Abstract:

Accurate identification of transcription factor binding sites (TFBSs) at single-nucleotide resolution remains a central challenge in deciphering gene expression regulatory networks. To improve the performance of existing computational models for predicting TFBSs across different cell types, we presented a deep learning model integrating channel and spatial attention mechanisms. In this study, we trained and tested the model using a comprehensive dataset that included ChIP-seq data from 51 groups, involving 10 core transcription factors (e.g., CTCF, EGR1, FOXA1) across 13 human cell lines (e.g., A549, GM12878, H1-hESC), and DNase-seq data from 13 datasets. The results demonstrated that this model exhibited superior performance across 23 TF-cell type combinations, achieving a mean area under the receiver operating characteristic curve (AUROC) of 0.986, with 91% of samples yielding an AUROC above 0.970. Additionally, the mean area under the precision-recall curve (AUPRC) reached 0.169, over 1,000-fold higher than the random baseline 0.000156. When compared to state-of-the-art models in the field, such as FactorNet, Leopard, and DeepGRN, our model outperformed them in terms of AUROC on nine shared TF-cell type datasets. Visualization analyses further confirmed that our model enabled accurate identification of cell-type-specific TFBSs. This study provides an efficient computational framework for precise cross-cell-type TFBS prediction, thereby facilitating in-depth investigations into gene expression regulatory mechanisms and the molecular pathogenesis of related diseases.

Key words: transcription factor binding sites, attention mechanism, deep learning, single-nucleotide resolution, cross- cell prediction