The association between miRNAs and diseases is crucial forunderstanding 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.

Metabolites, the end products of biological processes, hold immense potential as biomarkers for elucidating disease mechanisms. Given the costs associated with wet-lab experiments, computational methods offer a viable solution to narrow the search for candidate metabolites. However, sparse disease-metabolite associations pose a challenge to current deep learning-based methods, limiting comprehensive feature learning. To address this, we propose a novel Weighted Graph Convolutional Network approach based on Community-Driven Link Completion (WGCNCDLC) to improve the predictive accuracy of disease-metabolite associations. Specifically, we construct disease and metabolite similarity networks to complete potential links between same-type nodes. Furthermore, we partition the similarity networks into multiple communities and calculate the connection strengths between disease and metabolite communities to enrich the sparse links between different-type nodes. The completed network enables richer feature capture in subsequent weighted graph convolutional network encoding. Experimental results on two datasets show our WGCNCDLC model outperforms nine state-of-the-art algorithms. Case studies on Alzheimer's disease and asthma further validate the model's reliability as a predictive tool for potential metabolite discovery.

The study of the relationship between circular RNA (circRNA) and disease is crucial for understanding the mechanisms underlying disease onset. However, relying on biological experiments to explore all potential connections between circRNAs and diseases is both time-consuming and labor-intensive. While various prediction methods have been proposed, they still possess certain limitations in their ability to extract deep features. In this study, we introduce an innovative computational framework called Graph Isomorphism Networks and Graph Sampling Aggregation for predicting unknown circRNA-disease associations (GINSACDA). Specifically, GINSACDA first computes the Gaussian interactive profile kernel (GIP) similarity and functional similarity of circRNAs, as well as the GIP similarity and semantic similarity of diseases, serving as global features. Then, node labels extracted from seven-hop subgraphs connected to the target nodes are used as local features, which are fused with the global features. Next, the fused features are input into a Graph Isomorphism Network (GIN) for feature extraction and combined with the Graph Sampling Aggregation (GraphSAGE) method to extract deeper hidden features. Finally, we employed a fully connected layer to compute the prediction scores. The results of five-fold cross-validation conducted on two datasets indicate that GINSACDA outperforms five other state-of-the-art models. Additionally, we conducted case studies on hepatocellular carcinoma and breast cancer to further validate the superior predictive capabilities of our model.

Increasing biological research indicates that the expression levels of circRNAs fluctuate during the onset of various diseases, making them potential biomarkers for multiple conditions. Although numerous artificial intelligence-based computational methods are currently employed for circRNA-disease associations prediction, these methods often rely on a single objective function, which can lead to suboptimal prediction accuracy. To date, no method has designed a set of multi-objective functions specifically for the circRNA-disease prediction problem and optimized it using a non-dominated sorting genetic algorithm. This paper introduces a novel approach by utilizing multi-objective functions and an improved non-dominated sorting genetic algorithm (ICDNSGA) to identify potential associations of circRNA-disease. The method constructs a solution space through matrix factorization and network community characteristics, designing four distinct objective functions optimized via the enhanced multi-objective non-dominated sorting genetic algorithm. ICDNSGA incorporates a population-based adaptive normalization strategy, improving algorithm convergence and solution diversity. Experimental results show that ICDNSGA outperforms pure matrix factorization methods, non-dominated sorting genetic algorithms and other machine learning techniques in predictive performance. Additionally, the prediction results can be validated through existing research and biological analyses, underscoring ICDNSGA's potential as a valuable tool for biomedical experimentation.

As a series of single-stranded RNAs, circRNAs have been implicated in numerous diseases and can serve as valuable biomarkers for disease therapy and prevention. However, traditional biological experiments demand significant time and effort. Therefore, various computational methods have been proposed to address this limitation, but how to extract features more comprehensively remains a challenge that needs further attention in the future. In this study, we propose a unique approach to predict circRNA-disease associations based on resistance distance and graph attention network (RDGAN). First, the associations of circRNA and disease are obtained by fusing multiple databases, and resistance distance as a similarity matrix is used to further deal with the sparse of the similarity matrices. Then the circRNA-disease heterogeneous network is constructed based on the similiarity of circRNA-circRNA, disease-disease and the known circRNA-disease adjacency matric. Second, leveraging the three neural network modules-ResGatedGraphConv, GAT and MFConv-we gather node feature embeddings collected from the heterogeneous network. Subsequently, all the characteristics are supplied to the self-attention mechanism to predict new potential connections. Finally, our model obtains a remarkable AUC value of 0.9630 through five-fold cross-validation, surpassing the predictive performance of the other eight state-of-the-art models.

How to accurately identify influential nodes in complex networks remains a challenge due to the increasing network scale, complex topology, and dynamic network behaviors. Although a variety of approaches have been proposed, most researchers have only considered single or limited dimensions of nodes to varying degrees. In this article, a new centrality model [semi-local and global centrality (SLGC)] built on the semi-local and global structure of the node is proposed, which can capture a larger scope and richer information to evaluate how crucial a node is. First, the generalized energy is defined based on the generalized matrix, and then, the first-order and second-order generalized energy entropies are constructed by integrating information entropy and generalized energy to reflect semi-local influence (SLI). Second, the global influence (GI) is constructed based on the clustering coefficients of nodes and the distance between nodes, and finally, the total influence of nodes is derived from the above two aspects. The SLGC is contrasted with seven benchmark methods on nine real networks to assess the algorithm’s performance, and the data indicated that the SLGC has good effectiveness and versatility in monotonicity, resolution, accuracy, and top-10 nodes.