Background: Bulk RNA-seq is a cost-effective method for measuring average gene expression in tissue samples, but its lack of single-cell resolution limits the understanding of cellular heterogeneity. Computational deconvolution aims to infer cell-type proportions from bulk RNA-seq data; however, the accuracy of existing methods needs improvement, especially in complex tissues. Methods: In this study, we introduce DiffFormer, a novel deconvolution model that, for the first time, integrates a conditional diffusion model with a Transformer architecture. We systematically evaluated DiffFormer on four pseudo-bulk datasets and validated it on a gold-standard real-world dataset with FACS-based ground truth. Results: DiffFormer demonstrated consistent and strong performance across all test datasets, outperforming existing methods and a baseline MLP-based diffusion model (DiffMLP). For instance, on the pbmc3k dataset, DiffFormer reduced the Root Mean Square Error (RMSE) from 0.1060 to 0.0120 compared to DiffMLP. This advantage was further confirmed on the real-world dataset, where DiffFormer achieved the highest Pearson Correlation Coefficient (PCC). Conclusions: This work provides a high-precision, reproducible tool for cellular deconvolution. Crucially, the direct comparison with an MLP-based diffusion model provides definitive evidence that the Transformer architecture is key to its success, highlighting the potential of such models for solving complex bioinformatics problems.

Background: Low sequencing depth causes molecular capture loss and zero inflation in scRNA-seq data, reducing the reliability of downstream analyses. Methods: We propose DepthDiff, a depth-conditional expression enhancement model trained with diffusion-based denoising. Using low-depth expression profiles and sequencing depth ratio as conditions, DepthDiff learns a supervised residual mapping to paired high-depth references. During inference, it requires only a single-step forward prediction. Fixed-UMI downsampling was used to construct benchmarks across three public datasets. Results: DepthDiff outperformed supervised baselines, MAGIC, and unenhanced low-depth data in expression reconstruction and biological signal preservation. Ablation experiments showed that x0 prediction and cosine noise scheduling were important for stability, while full reverse diffusion sampling provided no additional benefit. Cross-dataset transfer and CITE-seq validation supported the generalizability and biological relevance of the recovered signals. Conclusions: DepthDiff is an efficient supervised framework for low-depth scRNA-seq expression enhancement, with gains mainly driven by diffusion-based denoising training rather than generative reverse sampling.

Over-representation analysis (ORA) is widely used to interpret selected-gene lists, including differentially expressed genes and cell-type marker genes. Unlike ranked-list gene-set enrichment analysis (GSEA), ORA tests whether selected genes are over-represented in predefined gene sets under an explicit background. Repeated analysis of large gene-set libraries, however, can be computationally costly. GPU-accelerated methods such as rapidGSEA mainly target ranked-list GSEA and do not address repeated exact-ORA processing. We developed paraORA to accelerate repeated ORA by preparing a gene-set library once and reusing its compressed representation across queries. In the GPU-exact path, overlap counting, exact hypergeometric testing, Benjamini–Hochberg correction, and odds-ratio calculation are performed on the GPU. Across 425 numerical validation cases on three GPU models, paraORA agreed with CPU-reference results within the prespecified tolerance while preserving statistical decisions and rankings. On an RTX 4090, the CPU/GPU runtime ratio for the complete in-memory workflow increased from 1.89 for one query to 60.10 for 100 queries against 141,002 gene sets, including one-time preparation but excluding file writing. In a separate complete file-to-file benchmark, GPU-exact was 2.55-fold faster than CPU execution. An exploratory limma reanalysis showed direction-specific Hallmark enrichment. paraORA is available as a web service and an open-source package.

Background/Objectives: Large language models and automated analytical tools show potential for biomedical text understanding, knowledge integration, and single-cell data interpretation, but interpretations of specific gene regulatory relationships still require empirical grounding. For TF–target relationships, one relevant constraint is whether curated regulatory edges show detectable expression-level statistical evidence in the single-cell data being interpreted, because such evidence may vary across cellular states, conditions, and regulatory mechanisms. We therefore evaluated the statistical visibility of known transcription factor (TF)–target relationships in single-cell expression space to aid the interpretation of regulatory inference and automated-tool outputs. Methods: We performed two complementary analyses: first, testing whether curated TF–target edges showed stronger pair-level associations than matched background gene pairs; and second, assessing whether high-scoring TF–gene pairs corresponded to existing regulatory knowledge and whether their target genes showed pathway coherence. Results: Across four peripheral blood mononuclear cell (PBMC) datasets, four adult tissues, and seven adult cell types, curated TF–target relationships showed weak but reproducible statistical visibility rather than strong separation. For all five association metrics, the mean area under the precision–recall curve (AUPRC) was only slightly above the random-ranking baseline of 0.5. High-scoring pairs were more often supported by curated resources, and their target genes showed context-specific Hallmark pathway coherence. Conclusions: Single-cell expression associations can provide useful but limited statistical clues for TF–target regulation and should be interpreted as complementary rather than definitive regulatory evidence.

Background: RNA degradation in clinically archived samples introduces systematic biases into RNA-seq data, limiting the accuracy of downstream analyses. Developing computational methods for high-fidelity transcriptome restoration is therefore of critical importance. Methods: We introduce DiffRepairer, a deep learning model that combines a Transformer architecture with a conditional diffusion model framework to reverse the effects of RNA degradation. The model is trained on “degraded-original” paired data, generated via a comprehensive simulation pipeline, to learn a direct, one-step repair mapping. Results: Across five diverse pseudo-degraded datasets, DiffRepairer demonstrated stable and superior performance, outperforming traditional statistical methods (e.g., CQN) and standard deep learning models (e.g., VAE) on key technical and biological metrics. Conclusion: DiffRepairer is a validated, high-precision tool for transcriptome repair that effectively restores biologically meaningful signals from degraded RNA-seq data, highlighting the potential of advanced generative models in bioinformatics.