Figure 3.

The overall framework of TransPhos. The original sequence is converted into a set of feature vectors with different window sizes through an embedding layer. Here, we set 2 different window sizes: 51 and 33. The sequence features are further represented by the encoder, and then the high-dimensional features are extracted through several densely connected convolutional neural networks (DC-CNN) blocks. After the activation function, the representations obtained by several DC-CNN blocks are concatenated by intra-block connectivity layer (Inter-BCL) and converted to a one-dimensional tensor by a flatten layer. After a full connection (FC) layer, the phosphorylation prediction is finally generated by the SoftMax function.