Identification of DNA-protein binding sites by bootstrap multiple convolutional neural networks on sequence information

Abstract Identification of DNA–protein binding sites in protein sequence plays an essential role in a wide variety of biological processes. In particular, there are huge volumes of protein sequences accumulated in the post-genomic era. In this study, we propose a new prediction approach appropriate for imbalanced DNA–protein binding sites data. Specifically, motivated by the imbalanced problem of the distribution of DNA–protein binding and non-binding sites, we employ the Adaptive Synthetic Sampling (ADASYN) approach to over-sample the positive data and Bootstrap strategy to under-sample the negative data to balance the number of the binding and non-binding samples. Furthermore, we employ the three types of features: the position specific scoring matrix, one-hot encoding and predicted solvent accessibility, to encode the sequence-based feature of each protein residue. In addition, we design an ensemble convolutional neural network classifier to handle the imbalance problem between binding and non-binding sites in protein sequence. Extensive experiments were conducted on the real DNA–protein binding sites dataset, PDNA-543, PDNA-224 and PDNA-316, in order to validate the effectiveness of our method on predicting the binding sites by ten-fold cross-validation metric. The experimental results demonstrate that our method achieves a high prediction performance and outperforms the state-of-the-art sequence-based DNA–protein binding sites predictors in terms of the Sensitivity, Specificity, Accuracy, Precision and Mathew’s Correlation Coefficient ( M C C ). Our method can obtain the M C C values of 0.63, 0.48 and 0.67 on PDNA-543, PDNA-224 and PDNA-316 datasets, respectively. Compared with the state-of-the art prediction models, the M C C values for our method are increased by at least 0.24, 0.13 and 0.23 on PDNA-543, PDNA-224 and PDNA-316 datasets, respectively.