Batch effects remain a major challenge in single-cell RNA-seq data, which often mask true biological variation across datasets. Most existing correction methods rely on pairwise mutual alignment or statistical integration, which can be inappropriate when batch sizes are imbalanced or when a consistent reference domain is desired. To address this problem, we propose scStarCorrect, a StarGAN-based adversarial model for reference-guided batch correction in single-cell data. The model includes a shared variational autoencoder conditioned on batch labels, a domain-aware discriminator that enforces alignment to a fixed reference batch. Unlike traditional mutual alignment methods that suffer from imbalance, scStarCorrect performs one-way batch correction by projecting all batches into the more accurate reference domain, enabling consistent and interpretable integration. We jointly optimize reconstruction loss, adversarial loss, domain classification loss, and variational regularization to learn batch-invariant embeddings. Experimental results on real single-cell datasets demonstrate that scStarCorrect effectively removes batch-specific variation while preserving cell type structure, offering a robust solution for multi-batch alignment and downstream analyses.

Predicting genes associated with diseases is of great significance for the early diagnosis and efficient treatment of neurodegenerative diseases. However, with the increasing complexity of omics data, it is difficult to integrate many aspects of relationships and effectively mine the genes associated with neurodegenerative diseases. To address this problem, we propose a novel multi-view graph embedding model, named TMGE, which can effectively integrate multiple networks to learn gene representations for prediction task of genes associated with neurodegenerative diseases. First, we design view-specific and view-shared encoders to extract view-specific content and capture consistent information across views; this enables the separation and preservation of diverse information within multi-view data for more accurate representations. Secondly, we introduce cross-linked encoders to enhance the integration of multi-view information, enabling the imputation of missing data across views by capturing cross-view dependencies. Furthermore, we employ adversarial learning to reduce the distribution discrepancy between diverse views at the global level, and then optimize the mean square error of local pairs with the help of the cross-linked encoder structure to achieve local alignment. We conduct experiments to verify the effectiveness of TMGE for predicting genes associated with neurodegenerative diseases using two disease datasets of Alzheimer and Huntington.

As an essential step in drug intervention discovery, predicting the drug-disease associations (DDAs) explores the potential therapeutic associations in given dugs and diseases. Since the various links in drugs and diseases contain high-order relations and complex therapeutic semantics, Graph Neural Networks (GNNs) have been introduced to DDA predictions and achieved great success. However, most previous approaches require the nodes of given drugs and diseases to have smooth attributes, which is difficult to meet in practical applications. Besides, GNN-based models suffer from the problem of semantic confusion for DDA prediction in heterogeneous graph. These challenges limit the model validity to discover therapeutic semantics in drug-disease networks. To address these challenges in DDA, we propose a novel graph neural network model called HSAGNN to augment node semantics hierarchically with three key steps by applying semantic-guided idea of SGNN method, including topological embedding learning, attribute completion, and semantic-guided aggregation. HSAGNN first learns the topological embedding and adopts the learned topological relationships to complete missing attributes with attention mechanism, which allows the node to contain richer information for neighbor aggregation. Then, the model aggregates the neighbor information with semantic-guided aggregation in both node and semantic levels. Here, HSAGNN injects the learned common knowledge as jumping knowledge to alleviate the semantic confusion. We evaluate the model in DDA tasks with various baselines and explore the model validity with extensive studies. The experimental results show that HSAGNN can discover the potential therapeutic associations by augmented semantics.

Due to feature redundancy in high dimensional data, the unsupervised feature selection methods for dimension reduction have attracted considerable attention. The current feature selection frameworks consider the global information of data, but ignore mutual screening of global and local information, and there have been no important breakthroughs on this approach research recently. We propose a novel feature selection method based on iterative optimization between the pseudo label matrix from spectral clustering and the local projection information (SNUFS), and then prove the convergence of the method. The pseudo label matrix and the local projection are designed in objective function for mutual screening and guiding the regularization feature selection by iterative approach. Our method selects features most relevant to the pseudo label and preserves the local structure of original data from feature selection matrix, where alternate iteration of different optimization items including pseudo label matrix achieve mutual screening. For this method, we give the objective function, iterative optimization fusion approach and convergence analysis in detail. Furthermore, we use K-Nearest Neighbor (KNN) and K-means to implement locality preserving projection and obtain two specific algorithms. Experiments on four real-world datasets in different fields demonstrate that our algorithms can effectively improve the accuracy of feature selection. In particular, our algorithm of KNN implementation is more effective and outperforms other major algorithms.

In recent years, a number of computational approaches have been proposed to effectively integrate multiple heterogeneous biological networks, and have shown impressive performance for inferring gene function. However, the previous methods do not fully represent the critical neighborhood relationship between genes during the feature learning process. Furthermore, it is difficult to accurately estimate the contributions of different views for multi-view integration. In this paper, we propose MGEGFP, a multi-view graph embedding method based on adaptive estimation with Graph Convolutional Network (GCN), to learn high-quality gene representations among multiple interaction networks for function prediction. First, we design a dual-channel GCN encoder to disentangle the view-specific information and the consensus pattern across diverse networks. By the aid of disentangled representations, we develop a multi-gate module to adaptively estimate the contributions of different views during each reconstruction process and make full use of the multiplexity advantages, where a diversity preservation constraint is designed to prevent the over-fitting problem. To validate the effectiveness of our model, we conduct experiments on networks from the STRING database for both yeast and human datasets, and compare the performance with seven state-of-the-art methods in five evaluation metrics. Moreover, the ablation study manifests the important contribution of the designed dual-channel encoder, multi-gate module and the diversity preservation constraint in MGEGFP. The experimental results confirm the superiority of our proposed method and suggest that MGEGFP can be a useful tool for gene function prediction.