Briefings in bioinformatics最新文献

Cancer classification is pivotal for precision oncology, yet traditional methods struggle with the molecular heterogeneity of tumors. Our study introduces a self-attention based Conv1D machine learning network designed for panel capture sequencing data, which is more commonly used in clinical settings. Combining clinical capture sequencing data and The Cancer Genome Atlas data, we achieved an overall classification accuracy of over 90%, with precision rates reaching 100% for cervical and gastric cancers. Additionally, recall rates were highest at 95.79% for gastric cancer and lowest at 77.46% for cervical cancer, demonstrating robust performance across various cancer types. The model identified key genes such as C3orf36, JHY, and TASP1, showing significant differences in mutation counts across cancers. High-impact gene enrichment analysis highlighted critical pathways like acute myeloid leukemia and adipocytokine signaling. This approach not only significantly improves the precision of cancer classification, demonstrating the potential for clinical application, but also enhances our understanding of cancer biology.

Comprehensive pan-domain metagenomic classification is increasingly constrained by the memory and runtime costs of building and querying the rapidly expanding reference genome space. We introduce Kun-peng, a taxonomic classifier powered by an intelligent block-partitioned database structure and optimized search strategies, enabling ultra-scalable, memory-efficient pan-domain profiling. Using the Critical Assessment of Metagenome Interpretation II benchmark, Kun-peng substantially reduces the memory usage of database-building and querying by up to 24-fold, and accelerates sample classification by up to 4.73-fold compared with Kraken2. Kun-peng achieves competitive accuracy with fewer false positives than Kraken2, Centrifuger, and even KrakenUniq, while maintaining consistently high sensitivity across diverse datasets. In a real-world evaluation of 586 metagenomic samples spanning air, water, soil, and human-associated environments, we performed classification using a 4.3 TB pan-domain database comprising 204,477 genomes, which was built by Kun-peng with only 4.1 GB peak memory. Kun-peng processed each sample in 0.2-11.2 min with 4.0-35.4 GB peak memory, corresponding to a 54-473-fold reduction in memory usage relative to Kraken2. Compared with Sylph, Kun-peng achieved up to a 46-fold speedup while requiring 21-fold less memory. Kun-peng classified 69.8%-94.3% of reads, improving coverage by 20%-60% over the standard Kraken2 database with 62,026 genomes. This improvement reflects expanded reference coverage, although a small fraction of false positives is inherent to k-mer-based methods. Overall, Kun-peng effectively eliminates the long-standing memory bottleneck in pan-domain database building and classification, enabling rapid and scalable pan-domain taxonomic analysis of complex environmental, ecological, and exposomic sequencing datasets.

Spatially variable genes (SVGs) are essential for elucidating tissue organization within spatially resolved transcriptomics. While a number of computational methods have been developed for SVG identification, their reliance on algorithm-specific assumptions, such as predefined kernel functions or spatial neighborhood graphs, often results in substantial variability in sensitivity and inflated false discovery rates (FDRs) across heterogeneous datasets. To address this challenge, we here develop Castl, an ensemble-based framework for SVG identification that integrates multiple detection methods through statistically designed aggregation modules. Comprehensive evaluations on both simulated and real-world data demonstrate that Castl consistently identifies biologically meaningful spatial expression patterns, mitigates method-specific biases and effectively controls FDRs across various biological contexts, resolutions, and spatial technologies. This flexible, assumption-free framework offers a robust and standardized foundation for spatially informed feature discovery in complex biological systems.

The development of single-cell RNA sequencing (scRNA-seq) technology provides unprecedented opportunities for elucidating cell heterogeneity and gene expression. Identifying and discovering cell types through cell clustering is a crucial step in analyzing scRNA-seq data. However, the high-dimensionality nature and frequent dropout events of the data raise great challenges for cell clustering. Here, we propose a novel contrastive clustering framework called scSCCNIA (Similarity-matrix-based Contrastive Clustering with Neighbor Information Aggregation), for the accurate identification of cell clusters from scRNA-seq data. scSCCNIA adopts a Laplacian filter to conduct neighbor information aggregation, constructs different graph views by using special un-shared parameters Siamese encoders for data augmentation, and learns the latent low-dimensional embedding representations via similarity-matrix-based contrastive learning. Comparative analyses of multiple scRNA-seq datasets from different platforms and with varying cell numbers demonstrate that scSCCNIA outperforms existing methods in terms of cell clustering and marker gene identification. Furthermore, scSCCNIA reveals the heterogeneity and functional specificity of various cell types through Gene Ontology terms and Kyoto Encyclopedia of Genes and Genomes enrichment analyses. Overall, scSCCNIA is an effective algorithm for learning latent features from scRNA-seq data, enhancing cell type identification accuracy and facilitating downstream analyses of scRNA-seq data.

G protein-coupled receptors (GPCRs) are among the most important drug targets, and peptide therapeutics are rapidly emerging. However, accurate prediction of peptide-GPCR interactions (PepGI) remains challenging due to the scarcity of high-quality data and the poor generalization of existing drug-target interaction (DTI) models, which are largely trained on small molecule data. Here, we introduce a progressive fine-tuning framework with a dynamic parameter selection strategy that adaptively selects critical fine-tuning parameters using Fisher information. Our method begins with pretraining on a large small molecule-GPCR dataset, followed by intermediate fine-tuning on peptide-target data to alleviate the representation mismatch across heterogeneous ligand modalities. Finally, the task-specific fine-tuning is performed on the low-resource PepGI scenario. Extensive experiments show that our approach significantly outperforms baselines across multiple evaluation metrics, and exhibits robust generalization under few-shot and practical cold-start settings. Overall, this work offers an effective solution for low-resource peptide-GPCR prediction and presents a transferable framework for cross-structure DTI modeling.

Genome-wide association studies (GWASs) have been conducted primarily in European (EUR) populations, limiting insights into underrepresented groups such as East Asian (EAS), but cross-ancestry GWASs have demonstrated high trans-ethnic genetic similarity between EUR and non-EUR populations. To enhance association analysis power in EAS populations, we propose tranScore, a novel summary-statistics-based transfer learning method that leverages trans-ethnic genetic similarity through hierarchical modeling. By considering EUR as auxiliary population, tranScore performs joint testing of genetic effects in auxiliary and target populations via well-established P-value combination procedures. Simulations demonstrate that tranScore maintains control of type I error rates and provides substantial power gains for diverse genetic architectures, showing robustness against various challenges including incomplete SNP overlap and effect heterogeneity. In the real-data application of eight diseases from the China Kadoorie Biobank (CKB), after incorporating the genetic information of the EUR population, tranScore identified significantly more genes than the traditional score test which ignored such information. Approximately 41.9% of discovered genes were replicated in the Biobank Japan cohort. Overall, tranScore represents a flexible and powerful statistical approach for association analysis of complex diseases and traits through transfer learning of shared genetic similarities between the auxiliary and target populations.

Drug repurposing provides a cost-effective and time-efficient strategy to accelerate therapeutic discovery, yet most computational approaches fail to capture the multi-scale biomedical mechanisms underlying drug-disease associations, limiting interpretability. We introduce BioMNEDR (mechanism-guided network embedding for drug repurposing) that integrates heterogeneous biomedical networks through biologically curated meta-paths. BioMNEDR generates low-dimensional embeddings preserving protein-protein interactions and functional hierarchies. It further integrates multi-path predictions through an XGBoost classifier. The framework achieves state-of-the-art performance, consistently surpassing strong baselines across AUROC, AUPR, recall, and F1-score, while maintaining a balanced trade-off in precision. Case studies further highlight its practical utility, demonstrating the ability to rediscover approved drugs and prioritize promising candidates, such as cromoglicic acid for Alzheimer's disease. By explicitly modeling multi-scale mechanisms, BioMNEDR enhances both predictive accuracy and biomedical interpretability, offering a robust computational framework for systematic drug repurposing.

As genome-wide association studies (GWAS) studies move from array-based genotyping to whole exome and genome sequencing, there is a significant increase in cost. Applying an appropriate technique for the selection of which controls to include, in large studies where more potential controls are available than needed for the study, may be a useful technique for minimizing resource intensity whilst maintaining statistical power. We evaluated three control selection strategies in prostate cancer GWAS using 15 250 UK Biobank cases: (a) all controls, (b) matched controls, and (c) random selection. Both (b) and (c) achieved comparable power in detecting significant loci relative to (a), but matched controls (b) showed greater consistency in identifying leading single nucleotide polymorphisms (SNPs). However, using (b) matched controls reduced discovery power by ~30% compared with (a) all controls, highlighting a trade-off. Matching controls (1:4 ratio) offers a cost-effective approach for targeted SNP analysis across phenotypes but may miss novel associations.

RNA-based technologies have demonstrated significant potential for diverse applications, ranging from vaccination to gene editing. However, their widespread adoption is limited by the critical challenge of efficient delivery. Lipid nanoparticles (LNPs) have emerged as a widely utilized RNA delivery system, yet their formulation design and optimization primarily rely on empirical trial-and-error, which is labor-intensive, time-consuming, and cost-prohibitive, thus hindering the rapid development of RNA therapeutics. To facilitate the early-stage design and optimization of LNPs for enhanced delivery efficiency, in this study, we construct LNPs-TE, a benchmark dataset comprising over 10 000 experimentally measured transfection efficiency (TE) values, and introduce LNPs integrated feature fusion Transformer (LIFT), a deep learning framework for LNPs TE prediction. Comprehensive experiments demonstrate that LIFT effectively integrates multidimensional molecular representations of ionizable lipids, the key component in LNPs formulation, achieving superior predictive performance, with an average Pearson correlation coefficient of 0.845 for regression and an area under the receiver operating characteristic curve (AUC-ROC) of 0.818 for multi-class classification across multiple datasets. Through scaffold-based splitting and activity cliff tasks, we further validated the exceptional generalization ability and robustness of LIFT, which achieved over a 10% improvement in the coefficient of determination (R2) compared with state-of-the-art baseline models, highlighting its potential as a practical and stable approach for the virtual screening of efficient LNPs formulation. The relevant data, model and code are made publicly available at https://github.com/U12458/LIFT.