"transcription factor binding site prediction"

"Accurate prediction of transcription factor binding sites (TFBSs) is crucial for understanding gene regulatory mechanisms. However, existing computational methods typically rely on linear sequence modeling, overlooking the potential complex structures and semantic information within DNA sequences, which limits their ability to capture key binding regions. Therefore, we propose MSGSNet, a computational framework for TFBS prediction that integrates linear sequence features with graph topological structures. MSGSNet encodes DNA sequences using stacked codon encoding and extracts deep semantic features via multi-scale convolutions and Bi-LSTM, with hierarchical bidirectional attention enhancing regulatory feature learning. To further capture topological dependencies among k-mers, MSGSNet constructs multi-scale De Bruijn graphs and leverages graph convolutional and attention mechanisms to effectively learn graph structural representations. Experimental results on 165 ChIP-seq (chromatin immunoprecipitation sequencing) benchmark datasets demonstrate that MSGSNet outperforms existing computational methods. Cross-cell line and cross-transcription factor evaluations further demonstrate that MSGSNet maintains superior robustness under conditions of high data heterogeneity and weak peak signals. In addition, we conducted both global and local interpretability analyses of the latent representations learned by MSGSNet using SHAP values and LIME, and further examined the sequence regions attended to by the model through hierarchical bidirectional attention analysis. The results demonstrate that MSGSNet can effectively identify biologically meaningful core motifs and long-range contextual regions."