BLSTM Architecture

As Figure 4 shows, a sequence of numbers with the sequence encoding scheme with the length 198 followed by the body of the structure is used as input to BLSTM. It is mainly composed of an embedding layer, a bidirectional LSTM, a dropout layer, the rectified linear unit (relu), a dropout layer, and sigmoid activation functions. In the architecture, the main purpose of embedding term training is to incorporate into the model to form an end-to-end structure, and the vector trained by the embedding layer can better adapt to the corresponding tasks (Kleinjan and Lettice, 2008; Liu G. et al., 2016, 2018; Liu et al., 2017; Zhang T. et al., 2020). The recurrent neural network (RNN) is a network of nerves that processes sequential data. Compared with the ordinary neural network, it can process the sequence variation data (Zou et al., 2016; Zacher et al., 2017). Long short-term memory (LSTM) is a special RNN, and it is mainly used to solve the problem of gradient explosion and disappearance. In short, LSTM performs better than normal RNN if the sequence is long (Liu et al., 2019; Zou et al., 2019; Yang et al., 2020; Zhang Z. Y. et al., 2020). Bidirectional LSTM is equivalent to the LSTM upgraded version, which means that time sequence data are used to input history and future data simultaneously. In contrast to time sequence, two cyclic neural networks are connected to the same output, and the output layer can obtain historical and future information at the same time (Bian et al., 2014; Goldberg and Levy, 2014; Juntao and Zou, unpublished; Tang et al., 2014). The function of dropout layer is preventing model overfitting. In addition, after relu and sigmoid layers (Gers et al., 1999; Graves and Schmidhuber, 2005; Sundermeyer et al., 2012; Zaremba et al., 2014; Huang et al., 2015; Xingjian et al., 2015; Li and Liu, 2020; Sherstinsky, 2020), a probability of whether the sequence is an enhancer or not can be calculated.

Architecture of the BLSTM model.