tyonas9 / pdbp-fusion

Prediction of DNA binding proteins using local features and long-term dependencies with primary sequences based on deep learning

Guobin Li 1, Xiuquan Du 2, Xinlu Li 1, Le Zou 1, Guanhong Zhang 1 and Zhize Wu 1

1 School of Artificial Intelligence and Big Data, Hefei University, Hefei 230601, China 2 School of Computer Science and Technology, Anhui University, Hefei 230601, China

Abstract

DNA-binding proteins (DBPs) play pivotal roles in many biological functions such as alternative splicing, RNA editing, and methylation. Many researchers have proposed traditional computational methods to predict DBPs. However, these methods either rely on manual feature extraction or fail to capture long-term dependencies in the amino acid sequence. This paper proposes the PDBP-Fusion method based on deep learning (DL) to identify DBPs only from primary sequences. The proposed approach predicts DBPs based on the fusion of local characteristics and long-term dependencies. A convolutional neural network (CNN) is used to complete local feature learning, and a bi-directional long-short term memory network (Bi-LSTM) is used to capture critical long-term dependencies in context. Furthermore, feature extraction, model training, and model prediction are achieved simultaneously. The proposed PDBP-Fusion approach can predict DBPs with 86.90% sensitivity, 78.20% specificity, 82.57% accuracy, and 0.656 MCC on the PDB14189 benchmark dataset. The MCC of the authors' proposed methods was increased by at least 8% compared with other advanced prediction models in the literature. Moreover, the proposed method outperforms most existing classifiers on the PDB2272 independent dataset. The proposed PDBP-Fusion approach can be used to predict DBPs from sequences effectively; the online server is at http://119.45.144.26:8080/PDBP-Fusion/.

Dataset: PDB14189 （PDB14189_N.txt+PDB14189_P.txt） A dataset obtained from Ma et al. [18] referred to as PDB14189 was used as the benchmark dataset. The PDB14189 dataset consists of 7129 DBPs (positive samples) and 7060 non-DBPs (negative samples). They all came from the UniProt database [46]. The dataset is the same as the MsDBP [45].

In addition, an additional independent test dataset, PDB2272, was used to compare the performance of the proposed model with other existing prediction methods [19][20][21][45]. (1) to obtain the original dataset, consisting of 1153 DBPs and 1153 non-DBPs, was obtained from Swiss-Prot. (2) Sequences that had more than 25% similarity were removed, and sequences with irregular characters ("X" or "Z") were filtered out. Finally, the PDB2272 dataset contained 1153 DBPs and 1119 non-DBPs.

Sourecode usage: