A deep learning-based computational approach for discrimination of DNA N6-methyladenosine sites by fusing heterogeneous features

Abstract N6-methyladenine is post-replication modifications, which take place in the extensive range of DNA sequences and involved with a large number of different bioprocesses such as DNA repair, replication, cellular defense, and transcription in prokaryotes. Recently, various computational models were established to predict N6-methyladenine sites within DNAs. However, one of the main issues in the precise prediction of N6-methyladenine is the extraction of those features, which clearly define the characteristics of N6-methyladenine sites. In this method, input sequences of DNA are expressed by one-hot representation in order to allow progressive convolution layers. To exhibit the hidden information from the recognized sequences, the convolution neural network (CNN) model is applied to automatically learn the abstract features. Then, we apply the tri-nucleotide Composition (TNC) feature extraction technique and concatenate with CNN features. Our proposed model achieved 98.05% accuracy for the S1 benchmark dataset and 89.22% accuracy for the S2 benchmark dataset. The classification rates demonstrated that the developed approach performed better compared to existing approaches in terms of all the evaluation measures. It is expected that the developed intelligent approach might be played a leading and progressive role for academia as well as industrial research in the area of genomics prediction. The code cv is attached here.