DRBPPred-GAT: Accurate prediction of DNA-binding proteins and RNA-binding proteins based on graph multi-head attention network

Abstract

DNA-binding proteins (DBPs) and RNA-binding proteins (RBPs) play a crucial role in regulating various cellular functions that are closely associated with human disease. However, traditional experimental methods for predicting these proteins are often inefficient and limited by experimental conditions. Recently, deep learning methods have emerged as a promising research frontier for predicting DBPs and RBPs. In this paper, we have developed a novel model, called DRBPPred-GAT, for predicting DBPs and RBPs based on a graph multi-head attention network. The model consists of three main steps: (i) Protein information is fully captured by fusing eight different types of features. (ii) Autoencoder (AE) is used to remove irrelevant features from the multi-information fusion stage. (iii) Finally, a graph convolutional neural network with multi-head attention is used to predict DBPs and RBPs. Under 10-fold cross-validation, DRBPPred-GAT outperformed other benchmarked methods in terms of accuracy (ACC) and area under the receiver operating characteristic curve (AUC). On the training dataset, our model achieved ACC values of 84.32% for DBPs and 83.60% for RBPs prediction, with AUC values of 0.9219 and 0.9040 respectively. On the PDB186 dataset, the model achieved an ACC value of 76.97% with an AUC of 0.8063. The model also demonstrated high prediction accuracy on Human, S. cerevisiae, and A. thaliana, with ACC values of 76.45%, 79.96%, and 77.02% respectively. In summary, our results suggest that DRBPPred-GAT is highly accurate and effective for predicting DBPs and RBPs.