, scSCCNIA is an effective algorithm for learning latent features from scRNA-seq data, enhancing cell type identification accuracy and facilitating downstream analyses of scRNA-seq data.

Since genomics was proposed, the exploration of genes has been the focus of research. The emergence of single-cell RNA sequencing (scRNA-seq) technology makes it possible to explore gene expression at the single-cell level. Due to the limitations of sequencing technology, the data contains a lot of noise. At the same time, it also has the characteristics of highdimensional and sparse. Clustering is a common method of analyzing scRNA-seq data. This paper proposes a novel singlecell clustering method called Robust Manifold Nonnegative LowRank Representation with Adaptive Total-Variation Regularization (MLRR-ATV). The Adaptive Total-Variation (ATV) regularization is introduced into Low-Rank Representation (LRR) model to reduce the influence of noise through gradient learning. Then, the linear and nonlinear manifold structures in the data are learned through Euclidean distance and cosine similarity, and more valuable information is retained. Because the model is non-convex, we use the Alternating Direction Method of Multipliers (ADMM) to optimize the model. We tested the performance of the MLRRATV model on eight real scRNA-seq datasets and selected nine state-of-the-art methods as comparison methods. The experimental results show that the performance of the MLRRATV model is better than the other nine methods.