Fig. 1

Framework of CMIC. Given a DNA sequence s, multiple variable-length k-mer sequences of s are generated first. The input layer of the neural network takes a variable-length k-mer sequence of as the input. The second layer is an embedding layer in which each k-mer of the inputted sequence is in order mapped to the corresponding embedding vector. The initial embedding vectors in the embedding layer are created using our new embedding vector generation method called splitDNA2vec. The third layer is a BiGRU layer comprising forward and backward directional GRU neural networks. The two outputs of the previous layer are concatenated into a single vector, and inputted into the next fully connected layer followed by a sigmoid activation function. The output is interpreted as an estimation of the probability that S is unmethylated