BACKGROUND : Identifying drug–target interactions (DTIs) is essential in drug discovery and repositioning. Recently, deep learning has become the mainstream methodology for DTI prediction. However, the scarcity of three-dimensional structural data has forced almost all methods to predict drug-target interactions with low-dimensional data, thereby constraining their overall performance. METHODS : In tackling this challenge, we introduce a novel approach, CLSF-DTI. CLSF-DTI incorporates high-dimensional structural and functional information into the drug and protein features through contrastive learning during the feature extraction stage. This ensures that the model no longer solely focuses on sequence information, leading to a more precise modeling outcome. RESULTS : Experiments on five benchmark datasets demonstrate that CLSF-DTI achieves the best overall performance among five state-of-the-art baselines. Through ablation studies, we further prove that the contrastive module enhances the predictive performance and generalization ability of CLSF-DTI. Moreover, CLSF-DTI successfully identified some ligands for the protein PKA-Cα in drug screening experiments. CONCLUSIONS : This study proposed a contrastive learning model CLSF-DTI that integrates structural and functional similarity. It outperforms existing methods in drug-target interaction prediction and has stronger generalization ability. However, the handling of unbalanced data and long-distance dependencies still needs to be improved in the future. The data and source code are available at https://github.com/ZhangLab312/CLSF_DTI .

Single-cell RNA sequencing (scRNA-seq) is entering an era of foundation models that accept the complete gene atlas as input, yet most current datasets cover only 10-12 k genes and contain numerous technical zeros, severely limiting the generalization of these models in downstream tasks. To address this, we pioneer the gene-completion task for scRNA-seq and present SAD, a diffusion-based framework tailored to extremely sparse data, capable of completing genes and correcting sparsity bias under high missing rates. Unlike imputation or reconstruction methods that rely on the i.i.d. assumption, SAD's completion paradigm can generate gene entries originally absent from the expression profile, be aware of and rectify sparsity-distribution bias, and supply foundation models with consistent, reliable inputs of more than 30 k genes. Extensive benchmarks show that SAD significantly outperforms existing methods across multiple completion metrics, particularly in extreme scenarios with missing rates above 80%. This provides a data foundation for reusing missing scRNA-seq information and for precision-medicine applications.

Single-cell type annotation is crucial for understanding the tumor microenvironment. With advances in single-cell sequencing, researchers can now analyze the functional characteristics of different cell types at the single-cell level. However, high-dimensional sparsity and technical noise in single-cell data pose challenges for existing annotation methods. Traditional approaches rely mainly on transcriptomic data, neglecting complementary multi-omics information, leading to information loss and reduced accuracy. To address these limitations, this study proposes a novel single-cell annotation method based on multimodal alignment. By designing two alignment strategies—inter-modal feature alignment and inter-model knowledge alignment—the method integrates scRNA-seq and scATAC-seq data. The inter-modal feature alignment strategy corrects distributional discrepancies between modalities, enabling seamless mapping in the latent space and more accurately capturing multi-omics characteristics. The inter-model knowledge alignment strategy transfers structured knowledge from a teacher model to a student model, balancing computational efficiency and annotation performance. This approach reduces reliance on single modalities in traditional methods and enhances adaptability to diverse experimental conditions. Experiments on public datasets show that the proposed method outperforms state-of-the-art single-omics models across multiple benchmarks, demonstrating strong annotation capabilities and superior generalization to rare cell types.

Multiomics integration of single-cell datasets generated from multiple omics technologies is crucial for defining cellular heterogeneity. Mosaic data contains any combination of different modalities and batches of data, which brings challenges such as incomplete reference samples, non-overlapping features, and batch effects, making it difficult to extract information from them. Most existing methods require a reference batch that contains all modalities, which limits the versatility of mosaic data. In this study, we proposed a single-cell mosaic data integration method that guides variational cross-modal fusion via regulation graphs (GVMOS). Based on mosaic data, GVMOS first generates a regulatory graph of cross-modal feature interactions. Then, a coupled-trained dual variational autoencoder module is used to extract the complementary and specific information in different modal data and the cross-modal feature topology information in the regulatory graph. By integrating modality information with feature information, GVMOS can guide modality integration from the perspective of spatial semantics. Experiments on three bimodal paired datasets and five bimodal mosaic combinations show that GVMOS significantly improves the integration performance of existing methods, with an average improvement of 3% in overall scores.

Discovering effective combinations of anticancer drugs is crucial for improving cancer treatment strategies. The accumulation of drug information and cell line data contributes to the development of effective deep learning prediction models for drug synergy. However, selecting high-quality data sources and designing appropriate methods remains a challenge. This article proposes an attention-based multimodal heterogeneous information fusion network MHFSyn for predicting drug synergy. MHFSyn uses multiple feature extractors to extract different modality features, and captures cross modal interactions and structural information through an attention fusion network, thereby achieving effective fusion of multimodal data and improving the overall performance and generalization ability of the model. Five fold cross-validation and leave-one-out cross-validation experiments show that MHFSyn had the best overall performance compared to the six comparison methods.Our code is available at https://github.com/ZhangLab312/MHFSyn.

The rapid accumulation of single-cell RNA sequencing data has provided unprecedented computational resources for cell type annotation, significantly advancing our understanding of cellular heterogeneity. Leveraging gene expression profiles derived from transcriptomic data, researchers can accurately infer cell types, sparking the development of numerous innovative annotation methods. These methods utilize a range of strategies, including marker genes, correlation-based matching, and supervised learning, to classify cell types. In this review, we systematically examine these annotation approaches based on transcriptomics-specific gene expression profiles and provide a comprehensive comparison and categorization of these methods. Furthermore, we focus on the main challenges in the annotation process, especially the long-tail distribution problem arising from data imbalance in rare cell types. We discuss the potential of deep learning techniques to address these issues and enhance model capability in recognizing novel cell types within an open-world framework.

The Gene Regulatory Networks (GRNs) ensure the stability of cellular states, preserving specific phenotypes and functions throughout the differentiation process. However, current tools still need improvement to effectively integrate multi-omics data and infer GRNs for particular cell types. We introduce CLMOGRI, a multi-omics TF-gene regulatory network inference framework based on heterogeneous networks and contrastive learning, designed to integrate multi-omics data for GRN inference. Through random walk techniques, CLMOGRI embeds multi-omics data into a unified feature space and extracts similar features between nodes. It then measures node similarity and predicts node relationships by contrastive learning. Finally, it includes a regulatory network interpreter to identify critical nodes and modules in GRNs, offering an analytical method for understanding complex interactions within biological systems. CLMOGRI surpasses existing baseline methods in terms of Area Under the Precision-Recall Curve (AUPR) and F-Score metrics, indicating its efficacy in capturing multi-omics information for GRN inference. It also reveals vital nodes and modules within the gene regulatory network, improving the interpretability of CLMOGRI and the utility of GRNs.

Self-supervised learning has succeeded significantly in cell type annotation based on transcriptomic data. However, existing methods encode transcriptomic data as gene feature representations, limiting self-supervised learning to the gene level. This paper proposes a novel self-supervised distillation learning framework, scTCHCN, which designs a cell-specific highly correlated network. Firstly, this framework integrates the cell-specific highly correlated network with single-cell transcriptomic data to extract cell and gene features. ScTCHCN promotes interactions between the network and the transcriptomic data features during pre-training by constructing the cell-specific highly correlated network. Additionally, a self-supervised paradigm with distillation learning is proposed to enhance the global feature correlation between the two. Ultimately, the pre-trained student model aggregates deep feature constraints and complementary information at the single-cell and gene levels. It combines them with the prediction layer to construct the final cell type prediction model. Experimental results demonstrate that scTCHCN excels in cell type annotation and rare cell identification tasks, showcasing its potential for other applications. The source code and additional information for this study are available at https://github.com/ZhangLab312/scTCHCN.

Accurate identification of cell types is a pivotal and intricate task in scRNA-seq data analysis. Recently, significant strides have been made in cell type annotation of scRNA-seq data using pre-trained language models (PLMs). This method has surmounted the constraints of conventional approaches regarding precision, robustness, and generalization. However, the fine-tuning process of large-scale pre-trained models incurs substantial computational expenses. To tackle this issue, a promising avenue of research has emerged, proposing parameter-efficient fine-tuning techniques for PLMs. These techniques concentrate on fine-tuning only a small portion of the model parameters while attaining comparable performance. In this study, we extensively research parameter-efficient fine-tuning methods for scRNA-seq cell type annotation, employing scBERT as the backbone. We scrutinize the performance and compatibility of various parameter-efficient fine-tuning methodologies across multiple datasets. Through comprehensive analysis, we demonstrate the remarkable performance of parameter-efficient fine-tuning methods in cell type annotation. Hopefully, this study can inspire new thinking in analyzing scRNA-seq data.

Accurately predicting cancer prognosis remains challenging, owing to the combination of computational and practical challenges. This study proposes KDProg, a knowledge distillation-based graph learning framework for predicting cancer prognosis and exploring downstream tasks. The framework includes a novel feature distillation paradigm that compresses a multi-layer complex teacher model to a single-layer simple student by using the teacher model’s middle-layer feature representations and outputs as supervision information to improve the student model’s performance. In addition, instead of introducing a unified temperature hyperparameter, KDProg adopts a novel strategy to parameterize the distillation temperature and combine it with the Cox partial log-likelihood function. So the model can learn the appropriate temperature. Furthermore, considering multi-omics data of patients are often complex to obtain in practical cancer prognosis, this paper uses different input data for the teacher and student models, respectively. The input data for the teacher model are multi-omics data (mRNA, CNV, and DNA methylation), clinical data, and KEGG pathways. The input data for the student model are mRNA, clinical data, and KEGG pathways. Extensive experiments on 15 real-world datasets from TCGA demonstrated the effectiveness and efficiency of the proposed method in predicting cancer prognosis. The results suggest that the proposed model can guide clinical decision-making.

Accurately predicting cancer driver genes remains challenging due to the increasing size and complexity of cancer genomic data. In this study, HGTDG is proposed, a heterogeneous graph transformer framework for predicting cancer driver genes and exploring downstream tasks. The framework includes a heterogeneous graph construction module that constructs a gene-protein heterogeneous network based on KEGG pathways and the protein-protein interactions from the STRING database. In addition, the framework introduces a novel heterogeneous graph transformer module that uses multi-head attention mechanisms for gene node embedding. The transformer module can capture dedicated representations for genes and edges. Finally, the generated gene embeddings are fed into the classification module to classify genes into driver and non-driver genes. The experiment results show that HGTDG outperforms the state-of-the-art methods regarding the area under the receiver operating characteristic curves (AUROC) and the area under the precision-recall curves (AUPRC).

Enhancers are crucial cis-regulatory elements that control gene expression in a cell-type-specific manner. Despite extensive genetic and computational studies, accurately predicting enhancer activity in different cell types remains a challenge, and the grammar of enhancers is still poorly understood. Here, we present HEAP (high-resolution enhancer activity prediction), an explainable deep learning framework for predicting enhancers and exploring enhancer grammar. The framework includes three modules that use grammar-based reasoning for enhancer prediction. The algorithm can incorporate DNA sequences and epigenetic modifications to obtain better accuracy. We use a novel two-step multi-task learning method, task adaptive parameter sharing (TAPS), to efficiently predict enhancers in different cell types. We first train a shared model with all cell-type datasets. Then we adapt to specific tasks by adding several task-specific subset layers. Experiments demonstrate that HEAP outperforms published methods and showcases the effectiveness of the TAPS, especially for those with limited training samples. Notably, the explainable framework HEAP utilizes post-hoc interpretation to provide insights into the prediction mechanisms from three perspectives: data, model architecture and algorithm, leading to a better understanding of model decisions and enhancer grammar. To the best of our knowledge, HEAP will be a valuable tool for insight into the complex mechanisms of enhancer activity.

MOTIVATION: Multiple sequence alignment (MSA) is one of the hotspots of current research and is commonly used in sequence analysis scenarios. However, there is no lasting solution for MSA because it is a Nondeterministic Polynomially complete problem, and the existing methods still have room to improve the accuracy. RESULTS: We propose Deep reinforcement learning with Positional encoding and self-Attention for MSA, based on deep reinforcement learning, to enhance the accuracy of the alignment Specifically, inspired by the translation technique in natural language processing, we introduce self-attention and positional encoding to improve accuracy and reliability. Firstly, positional encoding encodes the position of the sequence to prevent the loss of nucleotide position information. Secondly, the self-attention model is used to extract the key features of the sequence. Then input the features into a multi-layer perceptron, which can calculate the insertion position of the gap according to the features. In addition, a novel reinforcement learning environment is designed to convert the classic progressive alignment into progressive column alignment, gradually generating each column's sub-alignment. Finally, merge the sub-alignment into the complete alignment. Extensive experiments based on several datasets validate our method's effectiveness for MSA, outperforming some state-of-the-art methods in terms of the Sum-of-pairs and Column scores. AVAILABILITY AND IMPLEMENTATION: The process is implemented in Python and available as open-source software from https://github.com/ZhangLab312/DPAMSA.

MOTIVATION: Transcription factor (TF) binds to conservative DNA binding sites in different cellular environments and development stages by physical interaction with interdependent nucleotides. However, systematic computational characterization of the relationship between higher-order nucleotide dependency and TF-DNA binding mechanism in diverse cell types remains challenging. RESULTS: Here, we propose a novel multi-task learning framework HAMPLE to simultaneously predict TF binding sites (TFBS) in distinct cell types by characterizing higher-order nucleotide dependencies. Specifically, HAMPLE first represents a DNA sequence through three higher-order nucleotide dependencies, including k-mer encoding, DNA shape and histone modification. Then, HAMPLE uses the customized gate control and the channel attention convolutional architecture to further capture cell-type-specific and cell-type-shared DNA binding motifs and epigenomic languages. Finally, HAMPLE exploits the joint loss function to optimize the TFBS prediction for different cell types in an end-to-end manner. Extensive experimental results on seven datasets demonstrate that HAMPLE significantly outperforms the state-of-the-art approaches in terms of auROC. In addition, feature importance analysis illustrates that k-mer encoding, DNA shape, and histone modification have predictive power for TF-DNA binding in different cellular environments and are complementary to each other. Furthermore, ablation study, and interpretable analysis validate the effectiveness of the customized gate control and the channel attention convolutional architecture in characterizing higher-order nucleotide dependencies. AVAILABILITY AND IMPLEMENTATION: The source code is available at https://github.com/ZhangLab312/Hample.

Cell type-specific gene expressions during development or in disease are regulated by interactions between transcription factors (TFs) and their binding sites. Recently, many deep learning approaches have been developed to characterize TF-DNA binding within a population of cells. However, determining TF binding sites (TFBSs) in single cells remains challenging due to the sparsity of data. Here, we propose a multi-stage transfer learning framework called STAPLE for single-cell TF-DNA binding prediction and analysis. Specifically, we design the Cell Type Learning to capture the relationship between different TF-DNA binding events in the same cell type. Meanwhile, we present Individual Learning to extract common motif and chromatin accessibility features of a particular binding event in a cellular population. In addition, we leverage Single-cell Learning to annotate TFBSs in each cell without any supervised label. Extensive experiments based on 570 single-cell datasets validate the effectiveness of our framework for considering cellular heterogeneity, outperforming current methods. This work can provide new insight into the relationship between TF-DNA binding and cellular heterogeneity. The source code of STAPLE can be found at https://github.com/ZhangLab312/STAPLE.

Interpreting the regulatory code of gene expression and further understanding the functionality of noncoding variants on transcriptional effect is a crucial challenge. However, this remains difficult due to the complex association between SNPs and chromatin state. Here, we develop a meta-learning-based framework, U-TransNet, that can accurately predict the TF-DNA binding based on multiple chromatin features. Motivated by ab initio, our proposed framework contain two steps. (i) Meta-learning strategy is applied to predict chromatin profiles from DNA sequence. (ii) DNA sequence and all of these predicted chromatin features are used to predict TF-DNA binding affinity. Experiments demonstrate that U-TransNet has excellent performance, achieving significant improvements over existing methods in predicting base-resolution TF-DNA binding signals, TF binding sites, and motifs. We also demonstrate that integrating the more extended TFBS flank regions is a potential path to better understanding gene transcription. In addition, U-TransNet is applied to infer the effects of variants on TF-DNA binding affinity via in silico mutagenesis, and further to identify cell type-specific functional variants via comparing different cells. To the best of the authors’ knowledge, U-TransNet provides an efficient end-to-end computational framework for deciphering cis-regulator evolution.

The discovery of putative transcription factor binding sites (TFBSs) is important for understanding the underlying binding mechanism and cellular functions. Recently, many computational methods have been proposed to jointly account for DNA sequence and shape properties in TFBSs prediction. However, these methods fail to fully utilize the latent features derived from both sequence and shape profiles and have limitation in interpretability and knowledge discovery. To this end, we present a novel Deep Convolution Attention network combining Sequence and Shape, dubbed as D-SSCA, for precisely predicting putative TFBSs. Experiments conducted on 165 ENCODE ChIP-seq datasets reveal that D-SSCA significantly outperforms several state-of-the-art methods in predicting TFBSs, and justify the utility of channel attention module for feature refinements. Besides, the thorough analysis about the contribution of five shapes to TFBSs prediction demonstrates that shape features can improve the predictive power for transcription factors-DNA binding. Furthermore, D-SSCA can realize the cross-cell line prediction of TFBSs, indicating the occupancy of common interplay patterns concerning both sequence and shape across various cell lines. The source code of D-SSCA can be found at https://github.com/MoonLord0525/.

Transcription factors (TFs) binding sites prediction and analysis are vital for comprehending cis-regulatory mechanisms. Recently, several deep learning-based methods have shown outstanding performance on TFs binding sites (TFBSs) recognition by leveraging solely base-pair arrangement of regulatory sequences. Except for the aforementioned genomic features, the epigenomics signature represented by the histone modification is also a critical factor related to TFs-DNA binding. We present a multi-omics based hybrid neural network, dubbed as BHSite, for TFBSs prediction by adaptively integrating base-pair arrangements and histone modification signatures. Experiments over 196 ChIP-seq datasets demonstrate that BHSite significantly outperforms several state-of-the-art methods in TFBSs prediction. Besides, studies of the relative importance of histone modification signatures prove that diverse signatures complement each other. Furthermore, visualization analysis of Squeeze-and-Excitation Network reveals the contribution of multi-omics latent features concerning different cell types to TFBS prediction. Thus, BHSite improves both performance and interpretability by combining the multi-omic features into deep learning architecture.

Transcription factors (TFs) are essential proteins in regulating the spatiotemporal expression of genes. It is crucial to infer the potential transcription factor binding sites (TFBSs) with high resolution to promote biology and realize precision medicine. Recently, deep learning-based models have shown exemplary performance in the prediction of TFBSs at the base-pair level. However, the previous models fail to integrate nucleotide position information and semantic information without noisy responses. Thus, there is still room for improvement. Moreover, both the inner mechanism and prediction results of these models are challenging to interpret. To this end, the Deep Attentive Encoder-Decoder Neural Network (D-AEDNet) is developed to identify the location of TFs-DNA binding sites in DNA sequences. In particular, our model adopts Skip Architecture to leverage the nucleotide position information in the encoder and removes noisy responses in the information fusion process by Attention Gate. Simultaneously, the Transcription Factor Motif Discovery based on Sliding Window (TF-MoDSW), an approach to discover TFs-DNA binding motifs by utilizing the output of neural networks, is proposed to understand the biological meaning of the predicted result. On ChIP-exo datasets, experimental results show that D-AEDNet has better performance than competing methods. Besides, we authenticate that Attention Gate can improve the interpretability of our model by ways of visualization analysis. Furthermore, we confirm that ability of D-AEDNet to learn TFs-DNA binding motifs outperform the state-of-the-art methods and availability of TF-MoDSW to discover biological sequence motifs in TFs-DNA interaction by conducting experiment on ChIP-seq datasets.