Fig. 2

The evolutionary information model uses 2D Conv and CapsNet layers. First, the PSSM matrix is processed as the matrix with the same shape. These matrixes are fed to the 2D Conv layer (filters: 256, kernel size: (5,5), strides: 1, padding: valid, activation: relu) and average pooling layer (pool size: (2,2) and the remainder as default settings) to extract the rich and effective features. Next, the 2D Conv layer(filters: 128, kernel__size: (5,5), strides: 1, padding: ’valid’, activation: ’relu’) and average pooling layer(pool__size: (2,2)) are reused to extract more advanced features. The PrimaryCaps layer is the convolutional capsule layer, which has size 1 ×1 convolution kernels and 4 channels of 16D capsules. The ProteinCaps layer has eight 16D capsules to represent one of membrane protein types. Finally, the L2-norm of each capsule vector is calculated to indicate the probability of each type