Background: Marker genes are interpretable summaries of single-cell clusters,but evaluating semantic representations derived from them is vulnerable tolanguage-model leakage, privileged candidate information, and post hoc methodselection. Results: We introduce a staged evaluation framework that preservesnegative findings, closes LLM and deterministic-development branches underexplicit gates, and tests frozen representations on external datasets registeredbefore performance evaluation. DeepSeek free-text reasoning, constrained ontol-ogy selection, and semantic recovery did not demonstrate model-specific valueafter leakage and candidate-universe controls. Four frozen deterministic GeneOntology representations were then evaluated on three independent externaldatasets. All four improved adjusted Rand index over gene-name embeddingson each external dataset. MSOA Top-12 achieved the largest mean Delta ARI(+0.1542), while the strongest representation differed across biological settings.Gains persisted across seven hierarchical settings and 50-seed K-means analyses. Conclusions: The signal that persisted under frozen external validationcame from explicit structured biological knowledge evaluated through a lockedprotocol, not unconstrained generated interpretation or post hoc tuning.

The rapid development of sequence-based genomic models, such as DNA language models and sequence-to-activity models, has highlighted the urgent need for standardized benchmarks to systematically evaluate their performance across diverse biological tasks. Existing benchmark datasets are typically limited in scope, focusing on specific aspects of genome regulation or a limited set of prediction tasks. Here, we present GenBench, a comprehensive benchmark dataset for genomic modeling. GenBench integrates eight tasks spanning four key dimensions of genome function: 1) sequence composition (exon–intron detection, splice site detection), 2) sequence activity (gene expression prediction, MPRA-based cis-regulatory activity), 3) regulatory element identification (enhancer classification, promoter prediction), and 4) noncoding variant effect prediction (MPRA SNP perturbations, eQTL mapping). These tasks encompass binary, multi-class, regression, and long-range prediction settings, providing a broad and challenging evaluation framework. GenBench offers a unified benchmark dataset that comprehensively reflects sequence composition, activity, regulation, and variation, and we anticipate it will serve as a foundation for the development and comparison of future genomic models.

Background: Pathology reports contain clinically important information for cancer diagnosis, staging, and outcome research, but their unstructured format limits scalable reuse. We evaluated whether DeepSeek can support pathology report understanding across cancer type extraction, AJCC stage prediction, and prognosis classification. Methods: We conducted an independent benchmark on PathRep-Bench style pathology report tasks derived from The Cancer Genome Atlas. The dataset contained 9,523 reports split into 7,618 training, 953 validation, and 952 test reports. DeepSeek V4 Flash and DeepSeek V4 Pro were evaluated using JSON-constrained prompting. Cancer type identification used non-reasoning prompting, whereas AJCC staging and prognosis prompting used reasoning-mode prompting. We also trained local TF-IDF logistic regression prognosis models and tested whether DeepSeek-extracted structured variables improved supervised prognosis modeling. Results: DeepSeek V4 Flash achieved 0.9800 accuracy (95% CI, 0.9695--0.9884) and 0.9778 macro F1 (95% CI, 0.9660--0.9867) for cancer type identification. For AJCC stage prediction, DeepSeek V4 Flash achieved 0.8165 accuracy (95% CI, 0.7845--0.8468) and 0.7841 macro F1 (95% CI, 0.7421--0.8201) on 594 labeled test reports. DeepSeek V4 Pro did not significantly improve paired accuracy over Flash for either task and produced 14 empty final outputs during AJCC staging. DeepSeek prognosis prompting was weak on a 100-report subset, improving from 0.5083 to 0.5647 macro F1 with eight examples. A supervised TF-IDF text-only logistic regression classifier achieved the highest prognosis point estimate on the full test set, with 0.8571 accuracy and 0.8543 macro F1. Adding DeepSeek-extracted structured variables did not produce a statistically significant improvement. Conclusions: DeepSeek is effective for pathology information extraction and provides strong AJCC staging support, but prognosis prediction is better framed as supervised outcome modeling under the evaluated framework. These findings support a hybrid architecture in which DeepSeek performs extraction and staging support, while supervised text-based models handle prognosis prediction.

MOTIVATION:MicroRNAs are key post-transcriptional regulators whose dysregulation is associated with complex human diseases. Computational prediction can prioritize candidate miRNA-disease associations, but reliable evaluation is complicated by sparse labels, cold-start entities, limited biological context in bipartite graphs, and leakage when association-derived features are constructed before data splitting. RESULTS:We present RGCNMDA, a leakage-controlled multi-view framework that integrates global latent structure, local profiles and similarities, and pathway context. Within every fold, interaction profiles, GIP similarities, PCA inputs, MDMF factors, and miRNA-disease graph edges are reconstructed exclusively from training positives. Four independent factorized encoders transform the miRNA and disease interaction profiles and GIP similarities, while a fold-local MDMF branch captures global latent structure. These representations are integrated with a pathway-bridged graph containing miRNA, disease, and pathway nodes connected by six directed relation types. A relational graph convolutional network performs type- and direction-specific message passing, and node-wise gates adaptively fuse graph, MDMF, and combined profile and similarity representations before an MLP pair decoder scores candidate associations. On HMDD v4.0, RGCNMDA achieved AUCs of 0.9589, 0.9109, and 0.8836 under random, cold-disease, and cold-miRNA evaluation, respectively; on the processed independent-source RNADisease v4.0 benchmark, the corresponding values were 0.9580, 0.8446, and 0.8913. Same-protocol baseline comparisons and diagnostic analyses showed that the benefits of RGCNMDA were setting dependent, with the strongest pathway-related improvement under cold-disease evaluation. These results support the robustness of leakage-controlled multi-view learning across standard and cold-start evaluation settings. AVAILABILITY AND IMPLEMENTATION:https://github.com/hnuchao/pathway-RGCNMDA.

Accurate classification of cancer subtypes is critical for precision oncology, yet existing methods often fail to integrate multi-omic data while providing biologically interpretable insights. This study presents a novel hybrid deep learning framework for cancer classification that synergistically combines graph neural networks (GNNs) and transformers to model complex interactions across genomic, epigenetic, and functional biological layers. Our approach introduces three key innovations: (1) a heterogeneous graph integrating gene expression, DNA methylation, and pathway nodes with biologically meaningful edges (protein interactions, regulatory relationships); (2) discrete wavelet transforms for spatial-aware dimensionality reduction of methylation data, preserving critical regional patterns; and (3) pathway-guided attention mechanisms that prioritize oncogenic signaling pathways while ensuring interpretability. Experiments demonstrate superior performance over state-of-the-art methods in the classification of 23 cancer types and of breast-cancer subtypes. By bridging artificial intelligence (AI) with cancer biology, our framework enables interpretable multi-omic integration, offering clinicians pathway-level explanations for predictions while maintaining cross-platform robustness. This work advances precision oncology by providing both accurate classification and actionable biological insights, paving the way for improved therapeutic strategies.

Gene regulatory networks (GRNs) are crucial for revealing gene interactions and understanding cellular biological mechanisms. However, the high dimensionality and nonlinearity of gene expression data make accurate inference and reconstruction of large-scale GRNs a core computational challenge in systems biology. This study introduces a novel approach, termed MRF-XGBLC, for the reconstruction of large-scale Gene Regulatory Networks (GRNs) utilizing steady-state and time-series gene expression data through nonlinear ordinary differential equations. Firstly, MRF-XGBLC uses the maximum information coefficient (MIC) for dimensionality reduction and eliminates redundant regulatory relationships by calculating the MIC between factors as a prior step in model processing. Furthermore, recognizing the superior performance of the Lasso-Cox model in survival analysis, the feature fusion algorithm of this paper incorporates a hybrid model of XGBoost (eXtreme Gradient Boosting), RF (Random Forest), and Lasso-Cox (Least Absolute Shrinkage and Selection Operator-Cox proportional hazards regression model) integration to effectively train the nonlinear ordinary differential equations, thus improving the accuracy and stability of the inference algorithm. Extensive experiments on datasets of varying sizes demonstrate significant improvements over state-of-the-art methods. Cross-validation experiments on real gene datasets confirm the robustness and effectiveness of MRF-XGBLC.

Breast cancer (BC) is one of the most common causes of mortality among women. However, early detection of BC can effectively improve the treatment outcomes. Computer-aided diagnosis (CAD) systems can be utilized clinical specialists for accurate diagnosis of BC in its early stages. Due to their superior classification performance, deep learning (DL) methods have been extensively used in CAD systems. The classification accuracy of a DL model mainly depends on the parameters, such as weights and biases, of the deep neural network (DNN), which are optimized during the training phase. The training of DL models has been carried out by gradient-based techniques, e.g., stochastic gradient descent with momentum (SGDM) and adaptive momentum estimation (ADAM), and metaheuristic techniques, e.g., genetic algorithms (GA) and particle swarm optimization (PSO). However, these techniques suffer from frequent stagnation in local optima due to the huge search space, which can lead to sub-optimal DL performance. This article proposes a hybrid optimization algorithm, based on incorporating a simple gradient search mechanism into a metaheuristic technique, multi-verse optimizer (MVO), to facilitate the search for global optimal solution in the high-dimensional search space of DL models. A DL model for BC diagnosis is developed based on a three-hidden-layer DNN whose parameters are optimized using the proposed hybrid optimizer. Experimental analysis is carried out on the Wisconsin breast cancer dataset (WBCD) and the Wisconsin Diagnosis Breast Cancer (WDBC) dataset, each is divided into 70% for training and 30% for testing. For comparison reasons, similar DL models trained using various optimizers, including gradient-based, metaheuristic, and recently-proposed hybrid optimization algorithms, are also analyzed. The results demonstrate the superior performance of our optimizer in terms of attaining the most accurate DL model in the fastest convergence rate. The proposed model achieves outstanding metrics, including accuracy at 93.5%, precision at 88.06%, specificity at 93.06%, sensitivity at 95.64%, F1 score at 91.67%, and Matthew’s correlation coefficient (MCC) at 87.14% on WBCD, and accuracy at 96.73%, precision at 93.38%, specificity at 95.83%, sensitivity at 98.25%, F1 score at 95.75%, and MCC at 93.18% on WDBC, in just six epochs. This research significantly contributes to advancing CAD systems for BC, emphasizing the potential benefits of the proposed optimizer in medical classification domains.

Single-cell RNA sequencing (scRNA-seq) technology offers unprecedented opportunities for inferring gene regulatory networks (GRNs) at the genome level. However, scRNA-seq data is highly sparse and has a low signal-to-noise ratio with significant dropout. Many unsupervised or self-supervised models have been proposed to infer GRNs from large RNA-seq datasets, but few are suitable for scRNA-seq data. Recent research confirms that transcription factor (TF)-DNA binding data enables supervised GRN inference. In this paper, we propose a novel framework called GRNNLink, which leverages known GRNs to infer potential regulatory relationships between genes. First, we preprocess the raw scRNA-seq data. Then, we introduce an interactive graph encoder based on a graph recurrent neural network (GRNN) to refine gene features by capturing the correlations between network nodes. Finally, matrix completion is performed using the node correlation features to predict GRNs. To evaluate model performance, we compare GRNNLink with six existing GRN reconstruction methods across seven scRNA-seq datasets. The results demonstrate that our method exhibits high robustness and accuracy.

Temporal Action Detection (TAD) is a crucial task in video understanding. Its primary objective is to accurately identify the semantic labels of each action instance in an untrimmed video, along with their temporal range. This paper constructs the Temporal Relation-aware Global Attention Network (TRGA-Net), which is a long-term temporal context modelling network. The model comprises video preprocessing, spatiotemporal feature extraction, temporal context modelling, and temporal action detection header. TRGA-Net introduces a temporal context modelling-based temporal channel global attention module to efficiently perform long-term temporal context modelling. Experiments were conducted on the ActivityNet and THUMOS14 datasets to evaluate the performance of TRGA-Net. The results demonstrate better mean average precision (mAP) metrics than the previously proposed temporal detection model, verifying the usefulness of TRGA-Net for temporal context modelling.

BACKGROUND:Single-cell RNA sequencing (scRNA-seq) technology has emerged as a crucial tool for studying cellular heterogeneity. However, dropouts are inherent to the sequencing process, known as dropout events, posing challenges in downstream analysis and interpretation. Imputing dropout data becomes a critical concern in scRNA-seq data analysis. Present imputation methods predominantly rely on statistical or machine learning approaches, often overlooking inter-sample correlations. RESULTS:To address this limitation, We introduced SAE-Impute, a new computational method for imputing single-cell data by combining subspace regression and auto-encoders for enhancing the accuracy and reliability of the imputation process. Specifically, SAE-Impute assesses sample correlations via subspace regression, predicts potential dropout values, and then leverages these predictions within an autoencoder framework for interpolation. To validate the performance of SAE-Impute, we systematically conducted experiments on both simulated and real scRNA-seq datasets. These results highlight that SAE-Impute effectively reduces false negative signals in single-cell data and enhances the retrieval of dropout values, gene-gene and cell-cell correlations. Finally, We also conducted several downstream analyses on the imputed single-cell RNA sequencing (scRNA-seq) data, including the identification of differential gene expression, cell clustering and visualization, and cell trajectory construction. CONCLUSIONS:These results once again demonstrate that SAE-Impute is able to effectively reduce the droupouts in single-cell dataset, thereby improving the functional interpretability of the data.

With the advancement of new technologies, a huge amount of high dimensional data is being generated which is opening new opportunities and challenges to the study of cancer and diseases. In particular, distinguishing the patient-specific key components and modules which drive tumorigenesis is necessary to analyze. A complex disease generally does not initiate from the dysregulation of a single component but it is the result of the dysfunction of a group of components and networks which differs from patient to patient. However, a patient-specific network is required to understand the disease and its molecular mechanism. We address this requirement by constructing a patient-specific network by sample-specific network theory with integrating cancer-specific differentially expressed genes and elite genes. By elucidating patient-specific networks, it can identify the regulatory modules, driver genes as well as personalized disease networks which can lead to personalized drug design. This method can provide insight into how genes are associating with each other and characterized the patient-specific disease subtypes. The results show that this method can be beneficial for the detection of patient-specific differential modules and interaction between genes. Extensive analysis using existing literature, gene enrichment and survival analysis for three cancer types STAD, PAAD and LUAD shows the effectiveness of this method over other existing methods. In addition, this method can be useful for personalized therapeutics and drug design. This methodology is implemented in the R language and is available at https://github.com/riasatazim/PatientSpecificRNANetwork.

Due to the anonymity and imperfect supervision of the blockchain, criminal acts committed by criminals on the blockchain are difficult to be investigated. In recent years, scams based on district smart contracts have emerged one after another, among which the losses caused by Ponzi schemes have reached millions of dollars. However, there is little research on smart contract fraud identification at present, and for fraud detection, information utilization is not comprehensive, only based on a single feature or simply fused multiple features directly, without considering the duplication and connection between features. In order to solve these problems, in this paper, we propose a multi feature fusion scheme identification model (MFFSI) for smart contracts. Our contributions mainly include the following two points: 1) In terms of information use, the operands containing the information about the jump relationship of the opcode execution are retained; 2) In feature fusion, the operation code (opcode) and application binary interface (ABI) sequence features are extracted, and attention modules are used to guide the fusion of features to alleviate the interference of irrelevant features on classification. The results show that the model proposed in this paper has a good detection effect. The F1 score is higher than 86%, which is better than the previous.

Numerous experiments have demonstrated that long non-coding RNA (lncRNA) play an important role in various systems of the human body. LncRNA deletions or mutations can cause human disease. The prediction of lncRNA-disease associations is conducive to the diagnosis and prevention of complex diseases. As we all know, it is a time-consuming and expensive process to predict lncRNA-disease associations via biological experiments. However, the computation methods can effectively discover lncRNA-disease associations with less human and material resources. In this paper, we propose a neural network-based matrix factorization model to predict lncRNA-disease associations, which is called NeuMFLDA. NeuMFLDA first converts the one-hot encoding of disease or lncRNA into word vector via the embedding layer. Then combined with the memorization of the conventional matrix factorization and the generalization of the multi-layer perceptron, the lncRNA-disease associations can be predicted more accurately. In addition, as opposed to conventional pointwise loss function, a new pairwise loss function is proposed to update our model parameters. Our new loss function optimizes the model from the perspective of ranking priority, which is more in line with the solution to the lncRNA-disease associations prediction task. Experiments show that NeuMFLDA reaches average AUCs of 0.904± 0.003 and 0.918± 0.002 in the framework of 5-fold cross validation and Leave-one-out cross validation, which is superior to three the-state-of-art methods. In case studies, 9, 9 and 8 out of top-10 candidate lncRNAs are verified by recently published literatures for hepatocelluar carcinoma, kidney cancer and ovarian cancer, respectively. In short, NeuMFLDA is an effective tool for predicting lncRNA-disease associations.

BACKGROUND:Single-cell RNA-sequencing enables the opportunity to investigate cell heterogeneity, discover new types of cells and to perform transcriptomic reconstruction at a single-cell resolution. Due to technical inadequacy, the presence of dropout events hinders the downstream and differential expression analysis. Therefore, it demands an efficient and accurate approach to recover the true gene expression. To fill the gap, we present a novel Single-cell RNA dropout imputation method to retrieve the original gene expression of the genes with excessive zero and near-zero counts.RESULT:Here we have developed CDSImpute (Correlation Distance Similarity Imputation) to identify dropouts induced in scRNA-seq data rather than biological zeros and recover true gene expression. By taking into consideration correlation and negative distance between cells, a similar cell list has been created and by borrowing the gene expression from similar cells dropout has been detected and corrected simultaneously. The improvement is consistent with simulation data and several publicly available scRNA-seq datasets. The clustering accuracy of CDSImpute is evaluated by adjusted rand index on Kolod, Pollen and Usoskin datasets are 1.00, 0.79 and 0.34 respectively. CDSImpute achieves improved performance compared to the three existing methods evaluated by precise cell-type identification and differentially expressed gene detection from scRNA-seq Data.CONCLUSION:CDSImpute is a novel effective method to impute the dropout events of a scRNA-seq expression matrix. The package is implemented in the R language and is available at https://github.com/riasatazim/CDSImpute.

Single-cell multi-omics sequencing technologies allow simultaneous measurement of transcriptome and epigenome profiles in the same cell, providing unprecedented opportunities to dissect cell heterogeneity. Despite great efforts, conjoint analysis of single-cell multi-omics data still suffers from sparsity, high dimensionality and binary. In this study, we present a heterogeneous graph cross-omics attention model (scHGA), a computational tool based on a heterogeneous graph neural network combining two attention mechanisms to jointly analyze single-cell multi-omics data based on different protocols data, including SNARE-seq, scMT-seq and sci-CAR. To avoid the cell heterogeneity of single-omics data, scHGA automatically learns a cell association graph to capture neighbor information. The latent representation of aggregated cells generated by hierarchical attention can fuse knowledge across different omics to dissect cellular heterogeneity, providing a better scheme to characterize the features of cells. scHGA is an effective exploration of graph neural networks in single-cell multi-omics analysis, providing new insights into the understanding of single-cell sequencing data.

MicroRNA (miRNA) is a class of non-coding single-stranded RNA molecules encoded by endogenous genes with a length of about 22 nucleotides. MiRNAs have been successfully identified as differentially expressed in various cancers. There is evidence that disorders of miRNAs are associated with a variety of complex diseases. Therefore, inferring potential miRNA-disease associations (MDAs) is very important for understanding the aetiology and pathogenesis of many diseases and is useful to disease diagnosis, prognosis and treatment. First, We creatively fused multiple similarity subnetworks from multi-sources for miRNAs, genes and diseases by multiplexing technology, respectively. Then, three multiplexed biological subnetworks are connected through the extended binary association to form a tripartite complete heterogeneous multiplexed network (Tri-HM). Finally, because the constructed Tri-HM network can retain subnetworks’ original topology and biological functions and expands the binary association and dependence between the three biological entities, rich neighbourhood information is obtained iteratively from neighbours by a non-equilibrium random walk. Through cross-validation, our tri-HM-RWR model obtained an AUC value of 0.8657, and an AUPR value of 0.2139 in the global 5-fold cross-validation, which shows that our model can more fully speculate disease-related miRNAs.

Background: Recently, ample researches show that microRNAs (miRNAs) not only interact with coding genes but interact with a pool of different RNAs. Those RNAs are called miRNA sponges, including long non-coding RNAs (lncRNAs), circular RNA, pseudogenes and various messenger RNAs. Understanding regulatory networks of miRNA sponges can better help researchers to study the mechanisms of breast cancers. Objective: We develop a new method to explore miRNA sponge networks of breast cancer by combining miRNAdisease-lncRNA and miRNA-target networks (MSNMDL). Method: Firstly, MSNMDL infers miRNA-lncRNA functional similarity networks from miRNAdisease- lncRNA networks. Secondly, MSNMDL forms lncRNA-target networks by using lncRNA to replace the role of matched miRNA in miRNA-target networks according to the lncRNA-miRNA pair of miRNA-lncRNA functional similarity networks. And MSNMDL only retains the genes of breast cancer in lncRNA-target networks to construct candidate miRNA sponge networks. Thirdly, MSNMDL merges these candidate miRNA sponge networks with other miRNA sponge interactions and then selects top-hub lncRNA and its interactions to construct miRNA sponge networks. Results: MSNMDL is superior to other methods in terms of biological significance and its identified modules might act as module signatures for prognostication of breast cancer. Conclusion: MiRNA sponge networks identified by MSNMDL are biologically significant and are closely associated with breast cancer, which makes MSNMDL a promising way for researchers to study the pathogenesis of breast cancer.

A major challenge in cancer genomics is to identify cancer driver genes and modules. Most existing methods to identify cancer driver modules (iCDM) identify groups of genes whose somatic mutational patterns exhibit either mutual exclusivity or high coverage of patient samples, without considering other biological information from multiomics data sets. Here we integrate mutual exclusivity, coverage, and protein-protein interaction information to construct an edge-weighted network, and present a graph clustering approach based on symmetric non-negative matrix factorization to iCDM. iCDM was tested on pan-cancer data and the results were compared with those from several advanced computational methods. Our approach outperformed other methods in recovering known cancer driver modules, and the identified driver modules showed high accuracy in classifying normal and tumor samples.

The recent development of a high throughput single-cell RNA sequence devises the opportunity to study entire transcriptomes in the smallest detail. It also leads to the characterization of molecules and subtypes of a cell. Cancer epigenetics induced not only from individual molecules but also from the dysfunction of the system and the coupling effect of genes. While rapid advances are being made in the development of tools for single-cell RNA-seq data analysis, few slants are noticed in the potential advantages of single-cell network construction. Here, we used network perturbation theory with significant analysis to develop a cell-specific network that provides an insight into gene-gene association based on molecular expressions in a single-cell resolution. Besides, using this method, we can characterize each cell by inspecting how genes are connected and can identify the hub genes using network degree theory. Pathway & Gene enrichment analysis of the identified cell-specific high network degree genes supported the effectiveness of this method. This method could be beneficial for personalized drug design and even therapeutics.

Clustering gene expression profiles is a fundamental task in the genome and biomedical research. With the development of RNA-seq and gene chip technology, mass gene expression profile data has been generated, which puts forward two requirements for related research of gene expression profile: i) accurate analysis of drug R&D requires high accuracy of similarity analysis, ii) large-scale analysis of data requires as little running time as possible. We propose a faster, more accurate method called DeepCDNet, which is based on the framework of the Siamese network. DeepCDNet uses the DenseNet structure and optimized loss function to achieve rapid convergence, and the similarity between expression spectra is calculated by a cosine function. The experiment results show that: i) our method breaks through the limitation of high dimensions of gene expression profile and can quickly and accurately learn the required gene characteristics, ii) The accuracy of our method in similarity analysis is greatly improved, iii) as the dimension of data increases, the advantage of our method on time cost gradually becomes more prominent, and time consumption is less.