Fig 7. Overall schema for extracting the local patterns from the whole protein sequence.

First, we transformed the protein sequence to an embedding vector with a fixed size, and the margins were padded, which are marked as $ and the corresponding embedding vectors. Second, we executed convolution along the sequence. Third, for each filter of window size, we pooled the max value. By concatenating all of the max-pooling values, we built a protein feature vector whose dimension is multiplying the number (#) of filters by the number (#) of windows.