DBP-CNN: Deep learning-based prediction of DNA-binding proteins by coupling discrete cosine transform with two-dimensional convolutional neural network

Abstract

To improve the prediction of DNA-binding Proteins (DBPs), this paper presents a deep learning-based method, named DBP-CNN. To efficiently extract the important features, we design a novel feature descriptor namely position-specific scoring matrix-tetra slices-discrete cosine transform (PSSM-TS-DCT). PSSM-TS-DCT explores the local features using tetra-slices notion with PSSM and captures decisive information by a compression scheme called DCT. The conventional feature descriptors such as DDE (dipeptide deviation from expected mean) and BiPSSM (bigram position-specific scoring matrix) are also used for feature extraction. The feature vectors of these feature descriptors are provided to RF (random forest), ERT (extremely randomized trees), XGB (eXtreme gradient boosting), and 2D CNN (two-dimensional convolutional neural network) classifiers. Our proposed feature descriptor (PSSM-TS-DCT) performs better than DDE and BiPSSM on all four classifiers. Similarly, among all classifiers, 2D CNN with PSSM-TS-DCT attains 2.80% and 0.92% higher accuracies than the recent predictor on both training and independent datasets, respectively. The experimental results show that our novel method (DBP-CNN) can predict DBPs more accurately than existing predictors in the literature.