Background: Single-cell annotation plays a crucial role in the analysis of single-cell genomics data. Despite the existence of numerous single-cell annotation algorithms, a comprehensive tool for integrating and comparing these algorithms is also lacking. Methods: This study meticulously investigated a plethora of widely adopted single-cell annotation algorithms. Ten single-cell annotation algorithms were selected based on the classification of either reference dataset-dependent or marker gene-dependent approaches. These algorithms included SingleR, Seurat, sciBet, scmap, CHETAH, scSorter, sc.type, cellID, scCATCH, and SCINA. Building upon these algorithms, we developed an R package named scAnnoX for the integration and comparative analysis of single-cell annotation algorithms. Results: The development of the scAnnoX software package provides a cohesive framework for annotating cells in scRNA-seq data, enabling researchers to more efficiently perform comparative analyses among the cell type annotations contained in scRNA-seq datasets. The integrated environment of scAnnoX streamlines the testing, evaluation, and comparison processes among various algorithms. Among the ten annotation tools evaluated, SingleR, Seurat, sciBet, and scSorter emerged as top-performing algorithms in terms of prediction accuracy, with SingleR and sciBet demonstrating particularly superior performance, offering guidance for users. Interested parties can access the scAnnoX package at https://github.com/XQ-hub/scAnnoX.

Introduction: Tertiary lymphoid structures (TLSs) are ectopic lymphoid formations that arise in non-lymphoid tissues due to chronic inflammation. The pivotal function of TLSs in regulating tumor invasion and metastasis has been established across several cancers, such as lung cancer, liver cancer, and melanoma, with a positive correlation between increased TLS presence and improved prognosis. Nevertheless, the current research about the clinical significance of TLSs in breast cancer remains limited.Methods: In our investigation, we discovered TLS-critical genes that may impact the prognosis of breast cancer patients, and categorized breast cancer into three distinct subtypes based on critical gene expression profiles, each exhibiting substantial differences in prognosis (p = 0.0046, log-rank test), with Cluster 1 having the best prognosis, followed by Cluster 2, and Cluster 3 having the worst prognosis. We explored the impact of the heterogeneity of these subtypes on patient prognosis, the differences in the molecular mechanism, and their responses to drug therapy and immunotherapy. In addition, we designed a machine learning-based classification model, unveiling highly consistent prognostic distinctions in several externally independent cohorts.Results: A notable marker gene CXCL13 was identified in Cluster 3, potentially pivotal in enhancing patient prognosis. At the single-cell resolution, we delved into the adverse prognosis of Cluster 3, observing an enhanced interaction between fibroblasts, myeloid cells, and basal cells, influencing patient prognosis. Furthermore, we identified several significantly upregulated genes (CD46, JAG1, IL6, and IL6R) that may positively correlate with cancer cells' survival and invasive capabilities in this subtype.Discussion: Our study is a robust foundation for precision medicine and personalized therapy, presenting a novel perspective for the contemporary classification of breast cancer.

Spatial transcriptomics technologies have revolutionized our understanding of tissue heterogeneity by providing gene expression data with spatial context. However, due to the limited resolution of some of these techniques, multiple cell types are often captured at each spatial location. Several spatial deconvolution methods have been developed in recent years to infer cell type compositions from these mixed spots, but their results can be noisy and spatially inconsistent. In this study, we introduce a spatial smoothing method, Spatialsmooth based on convolutional autoencoder, in which several spatial deconvolution tools are integrated to obtain the deconvolution results, the spatial location information is fully utilized using positional encoding, and the deconvolved cell type compositions are smoothed using a convolutional autoencoder to optimize the distribution of different cell type compositions in space. This method integrates multiple spatial deconvolutions tools and fully utilizes spatial location information and cell type composition inferred by multiple deconvolution algorithms to produce smooth, biologically plausible cell type distributions. The smoothing effect was measured by using three spatial metrics, namely, Moran's I, Geary's C, and Total Variation, on multiple pairs of datasets. pancreatic ductal adenocarcinomas (PDAC) dataset, Spatialsmooth reached 0.52 on the Moran's I metric (92% higher than the Redeconve method), the Geary's C metric was reduced to 0.47 (a 45% improvement compared to RCTD), and Total Variation was reduced by 22%. Applied to benchmark data, multiple cell types and molecular markers with different spatial localizations were identified, closer to the spatial projection of marker genes than other deconvolution tools. ( https://github.com/njjyxl/Spatialsmooth ).

Polyamine metabolism is involved in several cellular processes in organisms and is closely associated with tumorigenesis, progression, and metastasis. In this study, using breast cancer data from The Cancer Genome Atlas (TCGA) database as well as polyamine metabolism-related genes obtained from previous studies, key genes were screened using LASSO regression, and a prognostic label was created, which was validated using an independent cohort, Metabric. The breast cancer cohort of this prognostic label TCGA was divided into high-risk and low-risk groups, and the prognosis of high-risk and low-risk groups was significantly different. The biological indicators of high-risk and low-risk groups were significantly different, and the prediction results of this model were validated in the Metabric cohort. Our model can effectively predict the prognosis of breast cancer patients based on the characteristics of polyamine metabolism-related genes, which are also related to immune cell infiltration and immunotherapy response, and have some potential for clinical application.

Spatial transcriptomics technologies profile gene expression across tissue sections while retaining spatial information, yet most platforms capture signals from multiple cells per measurement location, requiring computational methods to determine the underlying cellular composition. Current deconvolution approaches either depend on single-cell reference atlases—which may be unavailable for rare tissues or specific disease contexts—or employ unsupervised methods that struggle with complex spatial patterns and typically need manual specification of how many cell types to identify. We developed Attention-Guided Enhanced Deconvolution (AGED), a two-stage framework combining probabilistic modeling with neural attention architectures for reference-free analysis. The first stage uses a Performer-based network with linear-complexity attention to systematically evaluate models with different numbers of cell types, automatically selecting the optimal configuration through composite scoring of reconstruction quality, cluster separation, and topic diversity. The second stage, Attention-Guide, stars with hierarchical probabilistic initialization. It progressively refines cell type features through multiple attention mechanisms: cross-attention based on statistical priors, spatial attention aggregating neighborhood information, and collaborative attention capturing theme-gene associations. A dynamic gating mechanism enables the model to balance global statistical patterns with locally data-driven features, while regularization promotes sparse solutions consistent with biological reality—each position containing only a few dominant cell types. Testing on Mouse Olfactory Bulb (MOB) tissue, AGED automatically identified four anatomical structures and achieved superior reconstruction performance (r=0.86) with characteristic sparsity patterns. When similarly applied to human pancreatic ductal adenocarcinoma (PDAC) and human thymus tissue, it revealed detailed relationships between dissected structures and cell types. The learned cell type distributions aligned well with established neuroanatomical boundaries and known molecular markers, indicating the attention-based refinement maintains biological interpretability throughout training. This framework offers a practical solution for spatial transcriptomics analysis across diverse experimental systems without requiring matched single-cell data.

Background: Cancer classification using gene expression data presents significant challenges due to high dimensionality and complex biological patterns. While integrated learning approaches show promise, current methods often fail to capture complex model interactions and lack automated optimization frameworks. Methods: We developed a novel hybrid machine learning framework combining traditional machine learning models with a Transformer-based stacked architecture for cancer classification. The framework employs a parallel optimization strategy for multiple base models (Random Forest, Gradient Boosting, eXtreme Gradient Boosting, and support vector machines) and utilizes a Transformer-based meta-learner for ensemble prediction. We evaluated the framework on a comprehensive cancer gene expression dataset containing 3,181 samples across five cancer types. Results: The proposed framework achieved superior performance with over 99.7% accuracy across all cancer types. The Transformer-based meta-learner demonstrated particular effectiveness in handling complex cases, showing significant improvements in classification accuracy for rare cancer subtypes compared to traditional methods. Discussion: Our hybrid framework successfully leverages the strengths of traditional machine learning algorithms and the advanced pattern recognition capabilities of Transformer architecture to provide a robust and interpretable solution for cancer classification. The research results indicate that this framework has potential application value in the fields of cancer diagnosis and model interpretability, marking an important advancement in the integration of learning technologies in the field of cancer classification.