Identifying the molecular mechanisms related to genomic breakage is an important goal of cancer mechanism studies. Among diverse locations of structural variants, fusion genes, which have the breakpoints in the gene bodies and are typically identified from the split reads of RNA-seq data, can provide a highlighted structural variant resource for studying the genomic breakages with expression and potential pathogenic impacts. In this study, we developed FusionAI, which utilizes deep learning to predict gene fusion breakpoints based on DNA sequence and let us identify fusion breakage code and genomic context. FusionAI leverages the known fusion breakpoints to provide a prediction model of the fusion genes from the primary genomic sequences via deep learning, thereby helping researchers a more accurate selection of fusion genes and better understand genomic breakage.

Understanding disease progression and sophisticated tumor ecosystems is imperative for investigating tumorigenesis mechanisms and developing novel prevention strategies. Here, we dissected heterogeneous microenvironments during malignant transitions by leveraging data from 1396 samples spanning 13 major tissues. Within transitional stem-like subpopulations highly enriched in precancers and cancers, we identified 30 recurring cellular states strongly linked to malignancy, including hypoxia and epithelial senescence, revealing a high degree of plasticity in epithelial stem cells. By characterizing dynamics in stem-cell crosstalk with the microenvironment along the pseudotime axis, we found differential roles of ANXA1 at different stages of tumor development. In precancerous stages, reduced ANXA1 levels promoted monocyte differentiation toward M1 macrophages and inflammatory responses, whereas during malignant progression, upregulated ANXA1 fostered M2 macrophage polarization and cancer-associated fibroblast transformation by increasing TGF-β production. Our spatiotemporal analysis further provided insights into mechanisms responsible for immunosuppression and a potential target to control evolution of precancer and mitigate the risk for cancer development.

In recent years, the integration of single‐cell multi‐omics data has provided a more comprehensive understanding of cell functions and internal regulatory mechanisms from a non‐single omics perspective, but it still suffers many challenges, such as omics‐variance, sparsity, cell heterogeneity, and confounding factors. As it is known, the cell cycle is regarded as a confounder when analyzing other factors in single‐cell RNA‐seq data, but it is not clear how it will work on the integrated single‐cell multi‐omics data. Here, a cell cycle‐aware network (CCAN) is developed to remove cell cycle effects from the integrated single‐cell multi‐omics data while keeping the cell type‐specific variations. This is the first computational model to study the cell‐cycle effects in the integration of single‐cell multi‐omics data. Validations on several benchmark datasets show the outstanding performance of CCAN in a variety of downstream analyses and applications, including removing cell cycle effects and batch effects of scRNA‐seq datasets from different protocols, integrating paired and unpaired scRNA‐seq and scATAC‐seq data, accurately transferring cell type labels from scRNA‐seq to scATAC‐seq data, and characterizing the differentiation process from hematopoietic stem cells to different lineages in the integration of differentiation data.

Single-cell RNA-sequencing (scRNA-seq) enables the characterization of transcriptomic profiles at the single-cell resolution with increasingly high throughput. However, it suffers from many sources of technical noises, including insufficient mRNA molecules that lead to excess false zero values, termed dropouts. Computational approaches have been proposed to recover the biologically meaningful expression by borrowing information from similar cells in the observed dataset. However, these methods suffer from oversmoothing and removal of natural cell-to-cell stochasticity in gene expression. Here, we propose the generative adversarial networks (GANs) for scRNA-seq imputation (scIGANs), which uses generated cells rather than observed cells to avoid these limitations and balances the performance between major and rare cell populations. Evaluations based on a variety of simulated and real scRNA-seq datasets show that scIGANs is effective for dropout imputation and enhances various downstream analysis. ScIGANs is robust to small datasets that have very few genes with low expression and/or cell-to-cell variance. ScIGANs works equally well on datasets from different scRNA-seq protocols and is scalable to datasets with over 100,000 cells. We demonstrated in many ways with compelling evidence that scIGANs is not only an application of GANs in omics data but also represents a competing imputation method for the scRNA-seq data.

The emergence of RNA velocity has enriched our understanding of the dynamic transcriptional landscape within individual cells. In light of this breakthrough, we embarked on integrating RNA velocity with cellular pseudotime inference, aiming to improve the prediction of cell orders along biological trajectories beyond existing methods. Here, we developed LVPT, a novel method for pseudotime and trajectory inference. LVPT introduces a lazy probability to indicate the probability that the cell stays in the original state and calculates the transition matrix based on RNA velocity to provide the probability and direction of cell differentiation. LVPT shows better and comparable performance of pseudotime inference compared with other existing methods on both simulated datasets with different structures and real datasets. The validation results were consistent with prior knowledge, indicating that LVPT is an accurate and efficient method for pseudotime inference.

Large language models (LLMs) are deep learning-based artificial intelligence models that have achieved remarkable success in natural language processing. Typically composed of neural networks with billions of parameters, they are trained on massive unlabeled datasets using self-supervised or semi-supervised learning. Beyond language, LLMs hold immense potential for addressing complex bioinformatics challenges. This review provides a comprehensive overview of transformer-based model applications in genomics, transcriptomics, proteomics, drug discovery, and single-cell analysis. We discuss critical components, including tokenization strategies for diverse biological data, transformer architectures, attention mechanisms, and pretraining approaches. We also survey currently available foundation models and their downstream applications across bioinformatics domains. Finally, we highlight major challenges that remain insufficiently addressed in prior reviews and outline future perspectives and design principles for next-generation biological language models, offering practical guidance for both users and developers.