Bidirectional LSTM and CNN model for predicting DNA-protein binding

Abstract

The invention provides a bidirectional LSTM and CNN model for predicting DNA-protein binding. The model includes an input layer, a BLSTM layer, a convolutional layer, a maximum pooling layer, a full connection layer, and an output layer. Each input sequence is expressed as a four-line binary matrix by the input layer through single thermal coding. In the BLSTM layer, each LSTM model in a previouslayer will receive information of interest on DNA from an input sequence, and encode and interpret contributions from past historical information to a hidden state; then, the BLSTM module is propagated to the next BLSTM module, wherein a matrix scanned and input by each convolution kernel in the convolution layer is used for motif discovery, and information with different intensities is associatedwith potential sequence patterns; the maximum pooling layer is used for maximizing an output signal of each convolution kernel to form a complete sequence; the output layer performs non-linear conversion to determine DNA-protein binding feature information.

Description

Technical field [0001] The present invention relates to the field of deep learning and biological information technology, in particular to a two-way LSTM and CNN model for predicting DNA-protein binding. Background technique [0002] Accurately establishing a specific model of transcription factor (TF) sequence is a basic problem for understanding genome function and evolution. In particular, the binding properties of transcription factors will have a decisive effect on downstream gene expression. With the development of high-throughput sequencing technology, the ENCODE project provides the binding specificity of 187 TFs in the whole genome of 98 cells. According to transcription factor binding sites, a binary classification problem of a sequence can be defined, that is, according to whether TF is bound, it is divided into positive samples and negative samples. By establishing a binary classification model of sequences, the binding sites of new samples can be predicted. [0003]...

AI Technical Summary

Problems solved by technology

[0004] However, existing solutions have the following disadvantages: the framework of the neural network will greatly affect the system performance of the model, and at the same time perform worse in capturing the position and dynamics of the probe sequence in order to predict the TF binding site from the DNA sequence

Therefore, the accuracy of existing computational methods in predicting DNA-protein binding needs to be improved

Images