Figure 1.

Illustration of the CNN-concat-filters model architecture for base-level pre-miRNAs classification. In CNN-concat-filters model, four kinds of filters, each of which has 32 filters, with the same width and different lengths (3, 4, 5 and 6) are employed (Conv3-32, Conv4-32, Conv5-32 and Conv6-32). In the convolution layer, each filter performs convolution on the sequence matrix and generates a feature map. The max-pooling operation then takes the largest number of each feature map. All the features are concatenated to form a 128-long feature vector for the penultimate fully-connected layer. The final layer is the softmax output which gives the probability of each classification. The shapes of the tensors as indicated in parentheses are given by height × width × channels.