Current Bioinformatics - Online First

Considering the heterogeneity of proteins across diverse cell types and states, studying protein thermostability at the single-cell level enables a more profound comprehension of cellular function and the mechanisms underlying disease progression.

In this study, we constructed classification and regression models to predict the thermostability difference of homologous protein pairs by integrating implicit features extracted from protein sequences using eight language models, including ProtBERT, AminoBERT, and ProtT5-XL, with explicit sequence features that are manually computed.

Our results demonstrate that the fusion of explicit and implicit features significantly enhances prediction performance. In classification tasks, the combination of implicit features extracted by AminoBERT and the optimal explicit feature set achieves an accuracy of 87.1%. In regression tasks, the combination of implicit features extracted by Word2vec and the optimal explicit feature set yields a PCC of 0.864 and a R2 of 0.742, which is better than previously reported results.

This study reveals the complementary strengths of language models and handcrafted features in predicting protein thermostability. Combining both types of features significantly improves the performance of classification and regression models and helps identify key factors affecting protein stability. However, the study is limited by its reliance on existing datasets, which may reduce its ability to generalize to novel or rare protein families.

The integration of implicit and explicit sequence features enables a more comprehensive representation of protein sequences and facilitates the identification of factors influencing the thermostability of orthologous proteins.

CircRNA, with its covalently closed circular structure, plays key roles in biological functions and diseases by interacting with RNA-binding proteins (RBPs) and microRNAs (miRNAs). However, existing computational methods struggle to capture secondary structure features.

We introduce CSGN, a graph neural network model that predicts circRNA-RBP interactions using secondary structure information. CSGN enhances physicochemical feature encodings by incorporating pseudo-secondary structures from thermodynamic models and utilizes graph convolutional networks (GCNs) for feature extraction. It also integrates Doc2Vec embeddings and employs CNNs, BiGRUs, and MLPs for efficient feature representation.

CSGN outperforms existing models across 16 datasets. Ablation studies confirm the significance of RNA secondary structure and GCNs in improving prediction accuracy. Principal component analysis further highlights CSGN's strength in feature extraction.

CSGN advances circRNA-RBP prediction by integrating GCNs and Doc2Vec, though global structural constraints remain. Future work should address longer-sequence modeling and experimental validation.

CSGN effectively improves circRNA-RBP interaction prediction, demonstrating superior performance through the integration of RNA secondary structure and GCNs.

Microbes have increasingly become critical new drug targets in human health. However, the paucity of known microbe-drug association data hinders drug discovery. Predicting potential microbe-drug associations can complement traditional experiments and accelerate drug development, making it crucial to develop efficient computational methods.

We proposed GAALSMDA, a graph attention-based fusion network. First, a microbe-drug heterogeneous network and feature matrix were constructed by integrating multiple similarities of microbes and drugs. Graph Attention Network (GAT) was used to mine low-dimensional features of microbes and drugs. Then, dual attention mechanism (CBAM) and Bidirectional Long Short-Term Memory (BiLSTM) were applied to fuse local and global features. Finally, a classifier output the likelihood scores of associations.

The experimental results indicated that the AUC and AUPR evaluation indices of the model reached 0.9900±0.0011, 0.9958±0.0015 and 0.9492±0.0051, 0.9668±0.0042 in MDAD and aBiofilm datasets, respectively, and the prediction performance was significantly superior to that of existing prediction methods.

The outstanding performance highlights GAALSMDA's ability to process sparse data and integrate multi-source information, addressing the limitations of previous models in terms of insufficient feature fusion. However, the similarity calculations of GIP and HIP may introduce parameter uncertainty, which still needs further optimization.

Our model demonstrates effectiveness and reliability in accurately inferring potential microbe-drug associations.

Single-cell multi-omics technologies provide a comprehensive view of cellular states and transcriptional regulatory mechanisms by integrating diverse omics data. However, their complexity and heterogeneity present significant analytical challenges, particularly in understanding neurodegenerative disorders such as Alzheimer's disease (AD), an irreversible and progressive condition.

This study introduces the Multi-Omics Attention Graphical Convolutional Networks (MOAGCN), a novel multilayer deep learning model that addresses the heterogeneity in single-cell multi-omics data to enhance the analysis of multi-omics datasets and uncover potential mechanisms underlying AD. MOAGCN combines Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) to simultaneously capture local cellular connectivity and dynamically weight cell-to-cell interactions. The model was applied to AD-related single-cell RNA-seq and ATAC-seq datasets to identify significant gene expression and epigenetic alterations. It was further validated on datasets including DNA methylation, mRNA, and miRNA from other diseases. The model's performance was compared with conventional methods using metrics such as AUC, accuracy, and F1 scores.

MOAGCN effectively revealed key gene regulatory and protein interaction networks associated with AD, identifying significant changes in gene expression and epigenetic markers. In comparative validation across multiple datasets, MOAGCN outperformed traditional approaches in feature extraction and classification, achieving higher AUC, accuracy, and F1 scores. These results demonstrate its robustness in minimizing false positives and negatives while accurately identifying relevant features.

By testing in the classification of cell types and disease samples, MOAGCN achieved remarkable results, showing that its performance outperformed eight leading algorithms in multi-omics data classification tasks. Further analysis of MOAGCN's accuracy revealed a 95% confidence interval for its performance, reinforcing the model's robustness and stability across different datasets.

MOAGCN presents a robust and adaptable framework for integrating single-cell multi-omics data, addressing the challenges of complexity and heterogeneity. Its application to AD datasets highlights its potential to uncover regulatory mechanisms and bio-signals, advancing our understanding of complex diseases. This innovative approach holds promise for broad applications in multi-omics data analysis, particularly in elucidating mechanisms underlying neurodegenerative disorders.

Associations of abnormally expressed miRNAs with disease development have long been investigated in the biomedical field. The association types are diverse and complex, including circulation type, epigenetic type, target type, and genetic type, as well as various unknown associations and possibly novel association types. However, most current studies focus on the yes/no binary prediction of miRNA–disease associations. Algorithms for multi-type prediction or novel-type discovery of these associations are less developed.

Graph convolution and attention mechanisms, integrated within a deep learning framework, form the basis of deepMDpred. In the first step, deepMDpred employs the ViennaRNA tool to derive sequence and functional features of miRNAs by calculating base pairs, minimum free energy, and other relevant properties. In the second step, disease features are extracted using a Graph Convolutional Network (GCN) combined with attention learning, enabling the adaptive capture of the importance of different node features. Finally, a nonlinear fully connected layer (NFCL) is applied to generate the embedding vectors for both diseases and miRNAs.

In five-fold cross-validation, the model achieved high predictive performance for multi-type miRNA–disease associations. For task 1, the average AUC across the four predicted types exceeded 85%, with the genetics type achieving an accuracy of 0.919. For tasks 2 and 3, the average AUC exceeded 80%, and for the un-association type, the AUC reached 0.894. Validation using the HMDD v2.0 and HMDD v3.2 databases confirmed the robustness of the model, while additional case studies with the HMDD v3.2 and HMDD v4.0 databases demonstrated its applicability. Furthermore, investigations in breast and liver cancers supported the method’s capability to identify novel miRNA–disease associations.

The findings of this study demonstrate the potential of DeepMDpred as a novel and effective approach for predicting multi-type associations between miRNAs and diseases. Validation across multiple databases, along with successful application in case studies on breast and liver cancers, underscores the generalizability and practical utility of this approach. The framework also offers a pathway for identifying novel associations, which may accelerate the discovery of biomarkers and therapeutic targets in complex diseases such as cancer. Nonetheless, certain limitations remain. Although the model achieves strong performance on curated datasets, its robustness in real-world noisy datasets and its applicability to rare diseases require further investigation. Future research should also consider integrating additional data modalities, including epigenetic modifications and clinical phenotypes, to improve predictive accuracy further and broaden the scope of application.

DeepMDpred is an effective method that combines graph convolution and attention learning for the multi-type prediction of miRNA-disease associations. It provides a better ability to identify new association types between diseases and miRNAs, as well as broader applicability to unveil associated miRNAs with new diseases.

Genetic correlation plays a pivotal role in elucidating the shared genetic architecture underlying complex traits and diseases. Local genetic correlation can efficiently pinpoint specific genomic regions, thereby enhancing the precision of gene correlation analysis. However, accurate estimation of local genetic correlations remains challenging owing to linkage disequilibrium in local genomic regions and overlapping study samples.

In this work, we propose a novel method called LO-HDL that is based on high-dimensional maximum likelihood estimation. LO-HDL constructs marginal statistics using the summary statistics of GWAS and combines the 1000 Genomes Project Phase 3 data as a reference panel.

To assess the statistical power of LO-HDL, we performed a comparative analysis of LO-HDL with other local genetic correlation estimation methods on simulations with three different degrees of sample overlap. In the case of the absence of sample overlaps, the LO-HDL method improves the statistical power for cases with high local genetic covariance. In the case of partial sample overlap and complete sample overlap, LO-HDL demonstrates an overall improvement in statistical power. As an application, we used LO-HDL to estimate local genetic associations between the four autoimmune disorders. We found that LO-HDL could identify 31 regions with significant associations.

The LO-HDL method can identify genes or genomic regions that jointly influence multiple complex traits, thereby revealing the shared genetic architecture among traits. This approach elucidates the genetic relationships between traits and provides a basis for interpreting their associations. In simulated data, when the local genetic covariance ranges between (0.002–0.004), the statistical power of LO-HDL is slightly lower than that of previous methods. However, LO-HDL demonstrates superior performance in study scenarios with partial or complete sample overlap, as well as in real GWAS data analyses. Through LO-HDL, researchers can more accurately pinpoint genetically correlated regions among diseases. For instance, the TRIM27 gene on chromosome 6 exhibits significant associations with four diseases and may serve as a potential therapeutic target in future treatments.

LO-HDL is a novel method for estimating local genetic correlations, which is based on high-dimensional maximum likelihood estimation. Through its application in simulated datasets and four autoimmune diseases, LO-HDL improves the accuracy of estimating local genetic correlations, which has applicability for revealing relationships between genetic variants and specific traits or diseases.

The automatic classification of heart sound signals offers an economical and convenient approach for early diagnosis of cardiac diseases. By leveraging technological advancements, this method facilitates early detection and management of heart conditions, which is critical for improving patient outcomes.

To address the challenges in analyzing complex heart sound signals, we introduce a novel method utilizing EasyMKL-enhanced kernel partial least squares (KPLS). This approach begins with transforming segmented cardiac cycles into the time-frequency domain using the short-time Fourier transform (STFT). The STFT representations are then mapped into a high-dimensional feature space using multiple kernel functions derived from Easy MKL, designed to capture and enhance the discriminative nonlinear relationships among various heart sound categories. The extracted features are classified using a Support Vector Machine (SVM) for datasets with balanced samples and an XGBoost classifier for those with imbalanced samples.

The proposed method was evaluated on two publicly available heart sound datasets, the PhysioNet/CinC Challenge 2016 and the Yaseen dataset. On the PhysioNet/CinC Challenge 2016 dataset, our method achieved a sensitivity of 0.9217, a specificity of 0.8950, and an overall score of 0.9084. On the Yaseen dataset, our method achieved an average recall of 0.9933, precision of 0.9930, and F1-score of 0.9930, demonstrating high classification accuracy across different heart sound categories. These results confirm the effectiveness of our approach in extracting discriminative features and improving classification performance.

The high performance across two diverse datasets confirms the generalizability and robustness of the proposed method. Notably, the EasyMKL-enhanced KPLS framework captures complex nonlinear patterns while maintaining interpretability—an essential attribute for clinical applications. Compared to traditional approaches, our method significantly improves feature discriminability, as evidenced by ablation studies. While minor misclassifications persist in acoustically similar classes, the model consistently outperforms baselines, highlighting its strong potential for deployment in real-world intelligent auscultation systems.

The experimental results confirm the superiority of our proposed method, demonstrating its potential as a powerful tool for the automatic classification of heart sound signals. This approach not only enhances the accuracy of cardiac disease diagnostics but also offers a robust framework for handling complex and nonlinear characteristics of heart sound data, promising significant contributions to clinical practices and research in cardiology.

Previous studies have extensively reported various feature selection methods for identifying cancer signatures using RNA expression profiles. However, these methods often produce unreliable signatures due to four key factors. First, classifiers other than regression models are always inappropriately applied in prognostic survival analysis. Second, the unknown distribution of samples can lead to the ineffective selection of regression models. Third, high-dimensional expression profiles with small sample sizes typically result in poor predictive performance of the selected regression model. Fourth, variable control is usually overlooked.

To solve these problems, we have proposed a novel feature selection framework using ensemble regression to identify cancer prognostic signatures. This framework utilizes ensemble regression to overcome the limitations of classification models, as classification models reduce survival time to categorical labels, losing the original continuous information. At the same time, it incorporates up-sampling techniques to increase sample size and uses a bagging strategy to randomly select samples and features, addressing the challenges posed by high-dimensional data and small sample sizes. Additionally, the framework controls for clinical variables to ensure stable feature selection and reliable prediction results.

Experimental results demonstrate the effectiveness of this method in addressing the issues mentioned, providing reliable prognostic signatures. The ensemble regression method significantly improves predictive performance, with robust adaptability to unknown sample distributions.

The proposed ensemble regression model outperforms classification and single regressors in prognostic survival analysis by preserving continuous survival information, adapting to sample distribution, and benefiting from controlled variables. Using TCGA-GBM data, six prognostic miRNAs were validated as reliable biomarkers, whereas mRNA-based models showed limited robustness due to high dimensionality and small sample size.

The proposed feature selection framework offers a robust approach to improving the identification of cancer prognostic signatures, enhancing predictive accuracy in prognostic survival analysis.