Cancer subtype classification requires capturing complex multi-omics interactions that conventional machine learning models often struggle to represent effectively. Graph Convolutional Networks (GCNs) leverage biological topologies but typically rely on fixed propagation mechanisms, which limit adaptability to diverse graph structures. This study introduces the Dynamic Bernstein Graph Convolutional Network (DB-GCN), a novel architecture that employs adaptive spectral propagation using Bernstein polynomials to enable topology-aware learning without eigendecomposition. DB-GCN supports single-omics (RNA) and multi-omics (RNA+CNV) inputs within a graph-based framework that represents genes as nodes and interactions as edges derived from gene-gene, protein-protein, and Co-expression networks. A dual-stream design combines a Bernstein graph stream with an omics multilayer perceptron to capture both local and global features. In pan-cancer experiments on 28 TCGA subtypes, DB-GCN achieves 86.05% $\pm$ 0.83 on STRING, 85.86% $\pm$ 0.98 on BioGRID, and 85.88% $\pm$ 0.71 on Co-expression for the 2,000-gene multi-omics setting. SHAP-based analysis identifies putative biomarker genes such as KLK11, OR4F15, and UBE2DNL, and 12 of the top 50 genes map to KEGG cancer pathways. These results indicate that DB-GCN provides an accurate and interpretable graph-based framework for pan-cancer subtype classification and biomarker discovery in precision oncology.
Automatic question-answering systems demonstrate valuable utility in the biomedical domain, improving the precision and efficiency of clinical decision-making significantly. Despite large-scale language models achieving notable success in general domains, even outperforming human-level performance in certain aspects, they are still faced with challenges such as data privacy and scarcity in the special domain. This study explores the method for efficient fine-tuning under resource-constrained conditions in the biomedical field. We propose a multi-stage fine-tuning approach that effectively improves the performance of pre-trained language models in biomedical question-answering tasks. Specially, A multi-prompt-based contrastive learning strategy and a multi-prompt self-consistency voting module are introduced, which improve the accuracy of QA tasks. The experiments on the PubMedQA dataset under reasoning-required settings indicate that our approach outperforms domain-specific pre-training models and achieves comparable performance with GPT-4, while the number of fine-tuned parameters is much less than the total parameters of the base model.
Drug combination therapy for complex diseases is currently widely utilized in clinical treatment. An increasing number of computational methods are being adopted to discover new drug combinations. However, most existing methods mainly focus on phenotypes, while ignoring the deep level interactions between drug pairs and cell lines. In this paper, we introduce a multimodal hypergraph representation learning approach named MHGSynergy, aimed at predicting drug synergies. MHGSynergy models the synergistic relationship as a hypergraph by considering the structure, targets, and physicochemical features of drugs. The nodes represent drugs and cell lines, and hyperedges capture synergistic triplets of drug pairs and cell lines. Different drug features are used as attributes of drug nodes in the hypergraph to construct three hypergraphs. Hypergraph neural networks are employed to update embedding features of drugs and cell lines, then input these embedding features into the channel attention fusion module to obtain a comprehensive representation. Finally, these representations are used to build predictive models. In both classification and regression tasks, MHGSynergy performs well on two benchmark datasets, and outperforming those baseline methods, and MHGSynergy has unique advantages when facing unknown drug pairs or cell lines. We believe that MHGSynergy is a valuable tool for discovering synergistic drug combinations.
Classifying drugs into symptomatic (SYM) and disease-modifying (DM) categories is essential for understanding their therapeutic effect and plays a key role in drug repurposing. While many computational approaches focus on drug-target prediction, they often ignore the nature of the drug's action on disease progression. This study proposes a graph-based strategy to classify drugs as SYM or DM based on their effect on disease treatment. We construct a heterogeneous network comprising genes, diseases, and drugs, and apply a guided shortest path traversal framework for drug effect classification. During this traversal, certain genes appear frequently in the shortest metapaths linking diseases and drugs. These recurrent genes are identified based on their frequency of occurrence in known drug-disease paths. For a new drug-disease pair, if the traversal path contains recurrent genes marked for a specific treatment type, we classify the drug accordingly. Over and above classifying the drugs, the proposed method incorporates the metapath-based framework to improve interpretability. Experimental results show that our model achieves significantly better classification accuracy compared to advanced machine learning and deep learning methods. A case study on multiple sclerosis further supports the biological relevance of our approach.
Compared with monotherapy, combination drug therapy plays a crucial role in clinical treatment. However, the exponential expansion of the drug combination space has rendered traditional exploration methods for synergistic drug combinations inadequate. Recently, numerous efficient and accurate computational approaches have been developed to predict anticancer drug synergy, particularly those leveraging hypergraphs to model the multifaceted relationships between drug combinations and cell lines, which have demonstrated remarkable potential. Nevertheless, existing hypergraph-based methods fail to account for the heterogeneity of anticancer synergy hypergraphs and overlook the underlying similarities among drugs and cell lines, thereby limiting their ability to fully capture the complex interactions between drug combinations and cell lines. To address these limitations, we propose an Adaptive Heterogeneous Hypergraph Representation Learning Method (HHGSynergy) for predicting anticancer drug synergy, enabling more precise identification of synergistic drug combinations. Specifically, our framework first constructs drug/cell line similarity-based synergy hypergraphs based on the foundational anticancer synergy hypergraph, thereby establishing a comprehensive heterogeneous hypergraph. Next, a node importance calculation module is employed to learn both local and global importance weights of nodes, effectively capturing the structural characteristics of the hypergraph. Finally, a type-specific multi-head attention mechanism is utilized to iteratively update node embeddings, adaptively learning the significance of heterogeneous hyperedges. Experimental results demonstrate that HHGSynergy achieves state-of-the-art performance in both classification and regression tasks across diverse experimental scenarios, outperforming existing leading models. Case studies further underscore its potential for discovering novel synergistic drug combinations.
2'-O-methylation (2OM) of ribose is a widespread RNA modification that significantly impacts RNA stability, structure, and function. Accurately predicting 2OM sites is crucial for understanding RNA's biological functions and related pathologies. Traditional detection methods pose challenges such as resource intensiveness, potential RNA sample damage, and high costs. However, recent advancements in machine learning, particularly deep learning techniques, offer rapid and cost-effective prediction solutions. In this study, we introduce DeepR2OM, a novel method integrating feature selection and deep learning for 2OM sites prediction. DeepR2OM encodes sequences using eight RNA descriptors, employs feature selection algorithms to reduce dimensions, and then utilizes a deep learning network for training. After evaluating various deep learning architectures, we selected Convolutional Neural Network (CNN), Multi-Head Self-Attention mechanism, and Deep Neural Network (DNN) as our final prediction models. Experimental results demonstrate DeepR2OM's effectiveness, achieving 87.1% accuracy (ACC), 85.5% recall rate (Recall), 87.9% precision (PRE), and a Matthews correlation coefficient (MCC) of 75.7% on an independent test set. This tool serves as a valuable resource for exploring the functional and bioinformatic aspects of 2OM sites.
This study investigates the asynchronous controllability of non-homogeneous Markov switch generalized asynchronous Boolean control networks (NMHGABCNs) and random switching signals in these networks, aiming to follow a non-homogeneous Markov process. The controllability of the proposed networks is achieved using the discrepancy between the Markov chain mode and the control mode. Assisted by the semi-tensor product (STP), the algebraic forms of the NMHGABCNs are obtained, and the sufficient and necessary criteria for their asynchronous controllability are derived. The effectiveness of controllability is demonstrated through two examples, which validate the theoretical results.
Single-cell RNA sequencing (scRNA-seq) enables unprecedented exploration of cellular heterogeneity, yet technical variations across datasets introduce pervasive batch effects that severely compromise integrative analysis. We present Graph-based Adversarial Latent Alignment (GALA), a novel batch correction framework that synergistically integrates weighted graph random walks with latent space adversarial training to robustly align scRNA-seq data while preserving critical biological signals. GALA employs a Weighted Graph Mutual Nearest Neighbor (WGMNN) module that achieves up to 125% improvement in cross-batch cell pairing diversity and 48% increase in coverage compared to conventional approaches, substantially enhancing the detection of biologically meaningful correspondences between batches. These optimized cell pairs guide adversarial training within a carefully designed low-dimensional latent space, generating batch-agnostic representations that simultaneously eliminate technical artifacts while faithfully preserving biological variability. Evaluated across five benchmark datasets representing diverse real-world scenarios-including identical cell types, non-identical cell types, and complex multi-batch settings-GALA demonstrates robust performance in batch correction and biological signal preservation, achieving high F1 scores of 0.91, 0.82, 0.88, 0.82, and 0.76, respectively. GALA outperforms established methods including Seurat v4, Harmony, and Scanorama, with notable advantages in challenging datasets with heterogeneous cell populations. Aggregating performance across all datasets, GALA achieves the highest overall integration score and best average rank while maintaining high computational efficiency, demonstrating consistent superiority and strong robustness across diverse scRNA-seq integration scenarios.
Machine learning has been successfully applied to drug combination prediction in recent years. However, in some situations, the class imbalance problem still shows highly negative impacts on the modeling process, which cannot be directly handled by traditional methods. In addition, the interpretability of models is another key point for biological and medical experts. In this study, a clustering-based oblique decision tree (MOCT) algorithm is proposed to extract interpretable knowledge for the multi-class datasets. It firstly clusters samples of different classes, and then a proper feature subspace is generated to split data and forms a nonleaf node. Unlike traditional decision trees, our MOCT only grows one none-leaf node in each layer to generate a concise tree structure. Datasets of drug combinations were collected from three cell lines with three classes (Additive, Antagonism, and Synergy) in experiments, and the results show that our MOCT algorithm is superior to other methods with better interpretability.
The association between miRNAs and diseases is crucial for understanding pathological mechanisms. However, existing methods often struggle to capture deep topological structures and suffer significant performance degradation under sparse associations. To address these challenges, we propose FKAMHV, a novel framework that integrates Fast Kolmogorov-Arnold Networks (FastKAN) and Multi-Head Hypergraph Convolution Networks (Multi-Head HGCN) to extract deep topological features, and incorporates $\beta$-Variational Autoencoder ($\beta$-VAE) to uncover latent associations. First, we construct heterogeneous networks based on functional, semantic, and Gaussian kernel similarities, and generate miRNA and disease specific hypergraphs using the K-nearest neighbors (KNN) algorithm. Next, FastKAN is employed to perform nonlinear modeling of node features, enhancing their representational capacity, and is combined with Multi-Head HGCN to strengthen the joint representation of structural and attribute information. To extract high-order topological features while avoiding redundancy, we introduce edge attention weights, a channel-wise squeeze-and-excitation (SE) mechanism, and a Jumping Knowledge strategy in the Multi-Head HGCN. Finally, $\beta$-VAE is used to model latent association distributions, where the $\beta$ coefficient serves as a regularization factor to balance reconstruction accuracy and latent space disentanglement, thereby improving generalization under sparse conditions. A learnable coefficient is introduced to fuse the outputs of the two modules for final prediction. Experimental results show that FKAMHV outperforms existing methods in terms of AUC and AUPR, and also achieves strong performance under sparse scenarios.
Single-cell multimodal data can simultaneously provide cellular features at different levels, such as gene expression, chromatin accessibility, and spatial location. The integration of multimodal data can efficiently utilize the information from various views, thereby enhancing the reliability and accuracy of cellular research. The current integration methods mainly focus on obtaining the representation of cells but neglect the representation of genes, not beneficial to cell typespecific gene module analyses. Besides, some integration algorithms only can integrate multi-omics data and cannot be applied to spatial transcriptome data for integrating transcriptomic data and spatial location. To this end, we propose MISF, a Multimodal data Integration algorithm based on adaptive Similarity network learning and matrix Factorization. MISF integrates multimodal data and learns the lower-dimensional representations of cells and genes. We validate the feasibility of MISF on multiple single-cell multi-omics data and spatial transcriptome data and compare it with the existing multi-omics data integration methods as well as spatial transcriptome data analysis algorithms. The results demonstrate that MISF can effectively integrate multimodal data, localize different types of cell clusters, and outline cellular spatial distribution pattern. Furthermore, MISF facilitates cell clustering and cell type-specific gene module analyses, providing new insights for the study of cellular heterogeneity.
The execution of most protein functions relies on the binding of ion ligands. Therefore,The accurate prediction of binding residues of proteins and ion ligands plays a crucial role in the execution of protein functions. However, due to the small sample size and class imbalance of small sample ion ligands, enhancing the prediction accuracy of binding residues remains challenging. This paper introduces a novel integrated algorithm, SG-DCNN, which combines Generative Adversarial Network (GAN) and Self-Attention within the Deep Convolutional Neural Network (DCNN) framework. Firstly, the theoretical analysis assesses the effectiveness of SG-DCNN in enhancing prediction accuracy. Subsequently, eight small sample ion ligands are chosen for experimental validation. The results demonstrate that SG-DCNN achieves an independent testing accuracy of 78.3% for Sn and a Matthews correlation coefficient (Mcc) of 0.23, outperforming previous prediction methods. The programs and materials utilized in this study can be accessed at (https://github.com/wahah0414/SG-DCNN.git).
Lysine crotonylation (Kcr) is an important protein modification occurring after translation in biology, serving an essential function in a range of biological processes in both plants and animals, including the regulation of gene expression, the maintenance of cellular metabolic balance, and the enhancement of photosynthesis. Exploring the detection of Kcr sites is essential for uncovering their biological functions. Nonetheless, conventional experimental approaches for detection are often time-consuming, expensive, and hindered by various technical constraints, making the precise identification of Kcr sites a significant challenge. This study seeks to develop a computational approach for the rapid and accurate prediction of Kcr sites in plant non-histone proteins. We introduce a novel deep learning framework named DeepNhKcr, which integrates the protein language model (ESM2) with a bidirectional long short-term memory (BiLSTM) network. To address the challenge of data imbalance, the model replaces the conventional cross-entropy loss with the focal loss function. In addition, DeepNhKcr combines advanced deep learning approaches with traditional protein encoding strategies to enable effective feature extraction and integration. This method not only significantly boosts the accuracy of predicting Kcr sites in non-histone proteins of plants. but also provides interpretability, shedding light on the potential links between key sequence characteristics and their biological roles. DeepNhKcr delivers outstanding results, surpassing existing machine learning and deep learning models, and demonstrating excellent performance in both five-fold cross-validation and independent test experiments. Moreover, the model integrates interpretability analysis techniques to investigate the connections between important sequence features and their biological roles. DeepNhKcr acts as a powerful method for detecting Kcr sites in plant non-histone proteins and is anticipated to greatly advance future studies in plant Kcr site prediction.
In the past few decades, several methods, ranging from statistical algorithms to deep learning models, have been proposed for driver gene identification. Although recent graph neural network-based methods have achieved progress using biological networks, the static and incomplete nature of networks can compromise the predictive power of the models. Here, we introduce a multimodal approach for driver gene prediction called DriverMONI that leverages pairs of complementary information from multiomics data and biological networks. In particular, DriverMONI utilizes condition-specific subnetworks of protein-protein interaction for generating input graphs and node attributes for graph attention network, which makes the model condition-specific and improves prediction accuracy. We empirically show the importance of multimodality and the ability of DriverMONI to mitigate the incompleteness of protein-protein interaction networks. In addition, comparative analysis showed that DriverMONI achieved the best performance among competing methods including graph neural network-based models using The Cancer Genome Atlas data, while having a strong consensus with other methods.
Comprehending biological reproduction and cellular metabolism is facilitated by the Enzyme Commission, which matches protein sequences to the biochemical reactions they catalyse through EC numbers. In recent years, several methods have been proposed for predicting enzyme function. However, these methods still encounter challenges. Firstly, traditional methods for manually designing enzyme features are complex and cumbersome, lacking an effective generalized method for embedding enzyme sequences. Secondly, the distribution gap between different enzymes is significant, which resulting in existing methods struggling to predict multilevel enzyme functions. Thirdly, traditional enzyme function prediction models only extract single view feature of enzyme, so there is still room for further improving the ability of these models to extract enzyme data. To address these challenges, a new multilevel enzyme function prediction model (SMENET) based on multi-view semantics is proposed. This method uses protein large language model to extract semantic information. Subsequently, this semantic information is fed into multiple information extraction network modules, followed by using Biologic Sematic Attention to integrate these views' information. Finally, a multi-view adaptive fusion network is designed to extract the best common representation between multiple semantic views. Extensive experiments were conducted on multiple datasets to validate the effectiveness of SMENET.
Researchers in bioinformatics use protein function prediction to classify proteins, study their roles in disease mechanisms. Many mechanisms have been proposed to address the significant gap between the rapid increase in identified proteins and the slower growth in annotated protein functions. While PPI network data can assist in protein function prediction, it is often overlooked. Moreover, existing approaches primarily rely on node2vec embeddings of PPI data, neglecting valuable information about the neighborhood relationships of proteins. In this paper, we proposed a novel multi-modal approach that leverages a Graph Convolutional Network (GCN) to analyze PPI network data and explore neighboring proteins. Additionally, we employ Multi-Head Self-Attention combined with a Convolutional Bi-LSTM on protein sequence data to enhance the prediction of protein functions. To evaluate the effectiveness of our method, we utilize benchmark datasets, specifically from homo sapiens, such as String database for PPI network information and UniprotKB for protein sequences, ensuring a comprehensive and robust analysis. Extensive experiments demonstrate that the proposed method, Multi-DeepProtGraphGO, achieves improvements of +18.28%, +4.56%, and +6.92% in terms of the Fmax score for the BP, CC, and MF sub-ontologies, respectively, compared to the state-of-the-art method CrossPredGO.
Accurate classification of cancer subtypes from microarray gene expression data remains challenging due to high dimensionality, class imbalance, and limited sample sizes. To address these issues, we propose GexpNet, a novel deep learning architecture specifically designed for gene expression analysis. GexpNet integrates an initial convolutional layer with dynamic multi-head residual blocks that employ heterogeneous kernel sizes and activation functions. These are followed by an adaptive dense classifier, enabling effective multi-scale feature extraction and robust classification. To reduce noise and overfitting, a customized preprocessing pipeline is introduced, comprising median imputation, quantile normalization, and supervised feature selection based on logistic regression. Evaluated on four publicly available datasets, GexpNet achieves 98.99% accuracy (95% CI: 98.57-99.43% ) and 98.26% F1-score (95% CI: 97.51-99.01% ) on the Mendeley dataset, outperforming prior deep learning models by up to 1.99% . Across three additional datasets, GexpNet consistently exceeds 99% accuracy with minimal variance in 10-fold cross-validation, confirming its robustness and generalizability. Overall, these results highlight GexpNet as a robust, scalable, and generalizable framework for precision oncology applications.
Cyclic peptides represent a rapidly growing class of therapeutics, yet their development is often hindered by the challenge of predicting cell membrane permeability, a critical determinant of drug efficacy. Existing computational methods often struggle to integrate the diverse structural information inherent in these complex molecules, resulting in suboptimal predictive accuracy. Here, we introduce MCPerm, a multi-modal deep learning framework that synergistically integrates 1D SMILES, 2D topological, and 3D geometric information through a novel modality share and contrastive learning strategy to accurately predict cyclic peptide permeability. MCPerm fine-tunes a pretrained peptide language model for SMILES encoding and uses a parameter-sharing graph transformer for structural representation, while a dual contrastive learning mechanism enforces representational consistency both within and between modalities. On the benchmark PAMPA dataset, MCPerm achieves state-of-the-art performance, significantly outperforming leading methods. We further demonstrate its robustness and competitive transferability across three independent assays (Caco-2, MDCK, and RRCK). Our work presents a robust in silico framework that holds potential to accelerate the rational design and discovery of cell-permeable cyclic peptide drugs. Furthermore, to move beyond predictive accuracy, we introduced an attention-based visualization analysis. The results demonstrate that our model is not a “black box”; it has learned key chemical principles governing cyclic peptide permeability.
Interactions among circular RNAs (circRNAs) and RNA-binding proteins (RBPs) involve almost all stages of the circRNA life cycle. Therefore, circRNA-RBP binding site identification is extremely important for the regulation of human diseases. Various approaches have been used to identify RBP binding sites on circRNAs. Sadly, these approaches are frequently constrained by insufficient feature learning and poor scalability. As a result, we provide a novel model named iDeep-cancer that predicts circRNA-RBP interactions solely using circRNA sequences. A hybrid deep learning model and feature encoding are used in the iDeep-cancer technique. In order to create the feature space, feature encoding uses four feature extraction techniques while accounting for the chemical makeup of circRNA sequences. The hybrid network includes an improved dense convolutional network (DenseNet), a bidirectional gated recurrent unit (BiGRU), and a self-attention mechanism (Self-attention). DenseNet is used to learn high-level localized features, and the combination of BiGRU and self-attention captures long-term dependencies in sequences. We conducted ablation tests and compared iDeep-cancer with other cutting-edge techniques on 13 datasets in order to verify its efficacy. Findings indicate that iDeep-cancer performs better than current techniques.
