Long non-coding RNAs (lncRNAs) play a pivotal role in regulating biological processes, primarily through lncRNA-Protein Interactions (LPI), which are crucial for understanding disease mechanisms and cellular functions. While computational methods exist for predicting LPI, they suffer from suboptimal feature integration, limited generalizability, and inadequate network modeling. To address these challenges, we propose MAKGLPIN (Multi-Architecture, Multi-Kernel Graph-based LPI Network), a novel deep learning framework that integrates sequence features, structural information, and physicochemical properties of lncRNAs and proteins. MAKGLPIN employs a multi-layer Graph Convolutional Network (GCN) enhanced with a self-attention mechanism to capture complex interaction patterns, alongside a Graph Attention Network (GAT) to prioritize critical interactions while preserving long-range dependencies. Extensive evaluations demonstrate that MAKGLPIN outperforms existing methods in robustness, generalizability, and predictive precision across benchmark datasets. Our model provides a computationally efficient and biologically interpretable solution, advancing the study of LPI and their applications in disease research.
Musical protein (MPro), the protein defined by music (namely biologization of music), is a reverse operation of the traditional life music that addresses the sonification of biomolecule to musical audio. We previously described a computational strategy for MPro facilitation, which maps the time sequence of music onto the spatial architecture of protein. Here, we further extend MPro to musical peptide (MPep) by fragmentizing MPro protein at either music time sequence level or protein residue sequence level, which represents a new technique to de novo generate distinct peptidic big data with intra-sequence organization, thus particularly suitable for AI-oriented tasks. On the other side, DeepPepQSAR methodology is raised as an integration of deep learning and traditional peptide quantitative structure–activity relationship (PepQSAR) in big-data bioactive peptide discovery, and we herein practice it on AI-driven high-throughput antimicrobial peptide (AMP) screening. Both the structural properties of length-varying peptide sequences and the bacterial species tested for these peptides are visually characterized by using PepImage representation, which is then fed into a dual-path convolutional neural network (CNN) to train a predictive DeepPepQSAR regression scoring function for further discovering potent AMPs from millions of MPep candidates. Consequently, 32 MPep peptides are identified to have a potential antibacterial profile, in which 4 and 12 peptides exhibit high and moderate activities, respectively, that are generally comparable with the sophisticated AMP peptide KR-12, but lower than the FDA-approved antibiotic Levofloxacin.
Self-supervised and context-aware representation learning offers a powerful strategy for extracting robust features from high-dimensional, sparse, and noisy single-cell data. Here, we present scDGCN, a graph-based deep clustering framework for single-cell RNA sequencing (scRNA-seq) analysis. scDGCN models intercellular similarity as a graph and leverages local neighborhood context to guide expression reconstruction, thereby learning robust cell representations. The method integrates a network enhancement module to denoise the cell similarity graph, a neighborhood-enhanced autoencoder that reconstructs gene expression using topological neighborhood context, and a graph neural network to preserve structural semantics among cells. By embedding each cell’s expression profile within its neighbors, scDGCN captures subtle transcriptional variation and identifies cell subtypes without labeled references. Across ten scRNA-seq datasets from diverse human and mouse tissues, scDGCN achieves competitive adjusted Rand index (ARI), normalized mutual information (NMI), and clustering accuracy (ACC), supporting unsupervised identification of cellular subpopulations. These results demonstrate that combining graph denoising, neighborhood-enhanced context reconstruction, and graph-structured representation learning improves unsupervised single-cell clustering and facilitates high-resolution cell atlas construction.
PIWI-interacting RNAs (piRNAs) maintain genomic stability by interacting with PIWI proteins and have been implicated in various human diseases. However, wet-lab experiments remain limited in throughput, creating an urgent need for computational methods to systematically identify piRNA-disease associations. Although numerous computational approaches have been proposed, they still face major challenges such as limited feature representations and insufficient integration of biological knowledge. This study presents LLM4PDA, a large language model-augmented framework for piRNA-disease association prediction. PiRNA sequences are encoded using RNA-FM, whereas disease representations are derived from LLM-generated disease descriptions guided by a reusable, structured prompt template and encoded using OpenAI’s text-embedding models. These representations are mapped into a common latent space through an MLP-based alignment module, enabling association scoring for candidate piRNA-disease pairs. Separately from the predictive model, a post-prediction interpretive module organizes high-confidence predictions into structured, knowledge-informed biological summaries to support hypothesis generation and guide downstream experimental validation. Five-fold cross-validation demonstrates that LLM4PDA achieves area under the curve (AUC) values of 0.984 and 0.985 on two benchmark datasets, outperforming eight state-of-the-art methods. Case studies further illustrate how the interpretive module facilitates structured synthesis of biological context for selected high-confidence associations.
A growing body of evidence indicates that microRNAs (miRNAs) play essential roles in cancer progression, therapy, and drug resistance. They are increasingly recognized as promising therapeutic targets for overcoming cancer drug resistance. However, existing computational methods for predicting miRNA-drug associations are often limited by sparse labeled associations, reliance on single-modality data, and insufficient ability to capture long-range dependencies in sequences. In this study, we propose a novel deep learning framework, named MDCL, that integrates multi-source information fusion with collaborative contrastive learning. MDCL uses a multi-scale convolutional neural network (multi-scale CNN) to learn sequence embeddings at different granularities from miRNA sequences and drug SMILES strings. In parallel, frequency chaos game representation (FCGR) and multiple molecular fingerprints are used to characterize the two-dimensional structural patterns of miRNAs and the physicochemical properties of drugs, respectively. To effectively integrate multi-view information, the model introduces a cross-view attention mechanism for feature enhancement and constructs a local–global alignment loss within a collaborative contrastive learning (CCL) framework, substantially improving feature robustness and discriminability. We evaluated MDCL on benchmark datasets manually curated from the latest database release. Extensive five-fold cross-validation experiments, repeated five times with different random seeds, show that MDCL consistently outperforms state-of-the-art methods, achieving an AUC of 0.9450 and an AUPR of 0.9281 for miRNA-drug resistance (MDR) association prediction, and an AUC of 0.9379 and an AUPR of 0.9293 for miRNA-drug sensitivity (MDS) association prediction. MDCL also attains the best F1-score and Matthews correlation coefficient (MCC) among all compared methods. Moreover, under a strict independent-node cold-start setting, in which both miRNAs and drugs in the test set are unseen during training, MDCL remains competitive in ranking-based metrics, demonstrating strong generalization ability. In addition, a case study of docetaxel further demonstrates the practical utility of the model for discovering novel miRNA-drug associations. The data and source code are available at https://github.com/scxdhr020629/MDCL .
