Tianjiao Zhang

Motivation: Cancer, a disease of high complexity. Identifying cancer driver genes is fundamental for elucidating oncogenesis and promoting precision medicine. Currently, most approaches mainly focus on homogeneous gene networks and single-omics data, thereby mainly identifying coding driver genes while ignoring non-coding driver genes. Result: Thus, we introduced GMHAN, a novel framework based on HAN. Firstly, we integrated the three types of omics data of genes and PPI network topology feature, together with the multi-dimensional features of miRNAs. Afterwards, we used heterogeneous graph attention networks to obtain deep feature embeddings of genes and miRNAs. Finally, the deep feature embeddings are input multilayer perceptron to obtain the probability that genes and miRNAs being cancer drivers. In a comparative evaluation against seven methods, GMHAN demonstrates better performance across both pan-cancer and cancer-specific datasets, achieving higher scores in AUC and AUPR. It has confirmed its effectiveness in identifying carcinogenic drivers. Availability and implementation: The source code of GMHAN is available at: https://github.com/mping315/GMHAN and https://doi.org/10.5281/zenodo.20154736.

Motivation: The rapid growth of spatial transcriptomics data holds potential for deep understanding of spatial specificity and tissue heterogeneity. Recognizing spatial domains is a fundamental step for deciphering tissue functional architecture and dissecting tissue heterogeneity. However, existing models typically define adjacency relations using static weights, which cannot dynamically adjust neighbor importance based on expression context, thereby limiting the accuracy and robustness of spatial domain recognition. Results: We propose spAttClu, a clustering model integrating spatially weighted graph attention with contrastive learning. It adaptively learns neighbor contributions in varying contexts through a distance-weighted graph attention mechanism and enhances embedding discriminability via multi-level contrastive learning. spAttClu demonstrates superior clustering performance on the DLPFC dataset. Moreover, it shows cross-platform adaptability and enables vertical/horizontal inte-gration of multiple tissue slices.

Cancer, as a complex disease with high heterogeneity, is gradually shifting from standardization to individualization. Current research indicates that integrating multi-omics data, including genome, transcriptome, and epigenome, can more comprehensively reveal the molecular features of tumors, thereby providing a more reliable basis for subtype classification. It is worth noting that driver genes, as key factors directly involved in tumorigenesis and development, when combined with multi-omics data, improve model interpretability. We proposed a cancer subtype classification model named MOTCS, which consists of four key steps: (1) The extraction of driver gene features from existing databases and the execution of preliminary feature selection for non-driver genes in different omics data; (2) the use of the transformer encoder to conduct feature representation learning for non-driver gene features and driver gene features, respectively; (3) the non-driver gene feature embeddings and driver gene feature embeddings of each omics learned are concatenated and input into the MLP classifier to obtain the classification probabilities of each subtype. (4) By integrating the above features through view association discovery networks, cross-omics feature representations are learned so as to achieve effective classification. The results show that MOTCS outperformed existing methods across all four datasets and that the introduction of driver genes will enhance the classification performance of MOTCS. MOTCS achieved precise classification of cancer subtypes by mining the feature embeddings of driver genes and non-driver gene features.

The high heterogeneity of cancer poses significant challenges for precision diagnosis, particularly in tasks such as rare subtype identification, early lesion detection, and tumor grading. Notably, cancer cell biological traits are closely correlated with cell death regulatory mechanisms, and accurate cancer region identification is a pivotal premise for exploring the cancer-cell death intrinsic association. Single-modality methods often struggle to balance sensitivity and accuracy, while existing general multimodal models are poorly adapted to the processing of pathological long texts and high-resolution images, leading to issues of semantic truncation and feature loss. To address these challenges, this study proposes CanLRHI, a multimodal pretraining model tailored for cancer pathology that focuses on the synergistic modeling of long pathological reports and high-resolution images to achieve comprehensive cross-modal alignment and accurate characterization of cancer regions. Experimental results on the CancerPath-170 K-v1 dataset, which contains 170 000 cancer pathology image–text pairs, demonstrate that CanLRHI significantly outperforms mainstream multimodal baselines across various tasks, including Zero-Shot classification and Few-Shot Fine-Tuning. This work provides an extensible technical framework for long-text-driven cross-modal representation learning in medical pathology, and further offers a reliable technical support for cell death-related cancer pathology research via high-precision cancer region detection.

Protein function prediction plays an important role in the field of biology. With the wide application of deep learning in the field of bioinformatics, more and more natural language processing (NLP) technologies are applied to the downstream tasks in the field of bioinformatics, and it has also shown excellent performance in protein function prediction. Protein-protein interaction (PPI) networks and other biological attributes contain rich information critical for annotating protein functions. However, existing deep learning networks still suffer from overfitting and noise issues, resulting in low accuracy in protein function prediction. Consequently, developing efficient models for protein function prediction remains a popular and challenging topic in the application of NLP in bioinformatics. In this study, we propose a novel protein function prediction model based on attention mechanisms, termed IMAGO. This model employs the Transformer pre-training process, integrating multi-head attention mechanisms and regularization techniques, and optimizes the loss function to effectively reduce overfitting and noise issues during training. It generates more robust embeddings, ultimately improving the accuracy of protein function prediction. Experimental results on human and mouse datasets indicate that our model surpasses other protein function prediction models across multiple metrics. Thus, this efficient, stable, and accurate deep learning model holds significant promise for protein function prediction.

Single-cell RNA sequencing enables transcriptome-wide analysis at single-cell resolution, offering unprecedented insights into cellular heterogeneity across biological conditions. However, accurately comparing transcriptomic distributions of cells from distinct biological states, such as healthy versus diseased individuals, remains challenging. To address this, we developed CTDP, a robust and interpretable computational framework that identifies disease phenotype-associated cell types of interest by integrating Lasso-regularized logistic regression with permutation testing. Through comprehensive evaluations on both simulated and real-world datasets, including melanoma immunotherapy, COVID-19 severity, and liver cirrhosis, CTDP consistently outperformed existing methods such as DA-seq, scDist, and PENCIL in both accuracy and robustness. In melanoma, CTDP uncovered immune-responsive clusters and revealed transcriptional regulators like PTPRC, CREM, and JUNB linked to immunotherapy efficacy. In COVID-19, it identified critical severity-associated cell types, such as B cells, NK cells, epithelial cells, and macrophages, which contribute to dysregulated immune responses and inflammation in severe cases. These results highlight CTDP’s power in uncovering disease-relevant cell populations and its potential to advance precision medicine through single-cell analysis.

Motivation: Identifying cell types that constitute complex tissue components using single-cell sequencing data is a critical issue in the field of biology. With the continuous advancement of sequencing technologies, the recognition of cell types has evolved from analyzing single-omics scRNA-seq data to integrating multi-omics single-cell data. However, existing methods for integrative analysis of high-dimensional multi-omics single-cell sequencing data have several limitations, including reliance on specific distribution assumptions of the data, sensitivity to noise, and clustering accuracy constrained by independent clustering methods. These issues have restricted improvements in the accuracy of cell type identification and hindered the application of such methods to large-scale datasets for cell type recognition. To address these challenges, we propose a novel method for aligning and integrating single-cell multi-omics data—scECDA. Results: The scECDA employs independently designed autoencoders that can autonomously learn the feature distributions of each omics dataset. By incorporating enhanced contrastive learning and differential attention mechanisms, the scECDA effectively reduces the interference of noise during data integration. The model design exhibits high flexibility, enabling adaptation to single-cell omics data generated by different technological platforms. It directly outputs integrated latent features and end-to-end cell clustering results. Through the analysis of the distribution of latent features, the scECDA can effectively identify key biological markers and precisely distinguish cell subtypes, recover cluster-specific motif and infer trajectory. The scECDA was applied to eight paired single-cell multi-omics datasets, covering data generated by 10X Multiome, CITE-seq, and TEA-seq technologies. Compared to eight state-of-the-art methods, scECDA demonstrated higher accuracy in cell clustering. Availability and implementation: The scECDA code is freely available at https://github.com/SuperheroBetter/scECDA

The rapid advancement of spatial transcriptomics has provided a critical data foundation for the high‐resolution characterization of tissue spatial domains. Traditional methods for spatial domain identification primarily rely on gene expression data from sampled spots in low‐resolution spatial transcriptomic data, often overlooking valuable information between spots that can be crucial for domain identification. Furthermore, these methods are limited by their focus on gene expression data from neighboring spots, without fully integrating prior knowledge of cell types within the tissue's spatial structure. To address these challenges, SGCD, a novel method for tissue spatial domain identification based on data interpolation and cell type deconvolution is proposed. SGCD utilizes interpolation techniques to estimate gene expression data for cells in the gaps between spots and applies deconvolution to extract cell type information from both spots and interstitial regions. By integrating gene expression, cell type, and spatial location data, SGCD achieves accurate delineation of complex spatial domains through graph contrastive learning. Evaluations on various publicly available datasets, including the human dorsolateral prefrontal cortex, mouse brain, pancreatic ductal adenocarcinoma, and breast cancer, demonstrate that SGCD significantly outperforms existing methods in both accuracy and detail, offering strong support for advancing the understanding of tissue functions and disease mechanisms.

Enhancer–promoter interaction (EPI) is a critical component of gene regulation. Accurately predicting EPIs across diverse cell types can advance our understanding of the molecular mechanisms behind transcriptional regulation and provide valuable insights into the onset and progression of related diseases. At present, large-scale genome-wide EPI predictions typically rely on computational approaches. However, most of these methods focus on predicting EPIs within a single cell line and lack a global perspective encompassing multiple cell lines. Furthermore, they often fail to fully account for the nonlinear relationships between features, leading to suboptimal prediction accuracy. In this study, we propose KansformerEPI, a global EPI prediction model designed for multiple cell lines. The model is built on Kansformer, an encoder that integrates KAN and Transformer, effectively capturing the nonlinear relationships among various epigenetic and sequence features. We utilized KansformerEPI to achieve cross-tissue prediction of EPIs across different cell types. This approach enhances the model’s scalability, eliminating the complexity of designing separate prediction models for individual tissues. As a result, our model is applicable to various tissues, thereby reducing dependency on extensive datasets. Experimental results demonstrate that KansformerEPI surpasses existing methods such as TransEPI, TargetFinder, and SPEID in both accuracy and stability of EPI predictions across datasets including HMEC, IMR90, K562, and NHEK.

The growth, development, and differentiation of multicellular organisms are primarily driven by intercellular communication, which coordinates the activities of diverse cell types. This cell-to-cell signaling is typically mediated by various types of protein–protein interactions, including ligand–receptor; receptor–receptor, and extracellular matrix–receptor interactions. Currently, computational methods for inferring ligand–receptor communication primarily depend on gene expression data of ligand–receptor pairs and spatial information of cells. Some approaches integrate protein complexes; transcription factors; or pathway information to construct cell communication networks. However, few methods consider the critical role of protein–protein interactions (PPIs) in intercellular communication networks, especially when predicting communication between different cell types in the absence of cell type information. These methods often rely on ligand–receptor pairs that lack PPI evidence, potentially compromising the accuracy of their predictions. To address this issue, we propose CellGAT, a framework that infers intercellular communication by integrating gene expression data of ligand–receptor pairs, PPI information, protein complex data, and experimentally validated pathway information. CellGAT not only builds a priori models but also uses node embedding algorithms and graph attention networks to build cell communication networks based on scRNA-seq (single-cell RNA sequencing) datasets and includes a built-in cell clustering algorithm. Through comparisons with various methods, CellGAT accurately predicts cell–cell communication (CCC) and analyzes its impact on downstream pathways; neighboring cells; and drug interventions.