An attention-based hybrid deep neural networks for accurate identification of transcription factor binding sites

摘要

Transcription factors (TF) control gene expression by binding to specific regions of DNA sequence. TF play an important role in various disease processes, and their identification helps in understanding underlying gene regulation leading to disease risk. Currently, the most powerful models used for the predicting binding sites between TF and DNA sequence from ChIP-Seq dataset are lagging in terms of good feature extraction capabilities. We propose two models named PCLAtt and TranAtt for the prediction of 690 TF-cell line pairs from DNA sequence data. PCLAtt consists of two sets of convolutional neural networks-bidirectional long short-term memory (CNN-BiLSTM) layers in parallel followed by a multi-head attention layer and weight-shared dense layer which all contribute towards extracting efficient features from DNA sequence. TranAtt consists of convolution layers of a pre-trained model along with a BiLSTM layer and attention layer. The convolutional layers of the model act as a motif scanner and the BiLSTM layer learns the regulatory grammar of the motifs. Further, the attention mechanism is applied to give more importance to those sequence regions of DNA that consist of transcription factor binding motifs thus resulting in better performance of the proposed models. PCLAtt outperformed other state-of-the-art methods like DeepSEA, DanQ, TBiNet and DeepATT in prediction of binding sites between TF and the DNA sequence.