TS-m6A-DL: Tissue-specific identification of N6-methyladenosine sites using a universal deep learning model

Abstract

The most communal post-transcriptional modification, N6-methyladenosine (m6A), is associated with a number of crucial biological processes. The precise detection of m6A sites around the genome is critical for revealing its regulatory function and providing new insights into drug design. Although both experimental and computational models for detecting m6A sites have been introduced, but these conventional methods are laborious and expensive. Furthermore, only a handful of these models are capable of detecting m6A sites in various tissues. Therefore, a more generic and optimized computational method for detecting m6A sites in different tissues is required. In this paper, we proposed a universal model using a deep neural network (DNN) and named it TS-m6A-DL, which can classify m6A sites in several tissues of humans (Homo sapiens), mice (Mus musculus), and rats (Rattus norvegicus). To extract RNA sequence features and to convert the input into numerical format for the network, we utilized one-hot-encoding method. The model was tested using fivefold cross-validation and its stability was measured using independent datasets. The proposed model, TS-m6A-DL, achieved accuracies in the range of 75-85% using the fivefold cross-validation method and 72-84% on the independent datasets. Finally, to authenticate the generalization of the model, we performed cross-species testing and proved the generalization ability by achieving state-of-the-art results.

Keywords: Binary-encoding; Deep neural network; Motif; N6-methyladenosine (m6A); Tissue-specific.

Graphical abstract

Fig. 1

Architecture of the proposed network, TS-m6A-DL. First step is to encode the sequence data to make it readable for the network. Then the encoded data is fed to the convolution layer; Conv1D(f,k,s), where f is the number of filters, k is the kernel size, and s is the stride. The convolution layer is followed by the pooling layer MaxPool(2), where 2 is the pooling size; which is then followed by the dropout layer Dropout(0.6), where 0.6 is the dropout rate.

Fig. 2

Comparison of the models; iRNA-m6A, m6A-TS-CNN, and TS-m6A-DNN in term of accuracy.

Fig. 3

Cross-species testing. Here, h,m, and r on x and y-axis represents the species human, mouse, and rat.

Fig. 4

The ACA-motif detected by our model, TS-m6A-DL.