Interdisciplinary Sciences: Computational Life Sciences最新文献

Emerging research continues to reveal the fundamental contributions of microbial communities to maintaining human physiological balance and advancing drug discovery. However, established wet-lab investigation techniques require significant time and resources. Contemporary research efforts have predominantly concentrated on establishing robust computational architectures to predict microbe-drug associations. Our research establishes a neural network architecture that synthesizes heterogeneous biological relationships with attentional factorization machines (HAFMMDA) to predict undiscovered microbe-drug linkages. The initial step involves assembling a heterogeneous network architecture integrating three key components: microbe similarity networks, drug similarity networks, and established microbe-drug interaction networks. HAFMMDA utilizes HIN2vec to extract feature representations of microbe-drug pairs. Finally, it combines second-order feature interactions and attention mechanism to perform comprehensive prediction. Five-fold cross-validation results confirmed excellent predictive performance with an AUC score of 0.9805, demonstrating statistically significant improvements over five contemporary baseline approaches. These findings corroborate HAFMMDA's effectiveness in uncovering verified drug-microorganism associations while simultaneously predicting innovative therapeutic-microbe relationships.

The emergence of novel viral diseases, with SARS-CoV-2 as a stark example, poses increasing threats to public health, causing significant global morbidity and mortality. Accurate identification and segmentation of viral imaging are crucial for tracking virus progression and mutations, and for devising new treatment strategies. Advanced virus recognition and segmentation models, utilizing high-performance networks like U-Net, have achieved notable success. However, these models struggle with multiple challenges, including limited labeled virus images, significant morphological variability, and indistinct boundaries. Consequently, this study introduces ViruSeg, based on the EVA-02 large language-image pre-trained model and data augmentation techniques, designed to efficiently perform virus segmentation tasks. Initially, the ViruSeg model employs data augmentation techniques like cutout and image fine-tuning to enrich electron microscope virus images, enhancing model generalization and effectively delineating virus boundaries and different forms. Secondly, ViruSeg utilizes the EVA-02 pre-trained model to learn a universal representation of virus images, enhancing adaptability to data scarcity. Finally, virus segmentation is conducted using the Cascade Mask R-CNN (CMR) model. Comprehensive evaluations on benchmark datasets demonstrate the superior performance of ViruSeg compared to advanced virus segmentation methods. We anticipate that the proposed solution will advance virology research and the development of treatments for related diseases. All dataset and code are available through https://github.com/xiachashuanghua/project .

 Multi-modal medical image registration aims to align images from different modalities to establish spatial correspondences. Although deep learning-based methods have shown great potential, the lack of explicit reference relations makes unsupervised multi-modal registration still a challenging task. In this paper, we propose a novel unsupervised dual-stream multi-modal registration framework (DSMR), which combines a dual-stream registration network with a refinement module. Unlike existing methods that treat multi-modal registration as a uni-modal problem using a translation network, DSMR leverages the moving, fixed and translated images to generate two deformation fields. Specifically, we first utilize a translation network to convert a moving image into a translated image similar to a fixed image. Then, we employ the dual-stream registration network to compute two deformation fields respectively: the initial deformation field generated from the fixed image and the moving image, and the translated deformation field generated from the translated image and the fixed image. The translated deformation field acts as a pseudo-ground truth to refine the initial deformation field and mitigate issues such as artificial features introduced by translation. Finally, we use the refinement module to enhance the deformation field by integrating registration errors and contextual information. Extensive experimental results show that our DSMR achieves exceptional performance, demonstrating its strong generalization in learning the spatial relationships between images from unsupervised modalities. The source code of this work is available at https://github.com/raylihaut/DSMR .

With the advantages of reducing biochemical experiments and enabling the rapid screening of potential druggable compounds, accurate computational methods are essential for predicting Drug-Target affinity (DTA). Current deep learning-based DTA prediction methods predominantly concentrate on single-modal information from drugs or targets. In this article, we propose a new multi-modal DTA prediction method, MGSDTA, to integrate graph features and sequence features of drug molecules and target proteins. We extract features from the drug molecular graphs and target protein graphs, meanwhile, we extract sequence features using continuous embeddings generated by advanced self-supervised pre-trained models, Mol2vec and ProtVec, for drug substructures and target subsequences respectively. Finally, they are integrated with a weighted fusion module for DTA prediction. Experiments on benchmark datasets indicate that the performance of MGSDTA exceeds single-modal methods based solely on sequences or graphs.

Identifying DNA binding sites remains a critical task in bioinformatics, with applications ranging from gene regulation studies to drug design. Although progress has been made in computational techniques, we still face challenges such as data complexity and prediction accuracy. In this paper, we introduce OptimDase, a new algorithm. It integrates feature encoding with optimum decision-making frameworks to improve DNA binding site prediction. OptimDase integrates multi-scale scanning and feature selection strategies, making it highly effective for both classification and regression tasks. Our experiments demonstrate that OptimDase achieves superior performance with an accuracy of 0.8943 in classification tasks and an RMSE of 0.0054 in regression tasks, outperforming existing algorithms in key evaluation metrics. These results highlight OptimDase's portability and robustness, making it an effective solution for identifying DNA binding sites and advancing the applications of drug design.

Recently increasing researches have discovered that circRNAs are remarkably reliable in organisms and play a crucial role as marker in many diseases. Although deep learning techniques has been universally applied to investigate the relationship of circRNA-disease, optimizing many parameters involved in these techniques for best performance has been a challenge. Therefore, we present, for the first time, a multi-objective particle swarm optimization algorithm to optimize the parameters in a graph attention network, ensuring that the model operates at peak efficiency. In addition, it also limits feature learning due to uneven distribution of different node types in heterogeneous graphs based on association relationships. We suggest a unique approach, MOPSOGAT, to overcome the aforementioned problems. MOPSOGAT is a method for predicting circRNA-disease associations utilizing the improved multi-objective particle swarm optimization (MOPSO) and the graph attention network. Initially, we obtain node sequences by utilizing multiple circRNA similarities and disease phenotypic similarities, and employing a heterogeneous graph with random walks incorporating jump and stay strategies. These sequences are then processed using word2vec to derive the neighbor vectors of the nodes, thus providing initial embeddings for circRNAs and diseases. Subsequently, in order to model convergence and diversity of the Pareto front solutions, an improved MOPSO algorithm is used to iteratively search for optimal solutions in the parameter space. After MOPSO optimization, parameters are fed into a graph attention network to further refine the model embedding. As a result, MOPSOGAT performs better than deep learning based methods, solely multi-objective optimization-based methods and machine learning-based ways. Moreover, the potential associations predicted by MOPSOGAT have been validated through case studies, further demonstrating the potential of MOPSOGAT in future biomedical research.

Backgrounds: During the development of new drugs, it is essential to assess their effectiveness and examine the potential mechanisms behind side effects. This process typically involves combining the analysis of drugs under development with relevant existing drugs to more precisely evaluate the effects of drugs and targets. The use of deep learning methods to analyze this problem is currently a research hotspot, but several limitations remain: (i) how to deepen the analysis from the molecular level to the atomic level and analyze the key substructures that affect interactions on the basis of pharmaceutical mechanisms; (ii) how to integrate biomedical analysis with deep learning methods to make it medically sound and enhance interpretability.

Methods: To address the limitations of existing research, based on Deep Graph Convolutional Network (Deep-GCN) and Bilinear Attention Network (BAN), we have constructed an interpretable deep learning framework, WDGBANDTI, to analyze and predict drug‒target interactions at the substructure level and enhance the prediction capability of the model with respect to unidentified target pairings by adding modules.

Results: For different application scenarios, we validated the model via several commonly used and highly covered datasets. We also selected several state-of-the-art computer methods as comparison objects, and our model demonstrates advantages in accuracy, sensitivity, specificity, and other deep learning features. More importantly, the model can identify the substructures that play a role in drug‒target interactions through BAN, highlighting its excellent interpretability.

Conclusion: In conclusion, we believe that our work will contribute to advancements in drug development and side effect experiments and provide meaningful guidance for drug design.

In the domain of Chinese clinical medical question-answering (QA), traditional Large Language Models (LLMs) encounter challenges such as hallucinations and difficulties in updating knowledge for knowledge-intensive tasks. To address these issues, this research presents a Chinese clinical medical QA model that integrates Retrieval-Augmented Generation (RAG) and a medical knowledge graph, named CMedRAGBot. First, a Chinese medical knowledge graph encompassing multiple entity types-including diseases, medications, and symptoms-is constructed. Based on this knowledge graph, a Named Entity Recognition (NER) model built on a Chinese-RoBERTa and BiGRU architecture is designed, with data augmentation strategies employed to enhance its generalization capability. In addition, prompt engineering techniques are used to implement intent recognition for user queries, mapping them to predefined intent categories. Finally, the aforementioned modules are integrated to form a complete Chinese clinical medical QA system. In the experimental evaluation, CMedRAGBot is deployed on five state-of-the-art LLMs (including ChatGPT-4o, ChatGPT-o1, DeepSeek-R1, Llama-3.3-70B-Instruct, and Gemini 2.0 Flash) and tested using specialized question banks derived from the Chinese Clinical Medical Qualification Examinations and Residency Standardization Training Examinations from 2000 to 2023. The results indicate that the integration of CMedRAGBot significantly improves the test accuracy of all models, with increases of up to approximately 10%. Furthermore, ablation experiments reveal that data augmentation enhances NER model's F1 score from 95.27% to 97.55%, while the inclusion of an intent recognition module markedly improves the model's ability to understand complex queries, thereby further boosting answer accuracy. Source code of the research is available at https://github.com/zhdongfang/CMedRAGBot .

Long non-coding RNAs (lncRNAs) play key regulatory roles in biological activities, making it crucial to accurately predict lncRNA-disease relationships for understanding disease pathophysiology and developing effective prevention and treatment strategies. Nevertheless, existing computational prediction methods are often limited by data sparsity and incompleteness, as well as the inadequate representation of node information. To mitigate these issues, a novel prediction model, termed DSGCNLDA, is proposed to enhance predictive performance. The proposed model integrates biological similarity features from multiple perspectives into a comprehensive similarity matrix using multi-view fusion learning. Utilizing this similarity matrix in conjunction with the adjacency matrix, a heterogeneous network of lncRNA-disease associations is constructed. Subsequently, feature extraction is conducted on the randomly masked heterogeneous network. During this process, a graph convolutional network (GCN) is employed as the encoder, and our proposed DualScope attention mechanism is introduced to more effectively capture complex topological relationships between nodes, thereby obtaining a comprehensive representation of the nodes. Finally, association prediction is made via a multi-layer perceptron (MLP). Experimental results on multiple public datasets show that DSGCNLDA performs strongly in lncRNA-disease association prediction. Ablation studies confirm its novelty, while case studies and generalization evaluations demonstrate its effectiveness in biomedical prediction tasks.

Copy number variation (CNV) is a major type of structural variation (SV) that plays critical roles in genetic diversity and disease. Currently, many CNV detection tools have been developed. Although each tool exhibits different advantages under specific scenarios, they still have disadvantages such as suboptimal sensitivity, imprecise breakpoint resolution, and reduced robustness in complex sequencing environments. Developing more effective CNV detection tools by building upon the strengths of existing tools presents a significant challenge in the field. To fully leverage the detection results of existing tools and improve the accuracy of CNV detection under complex sequencing conditions, a new method called SSLCNV (semi-supervised learning framework for CNV detection) is proposed. It combines consensus-based pseudo-labeling using density-based clustering. SSLCNV generates high-confidence pseudo-labels by intersecting CNV predictions from four representative tools (CNVkit, GROM-RD, Matchclips2, OTSUCNV) and uses these as core seeds for clustering. Additionally, SSLCNV introduces a new constraint z-score into the DBSCAN algorithm to enhance clustering accuracy. By leveraging the improved DBSCAN and incorporating reliable labels, SSLCNV effectively detects CNV from partially labeled and unlabeled data. Comprehensive evaluations on both simulated and real datasets demonstrate that SSLCNV consistently achieves superior F1-scores compared to existing tools across diverse sequencing depths and tumor purities. Importantly, it maintains robust performance under low-coverage conditions, yielding higher recall without a substantial loss in precision. SSLCNV offers a scalable and accurate solution for CNV detection, particularly advantageous in scenarios with complex genomic backgrounds.