DeepUTF: Locating transcription factor binding sites via interpretable dual-channel encoder-decoder structure

•A novel method (DeepUTF) to predict transcription factor binding sites (TFBSs).•DeepUTF captures other relevant patterns in the transcriptional regulatory network.•The feature learning process of the model is interpretable.•Excellent model predictions compared to TF-DNA complexes in the PDB.•DeepUT...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Pattern recognition Ročník 161; s. 111279
Hlavní autoři: Ding, Pengju, Wang, Jianxin, He, Shiyue, Gao, Xin, Yu, Xu, Yu, Bin
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.05.2025
Témata:
Improved encoder-decoder
Interpretability
Swin transformer
Transcription factor binding sites
Swin transformer
Transcription factor binding sites
Improved encoder-decoder
Interpretability
ISSN:0031-3203
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•A novel method (DeepUTF) to predict transcription factor binding sites (TFBSs).•DeepUTF captures other relevant patterns in the transcriptional regulatory network.•The feature learning process of the model is interpretable.•Excellent model predictions compared to TF-DNA complexes in the PDB.•DeepUTF improves prediction performance compared to state-of-the-art methods. The accurate location of transcription factor binding sites (TFBSs) is important for the design of synthetic biology components and the realization of precision medicine. Despite the growing use of deep learning for TFBSs prediction, model interpretability remains challenging. We introduce DeepUTF, a novel architecture integrating improved encoder-decoder, swin transformer, and parallel Bi-LSTM, which realizes precise localization of TFBSs and prediction of motifs. We elucidate the effectiveness of the swin transformer in capturing a wide range of dependencies and emphasizing the learning of critical features. Meanwhile, interpretability analysis of the model's output and predictions of TF-DNA binding motifs are conducted, providing a thorough exploration of the model's intrinsic mechanisms and feature learning process. Experiments conducted on 53 ChIP-seq datasets illustrate that DeepUTF surpasses several leading algorithms. The trained model accurately predicts direct and indirect TF-DNA binding motifs. Furthermore, comparisons with the PDB database validate the continuity and accuracy of predictions.
ISSN:0031-3203
DOI:10.1016/j.patcog.2024.111279