Advanced Intelligent Computing Technology and Applications

For graph $$G$$ G , the first Zagreb index $${M}_{1}\left(G\right)$$ M 1 G and second Zagreb index $${M}_{2}\left(G\right)$$ M 2 G are defined as: $${M}_{1}\left(G\right)={\sum }_{u\in V\left(G\right)}{d}_{G}^{2}\left(u\right)$$ M 1 G = ∑ u ∈ V G d G 2 u and $${M}_{2}\left(G\right)={\sum }_{u\upnu \in E\left(G\right)}{d}_{G}\left(u\right){d}_{G}\left(\upnu \right)$$ M 2 G = ∑ u ν ∈ E G d G u d G ν , where $${d}_{G}\left(v\right)$$ d G v denotes the degree of vertex $$v$$ v in $$G$$ G . The Hyper-Zagreb index $$HM\left(G\right)$$ H M G is defined as: $$HM\left(G\right)={\sum }_{u\nu \in E\left(G\right)}{\left[{d}_{G}\left(u\right)+{d}_{G}\left(\nu \right)\right]}^{2}$$ H M G = ∑ u ν ∈ E G d G u + d G ν 2 . In this paper, we introduce a novel class of F-multiplicity corona graphs and derive explicit formulas for their $${M}_{1}\left(G\right)$$ M 1 G , $${M}_{2}\left(G\right)$$ M 2 G and $$HM\left(G\right)$$ H M G . Additionally, we demonstrate the chemical relevance of these graphs through applications in molecular design, highlighting their potential for modeling complex chemical structures.

The computational prediction of mass spectra is a vital approach for identifying small-molecule structures. Existing methods have adopted approaches such as rule-based frameworks and deep learning models to predict mass spectra from molecular structures. However, due to the complex structure of molecules and the intricate fragmentation patterns in mass spectra, it is still challenging to effectively utilize the local and global features for modeling molecular fragmentation. The local features of molecules are vital for capturing the fine-grained structural details, while the global features are essential for understanding the complex interactions between substructures. In this work, we propose GraphMS, a novel model that integrates local and global graph features for mass spectral prediction. We first decompose the molecular graph into chemically meaningful substructures for extracting local features. A Transformer module is then applied to capture global features and model long-range dependencies between super nodes. Finally, the local and global features are fused adaptively to create a comprehensive molecular representation for mass spectrum prediction. Experiments on multiple datasets demonstrate that GraphMS outperforms existing methods, achieving higher accuracy and better generalization in mass spectrum prediction.

Crystals are essential constituents of a wide range of materials, encompassing both advanced technologies and everyday applications. Recently, deep learning-based methods for crystal property prediction have demonstrated remarkable performance, greatly facilitating the discovery of novel materials. However, these approaches typically concentrate on atom-wise interactions and often fail to account for periodicity, a fundamental characteristic of crystals. To address this limitation, we propose a novel plug-and-play component, the Periodicity Injection Module (PIM), which seamlessly incorporates periodicity into existing crystal models. Specifically, the PIM employs crystal-wise attention to ensure that the surroundings of unit cells at periodic distances remain identical, aligning with the definition of crystal periodicity. By capturing interactions among bases at the crystal level, the PIM complements and enhances the modeling of intra-crystal interactions. Extensive experiments on benchmark datasets demonstrate that our PIM significantly improves crystal property prediction.

Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder worldwide. Speech-based diagnostic approaches for PD have attracted increasing attention, with deep learning models demonstrating promising performance. In this paper, we propose a speech-based diagnostic model for PD, aiming to enhance the diagnostic accuracy using deep learning techniques. We adopt the SMOTE–ENN oversampling method to solve the data imbalance problem, and develop a hybrid model that integrates a Convolutional Neural Network (CNN) and Bi-directional Long and Short-Term Memory network (BiLSTM) to efficiently extract the speech features and capture temporal dependencies. Experimental results show that the proposed model achieves an accuracy of 95% on public datasets and outperforms traditional machine learning and other deep learning models in several evaluation metrics, validating the effectiveness of our network in Parkinson's disease diagnosis. These results validate the effectiveness of our approach and highlight its potential for high-precision early screening of PD, offering reliable technical support for clinical applications.

Spatial transcriptome RNA sequencing (stRNA-seq) can characterize the gene expression patterns of samples while preserving the original spatial position of tissue, thereby facilitating the understanding of tissue's biological characteristics and disease's pathogenesis. However, unlike single-cell RNA sequencing (scRNA-seq), stRNA-seq data of bulk samples do not have single-cell resolution and only measure the averaged gene expression of mixed cells. Therefore, it is necessary to perform deconvolution on the spatial transcriptomic data to infer the relative abundance of cell types and/or cell-type specific gene expression profiles (GEPs). Recently, deconvolution methods have been developed, while many of them cannot simultaneously estimate the cell type abundance at each spot and cell type-specific GEPs and may not utilize spatial localization information fully. Here, we propose a deconvolution method for spatial transcriptomic data, SpatialDSSC, to simultaneously estimate cell type-specific GEPs and cell type abundance. SpatialDSSC models on sample-sample and gene-gene similarities from gene expression and sample-sample similarity from spatial position and leverages scRNA-seq data or provided cell-type specific GEPs as references. By testing on multiple sets of simulated and real data and comparing with the existing deconvolution methods, we demonstrate the effectiveness and accuracy of SpatialDSSC in inferring cell type-specific GEPs and cell type abundance.

Cuproptosis, a novel type of cell death. However, the prognostic of cuproptosis-related genes (CRGs) in SKCM remains unexplored. In the study, through univariate Cox, Lasso, and multivariate Cox regression analyses, 8 CRGs were selected to build the prognostic prediction model. The area under the curve for the prediction of 5-year overall survival in the training dataset is 0.825, while in the test dataset, it is 0.75. Besides, low risk group exhibits a high abundance of immune cells. In conclusion, CRGs could effectively predict the prognosis of SKCM patients, and may provide novel insights into the cancer treatment.

Phylogenetic tree construction serves as a fundamental technique for elucidating evolutionary branching patterns and a core methodology for inferring species relationships. However, existing phylogenetic inference tools have inherent limitations: traditional alignment-based methods exhibit exponential growth in computational resource consumption as genomic data scales, while alignment-free approaches often rely on overly simplified assumptions, restricting their practical applicability. The trade-off between accuracy and efficiency remains a fundamental bottleneck in modern phylogenetic research. To address this challenge, we propose CS-Phylo, an innovative evolutionary distance estimation method that integrates closed-syncmer substrings with the MinHash algorithm, and leverages GPU computing to enhance computational efficiency. CS-Phylo employs a mathematically formalized anchoring mechanism to effectively select sequence features based on Closed-Syncmers, transforming long genomic sequences into compact CS-Sketch for efficient evolutionary distance matrix computation. Experimental results demonstrate that CS-Phylo exhibits significant advantages over existing tools in phylogenetic tree construction, achieving higher accuracy while maintaining computational efficiency. Additionally, by utilizing GPU acceleration, CS-Phylo effectively handles datasets of varying scales, making it a robust solution for large-scale phylogenetic analyses.

Spatial transcriptomics (ST) reveals the intricate dynamics of cell regulation and gene expression but is often limited by high costs. Leveraging artificial intelligence to predict spatial gene expression from histological images provides a more economical alternative. However, current approaches struggle to fully capture deep-level information embedded in histological images. In this study, we introduce DCGT, a deep learning method that combines Dynamic Convolutional networks and Graph Transformer to dynamically uncover potential molecular patterns within histological images. This approach bridges a connection between detailed histological image features and spatially resolved gene expression. Extensive evaluations using four different spatial transcriptomics datasets highlight the remarkable efficiency of DCGT for predicting spatial gene expression. Furthermore, DCGT effectively uncovers spatial contexts and accumulation markers within defined tissue regions.

Spatial domain identification plays a critical role in elucidating tissue function and intercellular dynamics by capturing cell type diversity, gene expression heterogeneity, and cell-cell interactions within their spatial context. The functionality of complex tissues is intrinsically linked to the spatial arrangement of distinct cell types, and recent advancements in spatial transcriptomics (ST) have been pivotal in uncovering this relationship. However, current methods often struggle with accuracy and computational efficiency when processing high-dimensional, complex datasets. To overcome these limitations, we propose the application of graph attention autoencoder for spatial domain identification. The graph attention mechanism enables the model to effectively capture both local and global gene expression dependencies, thereby enhancing the precision of spatial domain delineation. The results from experiments we conducted show that our method outperforms traditional methods across multiple datasets, offering a robust approach for spatial domain identification.

This paper proposes a multi-modal model, namely MMF2Drug, fusing the sequences and structures of proteins, physicochemical properties and structures of molecules, to generate high-quality molecules. MMF2Drug includes a Feature Extraction (FE) module, a Multi-Modal Fusion (MMF) module and a Conditional Molecule Generation (CMG) module. FE module is used to extract the different modal embeddings from proteins and molecules. MMF module aims to capture the consistency among various modalities by Iterative Multi-scale Channel Attention (IMsCA), which can assist the model in identifying key feature channels that share commonalities across different modalities, enhancing the alignment of features. CMG module combined with Generative Adversarial Network (GAN) can significantly improve capabilities in exploring molecular chemical space, generating molecules with novel structures. It uses the protein features as constraints to more accurately generate molecules for specific proteins. Experiments show that MMF2Drug has better performance for generating molecules. Case studies show that targeted drug molecules generated by MMF2Drug exhibit strong affinity for two key targets (KRAS and EGFR) in pancreatic cancer.

DTI prediction serves as a vital function in drug discovery and repositioning. In this paper, we propose a novel method, DFDGRU-DTI, which employs a random-walk–based embedding approach combined with a BiGRU neural network augmented by multi-head attention to predict drug–target interactions. First, the random walk embedding model is used to generate feature vectors for drugs and targets, capturing the semantic relationships between them. Then, a bidirectional GRU network is employed for sequence learning, and the inclusion of multi-head attention helps the model concentrate on the most relevant input features. The evaluation demonstrates that the introduced model delivers superior performance relative to numerous advanced drug–target interaction prediction techniques. This model exhibits significant application potential, providing valuable support for real-world drug development.

Effective segmentation of intracranial hemorrhage (ICH) is key to early diagnosis and treatment of stroke. Traditional segmentation methods rely on manual intervention, which is inefficient and susceptible to subjective factors, while existing deep learning models have limitations in handling ICH lesions of different shapes and sizes. Denoising Diffusion Probabilistic Models (DDPM) have demonstrated effectiveness across various vision tasks, such as image deblurring, super-resolution, and anomaly detection. Motivated by DDPM’s success, this paper introduces a new network, RDT-Net, built upon DDPM. We combined the model with the Multi-Scale Adaptive Rotated Convolution (MARC) module, which improves segmentation accuracy by dynamically adjusting kernel weights, eliminating redundant features, and integrating multiple features and angle information derived from the feature maps. In addition, RDT-Net ensures the uniqueness of stable generated masks through a dynamic conditional encoder with a deep stack transformer (DST), thereby enhancing segmentation accuracy. Experimental results show that RDT-Net outperforms other models significantly, achieving superior results in terms of the Dice coefficient, Jaccard index, and 95% Hausdorff distance, providing an effective and dependable solution for the complex task of ICH segmentation.

The interpretability of Magnetoencephalography (MEG) decoding models is crucial for advancing their applications. While current research predominantly focuses on interpreting individual models, systematic investigations into cross-model explanation comparison remain scarce, hindering advancements in both understanding neural mechanisms and optimizing model performance. This paper introduces a novel explanation comparison framework. First, we propose a joint feature attribution algorithm to reliably compute explanations across different models. Next, we quantify the similarity of explanations between models, based on within- and cross-sample relation metrics. Empirical evaluations on two MEG datasets reveal three key findings: (1) our joint attribution method effectively reduces explanation comparison errors; (2) the explanation similarity between different models correlates with their decoding performance; and (3) leveraging consensus features to refine underperforming models boosts classification accuracy by up to 4.37%, even surpassing original state-of-the-art models in specific scenarios. These results demonstrate that explanation comparison not only deepens our understanding of the neurophysiological knowledge derived from MEG, but also provides novel insights for improving these models.

Single-cell RNA sequencing (scRNA-seq) is an advanced technology used to study cellular heterogeneity and diversity. It allows gene expression analysis at the single-cell level, but it also confronts several challenges, including high noise levels, high dimensionality, and high sparsity of data. These characteristics increase the difficulty of effectively clustering cells from scRNA-seq data. Therefore, we propose a new deep clustering approach, scAFC, specially designed for scRNA-seq data. This method combines an Autoencoder (AE) and a Graph Attention Network (GAT), where the AE is responsible for extracting key features from the data, and the GAT utilizes the relationships between cells to enhance feature expression. Through an innovative adaptive fusion module, we integrate the outputs of the AE and GAT at different levels and use an attention mechanism to dynamically adjust the contribution of different features. Furthermore, scAFC employs a self-supervised learning optimization module to refine the computation of clustering centers, effectively enhancing the clustering accuracy and the generalization ability of the model. Extensive testing on 12 public scRNA-seq datasets shows that the scAFC method not only improves clustering accuracy but also demonstrates superior performance in handling the complexity of single-cell data, outperforming some of the state-of-the-art scRNA sequencing clustering methods. Code: https://github.com/728267035/scAFC.git .

Drug-Drug Interactions (DDIs) prediction is crucial for drug safety and clinical decision-making. While existing computational methods have made progress in DDI prediction, they often fail to fully integrate multiple drug features, particularly the biological relevance from protein sequences. In this paper, we propose BIOFUSE-DDI, a novel dual-source transformer framework that effectively combines knowledge graph information with protein sequence features for DDI prediction. Our model incorporates three key components: a Dual-Source Attention Mechanism for feature extraction, a Bimodal Interaction Enhancement module for feature fusion, and a Sequential Transformer Refinement for comprehensive interaction pattern learning. Extensive experiments on both binary-class and multi-class DDI prediction tasks demonstrate the superiority of our approach. For binary classification, BIOFUSE-DDI achieves 98.20% F1 score and 99.80% AUC, surpassing state-of-the-art methods by significant margins. In the more challenging multi-class prediction task, our model attains 96.01% macro-F1 and 97.55% mean-accuracy, establishing new benchmarks in DDI prediction. These results validate the effectiveness of integrating protein sequence features with knowledge graph information through our hierarchical fusion architecture.

To protect patient privacy, the ability of diffusion models to generate medical images from noise has become a key focus for enriching datasets. However, due to the high precision required for medical image anatomical structures, generative models designed for natural scenes fail to meet the stringent standards for medical logic. We propose MedMaskDiff, a Mamba-based semantic image synthesis model that generates medical images from masks, which takes full advantage of Mamba's capability to capture long-range medical features. Additionally, it utilizes an evolutionary condition-guide method to enhance the quality and medical logic of the generated target regions. MedMaskDiff outperforms other advanced methods in synthesizing liver CT, thyroid nodule ultrasound, and low-grade intraepithelial neoplasia microscopy images. By utilizing masks from other liver CT datasets for semantic synthesis and data augmentation, comparative experiments demonstrate that MedMaskDiff effectively safeguards patient privacy while enhancing downstream medical image segmentation tasks, significantly improving the performance of segmentation models. Our code is available on https://github.com/Jiacheng-Han/MedMaskDiff .

Due to the different clarity of the cells in different layers of the cellular images acquired by professional CT equipment, it is necessary to consume a large amount of manpower, material, and time costs to organize and archive the acquired data in case of too many layers, which seriously affects the work efficiency of the relevant workers. Currently, there is no mature solution for segmenting and sampling cell images from different CT layers and combining them into clearer images without changing the original data. In order to address the above problems, we propose a feasible and efficient solution for the whole process of combining CT images. Firstly, we propose and complete three feasible sampling methods for CT images, and use the four-channel representation method for relational retrieval and classification. At the same time, we use objective quantitative indexes to objectively select the cells with the highest clarity and combine them at a very low time cost. In our experiments, 20 layers of CT images with a resolution of 9391 × 9391 are sampled simultaneously, and the three proposed schemes can improve the average gradient by up to 148.0%, the variance by up to 110.0%, and the NIQE by up to 50.8% compared with the original images, and the total time consumed can be controlled within 68.9 s.

Piwi-interacting RNA (piRNA) is a key biomarker for complex disease diagnosis and prediction. Predicting piRNA-disease associations (PDA) is crucial for revealing their genetic mechanisms. In this study, a method PDA-PAGCN based on proxy attention graph convolutional network for predicting PDA. Firstly, a heterogeneous network was constructed based on the similarity and association information of piRNA and disease, which is then input into a graph convolutional network, and the feature dimensions are aligned through the group feature transformation module to obtain initial features. Subsequently, the Topk graph pooling method was employed to obtain feature subgraphs from these initial features. Finally, we fuse these feature subgraphs with the initial features using a proxy attention mechanism and calculate cosine similarity association scores to derive the final PDA reconstruction scores. The predictive performance of PDA-PAGCN is validated through five-fold cross-validation experiments, achieving an AUC of 0.9667 and an ACC of 0.9707. Case studies on two human diseases further confirm the reliability of PDA-PAGCN in practical applications. Therefore, PDA-PAGCN is proved to be effective in predicting hidden PDA.

Anticancer peptides (ACPs), which are able to specifically target and kill cancer cells, are promising cancer therapeutics. However, it requires significant time and cost to identify ACPs through biological experiments. To facilitate the ACP screening process, we propose the ACP prediction method CALM-AcPEP, a deep learning framework based on the ACmix module, Evolutionary Scale Modeling 2 (ESM2) and cross-attention. The ACmix module combines a convolution neural network and self-attention to recognize the original sequence representation, while the pre-trained ESM2 efficiently captures the evolutionary information of the peptide sequence. Then, the relationship between the original sequence and the evolutionary information is learned by the cross-attention mechanism, strengthening the representation of ACPs. The results of our study show that our proposed method is promising for the prediction of ACPs.

The identification of anticancer peptides (ACPs) has emerged as a critical research area due to their potential to revolutionize cancer treatment, offering more targeted and less toxic alternatives to traditional therapies. Most existing deep-learning models mainly focus on primary sequence information and basic physicochemical properties. They frequently neglect long-range dependencies and positional relationships within peptide sequences. In this study, we introduce ACP-TransLSTM, a novel framework for anticancer peptide prediction that employs a comprehensive feature extraction strategy, integrating amino acid composition, structural and physicochemical properties. By combining handcrafted descriptors with Transformer-based representations, the model better captures the diverse characteristics of peptides. Its architecture combines Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks, Bi-LSTM can process sequential data bidirectionally, integrating information from both past and future in the peptide sequence. We tested ACP-TransLSTM on six public datasets and the experimental results demonstrate that ACP-TransLSTM exhibits strong robustness and performs more competitive than compared methods, particularly in terms of AUC and ACC. For example, in terms of the average ACC and AUC across the six datasets, ACP-TransLSTM achieves 0.907 and 0.93 respectively, outperforming the best-performing compared method by at least 2.7% in average ACC and 1.6% in average AUC. The source code for ACP-TransLSTM is available at https://github.com/qdu-bioinfo/ACP-TransLSTM .

Due to the toxicity of traditional treatments and the emergence of resistance in targeted therapies, cancer therapies confront significant challenges. Synthetic lethality (SL) presents a promising avenue for precision medicine by selectively inhibiting genes that partner with tumor-specific mutations. Traditional SL prediction methods are constrained by binary interaction assumptions, making it challenging to capture complex gene-gene interactions and address the imbalance inherent in multimodal data. This study introduces the multimodal generative adversarial network integrating hypergraph and knowledge graph representations for SL prediction to overcome this limitation. This model innovatively incorporates the hypergraph to represent SL relationships and leverages graph neural networks to character higher-order gene interactions through hyperedge, thereby overcoming the limitations of traditional binary interaction assumptions. Additionally, a dynamic weight allocation mechanism is designed, utilizing attention networks to quantify the heterogeneous contributions of multimodal data, addressing the challenges posed by data imbalance. Finally, an adversarial training approach is employed, where the generator dynamically produces negative samples to mitigate issues related to sample scarcity and annotation noise. Experiments conducted on the SynLethDB 2.0 dataset demonstrate that MGHK4SL significantly outperforms existing methods based on traditional machine learning and graph neural networks in multimodal integration and higher-order relationship mining. This advancement offers interpretable biomarker networks for precision cancer therapy. Our code has been released at https://github.com/wyl20181914/MGHK4SL .

Magnetic Resonance Imaging (MRI)-derived brain networks are significant for detecting neurological disorders. Nevertheless, traditional node-centric methods usually focus excessively on capturing features on individual nodes (i.e., brain regions) but ignore the critical information of dynamic evolution between the edges (i.e., connections between brain regions), resulting in suboptimal identification performance. Regarding this, two key challenges are identified: 1) oversimplified edge dynamics cannot adequately capture neural information transmission, and 2) neglecting higher-order edge causal effects and structural asymmetry leads to insufficient analysis of the collaborative patterns across multiple brain regions. To overcome these challenges, this paper introduces a novel edge-centric fusion network (EdgeViewDet) for neurological disorder detection. First, based on conditional Granger causality analysis, an edge-centric effective connectivity network (EC-ECN) is designed to investigate high-order time-varying causal effects within sliding windows. Then, an edge-centric structural connectivity network (EC-SCN) based on graph diffusion is employed to capture structural lateralization abnormalities. Additionally, to efficiently fuse multiple structural-causal features, a directed spatiotemporal feature extraction module (DST-FE) is designed, which helps improve the feature discrimination by considering the underlying relations among structural-causal features in different edges. The superior disease detection performance of EdgeViewDet is validated through extensive experiments on two public datasets and one private dataset. The code of our work will be released upon acceptance.

Molecular graph representation learning has made important contributions in the field of drug discovery and design. Most of the existing works leverages graph neural networks (GNNs) as a backbone for encoding implicit molecular representations, which are honed through various self-supervised learning (SSL) pretext tasks. However, vanilla GNN encoders ignore the implicit chemical structure information and functions in molecular motifs, and existing works do not adequately capture various modalities such as attributes, semantics, and structures in molecules and motifs. Combining the various information mentioned above is still challenging. To address the above issues, we introduce property-aware motifs to both node-level and graph-level pre-training tasks. First, we propose a new node-level pre-training task MotifMask (MOM) which masks molecules by atom type to capture semantics in the motif range, and alleviates the negative migration problem of AttrMask. Further, we propose Three-Layer Augmented Graph Contrastive Learning (TECL) to comprehensively capture the structural information of different layers of the molecule. Finally, the complementary strengths of the semantic node-level task and the structural graph-level task are combined to form the MOMTERL framework. MOMTERL, through its design of local-global tasks, demonstrates better performance in molecular property prediction by effectively integrating different information.

Respiratory sound recognition is a challenging task. This paper presents an automatic classification model that combines convolutional neural network and audio spectrogram transformer. It utilizes respiratory sound recordings from 418 patients with respiratory diseases. Each respiratory cycle is transformed into a Mel-spectrogram via Fourier transform, which serves as input to the model. Our methodology implements a dual-perspective parallel extraction approach that identifies comprehensive and specific acoustic patterns. These representations merge through contrastive learning techniques and an adaptive fusion mechanism to enhance diagnostic precision. Performance evaluation reveals the framework attains accuracy rates of 69.50% and 92.01% in binary classification (normal/abnormal) on ICBHI 2017 and SPRSound datasets. For multi-class categorization, it reaches 63.27% and 90.63% respectively, surpassing contemporary methodologies. These findings validate the efficacy of our integrated approach for computerized respiratory acoustic analysis.

Accurately predicting drug-protein interactions (DPIs) is critical in computational biomedicine, as it facilitates the understanding of molecular pharmacodynamics and accelerates drug discovery. Existing prediction methods fall into two main categories, atomic-scale structure-based approaches, which exhibit strong cross-domain generalization but limited within-domain accuracy, and molecular-scale network-based methods, which perform well in within-domain scenarios but struggle with cross-domain generalization. A key challenge lies in effectively integrating multi-scale information to overcome the limitations of single-scale models. To address this, we propose MGDPI, a multi-scale representation learning framework enhanced by graph regularization. MGDPI introduces a graph regularization loss function based on network information to guide the learning of drug and protein representations from structural data. By seamlessly integrating multi-scale information, MGDPI significantly improves both generalization and predictive performance. Experimental results demonstrate that MGDPI outperforms state-of-the-art models in within-domain prediction accuracy under various imbalance scenarios on the BioSNAP dataset. Furthermore, MGDPI achieves superior performance across three independent cross-domain datasets, highlighting its exceptional generalizability and predictive power. These findings underscore MGDPI’s potential not only as a robust DPI prediction tool but also as an effective solution to data imbalance and cross-domain challenges.

Lung cancer histopathological diagnosis faces considerable challenges due to equipment-induced variations and staining inconsistencies across institutions. Although Weakly Supervised Object Localization (WSOL) and Source-Free Domain Adaptation (SFDA) methodologies have been investigated to address these issues, existing approaches lack sufficient optimization for Whole Slide Imaging (WSI), resulting in suboptimal performance. Faced with these limitations, we propose a novel SFDA framework tailored for WSOL in histopathology. Specifically, we make the following contributions: (1) We systematically compare SFDA approaches on multiple public histopathology datasets under WSOL settings, and demonstrate that generative-based black-box methods achieve the best performance in cross-domain localization and classification tasks. (2) We design a diffusion-driven training framework that integrates Denoising Diffusion Probabilistic Models with hybrid Vision Transformer architectures, effectively preserving semantic fidelity while adapting to domain shifts. (3) Our model establishes state-of-the-art performance across six benchmark datasets, highlighting its superior generalizability and robustness in real-world multi-center clinical scenarios. The experimental evaluation demonstrates that our method achieves highest precision in the optimal target domain configuration, representing a substantial 30% improvement over previous SFDA approaches and providing a more robust AI-assisted tool for advanced histopathological examination.

MicroRNAs (miRNAs), as central regulators of gene expression, have been demonstrated to be deeply involved in the pathogenic processes of various diseases. In current clinical practice, modulating microRNA expression through pharmacological intervention has become an important therapeutic approach for various diseases. However, the emergence of miRNA drug resistance during treatment can significantly compromise therapeutic efficacy. Therefore, accurate prediction of miRNA drug resistance not only provides a basis for developing personalized treatment regimens in clinical practice, but also effectively enhances disease treatment outcomes. The inherent complexity of miRNA-target interaction networks and the multifactorial nature of drug resistance mechanisms pose substantial challenges for conventional experimental approaches, which are often limited by high costs and low throughput. Fortunately, the rapid advancement of artificial intelligence technologies in recent years has opened new avenues to address these challenges through computational approaches based on machine learning and deep learning algorithms. In this paper, we propose a dual-channel model based on feature alignment, DCMFA, for miRNA drug resistance prediction. DCMFA enhances its data comprehension and analytical capabilities by integrating multimodal information to comprehensively capture enriched features between nodes, thereby improving its adaptability and generalization performance. Additionally, DCMFA employs a modular learning framework with two independent modules dedicated to processing distinct node feature groups. This architecture effectively prevents noise interference from irrelevant feature interactions while enabling each module to capture latent patterns within specific feature subsets, thereby facilitating cross-type feature alignment. Experimental results demonstrate that through five-fold cross-validation, DCMFA achieved impressive performance metrics: AUC (95.40%), ACC (91.32%), F1 (91.29%), Precision (91.57%), and AUPR (94.60%), outperforming state-of-the-art models by 1.19%, 3.51%, 3.48%, 3.20%, and 0.32% respectively.

Effective representation of whole slide images (WSIs) is essential for survival prediction tasks. Previous studies have primarily focused on multimodal approaches, exploring complex fusion techniques to integrate information from different modalities. However, these methods are met with several challenges: (1) Increasingly complex modality fusion techniques result in prolonged model training and inference times. (2) Histology datasets are typically small, making these complex models vulnerable to overfitting. To address these issues and improve a model’s ability to capture effective representations without increasing complexity, we introduce multi-task learning into survival prediction. Specifically, we propose a multi-task survival prediction framework that incorporates tumor staging classification as an auxiliary task, trained simultaneously with the survival prediction task. To the best of our knowledge, this is the first study to integrate tumor staging information into survival prediction. Our method was comprehensively evaluated through unimodal and multimodal experiments across five TCGA datasets. Most experiments demonstrated improved performance, with the best C-Index showing a 16.3% increase.

Skin cancer remains one of the most commonly diagnosed malignancies worldwide, where early and accurate detection is critical to improving patient outcomes. Dermoscopic image analysis has emerged as a critical tool for skin lesion classification. Traditional diagnostic methods rely heavily on visual assessment and histopathological analysis, which are time-consuming, subjective, and prone to inter-observer variability. Although recent computational approaches using convolutional neural networks (CNNs), hybrid models, and vision transformers have shown promise, they still struggle with key challenges such as class imbalance, inconsistent segmentation, and reduced generalizability across lesion types, limiting their clinical applicability. To overcome these limitations, we present an end-to-end framework that optimally combines preprocessing, segmentation, and classification to enhance performance. Our approach leverages a DermoGrabcut segmentation technique for more accurate segmentation. We employ a feature extraction approach that integrates Local Binary Patterns (LBP) and color histograms to capture essential textural and color features, thereby enhancing the performance of the subsequent classifier for more accurate predictions. We utilize classification test on Convolutional Neural Network (CNN) for robust classification, alongside comparative evaluations with Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Decision Trees (DT). We evaluate our model on the ISIC-2019 dataset, a comprehensive collection of dermoscopic images encompassing diverse lesion types and imbalanced class distribution. Results indicate that our proposed method using CNN model outperforms traditional classifiers and significantly outperforming existing methods, achieving high accuracy of 97.93%, a precision of 97.88%, and an F1-score of 96.80%. This approach successfully addresses the challenges of feature extraction and misclassification risk, providing an efficient and accurate solution for skin lesion classification for both clinical and telemedicine applications.

Viruses are of great diversity and variability and impact human society deeply and broadly. Advances in sequencing technologies enable easier detection and analysis of environmental samples, aiding clinical and virology research. Efficient viral classification remains a key challenge, with ongoing methodological improvements pursuing more reliable outputs. Time-consuming alignment-based methods are becoming obsolete with the explosion of data, while alignment-free especially machine learning methods take the lead. Along this line, we introduce VirB, a hierarchical classification method based on the latest BERT model ModernBERT, to classify the viral contigs or genomes at the order and family level. Integration of BPE and Transformer architecture with optimized embedding and attention strategies enables our model to process ultra-long sequences effectively, demonstrating superior performance over all compared methods in test cases. Experiments conducted with diverse real-world datasets confirmed the generalization power of the proposed model by presenting remarkable performance in predicting unseen sequences and sequences with noises.

Predicting drug-target interactions (DTI) is critical for drug discovery, influencing both the success and efficiency of drug development. Despite numerous methods exist, their predictive performance remains limited by irrelevant information and insufficient consideration of the spatial arrangement of proteins and drugs—both crucial for accurately modeling local interactions. To address these challenges, we propose FAPE-DTI, a deep learning model that integrates a focal attention network, a bilinear attention network, and relative positional encoding. The focal attention network leverages node importance scores to filter out irrelevant protein fragments, highlight key features, and reduce representational noise. The bilinear attention network further processes the refined protein and drug features to capture fine-grained pairwise interactions, constructing an atomic-level interaction graph that enhances the model’s structural awareness and interpretability. In addition, relative positional encoding—often overlooked in existing approaches—is incorporated into both the focal and bilinear attention modules to strengthen spatial interaction modeling. Experimental results show that FAPE-DTI consistently outperforms baseline models across multiple metrics on four benchmark datasets. Furthermore, case studies demonstrate that FAPE-DTI provides clearer insights into drug–target interaction regions at the atomic level, offering valuable interpretability to support rational drug design and development.

Drug-drug interactions (DDIs) refer to pharmacological and clinical responses to a drug combination, which are different from the known mode of actions of two drugs when used alone. Identifying potential DDIs is helpful for studying combination therapies and avoiding adverse effects that may occur when multiple drugs are used together. A number of models have been proposed to predict DDIs. However, identifying drug features and combining those features from multiple sources are still challenging. In this study, we propose a deep learning framework to identify potential DDIs, generate drug features from molecular view and DDI graph view, and fuse them together adaptively. In the molecular view, we take the atom attributes and molecular graphs into account, to reflect the chemical and topological properties of a drug molecule. In the DDI graph view, we concatenate those features from both chemical substructures and large language model, and design a soft-threshold dimensionality reduction network to retain essential features. In the feature fusion process, we design two adaptive parameters to concatenate those multi-view features. To systematically assess the efficacy of our proposed methodology, we performed comprehensive empirical evaluations across two benchmark datasets, employing comparative analysis against five contemporary cutting-edge approaches. The consistent experimental outcomes across all test scenarios substantiate the superior performance characteristics of our novel framework. In case studies, it shows the application value under realistic conditions.

In recent years, protein function prediction has made great breakthroughs in prediction performance. However, traditional protein function prediction methods mainly focus on single-species data, overlooking multi-species functional associations, which limits their generalizability to poorly annotated species. In addition, current deep learning methods are difficult to provide prediction basis in biology, which affects the application value in practical research. To solve the above problems, we propose a new interpretable multi-species protein function prediction model E-MSNGO, which constructs a heterogeneous network combining species and sequences by integrating sequence, structure and protein interaction (PPI) network information, and uses the graph attention mechanism in the model to efficiently propagate information. In addition, by calculating the functional similarity of proteins between species, the biological rationality of the prediction is improved, and the corresponding features are explained. Experimental results show that E-MSNGO has improved the corresponding indicators in multi-species protein function prediction, and can provide reasonable biological explanations, which improves the functional prediction ability of low-annotation species.

Channelrhodopsins (ChRs) are essential tools in optogenetics, playing a key role in analyzing and modulating neural circuits. However, traditional experimental methods for screening high performance ChR variants are costly, inefficient, and yield scarce data, making accurate function prediction and efficient screening challenging. To address this, we propose PRNet, a deep learning architecture based on contrastive ranking networks. It jointly encodes ChR variant sequences and structures, and uses a pairwise ranking strategy to transform regression into ranking, thereby highlighting critical features through pairwise comparisons. This expands the original 163 samples into 13,203 training pairs, effectively tackling the small-sample problem. PRNet employs depthwise separable 3D convolutions for efficient feature extraction, Bi-LSTM to model local sequence dependencies, and an SGA attention mechanism to capture residue relationships, significantly boosting prediction accuracy. Experimental results show that PRNet outperforms existing models, achieving prediction accuracies of 89%, 90%, and 91% for photocurrent strength, wavelength sensitivity of photocurrents, and off-kinetics, respectively. This study offers a novel solution to the precision issue in ChR protein prediction with small-samples, setting a new paradigm for protein engineering.

Accurate protein function prediction remains a fundamental challenge in bioinformatics, requiring computational methods that effectively translate amino acid sequences into functional annotations. We present DeepGO-ESM, a graph-based model that leverages the Evolutionary Scale Modeling (ESM) framework to improve protein function prediction based on the architecture of DeepGraphGO. In our approach, ESM transforms protein sequences into semantic embeddings, which are then utilized in two complementary graph-based prediction scenarios: a homogeneous protein-protein interaction graph where edges are directly inferred from embedding similarities, and a heterogeneous graph that integrates proteins with the directed acyclic structure of Gene Ontology terms. Through rigorous benchmark evaluations against other computational methods, our method demonstrates better predictive performance across multiple assessment metrics, particularly in scenarios with limited annotation data. Furthermore, our heterogeneous graph-based approach outperforms both conventional graph-based methods and state-of-the-art non-graph sequence models. The key innovations of DeepGO-ESM lie in: i) the automatic derivation of protein-protein interaction networks directly from ESM-generated embeddings, eliminating reliance on pre-existing interaction databases, and ii) the development of dual prediction models operating on both homogeneous protein graphs and heterogeneous protein-GO term networks. These contributions establish DeepGO-ESM as a powerful paradigm that bridges deep sequence modeling with structured biological knowledge, advancing computational approaches to functional genomics.

Single-cell RNA sequencing (scRNA-seq) has provided a large volume of data to discover cellular differences. Unsupervised clustering of scRNA-seq data is an important analytical method to identify cell subtypes. This paper introduces scGECA, a graph embedded representation learning method for clustering, which learns a low-dimensional data representation through an optimized graph attention network and uses a multilayer perceptron as a decoder to optimize the graph aggregation representation. As a result, we obtain a low-dimensional feature that is more suitable for single-cell clustering. Extensive experiments on multiple real-world datasets demonstrate that scGECA outperforms other state-of-the-art single-cell clustering methods in both clustering accuracy and robustness.

Predicting the structures of chemical and biomolecular products from reactants is a fundamental challenge in organic chemistry and biochemistry, essential for advancing synthetic pathway design and drug discovery. Traditional methods often fail to capture the intricate, multi-scale dynamics of these transformations, resulting in limited predictive accuracy and interpretability. In this work, we introduce ChemTransGNN +  +, a novel machine learning framework that integrates Graph Neural Networks (GNNs) and Transformers to predict chemical and biomolecular products from reactants with unprecedented accuracy. ChemTransGNN +  + incorporates four innovative components: Multi-Scale Chemical Attention (MSCA), which captures transformation patterns at atomic, functional group, and molecular levels; Reaction Narrative Flow (RNF) and Dynamic Graph Rewriting with Narrative Memory (DGR-NM), which model the temporal evolution of molecular graphs for chemically consistent predictions; and Visual Narrative Generator (VNG), which enhances interpretability by visualizing reaction pathways and key substructures. Evaluated on a dataset of 50,000 reaction pairs, ChemTransGNN +  + achieves state-of-the-art (SOTA) performance, with a Top-1 SMILES prediction accuracy of 87.3%, a Tanimoto similarity of 0.85, and a chemical validity of 96.8%, surpassing established methods such as Molecular Transformer and ReaMVP by 1.4% in accuracy and 1.7% in chemical validity. An ablation study validates the contributions of each component, while visualizations demonstrate the model’s ability to predict and interpret diverse reaction mechanisms, from dehydrations to cyclizations, across chemical and biomolecular systems. These results position ChemTransGNN +  + as a powerful tool for reaction product prediction, offering both high accuracy and mechanistic insights for applications in organic chemistry and biochemistry.

In the star alignment algorithm for multiple sequence alignment, all sequences are aligned directly to the central’star’ sequence without using a guide tree. This method greatly reduces computation time, making star alignment-based tools effective for aligning homologous sequences with high similarity. However, as sequence similarity decreases or the number of sequences increases, the algorithm’s accuracy drops significantly. In particular, “junk sequences” with very low similarity to the central star sequence tend to result in poor alignments, which can degrade the overall alignment quality. While realignment methods can greatly enhance the accuracy of alignments, there is currently a lack of approaches specifically tailored for star alignment tools. This study presents ReAlign-Star, a realignment method specifically designed for star alignment-based tools. The core of ReAlign-Star employs two key strategies—filtering out “junk sequences” and applying local vertical partitioning for realignment—to efficiently improve the quality of star alignments. Experiments on both simulated and real datasets demonstrate that ReAlign-Star significantly improves alignment accuracy in most cases, outperforming the initial alignments and extending the applicability of star alignment tools. The source code and test data for ReAlign-Star are available on GitHub ( https://github.com/malabz/ReAlign-Star ).

Multiple Sequence Alignment (MSA) is a core problem in computational biology, but traditional MSA tools struggle with high time and memory demands, particularly when handling large, dissimilar sequences. Existing methods, such as vertical segmentation, still face a trade-off between efficiency and accuracy.This study proposes FMAlign3, which addresses these challenges with two strategies. For large-scale similar sequences, we introduce a faster Star mode with parallel segmentation and a reconstructed core comparison algorithm, significantly boosting processing speed. For dissimilar sequences, we design the Pro mode with a horizontal segmentation strategy to improve accuracy while maintaining time efficiency.Experimental results show FMAlign3 achieves a 22-fold performance improvement over FMAlign2, with lower memory overhead and improved accuracy, particularly in dissimilar sequence alignment. Overall, the method improves performance by 20%. The program is available at: https://github.com/Aohy-github/FMAlign3 .

A ( $$k$$ k ]-tuple common substring (abbr. ( $$k$$ k ]-CSS) is a common subsequence of multiple given strings including at most $$k$$ k common substrings. This pattern of two strings is retrievable in quadratic time and linear space and even more, in subquadratic time and space if $$k$$ k is a constant. Motivated by computational biology applications in need of a ( $$k$$ k ]-CSS with substantially many consecutive matches, we propose to find a longest ( $$k$$ k ]-CSS of two strings whose substrings are of length at least $$l$$ l , of which the complexity is indefinite.We present a dynamic programming algorithm to find such a longest ( $$k$$ k ]-CSS of two strings whose lengths are $${n}_{1}$$ n 1 and $${n}_{2}$$ n 2 in $$O(k{n}_{1}{n}_{2})$$ O ( k n 1 n 2 ) time and space, the same complexity as without the length bound $$l$$ l . Through rolling array based dynamic programming to get the longest ( $$k$$ k ]-CSS length in advance, we present a divide-and-conquer algorithm to find such a longest ( $$k$$ k ]-CSS in $$O(k{n}_{1}{n}_{2})$$ O ( k n 1 n 2 ) time and $$O({n}_{1}+kl{n}_{2})$$ O ( n 1 + k l n 2 ) space, which is intended to work for two much longer given strings. We also present an algorithm to find such a longest ( $$2$$ 2 ]-CSS in $$O(n{log}^{2}n)$$ O ( n log 2 n ) time where $$n$$ n is the total length of input strings.

The prediction of transcription factor binding site (TFBS) is essential for under-standing the combination mechanism and cell function of transcription factors. Although there are already several algorithms to predict TFBS, there is still room for improvement in the prediction effect. This study proposes a new neural net-work model called DeepCatl, which combines DNA sequence and shape characteristics to improve the accuracy and reliability of TFBS prediction. DeepCatl uses the convolutional neural network (CNN), channel attention module, and the improved Transformer encoder to capture DNA characteristics with refinement and comprehensiveness. The DeepCatl model extracts the sequence features of DNA through convolutional neural networks and channel attention mechanism, and extracts the structural features of DNA through Transformer encoder and Bi-LSTM. In the end, based on the fusion of these characteristics, DeepCatl is calculated and predicts the location of TFBS. We verified the superior performance of the DeepCatl model on the 165 Encode Chip-seq dataset, and the results show a significant improvement over traditional methods. This deep learning model that combines the channel attention mechanism and the Transformer encoder provides a new approach for bioinformatics research, which can further study the complex mechanism of interaction with DNA. The DeepCatl model not only integrates the channel attention mechanism and transformer encoding, but also extracts features from DNA sequences and DNA structures to predict transcription factor binding sites (TFBS).

Due to variations in acquisition conditions, medical image data often exhibit substantial domain shift, which decreases the performance of detection models. Prior works suffer from suboptimal performance and time inefficiencies in cross-domain detection. Specifically, we designed a domain adaptation module to further process the features output by the prototype to complete the detection. Furthermore, we have restructured the backbone architecture through FasterNet. Using the proposed domain discriminator and gradient reversal layer, we conducted domain adversarial training for the module. Finally, we tested the proposed DA-YOLO model on brain tumor datasets collected under various conditions. The experimental results demonstrate that the proposed FDA-YOLO is well performed in domain adaptation, particularly on unseen datasets, outperforming YOLOv11 and its general variants. Compared to the baseline model, FDA-YOLO achieves notable improvements in scenarios with substantial cross-domain discrepancies, including a 9.5% increase in precision and a 4.44% boost in mAP@50, all without sacrificing computational efficiency.

Graph Neural Networks have found widespread application in the field of precision medicine, particularly for predicting drug responses in cell lines, which imposes greater demands on the interpretability of prediction results. However, current graph interpretability algorithms tend to emphasize generality and overlook the complex interactions in drug data, making it difficult to attribute predictions to individual factors and limiting their use in predicting cancer drug responses. In this paper, we propose CETExplainer, a novel post-hoc interpretability algorithm built upon a multi-relational graph neural network-based framework for drug response prediction. We model drug response data using a multi-relational graph and enhance feature representations through both contrastive learning and multi-task learning. Furthermore, we introduce an interpretability mechanism based on a controllable edge-type-specific weighting scheme. It considers the mutual information between subgraphs and predictions, proposing a structural scoring approach to provide fine-grained, intuitive explanations for predictive models. We also introduce a method for constructing ground truth based on real-world datasets to quantitatively evaluate the proposed interpretability algorithm. The experimental results achieved a prediction AUC of 0.942 interpretability precision of 0.7134, outperforming the baseline methods. Qualitative experiments further demonstrated that our model can capture meaningful structures, providing a promising solution to the black-box challenge in drug response prediction.