Skip to main content
Fig. 7 | BMC Bioinformatics

Fig. 7

From: Predicting environmentally responsive transgenerational differential DNA methylated regions (epimutations) in the genome using a hybrid deep-machine learning approach

Fig. 7

Simplified diagram of the hybrid model. A 1000 bp input DNA sequence is one-hot encoded using a 5 × 1000 binary matrix. A convolution layer transforms the input matrix into an output matrix, where each output represents a sequence motif. After each convolutional layer is a batch-normalization layer following by a ReLU transformer layer. The max-pooling down-samples the output matrix. This block is followed by another similar block consisting of a convolutional layer, following by batch-normalization, ReLU, and a max-pooling layer. The classification block begins by flattening the output matrix of the previous layer, followed by two fully connected dense layers with 256 and 128 nodes, and the last layer consists of two nodes, one for each label: DMR and non-DMR. After training the network, the output of the first convolution layer represents new features used to re-express the input sequence. The re-expressed training data can then be used as input to a traditional ML classifier

Back to article page