DeepCTF: transcription factor binding specificity prediction using DNA sequence plus shape in an attention-based deep learning model

Within the domain of molecular biology research, the intricate regulation of transcription continues to present a challenging yet imperative area of study. According to recent scientific studies, the nucleotide double helix shape is a major factor in improving the accuracy and comprehensibility of Transcription Factor Binding Sites (TFBSs). Despite the significant growth in computational methods aiming to concurrently incorporate both DNA sequence and DNA shape features, devising an effective model remains a challenging and unresolved issue. In this paper, we proposed a deep learning prediction model for TFBSs using attention mechanism, convolutional, and RNN-based networks by incorporating the DNA sequence and shape data. Attention mechanisms recognise the long-range dependencies but encounter challenges in focusing on local feature details. On the other hand, convolutional operations are proficient at extracting local features but may inadvertently neglect global information. Recurrent Neural Networks (RNNs) capture long-term dependencies within sequences. We demonstrate that the ability to predict TFBSs is greatly improved by our proposed technique, DeepCTF, using 12 in-vitro datasets collected from Protein Binding Microarray (PBMs) compared to the other state-of-the-art models.